xref: /freebsd/contrib/llvm-project/llvm/lib/Target/PowerPC/PPCISelLowering.cpp (revision 580d00f42fdd94ce43583cc45fe3f1d9fdff47d4) 
1 //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the PPCISelLowering class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "PPCISelLowering.h"
14 #include "MCTargetDesc/PPCPredicates.h"
15 #include "PPC.h"
16 #include "PPCCCState.h"
17 #include "PPCCallingConv.h"
18 #include "PPCFrameLowering.h"
19 #include "PPCInstrInfo.h"
20 #include "PPCMachineFunctionInfo.h"
21 #include "PPCPerfectShuffle.h"
22 #include "PPCRegisterInfo.h"
23 #include "PPCSubtarget.h"
24 #include "PPCTargetMachine.h"
25 #include "llvm/ADT/APFloat.h"
26 #include "llvm/ADT/APInt.h"
27 #include "llvm/ADT/ArrayRef.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/ADT/StringRef.h"
35 #include "llvm/ADT/StringSwitch.h"
36 #include "llvm/CodeGen/CallingConvLower.h"
37 #include "llvm/CodeGen/ISDOpcodes.h"
38 #include "llvm/CodeGen/MachineBasicBlock.h"
39 #include "llvm/CodeGen/MachineFrameInfo.h"
40 #include "llvm/CodeGen/MachineFunction.h"
41 #include "llvm/CodeGen/MachineInstr.h"
42 #include "llvm/CodeGen/MachineInstrBuilder.h"
43 #include "llvm/CodeGen/MachineJumpTableInfo.h"
44 #include "llvm/CodeGen/MachineLoopInfo.h"
45 #include "llvm/CodeGen/MachineMemOperand.h"
46 #include "llvm/CodeGen/MachineModuleInfo.h"
47 #include "llvm/CodeGen/MachineOperand.h"
48 #include "llvm/CodeGen/MachineRegisterInfo.h"
49 #include "llvm/CodeGen/RuntimeLibcalls.h"
50 #include "llvm/CodeGen/SelectionDAG.h"
51 #include "llvm/CodeGen/SelectionDAGNodes.h"
52 #include "llvm/CodeGen/TargetInstrInfo.h"
53 #include "llvm/CodeGen/TargetLowering.h"
54 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
55 #include "llvm/CodeGen/TargetRegisterInfo.h"
56 #include "llvm/CodeGen/ValueTypes.h"
57 #include "llvm/IR/CallingConv.h"
58 #include "llvm/IR/Constant.h"
59 #include "llvm/IR/Constants.h"
60 #include "llvm/IR/DataLayout.h"
61 #include "llvm/IR/DebugLoc.h"
62 #include "llvm/IR/DerivedTypes.h"
63 #include "llvm/IR/Function.h"
64 #include "llvm/IR/GlobalValue.h"
65 #include "llvm/IR/IRBuilder.h"
66 #include "llvm/IR/Instructions.h"
67 #include "llvm/IR/Intrinsics.h"
68 #include "llvm/IR/IntrinsicsPowerPC.h"
69 #include "llvm/IR/Module.h"
70 #include "llvm/IR/Type.h"
71 #include "llvm/IR/Use.h"
72 #include "llvm/IR/Value.h"
73 #include "llvm/MC/MCContext.h"
74 #include "llvm/MC/MCExpr.h"
75 #include "llvm/MC/MCRegisterInfo.h"
76 #include "llvm/MC/MCSectionXCOFF.h"
77 #include "llvm/MC/MCSymbolXCOFF.h"
78 #include "llvm/Support/AtomicOrdering.h"
79 #include "llvm/Support/BranchProbability.h"
80 #include "llvm/Support/Casting.h"
81 #include "llvm/Support/CodeGen.h"
82 #include "llvm/Support/CommandLine.h"
83 #include "llvm/Support/Compiler.h"
84 #include "llvm/Support/Debug.h"
85 #include "llvm/Support/ErrorHandling.h"
86 #include "llvm/Support/Format.h"
87 #include "llvm/Support/KnownBits.h"
88 #include "llvm/Support/MachineValueType.h"
89 #include "llvm/Support/MathExtras.h"
90 #include "llvm/Support/raw_ostream.h"
91 #include "llvm/Target/TargetMachine.h"
92 #include "llvm/Target/TargetOptions.h"
93 #include <algorithm>
94 #include <cassert>
95 #include <cstdint>
96 #include <iterator>
97 #include <list>
98 #include <optional>
99 #include <utility>
100 #include <vector>
101 
102 using namespace llvm;
103 
104 #define DEBUG_TYPE "ppc-lowering"
105 
106 static cl::opt<bool> DisablePPCPreinc("disable-ppc-preinc",
107 cl::desc("disable preincrement load/store generation on PPC"), cl::Hidden);
108 
109 static cl::opt<bool> DisableILPPref("disable-ppc-ilp-pref",
110 cl::desc("disable setting the node scheduling preference to ILP on PPC"), cl::Hidden);
111 
112 static cl::opt<bool> DisablePPCUnaligned("disable-ppc-unaligned",
113 cl::desc("disable unaligned load/store generation on PPC"), cl::Hidden);
114 
115 static cl::opt<bool> DisableSCO("disable-ppc-sco",
116 cl::desc("disable sibling call optimization on ppc"), cl::Hidden);
117 
118 static cl::opt<bool> DisableInnermostLoopAlign32("disable-ppc-innermost-loop-align32",
119 cl::desc("don't always align innermost loop to 32 bytes on ppc"), cl::Hidden);
120 
121 static cl::opt<bool> UseAbsoluteJumpTables("ppc-use-absolute-jumptables",
122 cl::desc("use absolute jump tables on ppc"), cl::Hidden);
123 
124 static cl::opt<bool> EnableQuadwordAtomics(
125     "ppc-quadword-atomics",
126     cl::desc("enable quadword lock-free atomic operations"), cl::init(false),
127     cl::Hidden);
128 
129 static cl::opt<bool>
130     DisablePerfectShuffle("ppc-disable-perfect-shuffle",
131                           cl::desc("disable vector permute decomposition"),
132                           cl::init(true), cl::Hidden);
133 
134 cl::opt<bool> DisableAutoPairedVecSt(
135     "disable-auto-paired-vec-st",
136     cl::desc("disable automatically generated 32byte paired vector stores"),
137     cl::init(true), cl::Hidden);
138 
139 STATISTIC(NumTailCalls, "Number of tail calls");
140 STATISTIC(NumSiblingCalls, "Number of sibling calls");
141 STATISTIC(ShufflesHandledWithVPERM,
142           "Number of shuffles lowered to a VPERM or XXPERM");
143 STATISTIC(NumDynamicAllocaProbed, "Number of dynamic stack allocation probed");
144 
145 static bool isNByteElemShuffleMask(ShuffleVectorSDNode *, unsigned, int);
146 
147 static SDValue widenVec(SelectionDAG &DAG, SDValue Vec, const SDLoc &dl);
148 
149 static const char AIXSSPCanaryWordName[] = "__ssp_canary_word";
150 
151 // FIXME: Remove this once the bug has been fixed!
152 extern cl::opt<bool> ANDIGlueBug;
153 
154 PPCTargetLowering::PPCTargetLowering(const PPCTargetMachine &TM,
155                                      const PPCSubtarget &STI)
156     : TargetLowering(TM), Subtarget(STI) {
157   // Initialize map that relates the PPC addressing modes to the computed flags
158   // of a load/store instruction. The map is used to determine the optimal
159   // addressing mode when selecting load and stores.
160   initializeAddrModeMap();
161   // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
162   // arguments are at least 4/8 bytes aligned.
163   bool isPPC64 = Subtarget.isPPC64();
164   setMinStackArgumentAlignment(isPPC64 ? Align(8) : Align(4));
165 
166   // Set up the register classes.
167   addRegisterClass(MVT::i32, &PPC::GPRCRegClass);
168   if (!useSoftFloat()) {
169     if (hasSPE()) {
170       addRegisterClass(MVT::f32, &PPC::GPRCRegClass);
171       // EFPU2 APU only supports f32
172       if (!Subtarget.hasEFPU2())
173         addRegisterClass(MVT::f64, &PPC::SPERCRegClass);
174     } else {
175       addRegisterClass(MVT::f32, &PPC::F4RCRegClass);
176       addRegisterClass(MVT::f64, &PPC::F8RCRegClass);
177     }
178   }
179 
180   // Match BITREVERSE to customized fast code sequence in the td file.
181   setOperationAction(ISD::BITREVERSE, MVT::i32, Legal);
182   setOperationAction(ISD::BITREVERSE, MVT::i64, Legal);
183 
184   // Sub-word ATOMIC_CMP_SWAP need to ensure that the input is zero-extended.
185   setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
186 
187   // Custom lower inline assembly to check for special registers.
188   setOperationAction(ISD::INLINEASM, MVT::Other, Custom);
189   setOperationAction(ISD::INLINEASM_BR, MVT::Other, Custom);
190 
191   // PowerPC has an i16 but no i8 (or i1) SEXTLOAD.
192   for (MVT VT : MVT::integer_valuetypes()) {
193     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
194     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Expand);
195   }
196 
197   if (Subtarget.isISA3_0()) {
198     setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Legal);
199     setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Legal);
200     setTruncStoreAction(MVT::f64, MVT::f16, Legal);
201     setTruncStoreAction(MVT::f32, MVT::f16, Legal);
202   } else {
203     // No extending loads from f16 or HW conversions back and forth.
204     setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
205     setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
206     setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
207     setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
208     setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
209     setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
210     setTruncStoreAction(MVT::f64, MVT::f16, Expand);
211     setTruncStoreAction(MVT::f32, MVT::f16, Expand);
212   }
213 
214   setTruncStoreAction(MVT::f64, MVT::f32, Expand);
215 
216   // PowerPC has pre-inc load and store's.
217   setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
218   setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
219   setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
220   setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
221   setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
222   setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
223   setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
224   setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
225   setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
226   setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
227   if (!Subtarget.hasSPE()) {
228     setIndexedLoadAction(ISD::PRE_INC, MVT::f32, Legal);
229     setIndexedLoadAction(ISD::PRE_INC, MVT::f64, Legal);
230     setIndexedStoreAction(ISD::PRE_INC, MVT::f32, Legal);
231     setIndexedStoreAction(ISD::PRE_INC, MVT::f64, Legal);
232   }
233 
234   // PowerPC uses ADDC/ADDE/SUBC/SUBE to propagate carry.
235   const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
236   for (MVT VT : ScalarIntVTs) {
237     setOperationAction(ISD::ADDC, VT, Legal);
238     setOperationAction(ISD::ADDE, VT, Legal);
239     setOperationAction(ISD::SUBC, VT, Legal);
240     setOperationAction(ISD::SUBE, VT, Legal);
241   }
242 
243   if (Subtarget.useCRBits()) {
244     setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
245 
246     if (isPPC64 || Subtarget.hasFPCVT()) {
247       setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i1, Promote);
248       AddPromotedToType(ISD::STRICT_SINT_TO_FP, MVT::i1,
249                         isPPC64 ? MVT::i64 : MVT::i32);
250       setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i1, Promote);
251       AddPromotedToType(ISD::STRICT_UINT_TO_FP, MVT::i1,
252                         isPPC64 ? MVT::i64 : MVT::i32);
253 
254       setOperationAction(ISD::SINT_TO_FP, MVT::i1, Promote);
255       AddPromotedToType (ISD::SINT_TO_FP, MVT::i1,
256                          isPPC64 ? MVT::i64 : MVT::i32);
257       setOperationAction(ISD::UINT_TO_FP, MVT::i1, Promote);
258       AddPromotedToType(ISD::UINT_TO_FP, MVT::i1,
259                         isPPC64 ? MVT::i64 : MVT::i32);
260 
261       setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i1, Promote);
262       AddPromotedToType(ISD::STRICT_FP_TO_SINT, MVT::i1,
263                         isPPC64 ? MVT::i64 : MVT::i32);
264       setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i1, Promote);
265       AddPromotedToType(ISD::STRICT_FP_TO_UINT, MVT::i1,
266                         isPPC64 ? MVT::i64 : MVT::i32);
267 
268       setOperationAction(ISD::FP_TO_SINT, MVT::i1, Promote);
269       AddPromotedToType(ISD::FP_TO_SINT, MVT::i1,
270                         isPPC64 ? MVT::i64 : MVT::i32);
271       setOperationAction(ISD::FP_TO_UINT, MVT::i1, Promote);
272       AddPromotedToType(ISD::FP_TO_UINT, MVT::i1,
273                         isPPC64 ? MVT::i64 : MVT::i32);
274     } else {
275       setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i1, Custom);
276       setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i1, Custom);
277       setOperationAction(ISD::SINT_TO_FP, MVT::i1, Custom);
278       setOperationAction(ISD::UINT_TO_FP, MVT::i1, Custom);
279     }
280 
281     // PowerPC does not support direct load/store of condition registers.
282     setOperationAction(ISD::LOAD, MVT::i1, Custom);
283     setOperationAction(ISD::STORE, MVT::i1, Custom);
284 
285     // FIXME: Remove this once the ANDI glue bug is fixed:
286     if (ANDIGlueBug)
287       setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
288 
289     for (MVT VT : MVT::integer_valuetypes()) {
290       setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
291       setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
292       setTruncStoreAction(VT, MVT::i1, Expand);
293     }
294 
295     addRegisterClass(MVT::i1, &PPC::CRBITRCRegClass);
296   }
297 
298   // Expand ppcf128 to i32 by hand for the benefit of llvm-gcc bootstrap on
299   // PPC (the libcall is not available).
300   setOperationAction(ISD::FP_TO_SINT, MVT::ppcf128, Custom);
301   setOperationAction(ISD::FP_TO_UINT, MVT::ppcf128, Custom);
302   setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::ppcf128, Custom);
303   setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::ppcf128, Custom);
304 
305   // We do not currently implement these libm ops for PowerPC.
306   setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand);
307   setOperationAction(ISD::FCEIL,  MVT::ppcf128, Expand);
308   setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand);
309   setOperationAction(ISD::FRINT,  MVT::ppcf128, Expand);
310   setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand);
311   setOperationAction(ISD::FREM, MVT::ppcf128, Expand);
312 
313   // PowerPC has no SREM/UREM instructions unless we are on P9
314   // On P9 we may use a hardware instruction to compute the remainder.
315   // When the result of both the remainder and the division is required it is
316   // more efficient to compute the remainder from the result of the division
317   // rather than use the remainder instruction. The instructions are legalized
318   // directly because the DivRemPairsPass performs the transformation at the IR
319   // level.
320   if (Subtarget.isISA3_0()) {
321     setOperationAction(ISD::SREM, MVT::i32, Legal);
322     setOperationAction(ISD::UREM, MVT::i32, Legal);
323     setOperationAction(ISD::SREM, MVT::i64, Legal);
324     setOperationAction(ISD::UREM, MVT::i64, Legal);
325   } else {
326     setOperationAction(ISD::SREM, MVT::i32, Expand);
327     setOperationAction(ISD::UREM, MVT::i32, Expand);
328     setOperationAction(ISD::SREM, MVT::i64, Expand);
329     setOperationAction(ISD::UREM, MVT::i64, Expand);
330   }
331 
332   // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
333   setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
334   setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
335   setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
336   setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
337   setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
338   setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
339   setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
340   setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
341 
342   // Handle constrained floating-point operations of scalar.
343   // TODO: Handle SPE specific operation.
344   setOperationAction(ISD::STRICT_FADD, MVT::f32, Legal);
345   setOperationAction(ISD::STRICT_FSUB, MVT::f32, Legal);
346   setOperationAction(ISD::STRICT_FMUL, MVT::f32, Legal);
347   setOperationAction(ISD::STRICT_FDIV, MVT::f32, Legal);
348   setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal);
349 
350   setOperationAction(ISD::STRICT_FADD, MVT::f64, Legal);
351   setOperationAction(ISD::STRICT_FSUB, MVT::f64, Legal);
352   setOperationAction(ISD::STRICT_FMUL, MVT::f64, Legal);
353   setOperationAction(ISD::STRICT_FDIV, MVT::f64, Legal);
354 
355   if (!Subtarget.hasSPE()) {
356     setOperationAction(ISD::STRICT_FMA, MVT::f32, Legal);
357     setOperationAction(ISD::STRICT_FMA, MVT::f64, Legal);
358   }
359 
360   if (Subtarget.hasVSX()) {
361     setOperationAction(ISD::STRICT_FRINT, MVT::f32, Legal);
362     setOperationAction(ISD::STRICT_FRINT, MVT::f64, Legal);
363   }
364 
365   if (Subtarget.hasFSQRT()) {
366     setOperationAction(ISD::STRICT_FSQRT, MVT::f32, Legal);
367     setOperationAction(ISD::STRICT_FSQRT, MVT::f64, Legal);
368   }
369 
370   if (Subtarget.hasFPRND()) {
371     setOperationAction(ISD::STRICT_FFLOOR, MVT::f32, Legal);
372     setOperationAction(ISD::STRICT_FCEIL,  MVT::f32, Legal);
373     setOperationAction(ISD::STRICT_FTRUNC, MVT::f32, Legal);
374     setOperationAction(ISD::STRICT_FROUND, MVT::f32, Legal);
375 
376     setOperationAction(ISD::STRICT_FFLOOR, MVT::f64, Legal);
377     setOperationAction(ISD::STRICT_FCEIL,  MVT::f64, Legal);
378     setOperationAction(ISD::STRICT_FTRUNC, MVT::f64, Legal);
379     setOperationAction(ISD::STRICT_FROUND, MVT::f64, Legal);
380   }
381 
382   // We don't support sin/cos/sqrt/fmod/pow
383   setOperationAction(ISD::FSIN , MVT::f64, Expand);
384   setOperationAction(ISD::FCOS , MVT::f64, Expand);
385   setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
386   setOperationAction(ISD::FREM , MVT::f64, Expand);
387   setOperationAction(ISD::FPOW , MVT::f64, Expand);
388   setOperationAction(ISD::FSIN , MVT::f32, Expand);
389   setOperationAction(ISD::FCOS , MVT::f32, Expand);
390   setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
391   setOperationAction(ISD::FREM , MVT::f32, Expand);
392   setOperationAction(ISD::FPOW , MVT::f32, Expand);
393 
394   // MASS transformation for LLVM intrinsics with replicating fast-math flag
395   // to be consistent to PPCGenScalarMASSEntries pass
396   if (TM.getOptLevel() == CodeGenOpt::Aggressive) {
397     setOperationAction(ISD::FSIN , MVT::f64, Custom);
398     setOperationAction(ISD::FCOS , MVT::f64, Custom);
399     setOperationAction(ISD::FPOW , MVT::f64, Custom);
400     setOperationAction(ISD::FLOG, MVT::f64, Custom);
401     setOperationAction(ISD::FLOG10, MVT::f64, Custom);
402     setOperationAction(ISD::FEXP, MVT::f64, Custom);
403     setOperationAction(ISD::FSIN , MVT::f32, Custom);
404     setOperationAction(ISD::FCOS , MVT::f32, Custom);
405     setOperationAction(ISD::FPOW , MVT::f32, Custom);
406     setOperationAction(ISD::FLOG, MVT::f32, Custom);
407     setOperationAction(ISD::FLOG10, MVT::f32, Custom);
408     setOperationAction(ISD::FEXP, MVT::f32, Custom);
409   }
410 
411   if (Subtarget.hasSPE()) {
412     setOperationAction(ISD::FMA  , MVT::f64, Expand);
413     setOperationAction(ISD::FMA  , MVT::f32, Expand);
414   } else {
415     setOperationAction(ISD::FMA  , MVT::f64, Legal);
416     setOperationAction(ISD::FMA  , MVT::f32, Legal);
417   }
418 
419   if (Subtarget.hasSPE())
420     setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
421 
422   setOperationAction(ISD::GET_ROUNDING, MVT::i32, Custom);
423 
424   // If we're enabling GP optimizations, use hardware square root
425   if (!Subtarget.hasFSQRT() &&
426       !(TM.Options.UnsafeFPMath && Subtarget.hasFRSQRTE() &&
427         Subtarget.hasFRE()))
428     setOperationAction(ISD::FSQRT, MVT::f64, Expand);
429 
430   if (!Subtarget.hasFSQRT() &&
431       !(TM.Options.UnsafeFPMath && Subtarget.hasFRSQRTES() &&
432         Subtarget.hasFRES()))
433     setOperationAction(ISD::FSQRT, MVT::f32, Expand);
434 
435   if (Subtarget.hasFCPSGN()) {
436     setOperationAction(ISD::FCOPYSIGN, MVT::f64, Legal);
437     setOperationAction(ISD::FCOPYSIGN, MVT::f32, Legal);
438   } else {
439     setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
440     setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
441   }
442 
443   if (Subtarget.hasFPRND()) {
444     setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
445     setOperationAction(ISD::FCEIL,  MVT::f64, Legal);
446     setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
447     setOperationAction(ISD::FROUND, MVT::f64, Legal);
448 
449     setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
450     setOperationAction(ISD::FCEIL,  MVT::f32, Legal);
451     setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
452     setOperationAction(ISD::FROUND, MVT::f32, Legal);
453   }
454 
455   // Prior to P10, PowerPC does not have BSWAP, but we can use vector BSWAP
456   // instruction xxbrd to speed up scalar BSWAP64.
457   if (Subtarget.isISA3_1()) {
458     setOperationAction(ISD::BSWAP, MVT::i32, Legal);
459     setOperationAction(ISD::BSWAP, MVT::i64, Legal);
460   } else {
461     setOperationAction(ISD::BSWAP, MVT::i32, Expand);
462     setOperationAction(
463         ISD::BSWAP, MVT::i64,
464         (Subtarget.hasP9Vector() && Subtarget.isPPC64()) ? Custom : Expand);
465   }
466 
467   // CTPOP or CTTZ were introduced in P8/P9 respectively
468   if (Subtarget.isISA3_0()) {
469     setOperationAction(ISD::CTTZ , MVT::i32  , Legal);
470     setOperationAction(ISD::CTTZ , MVT::i64  , Legal);
471   } else {
472     setOperationAction(ISD::CTTZ , MVT::i32  , Expand);
473     setOperationAction(ISD::CTTZ , MVT::i64  , Expand);
474   }
475 
476   if (Subtarget.hasPOPCNTD() == PPCSubtarget::POPCNTD_Fast) {
477     setOperationAction(ISD::CTPOP, MVT::i32  , Legal);
478     setOperationAction(ISD::CTPOP, MVT::i64  , Legal);
479   } else {
480     setOperationAction(ISD::CTPOP, MVT::i32  , Expand);
481     setOperationAction(ISD::CTPOP, MVT::i64  , Expand);
482   }
483 
484   // PowerPC does not have ROTR
485   setOperationAction(ISD::ROTR, MVT::i32   , Expand);
486   setOperationAction(ISD::ROTR, MVT::i64   , Expand);
487 
488   if (!Subtarget.useCRBits()) {
489     // PowerPC does not have Select
490     setOperationAction(ISD::SELECT, MVT::i32, Expand);
491     setOperationAction(ISD::SELECT, MVT::i64, Expand);
492     setOperationAction(ISD::SELECT, MVT::f32, Expand);
493     setOperationAction(ISD::SELECT, MVT::f64, Expand);
494   }
495 
496   // PowerPC wants to turn select_cc of FP into fsel when possible.
497   setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
498   setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
499 
500   // PowerPC wants to optimize integer setcc a bit
501   if (!Subtarget.useCRBits())
502     setOperationAction(ISD::SETCC, MVT::i32, Custom);
503 
504   if (Subtarget.hasFPU()) {
505     setOperationAction(ISD::STRICT_FSETCC, MVT::f32, Legal);
506     setOperationAction(ISD::STRICT_FSETCC, MVT::f64, Legal);
507     setOperationAction(ISD::STRICT_FSETCC, MVT::f128, Legal);
508 
509     setOperationAction(ISD::STRICT_FSETCCS, MVT::f32, Legal);
510     setOperationAction(ISD::STRICT_FSETCCS, MVT::f64, Legal);
511     setOperationAction(ISD::STRICT_FSETCCS, MVT::f128, Legal);
512   }
513 
514   // PowerPC does not have BRCOND which requires SetCC
515   if (!Subtarget.useCRBits())
516     setOperationAction(ISD::BRCOND, MVT::Other, Expand);
517 
518   setOperationAction(ISD::BR_JT,  MVT::Other, Expand);
519 
520   if (Subtarget.hasSPE()) {
521     // SPE has built-in conversions
522     setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Legal);
523     setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Legal);
524     setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Legal);
525     setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
526     setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
527     setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
528 
529     // SPE supports signaling compare of f32/f64.
530     setOperationAction(ISD::STRICT_FSETCCS, MVT::f32, Legal);
531     setOperationAction(ISD::STRICT_FSETCCS, MVT::f64, Legal);
532   } else {
533     // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
534     setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom);
535     setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
536 
537     // PowerPC does not have [U|S]INT_TO_FP
538     setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Expand);
539     setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Expand);
540     setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
541     setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
542   }
543 
544   if (Subtarget.hasDirectMove() && isPPC64) {
545     setOperationAction(ISD::BITCAST, MVT::f32, Legal);
546     setOperationAction(ISD::BITCAST, MVT::i32, Legal);
547     setOperationAction(ISD::BITCAST, MVT::i64, Legal);
548     setOperationAction(ISD::BITCAST, MVT::f64, Legal);
549     if (TM.Options.UnsafeFPMath) {
550       setOperationAction(ISD::LRINT, MVT::f64, Legal);
551       setOperationAction(ISD::LRINT, MVT::f32, Legal);
552       setOperationAction(ISD::LLRINT, MVT::f64, Legal);
553       setOperationAction(ISD::LLRINT, MVT::f32, Legal);
554       setOperationAction(ISD::LROUND, MVT::f64, Legal);
555       setOperationAction(ISD::LROUND, MVT::f32, Legal);
556       setOperationAction(ISD::LLROUND, MVT::f64, Legal);
557       setOperationAction(ISD::LLROUND, MVT::f32, Legal);
558     }
559   } else {
560     setOperationAction(ISD::BITCAST, MVT::f32, Expand);
561     setOperationAction(ISD::BITCAST, MVT::i32, Expand);
562     setOperationAction(ISD::BITCAST, MVT::i64, Expand);
563     setOperationAction(ISD::BITCAST, MVT::f64, Expand);
564   }
565 
566   // We cannot sextinreg(i1).  Expand to shifts.
567   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
568 
569   // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
570   // SjLj exception handling but a light-weight setjmp/longjmp replacement to
571   // support continuation, user-level threading, and etc.. As a result, no
572   // other SjLj exception interfaces are implemented and please don't build
573   // your own exception handling based on them.
574   // LLVM/Clang supports zero-cost DWARF exception handling.
575   setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
576   setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
577 
578   // We want to legalize GlobalAddress and ConstantPool nodes into the
579   // appropriate instructions to materialize the address.
580   setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
581   setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
582   setOperationAction(ISD::BlockAddress,  MVT::i32, Custom);
583   setOperationAction(ISD::ConstantPool,  MVT::i32, Custom);
584   setOperationAction(ISD::JumpTable,     MVT::i32, Custom);
585   setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
586   setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
587   setOperationAction(ISD::BlockAddress,  MVT::i64, Custom);
588   setOperationAction(ISD::ConstantPool,  MVT::i64, Custom);
589   setOperationAction(ISD::JumpTable,     MVT::i64, Custom);
590 
591   // TRAP is legal.
592   setOperationAction(ISD::TRAP, MVT::Other, Legal);
593 
594   // TRAMPOLINE is custom lowered.
595   setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
596   setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
597 
598   // VASTART needs to be custom lowered to use the VarArgsFrameIndex
599   setOperationAction(ISD::VASTART           , MVT::Other, Custom);
600 
601   if (Subtarget.is64BitELFABI()) {
602     // VAARG always uses double-word chunks, so promote anything smaller.
603     setOperationAction(ISD::VAARG, MVT::i1, Promote);
604     AddPromotedToType(ISD::VAARG, MVT::i1, MVT::i64);
605     setOperationAction(ISD::VAARG, MVT::i8, Promote);
606     AddPromotedToType(ISD::VAARG, MVT::i8, MVT::i64);
607     setOperationAction(ISD::VAARG, MVT::i16, Promote);
608     AddPromotedToType(ISD::VAARG, MVT::i16, MVT::i64);
609     setOperationAction(ISD::VAARG, MVT::i32, Promote);
610     AddPromotedToType(ISD::VAARG, MVT::i32, MVT::i64);
611     setOperationAction(ISD::VAARG, MVT::Other, Expand);
612   } else if (Subtarget.is32BitELFABI()) {
613     // VAARG is custom lowered with the 32-bit SVR4 ABI.
614     setOperationAction(ISD::VAARG, MVT::Other, Custom);
615     setOperationAction(ISD::VAARG, MVT::i64, Custom);
616   } else
617     setOperationAction(ISD::VAARG, MVT::Other, Expand);
618 
619   // VACOPY is custom lowered with the 32-bit SVR4 ABI.
620   if (Subtarget.is32BitELFABI())
621     setOperationAction(ISD::VACOPY            , MVT::Other, Custom);
622   else
623     setOperationAction(ISD::VACOPY            , MVT::Other, Expand);
624 
625   // Use the default implementation.
626   setOperationAction(ISD::VAEND             , MVT::Other, Expand);
627   setOperationAction(ISD::STACKSAVE         , MVT::Other, Expand);
628   setOperationAction(ISD::STACKRESTORE      , MVT::Other, Custom);
629   setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32  , Custom);
630   setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64  , Custom);
631   setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, MVT::i32, Custom);
632   setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, MVT::i64, Custom);
633   setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom);
634   setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom);
635 
636   // We want to custom lower some of our intrinsics.
637   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
638   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f64, Custom);
639   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::ppcf128, Custom);
640   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
641   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v2f64, Custom);
642 
643   // To handle counter-based loop conditions.
644   setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i1, Custom);
645 
646   setOperationAction(ISD::INTRINSIC_VOID, MVT::i8, Custom);
647   setOperationAction(ISD::INTRINSIC_VOID, MVT::i16, Custom);
648   setOperationAction(ISD::INTRINSIC_VOID, MVT::i32, Custom);
649   setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
650 
651   // Comparisons that require checking two conditions.
652   if (Subtarget.hasSPE()) {
653     setCondCodeAction(ISD::SETO, MVT::f32, Expand);
654     setCondCodeAction(ISD::SETO, MVT::f64, Expand);
655     setCondCodeAction(ISD::SETUO, MVT::f32, Expand);
656     setCondCodeAction(ISD::SETUO, MVT::f64, Expand);
657   }
658   setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
659   setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
660   setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
661   setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
662   setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
663   setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
664   setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
665   setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
666   setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
667   setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
668   setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
669   setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
670 
671   setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f32, Legal);
672   setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f64, Legal);
673 
674   if (Subtarget.has64BitSupport()) {
675     // They also have instructions for converting between i64 and fp.
676     setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i64, Custom);
677     setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i64, Expand);
678     setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i64, Custom);
679     setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i64, Expand);
680     setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
681     setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
682     setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
683     setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
684     // This is just the low 32 bits of a (signed) fp->i64 conversion.
685     // We cannot do this with Promote because i64 is not a legal type.
686     setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom);
687     setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
688 
689     if (Subtarget.hasLFIWAX() || Subtarget.isPPC64()) {
690       setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
691       setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Custom);
692     }
693   } else {
694     // PowerPC does not have FP_TO_UINT on 32-bit implementations.
695     if (Subtarget.hasSPE()) {
696       setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Legal);
697       setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
698     } else {
699       setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Expand);
700       setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
701     }
702   }
703 
704   // With the instructions enabled under FPCVT, we can do everything.
705   if (Subtarget.hasFPCVT()) {
706     if (Subtarget.has64BitSupport()) {
707       setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i64, Custom);
708       setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i64, Custom);
709       setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i64, Custom);
710       setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i64, Custom);
711       setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
712       setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
713       setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
714       setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
715     }
716 
717     setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom);
718     setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom);
719     setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Custom);
720     setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Custom);
721     setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
722     setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
723     setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
724     setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
725   }
726 
727   if (Subtarget.use64BitRegs()) {
728     // 64-bit PowerPC implementations can support i64 types directly
729     addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
730     // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
731     setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
732     // 64-bit PowerPC wants to expand i128 shifts itself.
733     setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
734     setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
735     setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
736   } else {
737     // 32-bit PowerPC wants to expand i64 shifts itself.
738     setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
739     setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
740     setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
741   }
742 
743   // PowerPC has better expansions for funnel shifts than the generic
744   // TargetLowering::expandFunnelShift.
745   if (Subtarget.has64BitSupport()) {
746     setOperationAction(ISD::FSHL, MVT::i64, Custom);
747     setOperationAction(ISD::FSHR, MVT::i64, Custom);
748   }
749   setOperationAction(ISD::FSHL, MVT::i32, Custom);
750   setOperationAction(ISD::FSHR, MVT::i32, Custom);
751 
752   if (Subtarget.hasVSX()) {
753     setOperationAction(ISD::FMAXNUM_IEEE, MVT::f64, Legal);
754     setOperationAction(ISD::FMAXNUM_IEEE, MVT::f32, Legal);
755     setOperationAction(ISD::FMINNUM_IEEE, MVT::f64, Legal);
756     setOperationAction(ISD::FMINNUM_IEEE, MVT::f32, Legal);
757   }
758 
759   if (Subtarget.hasAltivec()) {
760     for (MVT VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
761       setOperationAction(ISD::SADDSAT, VT, Legal);
762       setOperationAction(ISD::SSUBSAT, VT, Legal);
763       setOperationAction(ISD::UADDSAT, VT, Legal);
764       setOperationAction(ISD::USUBSAT, VT, Legal);
765     }
766     // First set operation action for all vector types to expand. Then we
767     // will selectively turn on ones that can be effectively codegen'd.
768     for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
769       // add/sub are legal for all supported vector VT's.
770       setOperationAction(ISD::ADD, VT, Legal);
771       setOperationAction(ISD::SUB, VT, Legal);
772 
773       // For v2i64, these are only valid with P8Vector. This is corrected after
774       // the loop.
775       if (VT.getSizeInBits() <= 128 && VT.getScalarSizeInBits() <= 64) {
776         setOperationAction(ISD::SMAX, VT, Legal);
777         setOperationAction(ISD::SMIN, VT, Legal);
778         setOperationAction(ISD::UMAX, VT, Legal);
779         setOperationAction(ISD::UMIN, VT, Legal);
780       }
781       else {
782         setOperationAction(ISD::SMAX, VT, Expand);
783         setOperationAction(ISD::SMIN, VT, Expand);
784         setOperationAction(ISD::UMAX, VT, Expand);
785         setOperationAction(ISD::UMIN, VT, Expand);
786       }
787 
788       if (Subtarget.hasVSX()) {
789         setOperationAction(ISD::FMAXNUM, VT, Legal);
790         setOperationAction(ISD::FMINNUM, VT, Legal);
791       }
792 
793       // Vector instructions introduced in P8
794       if (Subtarget.hasP8Altivec() && (VT.SimpleTy != MVT::v1i128)) {
795         setOperationAction(ISD::CTPOP, VT, Legal);
796         setOperationAction(ISD::CTLZ, VT, Legal);
797       }
798       else {
799         setOperationAction(ISD::CTPOP, VT, Expand);
800         setOperationAction(ISD::CTLZ, VT, Expand);
801       }
802 
803       // Vector instructions introduced in P9
804       if (Subtarget.hasP9Altivec() && (VT.SimpleTy != MVT::v1i128))
805         setOperationAction(ISD::CTTZ, VT, Legal);
806       else
807         setOperationAction(ISD::CTTZ, VT, Expand);
808 
809       // We promote all shuffles to v16i8.
810       setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
811       AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
812 
813       // We promote all non-typed operations to v4i32.
814       setOperationAction(ISD::AND   , VT, Promote);
815       AddPromotedToType (ISD::AND   , VT, MVT::v4i32);
816       setOperationAction(ISD::OR    , VT, Promote);
817       AddPromotedToType (ISD::OR    , VT, MVT::v4i32);
818       setOperationAction(ISD::XOR   , VT, Promote);
819       AddPromotedToType (ISD::XOR   , VT, MVT::v4i32);
820       setOperationAction(ISD::LOAD  , VT, Promote);
821       AddPromotedToType (ISD::LOAD  , VT, MVT::v4i32);
822       setOperationAction(ISD::SELECT, VT, Promote);
823       AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
824       setOperationAction(ISD::VSELECT, VT, Legal);
825       setOperationAction(ISD::SELECT_CC, VT, Promote);
826       AddPromotedToType (ISD::SELECT_CC, VT, MVT::v4i32);
827       setOperationAction(ISD::STORE, VT, Promote);
828       AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
829 
830       // No other operations are legal.
831       setOperationAction(ISD::MUL , VT, Expand);
832       setOperationAction(ISD::SDIV, VT, Expand);
833       setOperationAction(ISD::SREM, VT, Expand);
834       setOperationAction(ISD::UDIV, VT, Expand);
835       setOperationAction(ISD::UREM, VT, Expand);
836       setOperationAction(ISD::FDIV, VT, Expand);
837       setOperationAction(ISD::FREM, VT, Expand);
838       setOperationAction(ISD::FNEG, VT, Expand);
839       setOperationAction(ISD::FSQRT, VT, Expand);
840       setOperationAction(ISD::FLOG, VT, Expand);
841       setOperationAction(ISD::FLOG10, VT, Expand);
842       setOperationAction(ISD::FLOG2, VT, Expand);
843       setOperationAction(ISD::FEXP, VT, Expand);
844       setOperationAction(ISD::FEXP2, VT, Expand);
845       setOperationAction(ISD::FSIN, VT, Expand);
846       setOperationAction(ISD::FCOS, VT, Expand);
847       setOperationAction(ISD::FABS, VT, Expand);
848       setOperationAction(ISD::FFLOOR, VT, Expand);
849       setOperationAction(ISD::FCEIL,  VT, Expand);
850       setOperationAction(ISD::FTRUNC, VT, Expand);
851       setOperationAction(ISD::FRINT,  VT, Expand);
852       setOperationAction(ISD::FNEARBYINT, VT, Expand);
853       setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
854       setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
855       setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
856       setOperationAction(ISD::MULHU, VT, Expand);
857       setOperationAction(ISD::MULHS, VT, Expand);
858       setOperationAction(ISD::UMUL_LOHI, VT, Expand);
859       setOperationAction(ISD::SMUL_LOHI, VT, Expand);
860       setOperationAction(ISD::UDIVREM, VT, Expand);
861       setOperationAction(ISD::SDIVREM, VT, Expand);
862       setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
863       setOperationAction(ISD::FPOW, VT, Expand);
864       setOperationAction(ISD::BSWAP, VT, Expand);
865       setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
866       setOperationAction(ISD::ROTL, VT, Expand);
867       setOperationAction(ISD::ROTR, VT, Expand);
868 
869       for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) {
870         setTruncStoreAction(VT, InnerVT, Expand);
871         setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
872         setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
873         setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
874       }
875     }
876     setOperationAction(ISD::SELECT_CC, MVT::v4i32, Expand);
877     if (!Subtarget.hasP8Vector()) {
878       setOperationAction(ISD::SMAX, MVT::v2i64, Expand);
879       setOperationAction(ISD::SMIN, MVT::v2i64, Expand);
880       setOperationAction(ISD::UMAX, MVT::v2i64, Expand);
881       setOperationAction(ISD::UMIN, MVT::v2i64, Expand);
882     }
883 
884     // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
885     // with merges, splats, etc.
886     setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
887 
888     // Vector truncates to sub-word integer that fit in an Altivec/VSX register
889     // are cheap, so handle them before they get expanded to scalar.
890     setOperationAction(ISD::TRUNCATE, MVT::v8i8, Custom);
891     setOperationAction(ISD::TRUNCATE, MVT::v4i8, Custom);
892     setOperationAction(ISD::TRUNCATE, MVT::v2i8, Custom);
893     setOperationAction(ISD::TRUNCATE, MVT::v4i16, Custom);
894     setOperationAction(ISD::TRUNCATE, MVT::v2i16, Custom);
895 
896     setOperationAction(ISD::AND   , MVT::v4i32, Legal);
897     setOperationAction(ISD::OR    , MVT::v4i32, Legal);
898     setOperationAction(ISD::XOR   , MVT::v4i32, Legal);
899     setOperationAction(ISD::LOAD  , MVT::v4i32, Legal);
900     setOperationAction(ISD::SELECT, MVT::v4i32,
901                        Subtarget.useCRBits() ? Legal : Expand);
902     setOperationAction(ISD::STORE , MVT::v4i32, Legal);
903     setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v4i32, Legal);
904     setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v4i32, Legal);
905     setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i32, Legal);
906     setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i32, Legal);
907     setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
908     setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
909     setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
910     setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
911     setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
912     setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
913     setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
914     setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
915 
916     // Custom lowering ROTL v1i128 to VECTOR_SHUFFLE v16i8.
917     setOperationAction(ISD::ROTL, MVT::v1i128, Custom);
918     // With hasAltivec set, we can lower ISD::ROTL to vrl(b|h|w).
919     if (Subtarget.hasAltivec())
920       for (auto VT : {MVT::v4i32, MVT::v8i16, MVT::v16i8})
921         setOperationAction(ISD::ROTL, VT, Legal);
922     // With hasP8Altivec set, we can lower ISD::ROTL to vrld.
923     if (Subtarget.hasP8Altivec())
924       setOperationAction(ISD::ROTL, MVT::v2i64, Legal);
925 
926     addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
927     addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
928     addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
929     addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
930 
931     setOperationAction(ISD::MUL, MVT::v4f32, Legal);
932     setOperationAction(ISD::FMA, MVT::v4f32, Legal);
933 
934     if (Subtarget.hasVSX()) {
935       setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
936       setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
937       setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
938     }
939 
940     if (Subtarget.hasP8Altivec())
941       setOperationAction(ISD::MUL, MVT::v4i32, Legal);
942     else
943       setOperationAction(ISD::MUL, MVT::v4i32, Custom);
944 
945     if (Subtarget.isISA3_1()) {
946       setOperationAction(ISD::MUL, MVT::v2i64, Legal);
947       setOperationAction(ISD::MULHS, MVT::v2i64, Legal);
948       setOperationAction(ISD::MULHU, MVT::v2i64, Legal);
949       setOperationAction(ISD::MULHS, MVT::v4i32, Legal);
950       setOperationAction(ISD::MULHU, MVT::v4i32, Legal);
951       setOperationAction(ISD::UDIV, MVT::v2i64, Legal);
952       setOperationAction(ISD::SDIV, MVT::v2i64, Legal);
953       setOperationAction(ISD::UDIV, MVT::v4i32, Legal);
954       setOperationAction(ISD::SDIV, MVT::v4i32, Legal);
955       setOperationAction(ISD::UREM, MVT::v2i64, Legal);
956       setOperationAction(ISD::SREM, MVT::v2i64, Legal);
957       setOperationAction(ISD::UREM, MVT::v4i32, Legal);
958       setOperationAction(ISD::SREM, MVT::v4i32, Legal);
959       setOperationAction(ISD::UREM, MVT::v1i128, Legal);
960       setOperationAction(ISD::SREM, MVT::v1i128, Legal);
961       setOperationAction(ISD::UDIV, MVT::v1i128, Legal);
962       setOperationAction(ISD::SDIV, MVT::v1i128, Legal);
963       setOperationAction(ISD::ROTL, MVT::v1i128, Legal);
964     }
965 
966     setOperationAction(ISD::MUL, MVT::v8i16, Legal);
967     setOperationAction(ISD::MUL, MVT::v16i8, Custom);
968 
969     setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
970     setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
971 
972     setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
973     setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
974     setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
975     setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
976 
977     // Altivec does not contain unordered floating-point compare instructions
978     setCondCodeAction(ISD::SETUO, MVT::v4f32, Expand);
979     setCondCodeAction(ISD::SETUEQ, MVT::v4f32, Expand);
980     setCondCodeAction(ISD::SETO,   MVT::v4f32, Expand);
981     setCondCodeAction(ISD::SETONE, MVT::v4f32, Expand);
982 
983     if (Subtarget.hasVSX()) {
984       setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f64, Legal);
985       setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Legal);
986       if (Subtarget.hasP8Vector()) {
987         setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Legal);
988         setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Legal);
989       }
990       if (Subtarget.hasDirectMove() && isPPC64) {
991         setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Legal);
992         setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Legal);
993         setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Legal);
994         setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Legal);
995         setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Legal);
996         setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Legal);
997         setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
998         setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
999       }
1000       setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Legal);
1001 
1002       // The nearbyint variants are not allowed to raise the inexact exception
1003       // so we can only code-gen them with unsafe math.
1004       if (TM.Options.UnsafeFPMath) {
1005         setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1006         setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
1007       }
1008 
1009       setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
1010       setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
1011       setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
1012       setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
1013       setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
1014       setOperationAction(ISD::FROUND, MVT::v2f64, Legal);
1015       setOperationAction(ISD::FROUND, MVT::f64, Legal);
1016       setOperationAction(ISD::FRINT, MVT::f64, Legal);
1017 
1018       setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
1019       setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
1020       setOperationAction(ISD::FROUND, MVT::v4f32, Legal);
1021       setOperationAction(ISD::FROUND, MVT::f32, Legal);
1022       setOperationAction(ISD::FRINT, MVT::f32, Legal);
1023 
1024       setOperationAction(ISD::MUL, MVT::v2f64, Legal);
1025       setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1026 
1027       setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
1028       setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
1029 
1030       // Share the Altivec comparison restrictions.
1031       setCondCodeAction(ISD::SETUO, MVT::v2f64, Expand);
1032       setCondCodeAction(ISD::SETUEQ, MVT::v2f64, Expand);
1033       setCondCodeAction(ISD::SETO,   MVT::v2f64, Expand);
1034       setCondCodeAction(ISD::SETONE, MVT::v2f64, Expand);
1035 
1036       setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
1037       setOperationAction(ISD::STORE, MVT::v2f64, Legal);
1038 
1039       setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
1040 
1041       if (Subtarget.hasP8Vector())
1042         addRegisterClass(MVT::f32, &PPC::VSSRCRegClass);
1043 
1044       addRegisterClass(MVT::f64, &PPC::VSFRCRegClass);
1045 
1046       addRegisterClass(MVT::v4i32, &PPC::VSRCRegClass);
1047       addRegisterClass(MVT::v4f32, &PPC::VSRCRegClass);
1048       addRegisterClass(MVT::v2f64, &PPC::VSRCRegClass);
1049 
1050       if (Subtarget.hasP8Altivec()) {
1051         setOperationAction(ISD::SHL, MVT::v2i64, Legal);
1052         setOperationAction(ISD::SRA, MVT::v2i64, Legal);
1053         setOperationAction(ISD::SRL, MVT::v2i64, Legal);
1054 
1055         // 128 bit shifts can be accomplished via 3 instructions for SHL and
1056         // SRL, but not for SRA because of the instructions available:
1057         // VS{RL} and VS{RL}O. However due to direct move costs, it's not worth
1058         // doing
1059         setOperationAction(ISD::SHL, MVT::v1i128, Expand);
1060         setOperationAction(ISD::SRL, MVT::v1i128, Expand);
1061         setOperationAction(ISD::SRA, MVT::v1i128, Expand);
1062 
1063         setOperationAction(ISD::SETCC, MVT::v2i64, Legal);
1064       }
1065       else {
1066         setOperationAction(ISD::SHL, MVT::v2i64, Expand);
1067         setOperationAction(ISD::SRA, MVT::v2i64, Expand);
1068         setOperationAction(ISD::SRL, MVT::v2i64, Expand);
1069 
1070         setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
1071 
1072         // VSX v2i64 only supports non-arithmetic operations.
1073         setOperationAction(ISD::ADD, MVT::v2i64, Expand);
1074         setOperationAction(ISD::SUB, MVT::v2i64, Expand);
1075       }
1076 
1077       if (Subtarget.isISA3_1())
1078         setOperationAction(ISD::SETCC, MVT::v1i128, Legal);
1079       else
1080         setOperationAction(ISD::SETCC, MVT::v1i128, Expand);
1081 
1082       setOperationAction(ISD::LOAD, MVT::v2i64, Promote);
1083       AddPromotedToType (ISD::LOAD, MVT::v2i64, MVT::v2f64);
1084       setOperationAction(ISD::STORE, MVT::v2i64, Promote);
1085       AddPromotedToType (ISD::STORE, MVT::v2i64, MVT::v2f64);
1086 
1087       setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
1088 
1089       setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2i64, Legal);
1090       setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2i64, Legal);
1091       setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v2i64, Legal);
1092       setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v2i64, Legal);
1093       setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1094       setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1095       setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1096       setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1097 
1098       // Custom handling for partial vectors of integers converted to
1099       // floating point. We already have optimal handling for v2i32 through
1100       // the DAG combine, so those aren't necessary.
1101       setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2i8, Custom);
1102       setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i8, Custom);
1103       setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2i16, Custom);
1104       setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i16, Custom);
1105       setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2i8, Custom);
1106       setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i8, Custom);
1107       setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2i16, Custom);
1108       setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i16, Custom);
1109       setOperationAction(ISD::UINT_TO_FP, MVT::v2i8, Custom);
1110       setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
1111       setOperationAction(ISD::UINT_TO_FP, MVT::v2i16, Custom);
1112       setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
1113       setOperationAction(ISD::SINT_TO_FP, MVT::v2i8, Custom);
1114       setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Custom);
1115       setOperationAction(ISD::SINT_TO_FP, MVT::v2i16, Custom);
1116       setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
1117 
1118       setOperationAction(ISD::FNEG, MVT::v4f32, Legal);
1119       setOperationAction(ISD::FNEG, MVT::v2f64, Legal);
1120       setOperationAction(ISD::FABS, MVT::v4f32, Legal);
1121       setOperationAction(ISD::FABS, MVT::v2f64, Legal);
1122       setOperationAction(ISD::FCOPYSIGN, MVT::v4f32, Legal);
1123       setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Legal);
1124 
1125       setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
1126       setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
1127 
1128       // Handle constrained floating-point operations of vector.
1129       // The predictor is `hasVSX` because altivec instruction has
1130       // no exception but VSX vector instruction has.
1131       setOperationAction(ISD::STRICT_FADD, MVT::v4f32, Legal);
1132       setOperationAction(ISD::STRICT_FSUB, MVT::v4f32, Legal);
1133       setOperationAction(ISD::STRICT_FMUL, MVT::v4f32, Legal);
1134       setOperationAction(ISD::STRICT_FDIV, MVT::v4f32, Legal);
1135       setOperationAction(ISD::STRICT_FMA, MVT::v4f32, Legal);
1136       setOperationAction(ISD::STRICT_FSQRT, MVT::v4f32, Legal);
1137       setOperationAction(ISD::STRICT_FMAXNUM, MVT::v4f32, Legal);
1138       setOperationAction(ISD::STRICT_FMINNUM, MVT::v4f32, Legal);
1139       setOperationAction(ISD::STRICT_FRINT, MVT::v4f32, Legal);
1140       setOperationAction(ISD::STRICT_FFLOOR, MVT::v4f32, Legal);
1141       setOperationAction(ISD::STRICT_FCEIL,  MVT::v4f32, Legal);
1142       setOperationAction(ISD::STRICT_FTRUNC, MVT::v4f32, Legal);
1143       setOperationAction(ISD::STRICT_FROUND, MVT::v4f32, Legal);
1144 
1145       setOperationAction(ISD::STRICT_FADD, MVT::v2f64, Legal);
1146       setOperationAction(ISD::STRICT_FSUB, MVT::v2f64, Legal);
1147       setOperationAction(ISD::STRICT_FMUL, MVT::v2f64, Legal);
1148       setOperationAction(ISD::STRICT_FDIV, MVT::v2f64, Legal);
1149       setOperationAction(ISD::STRICT_FMA, MVT::v2f64, Legal);
1150       setOperationAction(ISD::STRICT_FSQRT, MVT::v2f64, Legal);
1151       setOperationAction(ISD::STRICT_FMAXNUM, MVT::v2f64, Legal);
1152       setOperationAction(ISD::STRICT_FMINNUM, MVT::v2f64, Legal);
1153       setOperationAction(ISD::STRICT_FRINT, MVT::v2f64, Legal);
1154       setOperationAction(ISD::STRICT_FFLOOR, MVT::v2f64, Legal);
1155       setOperationAction(ISD::STRICT_FCEIL,  MVT::v2f64, Legal);
1156       setOperationAction(ISD::STRICT_FTRUNC, MVT::v2f64, Legal);
1157       setOperationAction(ISD::STRICT_FROUND, MVT::v2f64, Legal);
1158 
1159       addRegisterClass(MVT::v2i64, &PPC::VSRCRegClass);
1160       addRegisterClass(MVT::f128, &PPC::VRRCRegClass);
1161 
1162       for (MVT FPT : MVT::fp_valuetypes())
1163         setLoadExtAction(ISD::EXTLOAD, MVT::f128, FPT, Expand);
1164 
1165       // Expand the SELECT to SELECT_CC
1166       setOperationAction(ISD::SELECT, MVT::f128, Expand);
1167 
1168       setTruncStoreAction(MVT::f128, MVT::f64, Expand);
1169       setTruncStoreAction(MVT::f128, MVT::f32, Expand);
1170 
1171       // No implementation for these ops for PowerPC.
1172       setOperationAction(ISD::FSIN, MVT::f128, Expand);
1173       setOperationAction(ISD::FCOS, MVT::f128, Expand);
1174       setOperationAction(ISD::FPOW, MVT::f128, Expand);
1175       setOperationAction(ISD::FPOWI, MVT::f128, Expand);
1176       setOperationAction(ISD::FREM, MVT::f128, Expand);
1177     }
1178 
1179     if (Subtarget.hasP8Altivec()) {
1180       addRegisterClass(MVT::v2i64, &PPC::VRRCRegClass);
1181       addRegisterClass(MVT::v1i128, &PPC::VRRCRegClass);
1182     }
1183 
1184     if (Subtarget.hasP9Vector()) {
1185       setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1186       setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1187 
1188       // 128 bit shifts can be accomplished via 3 instructions for SHL and
1189       // SRL, but not for SRA because of the instructions available:
1190       // VS{RL} and VS{RL}O.
1191       setOperationAction(ISD::SHL, MVT::v1i128, Legal);
1192       setOperationAction(ISD::SRL, MVT::v1i128, Legal);
1193       setOperationAction(ISD::SRA, MVT::v1i128, Expand);
1194 
1195       setOperationAction(ISD::FADD, MVT::f128, Legal);
1196       setOperationAction(ISD::FSUB, MVT::f128, Legal);
1197       setOperationAction(ISD::FDIV, MVT::f128, Legal);
1198       setOperationAction(ISD::FMUL, MVT::f128, Legal);
1199       setOperationAction(ISD::FP_EXTEND, MVT::f128, Legal);
1200 
1201       setOperationAction(ISD::FMA, MVT::f128, Legal);
1202       setCondCodeAction(ISD::SETULT, MVT::f128, Expand);
1203       setCondCodeAction(ISD::SETUGT, MVT::f128, Expand);
1204       setCondCodeAction(ISD::SETUEQ, MVT::f128, Expand);
1205       setCondCodeAction(ISD::SETOGE, MVT::f128, Expand);
1206       setCondCodeAction(ISD::SETOLE, MVT::f128, Expand);
1207       setCondCodeAction(ISD::SETONE, MVT::f128, Expand);
1208 
1209       setOperationAction(ISD::FTRUNC, MVT::f128, Legal);
1210       setOperationAction(ISD::FRINT, MVT::f128, Legal);
1211       setOperationAction(ISD::FFLOOR, MVT::f128, Legal);
1212       setOperationAction(ISD::FCEIL, MVT::f128, Legal);
1213       setOperationAction(ISD::FNEARBYINT, MVT::f128, Legal);
1214       setOperationAction(ISD::FROUND, MVT::f128, Legal);
1215 
1216       setOperationAction(ISD::FP_ROUND, MVT::f64, Legal);
1217       setOperationAction(ISD::FP_ROUND, MVT::f32, Legal);
1218       setOperationAction(ISD::BITCAST, MVT::i128, Custom);
1219 
1220       // Handle constrained floating-point operations of fp128
1221       setOperationAction(ISD::STRICT_FADD, MVT::f128, Legal);
1222       setOperationAction(ISD::STRICT_FSUB, MVT::f128, Legal);
1223       setOperationAction(ISD::STRICT_FMUL, MVT::f128, Legal);
1224       setOperationAction(ISD::STRICT_FDIV, MVT::f128, Legal);
1225       setOperationAction(ISD::STRICT_FMA, MVT::f128, Legal);
1226       setOperationAction(ISD::STRICT_FSQRT, MVT::f128, Legal);
1227       setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f128, Legal);
1228       setOperationAction(ISD::STRICT_FP_ROUND, MVT::f64, Legal);
1229       setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal);
1230       setOperationAction(ISD::STRICT_FRINT, MVT::f128, Legal);
1231       setOperationAction(ISD::STRICT_FNEARBYINT, MVT::f128, Legal);
1232       setOperationAction(ISD::STRICT_FFLOOR, MVT::f128, Legal);
1233       setOperationAction(ISD::STRICT_FCEIL, MVT::f128, Legal);
1234       setOperationAction(ISD::STRICT_FTRUNC, MVT::f128, Legal);
1235       setOperationAction(ISD::STRICT_FROUND, MVT::f128, Legal);
1236       setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
1237       setOperationAction(ISD::BSWAP, MVT::v8i16, Legal);
1238       setOperationAction(ISD::BSWAP, MVT::v4i32, Legal);
1239       setOperationAction(ISD::BSWAP, MVT::v2i64, Legal);
1240       setOperationAction(ISD::BSWAP, MVT::v1i128, Legal);
1241     } else if (Subtarget.hasVSX()) {
1242       setOperationAction(ISD::LOAD, MVT::f128, Promote);
1243       setOperationAction(ISD::STORE, MVT::f128, Promote);
1244 
1245       AddPromotedToType(ISD::LOAD, MVT::f128, MVT::v4i32);
1246       AddPromotedToType(ISD::STORE, MVT::f128, MVT::v4i32);
1247 
1248       // Set FADD/FSUB as libcall to avoid the legalizer to expand the
1249       // fp_to_uint and int_to_fp.
1250       setOperationAction(ISD::FADD, MVT::f128, LibCall);
1251       setOperationAction(ISD::FSUB, MVT::f128, LibCall);
1252 
1253       setOperationAction(ISD::FMUL, MVT::f128, Expand);
1254       setOperationAction(ISD::FDIV, MVT::f128, Expand);
1255       setOperationAction(ISD::FNEG, MVT::f128, Expand);
1256       setOperationAction(ISD::FABS, MVT::f128, Expand);
1257       setOperationAction(ISD::FSQRT, MVT::f128, Expand);
1258       setOperationAction(ISD::FMA, MVT::f128, Expand);
1259       setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
1260 
1261       // Expand the fp_extend if the target type is fp128.
1262       setOperationAction(ISD::FP_EXTEND, MVT::f128, Expand);
1263       setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f128, Expand);
1264 
1265       // Expand the fp_round if the source type is fp128.
1266       for (MVT VT : {MVT::f32, MVT::f64}) {
1267         setOperationAction(ISD::FP_ROUND, VT, Custom);
1268         setOperationAction(ISD::STRICT_FP_ROUND, VT, Custom);
1269       }
1270 
1271       setOperationAction(ISD::SETCC, MVT::f128, Custom);
1272       setOperationAction(ISD::STRICT_FSETCC, MVT::f128, Custom);
1273       setOperationAction(ISD::STRICT_FSETCCS, MVT::f128, Custom);
1274       setOperationAction(ISD::BR_CC, MVT::f128, Expand);
1275 
1276       // Lower following f128 select_cc pattern:
1277       // select_cc x, y, tv, fv, cc -> select_cc (setcc x, y, cc), 0, tv, fv, NE
1278       setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
1279 
1280       // We need to handle f128 SELECT_CC with integer result type.
1281       setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
1282       setOperationAction(ISD::SELECT_CC, MVT::i64, isPPC64 ? Custom : Expand);
1283     }
1284 
1285     if (Subtarget.hasP9Altivec()) {
1286       if (Subtarget.isISA3_1()) {
1287         setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
1288         setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Legal);
1289         setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Legal);
1290         setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
1291       } else {
1292         setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1293         setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1294       }
1295       setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8,  Legal);
1296       setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Legal);
1297       setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i32, Legal);
1298       setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8,  Legal);
1299       setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Legal);
1300       setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i32, Legal);
1301       setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i64, Legal);
1302 
1303       setOperationAction(ISD::ABDU, MVT::v16i8, Legal);
1304       setOperationAction(ISD::ABDU, MVT::v8i16, Legal);
1305       setOperationAction(ISD::ABDU, MVT::v4i32, Legal);
1306       setOperationAction(ISD::ABDS, MVT::v4i32, Legal);
1307     }
1308 
1309     if (Subtarget.hasP10Vector()) {
1310       setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
1311     }
1312   }
1313 
1314   if (Subtarget.pairedVectorMemops()) {
1315     addRegisterClass(MVT::v256i1, &PPC::VSRpRCRegClass);
1316     setOperationAction(ISD::LOAD, MVT::v256i1, Custom);
1317     setOperationAction(ISD::STORE, MVT::v256i1, Custom);
1318   }
1319   if (Subtarget.hasMMA()) {
1320     if (Subtarget.isISAFuture())
1321       addRegisterClass(MVT::v512i1, &PPC::WACCRCRegClass);
1322     else
1323       addRegisterClass(MVT::v512i1, &PPC::UACCRCRegClass);
1324     setOperationAction(ISD::LOAD, MVT::v512i1, Custom);
1325     setOperationAction(ISD::STORE, MVT::v512i1, Custom);
1326     setOperationAction(ISD::BUILD_VECTOR, MVT::v512i1, Custom);
1327   }
1328 
1329   if (Subtarget.has64BitSupport())
1330     setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
1331 
1332   if (Subtarget.isISA3_1())
1333     setOperationAction(ISD::SRA, MVT::v1i128, Legal);
1334 
1335   setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, isPPC64 ? Legal : Custom);
1336 
1337   if (!isPPC64) {
1338     setOperationAction(ISD::ATOMIC_LOAD,  MVT::i64, Expand);
1339     setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
1340   }
1341 
1342   if (shouldInlineQuadwordAtomics()) {
1343     setOperationAction(ISD::ATOMIC_LOAD, MVT::i128, Custom);
1344     setOperationAction(ISD::ATOMIC_STORE, MVT::i128, Custom);
1345     setOperationAction(ISD::INTRINSIC_VOID, MVT::i128, Custom);
1346   }
1347 
1348   setBooleanContents(ZeroOrOneBooleanContent);
1349 
1350   if (Subtarget.hasAltivec()) {
1351     // Altivec instructions set fields to all zeros or all ones.
1352     setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
1353   }
1354 
1355   setLibcallName(RTLIB::MULO_I128, nullptr);
1356   if (!isPPC64) {
1357     // These libcalls are not available in 32-bit.
1358     setLibcallName(RTLIB::SHL_I128, nullptr);
1359     setLibcallName(RTLIB::SRL_I128, nullptr);
1360     setLibcallName(RTLIB::SRA_I128, nullptr);
1361     setLibcallName(RTLIB::MUL_I128, nullptr);
1362     setLibcallName(RTLIB::MULO_I64, nullptr);
1363   }
1364 
1365   if (!isPPC64)
1366     setMaxAtomicSizeInBitsSupported(32);
1367   else if (shouldInlineQuadwordAtomics())
1368     setMaxAtomicSizeInBitsSupported(128);
1369   else
1370     setMaxAtomicSizeInBitsSupported(64);
1371 
1372   setStackPointerRegisterToSaveRestore(isPPC64 ? PPC::X1 : PPC::R1);
1373 
1374   // We have target-specific dag combine patterns for the following nodes:
1375   setTargetDAGCombine({ISD::ADD, ISD::SHL, ISD::SRA, ISD::SRL, ISD::MUL,
1376                        ISD::FMA, ISD::SINT_TO_FP, ISD::BUILD_VECTOR});
1377   if (Subtarget.hasFPCVT())
1378     setTargetDAGCombine(ISD::UINT_TO_FP);
1379   setTargetDAGCombine({ISD::LOAD, ISD::STORE, ISD::BR_CC});
1380   if (Subtarget.useCRBits())
1381     setTargetDAGCombine(ISD::BRCOND);
1382   setTargetDAGCombine({ISD::BSWAP, ISD::INTRINSIC_WO_CHAIN,
1383                        ISD::INTRINSIC_W_CHAIN, ISD::INTRINSIC_VOID});
1384 
1385   setTargetDAGCombine({ISD::SIGN_EXTEND, ISD::ZERO_EXTEND, ISD::ANY_EXTEND});
1386 
1387   setTargetDAGCombine({ISD::TRUNCATE, ISD::VECTOR_SHUFFLE});
1388 
1389   if (Subtarget.useCRBits()) {
1390     setTargetDAGCombine({ISD::TRUNCATE, ISD::SETCC, ISD::SELECT_CC});
1391   }
1392 
1393   if (Subtarget.hasP9Altivec()) {
1394     setTargetDAGCombine({ISD::VSELECT});
1395   }
1396 
1397   setLibcallName(RTLIB::LOG_F128, "logf128");
1398   setLibcallName(RTLIB::LOG2_F128, "log2f128");
1399   setLibcallName(RTLIB::LOG10_F128, "log10f128");
1400   setLibcallName(RTLIB::EXP_F128, "expf128");
1401   setLibcallName(RTLIB::EXP2_F128, "exp2f128");
1402   setLibcallName(RTLIB::SIN_F128, "sinf128");
1403   setLibcallName(RTLIB::COS_F128, "cosf128");
1404   setLibcallName(RTLIB::POW_F128, "powf128");
1405   setLibcallName(RTLIB::FMIN_F128, "fminf128");
1406   setLibcallName(RTLIB::FMAX_F128, "fmaxf128");
1407   setLibcallName(RTLIB::REM_F128, "fmodf128");
1408   setLibcallName(RTLIB::SQRT_F128, "sqrtf128");
1409   setLibcallName(RTLIB::CEIL_F128, "ceilf128");
1410   setLibcallName(RTLIB::FLOOR_F128, "floorf128");
1411   setLibcallName(RTLIB::TRUNC_F128, "truncf128");
1412   setLibcallName(RTLIB::ROUND_F128, "roundf128");
1413   setLibcallName(RTLIB::LROUND_F128, "lroundf128");
1414   setLibcallName(RTLIB::LLROUND_F128, "llroundf128");
1415   setLibcallName(RTLIB::RINT_F128, "rintf128");
1416   setLibcallName(RTLIB::LRINT_F128, "lrintf128");
1417   setLibcallName(RTLIB::LLRINT_F128, "llrintf128");
1418   setLibcallName(RTLIB::NEARBYINT_F128, "nearbyintf128");
1419   setLibcallName(RTLIB::FMA_F128, "fmaf128");
1420 
1421   // With 32 condition bits, we don't need to sink (and duplicate) compares
1422   // aggressively in CodeGenPrep.
1423   if (Subtarget.useCRBits()) {
1424     setHasMultipleConditionRegisters();
1425     setJumpIsExpensive();
1426   }
1427 
1428   setMinFunctionAlignment(Align(4));
1429 
1430   switch (Subtarget.getCPUDirective()) {
1431   default: break;
1432   case PPC::DIR_970:
1433   case PPC::DIR_A2:
1434   case PPC::DIR_E500:
1435   case PPC::DIR_E500mc:
1436   case PPC::DIR_E5500:
1437   case PPC::DIR_PWR4:
1438   case PPC::DIR_PWR5:
1439   case PPC::DIR_PWR5X:
1440   case PPC::DIR_PWR6:
1441   case PPC::DIR_PWR6X:
1442   case PPC::DIR_PWR7:
1443   case PPC::DIR_PWR8:
1444   case PPC::DIR_PWR9:
1445   case PPC::DIR_PWR10:
1446   case PPC::DIR_PWR_FUTURE:
1447     setPrefLoopAlignment(Align(16));
1448     setPrefFunctionAlignment(Align(16));
1449     break;
1450   }
1451 
1452   if (Subtarget.enableMachineScheduler())
1453     setSchedulingPreference(Sched::Source);
1454   else
1455     setSchedulingPreference(Sched::Hybrid);
1456 
1457   computeRegisterProperties(STI.getRegisterInfo());
1458 
1459   // The Freescale cores do better with aggressive inlining of memcpy and
1460   // friends. GCC uses same threshold of 128 bytes (= 32 word stores).
1461   if (Subtarget.getCPUDirective() == PPC::DIR_E500mc ||
1462       Subtarget.getCPUDirective() == PPC::DIR_E5500) {
1463     MaxStoresPerMemset = 32;
1464     MaxStoresPerMemsetOptSize = 16;
1465     MaxStoresPerMemcpy = 32;
1466     MaxStoresPerMemcpyOptSize = 8;
1467     MaxStoresPerMemmove = 32;
1468     MaxStoresPerMemmoveOptSize = 8;
1469   } else if (Subtarget.getCPUDirective() == PPC::DIR_A2) {
1470     // The A2 also benefits from (very) aggressive inlining of memcpy and
1471     // friends. The overhead of a the function call, even when warm, can be
1472     // over one hundred cycles.
1473     MaxStoresPerMemset = 128;
1474     MaxStoresPerMemcpy = 128;
1475     MaxStoresPerMemmove = 128;
1476     MaxLoadsPerMemcmp = 128;
1477   } else {
1478     MaxLoadsPerMemcmp = 8;
1479     MaxLoadsPerMemcmpOptSize = 4;
1480   }
1481 
1482   IsStrictFPEnabled = true;
1483 
1484   // Let the subtarget (CPU) decide if a predictable select is more expensive
1485   // than the corresponding branch. This information is used in CGP to decide
1486   // when to convert selects into branches.
1487   PredictableSelectIsExpensive = Subtarget.isPredictableSelectIsExpensive();
1488 }
1489 
1490 // *********************************** NOTE ************************************
1491 // For selecting load and store instructions, the addressing modes are defined
1492 // as ComplexPatterns in PPCInstrInfo.td, which are then utilized in the TD
1493 // patterns to match the load the store instructions.
1494 //
1495 // The TD definitions for the addressing modes correspond to their respective
1496 // Select<AddrMode>Form() function in PPCISelDAGToDAG.cpp. These functions rely
1497 // on SelectOptimalAddrMode(), which calls computeMOFlags() to compute the
1498 // address mode flags of a particular node. Afterwards, the computed address
1499 // flags are passed into getAddrModeForFlags() in order to retrieve the optimal
1500 // addressing mode. SelectOptimalAddrMode() then sets the Base and Displacement
1501 // accordingly, based on the preferred addressing mode.
1502 //
1503 // Within PPCISelLowering.h, there are two enums: MemOpFlags and AddrMode.
1504 // MemOpFlags contains all the possible flags that can be used to compute the
1505 // optimal addressing mode for load and store instructions.
1506 // AddrMode contains all the possible load and store addressing modes available
1507 // on Power (such as DForm, DSForm, DQForm, XForm, etc.)
1508 //
1509 // When adding new load and store instructions, it is possible that new address
1510 // flags may need to be added into MemOpFlags, and a new addressing mode will
1511 // need to be added to AddrMode. An entry of the new addressing mode (consisting
1512 // of the minimal and main distinguishing address flags for the new load/store
1513 // instructions) will need to be added into initializeAddrModeMap() below.
1514 // Finally, when adding new addressing modes, the getAddrModeForFlags() will
1515 // need to be updated to account for selecting the optimal addressing mode.
1516 // *****************************************************************************
1517 /// Initialize the map that relates the different addressing modes of the load
1518 /// and store instructions to a set of flags. This ensures the load/store
1519 /// instruction is correctly matched during instruction selection.
1520 void PPCTargetLowering::initializeAddrModeMap() {
1521   AddrModesMap[PPC::AM_DForm] = {
1522       // LWZ, STW
1523       PPC::MOF_ZExt | PPC::MOF_RPlusSImm16 | PPC::MOF_WordInt,
1524       PPC::MOF_ZExt | PPC::MOF_RPlusLo | PPC::MOF_WordInt,
1525       PPC::MOF_ZExt | PPC::MOF_NotAddNorCst | PPC::MOF_WordInt,
1526       PPC::MOF_ZExt | PPC::MOF_AddrIsSImm32 | PPC::MOF_WordInt,
1527       // LBZ, LHZ, STB, STH
1528       PPC::MOF_ZExt | PPC::MOF_RPlusSImm16 | PPC::MOF_SubWordInt,
1529       PPC::MOF_ZExt | PPC::MOF_RPlusLo | PPC::MOF_SubWordInt,
1530       PPC::MOF_ZExt | PPC::MOF_NotAddNorCst | PPC::MOF_SubWordInt,
1531       PPC::MOF_ZExt | PPC::MOF_AddrIsSImm32 | PPC::MOF_SubWordInt,
1532       // LHA
1533       PPC::MOF_SExt | PPC::MOF_RPlusSImm16 | PPC::MOF_SubWordInt,
1534       PPC::MOF_SExt | PPC::MOF_RPlusLo | PPC::MOF_SubWordInt,
1535       PPC::MOF_SExt | PPC::MOF_NotAddNorCst | PPC::MOF_SubWordInt,
1536       PPC::MOF_SExt | PPC::MOF_AddrIsSImm32 | PPC::MOF_SubWordInt,
1537       // LFS, LFD, STFS, STFD
1538       PPC::MOF_RPlusSImm16 | PPC::MOF_ScalarFloat | PPC::MOF_SubtargetBeforeP9,
1539       PPC::MOF_RPlusLo | PPC::MOF_ScalarFloat | PPC::MOF_SubtargetBeforeP9,
1540       PPC::MOF_NotAddNorCst | PPC::MOF_ScalarFloat | PPC::MOF_SubtargetBeforeP9,
1541       PPC::MOF_AddrIsSImm32 | PPC::MOF_ScalarFloat | PPC::MOF_SubtargetBeforeP9,
1542   };
1543   AddrModesMap[PPC::AM_DSForm] = {
1544       // LWA
1545       PPC::MOF_SExt | PPC::MOF_RPlusSImm16Mult4 | PPC::MOF_WordInt,
1546       PPC::MOF_SExt | PPC::MOF_NotAddNorCst | PPC::MOF_WordInt,
1547       PPC::MOF_SExt | PPC::MOF_AddrIsSImm32 | PPC::MOF_WordInt,
1548       // LD, STD
1549       PPC::MOF_RPlusSImm16Mult4 | PPC::MOF_DoubleWordInt,
1550       PPC::MOF_NotAddNorCst | PPC::MOF_DoubleWordInt,
1551       PPC::MOF_AddrIsSImm32 | PPC::MOF_DoubleWordInt,
1552       // DFLOADf32, DFLOADf64, DSTOREf32, DSTOREf64
1553       PPC::MOF_RPlusSImm16Mult4 | PPC::MOF_ScalarFloat | PPC::MOF_SubtargetP9,
1554       PPC::MOF_NotAddNorCst | PPC::MOF_ScalarFloat | PPC::MOF_SubtargetP9,
1555       PPC::MOF_AddrIsSImm32 | PPC::MOF_ScalarFloat | PPC::MOF_SubtargetP9,
1556   };
1557   AddrModesMap[PPC::AM_DQForm] = {
1558       // LXV, STXV
1559       PPC::MOF_RPlusSImm16Mult16 | PPC::MOF_Vector | PPC::MOF_SubtargetP9,
1560       PPC::MOF_NotAddNorCst | PPC::MOF_Vector | PPC::MOF_SubtargetP9,
1561       PPC::MOF_AddrIsSImm32 | PPC::MOF_Vector | PPC::MOF_SubtargetP9,
1562   };
1563   AddrModesMap[PPC::AM_PrefixDForm] = {PPC::MOF_RPlusSImm34 |
1564                                        PPC::MOF_SubtargetP10};
1565   // TODO: Add mapping for quadword load/store.
1566 }
1567 
1568 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1569 /// the desired ByVal argument alignment.
1570 static void getMaxByValAlign(Type *Ty, Align &MaxAlign, Align MaxMaxAlign) {
1571   if (MaxAlign == MaxMaxAlign)
1572     return;
1573   if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1574     if (MaxMaxAlign >= 32 &&
1575         VTy->getPrimitiveSizeInBits().getFixedValue() >= 256)
1576       MaxAlign = Align(32);
1577     else if (VTy->getPrimitiveSizeInBits().getFixedValue() >= 128 &&
1578              MaxAlign < 16)
1579       MaxAlign = Align(16);
1580   } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1581     Align EltAlign;
1582     getMaxByValAlign(ATy->getElementType(), EltAlign, MaxMaxAlign);
1583     if (EltAlign > MaxAlign)
1584       MaxAlign = EltAlign;
1585   } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1586     for (auto *EltTy : STy->elements()) {
1587       Align EltAlign;
1588       getMaxByValAlign(EltTy, EltAlign, MaxMaxAlign);
1589       if (EltAlign > MaxAlign)
1590         MaxAlign = EltAlign;
1591       if (MaxAlign == MaxMaxAlign)
1592         break;
1593     }
1594   }
1595 }
1596 
1597 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1598 /// function arguments in the caller parameter area.
1599 uint64_t PPCTargetLowering::getByValTypeAlignment(Type *Ty,
1600                                                   const DataLayout &DL) const {
1601   // 16byte and wider vectors are passed on 16byte boundary.
1602   // The rest is 8 on PPC64 and 4 on PPC32 boundary.
1603   Align Alignment = Subtarget.isPPC64() ? Align(8) : Align(4);
1604   if (Subtarget.hasAltivec())
1605     getMaxByValAlign(Ty, Alignment, Align(16));
1606   return Alignment.value();
1607 }
1608 
1609 bool PPCTargetLowering::useSoftFloat() const {
1610   return Subtarget.useSoftFloat();
1611 }
1612 
1613 bool PPCTargetLowering::hasSPE() const {
1614   return Subtarget.hasSPE();
1615 }
1616 
1617 bool PPCTargetLowering::preferIncOfAddToSubOfNot(EVT VT) const {
1618   return VT.isScalarInteger();
1619 }
1620 
1621 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
1622   switch ((PPCISD::NodeType)Opcode) {
1623   case PPCISD::FIRST_NUMBER:    break;
1624   case PPCISD::FSEL:            return "PPCISD::FSEL";
1625   case PPCISD::XSMAXC:          return "PPCISD::XSMAXC";
1626   case PPCISD::XSMINC:          return "PPCISD::XSMINC";
1627   case PPCISD::FCFID:           return "PPCISD::FCFID";
1628   case PPCISD::FCFIDU:          return "PPCISD::FCFIDU";
1629   case PPCISD::FCFIDS:          return "PPCISD::FCFIDS";
1630   case PPCISD::FCFIDUS:         return "PPCISD::FCFIDUS";
1631   case PPCISD::FCTIDZ:          return "PPCISD::FCTIDZ";
1632   case PPCISD::FCTIWZ:          return "PPCISD::FCTIWZ";
1633   case PPCISD::FCTIDUZ:         return "PPCISD::FCTIDUZ";
1634   case PPCISD::FCTIWUZ:         return "PPCISD::FCTIWUZ";
1635   case PPCISD::FP_TO_UINT_IN_VSR:
1636                                 return "PPCISD::FP_TO_UINT_IN_VSR,";
1637   case PPCISD::FP_TO_SINT_IN_VSR:
1638                                 return "PPCISD::FP_TO_SINT_IN_VSR";
1639   case PPCISD::FRE:             return "PPCISD::FRE";
1640   case PPCISD::FRSQRTE:         return "PPCISD::FRSQRTE";
1641   case PPCISD::FTSQRT:
1642     return "PPCISD::FTSQRT";
1643   case PPCISD::FSQRT:
1644     return "PPCISD::FSQRT";
1645   case PPCISD::STFIWX:          return "PPCISD::STFIWX";
1646   case PPCISD::VPERM:           return "PPCISD::VPERM";
1647   case PPCISD::XXSPLT:          return "PPCISD::XXSPLT";
1648   case PPCISD::XXSPLTI_SP_TO_DP:
1649     return "PPCISD::XXSPLTI_SP_TO_DP";
1650   case PPCISD::XXSPLTI32DX:
1651     return "PPCISD::XXSPLTI32DX";
1652   case PPCISD::VECINSERT:       return "PPCISD::VECINSERT";
1653   case PPCISD::XXPERMDI:        return "PPCISD::XXPERMDI";
1654   case PPCISD::XXPERM:
1655     return "PPCISD::XXPERM";
1656   case PPCISD::VECSHL:          return "PPCISD::VECSHL";
1657   case PPCISD::CMPB:            return "PPCISD::CMPB";
1658   case PPCISD::Hi:              return "PPCISD::Hi";
1659   case PPCISD::Lo:              return "PPCISD::Lo";
1660   case PPCISD::TOC_ENTRY:       return "PPCISD::TOC_ENTRY";
1661   case PPCISD::ATOMIC_CMP_SWAP_8: return "PPCISD::ATOMIC_CMP_SWAP_8";
1662   case PPCISD::ATOMIC_CMP_SWAP_16: return "PPCISD::ATOMIC_CMP_SWAP_16";
1663   case PPCISD::DYNALLOC:        return "PPCISD::DYNALLOC";
1664   case PPCISD::DYNAREAOFFSET:   return "PPCISD::DYNAREAOFFSET";
1665   case PPCISD::PROBED_ALLOCA:   return "PPCISD::PROBED_ALLOCA";
1666   case PPCISD::GlobalBaseReg:   return "PPCISD::GlobalBaseReg";
1667   case PPCISD::SRL:             return "PPCISD::SRL";
1668   case PPCISD::SRA:             return "PPCISD::SRA";
1669   case PPCISD::SHL:             return "PPCISD::SHL";
1670   case PPCISD::SRA_ADDZE:       return "PPCISD::SRA_ADDZE";
1671   case PPCISD::CALL:            return "PPCISD::CALL";
1672   case PPCISD::CALL_NOP:        return "PPCISD::CALL_NOP";
1673   case PPCISD::CALL_NOTOC:      return "PPCISD::CALL_NOTOC";
1674   case PPCISD::CALL_RM:
1675     return "PPCISD::CALL_RM";
1676   case PPCISD::CALL_NOP_RM:
1677     return "PPCISD::CALL_NOP_RM";
1678   case PPCISD::CALL_NOTOC_RM:
1679     return "PPCISD::CALL_NOTOC_RM";
1680   case PPCISD::MTCTR:           return "PPCISD::MTCTR";
1681   case PPCISD::BCTRL:           return "PPCISD::BCTRL";
1682   case PPCISD::BCTRL_LOAD_TOC:  return "PPCISD::BCTRL_LOAD_TOC";
1683   case PPCISD::BCTRL_RM:
1684     return "PPCISD::BCTRL_RM";
1685   case PPCISD::BCTRL_LOAD_TOC_RM:
1686     return "PPCISD::BCTRL_LOAD_TOC_RM";
1687   case PPCISD::RET_FLAG:        return "PPCISD::RET_FLAG";
1688   case PPCISD::READ_TIME_BASE:  return "PPCISD::READ_TIME_BASE";
1689   case PPCISD::EH_SJLJ_SETJMP:  return "PPCISD::EH_SJLJ_SETJMP";
1690   case PPCISD::EH_SJLJ_LONGJMP: return "PPCISD::EH_SJLJ_LONGJMP";
1691   case PPCISD::MFOCRF:          return "PPCISD::MFOCRF";
1692   case PPCISD::MFVSR:           return "PPCISD::MFVSR";
1693   case PPCISD::MTVSRA:          return "PPCISD::MTVSRA";
1694   case PPCISD::MTVSRZ:          return "PPCISD::MTVSRZ";
1695   case PPCISD::SINT_VEC_TO_FP:  return "PPCISD::SINT_VEC_TO_FP";
1696   case PPCISD::UINT_VEC_TO_FP:  return "PPCISD::UINT_VEC_TO_FP";
1697   case PPCISD::SCALAR_TO_VECTOR_PERMUTED:
1698     return "PPCISD::SCALAR_TO_VECTOR_PERMUTED";
1699   case PPCISD::ANDI_rec_1_EQ_BIT:
1700     return "PPCISD::ANDI_rec_1_EQ_BIT";
1701   case PPCISD::ANDI_rec_1_GT_BIT:
1702     return "PPCISD::ANDI_rec_1_GT_BIT";
1703   case PPCISD::VCMP:            return "PPCISD::VCMP";
1704   case PPCISD::VCMP_rec:        return "PPCISD::VCMP_rec";
1705   case PPCISD::LBRX:            return "PPCISD::LBRX";
1706   case PPCISD::STBRX:           return "PPCISD::STBRX";
1707   case PPCISD::LFIWAX:          return "PPCISD::LFIWAX";
1708   case PPCISD::LFIWZX:          return "PPCISD::LFIWZX";
1709   case PPCISD::LXSIZX:          return "PPCISD::LXSIZX";
1710   case PPCISD::STXSIX:          return "PPCISD::STXSIX";
1711   case PPCISD::VEXTS:           return "PPCISD::VEXTS";
1712   case PPCISD::LXVD2X:          return "PPCISD::LXVD2X";
1713   case PPCISD::STXVD2X:         return "PPCISD::STXVD2X";
1714   case PPCISD::LOAD_VEC_BE:     return "PPCISD::LOAD_VEC_BE";
1715   case PPCISD::STORE_VEC_BE:    return "PPCISD::STORE_VEC_BE";
1716   case PPCISD::ST_VSR_SCAL_INT:
1717                                 return "PPCISD::ST_VSR_SCAL_INT";
1718   case PPCISD::COND_BRANCH:     return "PPCISD::COND_BRANCH";
1719   case PPCISD::BDNZ:            return "PPCISD::BDNZ";
1720   case PPCISD::BDZ:             return "PPCISD::BDZ";
1721   case PPCISD::MFFS:            return "PPCISD::MFFS";
1722   case PPCISD::FADDRTZ:         return "PPCISD::FADDRTZ";
1723   case PPCISD::TC_RETURN:       return "PPCISD::TC_RETURN";
1724   case PPCISD::CR6SET:          return "PPCISD::CR6SET";
1725   case PPCISD::CR6UNSET:        return "PPCISD::CR6UNSET";
1726   case PPCISD::PPC32_GOT:       return "PPCISD::PPC32_GOT";
1727   case PPCISD::PPC32_PICGOT:    return "PPCISD::PPC32_PICGOT";
1728   case PPCISD::ADDIS_GOT_TPREL_HA: return "PPCISD::ADDIS_GOT_TPREL_HA";
1729   case PPCISD::LD_GOT_TPREL_L:  return "PPCISD::LD_GOT_TPREL_L";
1730   case PPCISD::ADD_TLS:         return "PPCISD::ADD_TLS";
1731   case PPCISD::ADDIS_TLSGD_HA:  return "PPCISD::ADDIS_TLSGD_HA";
1732   case PPCISD::ADDI_TLSGD_L:    return "PPCISD::ADDI_TLSGD_L";
1733   case PPCISD::GET_TLS_ADDR:    return "PPCISD::GET_TLS_ADDR";
1734   case PPCISD::ADDI_TLSGD_L_ADDR: return "PPCISD::ADDI_TLSGD_L_ADDR";
1735   case PPCISD::TLSGD_AIX:       return "PPCISD::TLSGD_AIX";
1736   case PPCISD::ADDIS_TLSLD_HA:  return "PPCISD::ADDIS_TLSLD_HA";
1737   case PPCISD::ADDI_TLSLD_L:    return "PPCISD::ADDI_TLSLD_L";
1738   case PPCISD::GET_TLSLD_ADDR:  return "PPCISD::GET_TLSLD_ADDR";
1739   case PPCISD::ADDI_TLSLD_L_ADDR: return "PPCISD::ADDI_TLSLD_L_ADDR";
1740   case PPCISD::ADDIS_DTPREL_HA: return "PPCISD::ADDIS_DTPREL_HA";
1741   case PPCISD::ADDI_DTPREL_L:   return "PPCISD::ADDI_DTPREL_L";
1742   case PPCISD::PADDI_DTPREL:
1743     return "PPCISD::PADDI_DTPREL";
1744   case PPCISD::VADD_SPLAT:      return "PPCISD::VADD_SPLAT";
1745   case PPCISD::SC:              return "PPCISD::SC";
1746   case PPCISD::CLRBHRB:         return "PPCISD::CLRBHRB";
1747   case PPCISD::MFBHRBE:         return "PPCISD::MFBHRBE";
1748   case PPCISD::RFEBB:           return "PPCISD::RFEBB";
1749   case PPCISD::XXSWAPD:         return "PPCISD::XXSWAPD";
1750   case PPCISD::SWAP_NO_CHAIN:   return "PPCISD::SWAP_NO_CHAIN";
1751   case PPCISD::BUILD_FP128:     return "PPCISD::BUILD_FP128";
1752   case PPCISD::BUILD_SPE64:     return "PPCISD::BUILD_SPE64";
1753   case PPCISD::EXTRACT_SPE:     return "PPCISD::EXTRACT_SPE";
1754   case PPCISD::EXTSWSLI:        return "PPCISD::EXTSWSLI";
1755   case PPCISD::LD_VSX_LH:       return "PPCISD::LD_VSX_LH";
1756   case PPCISD::FP_EXTEND_HALF:  return "PPCISD::FP_EXTEND_HALF";
1757   case PPCISD::MAT_PCREL_ADDR:  return "PPCISD::MAT_PCREL_ADDR";
1758   case PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR:
1759     return "PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR";
1760   case PPCISD::TLS_LOCAL_EXEC_MAT_ADDR:
1761     return "PPCISD::TLS_LOCAL_EXEC_MAT_ADDR";
1762   case PPCISD::ACC_BUILD:       return "PPCISD::ACC_BUILD";
1763   case PPCISD::PAIR_BUILD:      return "PPCISD::PAIR_BUILD";
1764   case PPCISD::EXTRACT_VSX_REG: return "PPCISD::EXTRACT_VSX_REG";
1765   case PPCISD::XXMFACC:         return "PPCISD::XXMFACC";
1766   case PPCISD::LD_SPLAT:        return "PPCISD::LD_SPLAT";
1767   case PPCISD::ZEXT_LD_SPLAT:   return "PPCISD::ZEXT_LD_SPLAT";
1768   case PPCISD::SEXT_LD_SPLAT:   return "PPCISD::SEXT_LD_SPLAT";
1769   case PPCISD::FNMSUB:          return "PPCISD::FNMSUB";
1770   case PPCISD::STRICT_FADDRTZ:
1771     return "PPCISD::STRICT_FADDRTZ";
1772   case PPCISD::STRICT_FCTIDZ:
1773     return "PPCISD::STRICT_FCTIDZ";
1774   case PPCISD::STRICT_FCTIWZ:
1775     return "PPCISD::STRICT_FCTIWZ";
1776   case PPCISD::STRICT_FCTIDUZ:
1777     return "PPCISD::STRICT_FCTIDUZ";
1778   case PPCISD::STRICT_FCTIWUZ:
1779     return "PPCISD::STRICT_FCTIWUZ";
1780   case PPCISD::STRICT_FCFID:
1781     return "PPCISD::STRICT_FCFID";
1782   case PPCISD::STRICT_FCFIDU:
1783     return "PPCISD::STRICT_FCFIDU";
1784   case PPCISD::STRICT_FCFIDS:
1785     return "PPCISD::STRICT_FCFIDS";
1786   case PPCISD::STRICT_FCFIDUS:
1787     return "PPCISD::STRICT_FCFIDUS";
1788   case PPCISD::LXVRZX:          return "PPCISD::LXVRZX";
1789   case PPCISD::STORE_COND:
1790     return "PPCISD::STORE_COND";
1791   }
1792   return nullptr;
1793 }
1794 
1795 EVT PPCTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &C,
1796                                           EVT VT) const {
1797   if (!VT.isVector())
1798     return Subtarget.useCRBits() ? MVT::i1 : MVT::i32;
1799 
1800   return VT.changeVectorElementTypeToInteger();
1801 }
1802 
1803 bool PPCTargetLowering::enableAggressiveFMAFusion(EVT VT) const {
1804   assert(VT.isFloatingPoint() && "Non-floating-point FMA?");
1805   return true;
1806 }
1807 
1808 //===----------------------------------------------------------------------===//
1809 // Node matching predicates, for use by the tblgen matching code.
1810 //===----------------------------------------------------------------------===//
1811 
1812 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
1813 static bool isFloatingPointZero(SDValue Op) {
1814   if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1815     return CFP->getValueAPF().isZero();
1816   else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1817     // Maybe this has already been legalized into the constant pool?
1818     if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
1819       if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1820         return CFP->getValueAPF().isZero();
1821   }
1822   return false;
1823 }
1824 
1825 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode.  Return
1826 /// true if Op is undef or if it matches the specified value.
1827 static bool isConstantOrUndef(int Op, int Val) {
1828   return Op < 0 || Op == Val;
1829 }
1830 
1831 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
1832 /// VPKUHUM instruction.
1833 /// The ShuffleKind distinguishes between big-endian operations with
1834 /// two different inputs (0), either-endian operations with two identical
1835 /// inputs (1), and little-endian operations with two different inputs (2).
1836 /// For the latter, the input operands are swapped (see PPCInstrAltivec.td).
1837 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
1838                                SelectionDAG &DAG) {
1839   bool IsLE = DAG.getDataLayout().isLittleEndian();
1840   if (ShuffleKind == 0) {
1841     if (IsLE)
1842       return false;
1843     for (unsigned i = 0; i != 16; ++i)
1844       if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
1845         return false;
1846   } else if (ShuffleKind == 2) {
1847     if (!IsLE)
1848       return false;
1849     for (unsigned i = 0; i != 16; ++i)
1850       if (!isConstantOrUndef(N->getMaskElt(i), i*2))
1851         return false;
1852   } else if (ShuffleKind == 1) {
1853     unsigned j = IsLE ? 0 : 1;
1854     for (unsigned i = 0; i != 8; ++i)
1855       if (!isConstantOrUndef(N->getMaskElt(i),    i*2+j) ||
1856           !isConstantOrUndef(N->getMaskElt(i+8),  i*2+j))
1857         return false;
1858   }
1859   return true;
1860 }
1861 
1862 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
1863 /// VPKUWUM instruction.
1864 /// The ShuffleKind distinguishes between big-endian operations with
1865 /// two different inputs (0), either-endian operations with two identical
1866 /// inputs (1), and little-endian operations with two different inputs (2).
1867 /// For the latter, the input operands are swapped (see PPCInstrAltivec.td).
1868 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
1869                                SelectionDAG &DAG) {
1870   bool IsLE = DAG.getDataLayout().isLittleEndian();
1871   if (ShuffleKind == 0) {
1872     if (IsLE)
1873       return false;
1874     for (unsigned i = 0; i != 16; i += 2)
1875       if (!isConstantOrUndef(N->getMaskElt(i  ),  i*2+2) ||
1876           !isConstantOrUndef(N->getMaskElt(i+1),  i*2+3))
1877         return false;
1878   } else if (ShuffleKind == 2) {
1879     if (!IsLE)
1880       return false;
1881     for (unsigned i = 0; i != 16; i += 2)
1882       if (!isConstantOrUndef(N->getMaskElt(i  ),  i*2) ||
1883           !isConstantOrUndef(N->getMaskElt(i+1),  i*2+1))
1884         return false;
1885   } else if (ShuffleKind == 1) {
1886     unsigned j = IsLE ? 0 : 2;
1887     for (unsigned i = 0; i != 8; i += 2)
1888       if (!isConstantOrUndef(N->getMaskElt(i  ),  i*2+j)   ||
1889           !isConstantOrUndef(N->getMaskElt(i+1),  i*2+j+1) ||
1890           !isConstantOrUndef(N->getMaskElt(i+8),  i*2+j)   ||
1891           !isConstantOrUndef(N->getMaskElt(i+9),  i*2+j+1))
1892         return false;
1893   }
1894   return true;
1895 }
1896 
1897 /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a
1898 /// VPKUDUM instruction, AND the VPKUDUM instruction exists for the
1899 /// current subtarget.
1900 ///
1901 /// The ShuffleKind distinguishes between big-endian operations with
1902 /// two different inputs (0), either-endian operations with two identical
1903 /// inputs (1), and little-endian operations with two different inputs (2).
1904 /// For the latter, the input operands are swapped (see PPCInstrAltivec.td).
1905 bool PPC::isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
1906                                SelectionDAG &DAG) {
1907   const PPCSubtarget &Subtarget = DAG.getSubtarget<PPCSubtarget>();
1908   if (!Subtarget.hasP8Vector())
1909     return false;
1910 
1911   bool IsLE = DAG.getDataLayout().isLittleEndian();
1912   if (ShuffleKind == 0) {
1913     if (IsLE)
1914       return false;
1915     for (unsigned i = 0; i != 16; i += 4)
1916       if (!isConstantOrUndef(N->getMaskElt(i  ),  i*2+4) ||
1917           !isConstantOrUndef(N->getMaskElt(i+1),  i*2+5) ||
1918           !isConstantOrUndef(N->getMaskElt(i+2),  i*2+6) ||
1919           !isConstantOrUndef(N->getMaskElt(i+3),  i*2+7))
1920         return false;
1921   } else if (ShuffleKind == 2) {
1922     if (!IsLE)
1923       return false;
1924     for (unsigned i = 0; i != 16; i += 4)
1925       if (!isConstantOrUndef(N->getMaskElt(i  ),  i*2) ||
1926           !isConstantOrUndef(N->getMaskElt(i+1),  i*2+1) ||
1927           !isConstantOrUndef(N->getMaskElt(i+2),  i*2+2) ||
1928           !isConstantOrUndef(N->getMaskElt(i+3),  i*2+3))
1929         return false;
1930   } else if (ShuffleKind == 1) {
1931     unsigned j = IsLE ? 0 : 4;
1932     for (unsigned i = 0; i != 8; i += 4)
1933       if (!isConstantOrUndef(N->getMaskElt(i  ),  i*2+j)   ||
1934           !isConstantOrUndef(N->getMaskElt(i+1),  i*2+j+1) ||
1935           !isConstantOrUndef(N->getMaskElt(i+2),  i*2+j+2) ||
1936           !isConstantOrUndef(N->getMaskElt(i+3),  i*2+j+3) ||
1937           !isConstantOrUndef(N->getMaskElt(i+8),  i*2+j)   ||
1938           !isConstantOrUndef(N->getMaskElt(i+9),  i*2+j+1) ||
1939           !isConstantOrUndef(N->getMaskElt(i+10), i*2+j+2) ||
1940           !isConstantOrUndef(N->getMaskElt(i+11), i*2+j+3))
1941         return false;
1942   }
1943   return true;
1944 }
1945 
1946 /// isVMerge - Common function, used to match vmrg* shuffles.
1947 ///
1948 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
1949                      unsigned LHSStart, unsigned RHSStart) {
1950   if (N->getValueType(0) != MVT::v16i8)
1951     return false;
1952   assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
1953          "Unsupported merge size!");
1954 
1955   for (unsigned i = 0; i != 8/UnitSize; ++i)     // Step over units
1956     for (unsigned j = 0; j != UnitSize; ++j) {   // Step over bytes within unit
1957       if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
1958                              LHSStart+j+i*UnitSize) ||
1959           !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
1960                              RHSStart+j+i*UnitSize))
1961         return false;
1962     }
1963   return true;
1964 }
1965 
1966 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
1967 /// a VMRGL* instruction with the specified unit size (1,2 or 4 bytes).
1968 /// The ShuffleKind distinguishes between big-endian merges with two
1969 /// different inputs (0), either-endian merges with two identical inputs (1),
1970 /// and little-endian merges with two different inputs (2).  For the latter,
1971 /// the input operands are swapped (see PPCInstrAltivec.td).
1972 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
1973                              unsigned ShuffleKind, SelectionDAG &DAG) {
1974   if (DAG.getDataLayout().isLittleEndian()) {
1975     if (ShuffleKind == 1) // unary
1976       return isVMerge(N, UnitSize, 0, 0);
1977     else if (ShuffleKind == 2) // swapped
1978       return isVMerge(N, UnitSize, 0, 16);
1979     else
1980       return false;
1981   } else {
1982     if (ShuffleKind == 1) // unary
1983       return isVMerge(N, UnitSize, 8, 8);
1984     else if (ShuffleKind == 0) // normal
1985       return isVMerge(N, UnitSize, 8, 24);
1986     else
1987       return false;
1988   }
1989 }
1990 
1991 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
1992 /// a VMRGH* instruction with the specified unit size (1,2 or 4 bytes).
1993 /// The ShuffleKind distinguishes between big-endian merges with two
1994 /// different inputs (0), either-endian merges with two identical inputs (1),
1995 /// and little-endian merges with two different inputs (2).  For the latter,
1996 /// the input operands are swapped (see PPCInstrAltivec.td).
1997 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
1998                              unsigned ShuffleKind, SelectionDAG &DAG) {
1999   if (DAG.getDataLayout().isLittleEndian()) {
2000     if (ShuffleKind == 1) // unary
2001       return isVMerge(N, UnitSize, 8, 8);
2002     else if (ShuffleKind == 2) // swapped
2003       return isVMerge(N, UnitSize, 8, 24);
2004     else
2005       return false;
2006   } else {
2007     if (ShuffleKind == 1) // unary
2008       return isVMerge(N, UnitSize, 0, 0);
2009     else if (ShuffleKind == 0) // normal
2010       return isVMerge(N, UnitSize, 0, 16);
2011     else
2012       return false;
2013   }
2014 }
2015 
2016 /**
2017  * Common function used to match vmrgew and vmrgow shuffles
2018  *
2019  * The indexOffset determines whether to look for even or odd words in
2020  * the shuffle mask. This is based on the of the endianness of the target
2021  * machine.
2022  *   - Little Endian:
2023  *     - Use offset of 0 to check for odd elements
2024  *     - Use offset of 4 to check for even elements
2025  *   - Big Endian:
2026  *     - Use offset of 0 to check for even elements
2027  *     - Use offset of 4 to check for odd elements
2028  * A detailed description of the vector element ordering for little endian and
2029  * big endian can be found at
2030  * http://www.ibm.com/developerworks/library/l-ibm-xl-c-cpp-compiler/index.html
2031  * Targeting your applications - what little endian and big endian IBM XL C/C++
2032  * compiler differences mean to you
2033  *
2034  * The mask to the shuffle vector instruction specifies the indices of the
2035  * elements from the two input vectors to place in the result. The elements are
2036  * numbered in array-access order, starting with the first vector. These vectors
2037  * are always of type v16i8, thus each vector will contain 16 elements of size
2038  * 8. More info on the shuffle vector can be found in the
2039  * http://llvm.org/docs/LangRef.html#shufflevector-instruction
2040  * Language Reference.
2041  *
2042  * The RHSStartValue indicates whether the same input vectors are used (unary)
2043  * or two different input vectors are used, based on the following:
2044  *   - If the instruction uses the same vector for both inputs, the range of the
2045  *     indices will be 0 to 15. In this case, the RHSStart value passed should
2046  *     be 0.
2047  *   - If the instruction has two different vectors then the range of the
2048  *     indices will be 0 to 31. In this case, the RHSStart value passed should
2049  *     be 16 (indices 0-15 specify elements in the first vector while indices 16
2050  *     to 31 specify elements in the second vector).
2051  *
2052  * \param[in] N The shuffle vector SD Node to analyze
2053  * \param[in] IndexOffset Specifies whether to look for even or odd elements
2054  * \param[in] RHSStartValue Specifies the starting index for the righthand input
2055  * vector to the shuffle_vector instruction
2056  * \return true iff this shuffle vector represents an even or odd word merge
2057  */
2058 static bool isVMerge(ShuffleVectorSDNode *N, unsigned IndexOffset,
2059                      unsigned RHSStartValue) {
2060   if (N->getValueType(0) != MVT::v16i8)
2061     return false;
2062 
2063   for (unsigned i = 0; i < 2; ++i)
2064     for (unsigned j = 0; j < 4; ++j)
2065       if (!isConstantOrUndef(N->getMaskElt(i*4+j),
2066                              i*RHSStartValue+j+IndexOffset) ||
2067           !isConstantOrUndef(N->getMaskElt(i*4+j+8),
2068                              i*RHSStartValue+j+IndexOffset+8))
2069         return false;
2070   return true;
2071 }
2072 
2073 /**
2074  * Determine if the specified shuffle mask is suitable for the vmrgew or
2075  * vmrgow instructions.
2076  *
2077  * \param[in] N The shuffle vector SD Node to analyze
2078  * \param[in] CheckEven Check for an even merge (true) or an odd merge (false)
2079  * \param[in] ShuffleKind Identify the type of merge:
2080  *   - 0 = big-endian merge with two different inputs;
2081  *   - 1 = either-endian merge with two identical inputs;
2082  *   - 2 = little-endian merge with two different inputs (inputs are swapped for
2083  *     little-endian merges).
2084  * \param[in] DAG The current SelectionDAG
2085  * \return true iff this shuffle mask
2086  */
2087 bool PPC::isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven,
2088                               unsigned ShuffleKind, SelectionDAG &DAG) {
2089   if (DAG.getDataLayout().isLittleEndian()) {
2090     unsigned indexOffset = CheckEven ? 4 : 0;
2091     if (ShuffleKind == 1) // Unary
2092       return isVMerge(N, indexOffset, 0);
2093     else if (ShuffleKind == 2) // swapped
2094       return isVMerge(N, indexOffset, 16);
2095     else
2096       return false;
2097   }
2098   else {
2099     unsigned indexOffset = CheckEven ? 0 : 4;
2100     if (ShuffleKind == 1) // Unary
2101       return isVMerge(N, indexOffset, 0);
2102     else if (ShuffleKind == 0) // Normal
2103       return isVMerge(N, indexOffset, 16);
2104     else
2105       return false;
2106   }
2107   return false;
2108 }
2109 
2110 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
2111 /// amount, otherwise return -1.
2112 /// The ShuffleKind distinguishes between big-endian operations with two
2113 /// different inputs (0), either-endian operations with two identical inputs
2114 /// (1), and little-endian operations with two different inputs (2).  For the
2115 /// latter, the input operands are swapped (see PPCInstrAltivec.td).
2116 int PPC::isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
2117                              SelectionDAG &DAG) {
2118   if (N->getValueType(0) != MVT::v16i8)
2119     return -1;
2120 
2121   ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2122 
2123   // Find the first non-undef value in the shuffle mask.
2124   unsigned i;
2125   for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
2126     /*search*/;
2127 
2128   if (i == 16) return -1;  // all undef.
2129 
2130   // Otherwise, check to see if the rest of the elements are consecutively
2131   // numbered from this value.
2132   unsigned ShiftAmt = SVOp->getMaskElt(i);
2133   if (ShiftAmt < i) return -1;
2134 
2135   ShiftAmt -= i;
2136   bool isLE = DAG.getDataLayout().isLittleEndian();
2137 
2138   if ((ShuffleKind == 0 && !isLE) || (ShuffleKind == 2 && isLE)) {
2139     // Check the rest of the elements to see if they are consecutive.
2140     for (++i; i != 16; ++i)
2141       if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
2142         return -1;
2143   } else if (ShuffleKind == 1) {
2144     // Check the rest of the elements to see if they are consecutive.
2145     for (++i; i != 16; ++i)
2146       if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
2147         return -1;
2148   } else
2149     return -1;
2150 
2151   if (isLE)
2152     ShiftAmt = 16 - ShiftAmt;
2153 
2154   return ShiftAmt;
2155 }
2156 
2157 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
2158 /// specifies a splat of a single element that is suitable for input to
2159 /// one of the splat operations (VSPLTB/VSPLTH/VSPLTW/XXSPLTW/LXVDSX/etc.).
2160 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
2161   EVT VT = N->getValueType(0);
2162   if (VT == MVT::v2i64 || VT == MVT::v2f64)
2163     return EltSize == 8 && N->getMaskElt(0) == N->getMaskElt(1);
2164 
2165   assert(VT == MVT::v16i8 && isPowerOf2_32(EltSize) &&
2166          EltSize <= 8 && "Can only handle 1,2,4,8 byte element sizes");
2167 
2168   // The consecutive indices need to specify an element, not part of two
2169   // different elements.  So abandon ship early if this isn't the case.
2170   if (N->getMaskElt(0) % EltSize != 0)
2171     return false;
2172 
2173   // This is a splat operation if each element of the permute is the same, and
2174   // if the value doesn't reference the second vector.
2175   unsigned ElementBase = N->getMaskElt(0);
2176 
2177   // FIXME: Handle UNDEF elements too!
2178   if (ElementBase >= 16)
2179     return false;
2180 
2181   // Check that the indices are consecutive, in the case of a multi-byte element
2182   // splatted with a v16i8 mask.
2183   for (unsigned i = 1; i != EltSize; ++i)
2184     if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
2185       return false;
2186 
2187   for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
2188     if (N->getMaskElt(i) < 0) continue;
2189     for (unsigned j = 0; j != EltSize; ++j)
2190       if (N->getMaskElt(i+j) != N->getMaskElt(j))
2191         return false;
2192   }
2193   return true;
2194 }
2195 
2196 /// Check that the mask is shuffling N byte elements. Within each N byte
2197 /// element of the mask, the indices could be either in increasing or
2198 /// decreasing order as long as they are consecutive.
2199 /// \param[in] N the shuffle vector SD Node to analyze
2200 /// \param[in] Width the element width in bytes, could be 2/4/8/16 (HalfWord/
2201 /// Word/DoubleWord/QuadWord).
2202 /// \param[in] StepLen the delta indices number among the N byte element, if
2203 /// the mask is in increasing/decreasing order then it is 1/-1.
2204 /// \return true iff the mask is shuffling N byte elements.
2205 static bool isNByteElemShuffleMask(ShuffleVectorSDNode *N, unsigned Width,
2206                                    int StepLen) {
2207   assert((Width == 2 || Width == 4 || Width == 8 || Width == 16) &&
2208          "Unexpected element width.");
2209   assert((StepLen == 1 || StepLen == -1) && "Unexpected element width.");
2210 
2211   unsigned NumOfElem = 16 / Width;
2212   unsigned MaskVal[16]; //  Width is never greater than 16
2213   for (unsigned i = 0; i < NumOfElem; ++i) {
2214     MaskVal[0] = N->getMaskElt(i * Width);
2215     if ((StepLen == 1) && (MaskVal[0] % Width)) {
2216       return false;
2217     } else if ((StepLen == -1) && ((MaskVal[0] + 1) % Width)) {
2218       return false;
2219     }
2220 
2221     for (unsigned int j = 1; j < Width; ++j) {
2222       MaskVal[j] = N->getMaskElt(i * Width + j);
2223       if (MaskVal[j] != MaskVal[j-1] + StepLen) {
2224         return false;
2225       }
2226     }
2227   }
2228 
2229   return true;
2230 }
2231 
2232 bool PPC::isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
2233                           unsigned &InsertAtByte, bool &Swap, bool IsLE) {
2234   if (!isNByteElemShuffleMask(N, 4, 1))
2235     return false;
2236 
2237   // Now we look at mask elements 0,4,8,12
2238   unsigned M0 = N->getMaskElt(0) / 4;
2239   unsigned M1 = N->getMaskElt(4) / 4;
2240   unsigned M2 = N->getMaskElt(8) / 4;
2241   unsigned M3 = N->getMaskElt(12) / 4;
2242   unsigned LittleEndianShifts[] = { 2, 1, 0, 3 };
2243   unsigned BigEndianShifts[] = { 3, 0, 1, 2 };
2244 
2245   // Below, let H and L be arbitrary elements of the shuffle mask
2246   // where H is in the range [4,7] and L is in the range [0,3].
2247   // H, 1, 2, 3 or L, 5, 6, 7
2248   if ((M0 > 3 && M1 == 1 && M2 == 2 && M3 == 3) ||
2249       (M0 < 4 && M1 == 5 && M2 == 6 && M3 == 7)) {
2250     ShiftElts = IsLE ? LittleEndianShifts[M0 & 0x3] : BigEndianShifts[M0 & 0x3];
2251     InsertAtByte = IsLE ? 12 : 0;
2252     Swap = M0 < 4;
2253     return true;
2254   }
2255   // 0, H, 2, 3 or 4, L, 6, 7
2256   if ((M1 > 3 && M0 == 0 && M2 == 2 && M3 == 3) ||
2257       (M1 < 4 && M0 == 4 && M2 == 6 && M3 == 7)) {
2258     ShiftElts = IsLE ? LittleEndianShifts[M1 & 0x3] : BigEndianShifts[M1 & 0x3];
2259     InsertAtByte = IsLE ? 8 : 4;
2260     Swap = M1 < 4;
2261     return true;
2262   }
2263   // 0, 1, H, 3 or 4, 5, L, 7
2264   if ((M2 > 3 && M0 == 0 && M1 == 1 && M3 == 3) ||
2265       (M2 < 4 && M0 == 4 && M1 == 5 && M3 == 7)) {
2266     ShiftElts = IsLE ? LittleEndianShifts[M2 & 0x3] : BigEndianShifts[M2 & 0x3];
2267     InsertAtByte = IsLE ? 4 : 8;
2268     Swap = M2 < 4;
2269     return true;
2270   }
2271   // 0, 1, 2, H or 4, 5, 6, L
2272   if ((M3 > 3 && M0 == 0 && M1 == 1 && M2 == 2) ||
2273       (M3 < 4 && M0 == 4 && M1 == 5 && M2 == 6)) {
2274     ShiftElts = IsLE ? LittleEndianShifts[M3 & 0x3] : BigEndianShifts[M3 & 0x3];
2275     InsertAtByte = IsLE ? 0 : 12;
2276     Swap = M3 < 4;
2277     return true;
2278   }
2279 
2280   // If both vector operands for the shuffle are the same vector, the mask will
2281   // contain only elements from the first one and the second one will be undef.
2282   if (N->getOperand(1).isUndef()) {
2283     ShiftElts = 0;
2284     Swap = true;
2285     unsigned XXINSERTWSrcElem = IsLE ? 2 : 1;
2286     if (M0 == XXINSERTWSrcElem && M1 == 1 && M2 == 2 && M3 == 3) {
2287       InsertAtByte = IsLE ? 12 : 0;
2288       return true;
2289     }
2290     if (M0 == 0 && M1 == XXINSERTWSrcElem && M2 == 2 && M3 == 3) {
2291       InsertAtByte = IsLE ? 8 : 4;
2292       return true;
2293     }
2294     if (M0 == 0 && M1 == 1 && M2 == XXINSERTWSrcElem && M3 == 3) {
2295       InsertAtByte = IsLE ? 4 : 8;
2296       return true;
2297     }
2298     if (M0 == 0 && M1 == 1 && M2 == 2 && M3 == XXINSERTWSrcElem) {
2299       InsertAtByte = IsLE ? 0 : 12;
2300       return true;
2301     }
2302   }
2303 
2304   return false;
2305 }
2306 
2307 bool PPC::isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
2308                                bool &Swap, bool IsLE) {
2309   assert(N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8");
2310   // Ensure each byte index of the word is consecutive.
2311   if (!isNByteElemShuffleMask(N, 4, 1))
2312     return false;
2313 
2314   // Now we look at mask elements 0,4,8,12, which are the beginning of words.
2315   unsigned M0 = N->getMaskElt(0) / 4;
2316   unsigned M1 = N->getMaskElt(4) / 4;
2317   unsigned M2 = N->getMaskElt(8) / 4;
2318   unsigned M3 = N->getMaskElt(12) / 4;
2319 
2320   // If both vector operands for the shuffle are the same vector, the mask will
2321   // contain only elements from the first one and the second one will be undef.
2322   if (N->getOperand(1).isUndef()) {
2323     assert(M0 < 4 && "Indexing into an undef vector?");
2324     if (M1 != (M0 + 1) % 4 || M2 != (M1 + 1) % 4 || M3 != (M2 + 1) % 4)
2325       return false;
2326 
2327     ShiftElts = IsLE ? (4 - M0) % 4 : M0;
2328     Swap = false;
2329     return true;
2330   }
2331 
2332   // Ensure each word index of the ShuffleVector Mask is consecutive.
2333   if (M1 != (M0 + 1) % 8 || M2 != (M1 + 1) % 8 || M3 != (M2 + 1) % 8)
2334     return false;
2335 
2336   if (IsLE) {
2337     if (M0 == 0 || M0 == 7 || M0 == 6 || M0 == 5) {
2338       // Input vectors don't need to be swapped if the leading element
2339       // of the result is one of the 3 left elements of the second vector
2340       // (or if there is no shift to be done at all).
2341       Swap = false;
2342       ShiftElts = (8 - M0) % 8;
2343     } else if (M0 == 4 || M0 == 3 || M0 == 2 || M0 == 1) {
2344       // Input vectors need to be swapped if the leading element
2345       // of the result is one of the 3 left elements of the first vector
2346       // (or if we're shifting by 4 - thereby simply swapping the vectors).
2347       Swap = true;
2348       ShiftElts = (4 - M0) % 4;
2349     }
2350 
2351     return true;
2352   } else {                                          // BE
2353     if (M0 == 0 || M0 == 1 || M0 == 2 || M0 == 3) {
2354       // Input vectors don't need to be swapped if the leading element
2355       // of the result is one of the 4 elements of the first vector.
2356       Swap = false;
2357       ShiftElts = M0;
2358     } else if (M0 == 4 || M0 == 5 || M0 == 6 || M0 == 7) {
2359       // Input vectors need to be swapped if the leading element
2360       // of the result is one of the 4 elements of the right vector.
2361       Swap = true;
2362       ShiftElts = M0 - 4;
2363     }
2364 
2365     return true;
2366   }
2367 }
2368 
2369 bool static isXXBRShuffleMaskHelper(ShuffleVectorSDNode *N, int Width) {
2370   assert(N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8");
2371 
2372   if (!isNByteElemShuffleMask(N, Width, -1))
2373     return false;
2374 
2375   for (int i = 0; i < 16; i += Width)
2376     if (N->getMaskElt(i) != i + Width - 1)
2377       return false;
2378 
2379   return true;
2380 }
2381 
2382 bool PPC::isXXBRHShuffleMask(ShuffleVectorSDNode *N) {
2383   return isXXBRShuffleMaskHelper(N, 2);
2384 }
2385 
2386 bool PPC::isXXBRWShuffleMask(ShuffleVectorSDNode *N) {
2387   return isXXBRShuffleMaskHelper(N, 4);
2388 }
2389 
2390 bool PPC::isXXBRDShuffleMask(ShuffleVectorSDNode *N) {
2391   return isXXBRShuffleMaskHelper(N, 8);
2392 }
2393 
2394 bool PPC::isXXBRQShuffleMask(ShuffleVectorSDNode *N) {
2395   return isXXBRShuffleMaskHelper(N, 16);
2396 }
2397 
2398 /// Can node \p N be lowered to an XXPERMDI instruction? If so, set \p Swap
2399 /// if the inputs to the instruction should be swapped and set \p DM to the
2400 /// value for the immediate.
2401 /// Specifically, set \p Swap to true only if \p N can be lowered to XXPERMDI
2402 /// AND element 0 of the result comes from the first input (LE) or second input
2403 /// (BE). Set \p DM to the calculated result (0-3) only if \p N can be lowered.
2404 /// \return true iff the given mask of shuffle node \p N is a XXPERMDI shuffle
2405 /// mask.
2406 bool PPC::isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &DM,
2407                                bool &Swap, bool IsLE) {
2408   assert(N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8");
2409 
2410   // Ensure each byte index of the double word is consecutive.
2411   if (!isNByteElemShuffleMask(N, 8, 1))
2412     return false;
2413 
2414   unsigned M0 = N->getMaskElt(0) / 8;
2415   unsigned M1 = N->getMaskElt(8) / 8;
2416   assert(((M0 | M1) < 4) && "A mask element out of bounds?");
2417 
2418   // If both vector operands for the shuffle are the same vector, the mask will
2419   // contain only elements from the first one and the second one will be undef.
2420   if (N->getOperand(1).isUndef()) {
2421     if ((M0 | M1) < 2) {
2422       DM = IsLE ? (((~M1) & 1) << 1) + ((~M0) & 1) : (M0 << 1) + (M1 & 1);
2423       Swap = false;
2424       return true;
2425     } else
2426       return false;
2427   }
2428 
2429   if (IsLE) {
2430     if (M0 > 1 && M1 < 2) {
2431       Swap = false;
2432     } else if (M0 < 2 && M1 > 1) {
2433       M0 = (M0 + 2) % 4;
2434       M1 = (M1 + 2) % 4;
2435       Swap = true;
2436     } else
2437       return false;
2438 
2439     // Note: if control flow comes here that means Swap is already set above
2440     DM = (((~M1) & 1) << 1) + ((~M0) & 1);
2441     return true;
2442   } else { // BE
2443     if (M0 < 2 && M1 > 1) {
2444       Swap = false;
2445     } else if (M0 > 1 && M1 < 2) {
2446       M0 = (M0 + 2) % 4;
2447       M1 = (M1 + 2) % 4;
2448       Swap = true;
2449     } else
2450       return false;
2451 
2452     // Note: if control flow comes here that means Swap is already set above
2453     DM = (M0 << 1) + (M1 & 1);
2454     return true;
2455   }
2456 }
2457 
2458 
2459 /// getSplatIdxForPPCMnemonics - Return the splat index as a value that is
2460 /// appropriate for PPC mnemonics (which have a big endian bias - namely
2461 /// elements are counted from the left of the vector register).
2462 unsigned PPC::getSplatIdxForPPCMnemonics(SDNode *N, unsigned EltSize,
2463                                          SelectionDAG &DAG) {
2464   ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2465   assert(isSplatShuffleMask(SVOp, EltSize));
2466   EVT VT = SVOp->getValueType(0);
2467 
2468   if (VT == MVT::v2i64 || VT == MVT::v2f64)
2469     return DAG.getDataLayout().isLittleEndian() ? 1 - SVOp->getMaskElt(0)
2470                                                 : SVOp->getMaskElt(0);
2471 
2472   if (DAG.getDataLayout().isLittleEndian())
2473     return (16 / EltSize) - 1 - (SVOp->getMaskElt(0) / EltSize);
2474   else
2475     return SVOp->getMaskElt(0) / EltSize;
2476 }
2477 
2478 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
2479 /// by using a vspltis[bhw] instruction of the specified element size, return
2480 /// the constant being splatted.  The ByteSize field indicates the number of
2481 /// bytes of each element [124] -> [bhw].
2482 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2483   SDValue OpVal;
2484 
2485   // If ByteSize of the splat is bigger than the element size of the
2486   // build_vector, then we have a case where we are checking for a splat where
2487   // multiple elements of the buildvector are folded together into a single
2488   // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
2489   unsigned EltSize = 16/N->getNumOperands();
2490   if (EltSize < ByteSize) {
2491     unsigned Multiple = ByteSize/EltSize;   // Number of BV entries per spltval.
2492     SDValue UniquedVals[4];
2493     assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
2494 
2495     // See if all of the elements in the buildvector agree across.
2496     for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2497       if (N->getOperand(i).isUndef()) continue;
2498       // If the element isn't a constant, bail fully out.
2499       if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
2500 
2501       if (!UniquedVals[i&(Multiple-1)].getNode())
2502         UniquedVals[i&(Multiple-1)] = N->getOperand(i);
2503       else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
2504         return SDValue();  // no match.
2505     }
2506 
2507     // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
2508     // either constant or undef values that are identical for each chunk.  See
2509     // if these chunks can form into a larger vspltis*.
2510 
2511     // Check to see if all of the leading entries are either 0 or -1.  If
2512     // neither, then this won't fit into the immediate field.
2513     bool LeadingZero = true;
2514     bool LeadingOnes = true;
2515     for (unsigned i = 0; i != Multiple-1; ++i) {
2516       if (!UniquedVals[i].getNode()) continue;  // Must have been undefs.
2517 
2518       LeadingZero &= isNullConstant(UniquedVals[i]);
2519       LeadingOnes &= isAllOnesConstant(UniquedVals[i]);
2520     }
2521     // Finally, check the least significant entry.
2522     if (LeadingZero) {
2523       if (!UniquedVals[Multiple-1].getNode())
2524         return DAG.getTargetConstant(0, SDLoc(N), MVT::i32);  // 0,0,0,undef
2525       int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
2526       if (Val < 16)                                   // 0,0,0,4 -> vspltisw(4)
2527         return DAG.getTargetConstant(Val, SDLoc(N), MVT::i32);
2528     }
2529     if (LeadingOnes) {
2530       if (!UniquedVals[Multiple-1].getNode())
2531         return DAG.getTargetConstant(~0U, SDLoc(N), MVT::i32); // -1,-1,-1,undef
2532       int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
2533       if (Val >= -16)                            // -1,-1,-1,-2 -> vspltisw(-2)
2534         return DAG.getTargetConstant(Val, SDLoc(N), MVT::i32);
2535     }
2536 
2537     return SDValue();
2538   }
2539 
2540   // Check to see if this buildvec has a single non-undef value in its elements.
2541   for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2542     if (N->getOperand(i).isUndef()) continue;
2543     if (!OpVal.getNode())
2544       OpVal = N->getOperand(i);
2545     else if (OpVal != N->getOperand(i))
2546       return SDValue();
2547   }
2548 
2549   if (!OpVal.getNode()) return SDValue();  // All UNDEF: use implicit def.
2550 
2551   unsigned ValSizeInBytes = EltSize;
2552   uint64_t Value = 0;
2553   if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
2554     Value = CN->getZExtValue();
2555   } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
2556     assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
2557     Value = FloatToBits(CN->getValueAPF().convertToFloat());
2558   }
2559 
2560   // If the splat value is larger than the element value, then we can never do
2561   // this splat.  The only case that we could fit the replicated bits into our
2562   // immediate field for would be zero, and we prefer to use vxor for it.
2563   if (ValSizeInBytes < ByteSize) return SDValue();
2564 
2565   // If the element value is larger than the splat value, check if it consists
2566   // of a repeated bit pattern of size ByteSize.
2567   if (!APInt(ValSizeInBytes * 8, Value).isSplat(ByteSize * 8))
2568     return SDValue();
2569 
2570   // Properly sign extend the value.
2571   int MaskVal = SignExtend32(Value, ByteSize * 8);
2572 
2573   // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
2574   if (MaskVal == 0) return SDValue();
2575 
2576   // Finally, if this value fits in a 5 bit sext field, return it
2577   if (SignExtend32<5>(MaskVal) == MaskVal)
2578     return DAG.getTargetConstant(MaskVal, SDLoc(N), MVT::i32);
2579   return SDValue();
2580 }
2581 
2582 //===----------------------------------------------------------------------===//
2583 //  Addressing Mode Selection
2584 //===----------------------------------------------------------------------===//
2585 
2586 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
2587 /// or 64-bit immediate, and if the value can be accurately represented as a
2588 /// sign extension from a 16-bit value.  If so, this returns true and the
2589 /// immediate.
2590 bool llvm::isIntS16Immediate(SDNode *N, int16_t &Imm) {
2591   if (!isa<ConstantSDNode>(N))
2592     return false;
2593 
2594   Imm = (int16_t)cast<ConstantSDNode>(N)->getZExtValue();
2595   if (N->getValueType(0) == MVT::i32)
2596     return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
2597   else
2598     return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
2599 }
2600 bool llvm::isIntS16Immediate(SDValue Op, int16_t &Imm) {
2601   return isIntS16Immediate(Op.getNode(), Imm);
2602 }
2603 
2604 /// Used when computing address flags for selecting loads and stores.
2605 /// If we have an OR, check if the LHS and RHS are provably disjoint.
2606 /// An OR of two provably disjoint values is equivalent to an ADD.
2607 /// Most PPC load/store instructions compute the effective address as a sum,
2608 /// so doing this conversion is useful.
2609 static bool provablyDisjointOr(SelectionDAG &DAG, const SDValue &N) {
2610   if (N.getOpcode() != ISD::OR)
2611     return false;
2612   KnownBits LHSKnown = DAG.computeKnownBits(N.getOperand(0));
2613   if (!LHSKnown.Zero.getBoolValue())
2614     return false;
2615   KnownBits RHSKnown = DAG.computeKnownBits(N.getOperand(1));
2616   return (~(LHSKnown.Zero | RHSKnown.Zero) == 0);
2617 }
2618 
2619 /// SelectAddressEVXRegReg - Given the specified address, check to see if it can
2620 /// be represented as an indexed [r+r] operation.
2621 bool PPCTargetLowering::SelectAddressEVXRegReg(SDValue N, SDValue &Base,
2622                                                SDValue &Index,
2623                                                SelectionDAG &DAG) const {
2624   for (SDNode *U : N->uses()) {
2625     if (MemSDNode *Memop = dyn_cast<MemSDNode>(U)) {
2626       if (Memop->getMemoryVT() == MVT::f64) {
2627           Base = N.getOperand(0);
2628           Index = N.getOperand(1);
2629           return true;
2630       }
2631     }
2632   }
2633   return false;
2634 }
2635 
2636 /// isIntS34Immediate - This method tests if value of node given can be
2637 /// accurately represented as a sign extension from a 34-bit value.  If so,
2638 /// this returns true and the immediate.
2639 bool llvm::isIntS34Immediate(SDNode *N, int64_t &Imm) {
2640   if (!isa<ConstantSDNode>(N))
2641     return false;
2642 
2643   Imm = (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
2644   return isInt<34>(Imm);
2645 }
2646 bool llvm::isIntS34Immediate(SDValue Op, int64_t &Imm) {
2647   return isIntS34Immediate(Op.getNode(), Imm);
2648 }
2649 
2650 /// SelectAddressRegReg - Given the specified addressed, check to see if it
2651 /// can be represented as an indexed [r+r] operation.  Returns false if it
2652 /// can be more efficiently represented as [r+imm]. If \p EncodingAlignment is
2653 /// non-zero and N can be represented by a base register plus a signed 16-bit
2654 /// displacement, make a more precise judgement by checking (displacement % \p
2655 /// EncodingAlignment).
2656 bool PPCTargetLowering::SelectAddressRegReg(
2657     SDValue N, SDValue &Base, SDValue &Index, SelectionDAG &DAG,
2658     MaybeAlign EncodingAlignment) const {
2659   // If we have a PC Relative target flag don't select as [reg+reg]. It will be
2660   // a [pc+imm].
2661   if (SelectAddressPCRel(N, Base))
2662     return false;
2663 
2664   int16_t Imm = 0;
2665   if (N.getOpcode() == ISD::ADD) {
2666     // Is there any SPE load/store (f64), which can't handle 16bit offset?
2667     // SPE load/store can only handle 8-bit offsets.
2668     if (hasSPE() && SelectAddressEVXRegReg(N, Base, Index, DAG))
2669         return true;
2670     if (isIntS16Immediate(N.getOperand(1), Imm) &&
2671         (!EncodingAlignment || isAligned(*EncodingAlignment, Imm)))
2672       return false; // r+i
2673     if (N.getOperand(1).getOpcode() == PPCISD::Lo)
2674       return false;    // r+i
2675 
2676     Base = N.getOperand(0);
2677     Index = N.getOperand(1);
2678     return true;
2679   } else if (N.getOpcode() == ISD::OR) {
2680     if (isIntS16Immediate(N.getOperand(1), Imm) &&
2681         (!EncodingAlignment || isAligned(*EncodingAlignment, Imm)))
2682       return false; // r+i can fold it if we can.
2683 
2684     // If this is an or of disjoint bitfields, we can codegen this as an add
2685     // (for better address arithmetic) if the LHS and RHS of the OR are provably
2686     // disjoint.
2687     KnownBits LHSKnown = DAG.computeKnownBits(N.getOperand(0));
2688 
2689     if (LHSKnown.Zero.getBoolValue()) {
2690       KnownBits RHSKnown = DAG.computeKnownBits(N.getOperand(1));
2691       // If all of the bits are known zero on the LHS or RHS, the add won't
2692       // carry.
2693       if (~(LHSKnown.Zero | RHSKnown.Zero) == 0) {
2694         Base = N.getOperand(0);
2695         Index = N.getOperand(1);
2696         return true;
2697       }
2698     }
2699   }
2700 
2701   return false;
2702 }
2703 
2704 // If we happen to be doing an i64 load or store into a stack slot that has
2705 // less than a 4-byte alignment, then the frame-index elimination may need to
2706 // use an indexed load or store instruction (because the offset may not be a
2707 // multiple of 4). The extra register needed to hold the offset comes from the
2708 // register scavenger, and it is possible that the scavenger will need to use
2709 // an emergency spill slot. As a result, we need to make sure that a spill slot
2710 // is allocated when doing an i64 load/store into a less-than-4-byte-aligned
2711 // stack slot.
2712 static void fixupFuncForFI(SelectionDAG &DAG, int FrameIdx, EVT VT) {
2713   // FIXME: This does not handle the LWA case.
2714   if (VT != MVT::i64)
2715     return;
2716 
2717   // NOTE: We'll exclude negative FIs here, which come from argument
2718   // lowering, because there are no known test cases triggering this problem
2719   // using packed structures (or similar). We can remove this exclusion if
2720   // we find such a test case. The reason why this is so test-case driven is
2721   // because this entire 'fixup' is only to prevent crashes (from the
2722   // register scavenger) on not-really-valid inputs. For example, if we have:
2723   //   %a = alloca i1
2724   //   %b = bitcast i1* %a to i64*
2725   //   store i64* a, i64 b
2726   // then the store should really be marked as 'align 1', but is not. If it
2727   // were marked as 'align 1' then the indexed form would have been
2728   // instruction-selected initially, and the problem this 'fixup' is preventing
2729   // won't happen regardless.
2730   if (FrameIdx < 0)
2731     return;
2732 
2733   MachineFunction &MF = DAG.getMachineFunction();
2734   MachineFrameInfo &MFI = MF.getFrameInfo();
2735 
2736   if (MFI.getObjectAlign(FrameIdx) >= Align(4))
2737     return;
2738 
2739   PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2740   FuncInfo->setHasNonRISpills();
2741 }
2742 
2743 /// Returns true if the address N can be represented by a base register plus
2744 /// a signed 16-bit displacement [r+imm], and if it is not better
2745 /// represented as reg+reg.  If \p EncodingAlignment is non-zero, only accept
2746 /// displacements that are multiples of that value.
2747 bool PPCTargetLowering::SelectAddressRegImm(
2748     SDValue N, SDValue &Disp, SDValue &Base, SelectionDAG &DAG,
2749     MaybeAlign EncodingAlignment) const {
2750   // FIXME dl should come from parent load or store, not from address
2751   SDLoc dl(N);
2752 
2753   // If we have a PC Relative target flag don't select as [reg+imm]. It will be
2754   // a [pc+imm].
2755   if (SelectAddressPCRel(N, Base))
2756     return false;
2757 
2758   // If this can be more profitably realized as r+r, fail.
2759   if (SelectAddressRegReg(N, Disp, Base, DAG, EncodingAlignment))
2760     return false;
2761 
2762   if (N.getOpcode() == ISD::ADD) {
2763     int16_t imm = 0;
2764     if (isIntS16Immediate(N.getOperand(1), imm) &&
2765         (!EncodingAlignment || isAligned(*EncodingAlignment, imm))) {
2766       Disp = DAG.getTargetConstant(imm, dl, N.getValueType());
2767       if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
2768         Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
2769         fixupFuncForFI(DAG, FI->getIndex(), N.getValueType());
2770       } else {
2771         Base = N.getOperand(0);
2772       }
2773       return true; // [r+i]
2774     } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
2775       // Match LOAD (ADD (X, Lo(G))).
2776       assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
2777              && "Cannot handle constant offsets yet!");
2778       Disp = N.getOperand(1).getOperand(0);  // The global address.
2779       assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
2780              Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
2781              Disp.getOpcode() == ISD::TargetConstantPool ||
2782              Disp.getOpcode() == ISD::TargetJumpTable);
2783       Base = N.getOperand(0);
2784       return true;  // [&g+r]
2785     }
2786   } else if (N.getOpcode() == ISD::OR) {
2787     int16_t imm = 0;
2788     if (isIntS16Immediate(N.getOperand(1), imm) &&
2789         (!EncodingAlignment || isAligned(*EncodingAlignment, imm))) {
2790       // If this is an or of disjoint bitfields, we can codegen this as an add
2791       // (for better address arithmetic) if the LHS and RHS of the OR are
2792       // provably disjoint.
2793       KnownBits LHSKnown = DAG.computeKnownBits(N.getOperand(0));
2794 
2795       if ((LHSKnown.Zero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
2796         // If all of the bits are known zero on the LHS or RHS, the add won't
2797         // carry.
2798         if (FrameIndexSDNode *FI =
2799               dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
2800           Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
2801           fixupFuncForFI(DAG, FI->getIndex(), N.getValueType());
2802         } else {
2803           Base = N.getOperand(0);
2804         }
2805         Disp = DAG.getTargetConstant(imm, dl, N.getValueType());
2806         return true;
2807       }
2808     }
2809   } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
2810     // Loading from a constant address.
2811 
2812     // If this address fits entirely in a 16-bit sext immediate field, codegen
2813     // this as "d, 0"
2814     int16_t Imm;
2815     if (isIntS16Immediate(CN, Imm) &&
2816         (!EncodingAlignment || isAligned(*EncodingAlignment, Imm))) {
2817       Disp = DAG.getTargetConstant(Imm, dl, CN->getValueType(0));
2818       Base = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO,
2819                              CN->getValueType(0));
2820       return true;
2821     }
2822 
2823     // Handle 32-bit sext immediates with LIS + addr mode.
2824     if ((CN->getValueType(0) == MVT::i32 ||
2825          (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) &&
2826         (!EncodingAlignment ||
2827          isAligned(*EncodingAlignment, CN->getZExtValue()))) {
2828       int Addr = (int)CN->getZExtValue();
2829 
2830       // Otherwise, break this down into an LIS + disp.
2831       Disp = DAG.getTargetConstant((short)Addr, dl, MVT::i32);
2832 
2833       Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, dl,
2834                                    MVT::i32);
2835       unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
2836       Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
2837       return true;
2838     }
2839   }
2840 
2841   Disp = DAG.getTargetConstant(0, dl, getPointerTy(DAG.getDataLayout()));
2842   if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N)) {
2843     Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
2844     fixupFuncForFI(DAG, FI->getIndex(), N.getValueType());
2845   } else
2846     Base = N;
2847   return true;      // [r+0]
2848 }
2849 
2850 /// Similar to the 16-bit case but for instructions that take a 34-bit
2851 /// displacement field (prefixed loads/stores).
2852 bool PPCTargetLowering::SelectAddressRegImm34(SDValue N, SDValue &Disp,
2853                                               SDValue &Base,
2854                                               SelectionDAG &DAG) const {
2855   // Only on 64-bit targets.
2856   if (N.getValueType() != MVT::i64)
2857     return false;
2858 
2859   SDLoc dl(N);
2860   int64_t Imm = 0;
2861 
2862   if (N.getOpcode() == ISD::ADD) {
2863     if (!isIntS34Immediate(N.getOperand(1), Imm))
2864       return false;
2865     Disp = DAG.getTargetConstant(Imm, dl, N.getValueType());
2866     if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0)))
2867       Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
2868     else
2869       Base = N.getOperand(0);
2870     return true;
2871   }
2872 
2873   if (N.getOpcode() == ISD::OR) {
2874     if (!isIntS34Immediate(N.getOperand(1), Imm))
2875       return false;
2876     // If this is an or of disjoint bitfields, we can codegen this as an add
2877     // (for better address arithmetic) if the LHS and RHS of the OR are
2878     // provably disjoint.
2879     KnownBits LHSKnown = DAG.computeKnownBits(N.getOperand(0));
2880     if ((LHSKnown.Zero.getZExtValue() | ~(uint64_t)Imm) != ~0ULL)
2881       return false;
2882     if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0)))
2883       Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
2884     else
2885       Base = N.getOperand(0);
2886     Disp = DAG.getTargetConstant(Imm, dl, N.getValueType());
2887     return true;
2888   }
2889 
2890   if (isIntS34Immediate(N, Imm)) { // If the address is a 34-bit const.
2891     Disp = DAG.getTargetConstant(Imm, dl, N.getValueType());
2892     Base = DAG.getRegister(PPC::ZERO8, N.getValueType());
2893     return true;
2894   }
2895 
2896   return false;
2897 }
2898 
2899 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
2900 /// represented as an indexed [r+r] operation.
2901 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
2902                                                 SDValue &Index,
2903                                                 SelectionDAG &DAG) const {
2904   // Check to see if we can easily represent this as an [r+r] address.  This
2905   // will fail if it thinks that the address is more profitably represented as
2906   // reg+imm, e.g. where imm = 0.
2907   if (SelectAddressRegReg(N, Base, Index, DAG))
2908     return true;
2909 
2910   // If the address is the result of an add, we will utilize the fact that the
2911   // address calculation includes an implicit add.  However, we can reduce
2912   // register pressure if we do not materialize a constant just for use as the
2913   // index register.  We only get rid of the add if it is not an add of a
2914   // value and a 16-bit signed constant and both have a single use.
2915   int16_t imm = 0;
2916   if (N.getOpcode() == ISD::ADD &&
2917       (!isIntS16Immediate(N.getOperand(1), imm) ||
2918        !N.getOperand(1).hasOneUse() || !N.getOperand(0).hasOneUse())) {
2919     Base = N.getOperand(0);
2920     Index = N.getOperand(1);
2921     return true;
2922   }
2923 
2924   // Otherwise, do it the hard way, using R0 as the base register.
2925   Base = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO,
2926                          N.getValueType());
2927   Index = N;
2928   return true;
2929 }
2930 
2931 template <typename Ty> static bool isValidPCRelNode(SDValue N) {
2932   Ty *PCRelCand = dyn_cast<Ty>(N);
2933   return PCRelCand && (PCRelCand->getTargetFlags() & PPCII::MO_PCREL_FLAG);
2934 }
2935 
2936 /// Returns true if this address is a PC Relative address.
2937 /// PC Relative addresses are marked with the flag PPCII::MO_PCREL_FLAG
2938 /// or if the node opcode is PPCISD::MAT_PCREL_ADDR.
2939 bool PPCTargetLowering::SelectAddressPCRel(SDValue N, SDValue &Base) const {
2940   // This is a materialize PC Relative node. Always select this as PC Relative.
2941   Base = N;
2942   if (N.getOpcode() == PPCISD::MAT_PCREL_ADDR)
2943     return true;
2944   if (isValidPCRelNode<ConstantPoolSDNode>(N) ||
2945       isValidPCRelNode<GlobalAddressSDNode>(N) ||
2946       isValidPCRelNode<JumpTableSDNode>(N) ||
2947       isValidPCRelNode<BlockAddressSDNode>(N))
2948     return true;
2949   return false;
2950 }
2951 
2952 /// Returns true if we should use a direct load into vector instruction
2953 /// (such as lxsd or lfd), instead of a load into gpr + direct move sequence.
2954 static bool usePartialVectorLoads(SDNode *N, const PPCSubtarget& ST) {
2955 
2956   // If there are any other uses other than scalar to vector, then we should
2957   // keep it as a scalar load -> direct move pattern to prevent multiple
2958   // loads.
2959   LoadSDNode *LD = dyn_cast<LoadSDNode>(N);
2960   if (!LD)
2961     return false;
2962 
2963   EVT MemVT = LD->getMemoryVT();
2964   if (!MemVT.isSimple())
2965     return false;
2966   switch(MemVT.getSimpleVT().SimpleTy) {
2967   case MVT::i64:
2968     break;
2969   case MVT::i32:
2970     if (!ST.hasP8Vector())
2971       return false;
2972     break;
2973   case MVT::i16:
2974   case MVT::i8:
2975     if (!ST.hasP9Vector())
2976       return false;
2977     break;
2978   default:
2979     return false;
2980   }
2981 
2982   SDValue LoadedVal(N, 0);
2983   if (!LoadedVal.hasOneUse())
2984     return false;
2985 
2986   for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end();
2987        UI != UE; ++UI)
2988     if (UI.getUse().get().getResNo() == 0 &&
2989         UI->getOpcode() != ISD::SCALAR_TO_VECTOR &&
2990         UI->getOpcode() != PPCISD::SCALAR_TO_VECTOR_PERMUTED)
2991       return false;
2992 
2993   return true;
2994 }
2995 
2996 /// getPreIndexedAddressParts - returns true by value, base pointer and
2997 /// offset pointer and addressing mode by reference if the node's address
2998 /// can be legally represented as pre-indexed load / store address.
2999 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3000                                                   SDValue &Offset,
3001                                                   ISD::MemIndexedMode &AM,
3002                                                   SelectionDAG &DAG) const {
3003   if (DisablePPCPreinc) return false;
3004 
3005   bool isLoad = true;
3006   SDValue Ptr;
3007   EVT VT;
3008   Align Alignment;
3009   if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3010     Ptr = LD->getBasePtr();
3011     VT = LD->getMemoryVT();
3012     Alignment = LD->getAlign();
3013   } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3014     Ptr = ST->getBasePtr();
3015     VT  = ST->getMemoryVT();
3016     Alignment = ST->getAlign();
3017     isLoad = false;
3018   } else
3019     return false;
3020 
3021   // Do not generate pre-inc forms for specific loads that feed scalar_to_vector
3022   // instructions because we can fold these into a more efficient instruction
3023   // instead, (such as LXSD).
3024   if (isLoad && usePartialVectorLoads(N, Subtarget)) {
3025     return false;
3026   }
3027 
3028   // PowerPC doesn't have preinc load/store instructions for vectors
3029   if (VT.isVector())
3030     return false;
3031 
3032   if (SelectAddressRegReg(Ptr, Base, Offset, DAG)) {
3033     // Common code will reject creating a pre-inc form if the base pointer
3034     // is a frame index, or if N is a store and the base pointer is either
3035     // the same as or a predecessor of the value being stored.  Check for
3036     // those situations here, and try with swapped Base/Offset instead.
3037     bool Swap = false;
3038 
3039     if (isa<FrameIndexSDNode>(Base) || isa<RegisterSDNode>(Base))
3040       Swap = true;
3041     else if (!isLoad) {
3042       SDValue Val = cast<StoreSDNode>(N)->getValue();
3043       if (Val == Base || Base.getNode()->isPredecessorOf(Val.getNode()))
3044         Swap = true;
3045     }
3046 
3047     if (Swap)
3048       std::swap(Base, Offset);
3049 
3050     AM = ISD::PRE_INC;
3051     return true;
3052   }
3053 
3054   // LDU/STU can only handle immediates that are a multiple of 4.
3055   if (VT != MVT::i64) {
3056     if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, std::nullopt))
3057       return false;
3058   } else {
3059     // LDU/STU need an address with at least 4-byte alignment.
3060     if (Alignment < Align(4))
3061       return false;
3062 
3063     if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, Align(4)))
3064       return false;
3065   }
3066 
3067   if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3068     // PPC64 doesn't have lwau, but it does have lwaux.  Reject preinc load of
3069     // sext i32 to i64 when addr mode is r+i.
3070     if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
3071         LD->getExtensionType() == ISD::SEXTLOAD &&
3072         isa<ConstantSDNode>(Offset))
3073       return false;
3074   }
3075 
3076   AM = ISD::PRE_INC;
3077   return true;
3078 }
3079 
3080 //===----------------------------------------------------------------------===//
3081 //  LowerOperation implementation
3082 //===----------------------------------------------------------------------===//
3083 
3084 /// Return true if we should reference labels using a PICBase, set the HiOpFlags
3085 /// and LoOpFlags to the target MO flags.
3086 static void getLabelAccessInfo(bool IsPIC, const PPCSubtarget &Subtarget,
3087                                unsigned &HiOpFlags, unsigned &LoOpFlags,
3088                                const GlobalValue *GV = nullptr) {
3089   HiOpFlags = PPCII::MO_HA;
3090   LoOpFlags = PPCII::MO_LO;
3091 
3092   // Don't use the pic base if not in PIC relocation model.
3093   if (IsPIC) {
3094     HiOpFlags |= PPCII::MO_PIC_FLAG;
3095     LoOpFlags |= PPCII::MO_PIC_FLAG;
3096   }
3097 }
3098 
3099 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
3100                              SelectionDAG &DAG) {
3101   SDLoc DL(HiPart);
3102   EVT PtrVT = HiPart.getValueType();
3103   SDValue Zero = DAG.getConstant(0, DL, PtrVT);
3104 
3105   SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
3106   SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
3107 
3108   // With PIC, the first instruction is actually "GR+hi(&G)".
3109   if (isPIC)
3110     Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
3111                      DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
3112 
3113   // Generate non-pic code that has direct accesses to the constant pool.
3114   // The address of the global is just (hi(&g)+lo(&g)).
3115   return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
3116 }
3117 
3118 static void setUsesTOCBasePtr(MachineFunction &MF) {
3119   PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
3120   FuncInfo->setUsesTOCBasePtr();
3121 }
3122 
3123 static void setUsesTOCBasePtr(SelectionDAG &DAG) {
3124   setUsesTOCBasePtr(DAG.getMachineFunction());
3125 }
3126 
3127 SDValue PPCTargetLowering::getTOCEntry(SelectionDAG &DAG, const SDLoc &dl,
3128                                        SDValue GA) const {
3129   const bool Is64Bit = Subtarget.isPPC64();
3130   EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
3131   SDValue Reg = Is64Bit ? DAG.getRegister(PPC::X2, VT)
3132                         : Subtarget.isAIXABI()
3133                               ? DAG.getRegister(PPC::R2, VT)
3134                               : DAG.getNode(PPCISD::GlobalBaseReg, dl, VT);
3135   SDValue Ops[] = { GA, Reg };
3136   return DAG.getMemIntrinsicNode(
3137       PPCISD::TOC_ENTRY, dl, DAG.getVTList(VT, MVT::Other), Ops, VT,
3138       MachinePointerInfo::getGOT(DAG.getMachineFunction()), std::nullopt,
3139       MachineMemOperand::MOLoad);
3140 }
3141 
3142 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
3143                                              SelectionDAG &DAG) const {
3144   EVT PtrVT = Op.getValueType();
3145   ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3146   const Constant *C = CP->getConstVal();
3147 
3148   // 64-bit SVR4 ABI and AIX ABI code are always position-independent.
3149   // The actual address of the GlobalValue is stored in the TOC.
3150   if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) {
3151     if (Subtarget.isUsingPCRelativeCalls()) {
3152       SDLoc DL(CP);
3153       EVT Ty = getPointerTy(DAG.getDataLayout());
3154       SDValue ConstPool = DAG.getTargetConstantPool(
3155           C, Ty, CP->getAlign(), CP->getOffset(), PPCII::MO_PCREL_FLAG);
3156       return DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, ConstPool);
3157     }
3158     setUsesTOCBasePtr(DAG);
3159     SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), 0);
3160     return getTOCEntry(DAG, SDLoc(CP), GA);
3161   }
3162 
3163   unsigned MOHiFlag, MOLoFlag;
3164   bool IsPIC = isPositionIndependent();
3165   getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag);
3166 
3167   if (IsPIC && Subtarget.isSVR4ABI()) {
3168     SDValue GA =
3169         DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), PPCII::MO_PIC_FLAG);
3170     return getTOCEntry(DAG, SDLoc(CP), GA);
3171   }
3172 
3173   SDValue CPIHi =
3174       DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), 0, MOHiFlag);
3175   SDValue CPILo =
3176       DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), 0, MOLoFlag);
3177   return LowerLabelRef(CPIHi, CPILo, IsPIC, DAG);
3178 }
3179 
3180 // For 64-bit PowerPC, prefer the more compact relative encodings.
3181 // This trades 32 bits per jump table entry for one or two instructions
3182 // on the jump site.
3183 unsigned PPCTargetLowering::getJumpTableEncoding() const {
3184   if (isJumpTableRelative())
3185     return MachineJumpTableInfo::EK_LabelDifference32;
3186 
3187   return TargetLowering::getJumpTableEncoding();
3188 }
3189 
3190 bool PPCTargetLowering::isJumpTableRelative() const {
3191   if (UseAbsoluteJumpTables)
3192     return false;
3193   if (Subtarget.isPPC64() || Subtarget.isAIXABI())
3194     return true;
3195   return TargetLowering::isJumpTableRelative();
3196 }
3197 
3198 SDValue PPCTargetLowering::getPICJumpTableRelocBase(SDValue Table,
3199                                                     SelectionDAG &DAG) const {
3200   if (!Subtarget.isPPC64() || Subtarget.isAIXABI())
3201     return TargetLowering::getPICJumpTableRelocBase(Table, DAG);
3202 
3203   switch (getTargetMachine().getCodeModel()) {
3204   case CodeModel::Small:
3205   case CodeModel::Medium:
3206     return TargetLowering::getPICJumpTableRelocBase(Table, DAG);
3207   default:
3208     return DAG.getNode(PPCISD::GlobalBaseReg, SDLoc(),
3209                        getPointerTy(DAG.getDataLayout()));
3210   }
3211 }
3212 
3213 const MCExpr *
3214 PPCTargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
3215                                                 unsigned JTI,
3216                                                 MCContext &Ctx) const {
3217   if (!Subtarget.isPPC64() || Subtarget.isAIXABI())
3218     return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
3219 
3220   switch (getTargetMachine().getCodeModel()) {
3221   case CodeModel::Small:
3222   case CodeModel::Medium:
3223     return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
3224   default:
3225     return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
3226   }
3227 }
3228 
3229 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
3230   EVT PtrVT = Op.getValueType();
3231   JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3232 
3233   // isUsingPCRelativeCalls() returns true when PCRelative is enabled
3234   if (Subtarget.isUsingPCRelativeCalls()) {
3235     SDLoc DL(JT);
3236     EVT Ty = getPointerTy(DAG.getDataLayout());
3237     SDValue GA =
3238         DAG.getTargetJumpTable(JT->getIndex(), Ty, PPCII::MO_PCREL_FLAG);
3239     SDValue MatAddr = DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA);
3240     return MatAddr;
3241   }
3242 
3243   // 64-bit SVR4 ABI and AIX ABI code are always position-independent.
3244   // The actual address of the GlobalValue is stored in the TOC.
3245   if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) {
3246     setUsesTOCBasePtr(DAG);
3247     SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
3248     return getTOCEntry(DAG, SDLoc(JT), GA);
3249   }
3250 
3251   unsigned MOHiFlag, MOLoFlag;
3252   bool IsPIC = isPositionIndependent();
3253   getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag);
3254 
3255   if (IsPIC && Subtarget.isSVR4ABI()) {
3256     SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3257                                         PPCII::MO_PIC_FLAG);
3258     return getTOCEntry(DAG, SDLoc(GA), GA);
3259   }
3260 
3261   SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
3262   SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
3263   return LowerLabelRef(JTIHi, JTILo, IsPIC, DAG);
3264 }
3265 
3266 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
3267                                              SelectionDAG &DAG) const {
3268   EVT PtrVT = Op.getValueType();
3269   BlockAddressSDNode *BASDN = cast<BlockAddressSDNode>(Op);
3270   const BlockAddress *BA = BASDN->getBlockAddress();
3271 
3272   // isUsingPCRelativeCalls() returns true when PCRelative is enabled
3273   if (Subtarget.isUsingPCRelativeCalls()) {
3274     SDLoc DL(BASDN);
3275     EVT Ty = getPointerTy(DAG.getDataLayout());
3276     SDValue GA = DAG.getTargetBlockAddress(BA, Ty, BASDN->getOffset(),
3277                                            PPCII::MO_PCREL_FLAG);
3278     SDValue MatAddr = DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA);
3279     return MatAddr;
3280   }
3281 
3282   // 64-bit SVR4 ABI and AIX ABI code are always position-independent.
3283   // The actual BlockAddress is stored in the TOC.
3284   if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) {
3285     setUsesTOCBasePtr(DAG);
3286     SDValue GA = DAG.getTargetBlockAddress(BA, PtrVT, BASDN->getOffset());
3287     return getTOCEntry(DAG, SDLoc(BASDN), GA);
3288   }
3289 
3290   // 32-bit position-independent ELF stores the BlockAddress in the .got.
3291   if (Subtarget.is32BitELFABI() && isPositionIndependent())
3292     return getTOCEntry(
3293         DAG, SDLoc(BASDN),
3294         DAG.getTargetBlockAddress(BA, PtrVT, BASDN->getOffset()));
3295 
3296   unsigned MOHiFlag, MOLoFlag;
3297   bool IsPIC = isPositionIndependent();
3298   getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag);
3299   SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag);
3300   SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag);
3301   return LowerLabelRef(TgtBAHi, TgtBALo, IsPIC, DAG);
3302 }
3303 
3304 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
3305                                               SelectionDAG &DAG) const {
3306   if (Subtarget.isAIXABI())
3307     return LowerGlobalTLSAddressAIX(Op, DAG);
3308 
3309   return LowerGlobalTLSAddressLinux(Op, DAG);
3310 }
3311 
3312 SDValue PPCTargetLowering::LowerGlobalTLSAddressAIX(SDValue Op,
3313                                                     SelectionDAG &DAG) const {
3314   GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3315 
3316   if (DAG.getTarget().useEmulatedTLS())
3317     report_fatal_error("Emulated TLS is not yet supported on AIX");
3318 
3319   SDLoc dl(GA);
3320   const GlobalValue *GV = GA->getGlobal();
3321   EVT PtrVT = getPointerTy(DAG.getDataLayout());
3322 
3323   // The general-dynamic model is the only access model supported for now, so
3324   // all the GlobalTLSAddress nodes are lowered with this model.
3325   // We need to generate two TOC entries, one for the variable offset, one for
3326   // the region handle. The global address for the TOC entry of the region
3327   // handle is created with the MO_TLSGDM_FLAG flag and the global address
3328   // for the TOC entry of the variable offset is created with MO_TLSGD_FLAG.
3329   SDValue VariableOffsetTGA =
3330       DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, PPCII::MO_TLSGD_FLAG);
3331   SDValue RegionHandleTGA =
3332       DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, PPCII::MO_TLSGDM_FLAG);
3333   SDValue VariableOffset = getTOCEntry(DAG, dl, VariableOffsetTGA);
3334   SDValue RegionHandle = getTOCEntry(DAG, dl, RegionHandleTGA);
3335   return DAG.getNode(PPCISD::TLSGD_AIX, dl, PtrVT, VariableOffset,
3336                      RegionHandle);
3337 }
3338 
3339 SDValue PPCTargetLowering::LowerGlobalTLSAddressLinux(SDValue Op,
3340                                                       SelectionDAG &DAG) const {
3341   // FIXME: TLS addresses currently use medium model code sequences,
3342   // which is the most useful form.  Eventually support for small and
3343   // large models could be added if users need it, at the cost of
3344   // additional complexity.
3345   GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3346   if (DAG.getTarget().useEmulatedTLS())
3347     return LowerToTLSEmulatedModel(GA, DAG);
3348 
3349   SDLoc dl(GA);
3350   const GlobalValue *GV = GA->getGlobal();
3351   EVT PtrVT = getPointerTy(DAG.getDataLayout());
3352   bool is64bit = Subtarget.isPPC64();
3353   const Module *M = DAG.getMachineFunction().getFunction().getParent();
3354   PICLevel::Level picLevel = M->getPICLevel();
3355 
3356   const TargetMachine &TM = getTargetMachine();
3357   TLSModel::Model Model = TM.getTLSModel(GV);
3358 
3359   if (Model == TLSModel::LocalExec) {
3360     if (Subtarget.isUsingPCRelativeCalls()) {
3361       SDValue TLSReg = DAG.getRegister(PPC::X13, MVT::i64);
3362       SDValue TGA = DAG.getTargetGlobalAddress(
3363           GV, dl, PtrVT, 0, (PPCII::MO_PCREL_FLAG | PPCII::MO_TPREL_FLAG));
3364       SDValue MatAddr =
3365           DAG.getNode(PPCISD::TLS_LOCAL_EXEC_MAT_ADDR, dl, PtrVT, TGA);
3366       return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TLSReg, MatAddr);
3367     }
3368 
3369     SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
3370                                                PPCII::MO_TPREL_HA);
3371     SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
3372                                                PPCII::MO_TPREL_LO);
3373     SDValue TLSReg = is64bit ? DAG.getRegister(PPC::X13, MVT::i64)
3374                              : DAG.getRegister(PPC::R2, MVT::i32);
3375 
3376     SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
3377     return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
3378   }
3379 
3380   if (Model == TLSModel::InitialExec) {
3381     bool IsPCRel = Subtarget.isUsingPCRelativeCalls();
3382     SDValue TGA = DAG.getTargetGlobalAddress(
3383         GV, dl, PtrVT, 0, IsPCRel ? PPCII::MO_GOT_TPREL_PCREL_FLAG : 0);
3384     SDValue TGATLS = DAG.getTargetGlobalAddress(
3385         GV, dl, PtrVT, 0,
3386         IsPCRel ? (PPCII::MO_TLS | PPCII::MO_PCREL_FLAG) : PPCII::MO_TLS);
3387     SDValue TPOffset;
3388     if (IsPCRel) {
3389       SDValue MatPCRel = DAG.getNode(PPCISD::MAT_PCREL_ADDR, dl, PtrVT, TGA);
3390       TPOffset = DAG.getLoad(MVT::i64, dl, DAG.getEntryNode(), MatPCRel,
3391                              MachinePointerInfo());
3392     } else {
3393       SDValue GOTPtr;
3394       if (is64bit) {
3395         setUsesTOCBasePtr(DAG);
3396         SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
3397         GOTPtr =
3398             DAG.getNode(PPCISD::ADDIS_GOT_TPREL_HA, dl, PtrVT, GOTReg, TGA);
3399       } else {
3400         if (!TM.isPositionIndependent())
3401           GOTPtr = DAG.getNode(PPCISD::PPC32_GOT, dl, PtrVT);
3402         else if (picLevel == PICLevel::SmallPIC)
3403           GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT);
3404         else
3405           GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT);
3406       }
3407       TPOffset = DAG.getNode(PPCISD::LD_GOT_TPREL_L, dl, PtrVT, TGA, GOTPtr);
3408     }
3409     return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TPOffset, TGATLS);
3410   }
3411 
3412   if (Model == TLSModel::GeneralDynamic) {
3413     if (Subtarget.isUsingPCRelativeCalls()) {
3414       SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
3415                                                PPCII::MO_GOT_TLSGD_PCREL_FLAG);
3416       return DAG.getNode(PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR, dl, PtrVT, TGA);
3417     }
3418 
3419     SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
3420     SDValue GOTPtr;
3421     if (is64bit) {
3422       setUsesTOCBasePtr(DAG);
3423       SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
3424       GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSGD_HA, dl, PtrVT,
3425                                    GOTReg, TGA);
3426     } else {
3427       if (picLevel == PICLevel::SmallPIC)
3428         GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT);
3429       else
3430         GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT);
3431     }
3432     return DAG.getNode(PPCISD::ADDI_TLSGD_L_ADDR, dl, PtrVT,
3433                        GOTPtr, TGA, TGA);
3434   }
3435 
3436   if (Model == TLSModel::LocalDynamic) {
3437     if (Subtarget.isUsingPCRelativeCalls()) {
3438       SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
3439                                                PPCII::MO_GOT_TLSLD_PCREL_FLAG);
3440       SDValue MatPCRel =
3441           DAG.getNode(PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR, dl, PtrVT, TGA);
3442       return DAG.getNode(PPCISD::PADDI_DTPREL, dl, PtrVT, MatPCRel, TGA);
3443     }
3444 
3445     SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
3446     SDValue GOTPtr;
3447     if (is64bit) {
3448       setUsesTOCBasePtr(DAG);
3449       SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
3450       GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSLD_HA, dl, PtrVT,
3451                            GOTReg, TGA);
3452     } else {
3453       if (picLevel == PICLevel::SmallPIC)
3454         GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT);
3455       else
3456         GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT);
3457     }
3458     SDValue TLSAddr = DAG.getNode(PPCISD::ADDI_TLSLD_L_ADDR, dl,
3459                                   PtrVT, GOTPtr, TGA, TGA);
3460     SDValue DtvOffsetHi = DAG.getNode(PPCISD::ADDIS_DTPREL_HA, dl,
3461                                       PtrVT, TLSAddr, TGA);
3462     return DAG.getNode(PPCISD::ADDI_DTPREL_L, dl, PtrVT, DtvOffsetHi, TGA);
3463   }
3464 
3465   llvm_unreachable("Unknown TLS model!");
3466 }
3467 
3468 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
3469                                               SelectionDAG &DAG) const {
3470   EVT PtrVT = Op.getValueType();
3471   GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
3472   SDLoc DL(GSDN);
3473   const GlobalValue *GV = GSDN->getGlobal();
3474 
3475   // 64-bit SVR4 ABI & AIX ABI code is always position-independent.
3476   // The actual address of the GlobalValue is stored in the TOC.
3477   if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) {
3478     if (Subtarget.isUsingPCRelativeCalls()) {
3479       EVT Ty = getPointerTy(DAG.getDataLayout());
3480       if (isAccessedAsGotIndirect(Op)) {
3481         SDValue GA = DAG.getTargetGlobalAddress(GV, DL, Ty, GSDN->getOffset(),
3482                                                 PPCII::MO_PCREL_FLAG |
3483                                                     PPCII::MO_GOT_FLAG);
3484         SDValue MatPCRel = DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA);
3485         SDValue Load = DAG.getLoad(MVT::i64, DL, DAG.getEntryNode(), MatPCRel,
3486                                    MachinePointerInfo());
3487         return Load;
3488       } else {
3489         SDValue GA = DAG.getTargetGlobalAddress(GV, DL, Ty, GSDN->getOffset(),
3490                                                 PPCII::MO_PCREL_FLAG);
3491         return DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA);
3492       }
3493     }
3494     setUsesTOCBasePtr(DAG);
3495     SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
3496     return getTOCEntry(DAG, DL, GA);
3497   }
3498 
3499   unsigned MOHiFlag, MOLoFlag;
3500   bool IsPIC = isPositionIndependent();
3501   getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag, GV);
3502 
3503   if (IsPIC && Subtarget.isSVR4ABI()) {
3504     SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT,
3505                                             GSDN->getOffset(),
3506                                             PPCII::MO_PIC_FLAG);
3507     return getTOCEntry(DAG, DL, GA);
3508   }
3509 
3510   SDValue GAHi =
3511     DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
3512   SDValue GALo =
3513     DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
3514 
3515   return LowerLabelRef(GAHi, GALo, IsPIC, DAG);
3516 }
3517 
3518 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3519   bool IsStrict = Op->isStrictFPOpcode();
3520   ISD::CondCode CC =
3521       cast<CondCodeSDNode>(Op.getOperand(IsStrict ? 3 : 2))->get();
3522   SDValue LHS = Op.getOperand(IsStrict ? 1 : 0);
3523   SDValue RHS = Op.getOperand(IsStrict ? 2 : 1);
3524   SDValue Chain = IsStrict ? Op.getOperand(0) : SDValue();
3525   EVT LHSVT = LHS.getValueType();
3526   SDLoc dl(Op);
3527 
3528   // Soften the setcc with libcall if it is fp128.
3529   if (LHSVT == MVT::f128) {
3530     assert(!Subtarget.hasP9Vector() &&
3531            "SETCC for f128 is already legal under Power9!");
3532     softenSetCCOperands(DAG, LHSVT, LHS, RHS, CC, dl, LHS, RHS, Chain,
3533                         Op->getOpcode() == ISD::STRICT_FSETCCS);
3534     if (RHS.getNode())
3535       LHS = DAG.getNode(ISD::SETCC, dl, Op.getValueType(), LHS, RHS,
3536                         DAG.getCondCode(CC));
3537     if (IsStrict)
3538       return DAG.getMergeValues({LHS, Chain}, dl);
3539     return LHS;
3540   }
3541 
3542   assert(!IsStrict && "Don't know how to handle STRICT_FSETCC!");
3543 
3544   if (Op.getValueType() == MVT::v2i64) {
3545     // When the operands themselves are v2i64 values, we need to do something
3546     // special because VSX has no underlying comparison operations for these.
3547     if (LHS.getValueType() == MVT::v2i64) {
3548       // Equality can be handled by casting to the legal type for Altivec
3549       // comparisons, everything else needs to be expanded.
3550       if (CC != ISD::SETEQ && CC != ISD::SETNE)
3551         return SDValue();
3552       SDValue SetCC32 = DAG.getSetCC(
3553           dl, MVT::v4i32, DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, LHS),
3554           DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, RHS), CC);
3555       int ShuffV[] = {1, 0, 3, 2};
3556       SDValue Shuff =
3557           DAG.getVectorShuffle(MVT::v4i32, dl, SetCC32, SetCC32, ShuffV);
3558       return DAG.getBitcast(MVT::v2i64,
3559                             DAG.getNode(CC == ISD::SETEQ ? ISD::AND : ISD::OR,
3560                                         dl, MVT::v4i32, Shuff, SetCC32));
3561     }
3562 
3563     // We handle most of these in the usual way.
3564     return Op;
3565   }
3566 
3567   // If we're comparing for equality to zero, expose the fact that this is
3568   // implemented as a ctlz/srl pair on ppc, so that the dag combiner can
3569   // fold the new nodes.
3570   if (SDValue V = lowerCmpEqZeroToCtlzSrl(Op, DAG))
3571     return V;
3572 
3573   if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS)) {
3574     // Leave comparisons against 0 and -1 alone for now, since they're usually
3575     // optimized.  FIXME: revisit this when we can custom lower all setcc
3576     // optimizations.
3577     if (C->isAllOnes() || C->isZero())
3578       return SDValue();
3579   }
3580 
3581   // If we have an integer seteq/setne, turn it into a compare against zero
3582   // by xor'ing the rhs with the lhs, which is faster than setting a
3583   // condition register, reading it back out, and masking the correct bit.  The
3584   // normal approach here uses sub to do this instead of xor.  Using xor exposes
3585   // the result to other bit-twiddling opportunities.
3586   if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
3587     EVT VT = Op.getValueType();
3588     SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, LHS, RHS);
3589     return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, dl, LHSVT), CC);
3590   }
3591   return SDValue();
3592 }
3593 
3594 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3595   SDNode *Node = Op.getNode();
3596   EVT VT = Node->getValueType(0);
3597   EVT PtrVT = getPointerTy(DAG.getDataLayout());
3598   SDValue InChain = Node->getOperand(0);
3599   SDValue VAListPtr = Node->getOperand(1);
3600   const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
3601   SDLoc dl(Node);
3602 
3603   assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
3604 
3605   // gpr_index
3606   SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
3607                                     VAListPtr, MachinePointerInfo(SV), MVT::i8);
3608   InChain = GprIndex.getValue(1);
3609 
3610   if (VT == MVT::i64) {
3611     // Check if GprIndex is even
3612     SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
3613                                  DAG.getConstant(1, dl, MVT::i32));
3614     SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
3615                                 DAG.getConstant(0, dl, MVT::i32), ISD::SETNE);
3616     SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
3617                                           DAG.getConstant(1, dl, MVT::i32));
3618     // Align GprIndex to be even if it isn't
3619     GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
3620                            GprIndex);
3621   }
3622 
3623   // fpr index is 1 byte after gpr
3624   SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
3625                                DAG.getConstant(1, dl, MVT::i32));
3626 
3627   // fpr
3628   SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
3629                                     FprPtr, MachinePointerInfo(SV), MVT::i8);
3630   InChain = FprIndex.getValue(1);
3631 
3632   SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
3633                                        DAG.getConstant(8, dl, MVT::i32));
3634 
3635   SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
3636                                         DAG.getConstant(4, dl, MVT::i32));
3637 
3638   // areas
3639   SDValue OverflowArea =
3640       DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr, MachinePointerInfo());
3641   InChain = OverflowArea.getValue(1);
3642 
3643   SDValue RegSaveArea =
3644       DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr, MachinePointerInfo());
3645   InChain = RegSaveArea.getValue(1);
3646 
3647   // select overflow_area if index > 8
3648   SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
3649                             DAG.getConstant(8, dl, MVT::i32), ISD::SETLT);
3650 
3651   // adjustment constant gpr_index * 4/8
3652   SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
3653                                     VT.isInteger() ? GprIndex : FprIndex,
3654                                     DAG.getConstant(VT.isInteger() ? 4 : 8, dl,
3655                                                     MVT::i32));
3656 
3657   // OurReg = RegSaveArea + RegConstant
3658   SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
3659                                RegConstant);
3660 
3661   // Floating types are 32 bytes into RegSaveArea
3662   if (VT.isFloatingPoint())
3663     OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
3664                          DAG.getConstant(32, dl, MVT::i32));
3665 
3666   // increase {f,g}pr_index by 1 (or 2 if VT is i64)
3667   SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3668                                    VT.isInteger() ? GprIndex : FprIndex,
3669                                    DAG.getConstant(VT == MVT::i64 ? 2 : 1, dl,
3670                                                    MVT::i32));
3671 
3672   InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
3673                               VT.isInteger() ? VAListPtr : FprPtr,
3674                               MachinePointerInfo(SV), MVT::i8);
3675 
3676   // determine if we should load from reg_save_area or overflow_area
3677   SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
3678 
3679   // increase overflow_area by 4/8 if gpr/fpr > 8
3680   SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
3681                                           DAG.getConstant(VT.isInteger() ? 4 : 8,
3682                                           dl, MVT::i32));
3683 
3684   OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
3685                              OverflowAreaPlusN);
3686 
3687   InChain = DAG.getTruncStore(InChain, dl, OverflowArea, OverflowAreaPtr,
3688                               MachinePointerInfo(), MVT::i32);
3689 
3690   return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo());
3691 }
3692 
3693 SDValue PPCTargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3694   assert(!Subtarget.isPPC64() && "LowerVACOPY is PPC32 only");
3695 
3696   // We have to copy the entire va_list struct:
3697   // 2*sizeof(char) + 2 Byte alignment + 2*sizeof(char*) = 12 Byte
3698   return DAG.getMemcpy(Op.getOperand(0), Op, Op.getOperand(1), Op.getOperand(2),
3699                        DAG.getConstant(12, SDLoc(Op), MVT::i32), Align(8),
3700                        false, true, false, MachinePointerInfo(),
3701                        MachinePointerInfo());
3702 }
3703 
3704 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
3705                                                   SelectionDAG &DAG) const {
3706   if (Subtarget.isAIXABI())
3707     report_fatal_error("ADJUST_TRAMPOLINE operation is not supported on AIX.");
3708 
3709   return Op.getOperand(0);
3710 }
3711 
3712 SDValue PPCTargetLowering::LowerINLINEASM(SDValue Op, SelectionDAG &DAG) const {
3713   MachineFunction &MF = DAG.getMachineFunction();
3714   PPCFunctionInfo &MFI = *MF.getInfo<PPCFunctionInfo>();
3715 
3716   assert((Op.getOpcode() == ISD::INLINEASM ||
3717           Op.getOpcode() == ISD::INLINEASM_BR) &&
3718          "Expecting Inline ASM node.");
3719 
3720   // If an LR store is already known to be required then there is not point in
3721   // checking this ASM as well.
3722   if (MFI.isLRStoreRequired())
3723     return Op;
3724 
3725   // Inline ASM nodes have an optional last operand that is an incoming Flag of
3726   // type MVT::Glue. We want to ignore this last operand if that is the case.
3727   unsigned NumOps = Op.getNumOperands();
3728   if (Op.getOperand(NumOps - 1).getValueType() == MVT::Glue)
3729     --NumOps;
3730 
3731   // Check all operands that may contain the LR.
3732   for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
3733     unsigned Flags = cast<ConstantSDNode>(Op.getOperand(i))->getZExtValue();
3734     unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
3735     ++i; // Skip the ID value.
3736 
3737     switch (InlineAsm::getKind(Flags)) {
3738     default:
3739       llvm_unreachable("Bad flags!");
3740     case InlineAsm::Kind_RegUse:
3741     case InlineAsm::Kind_Imm:
3742     case InlineAsm::Kind_Mem:
3743       i += NumVals;
3744       break;
3745     case InlineAsm::Kind_Clobber:
3746     case InlineAsm::Kind_RegDef:
3747     case InlineAsm::Kind_RegDefEarlyClobber: {
3748       for (; NumVals; --NumVals, ++i) {
3749         Register Reg = cast<RegisterSDNode>(Op.getOperand(i))->getReg();
3750         if (Reg != PPC::LR && Reg != PPC::LR8)
3751           continue;
3752         MFI.setLRStoreRequired();
3753         return Op;
3754       }
3755       break;
3756     }
3757     }
3758   }
3759 
3760   return Op;
3761 }
3762 
3763 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
3764                                                 SelectionDAG &DAG) const {
3765   if (Subtarget.isAIXABI())
3766     report_fatal_error("INIT_TRAMPOLINE operation is not supported on AIX.");
3767 
3768   SDValue Chain = Op.getOperand(0);
3769   SDValue Trmp = Op.getOperand(1); // trampoline
3770   SDValue FPtr = Op.getOperand(2); // nested function
3771   SDValue Nest = Op.getOperand(3); // 'nest' parameter value
3772   SDLoc dl(Op);
3773 
3774   EVT PtrVT = getPointerTy(DAG.getDataLayout());
3775   bool isPPC64 = (PtrVT == MVT::i64);
3776   Type *IntPtrTy = DAG.getDataLayout().getIntPtrType(*DAG.getContext());
3777 
3778   TargetLowering::ArgListTy Args;
3779   TargetLowering::ArgListEntry Entry;
3780 
3781   Entry.Ty = IntPtrTy;
3782   Entry.Node = Trmp; Args.push_back(Entry);
3783 
3784   // TrampSize == (isPPC64 ? 48 : 40);
3785   Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40, dl,
3786                                isPPC64 ? MVT::i64 : MVT::i32);
3787   Args.push_back(Entry);
3788 
3789   Entry.Node = FPtr; Args.push_back(Entry);
3790   Entry.Node = Nest; Args.push_back(Entry);
3791 
3792   // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
3793   TargetLowering::CallLoweringInfo CLI(DAG);
3794   CLI.setDebugLoc(dl).setChain(Chain).setLibCallee(
3795       CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3796       DAG.getExternalSymbol("__trampoline_setup", PtrVT), std::move(Args));
3797 
3798   std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
3799   return CallResult.second;
3800 }
3801 
3802 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3803   MachineFunction &MF = DAG.getMachineFunction();
3804   PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
3805   EVT PtrVT = getPointerTy(MF.getDataLayout());
3806 
3807   SDLoc dl(Op);
3808 
3809   if (Subtarget.isPPC64() || Subtarget.isAIXABI()) {
3810     // vastart just stores the address of the VarArgsFrameIndex slot into the
3811     // memory location argument.
3812     SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
3813     const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3814     return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
3815                         MachinePointerInfo(SV));
3816   }
3817 
3818   // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
3819   // We suppose the given va_list is already allocated.
3820   //
3821   // typedef struct {
3822   //  char gpr;     /* index into the array of 8 GPRs
3823   //                 * stored in the register save area
3824   //                 * gpr=0 corresponds to r3,
3825   //                 * gpr=1 to r4, etc.
3826   //                 */
3827   //  char fpr;     /* index into the array of 8 FPRs
3828   //                 * stored in the register save area
3829   //                 * fpr=0 corresponds to f1,
3830   //                 * fpr=1 to f2, etc.
3831   //                 */
3832   //  char *overflow_arg_area;
3833   //                /* location on stack that holds
3834   //                 * the next overflow argument
3835   //                 */
3836   //  char *reg_save_area;
3837   //               /* where r3:r10 and f1:f8 (if saved)
3838   //                * are stored
3839   //                */
3840   // } va_list[1];
3841 
3842   SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), dl, MVT::i32);
3843   SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), dl, MVT::i32);
3844   SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
3845                                             PtrVT);
3846   SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
3847                                  PtrVT);
3848 
3849   uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
3850   SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, dl, PtrVT);
3851 
3852   uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
3853   SDValue ConstStackOffset = DAG.getConstant(StackOffset, dl, PtrVT);
3854 
3855   uint64_t FPROffset = 1;
3856   SDValue ConstFPROffset = DAG.getConstant(FPROffset, dl, PtrVT);
3857 
3858   const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3859 
3860   // Store first byte : number of int regs
3861   SDValue firstStore =
3862       DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR, Op.getOperand(1),
3863                         MachinePointerInfo(SV), MVT::i8);
3864   uint64_t nextOffset = FPROffset;
3865   SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
3866                                   ConstFPROffset);
3867 
3868   // Store second byte : number of float regs
3869   SDValue secondStore =
3870       DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
3871                         MachinePointerInfo(SV, nextOffset), MVT::i8);
3872   nextOffset += StackOffset;
3873   nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
3874 
3875   // Store second word : arguments given on stack
3876   SDValue thirdStore = DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
3877                                     MachinePointerInfo(SV, nextOffset));
3878   nextOffset += FrameOffset;
3879   nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
3880 
3881   // Store third word : arguments given in registers
3882   return DAG.getStore(thirdStore, dl, FR, nextPtr,
3883                       MachinePointerInfo(SV, nextOffset));
3884 }
3885 
3886 /// FPR - The set of FP registers that should be allocated for arguments
3887 /// on Darwin and AIX.
3888 static const MCPhysReg FPR[] = {PPC::F1,  PPC::F2,  PPC::F3, PPC::F4, PPC::F5,
3889                                 PPC::F6,  PPC::F7,  PPC::F8, PPC::F9, PPC::F10,
3890                                 PPC::F11, PPC::F12, PPC::F13};
3891 
3892 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
3893 /// the stack.
3894 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
3895                                        unsigned PtrByteSize) {
3896   unsigned ArgSize = ArgVT.getStoreSize();
3897   if (Flags.isByVal())
3898     ArgSize = Flags.getByValSize();
3899 
3900   // Round up to multiples of the pointer size, except for array members,
3901   // which are always packed.
3902   if (!Flags.isInConsecutiveRegs())
3903     ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
3904 
3905   return ArgSize;
3906 }
3907 
3908 /// CalculateStackSlotAlignment - Calculates the alignment of this argument
3909 /// on the stack.
3910 static Align CalculateStackSlotAlignment(EVT ArgVT, EVT OrigVT,
3911                                          ISD::ArgFlagsTy Flags,
3912                                          unsigned PtrByteSize) {
3913   Align Alignment(PtrByteSize);
3914 
3915   // Altivec parameters are padded to a 16 byte boundary.
3916   if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 ||
3917       ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 ||
3918       ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64 ||
3919       ArgVT == MVT::v1i128 || ArgVT == MVT::f128)
3920     Alignment = Align(16);
3921 
3922   // ByVal parameters are aligned as requested.
3923   if (Flags.isByVal()) {
3924     auto BVAlign = Flags.getNonZeroByValAlign();
3925     if (BVAlign > PtrByteSize) {
3926       if (BVAlign.value() % PtrByteSize != 0)
3927         llvm_unreachable(
3928             "ByVal alignment is not a multiple of the pointer size");
3929 
3930       Alignment = BVAlign;
3931     }
3932   }
3933 
3934   // Array members are always packed to their original alignment.
3935   if (Flags.isInConsecutiveRegs()) {
3936     // If the array member was split into multiple registers, the first
3937     // needs to be aligned to the size of the full type.  (Except for
3938     // ppcf128, which is only aligned as its f64 components.)
3939     if (Flags.isSplit() && OrigVT != MVT::ppcf128)
3940       Alignment = Align(OrigVT.getStoreSize());
3941     else
3942       Alignment = Align(ArgVT.getStoreSize());
3943   }
3944 
3945   return Alignment;
3946 }
3947 
3948 /// CalculateStackSlotUsed - Return whether this argument will use its
3949 /// stack slot (instead of being passed in registers).  ArgOffset,
3950 /// AvailableFPRs, and AvailableVRs must hold the current argument
3951 /// position, and will be updated to account for this argument.
3952 static bool CalculateStackSlotUsed(EVT ArgVT, EVT OrigVT, ISD::ArgFlagsTy Flags,
3953                                    unsigned PtrByteSize, unsigned LinkageSize,
3954                                    unsigned ParamAreaSize, unsigned &ArgOffset,
3955                                    unsigned &AvailableFPRs,
3956                                    unsigned &AvailableVRs) {
3957   bool UseMemory = false;
3958 
3959   // Respect alignment of argument on the stack.
3960   Align Alignment =
3961       CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
3962   ArgOffset = alignTo(ArgOffset, Alignment);
3963   // If there's no space left in the argument save area, we must
3964   // use memory (this check also catches zero-sized arguments).
3965   if (ArgOffset >= LinkageSize + ParamAreaSize)
3966     UseMemory = true;
3967 
3968   // Allocate argument on the stack.
3969   ArgOffset += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
3970   if (Flags.isInConsecutiveRegsLast())
3971     ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
3972   // If we overran the argument save area, we must use memory
3973   // (this check catches arguments passed partially in memory)
3974   if (ArgOffset > LinkageSize + ParamAreaSize)
3975     UseMemory = true;
3976 
3977   // However, if the argument is actually passed in an FPR or a VR,
3978   // we don't use memory after all.
3979   if (!Flags.isByVal()) {
3980     if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
3981       if (AvailableFPRs > 0) {
3982         --AvailableFPRs;
3983         return false;
3984       }
3985     if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 ||
3986         ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 ||
3987         ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64 ||
3988         ArgVT == MVT::v1i128 || ArgVT == MVT::f128)
3989       if (AvailableVRs > 0) {
3990         --AvailableVRs;
3991         return false;
3992       }
3993   }
3994 
3995   return UseMemory;
3996 }
3997 
3998 /// EnsureStackAlignment - Round stack frame size up from NumBytes to
3999 /// ensure minimum alignment required for target.
4000 static unsigned EnsureStackAlignment(const PPCFrameLowering *Lowering,
4001                                      unsigned NumBytes) {
4002   return alignTo(NumBytes, Lowering->getStackAlign());
4003 }
4004 
4005 SDValue PPCTargetLowering::LowerFormalArguments(
4006     SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
4007     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
4008     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
4009   if (Subtarget.isAIXABI())
4010     return LowerFormalArguments_AIX(Chain, CallConv, isVarArg, Ins, dl, DAG,
4011                                     InVals);
4012   if (Subtarget.is64BitELFABI())
4013     return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins, dl, DAG,
4014                                        InVals);
4015   assert(Subtarget.is32BitELFABI());
4016   return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins, dl, DAG,
4017                                      InVals);
4018 }
4019 
4020 SDValue PPCTargetLowering::LowerFormalArguments_32SVR4(
4021     SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
4022     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
4023     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
4024 
4025   // 32-bit SVR4 ABI Stack Frame Layout:
4026   //              +-----------------------------------+
4027   //        +-->  |            Back chain             |
4028   //        |     +-----------------------------------+
4029   //        |     | Floating-point register save area |
4030   //        |     +-----------------------------------+
4031   //        |     |    General register save area     |
4032   //        |     +-----------------------------------+
4033   //        |     |          CR save word             |
4034   //        |     +-----------------------------------+
4035   //        |     |         VRSAVE save word          |
4036   //        |     +-----------------------------------+
4037   //        |     |         Alignment padding         |
4038   //        |     +-----------------------------------+
4039   //        |     |     Vector register save area     |
4040   //        |     +-----------------------------------+
4041   //        |     |       Local variable space        |
4042   //        |     +-----------------------------------+
4043   //        |     |        Parameter list area        |
4044   //        |     +-----------------------------------+
4045   //        |     |           LR save word            |
4046   //        |     +-----------------------------------+
4047   // SP-->  +---  |            Back chain             |
4048   //              +-----------------------------------+
4049   //
4050   // Specifications:
4051   //   System V Application Binary Interface PowerPC Processor Supplement
4052   //   AltiVec Technology Programming Interface Manual
4053 
4054   MachineFunction &MF = DAG.getMachineFunction();
4055   MachineFrameInfo &MFI = MF.getFrameInfo();
4056   PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
4057 
4058   EVT PtrVT = getPointerTy(MF.getDataLayout());
4059   // Potential tail calls could cause overwriting of argument stack slots.
4060   bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
4061                        (CallConv == CallingConv::Fast));
4062   const Align PtrAlign(4);
4063 
4064   // Assign locations to all of the incoming arguments.
4065   SmallVector<CCValAssign, 16> ArgLocs;
4066   PPCCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
4067                  *DAG.getContext());
4068 
4069   // Reserve space for the linkage area on the stack.
4070   unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
4071   CCInfo.AllocateStack(LinkageSize, PtrAlign);
4072   if (useSoftFloat())
4073     CCInfo.PreAnalyzeFormalArguments(Ins);
4074 
4075   CCInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4);
4076   CCInfo.clearWasPPCF128();
4077 
4078   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
4079     CCValAssign &VA = ArgLocs[i];
4080 
4081     // Arguments stored in registers.
4082     if (VA.isRegLoc()) {
4083       const TargetRegisterClass *RC;
4084       EVT ValVT = VA.getValVT();
4085 
4086       switch (ValVT.getSimpleVT().SimpleTy) {
4087         default:
4088           llvm_unreachable("ValVT not supported by formal arguments Lowering");
4089         case MVT::i1:
4090         case MVT::i32:
4091           RC = &PPC::GPRCRegClass;
4092           break;
4093         case MVT::f32:
4094           if (Subtarget.hasP8Vector())
4095             RC = &PPC::VSSRCRegClass;
4096           else if (Subtarget.hasSPE())
4097             RC = &PPC::GPRCRegClass;
4098           else
4099             RC = &PPC::F4RCRegClass;
4100           break;
4101         case MVT::f64:
4102           if (Subtarget.hasVSX())
4103             RC = &PPC::VSFRCRegClass;
4104           else if (Subtarget.hasSPE())
4105             // SPE passes doubles in GPR pairs.
4106             RC = &PPC::GPRCRegClass;
4107           else
4108             RC = &PPC::F8RCRegClass;
4109           break;
4110         case MVT::v16i8:
4111         case MVT::v8i16:
4112         case MVT::v4i32:
4113           RC = &PPC::VRRCRegClass;
4114           break;
4115         case MVT::v4f32:
4116           RC = &PPC::VRRCRegClass;
4117           break;
4118         case MVT::v2f64:
4119         case MVT::v2i64:
4120           RC = &PPC::VRRCRegClass;
4121           break;
4122       }
4123 
4124       SDValue ArgValue;
4125       // Transform the arguments stored in physical registers into
4126       // virtual ones.
4127       if (VA.getLocVT() == MVT::f64 && Subtarget.hasSPE()) {
4128         assert(i + 1 < e && "No second half of double precision argument");
4129         Register RegLo = MF.addLiveIn(VA.getLocReg(), RC);
4130         Register RegHi = MF.addLiveIn(ArgLocs[++i].getLocReg(), RC);
4131         SDValue ArgValueLo = DAG.getCopyFromReg(Chain, dl, RegLo, MVT::i32);
4132         SDValue ArgValueHi = DAG.getCopyFromReg(Chain, dl, RegHi, MVT::i32);
4133         if (!Subtarget.isLittleEndian())
4134           std::swap (ArgValueLo, ArgValueHi);
4135         ArgValue = DAG.getNode(PPCISD::BUILD_SPE64, dl, MVT::f64, ArgValueLo,
4136                                ArgValueHi);
4137       } else {
4138         Register Reg = MF.addLiveIn(VA.getLocReg(), RC);
4139         ArgValue = DAG.getCopyFromReg(Chain, dl, Reg,
4140                                       ValVT == MVT::i1 ? MVT::i32 : ValVT);
4141         if (ValVT == MVT::i1)
4142           ArgValue = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, ArgValue);
4143       }
4144 
4145       InVals.push_back(ArgValue);
4146     } else {
4147       // Argument stored in memory.
4148       assert(VA.isMemLoc());
4149 
4150       // Get the extended size of the argument type in stack
4151       unsigned ArgSize = VA.getLocVT().getStoreSize();
4152       // Get the actual size of the argument type
4153       unsigned ObjSize = VA.getValVT().getStoreSize();
4154       unsigned ArgOffset = VA.getLocMemOffset();
4155       // Stack objects in PPC32 are right justified.
4156       ArgOffset += ArgSize - ObjSize;
4157       int FI = MFI.CreateFixedObject(ArgSize, ArgOffset, isImmutable);
4158 
4159       // Create load nodes to retrieve arguments from the stack.
4160       SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
4161       InVals.push_back(
4162           DAG.getLoad(VA.getValVT(), dl, Chain, FIN, MachinePointerInfo()));
4163     }
4164   }
4165 
4166   // Assign locations to all of the incoming aggregate by value arguments.
4167   // Aggregates passed by value are stored in the local variable space of the
4168   // caller's stack frame, right above the parameter list area.
4169   SmallVector<CCValAssign, 16> ByValArgLocs;
4170   CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
4171                       ByValArgLocs, *DAG.getContext());
4172 
4173   // Reserve stack space for the allocations in CCInfo.
4174   CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrAlign);
4175 
4176   CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4_ByVal);
4177 
4178   // Area that is at least reserved in the caller of this function.
4179   unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
4180   MinReservedArea = std::max(MinReservedArea, LinkageSize);
4181 
4182   // Set the size that is at least reserved in caller of this function.  Tail
4183   // call optimized function's reserved stack space needs to be aligned so that
4184   // taking the difference between two stack areas will result in an aligned
4185   // stack.
4186   MinReservedArea =
4187       EnsureStackAlignment(Subtarget.getFrameLowering(), MinReservedArea);
4188   FuncInfo->setMinReservedArea(MinReservedArea);
4189 
4190   SmallVector<SDValue, 8> MemOps;
4191 
4192   // If the function takes variable number of arguments, make a frame index for
4193   // the start of the first vararg value... for expansion of llvm.va_start.
4194   if (isVarArg) {
4195     static const MCPhysReg GPArgRegs[] = {
4196       PPC::R3, PPC::R4, PPC::R5, PPC::R6,
4197       PPC::R7, PPC::R8, PPC::R9, PPC::R10,
4198     };
4199     const unsigned NumGPArgRegs = std::size(GPArgRegs);
4200 
4201     static const MCPhysReg FPArgRegs[] = {
4202       PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
4203       PPC::F8
4204     };
4205     unsigned NumFPArgRegs = std::size(FPArgRegs);
4206 
4207     if (useSoftFloat() || hasSPE())
4208        NumFPArgRegs = 0;
4209 
4210     FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs));
4211     FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs));
4212 
4213     // Make room for NumGPArgRegs and NumFPArgRegs.
4214     int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
4215                 NumFPArgRegs * MVT(MVT::f64).getSizeInBits()/8;
4216 
4217     FuncInfo->setVarArgsStackOffset(
4218       MFI.CreateFixedObject(PtrVT.getSizeInBits()/8,
4219                             CCInfo.getNextStackOffset(), true));
4220 
4221     FuncInfo->setVarArgsFrameIndex(
4222         MFI.CreateStackObject(Depth, Align(8), false));
4223     SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
4224 
4225     // The fixed integer arguments of a variadic function are stored to the
4226     // VarArgsFrameIndex on the stack so that they may be loaded by
4227     // dereferencing the result of va_next.
4228     for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
4229       // Get an existing live-in vreg, or add a new one.
4230       Register VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
4231       if (!VReg)
4232         VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
4233 
4234       SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
4235       SDValue Store =
4236           DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo());
4237       MemOps.push_back(Store);
4238       // Increment the address by four for the next argument to store
4239       SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, dl, PtrVT);
4240       FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
4241     }
4242 
4243     // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
4244     // is set.
4245     // The double arguments are stored to the VarArgsFrameIndex
4246     // on the stack.
4247     for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
4248       // Get an existing live-in vreg, or add a new one.
4249       Register VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
4250       if (!VReg)
4251         VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
4252 
4253       SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
4254       SDValue Store =
4255           DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo());
4256       MemOps.push_back(Store);
4257       // Increment the address by eight for the next argument to store
4258       SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8, dl,
4259                                          PtrVT);
4260       FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
4261     }
4262   }
4263 
4264   if (!MemOps.empty())
4265     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
4266 
4267   return Chain;
4268 }
4269 
4270 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
4271 // value to MVT::i64 and then truncate to the correct register size.
4272 SDValue PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags,
4273                                              EVT ObjectVT, SelectionDAG &DAG,
4274                                              SDValue ArgVal,
4275                                              const SDLoc &dl) const {
4276   if (Flags.isSExt())
4277     ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
4278                          DAG.getValueType(ObjectVT));
4279   else if (Flags.isZExt())
4280     ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
4281                          DAG.getValueType(ObjectVT));
4282 
4283   return DAG.getNode(ISD::TRUNCATE, dl, ObjectVT, ArgVal);
4284 }
4285 
4286 SDValue PPCTargetLowering::LowerFormalArguments_64SVR4(
4287     SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
4288     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
4289     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
4290   // TODO: add description of PPC stack frame format, or at least some docs.
4291   //
4292   bool isELFv2ABI = Subtarget.isELFv2ABI();
4293   bool isLittleEndian = Subtarget.isLittleEndian();
4294   MachineFunction &MF = DAG.getMachineFunction();
4295   MachineFrameInfo &MFI = MF.getFrameInfo();
4296   PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
4297 
4298   assert(!(CallConv == CallingConv::Fast && isVarArg) &&
4299          "fastcc not supported on varargs functions");
4300 
4301   EVT PtrVT = getPointerTy(MF.getDataLayout());
4302   // Potential tail calls could cause overwriting of argument stack slots.
4303   bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
4304                        (CallConv == CallingConv::Fast));
4305   unsigned PtrByteSize = 8;
4306   unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
4307 
4308   static const MCPhysReg GPR[] = {
4309     PPC::X3, PPC::X4, PPC::X5, PPC::X6,
4310     PPC::X7, PPC::X8, PPC::X9, PPC::X10,
4311   };
4312   static const MCPhysReg VR[] = {
4313     PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
4314     PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
4315   };
4316 
4317   const unsigned Num_GPR_Regs = std::size(GPR);
4318   const unsigned Num_FPR_Regs = useSoftFloat() ? 0 : 13;
4319   const unsigned Num_VR_Regs = std::size(VR);
4320 
4321   // Do a first pass over the arguments to determine whether the ABI
4322   // guarantees that our caller has allocated the parameter save area
4323   // on its stack frame.  In the ELFv1 ABI, this is always the case;
4324   // in the ELFv2 ABI, it is true if this is a vararg function or if
4325   // any parameter is located in a stack slot.
4326 
4327   bool HasParameterArea = !isELFv2ABI || isVarArg;
4328   unsigned ParamAreaSize = Num_GPR_Regs * PtrByteSize;
4329   unsigned NumBytes = LinkageSize;
4330   unsigned AvailableFPRs = Num_FPR_Regs;
4331   unsigned AvailableVRs = Num_VR_Regs;
4332   for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
4333     if (Ins[i].Flags.isNest())
4334       continue;
4335 
4336     if (CalculateStackSlotUsed(Ins[i].VT, Ins[i].ArgVT, Ins[i].Flags,
4337                                PtrByteSize, LinkageSize, ParamAreaSize,
4338                                NumBytes, AvailableFPRs, AvailableVRs))
4339       HasParameterArea = true;
4340   }
4341 
4342   // Add DAG nodes to load the arguments or copy them out of registers.  On
4343   // entry to a function on PPC, the arguments start after the linkage area,
4344   // although the first ones are often in registers.
4345 
4346   unsigned ArgOffset = LinkageSize;
4347   unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
4348   SmallVector<SDValue, 8> MemOps;
4349   Function::const_arg_iterator FuncArg = MF.getFunction().arg_begin();
4350   unsigned CurArgIdx = 0;
4351   for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
4352     SDValue ArgVal;
4353     bool needsLoad = false;
4354     EVT ObjectVT = Ins[ArgNo].VT;
4355     EVT OrigVT = Ins[ArgNo].ArgVT;
4356     unsigned ObjSize = ObjectVT.getStoreSize();
4357     unsigned ArgSize = ObjSize;
4358     ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
4359     if (Ins[ArgNo].isOrigArg()) {
4360       std::advance(FuncArg, Ins[ArgNo].getOrigArgIndex() - CurArgIdx);
4361       CurArgIdx = Ins[ArgNo].getOrigArgIndex();
4362     }
4363     // We re-align the argument offset for each argument, except when using the
4364     // fast calling convention, when we need to make sure we do that only when
4365     // we'll actually use a stack slot.
4366     unsigned CurArgOffset;
4367     Align Alignment;
4368     auto ComputeArgOffset = [&]() {
4369       /* Respect alignment of argument on the stack.  */
4370       Alignment =
4371           CalculateStackSlotAlignment(ObjectVT, OrigVT, Flags, PtrByteSize);
4372       ArgOffset = alignTo(ArgOffset, Alignment);
4373       CurArgOffset = ArgOffset;
4374     };
4375 
4376     if (CallConv != CallingConv::Fast) {
4377       ComputeArgOffset();
4378 
4379       /* Compute GPR index associated with argument offset.  */
4380       GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize;
4381       GPR_idx = std::min(GPR_idx, Num_GPR_Regs);
4382     }
4383 
4384     // FIXME the codegen can be much improved in some cases.
4385     // We do not have to keep everything in memory.
4386     if (Flags.isByVal()) {
4387       assert(Ins[ArgNo].isOrigArg() && "Byval arguments cannot be implicit");
4388 
4389       if (CallConv == CallingConv::Fast)
4390         ComputeArgOffset();
4391 
4392       // ObjSize is the true size, ArgSize rounded up to multiple of registers.
4393       ObjSize = Flags.getByValSize();
4394       ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
4395       // Empty aggregate parameters do not take up registers.  Examples:
4396       //   struct { } a;
4397       //   union  { } b;
4398       //   int c[0];
4399       // etc.  However, we have to provide a place-holder in InVals, so
4400       // pretend we have an 8-byte item at the current address for that
4401       // purpose.
4402       if (!ObjSize) {
4403         int FI = MFI.CreateFixedObject(PtrByteSize, ArgOffset, true);
4404         SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
4405         InVals.push_back(FIN);
4406         continue;
4407       }
4408 
4409       // Create a stack object covering all stack doublewords occupied
4410       // by the argument.  If the argument is (fully or partially) on
4411       // the stack, or if the argument is fully in registers but the
4412       // caller has allocated the parameter save anyway, we can refer
4413       // directly to the caller's stack frame.  Otherwise, create a
4414       // local copy in our own frame.
4415       int FI;
4416       if (HasParameterArea ||
4417           ArgSize + ArgOffset > LinkageSize + Num_GPR_Regs * PtrByteSize)
4418         FI = MFI.CreateFixedObject(ArgSize, ArgOffset, false, true);
4419       else
4420         FI = MFI.CreateStackObject(ArgSize, Alignment, false);
4421       SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
4422 
4423       // Handle aggregates smaller than 8 bytes.
4424       if (ObjSize < PtrByteSize) {
4425         // The value of the object is its address, which differs from the
4426         // address of the enclosing doubleword on big-endian systems.
4427         SDValue Arg = FIN;
4428         if (!isLittleEndian) {
4429           SDValue ArgOff = DAG.getConstant(PtrByteSize - ObjSize, dl, PtrVT);
4430           Arg = DAG.getNode(ISD::ADD, dl, ArgOff.getValueType(), Arg, ArgOff);
4431         }
4432         InVals.push_back(Arg);
4433 
4434         if (GPR_idx != Num_GPR_Regs) {
4435           Register VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass);
4436           FuncInfo->addLiveInAttr(VReg, Flags);
4437           SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
4438           EVT ObjType = EVT::getIntegerVT(*DAG.getContext(), ObjSize * 8);
4439           SDValue Store =
4440               DAG.getTruncStore(Val.getValue(1), dl, Val, Arg,
4441                                 MachinePointerInfo(&*FuncArg), ObjType);
4442           MemOps.push_back(Store);
4443         }
4444         // Whether we copied from a register or not, advance the offset
4445         // into the parameter save area by a full doubleword.
4446         ArgOffset += PtrByteSize;
4447         continue;
4448       }
4449 
4450       // The value of the object is its address, which is the address of
4451       // its first stack doubleword.
4452       InVals.push_back(FIN);
4453 
4454       // Store whatever pieces of the object are in registers to memory.
4455       for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
4456         if (GPR_idx == Num_GPR_Regs)
4457           break;
4458 
4459         Register VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
4460         FuncInfo->addLiveInAttr(VReg, Flags);
4461         SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
4462         SDValue Addr = FIN;
4463         if (j) {
4464           SDValue Off = DAG.getConstant(j, dl, PtrVT);
4465           Addr = DAG.getNode(ISD::ADD, dl, Off.getValueType(), Addr, Off);
4466         }
4467         unsigned StoreSizeInBits = std::min(PtrByteSize, (ObjSize - j)) * 8;
4468         EVT ObjType = EVT::getIntegerVT(*DAG.getContext(), StoreSizeInBits);
4469         SDValue Store =
4470             DAG.getTruncStore(Val.getValue(1), dl, Val, Addr,
4471                               MachinePointerInfo(&*FuncArg, j), ObjType);
4472         MemOps.push_back(Store);
4473         ++GPR_idx;
4474       }
4475       ArgOffset += ArgSize;
4476       continue;
4477     }
4478 
4479     switch (ObjectVT.getSimpleVT().SimpleTy) {
4480     default: llvm_unreachable("Unhandled argument type!");
4481     case MVT::i1:
4482     case MVT::i32:
4483     case MVT::i64:
4484       if (Flags.isNest()) {
4485         // The 'nest' parameter, if any, is passed in R11.
4486         Register VReg = MF.addLiveIn(PPC::X11, &PPC::G8RCRegClass);
4487         ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
4488 
4489         if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1)
4490           ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
4491 
4492         break;
4493       }
4494 
4495       // These can be scalar arguments or elements of an integer array type
4496       // passed directly.  Clang may use those instead of "byval" aggregate
4497       // types to avoid forcing arguments to memory unnecessarily.
4498       if (GPR_idx != Num_GPR_Regs) {
4499         Register VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass);
4500         FuncInfo->addLiveInAttr(VReg, Flags);
4501         ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
4502 
4503         if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1)
4504           // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
4505           // value to MVT::i64 and then truncate to the correct register size.
4506           ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
4507       } else {
4508         if (CallConv == CallingConv::Fast)
4509           ComputeArgOffset();
4510 
4511         needsLoad = true;
4512         ArgSize = PtrByteSize;
4513       }
4514       if (CallConv != CallingConv::Fast || needsLoad)
4515         ArgOffset += 8;
4516       break;
4517 
4518     case MVT::f32:
4519     case MVT::f64:
4520       // These can be scalar arguments or elements of a float array type
4521       // passed directly.  The latter are used to implement ELFv2 homogenous
4522       // float aggregates.
4523       if (FPR_idx != Num_FPR_Regs) {
4524         unsigned VReg;
4525 
4526         if (ObjectVT == MVT::f32)
4527           VReg = MF.addLiveIn(FPR[FPR_idx],
4528                               Subtarget.hasP8Vector()
4529                                   ? &PPC::VSSRCRegClass
4530                                   : &PPC::F4RCRegClass);
4531         else
4532           VReg = MF.addLiveIn(FPR[FPR_idx], Subtarget.hasVSX()
4533                                                 ? &PPC::VSFRCRegClass
4534                                                 : &PPC::F8RCRegClass);
4535 
4536         ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
4537         ++FPR_idx;
4538       } else if (GPR_idx != Num_GPR_Regs && CallConv != CallingConv::Fast) {
4539         // FIXME: We may want to re-enable this for CallingConv::Fast on the P8
4540         // once we support fp <-> gpr moves.
4541 
4542         // This can only ever happen in the presence of f32 array types,
4543         // since otherwise we never run out of FPRs before running out
4544         // of GPRs.
4545         Register VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass);
4546         FuncInfo->addLiveInAttr(VReg, Flags);
4547         ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
4548 
4549         if (ObjectVT == MVT::f32) {
4550           if ((ArgOffset % PtrByteSize) == (isLittleEndian ? 4 : 0))
4551             ArgVal = DAG.getNode(ISD::SRL, dl, MVT::i64, ArgVal,
4552                                  DAG.getConstant(32, dl, MVT::i32));
4553           ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
4554         }
4555 
4556         ArgVal = DAG.getNode(ISD::BITCAST, dl, ObjectVT, ArgVal);
4557       } else {
4558         if (CallConv == CallingConv::Fast)
4559           ComputeArgOffset();
4560 
4561         needsLoad = true;
4562       }
4563 
4564       // When passing an array of floats, the array occupies consecutive
4565       // space in the argument area; only round up to the next doubleword
4566       // at the end of the array.  Otherwise, each float takes 8 bytes.
4567       if (CallConv != CallingConv::Fast || needsLoad) {
4568         ArgSize = Flags.isInConsecutiveRegs() ? ObjSize : PtrByteSize;
4569         ArgOffset += ArgSize;
4570         if (Flags.isInConsecutiveRegsLast())
4571           ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
4572       }
4573       break;
4574     case MVT::v4f32:
4575     case MVT::v4i32:
4576     case MVT::v8i16:
4577     case MVT::v16i8:
4578     case MVT::v2f64:
4579     case MVT::v2i64:
4580     case MVT::v1i128:
4581     case MVT::f128:
4582       // These can be scalar arguments or elements of a vector array type
4583       // passed directly.  The latter are used to implement ELFv2 homogenous
4584       // vector aggregates.
4585       if (VR_idx != Num_VR_Regs) {
4586         Register VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
4587         ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
4588         ++VR_idx;
4589       } else {
4590         if (CallConv == CallingConv::Fast)
4591           ComputeArgOffset();
4592         needsLoad = true;
4593       }
4594       if (CallConv != CallingConv::Fast || needsLoad)
4595         ArgOffset += 16;
4596       break;
4597     }
4598 
4599     // We need to load the argument to a virtual register if we determined
4600     // above that we ran out of physical registers of the appropriate type.
4601     if (needsLoad) {
4602       if (ObjSize < ArgSize && !isLittleEndian)
4603         CurArgOffset += ArgSize - ObjSize;
4604       int FI = MFI.CreateFixedObject(ObjSize, CurArgOffset, isImmutable);
4605       SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
4606       ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo());
4607     }
4608 
4609     InVals.push_back(ArgVal);
4610   }
4611 
4612   // Area that is at least reserved in the caller of this function.
4613   unsigned MinReservedArea;
4614   if (HasParameterArea)
4615     MinReservedArea = std::max(ArgOffset, LinkageSize + 8 * PtrByteSize);
4616   else
4617     MinReservedArea = LinkageSize;
4618 
4619   // Set the size that is at least reserved in caller of this function.  Tail
4620   // call optimized functions' reserved stack space needs to be aligned so that
4621   // taking the difference between two stack areas will result in an aligned
4622   // stack.
4623   MinReservedArea =
4624       EnsureStackAlignment(Subtarget.getFrameLowering(), MinReservedArea);
4625   FuncInfo->setMinReservedArea(MinReservedArea);
4626 
4627   // If the function takes variable number of arguments, make a frame index for
4628   // the start of the first vararg value... for expansion of llvm.va_start.
4629   // On ELFv2ABI spec, it writes:
4630   // C programs that are intended to be *portable* across different compilers
4631   // and architectures must use the header file <stdarg.h> to deal with variable
4632   // argument lists.
4633   if (isVarArg && MFI.hasVAStart()) {
4634     int Depth = ArgOffset;
4635 
4636     FuncInfo->setVarArgsFrameIndex(
4637       MFI.CreateFixedObject(PtrByteSize, Depth, true));
4638     SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
4639 
4640     // If this function is vararg, store any remaining integer argument regs
4641     // to their spots on the stack so that they may be loaded by dereferencing
4642     // the result of va_next.
4643     for (GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize;
4644          GPR_idx < Num_GPR_Regs; ++GPR_idx) {
4645       Register VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
4646       SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
4647       SDValue Store =
4648           DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo());
4649       MemOps.push_back(Store);
4650       // Increment the address by four for the next argument to store
4651       SDValue PtrOff = DAG.getConstant(PtrByteSize, dl, PtrVT);
4652       FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
4653     }
4654   }
4655 
4656   if (!MemOps.empty())
4657     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
4658 
4659   return Chain;
4660 }
4661 
4662 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
4663 /// adjusted to accommodate the arguments for the tailcall.
4664 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
4665                                    unsigned ParamSize) {
4666 
4667   if (!isTailCall) return 0;
4668 
4669   PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
4670   unsigned CallerMinReservedArea = FI->getMinReservedArea();
4671   int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
4672   // Remember only if the new adjustment is bigger.
4673   if (SPDiff < FI->getTailCallSPDelta())
4674     FI->setTailCallSPDelta(SPDiff);
4675 
4676   return SPDiff;
4677 }
4678 
4679 static bool isFunctionGlobalAddress(SDValue Callee);
4680 
4681 static bool callsShareTOCBase(const Function *Caller, SDValue Callee,
4682                               const TargetMachine &TM) {
4683   // It does not make sense to call callsShareTOCBase() with a caller that
4684   // is PC Relative since PC Relative callers do not have a TOC.
4685 #ifndef NDEBUG
4686   const PPCSubtarget *STICaller = &TM.getSubtarget<PPCSubtarget>(*Caller);
4687   assert(!STICaller->isUsingPCRelativeCalls() &&
4688          "PC Relative callers do not have a TOC and cannot share a TOC Base");
4689 #endif
4690 
4691   // Callee is either a GlobalAddress or an ExternalSymbol. ExternalSymbols
4692   // don't have enough information to determine if the caller and callee share
4693   // the same  TOC base, so we have to pessimistically assume they don't for
4694   // correctness.
4695   GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
4696   if (!G)
4697     return false;
4698 
4699   const GlobalValue *GV = G->getGlobal();
4700 
4701   // If the callee is preemptable, then the static linker will use a plt-stub
4702   // which saves the toc to the stack, and needs a nop after the call
4703   // instruction to convert to a toc-restore.
4704   if (!TM.shouldAssumeDSOLocal(*Caller->getParent(), GV))
4705     return false;
4706 
4707   // Functions with PC Relative enabled may clobber the TOC in the same DSO.
4708   // We may need a TOC restore in the situation where the caller requires a
4709   // valid TOC but the callee is PC Relative and does not.
4710   const Function *F = dyn_cast<Function>(GV);
4711   const GlobalAlias *Alias = dyn_cast<GlobalAlias>(GV);
4712 
4713   // If we have an Alias we can try to get the function from there.
4714   if (Alias) {
4715     const GlobalObject *GlobalObj = Alias->getAliaseeObject();
4716     F = dyn_cast<Function>(GlobalObj);
4717   }
4718 
4719   // If we still have no valid function pointer we do not have enough
4720   // information to determine if the callee uses PC Relative calls so we must
4721   // assume that it does.
4722   if (!F)
4723     return false;
4724 
4725   // If the callee uses PC Relative we cannot guarantee that the callee won't
4726   // clobber the TOC of the caller and so we must assume that the two
4727   // functions do not share a TOC base.
4728   const PPCSubtarget *STICallee = &TM.getSubtarget<PPCSubtarget>(*F);
4729   if (STICallee->isUsingPCRelativeCalls())
4730     return false;
4731 
4732   // If the GV is not a strong definition then we need to assume it can be
4733   // replaced by another function at link time. The function that replaces
4734   // it may not share the same TOC as the caller since the callee may be
4735   // replaced by a PC Relative version of the same function.
4736   if (!GV->isStrongDefinitionForLinker())
4737     return false;
4738 
4739   // The medium and large code models are expected to provide a sufficiently
4740   // large TOC to provide all data addressing needs of a module with a
4741   // single TOC.
4742   if (CodeModel::Medium == TM.getCodeModel() ||
4743       CodeModel::Large == TM.getCodeModel())
4744     return true;
4745 
4746   // Any explicitly-specified sections and section prefixes must also match.
4747   // Also, if we're using -ffunction-sections, then each function is always in
4748   // a different section (the same is true for COMDAT functions).
4749   if (TM.getFunctionSections() || GV->hasComdat() || Caller->hasComdat() ||
4750       GV->getSection() != Caller->getSection())
4751     return false;
4752   if (const auto *F = dyn_cast<Function>(GV)) {
4753     if (F->getSectionPrefix() != Caller->getSectionPrefix())
4754       return false;
4755   }
4756 
4757   return true;
4758 }
4759 
4760 static bool
4761 needStackSlotPassParameters(const PPCSubtarget &Subtarget,
4762                             const SmallVectorImpl<ISD::OutputArg> &Outs) {
4763   assert(Subtarget.is64BitELFABI());
4764 
4765   const unsigned PtrByteSize = 8;
4766   const unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
4767 
4768   static const MCPhysReg GPR[] = {
4769     PPC::X3, PPC::X4, PPC::X5, PPC::X6,
4770     PPC::X7, PPC::X8, PPC::X9, PPC::X10,
4771   };
4772   static const MCPhysReg VR[] = {
4773     PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
4774     PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
4775   };
4776 
4777   const unsigned NumGPRs = std::size(GPR);
4778   const unsigned NumFPRs = 13;
4779   const unsigned NumVRs = std::size(VR);
4780   const unsigned ParamAreaSize = NumGPRs * PtrByteSize;
4781 
4782   unsigned NumBytes = LinkageSize;
4783   unsigned AvailableFPRs = NumFPRs;
4784   unsigned AvailableVRs = NumVRs;
4785 
4786   for (const ISD::OutputArg& Param : Outs) {
4787     if (Param.Flags.isNest()) continue;
4788 
4789     if (CalculateStackSlotUsed(Param.VT, Param.ArgVT, Param.Flags, PtrByteSize,
4790                                LinkageSize, ParamAreaSize, NumBytes,
4791                                AvailableFPRs, AvailableVRs))
4792       return true;
4793   }
4794   return false;
4795 }
4796 
4797 static bool hasSameArgumentList(const Function *CallerFn, const CallBase &CB) {
4798   if (CB.arg_size() != CallerFn->arg_size())
4799     return false;
4800 
4801   auto CalleeArgIter = CB.arg_begin();
4802   auto CalleeArgEnd = CB.arg_end();
4803   Function::const_arg_iterator CallerArgIter = CallerFn->arg_begin();
4804 
4805   for (; CalleeArgIter != CalleeArgEnd; ++CalleeArgIter, ++CallerArgIter) {
4806     const Value* CalleeArg = *CalleeArgIter;
4807     const Value* CallerArg = &(*CallerArgIter);
4808     if (CalleeArg == CallerArg)
4809       continue;
4810 
4811     // e.g. @caller([4 x i64] %a, [4 x i64] %b) {
4812     //        tail call @callee([4 x i64] undef, [4 x i64] %b)
4813     //      }
4814     // 1st argument of callee is undef and has the same type as caller.
4815     if (CalleeArg->getType() == CallerArg->getType() &&
4816         isa<UndefValue>(CalleeArg))
4817       continue;
4818 
4819     return false;
4820   }
4821 
4822   return true;
4823 }
4824 
4825 // Returns true if TCO is possible between the callers and callees
4826 // calling conventions.
4827 static bool
4828 areCallingConvEligibleForTCO_64SVR4(CallingConv::ID CallerCC,
4829                                     CallingConv::ID CalleeCC) {
4830   // Tail calls are possible with fastcc and ccc.
4831   auto isTailCallableCC  = [] (CallingConv::ID CC){
4832       return  CC == CallingConv::C || CC == CallingConv::Fast;
4833   };
4834   if (!isTailCallableCC(CallerCC) || !isTailCallableCC(CalleeCC))
4835     return false;
4836 
4837   // We can safely tail call both fastcc and ccc callees from a c calling
4838   // convention caller. If the caller is fastcc, we may have less stack space
4839   // than a non-fastcc caller with the same signature so disable tail-calls in
4840   // that case.
4841   return CallerCC == CallingConv::C || CallerCC == CalleeCC;
4842 }
4843 
4844 bool PPCTargetLowering::IsEligibleForTailCallOptimization_64SVR4(
4845     SDValue Callee, CallingConv::ID CalleeCC, const CallBase *CB, bool isVarArg,
4846     const SmallVectorImpl<ISD::OutputArg> &Outs,
4847     const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
4848   bool TailCallOpt = getTargetMachine().Options.GuaranteedTailCallOpt;
4849 
4850   if (DisableSCO && !TailCallOpt) return false;
4851 
4852   // Variadic argument functions are not supported.
4853   if (isVarArg) return false;
4854 
4855   auto &Caller = DAG.getMachineFunction().getFunction();
4856   // Check that the calling conventions are compatible for tco.
4857   if (!areCallingConvEligibleForTCO_64SVR4(Caller.getCallingConv(), CalleeCC))
4858     return false;
4859 
4860   // Caller contains any byval parameter is not supported.
4861   if (any_of(Ins, [](const ISD::InputArg &IA) { return IA.Flags.isByVal(); }))
4862     return false;
4863 
4864   // Callee contains any byval parameter is not supported, too.
4865   // Note: This is a quick work around, because in some cases, e.g.
4866   // caller's stack size > callee's stack size, we are still able to apply
4867   // sibling call optimization. For example, gcc is able to do SCO for caller1
4868   // in the following example, but not for caller2.
4869   //   struct test {
4870   //     long int a;
4871   //     char ary[56];
4872   //   } gTest;
4873   //   __attribute__((noinline)) int callee(struct test v, struct test *b) {
4874   //     b->a = v.a;
4875   //     return 0;
4876   //   }
4877   //   void caller1(struct test a, struct test c, struct test *b) {
4878   //     callee(gTest, b); }
4879   //   void caller2(struct test *b) { callee(gTest, b); }
4880   if (any_of(Outs, [](const ISD::OutputArg& OA) { return OA.Flags.isByVal(); }))
4881     return false;
4882 
4883   // If callee and caller use different calling conventions, we cannot pass
4884   // parameters on stack since offsets for the parameter area may be different.
4885   if (Caller.getCallingConv() != CalleeCC &&
4886       needStackSlotPassParameters(Subtarget, Outs))
4887     return false;
4888 
4889   // All variants of 64-bit ELF ABIs without PC-Relative addressing require that
4890   // the caller and callee share the same TOC for TCO/SCO. If the caller and
4891   // callee potentially have different TOC bases then we cannot tail call since
4892   // we need to restore the TOC pointer after the call.
4893   // ref: https://bugzilla.mozilla.org/show_bug.cgi?id=973977
4894   // We cannot guarantee this for indirect calls or calls to external functions.
4895   // When PC-Relative addressing is used, the concept of the TOC is no longer
4896   // applicable so this check is not required.
4897   // Check first for indirect calls.
4898   if (!Subtarget.isUsingPCRelativeCalls() &&
4899       !isFunctionGlobalAddress(Callee) && !isa<ExternalSymbolSDNode>(Callee))
4900     return false;
4901 
4902   // Check if we share the TOC base.
4903   if (!Subtarget.isUsingPCRelativeCalls() &&
4904       !callsShareTOCBase(&Caller, Callee, getTargetMachine()))
4905     return false;
4906 
4907   // TCO allows altering callee ABI, so we don't have to check further.
4908   if (CalleeCC == CallingConv::Fast && TailCallOpt)
4909     return true;
4910 
4911   if (DisableSCO) return false;
4912 
4913   // If callee use the same argument list that caller is using, then we can
4914   // apply SCO on this case. If it is not, then we need to check if callee needs
4915   // stack for passing arguments.
4916   // PC Relative tail calls may not have a CallBase.
4917   // If there is no CallBase we cannot verify if we have the same argument
4918   // list so assume that we don't have the same argument list.
4919   if (CB && !hasSameArgumentList(&Caller, *CB) &&
4920       needStackSlotPassParameters(Subtarget, Outs))
4921     return false;
4922   else if (!CB && needStackSlotPassParameters(Subtarget, Outs))
4923     return false;
4924 
4925   return true;
4926 }
4927 
4928 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
4929 /// for tail call optimization. Targets which want to do tail call
4930 /// optimization should implement this function.
4931 bool
4932 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
4933                                                      CallingConv::ID CalleeCC,
4934                                                      bool isVarArg,
4935                                       const SmallVectorImpl<ISD::InputArg> &Ins,
4936                                                      SelectionDAG& DAG) const {
4937   if (!getTargetMachine().Options.GuaranteedTailCallOpt)
4938     return false;
4939 
4940   // Variable argument functions are not supported.
4941   if (isVarArg)
4942     return false;
4943 
4944   MachineFunction &MF = DAG.getMachineFunction();
4945   CallingConv::ID CallerCC = MF.getFunction().getCallingConv();
4946   if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
4947     // Functions containing by val parameters are not supported.
4948     for (unsigned i = 0; i != Ins.size(); i++) {
4949        ISD::ArgFlagsTy Flags = Ins[i].Flags;
4950        if (Flags.isByVal()) return false;
4951     }
4952 
4953     // Non-PIC/GOT tail calls are supported.
4954     if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
4955       return true;
4956 
4957     // At the moment we can only do local tail calls (in same module, hidden
4958     // or protected) if we are generating PIC.
4959     if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
4960       return G->getGlobal()->hasHiddenVisibility()
4961           || G->getGlobal()->hasProtectedVisibility();
4962   }
4963 
4964   return false;
4965 }
4966 
4967 /// isCallCompatibleAddress - Return the immediate to use if the specified
4968 /// 32-bit value is representable in the immediate field of a BxA instruction.
4969 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
4970   ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4971   if (!C) return nullptr;
4972 
4973   int Addr = C->getZExtValue();
4974   if ((Addr & 3) != 0 ||  // Low 2 bits are implicitly zero.
4975       SignExtend32<26>(Addr) != Addr)
4976     return nullptr;  // Top 6 bits have to be sext of immediate.
4977 
4978   return DAG
4979       .getConstant(
4980           (int)C->getZExtValue() >> 2, SDLoc(Op),
4981           DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout()))
4982       .getNode();
4983 }
4984 
4985 namespace {
4986 
4987 struct TailCallArgumentInfo {
4988   SDValue Arg;
4989   SDValue FrameIdxOp;
4990   int FrameIdx = 0;
4991 
4992   TailCallArgumentInfo() = default;
4993 };
4994 
4995 } // end anonymous namespace
4996 
4997 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
4998 static void StoreTailCallArgumentsToStackSlot(
4999     SelectionDAG &DAG, SDValue Chain,
5000     const SmallVectorImpl<TailCallArgumentInfo> &TailCallArgs,
5001     SmallVectorImpl<SDValue> &MemOpChains, const SDLoc &dl) {
5002   for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
5003     SDValue Arg = TailCallArgs[i].Arg;
5004     SDValue FIN = TailCallArgs[i].FrameIdxOp;
5005     int FI = TailCallArgs[i].FrameIdx;
5006     // Store relative to framepointer.
5007     MemOpChains.push_back(DAG.getStore(
5008         Chain, dl, Arg, FIN,
5009         MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)));
5010   }
5011 }
5012 
5013 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
5014 /// the appropriate stack slot for the tail call optimized function call.
5015 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG, SDValue Chain,
5016                                              SDValue OldRetAddr, SDValue OldFP,
5017                                              int SPDiff, const SDLoc &dl) {
5018   if (SPDiff) {
5019     // Calculate the new stack slot for the return address.
5020     MachineFunction &MF = DAG.getMachineFunction();
5021     const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>();
5022     const PPCFrameLowering *FL = Subtarget.getFrameLowering();
5023     bool isPPC64 = Subtarget.isPPC64();
5024     int SlotSize = isPPC64 ? 8 : 4;
5025     int NewRetAddrLoc = SPDiff + FL->getReturnSaveOffset();
5026     int NewRetAddr = MF.getFrameInfo().CreateFixedObject(SlotSize,
5027                                                          NewRetAddrLoc, true);
5028     EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
5029     SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
5030     Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
5031                          MachinePointerInfo::getFixedStack(MF, NewRetAddr));
5032   }
5033   return Chain;
5034 }
5035 
5036 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
5037 /// the position of the argument.
5038 static void
5039 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
5040                          SDValue Arg, int SPDiff, unsigned ArgOffset,
5041                      SmallVectorImpl<TailCallArgumentInfo>& TailCallArguments) {
5042   int Offset = ArgOffset + SPDiff;
5043   uint32_t OpSize = (Arg.getValueSizeInBits() + 7) / 8;
5044   int FI = MF.getFrameInfo().CreateFixedObject(OpSize, Offset, true);
5045   EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
5046   SDValue FIN = DAG.getFrameIndex(FI, VT);
5047   TailCallArgumentInfo Info;
5048   Info.Arg = Arg;
5049   Info.FrameIdxOp = FIN;
5050   Info.FrameIdx = FI;
5051   TailCallArguments.push_back(Info);
5052 }
5053 
5054 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
5055 /// stack slot. Returns the chain as result and the loaded frame pointers in
5056 /// LROpOut/FPOpout. Used when tail calling.
5057 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(
5058     SelectionDAG &DAG, int SPDiff, SDValue Chain, SDValue &LROpOut,
5059     SDValue &FPOpOut, const SDLoc &dl) const {
5060   if (SPDiff) {
5061     // Load the LR and FP stack slot for later adjusting.
5062     EVT VT = Subtarget.isPPC64() ? MVT::i64 : MVT::i32;
5063     LROpOut = getReturnAddrFrameIndex(DAG);
5064     LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo());
5065     Chain = SDValue(LROpOut.getNode(), 1);
5066   }
5067   return Chain;
5068 }
5069 
5070 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
5071 /// by "Src" to address "Dst" of size "Size".  Alignment information is
5072 /// specified by the specific parameter attribute. The copy will be passed as
5073 /// a byval function parameter.
5074 /// Sometimes what we are copying is the end of a larger object, the part that
5075 /// does not fit in registers.
5076 static SDValue CreateCopyOfByValArgument(SDValue Src, SDValue Dst,
5077                                          SDValue Chain, ISD::ArgFlagsTy Flags,
5078                                          SelectionDAG &DAG, const SDLoc &dl) {
5079   SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
5080   return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
5081                        Flags.getNonZeroByValAlign(), false, false, false,
5082                        MachinePointerInfo(), MachinePointerInfo());
5083 }
5084 
5085 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
5086 /// tail calls.
5087 static void LowerMemOpCallTo(
5088     SelectionDAG &DAG, MachineFunction &MF, SDValue Chain, SDValue Arg,
5089     SDValue PtrOff, int SPDiff, unsigned ArgOffset, bool isPPC64,
5090     bool isTailCall, bool isVector, SmallVectorImpl<SDValue> &MemOpChains,
5091     SmallVectorImpl<TailCallArgumentInfo> &TailCallArguments, const SDLoc &dl) {
5092   EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
5093   if (!isTailCall) {
5094     if (isVector) {
5095       SDValue StackPtr;
5096       if (isPPC64)
5097         StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
5098       else
5099         StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
5100       PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
5101                            DAG.getConstant(ArgOffset, dl, PtrVT));
5102     }
5103     MemOpChains.push_back(
5104         DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo()));
5105     // Calculate and remember argument location.
5106   } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
5107                                   TailCallArguments);
5108 }
5109 
5110 static void
5111 PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
5112                 const SDLoc &dl, int SPDiff, unsigned NumBytes, SDValue LROp,
5113                 SDValue FPOp,
5114                 SmallVectorImpl<TailCallArgumentInfo> &TailCallArguments) {
5115   // Emit a sequence of copyto/copyfrom virtual registers for arguments that
5116   // might overwrite each other in case of tail call optimization.
5117   SmallVector<SDValue, 8> MemOpChains2;
5118   // Do not flag preceding copytoreg stuff together with the following stuff.
5119   InFlag = SDValue();
5120   StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
5121                                     MemOpChains2, dl);
5122   if (!MemOpChains2.empty())
5123     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
5124 
5125   // Store the return address to the appropriate stack slot.
5126   Chain = EmitTailCallStoreFPAndRetAddr(DAG, Chain, LROp, FPOp, SPDiff, dl);
5127 
5128   // Emit callseq_end just before tailcall node.
5129   Chain = DAG.getCALLSEQ_END(Chain, NumBytes, 0, InFlag, dl);
5130   InFlag = Chain.getValue(1);
5131 }
5132 
5133 // Is this global address that of a function that can be called by name? (as
5134 // opposed to something that must hold a descriptor for an indirect call).
5135 static bool isFunctionGlobalAddress(SDValue Callee) {
5136   if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
5137     if (Callee.getOpcode() == ISD::GlobalTLSAddress ||
5138         Callee.getOpcode() == ISD::TargetGlobalTLSAddress)
5139       return false;
5140 
5141     return G->getGlobal()->getValueType()->isFunctionTy();
5142   }
5143 
5144   return false;
5145 }
5146 
5147 SDValue PPCTargetLowering::LowerCallResult(
5148     SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
5149     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
5150     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
5151   SmallVector<CCValAssign, 16> RVLocs;
5152   CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
5153                     *DAG.getContext());
5154 
5155   CCRetInfo.AnalyzeCallResult(
5156       Ins, (Subtarget.isSVR4ABI() && CallConv == CallingConv::Cold)
5157                ? RetCC_PPC_Cold
5158                : RetCC_PPC);
5159 
5160   // Copy all of the result registers out of their specified physreg.
5161   for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
5162     CCValAssign &VA = RVLocs[i];
5163     assert(VA.isRegLoc() && "Can only return in registers!");
5164 
5165     SDValue Val;
5166 
5167     if (Subtarget.hasSPE() && VA.getLocVT() == MVT::f64) {
5168       SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
5169                                       InFlag);
5170       Chain = Lo.getValue(1);
5171       InFlag = Lo.getValue(2);
5172       VA = RVLocs[++i]; // skip ahead to next loc
5173       SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
5174                                       InFlag);
5175       Chain = Hi.getValue(1);
5176       InFlag = Hi.getValue(2);
5177       if (!Subtarget.isLittleEndian())
5178         std::swap (Lo, Hi);
5179       Val = DAG.getNode(PPCISD::BUILD_SPE64, dl, MVT::f64, Lo, Hi);
5180     } else {
5181       Val = DAG.getCopyFromReg(Chain, dl,
5182                                VA.getLocReg(), VA.getLocVT(), InFlag);
5183       Chain = Val.getValue(1);
5184       InFlag = Val.getValue(2);
5185     }
5186 
5187     switch (VA.getLocInfo()) {
5188     default: llvm_unreachable("Unknown loc info!");
5189     case CCValAssign::Full: break;
5190     case CCValAssign::AExt:
5191       Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
5192       break;
5193     case CCValAssign::ZExt:
5194       Val = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), Val,
5195                         DAG.getValueType(VA.getValVT()));
5196       Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
5197       break;
5198     case CCValAssign::SExt:
5199       Val = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), Val,
5200                         DAG.getValueType(VA.getValVT()));
5201       Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
5202       break;
5203     }
5204 
5205     InVals.push_back(Val);
5206   }
5207 
5208   return Chain;
5209 }
5210 
5211 static bool isIndirectCall(const SDValue &Callee, SelectionDAG &DAG,
5212                            const PPCSubtarget &Subtarget, bool isPatchPoint) {
5213   // PatchPoint calls are not indirect.
5214   if (isPatchPoint)
5215     return false;
5216 
5217   if (isFunctionGlobalAddress(Callee) || isa<ExternalSymbolSDNode>(Callee))
5218     return false;
5219 
5220   // Darwin, and 32-bit ELF can use a BLA. The descriptor based ABIs can not
5221   // becuase the immediate function pointer points to a descriptor instead of
5222   // a function entry point. The ELFv2 ABI cannot use a BLA because the function
5223   // pointer immediate points to the global entry point, while the BLA would
5224   // need to jump to the local entry point (see rL211174).
5225   if (!Subtarget.usesFunctionDescriptors() && !Subtarget.isELFv2ABI() &&
5226       isBLACompatibleAddress(Callee, DAG))
5227     return false;
5228 
5229   return true;
5230 }
5231 
5232 // AIX and 64-bit ELF ABIs w/o PCRel require a TOC save/restore around calls.
5233 static inline bool isTOCSaveRestoreRequired(const PPCSubtarget &Subtarget) {
5234   return Subtarget.isAIXABI() ||
5235          (Subtarget.is64BitELFABI() && !Subtarget.isUsingPCRelativeCalls());
5236 }
5237 
5238 static unsigned getCallOpcode(PPCTargetLowering::CallFlags CFlags,
5239                               const Function &Caller, const SDValue &Callee,
5240                               const PPCSubtarget &Subtarget,
5241                               const TargetMachine &TM,
5242                               bool IsStrictFPCall = false) {
5243   if (CFlags.IsTailCall)
5244     return PPCISD::TC_RETURN;
5245 
5246   unsigned RetOpc = 0;
5247   // This is a call through a function pointer.
5248   if (CFlags.IsIndirect) {
5249     // AIX and the 64-bit ELF ABIs need to maintain the TOC pointer accross
5250     // indirect calls. The save of the caller's TOC pointer to the stack will be
5251     // inserted into the DAG as part of call lowering. The restore of the TOC
5252     // pointer is modeled by using a pseudo instruction for the call opcode that
5253     // represents the 2 instruction sequence of an indirect branch and link,
5254     // immediately followed by a load of the TOC pointer from the the stack save
5255     // slot into gpr2. For 64-bit ELFv2 ABI with PCRel, do not restore the TOC
5256     // as it is not saved or used.
5257     RetOpc = isTOCSaveRestoreRequired(Subtarget) ? PPCISD::BCTRL_LOAD_TOC
5258                                                  : PPCISD::BCTRL;
5259   } else if (Subtarget.isUsingPCRelativeCalls()) {
5260     assert(Subtarget.is64BitELFABI() && "PC Relative is only on ELF ABI.");
5261     RetOpc = PPCISD::CALL_NOTOC;
5262   } else if (Subtarget.isAIXABI() || Subtarget.is64BitELFABI())
5263     // The ABIs that maintain a TOC pointer accross calls need to have a nop
5264     // immediately following the call instruction if the caller and callee may
5265     // have different TOC bases. At link time if the linker determines the calls
5266     // may not share a TOC base, the call is redirected to a trampoline inserted
5267     // by the linker. The trampoline will (among other things) save the callers
5268     // TOC pointer at an ABI designated offset in the linkage area and the
5269     // linker will rewrite the nop to be a load of the TOC pointer from the
5270     // linkage area into gpr2.
5271     RetOpc = callsShareTOCBase(&Caller, Callee, TM) ? PPCISD::CALL
5272                                                     : PPCISD::CALL_NOP;
5273   else
5274     RetOpc = PPCISD::CALL;
5275   if (IsStrictFPCall) {
5276     switch (RetOpc) {
5277     default:
5278       llvm_unreachable("Unknown call opcode");
5279     case PPCISD::BCTRL_LOAD_TOC:
5280       RetOpc = PPCISD::BCTRL_LOAD_TOC_RM;
5281       break;
5282     case PPCISD::BCTRL:
5283       RetOpc = PPCISD::BCTRL_RM;
5284       break;
5285     case PPCISD::CALL_NOTOC:
5286       RetOpc = PPCISD::CALL_NOTOC_RM;
5287       break;
5288     case PPCISD::CALL:
5289       RetOpc = PPCISD::CALL_RM;
5290       break;
5291     case PPCISD::CALL_NOP:
5292       RetOpc = PPCISD::CALL_NOP_RM;
5293       break;
5294     }
5295   }
5296   return RetOpc;
5297 }
5298 
5299 static SDValue transformCallee(const SDValue &Callee, SelectionDAG &DAG,
5300                                const SDLoc &dl, const PPCSubtarget &Subtarget) {
5301   if (!Subtarget.usesFunctionDescriptors() && !Subtarget.isELFv2ABI())
5302     if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
5303       return SDValue(Dest, 0);
5304 
5305   // Returns true if the callee is local, and false otherwise.
5306   auto isLocalCallee = [&]() {
5307     const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
5308     const Module *Mod = DAG.getMachineFunction().getFunction().getParent();
5309     const GlobalValue *GV = G ? G->getGlobal() : nullptr;
5310 
5311     return DAG.getTarget().shouldAssumeDSOLocal(*Mod, GV) &&
5312            !isa_and_nonnull<GlobalIFunc>(GV);
5313   };
5314 
5315   // The PLT is only used in 32-bit ELF PIC mode.  Attempting to use the PLT in
5316   // a static relocation model causes some versions of GNU LD (2.17.50, at
5317   // least) to force BSS-PLT, instead of secure-PLT, even if all objects are
5318   // built with secure-PLT.
5319   bool UsePlt =
5320       Subtarget.is32BitELFABI() && !isLocalCallee() &&
5321       Subtarget.getTargetMachine().getRelocationModel() == Reloc::PIC_;
5322 
5323   const auto getAIXFuncEntryPointSymbolSDNode = [&](const GlobalValue *GV) {
5324     const TargetMachine &TM = Subtarget.getTargetMachine();
5325     const TargetLoweringObjectFile *TLOF = TM.getObjFileLowering();
5326     MCSymbolXCOFF *S =
5327         cast<MCSymbolXCOFF>(TLOF->getFunctionEntryPointSymbol(GV, TM));
5328 
5329     MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
5330     return DAG.getMCSymbol(S, PtrVT);
5331   };
5332 
5333   if (isFunctionGlobalAddress(Callee)) {
5334     const GlobalValue *GV = cast<GlobalAddressSDNode>(Callee)->getGlobal();
5335 
5336     if (Subtarget.isAIXABI()) {
5337       assert(!isa<GlobalIFunc>(GV) && "IFunc is not supported on AIX.");
5338       return getAIXFuncEntryPointSymbolSDNode(GV);
5339     }
5340     return DAG.getTargetGlobalAddress(GV, dl, Callee.getValueType(), 0,
5341                                       UsePlt ? PPCII::MO_PLT : 0);
5342   }
5343 
5344   if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
5345     const char *SymName = S->getSymbol();
5346     if (Subtarget.isAIXABI()) {
5347       // If there exists a user-declared function whose name is the same as the
5348       // ExternalSymbol's, then we pick up the user-declared version.
5349       const Module *Mod = DAG.getMachineFunction().getFunction().getParent();
5350       if (const Function *F =
5351               dyn_cast_or_null<Function>(Mod->getNamedValue(SymName)))
5352         return getAIXFuncEntryPointSymbolSDNode(F);
5353 
5354       // On AIX, direct function calls reference the symbol for the function's
5355       // entry point, which is named by prepending a "." before the function's
5356       // C-linkage name. A Qualname is returned here because an external
5357       // function entry point is a csect with XTY_ER property.
5358       const auto getExternalFunctionEntryPointSymbol = [&](StringRef SymName) {
5359         auto &Context = DAG.getMachineFunction().getMMI().getContext();
5360         MCSectionXCOFF *Sec = Context.getXCOFFSection(
5361             (Twine(".") + Twine(SymName)).str(), SectionKind::getMetadata(),
5362             XCOFF::CsectProperties(XCOFF::XMC_PR, XCOFF::XTY_ER));
5363         return Sec->getQualNameSymbol();
5364       };
5365 
5366       SymName = getExternalFunctionEntryPointSymbol(SymName)->getName().data();
5367     }
5368     return DAG.getTargetExternalSymbol(SymName, Callee.getValueType(),
5369                                        UsePlt ? PPCII::MO_PLT : 0);
5370   }
5371 
5372   // No transformation needed.
5373   assert(Callee.getNode() && "What no callee?");
5374   return Callee;
5375 }
5376 
5377 static SDValue getOutputChainFromCallSeq(SDValue CallSeqStart) {
5378   assert(CallSeqStart.getOpcode() == ISD::CALLSEQ_START &&
5379          "Expected a CALLSEQ_STARTSDNode.");
5380 
5381   // The last operand is the chain, except when the node has glue. If the node
5382   // has glue, then the last operand is the glue, and the chain is the second
5383   // last operand.
5384   SDValue LastValue = CallSeqStart.getValue(CallSeqStart->getNumValues() - 1);
5385   if (LastValue.getValueType() != MVT::Glue)
5386     return LastValue;
5387 
5388   return CallSeqStart.getValue(CallSeqStart->getNumValues() - 2);
5389 }
5390 
5391 // Creates the node that moves a functions address into the count register
5392 // to prepare for an indirect call instruction.
5393 static void prepareIndirectCall(SelectionDAG &DAG, SDValue &Callee,
5394                                 SDValue &Glue, SDValue &Chain,
5395                                 const SDLoc &dl) {
5396   SDValue MTCTROps[] = {Chain, Callee, Glue};
5397   EVT ReturnTypes[] = {MVT::Other, MVT::Glue};
5398   Chain = DAG.getNode(PPCISD::MTCTR, dl, ArrayRef(ReturnTypes, 2),
5399                       ArrayRef(MTCTROps, Glue.getNode() ? 3 : 2));
5400   // The glue is the second value produced.
5401   Glue = Chain.getValue(1);
5402 }
5403 
5404 static void prepareDescriptorIndirectCall(SelectionDAG &DAG, SDValue &Callee,
5405                                           SDValue &Glue, SDValue &Chain,
5406                                           SDValue CallSeqStart,
5407                                           const CallBase *CB, const SDLoc &dl,
5408                                           bool hasNest,
5409                                           const PPCSubtarget &Subtarget) {
5410   // Function pointers in the 64-bit SVR4 ABI do not point to the function
5411   // entry point, but to the function descriptor (the function entry point
5412   // address is part of the function descriptor though).
5413   // The function descriptor is a three doubleword structure with the
5414   // following fields: function entry point, TOC base address and
5415   // environment pointer.
5416   // Thus for a call through a function pointer, the following actions need
5417   // to be performed:
5418   //   1. Save the TOC of the caller in the TOC save area of its stack
5419   //      frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()).
5420   //   2. Load the address of the function entry point from the function
5421   //      descriptor.
5422   //   3. Load the TOC of the callee from the function descriptor into r2.
5423   //   4. Load the environment pointer from the function descriptor into
5424   //      r11.
5425   //   5. Branch to the function entry point address.
5426   //   6. On return of the callee, the TOC of the caller needs to be
5427   //      restored (this is done in FinishCall()).
5428   //
5429   // The loads are scheduled at the beginning of the call sequence, and the
5430   // register copies are flagged together to ensure that no other
5431   // operations can be scheduled in between. E.g. without flagging the
5432   // copies together, a TOC access in the caller could be scheduled between
5433   // the assignment of the callee TOC and the branch to the callee, which leads
5434   // to incorrect code.
5435 
5436   // Start by loading the function address from the descriptor.
5437   SDValue LDChain = getOutputChainFromCallSeq(CallSeqStart);
5438   auto MMOFlags = Subtarget.hasInvariantFunctionDescriptors()
5439                       ? (MachineMemOperand::MODereferenceable |
5440                          MachineMemOperand::MOInvariant)
5441                       : MachineMemOperand::MONone;
5442 
5443   MachinePointerInfo MPI(CB ? CB->getCalledOperand() : nullptr);
5444 
5445   // Registers used in building the DAG.
5446   const MCRegister EnvPtrReg = Subtarget.getEnvironmentPointerRegister();
5447   const MCRegister TOCReg = Subtarget.getTOCPointerRegister();
5448 
5449   // Offsets of descriptor members.
5450   const unsigned TOCAnchorOffset = Subtarget.descriptorTOCAnchorOffset();
5451   const unsigned EnvPtrOffset = Subtarget.descriptorEnvironmentPointerOffset();
5452 
5453   const MVT RegVT = Subtarget.isPPC64() ? MVT::i64 : MVT::i32;
5454   const Align Alignment = Subtarget.isPPC64() ? Align(8) : Align(4);
5455 
5456   // One load for the functions entry point address.
5457   SDValue LoadFuncPtr = DAG.getLoad(RegVT, dl, LDChain, Callee, MPI,
5458                                     Alignment, MMOFlags);
5459 
5460   // One for loading the TOC anchor for the module that contains the called
5461   // function.
5462   SDValue TOCOff = DAG.getIntPtrConstant(TOCAnchorOffset, dl);
5463   SDValue AddTOC = DAG.getNode(ISD::ADD, dl, RegVT, Callee, TOCOff);
5464   SDValue TOCPtr =
5465       DAG.getLoad(RegVT, dl, LDChain, AddTOC,
5466                   MPI.getWithOffset(TOCAnchorOffset), Alignment, MMOFlags);
5467 
5468   // One for loading the environment pointer.
5469   SDValue PtrOff = DAG.getIntPtrConstant(EnvPtrOffset, dl);
5470   SDValue AddPtr = DAG.getNode(ISD::ADD, dl, RegVT, Callee, PtrOff);
5471   SDValue LoadEnvPtr =
5472       DAG.getLoad(RegVT, dl, LDChain, AddPtr,
5473                   MPI.getWithOffset(EnvPtrOffset), Alignment, MMOFlags);
5474 
5475 
5476   // Then copy the newly loaded TOC anchor to the TOC pointer.
5477   SDValue TOCVal = DAG.getCopyToReg(Chain, dl, TOCReg, TOCPtr, Glue);
5478   Chain = TOCVal.getValue(0);
5479   Glue = TOCVal.getValue(1);
5480 
5481   // If the function call has an explicit 'nest' parameter, it takes the
5482   // place of the environment pointer.
5483   assert((!hasNest || !Subtarget.isAIXABI()) &&
5484          "Nest parameter is not supported on AIX.");
5485   if (!hasNest) {
5486     SDValue EnvVal = DAG.getCopyToReg(Chain, dl, EnvPtrReg, LoadEnvPtr, Glue);
5487     Chain = EnvVal.getValue(0);
5488     Glue = EnvVal.getValue(1);
5489   }
5490 
5491   // The rest of the indirect call sequence is the same as the non-descriptor
5492   // DAG.
5493   prepareIndirectCall(DAG, LoadFuncPtr, Glue, Chain, dl);
5494 }
5495 
5496 static void
5497 buildCallOperands(SmallVectorImpl<SDValue> &Ops,
5498                   PPCTargetLowering::CallFlags CFlags, const SDLoc &dl,
5499                   SelectionDAG &DAG,
5500                   SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
5501                   SDValue Glue, SDValue Chain, SDValue &Callee, int SPDiff,
5502                   const PPCSubtarget &Subtarget) {
5503   const bool IsPPC64 = Subtarget.isPPC64();
5504   // MVT for a general purpose register.
5505   const MVT RegVT = IsPPC64 ? MVT::i64 : MVT::i32;
5506 
5507   // First operand is always the chain.
5508   Ops.push_back(Chain);
5509 
5510   // If it's a direct call pass the callee as the second operand.
5511   if (!CFlags.IsIndirect)
5512     Ops.push_back(Callee);
5513   else {
5514     assert(!CFlags.IsPatchPoint && "Patch point calls are not indirect.");
5515 
5516     // For the TOC based ABIs, we have saved the TOC pointer to the linkage area
5517     // on the stack (this would have been done in `LowerCall_64SVR4` or
5518     // `LowerCall_AIX`). The call instruction is a pseudo instruction that
5519     // represents both the indirect branch and a load that restores the TOC
5520     // pointer from the linkage area. The operand for the TOC restore is an add
5521     // of the TOC save offset to the stack pointer. This must be the second
5522     // operand: after the chain input but before any other variadic arguments.
5523     // For 64-bit ELFv2 ABI with PCRel, do not restore the TOC as it is not
5524     // saved or used.
5525     if (isTOCSaveRestoreRequired(Subtarget)) {
5526       const MCRegister StackPtrReg = Subtarget.getStackPointerRegister();
5527 
5528       SDValue StackPtr = DAG.getRegister(StackPtrReg, RegVT);
5529       unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset();
5530       SDValue TOCOff = DAG.getIntPtrConstant(TOCSaveOffset, dl);
5531       SDValue AddTOC = DAG.getNode(ISD::ADD, dl, RegVT, StackPtr, TOCOff);
5532       Ops.push_back(AddTOC);
5533     }
5534 
5535     // Add the register used for the environment pointer.
5536     if (Subtarget.usesFunctionDescriptors() && !CFlags.HasNest)
5537       Ops.push_back(DAG.getRegister(Subtarget.getEnvironmentPointerRegister(),
5538                                     RegVT));
5539 
5540 
5541     // Add CTR register as callee so a bctr can be emitted later.
5542     if (CFlags.IsTailCall)
5543       Ops.push_back(DAG.getRegister(IsPPC64 ? PPC::CTR8 : PPC::CTR, RegVT));
5544   }
5545 
5546   // If this is a tail call add stack pointer delta.
5547   if (CFlags.IsTailCall)
5548     Ops.push_back(DAG.getConstant(SPDiff, dl, MVT::i32));
5549 
5550   // Add argument registers to the end of the list so that they are known live
5551   // into the call.
5552   for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
5553     Ops.push_back(DAG.getRegister(RegsToPass[i].first,
5554                                   RegsToPass[i].second.getValueType()));
5555 
5556   // We cannot add R2/X2 as an operand here for PATCHPOINT, because there is
5557   // no way to mark dependencies as implicit here.
5558   // We will add the R2/X2 dependency in EmitInstrWithCustomInserter.
5559   if ((Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) &&
5560        !CFlags.IsPatchPoint && !Subtarget.isUsingPCRelativeCalls())
5561     Ops.push_back(DAG.getRegister(Subtarget.getTOCPointerRegister(), RegVT));
5562 
5563   // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
5564   if (CFlags.IsVarArg && Subtarget.is32BitELFABI())
5565     Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
5566 
5567   // Add a register mask operand representing the call-preserved registers.
5568   const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
5569   const uint32_t *Mask =
5570       TRI->getCallPreservedMask(DAG.getMachineFunction(), CFlags.CallConv);
5571   assert(Mask && "Missing call preserved mask for calling convention");
5572   Ops.push_back(DAG.getRegisterMask(Mask));
5573 
5574   // If the glue is valid, it is the last operand.
5575   if (Glue.getNode())
5576     Ops.push_back(Glue);
5577 }
5578 
5579 SDValue PPCTargetLowering::FinishCall(
5580     CallFlags CFlags, const SDLoc &dl, SelectionDAG &DAG,
5581     SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass, SDValue Glue,
5582     SDValue Chain, SDValue CallSeqStart, SDValue &Callee, int SPDiff,
5583     unsigned NumBytes, const SmallVectorImpl<ISD::InputArg> &Ins,
5584     SmallVectorImpl<SDValue> &InVals, const CallBase *CB) const {
5585 
5586   if ((Subtarget.is64BitELFABI() && !Subtarget.isUsingPCRelativeCalls()) ||
5587       Subtarget.isAIXABI())
5588     setUsesTOCBasePtr(DAG);
5589 
5590   unsigned CallOpc =
5591       getCallOpcode(CFlags, DAG.getMachineFunction().getFunction(), Callee,
5592                     Subtarget, DAG.getTarget(), CB ? CB->isStrictFP() : false);
5593 
5594   if (!CFlags.IsIndirect)
5595     Callee = transformCallee(Callee, DAG, dl, Subtarget);
5596   else if (Subtarget.usesFunctionDescriptors())
5597     prepareDescriptorIndirectCall(DAG, Callee, Glue, Chain, CallSeqStart, CB,
5598                                   dl, CFlags.HasNest, Subtarget);
5599   else
5600     prepareIndirectCall(DAG, Callee, Glue, Chain, dl);
5601 
5602   // Build the operand list for the call instruction.
5603   SmallVector<SDValue, 8> Ops;
5604   buildCallOperands(Ops, CFlags, dl, DAG, RegsToPass, Glue, Chain, Callee,
5605                     SPDiff, Subtarget);
5606 
5607   // Emit tail call.
5608   if (CFlags.IsTailCall) {
5609     // Indirect tail call when using PC Relative calls do not have the same
5610     // constraints.
5611     assert(((Callee.getOpcode() == ISD::Register &&
5612              cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
5613             Callee.getOpcode() == ISD::TargetExternalSymbol ||
5614             Callee.getOpcode() == ISD::TargetGlobalAddress ||
5615             isa<ConstantSDNode>(Callee) ||
5616             (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) &&
5617            "Expecting a global address, external symbol, absolute value, "
5618            "register or an indirect tail call when PC Relative calls are "
5619            "used.");
5620     // PC Relative calls also use TC_RETURN as the way to mark tail calls.
5621     assert(CallOpc == PPCISD::TC_RETURN &&
5622            "Unexpected call opcode for a tail call.");
5623     DAG.getMachineFunction().getFrameInfo().setHasTailCall();
5624     return DAG.getNode(CallOpc, dl, MVT::Other, Ops);
5625   }
5626 
5627   std::array<EVT, 2> ReturnTypes = {{MVT::Other, MVT::Glue}};
5628   Chain = DAG.getNode(CallOpc, dl, ReturnTypes, Ops);
5629   DAG.addNoMergeSiteInfo(Chain.getNode(), CFlags.NoMerge);
5630   Glue = Chain.getValue(1);
5631 
5632   // When performing tail call optimization the callee pops its arguments off
5633   // the stack. Account for this here so these bytes can be pushed back on in
5634   // PPCFrameLowering::eliminateCallFramePseudoInstr.
5635   int BytesCalleePops = (CFlags.CallConv == CallingConv::Fast &&
5636                          getTargetMachine().Options.GuaranteedTailCallOpt)
5637                             ? NumBytes
5638                             : 0;
5639 
5640   Chain = DAG.getCALLSEQ_END(Chain, NumBytes, BytesCalleePops, Glue, dl);
5641   Glue = Chain.getValue(1);
5642 
5643   return LowerCallResult(Chain, Glue, CFlags.CallConv, CFlags.IsVarArg, Ins, dl,
5644                          DAG, InVals);
5645 }
5646 
5647 SDValue
5648 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
5649                              SmallVectorImpl<SDValue> &InVals) const {
5650   SelectionDAG &DAG                     = CLI.DAG;
5651   SDLoc &dl                             = CLI.DL;
5652   SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
5653   SmallVectorImpl<SDValue> &OutVals     = CLI.OutVals;
5654   SmallVectorImpl<ISD::InputArg> &Ins   = CLI.Ins;
5655   SDValue Chain                         = CLI.Chain;
5656   SDValue Callee                        = CLI.Callee;
5657   bool &isTailCall                      = CLI.IsTailCall;
5658   CallingConv::ID CallConv              = CLI.CallConv;
5659   bool isVarArg                         = CLI.IsVarArg;
5660   bool isPatchPoint                     = CLI.IsPatchPoint;
5661   const CallBase *CB                    = CLI.CB;
5662 
5663   if (isTailCall) {
5664     if (Subtarget.useLongCalls() && !(CB && CB->isMustTailCall()))
5665       isTailCall = false;
5666     else if (Subtarget.isSVR4ABI() && Subtarget.isPPC64())
5667       isTailCall = IsEligibleForTailCallOptimization_64SVR4(
5668           Callee, CallConv, CB, isVarArg, Outs, Ins, DAG);
5669     else
5670       isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
5671                                                      Ins, DAG);
5672     if (isTailCall) {
5673       ++NumTailCalls;
5674       if (!getTargetMachine().Options.GuaranteedTailCallOpt)
5675         ++NumSiblingCalls;
5676 
5677       // PC Relative calls no longer guarantee that the callee is a Global
5678       // Address Node. The callee could be an indirect tail call in which
5679       // case the SDValue for the callee could be a load (to load the address
5680       // of a function pointer) or it may be a register copy (to move the
5681       // address of the callee from a function parameter into a virtual
5682       // register). It may also be an ExternalSymbolSDNode (ex memcopy).
5683       assert((Subtarget.isUsingPCRelativeCalls() ||
5684               isa<GlobalAddressSDNode>(Callee)) &&
5685              "Callee should be an llvm::Function object.");
5686 
5687       LLVM_DEBUG(dbgs() << "TCO caller: " << DAG.getMachineFunction().getName()
5688                         << "\nTCO callee: ");
5689       LLVM_DEBUG(Callee.dump());
5690     }
5691   }
5692 
5693   if (!isTailCall && CB && CB->isMustTailCall())
5694     report_fatal_error("failed to perform tail call elimination on a call "
5695                        "site marked musttail");
5696 
5697   // When long calls (i.e. indirect calls) are always used, calls are always
5698   // made via function pointer. If we have a function name, first translate it
5699   // into a pointer.
5700   if (Subtarget.useLongCalls() && isa<GlobalAddressSDNode>(Callee) &&
5701       !isTailCall)
5702     Callee = LowerGlobalAddress(Callee, DAG);
5703 
5704   CallFlags CFlags(
5705       CallConv, isTailCall, isVarArg, isPatchPoint,
5706       isIndirectCall(Callee, DAG, Subtarget, isPatchPoint),
5707       // hasNest
5708       Subtarget.is64BitELFABI() &&
5709           any_of(Outs, [](ISD::OutputArg Arg) { return Arg.Flags.isNest(); }),
5710       CLI.NoMerge);
5711 
5712   if (Subtarget.isAIXABI())
5713     return LowerCall_AIX(Chain, Callee, CFlags, Outs, OutVals, Ins, dl, DAG,
5714                          InVals, CB);
5715 
5716   assert(Subtarget.isSVR4ABI());
5717   if (Subtarget.isPPC64())
5718     return LowerCall_64SVR4(Chain, Callee, CFlags, Outs, OutVals, Ins, dl, DAG,
5719                             InVals, CB);
5720   return LowerCall_32SVR4(Chain, Callee, CFlags, Outs, OutVals, Ins, dl, DAG,
5721                           InVals, CB);
5722 }
5723 
5724 SDValue PPCTargetLowering::LowerCall_32SVR4(
5725     SDValue Chain, SDValue Callee, CallFlags CFlags,
5726     const SmallVectorImpl<ISD::OutputArg> &Outs,
5727     const SmallVectorImpl<SDValue> &OutVals,
5728     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
5729     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals,
5730     const CallBase *CB) const {
5731   // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
5732   // of the 32-bit SVR4 ABI stack frame layout.
5733 
5734   const CallingConv::ID CallConv = CFlags.CallConv;
5735   const bool IsVarArg = CFlags.IsVarArg;
5736   const bool IsTailCall = CFlags.IsTailCall;
5737 
5738   assert((CallConv == CallingConv::C ||
5739           CallConv == CallingConv::Cold ||
5740           CallConv == CallingConv::Fast) && "Unknown calling convention!");
5741 
5742   const Align PtrAlign(4);
5743 
5744   MachineFunction &MF = DAG.getMachineFunction();
5745 
5746   // Mark this function as potentially containing a function that contains a
5747   // tail call. As a consequence the frame pointer will be used for dynamicalloc
5748   // and restoring the callers stack pointer in this functions epilog. This is
5749   // done because by tail calling the called function might overwrite the value
5750   // in this function's (MF) stack pointer stack slot 0(SP).
5751   if (getTargetMachine().Options.GuaranteedTailCallOpt &&
5752       CallConv == CallingConv::Fast)
5753     MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
5754 
5755   // Count how many bytes are to be pushed on the stack, including the linkage
5756   // area, parameter list area and the part of the local variable space which
5757   // contains copies of aggregates which are passed by value.
5758 
5759   // Assign locations to all of the outgoing arguments.
5760   SmallVector<CCValAssign, 16> ArgLocs;
5761   PPCCCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
5762 
5763   // Reserve space for the linkage area on the stack.
5764   CCInfo.AllocateStack(Subtarget.getFrameLowering()->getLinkageSize(),
5765                        PtrAlign);
5766   if (useSoftFloat())
5767     CCInfo.PreAnalyzeCallOperands(Outs);
5768 
5769   if (IsVarArg) {
5770     // Handle fixed and variable vector arguments differently.
5771     // Fixed vector arguments go into registers as long as registers are
5772     // available. Variable vector arguments always go into memory.
5773     unsigned NumArgs = Outs.size();
5774 
5775     for (unsigned i = 0; i != NumArgs; ++i) {
5776       MVT ArgVT = Outs[i].VT;
5777       ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
5778       bool Result;
5779 
5780       if (Outs[i].IsFixed) {
5781         Result = CC_PPC32_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
5782                                CCInfo);
5783       } else {
5784         Result = CC_PPC32_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
5785                                       ArgFlags, CCInfo);
5786       }
5787 
5788       if (Result) {
5789 #ifndef NDEBUG
5790         errs() << "Call operand #" << i << " has unhandled type "
5791              << EVT(ArgVT).getEVTString() << "\n";
5792 #endif
5793         llvm_unreachable(nullptr);
5794       }
5795     }
5796   } else {
5797     // All arguments are treated the same.
5798     CCInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4);
5799   }
5800   CCInfo.clearWasPPCF128();
5801 
5802   // Assign locations to all of the outgoing aggregate by value arguments.
5803   SmallVector<CCValAssign, 16> ByValArgLocs;
5804   CCState CCByValInfo(CallConv, IsVarArg, MF, ByValArgLocs, *DAG.getContext());
5805 
5806   // Reserve stack space for the allocations in CCInfo.
5807   CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrAlign);
5808 
5809   CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4_ByVal);
5810 
5811   // Size of the linkage area, parameter list area and the part of the local
5812   // space variable where copies of aggregates which are passed by value are
5813   // stored.
5814   unsigned NumBytes = CCByValInfo.getNextStackOffset();
5815 
5816   // Calculate by how many bytes the stack has to be adjusted in case of tail
5817   // call optimization.
5818   int SPDiff = CalculateTailCallSPDiff(DAG, IsTailCall, NumBytes);
5819 
5820   // Adjust the stack pointer for the new arguments...
5821   // These operations are automatically eliminated by the prolog/epilog pass
5822   Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl);
5823   SDValue CallSeqStart = Chain;
5824 
5825   // Load the return address and frame pointer so it can be moved somewhere else
5826   // later.
5827   SDValue LROp, FPOp;
5828   Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl);
5829 
5830   // Set up a copy of the stack pointer for use loading and storing any
5831   // arguments that may not fit in the registers available for argument
5832   // passing.
5833   SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
5834 
5835   SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
5836   SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
5837   SmallVector<SDValue, 8> MemOpChains;
5838 
5839   bool seenFloatArg = false;
5840   // Walk the register/memloc assignments, inserting copies/loads.
5841   // i - Tracks the index into the list of registers allocated for the call
5842   // RealArgIdx - Tracks the index into the list of actual function arguments
5843   // j - Tracks the index into the list of byval arguments
5844   for (unsigned i = 0, RealArgIdx = 0, j = 0, e = ArgLocs.size();
5845        i != e;
5846        ++i, ++RealArgIdx) {
5847     CCValAssign &VA = ArgLocs[i];
5848     SDValue Arg = OutVals[RealArgIdx];
5849     ISD::ArgFlagsTy Flags = Outs[RealArgIdx].Flags;
5850 
5851     if (Flags.isByVal()) {
5852       // Argument is an aggregate which is passed by value, thus we need to
5853       // create a copy of it in the local variable space of the current stack
5854       // frame (which is the stack frame of the caller) and pass the address of
5855       // this copy to the callee.
5856       assert((j < ByValArgLocs.size()) && "Index out of bounds!");
5857       CCValAssign &ByValVA = ByValArgLocs[j++];
5858       assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
5859 
5860       // Memory reserved in the local variable space of the callers stack frame.
5861       unsigned LocMemOffset = ByValVA.getLocMemOffset();
5862 
5863       SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
5864       PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(MF.getDataLayout()),
5865                            StackPtr, PtrOff);
5866 
5867       // Create a copy of the argument in the local area of the current
5868       // stack frame.
5869       SDValue MemcpyCall =
5870         CreateCopyOfByValArgument(Arg, PtrOff,
5871                                   CallSeqStart.getNode()->getOperand(0),
5872                                   Flags, DAG, dl);
5873 
5874       // This must go outside the CALLSEQ_START..END.
5875       SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall, NumBytes, 0,
5876                                                      SDLoc(MemcpyCall));
5877       DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
5878                              NewCallSeqStart.getNode());
5879       Chain = CallSeqStart = NewCallSeqStart;
5880 
5881       // Pass the address of the aggregate copy on the stack either in a
5882       // physical register or in the parameter list area of the current stack
5883       // frame to the callee.
5884       Arg = PtrOff;
5885     }
5886 
5887     // When useCRBits() is true, there can be i1 arguments.
5888     // It is because getRegisterType(MVT::i1) => MVT::i1,
5889     // and for other integer types getRegisterType() => MVT::i32.
5890     // Extend i1 and ensure callee will get i32.
5891     if (Arg.getValueType() == MVT::i1)
5892       Arg = DAG.getNode(Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
5893                         dl, MVT::i32, Arg);
5894 
5895     if (VA.isRegLoc()) {
5896       seenFloatArg |= VA.getLocVT().isFloatingPoint();
5897       // Put argument in a physical register.
5898       if (Subtarget.hasSPE() && Arg.getValueType() == MVT::f64) {
5899         bool IsLE = Subtarget.isLittleEndian();
5900         SDValue SVal = DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg,
5901                         DAG.getIntPtrConstant(IsLE ? 0 : 1, dl));
5902         RegsToPass.push_back(std::make_pair(VA.getLocReg(), SVal.getValue(0)));
5903         SVal = DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg,
5904                            DAG.getIntPtrConstant(IsLE ? 1 : 0, dl));
5905         RegsToPass.push_back(std::make_pair(ArgLocs[++i].getLocReg(),
5906                              SVal.getValue(0)));
5907       } else
5908         RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
5909     } else {
5910       // Put argument in the parameter list area of the current stack frame.
5911       assert(VA.isMemLoc());
5912       unsigned LocMemOffset = VA.getLocMemOffset();
5913 
5914       if (!IsTailCall) {
5915         SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
5916         PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(MF.getDataLayout()),
5917                              StackPtr, PtrOff);
5918 
5919         MemOpChains.push_back(
5920             DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo()));
5921       } else {
5922         // Calculate and remember argument location.
5923         CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
5924                                  TailCallArguments);
5925       }
5926     }
5927   }
5928 
5929   if (!MemOpChains.empty())
5930     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
5931 
5932   // Build a sequence of copy-to-reg nodes chained together with token chain
5933   // and flag operands which copy the outgoing args into the appropriate regs.
5934   SDValue InFlag;
5935   for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
5936     Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
5937                              RegsToPass[i].second, InFlag);
5938     InFlag = Chain.getValue(1);
5939   }
5940 
5941   // Set CR bit 6 to true if this is a vararg call with floating args passed in
5942   // registers.
5943   if (IsVarArg) {
5944     SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
5945     SDValue Ops[] = { Chain, InFlag };
5946 
5947     Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET, dl,
5948                         VTs, ArrayRef(Ops, InFlag.getNode() ? 2 : 1));
5949 
5950     InFlag = Chain.getValue(1);
5951   }
5952 
5953   if (IsTailCall)
5954     PrepareTailCall(DAG, InFlag, Chain, dl, SPDiff, NumBytes, LROp, FPOp,
5955                     TailCallArguments);
5956 
5957   return FinishCall(CFlags, dl, DAG, RegsToPass, InFlag, Chain, CallSeqStart,
5958                     Callee, SPDiff, NumBytes, Ins, InVals, CB);
5959 }
5960 
5961 // Copy an argument into memory, being careful to do this outside the
5962 // call sequence for the call to which the argument belongs.
5963 SDValue PPCTargetLowering::createMemcpyOutsideCallSeq(
5964     SDValue Arg, SDValue PtrOff, SDValue CallSeqStart, ISD::ArgFlagsTy Flags,
5965     SelectionDAG &DAG, const SDLoc &dl) const {
5966   SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
5967                         CallSeqStart.getNode()->getOperand(0),
5968                         Flags, DAG, dl);
5969   // The MEMCPY must go outside the CALLSEQ_START..END.
5970   int64_t FrameSize = CallSeqStart.getConstantOperandVal(1);
5971   SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall, FrameSize, 0,
5972                                                  SDLoc(MemcpyCall));
5973   DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
5974                          NewCallSeqStart.getNode());
5975   return NewCallSeqStart;
5976 }
5977 
5978 SDValue PPCTargetLowering::LowerCall_64SVR4(
5979     SDValue Chain, SDValue Callee, CallFlags CFlags,
5980     const SmallVectorImpl<ISD::OutputArg> &Outs,
5981     const SmallVectorImpl<SDValue> &OutVals,
5982     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
5983     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals,
5984     const CallBase *CB) const {
5985   bool isELFv2ABI = Subtarget.isELFv2ABI();
5986   bool isLittleEndian = Subtarget.isLittleEndian();
5987   unsigned NumOps = Outs.size();
5988   bool IsSibCall = false;
5989   bool IsFastCall = CFlags.CallConv == CallingConv::Fast;
5990 
5991   EVT PtrVT = getPointerTy(DAG.getDataLayout());
5992   unsigned PtrByteSize = 8;
5993 
5994   MachineFunction &MF = DAG.getMachineFunction();
5995 
5996   if (CFlags.IsTailCall && !getTargetMachine().Options.GuaranteedTailCallOpt)
5997     IsSibCall = true;
5998 
5999   // Mark this function as potentially containing a function that contains a
6000   // tail call. As a consequence the frame pointer will be used for dynamicalloc
6001   // and restoring the callers stack pointer in this functions epilog. This is
6002   // done because by tail calling the called function might overwrite the value
6003   // in this function's (MF) stack pointer stack slot 0(SP).
6004   if (getTargetMachine().Options.GuaranteedTailCallOpt && IsFastCall)
6005     MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
6006 
6007   assert(!(IsFastCall && CFlags.IsVarArg) &&
6008          "fastcc not supported on varargs functions");
6009 
6010   // Count how many bytes are to be pushed on the stack, including the linkage
6011   // area, and parameter passing area.  On ELFv1, the linkage area is 48 bytes
6012   // reserved space for [SP][CR][LR][2 x unused][TOC]; on ELFv2, the linkage
6013   // area is 32 bytes reserved space for [SP][CR][LR][TOC].
6014   unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
6015   unsigned NumBytes = LinkageSize;
6016   unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
6017 
6018   static const MCPhysReg GPR[] = {
6019     PPC::X3, PPC::X4, PPC::X5, PPC::X6,
6020     PPC::X7, PPC::X8, PPC::X9, PPC::X10,
6021   };
6022   static const MCPhysReg VR[] = {
6023     PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
6024     PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
6025   };
6026 
6027   const unsigned NumGPRs = std::size(GPR);
6028   const unsigned NumFPRs = useSoftFloat() ? 0 : 13;
6029   const unsigned NumVRs = std::size(VR);
6030 
6031   // On ELFv2, we can avoid allocating the parameter area if all the arguments
6032   // can be passed to the callee in registers.
6033   // For the fast calling convention, there is another check below.
6034   // Note: We should keep consistent with LowerFormalArguments_64SVR4()
6035   bool HasParameterArea = !isELFv2ABI || CFlags.IsVarArg || IsFastCall;
6036   if (!HasParameterArea) {
6037     unsigned ParamAreaSize = NumGPRs * PtrByteSize;
6038     unsigned AvailableFPRs = NumFPRs;
6039     unsigned AvailableVRs = NumVRs;
6040     unsigned NumBytesTmp = NumBytes;
6041     for (unsigned i = 0; i != NumOps; ++i) {
6042       if (Outs[i].Flags.isNest()) continue;
6043       if (CalculateStackSlotUsed(Outs[i].VT, Outs[i].ArgVT, Outs[i].Flags,
6044                                  PtrByteSize, LinkageSize, ParamAreaSize,
6045                                  NumBytesTmp, AvailableFPRs, AvailableVRs))
6046         HasParameterArea = true;
6047     }
6048   }
6049 
6050   // When using the fast calling convention, we don't provide backing for
6051   // arguments that will be in registers.
6052   unsigned NumGPRsUsed = 0, NumFPRsUsed = 0, NumVRsUsed = 0;
6053 
6054   // Avoid allocating parameter area for fastcc functions if all the arguments
6055   // can be passed in the registers.
6056   if (IsFastCall)
6057     HasParameterArea = false;
6058 
6059   // Add up all the space actually used.
6060   for (unsigned i = 0; i != NumOps; ++i) {
6061     ISD::ArgFlagsTy Flags = Outs[i].Flags;
6062     EVT ArgVT = Outs[i].VT;
6063     EVT OrigVT = Outs[i].ArgVT;
6064 
6065     if (Flags.isNest())
6066       continue;
6067 
6068     if (IsFastCall) {
6069       if (Flags.isByVal()) {
6070         NumGPRsUsed += (Flags.getByValSize()+7)/8;
6071         if (NumGPRsUsed > NumGPRs)
6072           HasParameterArea = true;
6073       } else {
6074         switch (ArgVT.getSimpleVT().SimpleTy) {
6075         default: llvm_unreachable("Unexpected ValueType for argument!");
6076         case MVT::i1:
6077         case MVT::i32:
6078         case MVT::i64:
6079           if (++NumGPRsUsed <= NumGPRs)
6080             continue;
6081           break;
6082         case MVT::v4i32:
6083         case MVT::v8i16:
6084         case MVT::v16i8:
6085         case MVT::v2f64:
6086         case MVT::v2i64:
6087         case MVT::v1i128:
6088         case MVT::f128:
6089           if (++NumVRsUsed <= NumVRs)
6090             continue;
6091           break;
6092         case MVT::v4f32:
6093           if (++NumVRsUsed <= NumVRs)
6094             continue;
6095           break;
6096         case MVT::f32:
6097         case MVT::f64:
6098           if (++NumFPRsUsed <= NumFPRs)
6099             continue;
6100           break;
6101         }
6102         HasParameterArea = true;
6103       }
6104     }
6105 
6106     /* Respect alignment of argument on the stack.  */
6107     auto Alignement =
6108         CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
6109     NumBytes = alignTo(NumBytes, Alignement);
6110 
6111     NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
6112     if (Flags.isInConsecutiveRegsLast())
6113       NumBytes = ((NumBytes + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
6114   }
6115 
6116   unsigned NumBytesActuallyUsed = NumBytes;
6117 
6118   // In the old ELFv1 ABI,
6119   // the prolog code of the callee may store up to 8 GPR argument registers to
6120   // the stack, allowing va_start to index over them in memory if its varargs.
6121   // Because we cannot tell if this is needed on the caller side, we have to
6122   // conservatively assume that it is needed.  As such, make sure we have at
6123   // least enough stack space for the caller to store the 8 GPRs.
6124   // In the ELFv2 ABI, we allocate the parameter area iff a callee
6125   // really requires memory operands, e.g. a vararg function.
6126   if (HasParameterArea)
6127     NumBytes = std::max(NumBytes, LinkageSize + 8 * PtrByteSize);
6128   else
6129     NumBytes = LinkageSize;
6130 
6131   // Tail call needs the stack to be aligned.
6132   if (getTargetMachine().Options.GuaranteedTailCallOpt && IsFastCall)
6133     NumBytes = EnsureStackAlignment(Subtarget.getFrameLowering(), NumBytes);
6134 
6135   int SPDiff = 0;
6136 
6137   // Calculate by how many bytes the stack has to be adjusted in case of tail
6138   // call optimization.
6139   if (!IsSibCall)
6140     SPDiff = CalculateTailCallSPDiff(DAG, CFlags.IsTailCall, NumBytes);
6141 
6142   // To protect arguments on the stack from being clobbered in a tail call,
6143   // force all the loads to happen before doing any other lowering.
6144   if (CFlags.IsTailCall)
6145     Chain = DAG.getStackArgumentTokenFactor(Chain);
6146 
6147   // Adjust the stack pointer for the new arguments...
6148   // These operations are automatically eliminated by the prolog/epilog pass
6149   if (!IsSibCall)
6150     Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl);
6151   SDValue CallSeqStart = Chain;
6152 
6153   // Load the return address and frame pointer so it can be move somewhere else
6154   // later.
6155   SDValue LROp, FPOp;
6156   Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl);
6157 
6158   // Set up a copy of the stack pointer for use loading and storing any
6159   // arguments that may not fit in the registers available for argument
6160   // passing.
6161   SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
6162 
6163   // Figure out which arguments are going to go in registers, and which in
6164   // memory.  Also, if this is a vararg function, floating point operations
6165   // must be stored to our stack, and loaded into integer regs as well, if
6166   // any integer regs are available for argument passing.
6167   unsigned ArgOffset = LinkageSize;
6168 
6169   SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
6170   SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
6171 
6172   SmallVector<SDValue, 8> MemOpChains;
6173   for (unsigned i = 0; i != NumOps; ++i) {
6174     SDValue Arg = OutVals[i];
6175     ISD::ArgFlagsTy Flags = Outs[i].Flags;
6176     EVT ArgVT = Outs[i].VT;
6177     EVT OrigVT = Outs[i].ArgVT;
6178 
6179     // PtrOff will be used to store the current argument to the stack if a
6180     // register cannot be found for it.
6181     SDValue PtrOff;
6182 
6183     // We re-align the argument offset for each argument, except when using the
6184     // fast calling convention, when we need to make sure we do that only when
6185     // we'll actually use a stack slot.
6186     auto ComputePtrOff = [&]() {
6187       /* Respect alignment of argument on the stack.  */
6188       auto Alignment =
6189           CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
6190       ArgOffset = alignTo(ArgOffset, Alignment);
6191 
6192       PtrOff = DAG.getConstant(ArgOffset, dl, StackPtr.getValueType());
6193 
6194       PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
6195     };
6196 
6197     if (!IsFastCall) {
6198       ComputePtrOff();
6199 
6200       /* Compute GPR index associated with argument offset.  */
6201       GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize;
6202       GPR_idx = std::min(GPR_idx, NumGPRs);
6203     }
6204 
6205     // Promote integers to 64-bit values.
6206     if (Arg.getValueType() == MVT::i32 || Arg.getValueType() == MVT::i1) {
6207       // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
6208       unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
6209       Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
6210     }
6211 
6212     // FIXME memcpy is used way more than necessary.  Correctness first.
6213     // Note: "by value" is code for passing a structure by value, not
6214     // basic types.
6215     if (Flags.isByVal()) {
6216       // Note: Size includes alignment padding, so
6217       //   struct x { short a; char b; }
6218       // will have Size = 4.  With #pragma pack(1), it will have Size = 3.
6219       // These are the proper values we need for right-justifying the
6220       // aggregate in a parameter register.
6221       unsigned Size = Flags.getByValSize();
6222 
6223       // An empty aggregate parameter takes up no storage and no
6224       // registers.
6225       if (Size == 0)
6226         continue;
6227 
6228       if (IsFastCall)
6229         ComputePtrOff();
6230 
6231       // All aggregates smaller than 8 bytes must be passed right-justified.
6232       if (Size==1 || Size==2 || Size==4) {
6233         EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
6234         if (GPR_idx != NumGPRs) {
6235           SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
6236                                         MachinePointerInfo(), VT);
6237           MemOpChains.push_back(Load.getValue(1));
6238           RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
6239 
6240           ArgOffset += PtrByteSize;
6241           continue;
6242         }
6243       }
6244 
6245       if (GPR_idx == NumGPRs && Size < 8) {
6246         SDValue AddPtr = PtrOff;
6247         if (!isLittleEndian) {
6248           SDValue Const = DAG.getConstant(PtrByteSize - Size, dl,
6249                                           PtrOff.getValueType());
6250           AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
6251         }
6252         Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
6253                                                           CallSeqStart,
6254                                                           Flags, DAG, dl);
6255         ArgOffset += PtrByteSize;
6256         continue;
6257       }
6258       // Copy the object to parameter save area if it can not be entirely passed
6259       // by registers.
6260       // FIXME: we only need to copy the parts which need to be passed in
6261       // parameter save area. For the parts passed by registers, we don't need
6262       // to copy them to the stack although we need to allocate space for them
6263       // in parameter save area.
6264       if ((NumGPRs - GPR_idx) * PtrByteSize < Size)
6265         Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
6266                                                           CallSeqStart,
6267                                                           Flags, DAG, dl);
6268 
6269       // When a register is available, pass a small aggregate right-justified.
6270       if (Size < 8 && GPR_idx != NumGPRs) {
6271         // The easiest way to get this right-justified in a register
6272         // is to copy the structure into the rightmost portion of a
6273         // local variable slot, then load the whole slot into the
6274         // register.
6275         // FIXME: The memcpy seems to produce pretty awful code for
6276         // small aggregates, particularly for packed ones.
6277         // FIXME: It would be preferable to use the slot in the
6278         // parameter save area instead of a new local variable.
6279         SDValue AddPtr = PtrOff;
6280         if (!isLittleEndian) {
6281           SDValue Const = DAG.getConstant(8 - Size, dl, PtrOff.getValueType());
6282           AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
6283         }
6284         Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
6285                                                           CallSeqStart,
6286                                                           Flags, DAG, dl);
6287 
6288         // Load the slot into the register.
6289         SDValue Load =
6290             DAG.getLoad(PtrVT, dl, Chain, PtrOff, MachinePointerInfo());
6291         MemOpChains.push_back(Load.getValue(1));
6292         RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
6293 
6294         // Done with this argument.
6295         ArgOffset += PtrByteSize;
6296         continue;
6297       }
6298 
6299       // For aggregates larger than PtrByteSize, copy the pieces of the
6300       // object that fit into registers from the parameter save area.
6301       for (unsigned j=0; j<Size; j+=PtrByteSize) {
6302         SDValue Const = DAG.getConstant(j, dl, PtrOff.getValueType());
6303         SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
6304         if (GPR_idx != NumGPRs) {
6305           unsigned LoadSizeInBits = std::min(PtrByteSize, (Size - j)) * 8;
6306           EVT ObjType = EVT::getIntegerVT(*DAG.getContext(), LoadSizeInBits);
6307           SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, AddArg,
6308                                         MachinePointerInfo(), ObjType);
6309 
6310           MemOpChains.push_back(Load.getValue(1));
6311           RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
6312           ArgOffset += PtrByteSize;
6313         } else {
6314           ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
6315           break;
6316         }
6317       }
6318       continue;
6319     }
6320 
6321     switch (Arg.getSimpleValueType().SimpleTy) {
6322     default: llvm_unreachable("Unexpected ValueType for argument!");
6323     case MVT::i1:
6324     case MVT::i32:
6325     case MVT::i64:
6326       if (Flags.isNest()) {
6327         // The 'nest' parameter, if any, is passed in R11.
6328         RegsToPass.push_back(std::make_pair(PPC::X11, Arg));
6329         break;
6330       }
6331 
6332       // These can be scalar arguments or elements of an integer array type
6333       // passed directly.  Clang may use those instead of "byval" aggregate
6334       // types to avoid forcing arguments to memory unnecessarily.
6335       if (GPR_idx != NumGPRs) {
6336         RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
6337       } else {
6338         if (IsFastCall)
6339           ComputePtrOff();
6340 
6341         assert(HasParameterArea &&
6342                "Parameter area must exist to pass an argument in memory.");
6343         LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
6344                          true, CFlags.IsTailCall, false, MemOpChains,
6345                          TailCallArguments, dl);
6346         if (IsFastCall)
6347           ArgOffset += PtrByteSize;
6348       }
6349       if (!IsFastCall)
6350         ArgOffset += PtrByteSize;
6351       break;
6352     case MVT::f32:
6353     case MVT::f64: {
6354       // These can be scalar arguments or elements of a float array type
6355       // passed directly.  The latter are used to implement ELFv2 homogenous
6356       // float aggregates.
6357 
6358       // Named arguments go into FPRs first, and once they overflow, the
6359       // remaining arguments go into GPRs and then the parameter save area.
6360       // Unnamed arguments for vararg functions always go to GPRs and
6361       // then the parameter save area.  For now, put all arguments to vararg
6362       // routines always in both locations (FPR *and* GPR or stack slot).
6363       bool NeedGPROrStack = CFlags.IsVarArg || FPR_idx == NumFPRs;
6364       bool NeededLoad = false;
6365 
6366       // First load the argument into the next available FPR.
6367       if (FPR_idx != NumFPRs)
6368         RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
6369 
6370       // Next, load the argument into GPR or stack slot if needed.
6371       if (!NeedGPROrStack)
6372         ;
6373       else if (GPR_idx != NumGPRs && !IsFastCall) {
6374         // FIXME: We may want to re-enable this for CallingConv::Fast on the P8
6375         // once we support fp <-> gpr moves.
6376 
6377         // In the non-vararg case, this can only ever happen in the
6378         // presence of f32 array types, since otherwise we never run
6379         // out of FPRs before running out of GPRs.
6380         SDValue ArgVal;
6381 
6382         // Double values are always passed in a single GPR.
6383         if (Arg.getValueType() != MVT::f32) {
6384           ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
6385 
6386         // Non-array float values are extended and passed in a GPR.
6387         } else if (!Flags.isInConsecutiveRegs()) {
6388           ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
6389           ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal);
6390 
6391         // If we have an array of floats, we collect every odd element
6392         // together with its predecessor into one GPR.
6393         } else if (ArgOffset % PtrByteSize != 0) {
6394           SDValue Lo, Hi;
6395           Lo = DAG.getNode(ISD::BITCAST, dl, MVT::i32, OutVals[i - 1]);
6396           Hi = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
6397           if (!isLittleEndian)
6398             std::swap(Lo, Hi);
6399           ArgVal = DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
6400 
6401         // The final element, if even, goes into the first half of a GPR.
6402         } else if (Flags.isInConsecutiveRegsLast()) {
6403           ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
6404           ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal);
6405           if (!isLittleEndian)
6406             ArgVal = DAG.getNode(ISD::SHL, dl, MVT::i64, ArgVal,
6407                                  DAG.getConstant(32, dl, MVT::i32));
6408 
6409         // Non-final even elements are skipped; they will be handled
6410         // together the with subsequent argument on the next go-around.
6411         } else
6412           ArgVal = SDValue();
6413 
6414         if (ArgVal.getNode())
6415           RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], ArgVal));
6416       } else {
6417         if (IsFastCall)
6418           ComputePtrOff();
6419 
6420         // Single-precision floating-point values are mapped to the
6421         // second (rightmost) word of the stack doubleword.
6422         if (Arg.getValueType() == MVT::f32 &&
6423             !isLittleEndian && !Flags.isInConsecutiveRegs()) {
6424           SDValue ConstFour = DAG.getConstant(4, dl, PtrOff.getValueType());
6425           PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
6426         }
6427 
6428         assert(HasParameterArea &&
6429                "Parameter area must exist to pass an argument in memory.");
6430         LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
6431                          true, CFlags.IsTailCall, false, MemOpChains,
6432                          TailCallArguments, dl);
6433 
6434         NeededLoad = true;
6435       }
6436       // When passing an array of floats, the array occupies consecutive
6437       // space in the argument area; only round up to the next doubleword
6438       // at the end of the array.  Otherwise, each float takes 8 bytes.
6439       if (!IsFastCall || NeededLoad) {
6440         ArgOffset += (Arg.getValueType() == MVT::f32 &&
6441                       Flags.isInConsecutiveRegs()) ? 4 : 8;
6442         if (Flags.isInConsecutiveRegsLast())
6443           ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
6444       }
6445       break;
6446     }
6447     case MVT::v4f32:
6448     case MVT::v4i32:
6449     case MVT::v8i16:
6450     case MVT::v16i8:
6451     case MVT::v2f64:
6452     case MVT::v2i64:
6453     case MVT::v1i128:
6454     case MVT::f128:
6455       // These can be scalar arguments or elements of a vector array type
6456       // passed directly.  The latter are used to implement ELFv2 homogenous
6457       // vector aggregates.
6458 
6459       // For a varargs call, named arguments go into VRs or on the stack as
6460       // usual; unnamed arguments always go to the stack or the corresponding
6461       // GPRs when within range.  For now, we always put the value in both
6462       // locations (or even all three).
6463       if (CFlags.IsVarArg) {
6464         assert(HasParameterArea &&
6465                "Parameter area must exist if we have a varargs call.");
6466         // We could elide this store in the case where the object fits
6467         // entirely in R registers.  Maybe later.
6468         SDValue Store =
6469             DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo());
6470         MemOpChains.push_back(Store);
6471         if (VR_idx != NumVRs) {
6472           SDValue Load =
6473               DAG.getLoad(MVT::v4f32, dl, Store, PtrOff, MachinePointerInfo());
6474           MemOpChains.push_back(Load.getValue(1));
6475           RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
6476         }
6477         ArgOffset += 16;
6478         for (unsigned i=0; i<16; i+=PtrByteSize) {
6479           if (GPR_idx == NumGPRs)
6480             break;
6481           SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
6482                                    DAG.getConstant(i, dl, PtrVT));
6483           SDValue Load =
6484               DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo());
6485           MemOpChains.push_back(Load.getValue(1));
6486           RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
6487         }
6488         break;
6489       }
6490 
6491       // Non-varargs Altivec params go into VRs or on the stack.
6492       if (VR_idx != NumVRs) {
6493         RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
6494       } else {
6495         if (IsFastCall)
6496           ComputePtrOff();
6497 
6498         assert(HasParameterArea &&
6499                "Parameter area must exist to pass an argument in memory.");
6500         LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
6501                          true, CFlags.IsTailCall, true, MemOpChains,
6502                          TailCallArguments, dl);
6503         if (IsFastCall)
6504           ArgOffset += 16;
6505       }
6506 
6507       if (!IsFastCall)
6508         ArgOffset += 16;
6509       break;
6510     }
6511   }
6512 
6513   assert((!HasParameterArea || NumBytesActuallyUsed == ArgOffset) &&
6514          "mismatch in size of parameter area");
6515   (void)NumBytesActuallyUsed;
6516 
6517   if (!MemOpChains.empty())
6518     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
6519 
6520   // Check if this is an indirect call (MTCTR/BCTRL).
6521   // See prepareDescriptorIndirectCall and buildCallOperands for more
6522   // information about calls through function pointers in the 64-bit SVR4 ABI.
6523   if (CFlags.IsIndirect) {
6524     // For 64-bit ELFv2 ABI with PCRel, do not save the TOC of the
6525     // caller in the TOC save area.
6526     if (isTOCSaveRestoreRequired(Subtarget)) {
6527       assert(!CFlags.IsTailCall && "Indirect tails calls not supported");
6528       // Load r2 into a virtual register and store it to the TOC save area.
6529       setUsesTOCBasePtr(DAG);
6530       SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
6531       // TOC save area offset.
6532       unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset();
6533       SDValue PtrOff = DAG.getIntPtrConstant(TOCSaveOffset, dl);
6534       SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
6535       Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr,
6536                            MachinePointerInfo::getStack(
6537                                DAG.getMachineFunction(), TOCSaveOffset));
6538     }
6539     // In the ELFv2 ABI, R12 must contain the address of an indirect callee.
6540     // This does not mean the MTCTR instruction must use R12; it's easier
6541     // to model this as an extra parameter, so do that.
6542     if (isELFv2ABI && !CFlags.IsPatchPoint)
6543       RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee));
6544   }
6545 
6546   // Build a sequence of copy-to-reg nodes chained together with token chain
6547   // and flag operands which copy the outgoing args into the appropriate regs.
6548   SDValue InFlag;
6549   for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
6550     Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
6551                              RegsToPass[i].second, InFlag);
6552     InFlag = Chain.getValue(1);
6553   }
6554 
6555   if (CFlags.IsTailCall && !IsSibCall)
6556     PrepareTailCall(DAG, InFlag, Chain, dl, SPDiff, NumBytes, LROp, FPOp,
6557                     TailCallArguments);
6558 
6559   return FinishCall(CFlags, dl, DAG, RegsToPass, InFlag, Chain, CallSeqStart,
6560                     Callee, SPDiff, NumBytes, Ins, InVals, CB);
6561 }
6562 
6563 // Returns true when the shadow of a general purpose argument register
6564 // in the parameter save area is aligned to at least 'RequiredAlign'.
6565 static bool isGPRShadowAligned(MCPhysReg Reg, Align RequiredAlign) {
6566   assert(RequiredAlign.value() <= 16 &&
6567          "Required alignment greater than stack alignment.");
6568   switch (Reg) {
6569   default:
6570     report_fatal_error("called on invalid register.");
6571   case PPC::R5:
6572   case PPC::R9:
6573   case PPC::X3:
6574   case PPC::X5:
6575   case PPC::X7:
6576   case PPC::X9:
6577     // These registers are 16 byte aligned which is the most strict aligment
6578     // we can support.
6579     return true;
6580   case PPC::R3:
6581   case PPC::R7:
6582   case PPC::X4:
6583   case PPC::X6:
6584   case PPC::X8:
6585   case PPC::X10:
6586     // The shadow of these registers in the PSA is 8 byte aligned.
6587     return RequiredAlign <= 8;
6588   case PPC::R4:
6589   case PPC::R6:
6590   case PPC::R8:
6591   case PPC::R10:
6592     return RequiredAlign <= 4;
6593   }
6594 }
6595 
6596 static bool CC_AIX(unsigned ValNo, MVT ValVT, MVT LocVT,
6597                    CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
6598                    CCState &S) {
6599   AIXCCState &State = static_cast<AIXCCState &>(S);
6600   const PPCSubtarget &Subtarget = static_cast<const PPCSubtarget &>(
6601       State.getMachineFunction().getSubtarget());
6602   const bool IsPPC64 = Subtarget.isPPC64();
6603   const Align PtrAlign = IsPPC64 ? Align(8) : Align(4);
6604   const MVT RegVT = IsPPC64 ? MVT::i64 : MVT::i32;
6605 
6606   if (ValVT == MVT::f128)
6607     report_fatal_error("f128 is unimplemented on AIX.");
6608 
6609   if (ArgFlags.isNest())
6610     report_fatal_error("Nest arguments are unimplemented.");
6611 
6612   static const MCPhysReg GPR_32[] = {// 32-bit registers.
6613                                      PPC::R3, PPC::R4, PPC::R5, PPC::R6,
6614                                      PPC::R7, PPC::R8, PPC::R9, PPC::R10};
6615   static const MCPhysReg GPR_64[] = {// 64-bit registers.
6616                                      PPC::X3, PPC::X4, PPC::X5, PPC::X6,
6617                                      PPC::X7, PPC::X8, PPC::X9, PPC::X10};
6618 
6619   static const MCPhysReg VR[] = {// Vector registers.
6620                                  PPC::V2,  PPC::V3,  PPC::V4,  PPC::V5,
6621                                  PPC::V6,  PPC::V7,  PPC::V8,  PPC::V9,
6622                                  PPC::V10, PPC::V11, PPC::V12, PPC::V13};
6623 
6624   if (ArgFlags.isByVal()) {
6625     if (ArgFlags.getNonZeroByValAlign() > PtrAlign)
6626       report_fatal_error("Pass-by-value arguments with alignment greater than "
6627                          "register width are not supported.");
6628 
6629     const unsigned ByValSize = ArgFlags.getByValSize();
6630 
6631     // An empty aggregate parameter takes up no storage and no registers,
6632     // but needs a MemLoc for a stack slot for the formal arguments side.
6633     if (ByValSize == 0) {
6634       State.addLoc(CCValAssign::getMem(ValNo, MVT::INVALID_SIMPLE_VALUE_TYPE,
6635                                        State.getNextStackOffset(), RegVT,
6636                                        LocInfo));
6637       return false;
6638     }
6639 
6640     const unsigned StackSize = alignTo(ByValSize, PtrAlign);
6641     unsigned Offset = State.AllocateStack(StackSize, PtrAlign);
6642     for (const unsigned E = Offset + StackSize; Offset < E;
6643          Offset += PtrAlign.value()) {
6644       if (unsigned Reg = State.AllocateReg(IsPPC64 ? GPR_64 : GPR_32))
6645         State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, RegVT, LocInfo));
6646       else {
6647         State.addLoc(CCValAssign::getMem(ValNo, MVT::INVALID_SIMPLE_VALUE_TYPE,
6648                                          Offset, MVT::INVALID_SIMPLE_VALUE_TYPE,
6649                                          LocInfo));
6650         break;
6651       }
6652     }
6653     return false;
6654   }
6655 
6656   // Arguments always reserve parameter save area.
6657   switch (ValVT.SimpleTy) {
6658   default:
6659     report_fatal_error("Unhandled value type for argument.");
6660   case MVT::i64:
6661     // i64 arguments should have been split to i32 for PPC32.
6662     assert(IsPPC64 && "PPC32 should have split i64 values.");
6663     [[fallthrough]];
6664   case MVT::i1:
6665   case MVT::i32: {
6666     const unsigned Offset = State.AllocateStack(PtrAlign.value(), PtrAlign);
6667     // AIX integer arguments are always passed in register width.
6668     if (ValVT.getFixedSizeInBits() < RegVT.getFixedSizeInBits())
6669       LocInfo = ArgFlags.isSExt() ? CCValAssign::LocInfo::SExt
6670                                   : CCValAssign::LocInfo::ZExt;
6671     if (unsigned Reg = State.AllocateReg(IsPPC64 ? GPR_64 : GPR_32))
6672       State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, RegVT, LocInfo));
6673     else
6674       State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, RegVT, LocInfo));
6675 
6676     return false;
6677   }
6678   case MVT::f32:
6679   case MVT::f64: {
6680     // Parameter save area (PSA) is reserved even if the float passes in fpr.
6681     const unsigned StoreSize = LocVT.getStoreSize();
6682     // Floats are always 4-byte aligned in the PSA on AIX.
6683     // This includes f64 in 64-bit mode for ABI compatibility.
6684     const unsigned Offset =
6685         State.AllocateStack(IsPPC64 ? 8 : StoreSize, Align(4));
6686     unsigned FReg = State.AllocateReg(FPR);
6687     if (FReg)
6688       State.addLoc(CCValAssign::getReg(ValNo, ValVT, FReg, LocVT, LocInfo));
6689 
6690     // Reserve and initialize GPRs or initialize the PSA as required.
6691     for (unsigned I = 0; I < StoreSize; I += PtrAlign.value()) {
6692       if (unsigned Reg = State.AllocateReg(IsPPC64 ? GPR_64 : GPR_32)) {
6693         assert(FReg && "An FPR should be available when a GPR is reserved.");
6694         if (State.isVarArg()) {
6695           // Successfully reserved GPRs are only initialized for vararg calls.
6696           // Custom handling is required for:
6697           //   f64 in PPC32 needs to be split into 2 GPRs.
6698           //   f32 in PPC64 needs to occupy only lower 32 bits of 64-bit GPR.
6699           State.addLoc(
6700               CCValAssign::getCustomReg(ValNo, ValVT, Reg, RegVT, LocInfo));
6701         }
6702       } else {
6703         // If there are insufficient GPRs, the PSA needs to be initialized.
6704         // Initialization occurs even if an FPR was initialized for
6705         // compatibility with the AIX XL compiler. The full memory for the
6706         // argument will be initialized even if a prior word is saved in GPR.
6707         // A custom memLoc is used when the argument also passes in FPR so
6708         // that the callee handling can skip over it easily.
6709         State.addLoc(
6710             FReg ? CCValAssign::getCustomMem(ValNo, ValVT, Offset, LocVT,
6711                                              LocInfo)
6712                  : CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
6713         break;
6714       }
6715     }
6716 
6717     return false;
6718   }
6719   case MVT::v4f32:
6720   case MVT::v4i32:
6721   case MVT::v8i16:
6722   case MVT::v16i8:
6723   case MVT::v2i64:
6724   case MVT::v2f64:
6725   case MVT::v1i128: {
6726     const unsigned VecSize = 16;
6727     const Align VecAlign(VecSize);
6728 
6729     if (!State.isVarArg()) {
6730       // If there are vector registers remaining we don't consume any stack
6731       // space.
6732       if (unsigned VReg = State.AllocateReg(VR)) {
6733         State.addLoc(CCValAssign::getReg(ValNo, ValVT, VReg, LocVT, LocInfo));
6734         return false;
6735       }
6736       // Vectors passed on the stack do not shadow GPRs or FPRs even though they
6737       // might be allocated in the portion of the PSA that is shadowed by the
6738       // GPRs.
6739       const unsigned Offset = State.AllocateStack(VecSize, VecAlign);
6740       State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
6741       return false;
6742     }
6743 
6744     const unsigned PtrSize = IsPPC64 ? 8 : 4;
6745     ArrayRef<MCPhysReg> GPRs = IsPPC64 ? GPR_64 : GPR_32;
6746 
6747     unsigned NextRegIndex = State.getFirstUnallocated(GPRs);
6748     // Burn any underaligned registers and their shadowed stack space until
6749     // we reach the required alignment.
6750     while (NextRegIndex != GPRs.size() &&
6751            !isGPRShadowAligned(GPRs[NextRegIndex], VecAlign)) {
6752       // Shadow allocate register and its stack shadow.
6753       unsigned Reg = State.AllocateReg(GPRs);
6754       State.AllocateStack(PtrSize, PtrAlign);
6755       assert(Reg && "Allocating register unexpectedly failed.");
6756       (void)Reg;
6757       NextRegIndex = State.getFirstUnallocated(GPRs);
6758     }
6759 
6760     // Vectors that are passed as fixed arguments are handled differently.
6761     // They are passed in VRs if any are available (unlike arguments passed
6762     // through ellipses) and shadow GPRs (unlike arguments to non-vaarg
6763     // functions)
6764     if (State.isFixed(ValNo)) {
6765       if (unsigned VReg = State.AllocateReg(VR)) {
6766         State.addLoc(CCValAssign::getReg(ValNo, ValVT, VReg, LocVT, LocInfo));
6767         // Shadow allocate GPRs and stack space even though we pass in a VR.
6768         for (unsigned I = 0; I != VecSize; I += PtrSize)
6769           State.AllocateReg(GPRs);
6770         State.AllocateStack(VecSize, VecAlign);
6771         return false;
6772       }
6773       // No vector registers remain so pass on the stack.
6774       const unsigned Offset = State.AllocateStack(VecSize, VecAlign);
6775       State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
6776       return false;
6777     }
6778 
6779     // If all GPRS are consumed then we pass the argument fully on the stack.
6780     if (NextRegIndex == GPRs.size()) {
6781       const unsigned Offset = State.AllocateStack(VecSize, VecAlign);
6782       State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
6783       return false;
6784     }
6785 
6786     // Corner case for 32-bit codegen. We have 2 registers to pass the first
6787     // half of the argument, and then need to pass the remaining half on the
6788     // stack.
6789     if (GPRs[NextRegIndex] == PPC::R9) {
6790       const unsigned Offset = State.AllocateStack(VecSize, VecAlign);
6791       State.addLoc(
6792           CCValAssign::getCustomMem(ValNo, ValVT, Offset, LocVT, LocInfo));
6793 
6794       const unsigned FirstReg = State.AllocateReg(PPC::R9);
6795       const unsigned SecondReg = State.AllocateReg(PPC::R10);
6796       assert(FirstReg && SecondReg &&
6797              "Allocating R9 or R10 unexpectedly failed.");
6798       State.addLoc(
6799           CCValAssign::getCustomReg(ValNo, ValVT, FirstReg, RegVT, LocInfo));
6800       State.addLoc(
6801           CCValAssign::getCustomReg(ValNo, ValVT, SecondReg, RegVT, LocInfo));
6802       return false;
6803     }
6804 
6805     // We have enough GPRs to fully pass the vector argument, and we have
6806     // already consumed any underaligned registers. Start with the custom
6807     // MemLoc and then the custom RegLocs.
6808     const unsigned Offset = State.AllocateStack(VecSize, VecAlign);
6809     State.addLoc(
6810         CCValAssign::getCustomMem(ValNo, ValVT, Offset, LocVT, LocInfo));
6811     for (unsigned I = 0; I != VecSize; I += PtrSize) {
6812       const unsigned Reg = State.AllocateReg(GPRs);
6813       assert(Reg && "Failed to allocated register for vararg vector argument");
6814       State.addLoc(
6815           CCValAssign::getCustomReg(ValNo, ValVT, Reg, RegVT, LocInfo));
6816     }
6817     return false;
6818   }
6819   }
6820   return true;
6821 }
6822 
6823 // So far, this function is only used by LowerFormalArguments_AIX()
6824 static const TargetRegisterClass *getRegClassForSVT(MVT::SimpleValueType SVT,
6825                                                     bool IsPPC64,
6826                                                     bool HasP8Vector,
6827                                                     bool HasVSX) {
6828   assert((IsPPC64 || SVT != MVT::i64) &&
6829          "i64 should have been split for 32-bit codegen.");
6830 
6831   switch (SVT) {
6832   default:
6833     report_fatal_error("Unexpected value type for formal argument");
6834   case MVT::i1:
6835   case MVT::i32:
6836   case MVT::i64:
6837     return IsPPC64 ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
6838   case MVT::f32:
6839     return HasP8Vector ? &PPC::VSSRCRegClass : &PPC::F4RCRegClass;
6840   case MVT::f64:
6841     return HasVSX ? &PPC::VSFRCRegClass : &PPC::F8RCRegClass;
6842   case MVT::v4f32:
6843   case MVT::v4i32:
6844   case MVT::v8i16:
6845   case MVT::v16i8:
6846   case MVT::v2i64:
6847   case MVT::v2f64:
6848   case MVT::v1i128:
6849     return &PPC::VRRCRegClass;
6850   }
6851 }
6852 
6853 static SDValue truncateScalarIntegerArg(ISD::ArgFlagsTy Flags, EVT ValVT,
6854                                         SelectionDAG &DAG, SDValue ArgValue,
6855                                         MVT LocVT, const SDLoc &dl) {
6856   assert(ValVT.isScalarInteger() && LocVT.isScalarInteger());
6857   assert(ValVT.getFixedSizeInBits() < LocVT.getFixedSizeInBits());
6858 
6859   if (Flags.isSExt())
6860     ArgValue = DAG.getNode(ISD::AssertSext, dl, LocVT, ArgValue,
6861                            DAG.getValueType(ValVT));
6862   else if (Flags.isZExt())
6863     ArgValue = DAG.getNode(ISD::AssertZext, dl, LocVT, ArgValue,
6864                            DAG.getValueType(ValVT));
6865 
6866   return DAG.getNode(ISD::TRUNCATE, dl, ValVT, ArgValue);
6867 }
6868 
6869 static unsigned mapArgRegToOffsetAIX(unsigned Reg, const PPCFrameLowering *FL) {
6870   const unsigned LASize = FL->getLinkageSize();
6871 
6872   if (PPC::GPRCRegClass.contains(Reg)) {
6873     assert(Reg >= PPC::R3 && Reg <= PPC::R10 &&
6874            "Reg must be a valid argument register!");
6875     return LASize + 4 * (Reg - PPC::R3);
6876   }
6877 
6878   if (PPC::G8RCRegClass.contains(Reg)) {
6879     assert(Reg >= PPC::X3 && Reg <= PPC::X10 &&
6880            "Reg must be a valid argument register!");
6881     return LASize + 8 * (Reg - PPC::X3);
6882   }
6883 
6884   llvm_unreachable("Only general purpose registers expected.");
6885 }
6886 
6887 //   AIX ABI Stack Frame Layout:
6888 //
6889 //   Low Memory +--------------------------------------------+
6890 //   SP   +---> | Back chain                                 | ---+
6891 //        |     +--------------------------------------------+    |
6892 //        |     | Saved Condition Register                   |    |
6893 //        |     +--------------------------------------------+    |
6894 //        |     | Saved Linkage Register                     |    |
6895 //        |     +--------------------------------------------+    | Linkage Area
6896 //        |     | Reserved for compilers                     |    |
6897 //        |     +--------------------------------------------+    |
6898 //        |     | Reserved for binders                       |    |
6899 //        |     +--------------------------------------------+    |
6900 //        |     | Saved TOC pointer                          | ---+
6901 //        |     +--------------------------------------------+
6902 //        |     | Parameter save area                        |
6903 //        |     +--------------------------------------------+
6904 //        |     | Alloca space                               |
6905 //        |     +--------------------------------------------+
6906 //        |     | Local variable space                       |
6907 //        |     +--------------------------------------------+
6908 //        |     | Float/int conversion temporary             |
6909 //        |     +--------------------------------------------+
6910 //        |     | Save area for AltiVec registers            |
6911 //        |     +--------------------------------------------+
6912 //        |     | AltiVec alignment padding                  |
6913 //        |     +--------------------------------------------+
6914 //        |     | Save area for VRSAVE register              |
6915 //        |     +--------------------------------------------+
6916 //        |     | Save area for General Purpose registers    |
6917 //        |     +--------------------------------------------+
6918 //        |     | Save area for Floating Point registers     |
6919 //        |     +--------------------------------------------+
6920 //        +---- | Back chain                                 |
6921 // High Memory  +--------------------------------------------+
6922 //
6923 //  Specifications:
6924 //  AIX 7.2 Assembler Language Reference
6925 //  Subroutine linkage convention
6926 
6927 SDValue PPCTargetLowering::LowerFormalArguments_AIX(
6928     SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
6929     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
6930     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
6931 
6932   assert((CallConv == CallingConv::C || CallConv == CallingConv::Cold ||
6933           CallConv == CallingConv::Fast) &&
6934          "Unexpected calling convention!");
6935 
6936   if (getTargetMachine().Options.GuaranteedTailCallOpt)
6937     report_fatal_error("Tail call support is unimplemented on AIX.");
6938 
6939   if (useSoftFloat())
6940     report_fatal_error("Soft float support is unimplemented on AIX.");
6941 
6942   const PPCSubtarget &Subtarget = DAG.getSubtarget<PPCSubtarget>();
6943 
6944   const bool IsPPC64 = Subtarget.isPPC64();
6945   const unsigned PtrByteSize = IsPPC64 ? 8 : 4;
6946 
6947   // Assign locations to all of the incoming arguments.
6948   SmallVector<CCValAssign, 16> ArgLocs;
6949   MachineFunction &MF = DAG.getMachineFunction();
6950   MachineFrameInfo &MFI = MF.getFrameInfo();
6951   PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
6952   AIXCCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
6953 
6954   const EVT PtrVT = getPointerTy(MF.getDataLayout());
6955   // Reserve space for the linkage area on the stack.
6956   const unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
6957   CCInfo.AllocateStack(LinkageSize, Align(PtrByteSize));
6958   CCInfo.AnalyzeFormalArguments(Ins, CC_AIX);
6959 
6960   SmallVector<SDValue, 8> MemOps;
6961 
6962   for (size_t I = 0, End = ArgLocs.size(); I != End; /* No increment here */) {
6963     CCValAssign &VA = ArgLocs[I++];
6964     MVT LocVT = VA.getLocVT();
6965     MVT ValVT = VA.getValVT();
6966     ISD::ArgFlagsTy Flags = Ins[VA.getValNo()].Flags;
6967     // For compatibility with the AIX XL compiler, the float args in the
6968     // parameter save area are initialized even if the argument is available
6969     // in register.  The caller is required to initialize both the register
6970     // and memory, however, the callee can choose to expect it in either.
6971     // The memloc is dismissed here because the argument is retrieved from
6972     // the register.
6973     if (VA.isMemLoc() && VA.needsCustom() && ValVT.isFloatingPoint())
6974       continue;
6975 
6976     auto HandleMemLoc = [&]() {
6977       const unsigned LocSize = LocVT.getStoreSize();
6978       const unsigned ValSize = ValVT.getStoreSize();
6979       assert((ValSize <= LocSize) &&
6980              "Object size is larger than size of MemLoc");
6981       int CurArgOffset = VA.getLocMemOffset();
6982       // Objects are right-justified because AIX is big-endian.
6983       if (LocSize > ValSize)
6984         CurArgOffset += LocSize - ValSize;
6985       // Potential tail calls could cause overwriting of argument stack slots.
6986       const bool IsImmutable =
6987           !(getTargetMachine().Options.GuaranteedTailCallOpt &&
6988             (CallConv == CallingConv::Fast));
6989       int FI = MFI.CreateFixedObject(ValSize, CurArgOffset, IsImmutable);
6990       SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
6991       SDValue ArgValue =
6992           DAG.getLoad(ValVT, dl, Chain, FIN, MachinePointerInfo());
6993       InVals.push_back(ArgValue);
6994     };
6995 
6996     // Vector arguments to VaArg functions are passed both on the stack, and
6997     // in any available GPRs. Load the value from the stack and add the GPRs
6998     // as live ins.
6999     if (VA.isMemLoc() && VA.needsCustom()) {
7000       assert(ValVT.isVector() && "Unexpected Custom MemLoc type.");
7001       assert(isVarArg && "Only use custom memloc for vararg.");
7002       // ValNo of the custom MemLoc, so we can compare it to the ValNo of the
7003       // matching custom RegLocs.
7004       const unsigned OriginalValNo = VA.getValNo();
7005       (void)OriginalValNo;
7006 
7007       auto HandleCustomVecRegLoc = [&]() {
7008         assert(I != End && ArgLocs[I].isRegLoc() && ArgLocs[I].needsCustom() &&
7009                "Missing custom RegLoc.");
7010         VA = ArgLocs[I++];
7011         assert(VA.getValVT().isVector() &&
7012                "Unexpected Val type for custom RegLoc.");
7013         assert(VA.getValNo() == OriginalValNo &&
7014                "ValNo mismatch between custom MemLoc and RegLoc.");
7015         MVT::SimpleValueType SVT = VA.getLocVT().SimpleTy;
7016         MF.addLiveIn(VA.getLocReg(),
7017                      getRegClassForSVT(SVT, IsPPC64, Subtarget.hasP8Vector(),
7018                                        Subtarget.hasVSX()));
7019       };
7020 
7021       HandleMemLoc();
7022       // In 64-bit there will be exactly 2 custom RegLocs that follow, and in
7023       // in 32-bit there will be 2 custom RegLocs if we are passing in R9 and
7024       // R10.
7025       HandleCustomVecRegLoc();
7026       HandleCustomVecRegLoc();
7027 
7028       // If we are targeting 32-bit, there might be 2 extra custom RegLocs if
7029       // we passed the vector in R5, R6, R7 and R8.
7030       if (I != End && ArgLocs[I].isRegLoc() && ArgLocs[I].needsCustom()) {
7031         assert(!IsPPC64 &&
7032                "Only 2 custom RegLocs expected for 64-bit codegen.");
7033         HandleCustomVecRegLoc();
7034         HandleCustomVecRegLoc();
7035       }
7036 
7037       continue;
7038     }
7039 
7040     if (VA.isRegLoc()) {
7041       if (VA.getValVT().isScalarInteger())
7042         FuncInfo->appendParameterType(PPCFunctionInfo::FixedType);
7043       else if (VA.getValVT().isFloatingPoint() && !VA.getValVT().isVector()) {
7044         switch (VA.getValVT().SimpleTy) {
7045         default:
7046           report_fatal_error("Unhandled value type for argument.");
7047         case MVT::f32:
7048           FuncInfo->appendParameterType(PPCFunctionInfo::ShortFloatingPoint);
7049           break;
7050         case MVT::f64:
7051           FuncInfo->appendParameterType(PPCFunctionInfo::LongFloatingPoint);
7052           break;
7053         }
7054       } else if (VA.getValVT().isVector()) {
7055         switch (VA.getValVT().SimpleTy) {
7056         default:
7057           report_fatal_error("Unhandled value type for argument.");
7058         case MVT::v16i8:
7059           FuncInfo->appendParameterType(PPCFunctionInfo::VectorChar);
7060           break;
7061         case MVT::v8i16:
7062           FuncInfo->appendParameterType(PPCFunctionInfo::VectorShort);
7063           break;
7064         case MVT::v4i32:
7065         case MVT::v2i64:
7066         case MVT::v1i128:
7067           FuncInfo->appendParameterType(PPCFunctionInfo::VectorInt);
7068           break;
7069         case MVT::v4f32:
7070         case MVT::v2f64:
7071           FuncInfo->appendParameterType(PPCFunctionInfo::VectorFloat);
7072           break;
7073         }
7074       }
7075     }
7076 
7077     if (Flags.isByVal() && VA.isMemLoc()) {
7078       const unsigned Size =
7079           alignTo(Flags.getByValSize() ? Flags.getByValSize() : PtrByteSize,
7080                   PtrByteSize);
7081       const int FI = MF.getFrameInfo().CreateFixedObject(
7082           Size, VA.getLocMemOffset(), /* IsImmutable */ false,
7083           /* IsAliased */ true);
7084       SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
7085       InVals.push_back(FIN);
7086 
7087       continue;
7088     }
7089 
7090     if (Flags.isByVal()) {
7091       assert(VA.isRegLoc() && "MemLocs should already be handled.");
7092 
7093       const MCPhysReg ArgReg = VA.getLocReg();
7094       const PPCFrameLowering *FL = Subtarget.getFrameLowering();
7095 
7096       if (Flags.getNonZeroByValAlign() > PtrByteSize)
7097         report_fatal_error("Over aligned byvals not supported yet.");
7098 
7099       const unsigned StackSize = alignTo(Flags.getByValSize(), PtrByteSize);
7100       const int FI = MF.getFrameInfo().CreateFixedObject(
7101           StackSize, mapArgRegToOffsetAIX(ArgReg, FL), /* IsImmutable */ false,
7102           /* IsAliased */ true);
7103       SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
7104       InVals.push_back(FIN);
7105 
7106       // Add live ins for all the RegLocs for the same ByVal.
7107       const TargetRegisterClass *RegClass =
7108           IsPPC64 ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
7109 
7110       auto HandleRegLoc = [&, RegClass, LocVT](const MCPhysReg PhysReg,
7111                                                unsigned Offset) {
7112         const Register VReg = MF.addLiveIn(PhysReg, RegClass);
7113         // Since the callers side has left justified the aggregate in the
7114         // register, we can simply store the entire register into the stack
7115         // slot.
7116         SDValue CopyFrom = DAG.getCopyFromReg(Chain, dl, VReg, LocVT);
7117         // The store to the fixedstack object is needed becuase accessing a
7118         // field of the ByVal will use a gep and load. Ideally we will optimize
7119         // to extracting the value from the register directly, and elide the
7120         // stores when the arguments address is not taken, but that will need to
7121         // be future work.
7122         SDValue Store = DAG.getStore(
7123             CopyFrom.getValue(1), dl, CopyFrom,
7124             DAG.getObjectPtrOffset(dl, FIN, TypeSize::Fixed(Offset)),
7125             MachinePointerInfo::getFixedStack(MF, FI, Offset));
7126 
7127         MemOps.push_back(Store);
7128       };
7129 
7130       unsigned Offset = 0;
7131       HandleRegLoc(VA.getLocReg(), Offset);
7132       Offset += PtrByteSize;
7133       for (; Offset != StackSize && ArgLocs[I].isRegLoc();
7134            Offset += PtrByteSize) {
7135         assert(ArgLocs[I].getValNo() == VA.getValNo() &&
7136                "RegLocs should be for ByVal argument.");
7137 
7138         const CCValAssign RL = ArgLocs[I++];
7139         HandleRegLoc(RL.getLocReg(), Offset);
7140         FuncInfo->appendParameterType(PPCFunctionInfo::FixedType);
7141       }
7142 
7143       if (Offset != StackSize) {
7144         assert(ArgLocs[I].getValNo() == VA.getValNo() &&
7145                "Expected MemLoc for remaining bytes.");
7146         assert(ArgLocs[I].isMemLoc() && "Expected MemLoc for remaining bytes.");
7147         // Consume the MemLoc.The InVal has already been emitted, so nothing
7148         // more needs to be done.
7149         ++I;
7150       }
7151 
7152       continue;
7153     }
7154 
7155     if (VA.isRegLoc() && !VA.needsCustom()) {
7156       MVT::SimpleValueType SVT = ValVT.SimpleTy;
7157       Register VReg =
7158           MF.addLiveIn(VA.getLocReg(),
7159                        getRegClassForSVT(SVT, IsPPC64, Subtarget.hasP8Vector(),
7160                                          Subtarget.hasVSX()));
7161       SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, VReg, LocVT);
7162       if (ValVT.isScalarInteger() &&
7163           (ValVT.getFixedSizeInBits() < LocVT.getFixedSizeInBits())) {
7164         ArgValue =
7165             truncateScalarIntegerArg(Flags, ValVT, DAG, ArgValue, LocVT, dl);
7166       }
7167       InVals.push_back(ArgValue);
7168       continue;
7169     }
7170     if (VA.isMemLoc()) {
7171       HandleMemLoc();
7172       continue;
7173     }
7174   }
7175 
7176   // On AIX a minimum of 8 words is saved to the parameter save area.
7177   const unsigned MinParameterSaveArea = 8 * PtrByteSize;
7178   // Area that is at least reserved in the caller of this function.
7179   unsigned CallerReservedArea =
7180       std::max(CCInfo.getNextStackOffset(), LinkageSize + MinParameterSaveArea);
7181 
7182   // Set the size that is at least reserved in caller of this function. Tail
7183   // call optimized function's reserved stack space needs to be aligned so
7184   // that taking the difference between two stack areas will result in an
7185   // aligned stack.
7186   CallerReservedArea =
7187       EnsureStackAlignment(Subtarget.getFrameLowering(), CallerReservedArea);
7188   FuncInfo->setMinReservedArea(CallerReservedArea);
7189 
7190   if (isVarArg) {
7191     FuncInfo->setVarArgsFrameIndex(
7192         MFI.CreateFixedObject(PtrByteSize, CCInfo.getNextStackOffset(), true));
7193     SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
7194 
7195     static const MCPhysReg GPR_32[] = {PPC::R3, PPC::R4, PPC::R5, PPC::R6,
7196                                        PPC::R7, PPC::R8, PPC::R9, PPC::R10};
7197 
7198     static const MCPhysReg GPR_64[] = {PPC::X3, PPC::X4, PPC::X5, PPC::X6,
7199                                        PPC::X7, PPC::X8, PPC::X9, PPC::X10};
7200     const unsigned NumGPArgRegs = std::size(IsPPC64 ? GPR_64 : GPR_32);
7201 
7202     // The fixed integer arguments of a variadic function are stored to the
7203     // VarArgsFrameIndex on the stack so that they may be loaded by
7204     // dereferencing the result of va_next.
7205     for (unsigned GPRIndex =
7206              (CCInfo.getNextStackOffset() - LinkageSize) / PtrByteSize;
7207          GPRIndex < NumGPArgRegs; ++GPRIndex) {
7208 
7209       const Register VReg =
7210           IsPPC64 ? MF.addLiveIn(GPR_64[GPRIndex], &PPC::G8RCRegClass)
7211                   : MF.addLiveIn(GPR_32[GPRIndex], &PPC::GPRCRegClass);
7212 
7213       SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
7214       SDValue Store =
7215           DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo());
7216       MemOps.push_back(Store);
7217       // Increment the address for the next argument to store.
7218       SDValue PtrOff = DAG.getConstant(PtrByteSize, dl, PtrVT);
7219       FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
7220     }
7221   }
7222 
7223   if (!MemOps.empty())
7224     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
7225 
7226   return Chain;
7227 }
7228 
7229 SDValue PPCTargetLowering::LowerCall_AIX(
7230     SDValue Chain, SDValue Callee, CallFlags CFlags,
7231     const SmallVectorImpl<ISD::OutputArg> &Outs,
7232     const SmallVectorImpl<SDValue> &OutVals,
7233     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
7234     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals,
7235     const CallBase *CB) const {
7236   // See PPCTargetLowering::LowerFormalArguments_AIX() for a description of the
7237   // AIX ABI stack frame layout.
7238 
7239   assert((CFlags.CallConv == CallingConv::C ||
7240           CFlags.CallConv == CallingConv::Cold ||
7241           CFlags.CallConv == CallingConv::Fast) &&
7242          "Unexpected calling convention!");
7243 
7244   if (CFlags.IsPatchPoint)
7245     report_fatal_error("This call type is unimplemented on AIX.");
7246 
7247   const PPCSubtarget &Subtarget = DAG.getSubtarget<PPCSubtarget>();
7248 
7249   MachineFunction &MF = DAG.getMachineFunction();
7250   SmallVector<CCValAssign, 16> ArgLocs;
7251   AIXCCState CCInfo(CFlags.CallConv, CFlags.IsVarArg, MF, ArgLocs,
7252                     *DAG.getContext());
7253 
7254   // Reserve space for the linkage save area (LSA) on the stack.
7255   // In both PPC32 and PPC64 there are 6 reserved slots in the LSA:
7256   //   [SP][CR][LR][2 x reserved][TOC].
7257   // The LSA is 24 bytes (6x4) in PPC32 and 48 bytes (6x8) in PPC64.
7258   const unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
7259   const bool IsPPC64 = Subtarget.isPPC64();
7260   const EVT PtrVT = getPointerTy(DAG.getDataLayout());
7261   const unsigned PtrByteSize = IsPPC64 ? 8 : 4;
7262   CCInfo.AllocateStack(LinkageSize, Align(PtrByteSize));
7263   CCInfo.AnalyzeCallOperands(Outs, CC_AIX);
7264 
7265   // The prolog code of the callee may store up to 8 GPR argument registers to
7266   // the stack, allowing va_start to index over them in memory if the callee
7267   // is variadic.
7268   // Because we cannot tell if this is needed on the caller side, we have to
7269   // conservatively assume that it is needed.  As such, make sure we have at
7270   // least enough stack space for the caller to store the 8 GPRs.
7271   const unsigned MinParameterSaveAreaSize = 8 * PtrByteSize;
7272   const unsigned NumBytes = std::max(LinkageSize + MinParameterSaveAreaSize,
7273                                      CCInfo.getNextStackOffset());
7274 
7275   // Adjust the stack pointer for the new arguments...
7276   // These operations are automatically eliminated by the prolog/epilog pass.
7277   Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl);
7278   SDValue CallSeqStart = Chain;
7279 
7280   SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
7281   SmallVector<SDValue, 8> MemOpChains;
7282 
7283   // Set up a copy of the stack pointer for loading and storing any
7284   // arguments that may not fit in the registers available for argument
7285   // passing.
7286   const SDValue StackPtr = IsPPC64 ? DAG.getRegister(PPC::X1, MVT::i64)
7287                                    : DAG.getRegister(PPC::R1, MVT::i32);
7288 
7289   for (unsigned I = 0, E = ArgLocs.size(); I != E;) {
7290     const unsigned ValNo = ArgLocs[I].getValNo();
7291     SDValue Arg = OutVals[ValNo];
7292     ISD::ArgFlagsTy Flags = Outs[ValNo].Flags;
7293 
7294     if (Flags.isByVal()) {
7295       const unsigned ByValSize = Flags.getByValSize();
7296 
7297       // Nothing to do for zero-sized ByVals on the caller side.
7298       if (!ByValSize) {
7299         ++I;
7300         continue;
7301       }
7302 
7303       auto GetLoad = [&](EVT VT, unsigned LoadOffset) {
7304         return DAG.getExtLoad(
7305             ISD::ZEXTLOAD, dl, PtrVT, Chain,
7306             (LoadOffset != 0)
7307                 ? DAG.getObjectPtrOffset(dl, Arg, TypeSize::Fixed(LoadOffset))
7308                 : Arg,
7309             MachinePointerInfo(), VT);
7310       };
7311 
7312       unsigned LoadOffset = 0;
7313 
7314       // Initialize registers, which are fully occupied by the by-val argument.
7315       while (LoadOffset + PtrByteSize <= ByValSize && ArgLocs[I].isRegLoc()) {
7316         SDValue Load = GetLoad(PtrVT, LoadOffset);
7317         MemOpChains.push_back(Load.getValue(1));
7318         LoadOffset += PtrByteSize;
7319         const CCValAssign &ByValVA = ArgLocs[I++];
7320         assert(ByValVA.getValNo() == ValNo &&
7321                "Unexpected location for pass-by-value argument.");
7322         RegsToPass.push_back(std::make_pair(ByValVA.getLocReg(), Load));
7323       }
7324 
7325       if (LoadOffset == ByValSize)
7326         continue;
7327 
7328       // There must be one more loc to handle the remainder.
7329       assert(ArgLocs[I].getValNo() == ValNo &&
7330              "Expected additional location for by-value argument.");
7331 
7332       if (ArgLocs[I].isMemLoc()) {
7333         assert(LoadOffset < ByValSize && "Unexpected memloc for by-val arg.");
7334         const CCValAssign &ByValVA = ArgLocs[I++];
7335         ISD::ArgFlagsTy MemcpyFlags = Flags;
7336         // Only memcpy the bytes that don't pass in register.
7337         MemcpyFlags.setByValSize(ByValSize - LoadOffset);
7338         Chain = CallSeqStart = createMemcpyOutsideCallSeq(
7339             (LoadOffset != 0)
7340                 ? DAG.getObjectPtrOffset(dl, Arg, TypeSize::Fixed(LoadOffset))
7341                 : Arg,
7342             DAG.getObjectPtrOffset(dl, StackPtr,
7343                                    TypeSize::Fixed(ByValVA.getLocMemOffset())),
7344             CallSeqStart, MemcpyFlags, DAG, dl);
7345         continue;
7346       }
7347 
7348       // Initialize the final register residue.
7349       // Any residue that occupies the final by-val arg register must be
7350       // left-justified on AIX. Loads must be a power-of-2 size and cannot be
7351       // larger than the ByValSize. For example: a 7 byte by-val arg requires 4,
7352       // 2 and 1 byte loads.
7353       const unsigned ResidueBytes = ByValSize % PtrByteSize;
7354       assert(ResidueBytes != 0 && LoadOffset + PtrByteSize > ByValSize &&
7355              "Unexpected register residue for by-value argument.");
7356       SDValue ResidueVal;
7357       for (unsigned Bytes = 0; Bytes != ResidueBytes;) {
7358         const unsigned N = PowerOf2Floor(ResidueBytes - Bytes);
7359         const MVT VT =
7360             N == 1 ? MVT::i8
7361                    : ((N == 2) ? MVT::i16 : (N == 4 ? MVT::i32 : MVT::i64));
7362         SDValue Load = GetLoad(VT, LoadOffset);
7363         MemOpChains.push_back(Load.getValue(1));
7364         LoadOffset += N;
7365         Bytes += N;
7366 
7367         // By-val arguments are passed left-justfied in register.
7368         // Every load here needs to be shifted, otherwise a full register load
7369         // should have been used.
7370         assert(PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) &&
7371                "Unexpected load emitted during handling of pass-by-value "
7372                "argument.");
7373         unsigned NumSHLBits = PtrVT.getSimpleVT().getSizeInBits() - (Bytes * 8);
7374         EVT ShiftAmountTy =
7375             getShiftAmountTy(Load->getValueType(0), DAG.getDataLayout());
7376         SDValue SHLAmt = DAG.getConstant(NumSHLBits, dl, ShiftAmountTy);
7377         SDValue ShiftedLoad =
7378             DAG.getNode(ISD::SHL, dl, Load.getValueType(), Load, SHLAmt);
7379         ResidueVal = ResidueVal ? DAG.getNode(ISD::OR, dl, PtrVT, ResidueVal,
7380                                               ShiftedLoad)
7381                                 : ShiftedLoad;
7382       }
7383 
7384       const CCValAssign &ByValVA = ArgLocs[I++];
7385       RegsToPass.push_back(std::make_pair(ByValVA.getLocReg(), ResidueVal));
7386       continue;
7387     }
7388 
7389     CCValAssign &VA = ArgLocs[I++];
7390     const MVT LocVT = VA.getLocVT();
7391     const MVT ValVT = VA.getValVT();
7392 
7393     switch (VA.getLocInfo()) {
7394     default:
7395       report_fatal_error("Unexpected argument extension type.");
7396     case CCValAssign::Full:
7397       break;
7398     case CCValAssign::ZExt:
7399       Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
7400       break;
7401     case CCValAssign::SExt:
7402       Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
7403       break;
7404     }
7405 
7406     if (VA.isRegLoc() && !VA.needsCustom()) {
7407       RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
7408       continue;
7409     }
7410 
7411     // Vector arguments passed to VarArg functions need custom handling when
7412     // they are passed (at least partially) in GPRs.
7413     if (VA.isMemLoc() && VA.needsCustom() && ValVT.isVector()) {
7414       assert(CFlags.IsVarArg && "Custom MemLocs only used for Vector args.");
7415       // Store value to its stack slot.
7416       SDValue PtrOff =
7417           DAG.getConstant(VA.getLocMemOffset(), dl, StackPtr.getValueType());
7418       PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
7419       SDValue Store =
7420           DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo());
7421       MemOpChains.push_back(Store);
7422       const unsigned OriginalValNo = VA.getValNo();
7423       // Then load the GPRs from the stack
7424       unsigned LoadOffset = 0;
7425       auto HandleCustomVecRegLoc = [&]() {
7426         assert(I != E && "Unexpected end of CCvalAssigns.");
7427         assert(ArgLocs[I].isRegLoc() && ArgLocs[I].needsCustom() &&
7428                "Expected custom RegLoc.");
7429         CCValAssign RegVA = ArgLocs[I++];
7430         assert(RegVA.getValNo() == OriginalValNo &&
7431                "Custom MemLoc ValNo and custom RegLoc ValNo must match.");
7432         SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
7433                                   DAG.getConstant(LoadOffset, dl, PtrVT));
7434         SDValue Load = DAG.getLoad(PtrVT, dl, Store, Add, MachinePointerInfo());
7435         MemOpChains.push_back(Load.getValue(1));
7436         RegsToPass.push_back(std::make_pair(RegVA.getLocReg(), Load));
7437         LoadOffset += PtrByteSize;
7438       };
7439 
7440       // In 64-bit there will be exactly 2 custom RegLocs that follow, and in
7441       // in 32-bit there will be 2 custom RegLocs if we are passing in R9 and
7442       // R10.
7443       HandleCustomVecRegLoc();
7444       HandleCustomVecRegLoc();
7445 
7446       if (I != E && ArgLocs[I].isRegLoc() && ArgLocs[I].needsCustom() &&
7447           ArgLocs[I].getValNo() == OriginalValNo) {
7448         assert(!IsPPC64 &&
7449                "Only 2 custom RegLocs expected for 64-bit codegen.");
7450         HandleCustomVecRegLoc();
7451         HandleCustomVecRegLoc();
7452       }
7453 
7454       continue;
7455     }
7456 
7457     if (VA.isMemLoc()) {
7458       SDValue PtrOff =
7459           DAG.getConstant(VA.getLocMemOffset(), dl, StackPtr.getValueType());
7460       PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
7461       MemOpChains.push_back(
7462           DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo()));
7463 
7464       continue;
7465     }
7466 
7467     if (!ValVT.isFloatingPoint())
7468       report_fatal_error(
7469           "Unexpected register handling for calling convention.");
7470 
7471     // Custom handling is used for GPR initializations for vararg float
7472     // arguments.
7473     assert(VA.isRegLoc() && VA.needsCustom() && CFlags.IsVarArg &&
7474            LocVT.isInteger() &&
7475            "Custom register handling only expected for VarArg.");
7476 
7477     SDValue ArgAsInt =
7478         DAG.getBitcast(MVT::getIntegerVT(ValVT.getSizeInBits()), Arg);
7479 
7480     if (Arg.getValueType().getStoreSize() == LocVT.getStoreSize())
7481       // f32 in 32-bit GPR
7482       // f64 in 64-bit GPR
7483       RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgAsInt));
7484     else if (Arg.getValueType().getFixedSizeInBits() <
7485              LocVT.getFixedSizeInBits())
7486       // f32 in 64-bit GPR.
7487       RegsToPass.push_back(std::make_pair(
7488           VA.getLocReg(), DAG.getZExtOrTrunc(ArgAsInt, dl, LocVT)));
7489     else {
7490       // f64 in two 32-bit GPRs
7491       // The 2 GPRs are marked custom and expected to be adjacent in ArgLocs.
7492       assert(Arg.getValueType() == MVT::f64 && CFlags.IsVarArg && !IsPPC64 &&
7493              "Unexpected custom register for argument!");
7494       CCValAssign &GPR1 = VA;
7495       SDValue MSWAsI64 = DAG.getNode(ISD::SRL, dl, MVT::i64, ArgAsInt,
7496                                      DAG.getConstant(32, dl, MVT::i8));
7497       RegsToPass.push_back(std::make_pair(
7498           GPR1.getLocReg(), DAG.getZExtOrTrunc(MSWAsI64, dl, MVT::i32)));
7499 
7500       if (I != E) {
7501         // If only 1 GPR was available, there will only be one custom GPR and
7502         // the argument will also pass in memory.
7503         CCValAssign &PeekArg = ArgLocs[I];
7504         if (PeekArg.isRegLoc() && PeekArg.getValNo() == PeekArg.getValNo()) {
7505           assert(PeekArg.needsCustom() && "A second custom GPR is expected.");
7506           CCValAssign &GPR2 = ArgLocs[I++];
7507           RegsToPass.push_back(std::make_pair(
7508               GPR2.getLocReg(), DAG.getZExtOrTrunc(ArgAsInt, dl, MVT::i32)));
7509         }
7510       }
7511     }
7512   }
7513 
7514   if (!MemOpChains.empty())
7515     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
7516 
7517   // For indirect calls, we need to save the TOC base to the stack for
7518   // restoration after the call.
7519   if (CFlags.IsIndirect) {
7520     assert(!CFlags.IsTailCall && "Indirect tail-calls not supported.");
7521     const MCRegister TOCBaseReg = Subtarget.getTOCPointerRegister();
7522     const MCRegister StackPtrReg = Subtarget.getStackPointerRegister();
7523     const MVT PtrVT = Subtarget.isPPC64() ? MVT::i64 : MVT::i32;
7524     const unsigned TOCSaveOffset =
7525         Subtarget.getFrameLowering()->getTOCSaveOffset();
7526 
7527     setUsesTOCBasePtr(DAG);
7528     SDValue Val = DAG.getCopyFromReg(Chain, dl, TOCBaseReg, PtrVT);
7529     SDValue PtrOff = DAG.getIntPtrConstant(TOCSaveOffset, dl);
7530     SDValue StackPtr = DAG.getRegister(StackPtrReg, PtrVT);
7531     SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
7532     Chain = DAG.getStore(
7533         Val.getValue(1), dl, Val, AddPtr,
7534         MachinePointerInfo::getStack(DAG.getMachineFunction(), TOCSaveOffset));
7535   }
7536 
7537   // Build a sequence of copy-to-reg nodes chained together with token chain
7538   // and flag operands which copy the outgoing args into the appropriate regs.
7539   SDValue InFlag;
7540   for (auto Reg : RegsToPass) {
7541     Chain = DAG.getCopyToReg(Chain, dl, Reg.first, Reg.second, InFlag);
7542     InFlag = Chain.getValue(1);
7543   }
7544 
7545   const int SPDiff = 0;
7546   return FinishCall(CFlags, dl, DAG, RegsToPass, InFlag, Chain, CallSeqStart,
7547                     Callee, SPDiff, NumBytes, Ins, InVals, CB);
7548 }
7549 
7550 bool
7551 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
7552                                   MachineFunction &MF, bool isVarArg,
7553                                   const SmallVectorImpl<ISD::OutputArg> &Outs,
7554                                   LLVMContext &Context) const {
7555   SmallVector<CCValAssign, 16> RVLocs;
7556   CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
7557   return CCInfo.CheckReturn(
7558       Outs, (Subtarget.isSVR4ABI() && CallConv == CallingConv::Cold)
7559                 ? RetCC_PPC_Cold
7560                 : RetCC_PPC);
7561 }
7562 
7563 SDValue
7564 PPCTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
7565                                bool isVarArg,
7566                                const SmallVectorImpl<ISD::OutputArg> &Outs,
7567                                const SmallVectorImpl<SDValue> &OutVals,
7568                                const SDLoc &dl, SelectionDAG &DAG) const {
7569   SmallVector<CCValAssign, 16> RVLocs;
7570   CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
7571                  *DAG.getContext());
7572   CCInfo.AnalyzeReturn(Outs,
7573                        (Subtarget.isSVR4ABI() && CallConv == CallingConv::Cold)
7574                            ? RetCC_PPC_Cold
7575                            : RetCC_PPC);
7576 
7577   SDValue Flag;
7578   SmallVector<SDValue, 4> RetOps(1, Chain);
7579 
7580   // Copy the result values into the output registers.
7581   for (unsigned i = 0, RealResIdx = 0; i != RVLocs.size(); ++i, ++RealResIdx) {
7582     CCValAssign &VA = RVLocs[i];
7583     assert(VA.isRegLoc() && "Can only return in registers!");
7584 
7585     SDValue Arg = OutVals[RealResIdx];
7586 
7587     switch (VA.getLocInfo()) {
7588     default: llvm_unreachable("Unknown loc info!");
7589     case CCValAssign::Full: break;
7590     case CCValAssign::AExt:
7591       Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
7592       break;
7593     case CCValAssign::ZExt:
7594       Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
7595       break;
7596     case CCValAssign::SExt:
7597       Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
7598       break;
7599     }
7600     if (Subtarget.hasSPE() && VA.getLocVT() == MVT::f64) {
7601       bool isLittleEndian = Subtarget.isLittleEndian();
7602       // Legalize ret f64 -> ret 2 x i32.
7603       SDValue SVal =
7604           DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg,
7605                       DAG.getIntPtrConstant(isLittleEndian ? 0 : 1, dl));
7606       Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), SVal, Flag);
7607       RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
7608       SVal = DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg,
7609                          DAG.getIntPtrConstant(isLittleEndian ? 1 : 0, dl));
7610       Flag = Chain.getValue(1);
7611       VA = RVLocs[++i]; // skip ahead to next loc
7612       Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), SVal, Flag);
7613     } else
7614       Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
7615     Flag = Chain.getValue(1);
7616     RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
7617   }
7618 
7619   RetOps[0] = Chain;  // Update chain.
7620 
7621   // Add the flag if we have it.
7622   if (Flag.getNode())
7623     RetOps.push_back(Flag);
7624 
7625   return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, RetOps);
7626 }
7627 
7628 SDValue
7629 PPCTargetLowering::LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op,
7630                                                 SelectionDAG &DAG) const {
7631   SDLoc dl(Op);
7632 
7633   // Get the correct type for integers.
7634   EVT IntVT = Op.getValueType();
7635 
7636   // Get the inputs.
7637   SDValue Chain = Op.getOperand(0);
7638   SDValue FPSIdx = getFramePointerFrameIndex(DAG);
7639   // Build a DYNAREAOFFSET node.
7640   SDValue Ops[2] = {Chain, FPSIdx};
7641   SDVTList VTs = DAG.getVTList(IntVT);
7642   return DAG.getNode(PPCISD::DYNAREAOFFSET, dl, VTs, Ops);
7643 }
7644 
7645 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op,
7646                                              SelectionDAG &DAG) const {
7647   // When we pop the dynamic allocation we need to restore the SP link.
7648   SDLoc dl(Op);
7649 
7650   // Get the correct type for pointers.
7651   EVT PtrVT = getPointerTy(DAG.getDataLayout());
7652 
7653   // Construct the stack pointer operand.
7654   bool isPPC64 = Subtarget.isPPC64();
7655   unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
7656   SDValue StackPtr = DAG.getRegister(SP, PtrVT);
7657 
7658   // Get the operands for the STACKRESTORE.
7659   SDValue Chain = Op.getOperand(0);
7660   SDValue SaveSP = Op.getOperand(1);
7661 
7662   // Load the old link SP.
7663   SDValue LoadLinkSP =
7664       DAG.getLoad(PtrVT, dl, Chain, StackPtr, MachinePointerInfo());
7665 
7666   // Restore the stack pointer.
7667   Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
7668 
7669   // Store the old link SP.
7670   return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo());
7671 }
7672 
7673 SDValue PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG &DAG) const {
7674   MachineFunction &MF = DAG.getMachineFunction();
7675   bool isPPC64 = Subtarget.isPPC64();
7676   EVT PtrVT = getPointerTy(MF.getDataLayout());
7677 
7678   // Get current frame pointer save index.  The users of this index will be
7679   // primarily DYNALLOC instructions.
7680   PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
7681   int RASI = FI->getReturnAddrSaveIndex();
7682 
7683   // If the frame pointer save index hasn't been defined yet.
7684   if (!RASI) {
7685     // Find out what the fix offset of the frame pointer save area.
7686     int LROffset = Subtarget.getFrameLowering()->getReturnSaveOffset();
7687     // Allocate the frame index for frame pointer save area.
7688     RASI = MF.getFrameInfo().CreateFixedObject(isPPC64? 8 : 4, LROffset, false);
7689     // Save the result.
7690     FI->setReturnAddrSaveIndex(RASI);
7691   }
7692   return DAG.getFrameIndex(RASI, PtrVT);
7693 }
7694 
7695 SDValue
7696 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
7697   MachineFunction &MF = DAG.getMachineFunction();
7698   bool isPPC64 = Subtarget.isPPC64();
7699   EVT PtrVT = getPointerTy(MF.getDataLayout());
7700 
7701   // Get current frame pointer save index.  The users of this index will be
7702   // primarily DYNALLOC instructions.
7703   PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
7704   int FPSI = FI->getFramePointerSaveIndex();
7705 
7706   // If the frame pointer save index hasn't been defined yet.
7707   if (!FPSI) {
7708     // Find out what the fix offset of the frame pointer save area.
7709     int FPOffset = Subtarget.getFrameLowering()->getFramePointerSaveOffset();
7710     // Allocate the frame index for frame pointer save area.
7711     FPSI = MF.getFrameInfo().CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
7712     // Save the result.
7713     FI->setFramePointerSaveIndex(FPSI);
7714   }
7715   return DAG.getFrameIndex(FPSI, PtrVT);
7716 }
7717 
7718 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7719                                                    SelectionDAG &DAG) const {
7720   MachineFunction &MF = DAG.getMachineFunction();
7721   // Get the inputs.
7722   SDValue Chain = Op.getOperand(0);
7723   SDValue Size  = Op.getOperand(1);
7724   SDLoc dl(Op);
7725 
7726   // Get the correct type for pointers.
7727   EVT PtrVT = getPointerTy(DAG.getDataLayout());
7728   // Negate the size.
7729   SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
7730                                 DAG.getConstant(0, dl, PtrVT), Size);
7731   // Construct a node for the frame pointer save index.
7732   SDValue FPSIdx = getFramePointerFrameIndex(DAG);
7733   SDValue Ops[3] = { Chain, NegSize, FPSIdx };
7734   SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
7735   if (hasInlineStackProbe(MF))
7736     return DAG.getNode(PPCISD::PROBED_ALLOCA, dl, VTs, Ops);
7737   return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops);
7738 }
7739 
7740 SDValue PPCTargetLowering::LowerEH_DWARF_CFA(SDValue Op,
7741                                                      SelectionDAG &DAG) const {
7742   MachineFunction &MF = DAG.getMachineFunction();
7743 
7744   bool isPPC64 = Subtarget.isPPC64();
7745   EVT PtrVT = getPointerTy(DAG.getDataLayout());
7746 
7747   int FI = MF.getFrameInfo().CreateFixedObject(isPPC64 ? 8 : 4, 0, false);
7748   return DAG.getFrameIndex(FI, PtrVT);
7749 }
7750 
7751 SDValue PPCTargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
7752                                                SelectionDAG &DAG) const {
7753   SDLoc DL(Op);
7754   return DAG.getNode(PPCISD::EH_SJLJ_SETJMP, DL,
7755                      DAG.getVTList(MVT::i32, MVT::Other),
7756                      Op.getOperand(0), Op.getOperand(1));
7757 }
7758 
7759 SDValue PPCTargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
7760                                                 SelectionDAG &DAG) const {
7761   SDLoc DL(Op);
7762   return DAG.getNode(PPCISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
7763                      Op.getOperand(0), Op.getOperand(1));
7764 }
7765 
7766 SDValue PPCTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
7767   if (Op.getValueType().isVector())
7768     return LowerVectorLoad(Op, DAG);
7769 
7770   assert(Op.getValueType() == MVT::i1 &&
7771          "Custom lowering only for i1 loads");
7772 
7773   // First, load 8 bits into 32 bits, then truncate to 1 bit.
7774 
7775   SDLoc dl(Op);
7776   LoadSDNode *LD = cast<LoadSDNode>(Op);
7777 
7778   SDValue Chain = LD->getChain();
7779   SDValue BasePtr = LD->getBasePtr();
7780   MachineMemOperand *MMO = LD->getMemOperand();
7781 
7782   SDValue NewLD =
7783       DAG.getExtLoad(ISD::EXTLOAD, dl, getPointerTy(DAG.getDataLayout()), Chain,
7784                      BasePtr, MVT::i8, MMO);
7785   SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewLD);
7786 
7787   SDValue Ops[] = { Result, SDValue(NewLD.getNode(), 1) };
7788   return DAG.getMergeValues(Ops, dl);
7789 }
7790 
7791 SDValue PPCTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
7792   if (Op.getOperand(1).getValueType().isVector())
7793     return LowerVectorStore(Op, DAG);
7794 
7795   assert(Op.getOperand(1).getValueType() == MVT::i1 &&
7796          "Custom lowering only for i1 stores");
7797 
7798   // First, zero extend to 32 bits, then use a truncating store to 8 bits.
7799 
7800   SDLoc dl(Op);
7801   StoreSDNode *ST = cast<StoreSDNode>(Op);
7802 
7803   SDValue Chain = ST->getChain();
7804   SDValue BasePtr = ST->getBasePtr();
7805   SDValue Value = ST->getValue();
7806   MachineMemOperand *MMO = ST->getMemOperand();
7807 
7808   Value = DAG.getNode(ISD::ZERO_EXTEND, dl, getPointerTy(DAG.getDataLayout()),
7809                       Value);
7810   return DAG.getTruncStore(Chain, dl, Value, BasePtr, MVT::i8, MMO);
7811 }
7812 
7813 // FIXME: Remove this once the ANDI glue bug is fixed:
7814 SDValue PPCTargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
7815   assert(Op.getValueType() == MVT::i1 &&
7816          "Custom lowering only for i1 results");
7817 
7818   SDLoc DL(Op);
7819   return DAG.getNode(PPCISD::ANDI_rec_1_GT_BIT, DL, MVT::i1, Op.getOperand(0));
7820 }
7821 
7822 SDValue PPCTargetLowering::LowerTRUNCATEVector(SDValue Op,
7823                                                SelectionDAG &DAG) const {
7824 
7825   // Implements a vector truncate that fits in a vector register as a shuffle.
7826   // We want to legalize vector truncates down to where the source fits in
7827   // a vector register (and target is therefore smaller than vector register
7828   // size).  At that point legalization will try to custom lower the sub-legal
7829   // result and get here - where we can contain the truncate as a single target
7830   // operation.
7831 
7832   // For example a trunc <2 x i16> to <2 x i8> could be visualized as follows:
7833   //   <MSB1|LSB1, MSB2|LSB2> to <LSB1, LSB2>
7834   //
7835   // We will implement it for big-endian ordering as this (where x denotes
7836   // undefined):
7837   //   < MSB1|LSB1, MSB2|LSB2, uu, uu, uu, uu, uu, uu> to
7838   //   < LSB1, LSB2, u, u, u, u, u, u, u, u, u, u, u, u, u, u>
7839   //
7840   // The same operation in little-endian ordering will be:
7841   //   <uu, uu, uu, uu, uu, uu, LSB2|MSB2, LSB1|MSB1> to
7842   //   <u, u, u, u, u, u, u, u, u, u, u, u, u, u, LSB2, LSB1>
7843 
7844   EVT TrgVT = Op.getValueType();
7845   assert(TrgVT.isVector() && "Vector type expected.");
7846   unsigned TrgNumElts = TrgVT.getVectorNumElements();
7847   EVT EltVT = TrgVT.getVectorElementType();
7848   if (!isOperationCustom(Op.getOpcode(), TrgVT) ||
7849       TrgVT.getSizeInBits() > 128 || !isPowerOf2_32(TrgNumElts) ||
7850       !isPowerOf2_32(EltVT.getSizeInBits()))
7851     return SDValue();
7852 
7853   SDValue N1 = Op.getOperand(0);
7854   EVT SrcVT = N1.getValueType();
7855   unsigned SrcSize = SrcVT.getSizeInBits();
7856   if (SrcSize > 256 ||
7857       !isPowerOf2_32(SrcVT.getVectorNumElements()) ||
7858       !isPowerOf2_32(SrcVT.getVectorElementType().getSizeInBits()))
7859     return SDValue();
7860   if (SrcSize == 256 && SrcVT.getVectorNumElements() < 2)
7861     return SDValue();
7862 
7863   unsigned WideNumElts = 128 / EltVT.getSizeInBits();
7864   EVT WideVT = EVT::getVectorVT(*DAG.getContext(), EltVT, WideNumElts);
7865 
7866   SDLoc DL(Op);
7867   SDValue Op1, Op2;
7868   if (SrcSize == 256) {
7869     EVT VecIdxTy = getVectorIdxTy(DAG.getDataLayout());
7870     EVT SplitVT =
7871         N1.getValueType().getHalfNumVectorElementsVT(*DAG.getContext());
7872     unsigned SplitNumElts = SplitVT.getVectorNumElements();
7873     Op1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, N1,
7874                       DAG.getConstant(0, DL, VecIdxTy));
7875     Op2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, N1,
7876                       DAG.getConstant(SplitNumElts, DL, VecIdxTy));
7877   }
7878   else {
7879     Op1 = SrcSize == 128 ? N1 : widenVec(DAG, N1, DL);
7880     Op2 = DAG.getUNDEF(WideVT);
7881   }
7882 
7883   // First list the elements we want to keep.
7884   unsigned SizeMult = SrcSize / TrgVT.getSizeInBits();
7885   SmallVector<int, 16> ShuffV;
7886   if (Subtarget.isLittleEndian())
7887     for (unsigned i = 0; i < TrgNumElts; ++i)
7888       ShuffV.push_back(i * SizeMult);
7889   else
7890     for (unsigned i = 1; i <= TrgNumElts; ++i)
7891       ShuffV.push_back(i * SizeMult - 1);
7892 
7893   // Populate the remaining elements with undefs.
7894   for (unsigned i = TrgNumElts; i < WideNumElts; ++i)
7895     // ShuffV.push_back(i + WideNumElts);
7896     ShuffV.push_back(WideNumElts + 1);
7897 
7898   Op1 = DAG.getNode(ISD::BITCAST, DL, WideVT, Op1);
7899   Op2 = DAG.getNode(ISD::BITCAST, DL, WideVT, Op2);
7900   return DAG.getVectorShuffle(WideVT, DL, Op1, Op2, ShuffV);
7901 }
7902 
7903 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
7904 /// possible.
7905 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
7906   ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
7907   EVT ResVT = Op.getValueType();
7908   EVT CmpVT = Op.getOperand(0).getValueType();
7909   SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
7910   SDValue TV  = Op.getOperand(2), FV  = Op.getOperand(3);
7911   SDLoc dl(Op);
7912 
7913   // Without power9-vector, we don't have native instruction for f128 comparison.
7914   // Following transformation to libcall is needed for setcc:
7915   // select_cc lhs, rhs, tv, fv, cc -> select_cc (setcc cc, x, y), 0, tv, fv, NE
7916   if (!Subtarget.hasP9Vector() && CmpVT == MVT::f128) {
7917     SDValue Z = DAG.getSetCC(
7918         dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), CmpVT),
7919         LHS, RHS, CC);
7920     SDValue Zero = DAG.getConstant(0, dl, Z.getValueType());
7921     return DAG.getSelectCC(dl, Z, Zero, TV, FV, ISD::SETNE);
7922   }
7923 
7924   // Not FP, or using SPE? Not a fsel.
7925   if (!CmpVT.isFloatingPoint() || !TV.getValueType().isFloatingPoint() ||
7926       Subtarget.hasSPE())
7927     return Op;
7928 
7929   SDNodeFlags Flags = Op.getNode()->getFlags();
7930 
7931   // We have xsmaxc[dq]p/xsminc[dq]p which are OK to emit even in the
7932   // presence of infinities.
7933   if (Subtarget.hasP9Vector() && LHS == TV && RHS == FV) {
7934     switch (CC) {
7935     default:
7936       break;
7937     case ISD::SETOGT:
7938     case ISD::SETGT:
7939       return DAG.getNode(PPCISD::XSMAXC, dl, Op.getValueType(), LHS, RHS);
7940     case ISD::SETOLT:
7941     case ISD::SETLT:
7942       return DAG.getNode(PPCISD::XSMINC, dl, Op.getValueType(), LHS, RHS);
7943     }
7944   }
7945 
7946   // We might be able to do better than this under some circumstances, but in
7947   // general, fsel-based lowering of select is a finite-math-only optimization.
7948   // For more information, see section F.3 of the 2.06 ISA specification.
7949   // With ISA 3.0
7950   if ((!DAG.getTarget().Options.NoInfsFPMath && !Flags.hasNoInfs()) ||
7951       (!DAG.getTarget().Options.NoNaNsFPMath && !Flags.hasNoNaNs()))
7952     return Op;
7953 
7954   // If the RHS of the comparison is a 0.0, we don't need to do the
7955   // subtraction at all.
7956   SDValue Sel1;
7957   if (isFloatingPointZero(RHS))
7958     switch (CC) {
7959     default: break;       // SETUO etc aren't handled by fsel.
7960     case ISD::SETNE:
7961       std::swap(TV, FV);
7962       [[fallthrough]];
7963     case ISD::SETEQ:
7964       if (LHS.getValueType() == MVT::f32)   // Comparison is always 64-bits
7965         LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
7966       Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
7967       if (Sel1.getValueType() == MVT::f32)   // Comparison is always 64-bits
7968         Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1);
7969       return DAG.getNode(PPCISD::FSEL, dl, ResVT,
7970                          DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), Sel1, FV);
7971     case ISD::SETULT:
7972     case ISD::SETLT:
7973       std::swap(TV, FV);  // fsel is natively setge, swap operands for setlt
7974       [[fallthrough]];
7975     case ISD::SETOGE:
7976     case ISD::SETGE:
7977       if (LHS.getValueType() == MVT::f32)   // Comparison is always 64-bits
7978         LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
7979       return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
7980     case ISD::SETUGT:
7981     case ISD::SETGT:
7982       std::swap(TV, FV);  // fsel is natively setge, swap operands for setlt
7983       [[fallthrough]];
7984     case ISD::SETOLE:
7985     case ISD::SETLE:
7986       if (LHS.getValueType() == MVT::f32)   // Comparison is always 64-bits
7987         LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
7988       return DAG.getNode(PPCISD::FSEL, dl, ResVT,
7989                          DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
7990     }
7991 
7992   SDValue Cmp;
7993   switch (CC) {
7994   default: break;       // SETUO etc aren't handled by fsel.
7995   case ISD::SETNE:
7996     std::swap(TV, FV);
7997     [[fallthrough]];
7998   case ISD::SETEQ:
7999     Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, Flags);
8000     if (Cmp.getValueType() == MVT::f32)   // Comparison is always 64-bits
8001       Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
8002     Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
8003     if (Sel1.getValueType() == MVT::f32)   // Comparison is always 64-bits
8004       Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1);
8005     return DAG.getNode(PPCISD::FSEL, dl, ResVT,
8006                        DAG.getNode(ISD::FNEG, dl, MVT::f64, Cmp), Sel1, FV);
8007   case ISD::SETULT:
8008   case ISD::SETLT:
8009     Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, Flags);
8010     if (Cmp.getValueType() == MVT::f32)   // Comparison is always 64-bits
8011       Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
8012     return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
8013   case ISD::SETOGE:
8014   case ISD::SETGE:
8015     Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, Flags);
8016     if (Cmp.getValueType() == MVT::f32)   // Comparison is always 64-bits
8017       Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
8018     return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
8019   case ISD::SETUGT:
8020   case ISD::SETGT:
8021     Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS, Flags);
8022     if (Cmp.getValueType() == MVT::f32)   // Comparison is always 64-bits
8023       Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
8024     return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
8025   case ISD::SETOLE:
8026   case ISD::SETLE:
8027     Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS, Flags);
8028     if (Cmp.getValueType() == MVT::f32)   // Comparison is always 64-bits
8029       Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
8030     return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
8031   }
8032   return Op;
8033 }
8034 
8035 static unsigned getPPCStrictOpcode(unsigned Opc) {
8036   switch (Opc) {
8037   default:
8038     llvm_unreachable("No strict version of this opcode!");
8039   case PPCISD::FCTIDZ:
8040     return PPCISD::STRICT_FCTIDZ;
8041   case PPCISD::FCTIWZ:
8042     return PPCISD::STRICT_FCTIWZ;
8043   case PPCISD::FCTIDUZ:
8044     return PPCISD::STRICT_FCTIDUZ;
8045   case PPCISD::FCTIWUZ:
8046     return PPCISD::STRICT_FCTIWUZ;
8047   case PPCISD::FCFID:
8048     return PPCISD::STRICT_FCFID;
8049   case PPCISD::FCFIDU:
8050     return PPCISD::STRICT_FCFIDU;
8051   case PPCISD::FCFIDS:
8052     return PPCISD::STRICT_FCFIDS;
8053   case PPCISD::FCFIDUS:
8054     return PPCISD::STRICT_FCFIDUS;
8055   }
8056 }
8057 
8058 static SDValue convertFPToInt(SDValue Op, SelectionDAG &DAG,
8059                               const PPCSubtarget &Subtarget) {
8060   SDLoc dl(Op);
8061   bool IsStrict = Op->isStrictFPOpcode();
8062   bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT ||
8063                   Op.getOpcode() == ISD::STRICT_FP_TO_SINT;
8064 
8065   // TODO: Any other flags to propagate?
8066   SDNodeFlags Flags;
8067   Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept());
8068 
8069   // For strict nodes, source is the second operand.
8070   SDValue Src = Op.getOperand(IsStrict ? 1 : 0);
8071   SDValue Chain = IsStrict ? Op.getOperand(0) : SDValue();
8072   assert(Src.getValueType().isFloatingPoint());
8073   if (Src.getValueType() == MVT::f32) {
8074     if (IsStrict) {
8075       Src =
8076           DAG.getNode(ISD::STRICT_FP_EXTEND, dl,
8077                       DAG.getVTList(MVT::f64, MVT::Other), {Chain, Src}, Flags);
8078       Chain = Src.getValue(1);
8079     } else
8080       Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
8081   }
8082   SDValue Conv;
8083   unsigned Opc = ISD::DELETED_NODE;
8084   switch (Op.getSimpleValueType().SimpleTy) {
8085   default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
8086   case MVT::i32:
8087     Opc = IsSigned ? PPCISD::FCTIWZ
8088                    : (Subtarget.hasFPCVT() ? PPCISD::FCTIWUZ : PPCISD::FCTIDZ);
8089     break;
8090   case MVT::i64:
8091     assert((IsSigned || Subtarget.hasFPCVT()) &&
8092            "i64 FP_TO_UINT is supported only with FPCVT");
8093     Opc = IsSigned ? PPCISD::FCTIDZ : PPCISD::FCTIDUZ;
8094   }
8095   if (IsStrict) {
8096     Opc = getPPCStrictOpcode(Opc);
8097     Conv = DAG.getNode(Opc, dl, DAG.getVTList(MVT::f64, MVT::Other),
8098                        {Chain, Src}, Flags);
8099   } else {
8100     Conv = DAG.getNode(Opc, dl, MVT::f64, Src);
8101   }
8102   return Conv;
8103 }
8104 
8105 void PPCTargetLowering::LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
8106                                                SelectionDAG &DAG,
8107                                                const SDLoc &dl) const {
8108   SDValue Tmp = convertFPToInt(Op, DAG, Subtarget);
8109   bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT ||
8110                   Op.getOpcode() == ISD::STRICT_FP_TO_SINT;
8111   bool IsStrict = Op->isStrictFPOpcode();
8112 
8113   // Convert the FP value to an int value through memory.
8114   bool i32Stack = Op.getValueType() == MVT::i32 && Subtarget.hasSTFIWX() &&
8115                   (IsSigned || Subtarget.hasFPCVT());
8116   SDValue FIPtr = DAG.CreateStackTemporary(i32Stack ? MVT::i32 : MVT::f64);
8117   int FI = cast<FrameIndexSDNode>(FIPtr)->getIndex();
8118   MachinePointerInfo MPI =
8119       MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);
8120 
8121   // Emit a store to the stack slot.
8122   SDValue Chain = IsStrict ? Tmp.getValue(1) : DAG.getEntryNode();
8123   Align Alignment(DAG.getEVTAlign(Tmp.getValueType()));
8124   if (i32Stack) {
8125     MachineFunction &MF = DAG.getMachineFunction();
8126     Alignment = Align(4);
8127     MachineMemOperand *MMO =
8128         MF.getMachineMemOperand(MPI, MachineMemOperand::MOStore, 4, Alignment);
8129     SDValue Ops[] = { Chain, Tmp, FIPtr };
8130     Chain = DAG.getMemIntrinsicNode(PPCISD::STFIWX, dl,
8131               DAG.getVTList(MVT::Other), Ops, MVT::i32, MMO);
8132   } else
8133     Chain = DAG.getStore(Chain, dl, Tmp, FIPtr, MPI, Alignment);
8134 
8135   // Result is a load from the stack slot.  If loading 4 bytes, make sure to
8136   // add in a bias on big endian.
8137   if (Op.getValueType() == MVT::i32 && !i32Stack) {
8138     FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
8139                         DAG.getConstant(4, dl, FIPtr.getValueType()));
8140     MPI = MPI.getWithOffset(Subtarget.isLittleEndian() ? 0 : 4);
8141   }
8142 
8143   RLI.Chain = Chain;
8144   RLI.Ptr = FIPtr;
8145   RLI.MPI = MPI;
8146   RLI.Alignment = Alignment;
8147 }
8148 
8149 /// Custom lowers floating point to integer conversions to use
8150 /// the direct move instructions available in ISA 2.07 to avoid the
8151 /// need for load/store combinations.
8152 SDValue PPCTargetLowering::LowerFP_TO_INTDirectMove(SDValue Op,
8153                                                     SelectionDAG &DAG,
8154                                                     const SDLoc &dl) const {
8155   SDValue Conv = convertFPToInt(Op, DAG, Subtarget);
8156   SDValue Mov = DAG.getNode(PPCISD::MFVSR, dl, Op.getValueType(), Conv);
8157   if (Op->isStrictFPOpcode())
8158     return DAG.getMergeValues({Mov, Conv.getValue(1)}, dl);
8159   else
8160     return Mov;
8161 }
8162 
8163 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
8164                                           const SDLoc &dl) const {
8165   bool IsStrict = Op->isStrictFPOpcode();
8166   bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT ||
8167                   Op.getOpcode() == ISD::STRICT_FP_TO_SINT;
8168   SDValue Src = Op.getOperand(IsStrict ? 1 : 0);
8169   EVT SrcVT = Src.getValueType();
8170   EVT DstVT = Op.getValueType();
8171 
8172   // FP to INT conversions are legal for f128.
8173   if (SrcVT == MVT::f128)
8174     return Subtarget.hasP9Vector() ? Op : SDValue();
8175 
8176   // Expand ppcf128 to i32 by hand for the benefit of llvm-gcc bootstrap on
8177   // PPC (the libcall is not available).
8178   if (SrcVT == MVT::ppcf128) {
8179     if (DstVT == MVT::i32) {
8180       // TODO: Conservatively pass only nofpexcept flag here. Need to check and
8181       // set other fast-math flags to FP operations in both strict and
8182       // non-strict cases. (FP_TO_SINT, FSUB)
8183       SDNodeFlags Flags;
8184       Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept());
8185 
8186       if (IsSigned) {
8187         SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::f64, Src,
8188                                  DAG.getIntPtrConstant(0, dl));
8189         SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::f64, Src,
8190                                  DAG.getIntPtrConstant(1, dl));
8191 
8192         // Add the two halves of the long double in round-to-zero mode, and use
8193         // a smaller FP_TO_SINT.
8194         if (IsStrict) {
8195           SDValue Res = DAG.getNode(PPCISD::STRICT_FADDRTZ, dl,
8196                                     DAG.getVTList(MVT::f64, MVT::Other),
8197                                     {Op.getOperand(0), Lo, Hi}, Flags);
8198           return DAG.getNode(ISD::STRICT_FP_TO_SINT, dl,
8199                              DAG.getVTList(MVT::i32, MVT::Other),
8200                              {Res.getValue(1), Res}, Flags);
8201         } else {
8202           SDValue Res = DAG.getNode(PPCISD::FADDRTZ, dl, MVT::f64, Lo, Hi);
8203           return DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Res);
8204         }
8205       } else {
8206         const uint64_t TwoE31[] = {0x41e0000000000000LL, 0};
8207         APFloat APF = APFloat(APFloat::PPCDoubleDouble(), APInt(128, TwoE31));
8208         SDValue Cst = DAG.getConstantFP(APF, dl, SrcVT);
8209         SDValue SignMask = DAG.getConstant(0x80000000, dl, DstVT);
8210         if (IsStrict) {
8211           // Sel = Src < 0x80000000
8212           // FltOfs = select Sel, 0.0, 0x80000000
8213           // IntOfs = select Sel, 0, 0x80000000
8214           // Result = fp_to_sint(Src - FltOfs) ^ IntOfs
8215           SDValue Chain = Op.getOperand(0);
8216           EVT SetCCVT =
8217               getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), SrcVT);
8218           EVT DstSetCCVT =
8219               getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), DstVT);
8220           SDValue Sel = DAG.getSetCC(dl, SetCCVT, Src, Cst, ISD::SETLT,
8221                                      Chain, true);
8222           Chain = Sel.getValue(1);
8223 
8224           SDValue FltOfs = DAG.getSelect(
8225               dl, SrcVT, Sel, DAG.getConstantFP(0.0, dl, SrcVT), Cst);
8226           Sel = DAG.getBoolExtOrTrunc(Sel, dl, DstSetCCVT, DstVT);
8227 
8228           SDValue Val = DAG.getNode(ISD::STRICT_FSUB, dl,
8229                                     DAG.getVTList(SrcVT, MVT::Other),
8230                                     {Chain, Src, FltOfs}, Flags);
8231           Chain = Val.getValue(1);
8232           SDValue SInt = DAG.getNode(ISD::STRICT_FP_TO_SINT, dl,
8233                                      DAG.getVTList(DstVT, MVT::Other),
8234                                      {Chain, Val}, Flags);
8235           Chain = SInt.getValue(1);
8236           SDValue IntOfs = DAG.getSelect(
8237               dl, DstVT, Sel, DAG.getConstant(0, dl, DstVT), SignMask);
8238           SDValue Result = DAG.getNode(ISD::XOR, dl, DstVT, SInt, IntOfs);
8239           return DAG.getMergeValues({Result, Chain}, dl);
8240         } else {
8241           // X>=2^31 ? (int)(X-2^31)+0x80000000 : (int)X
8242           // FIXME: generated code sucks.
8243           SDValue True = DAG.getNode(ISD::FSUB, dl, MVT::ppcf128, Src, Cst);
8244           True = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, True);
8245           True = DAG.getNode(ISD::ADD, dl, MVT::i32, True, SignMask);
8246           SDValue False = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Src);
8247           return DAG.getSelectCC(dl, Src, Cst, True, False, ISD::SETGE);
8248         }
8249       }
8250     }
8251 
8252     return SDValue();
8253   }
8254 
8255   if (Subtarget.hasDirectMove() && Subtarget.isPPC64())
8256     return LowerFP_TO_INTDirectMove(Op, DAG, dl);
8257 
8258   ReuseLoadInfo RLI;
8259   LowerFP_TO_INTForReuse(Op, RLI, DAG, dl);
8260 
8261   return DAG.getLoad(Op.getValueType(), dl, RLI.Chain, RLI.Ptr, RLI.MPI,
8262                      RLI.Alignment, RLI.MMOFlags(), RLI.AAInfo, RLI.Ranges);
8263 }
8264 
8265 // We're trying to insert a regular store, S, and then a load, L. If the
8266 // incoming value, O, is a load, we might just be able to have our load use the
8267 // address used by O. However, we don't know if anything else will store to
8268 // that address before we can load from it. To prevent this situation, we need
8269 // to insert our load, L, into the chain as a peer of O. To do this, we give L
8270 // the same chain operand as O, we create a token factor from the chain results
8271 // of O and L, and we replace all uses of O's chain result with that token
8272 // factor (see spliceIntoChain below for this last part).
8273 bool PPCTargetLowering::canReuseLoadAddress(SDValue Op, EVT MemVT,
8274                                             ReuseLoadInfo &RLI,
8275                                             SelectionDAG &DAG,
8276                                             ISD::LoadExtType ET) const {
8277   // Conservatively skip reusing for constrained FP nodes.
8278   if (Op->isStrictFPOpcode())
8279     return false;
8280 
8281   SDLoc dl(Op);
8282   bool ValidFPToUint = Op.getOpcode() == ISD::FP_TO_UINT &&
8283                        (Subtarget.hasFPCVT() || Op.getValueType() == MVT::i32);
8284   if (ET == ISD::NON_EXTLOAD &&
8285       (ValidFPToUint || Op.getOpcode() == ISD::FP_TO_SINT) &&
8286       isOperationLegalOrCustom(Op.getOpcode(),
8287                                Op.getOperand(0).getValueType())) {
8288 
8289     LowerFP_TO_INTForReuse(Op, RLI, DAG, dl);
8290     return true;
8291   }
8292 
8293   LoadSDNode *LD = dyn_cast<LoadSDNode>(Op);
8294   if (!LD || LD->getExtensionType() != ET || LD->isVolatile() ||
8295       LD->isNonTemporal())
8296     return false;
8297   if (LD->getMemoryVT() != MemVT)
8298     return false;
8299 
8300   // If the result of the load is an illegal type, then we can't build a
8301   // valid chain for reuse since the legalised loads and token factor node that
8302   // ties the legalised loads together uses a different output chain then the
8303   // illegal load.
8304   if (!isTypeLegal(LD->getValueType(0)))
8305     return false;
8306 
8307   RLI.Ptr = LD->getBasePtr();
8308   if (LD->isIndexed() && !LD->getOffset().isUndef()) {
8309     assert(LD->getAddressingMode() == ISD::PRE_INC &&
8310            "Non-pre-inc AM on PPC?");
8311     RLI.Ptr = DAG.getNode(ISD::ADD, dl, RLI.Ptr.getValueType(), RLI.Ptr,
8312                           LD->getOffset());
8313   }
8314 
8315   RLI.Chain = LD->getChain();
8316   RLI.MPI = LD->getPointerInfo();
8317   RLI.IsDereferenceable = LD->isDereferenceable();
8318   RLI.IsInvariant = LD->isInvariant();
8319   RLI.Alignment = LD->getAlign();
8320   RLI.AAInfo = LD->getAAInfo();
8321   RLI.Ranges = LD->getRanges();
8322 
8323   RLI.ResChain = SDValue(LD, LD->isIndexed() ? 2 : 1);
8324   return true;
8325 }
8326 
8327 // Given the head of the old chain, ResChain, insert a token factor containing
8328 // it and NewResChain, and make users of ResChain now be users of that token
8329 // factor.
8330 // TODO: Remove and use DAG::makeEquivalentMemoryOrdering() instead.
8331 void PPCTargetLowering::spliceIntoChain(SDValue ResChain,
8332                                         SDValue NewResChain,
8333                                         SelectionDAG &DAG) const {
8334   if (!ResChain)
8335     return;
8336 
8337   SDLoc dl(NewResChain);
8338 
8339   SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
8340                            NewResChain, DAG.getUNDEF(MVT::Other));
8341   assert(TF.getNode() != NewResChain.getNode() &&
8342          "A new TF really is required here");
8343 
8344   DAG.ReplaceAllUsesOfValueWith(ResChain, TF);
8345   DAG.UpdateNodeOperands(TF.getNode(), ResChain, NewResChain);
8346 }
8347 
8348 /// Analyze profitability of direct move
8349 /// prefer float load to int load plus direct move
8350 /// when there is no integer use of int load
8351 bool PPCTargetLowering::directMoveIsProfitable(const SDValue &Op) const {
8352   SDNode *Origin = Op.getOperand(Op->isStrictFPOpcode() ? 1 : 0).getNode();
8353   if (Origin->getOpcode() != ISD::LOAD)
8354     return true;
8355 
8356   // If there is no LXSIBZX/LXSIHZX, like Power8,
8357   // prefer direct move if the memory size is 1 or 2 bytes.
8358   MachineMemOperand *MMO = cast<LoadSDNode>(Origin)->getMemOperand();
8359   if (!Subtarget.hasP9Vector() && MMO->getSize() <= 2)
8360     return true;
8361 
8362   for (SDNode::use_iterator UI = Origin->use_begin(),
8363                             UE = Origin->use_end();
8364        UI != UE; ++UI) {
8365 
8366     // Only look at the users of the loaded value.
8367     if (UI.getUse().get().getResNo() != 0)
8368       continue;
8369 
8370     if (UI->getOpcode() != ISD::SINT_TO_FP &&
8371         UI->getOpcode() != ISD::UINT_TO_FP &&
8372         UI->getOpcode() != ISD::STRICT_SINT_TO_FP &&
8373         UI->getOpcode() != ISD::STRICT_UINT_TO_FP)
8374       return true;
8375   }
8376 
8377   return false;
8378 }
8379 
8380 static SDValue convertIntToFP(SDValue Op, SDValue Src, SelectionDAG &DAG,
8381                               const PPCSubtarget &Subtarget,
8382                               SDValue Chain = SDValue()) {
8383   bool IsSigned = Op.getOpcode() == ISD::SINT_TO_FP ||
8384                   Op.getOpcode() == ISD::STRICT_SINT_TO_FP;
8385   SDLoc dl(Op);
8386 
8387   // TODO: Any other flags to propagate?
8388   SDNodeFlags Flags;
8389   Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept());
8390 
8391   // If we have FCFIDS, then use it when converting to single-precision.
8392   // Otherwise, convert to double-precision and then round.
8393   bool IsSingle = Op.getValueType() == MVT::f32 && Subtarget.hasFPCVT();
8394   unsigned ConvOpc = IsSingle ? (IsSigned ? PPCISD::FCFIDS : PPCISD::FCFIDUS)
8395                               : (IsSigned ? PPCISD::FCFID : PPCISD::FCFIDU);
8396   EVT ConvTy = IsSingle ? MVT::f32 : MVT::f64;
8397   if (Op->isStrictFPOpcode()) {
8398     if (!Chain)
8399       Chain = Op.getOperand(0);
8400     return DAG.getNode(getPPCStrictOpcode(ConvOpc), dl,
8401                        DAG.getVTList(ConvTy, MVT::Other), {Chain, Src}, Flags);
8402   } else
8403     return DAG.getNode(ConvOpc, dl, ConvTy, Src);
8404 }
8405 
8406 /// Custom lowers integer to floating point conversions to use
8407 /// the direct move instructions available in ISA 2.07 to avoid the
8408 /// need for load/store combinations.
8409 SDValue PPCTargetLowering::LowerINT_TO_FPDirectMove(SDValue Op,
8410                                                     SelectionDAG &DAG,
8411                                                     const SDLoc &dl) const {
8412   assert((Op.getValueType() == MVT::f32 ||
8413           Op.getValueType() == MVT::f64) &&
8414          "Invalid floating point type as target of conversion");
8415   assert(Subtarget.hasFPCVT() &&
8416          "Int to FP conversions with direct moves require FPCVT");
8417   SDValue Src = Op.getOperand(Op->isStrictFPOpcode() ? 1 : 0);
8418   bool WordInt = Src.getSimpleValueType().SimpleTy == MVT::i32;
8419   bool Signed = Op.getOpcode() == ISD::SINT_TO_FP ||
8420                 Op.getOpcode() == ISD::STRICT_SINT_TO_FP;
8421   unsigned MovOpc = (WordInt && !Signed) ? PPCISD::MTVSRZ : PPCISD::MTVSRA;
8422   SDValue Mov = DAG.getNode(MovOpc, dl, MVT::f64, Src);
8423   return convertIntToFP(Op, Mov, DAG, Subtarget);
8424 }
8425 
8426 static SDValue widenVec(SelectionDAG &DAG, SDValue Vec, const SDLoc &dl) {
8427 
8428   EVT VecVT = Vec.getValueType();
8429   assert(VecVT.isVector() && "Expected a vector type.");
8430   assert(VecVT.getSizeInBits() < 128 && "Vector is already full width.");
8431 
8432   EVT EltVT = VecVT.getVectorElementType();
8433   unsigned WideNumElts = 128 / EltVT.getSizeInBits();
8434   EVT WideVT = EVT::getVectorVT(*DAG.getContext(), EltVT, WideNumElts);
8435 
8436   unsigned NumConcat = WideNumElts / VecVT.getVectorNumElements();
8437   SmallVector<SDValue, 16> Ops(NumConcat);
8438   Ops[0] = Vec;
8439   SDValue UndefVec = DAG.getUNDEF(VecVT);
8440   for (unsigned i = 1; i < NumConcat; ++i)
8441     Ops[i] = UndefVec;
8442 
8443   return DAG.getNode(ISD::CONCAT_VECTORS, dl, WideVT, Ops);
8444 }
8445 
8446 SDValue PPCTargetLowering::LowerINT_TO_FPVector(SDValue Op, SelectionDAG &DAG,
8447                                                 const SDLoc &dl) const {
8448   bool IsStrict = Op->isStrictFPOpcode();
8449   unsigned Opc = Op.getOpcode();
8450   SDValue Src = Op.getOperand(IsStrict ? 1 : 0);
8451   assert((Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP ||
8452           Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) &&
8453          "Unexpected conversion type");
8454   assert((Op.getValueType() == MVT::v2f64 || Op.getValueType() == MVT::v4f32) &&
8455          "Supports conversions to v2f64/v4f32 only.");
8456 
8457   // TODO: Any other flags to propagate?
8458   SDNodeFlags Flags;
8459   Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept());
8460 
8461   bool SignedConv = Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP;
8462   bool FourEltRes = Op.getValueType() == MVT::v4f32;
8463 
8464   SDValue Wide = widenVec(DAG, Src, dl);
8465   EVT WideVT = Wide.getValueType();
8466   unsigned WideNumElts = WideVT.getVectorNumElements();
8467   MVT IntermediateVT = FourEltRes ? MVT::v4i32 : MVT::v2i64;
8468 
8469   SmallVector<int, 16> ShuffV;
8470   for (unsigned i = 0; i < WideNumElts; ++i)
8471     ShuffV.push_back(i + WideNumElts);
8472 
8473   int Stride = FourEltRes ? WideNumElts / 4 : WideNumElts / 2;
8474   int SaveElts = FourEltRes ? 4 : 2;
8475   if (Subtarget.isLittleEndian())
8476     for (int i = 0; i < SaveElts; i++)
8477       ShuffV[i * Stride] = i;
8478   else
8479     for (int i = 1; i <= SaveElts; i++)
8480       ShuffV[i * Stride - 1] = i - 1;
8481 
8482   SDValue ShuffleSrc2 =
8483       SignedConv ? DAG.getUNDEF(WideVT) : DAG.getConstant(0, dl, WideVT);
8484   SDValue Arrange = DAG.getVectorShuffle(WideVT, dl, Wide, ShuffleSrc2, ShuffV);
8485 
8486   SDValue Extend;
8487   if (SignedConv) {
8488     Arrange = DAG.getBitcast(IntermediateVT, Arrange);
8489     EVT ExtVT = Src.getValueType();
8490     if (Subtarget.hasP9Altivec())
8491       ExtVT = EVT::getVectorVT(*DAG.getContext(), WideVT.getVectorElementType(),
8492                                IntermediateVT.getVectorNumElements());
8493 
8494     Extend = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, IntermediateVT, Arrange,
8495                          DAG.getValueType(ExtVT));
8496   } else
8497     Extend = DAG.getNode(ISD::BITCAST, dl, IntermediateVT, Arrange);
8498 
8499   if (IsStrict)
8500     return DAG.getNode(Opc, dl, DAG.getVTList(Op.getValueType(), MVT::Other),
8501                        {Op.getOperand(0), Extend}, Flags);
8502 
8503   return DAG.getNode(Opc, dl, Op.getValueType(), Extend);
8504 }
8505 
8506 SDValue PPCTargetLowering::LowerINT_TO_FP(SDValue Op,
8507                                           SelectionDAG &DAG) const {
8508   SDLoc dl(Op);
8509   bool IsSigned = Op.getOpcode() == ISD::SINT_TO_FP ||
8510                   Op.getOpcode() == ISD::STRICT_SINT_TO_FP;
8511   bool IsStrict = Op->isStrictFPOpcode();
8512   SDValue Src = Op.getOperand(IsStrict ? 1 : 0);
8513   SDValue Chain = IsStrict ? Op.getOperand(0) : DAG.getEntryNode();
8514 
8515   // TODO: Any other flags to propagate?
8516   SDNodeFlags Flags;
8517   Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept());
8518 
8519   EVT InVT = Src.getValueType();
8520   EVT OutVT = Op.getValueType();
8521   if (OutVT.isVector() && OutVT.isFloatingPoint() &&
8522       isOperationCustom(Op.getOpcode(), InVT))
8523     return LowerINT_TO_FPVector(Op, DAG, dl);
8524 
8525   // Conversions to f128 are legal.
8526   if (Op.getValueType() == MVT::f128)
8527     return Subtarget.hasP9Vector() ? Op : SDValue();
8528 
8529   // Don't handle ppc_fp128 here; let it be lowered to a libcall.
8530   if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
8531     return SDValue();
8532 
8533   if (Src.getValueType() == MVT::i1) {
8534     SDValue Sel = DAG.getNode(ISD::SELECT, dl, Op.getValueType(), Src,
8535                               DAG.getConstantFP(1.0, dl, Op.getValueType()),
8536                               DAG.getConstantFP(0.0, dl, Op.getValueType()));
8537     if (IsStrict)
8538       return DAG.getMergeValues({Sel, Chain}, dl);
8539     else
8540       return Sel;
8541   }
8542 
8543   // If we have direct moves, we can do all the conversion, skip the store/load
8544   // however, without FPCVT we can't do most conversions.
8545   if (Subtarget.hasDirectMove() && directMoveIsProfitable(Op) &&
8546       Subtarget.isPPC64() && Subtarget.hasFPCVT())
8547     return LowerINT_TO_FPDirectMove(Op, DAG, dl);
8548 
8549   assert((IsSigned || Subtarget.hasFPCVT()) &&
8550          "UINT_TO_FP is supported only with FPCVT");
8551 
8552   if (Src.getValueType() == MVT::i64) {
8553     SDValue SINT = Src;
8554     // When converting to single-precision, we actually need to convert
8555     // to double-precision first and then round to single-precision.
8556     // To avoid double-rounding effects during that operation, we have
8557     // to prepare the input operand.  Bits that might be truncated when
8558     // converting to double-precision are replaced by a bit that won't
8559     // be lost at this stage, but is below the single-precision rounding
8560     // position.
8561     //
8562     // However, if -enable-unsafe-fp-math is in effect, accept double
8563     // rounding to avoid the extra overhead.
8564     if (Op.getValueType() == MVT::f32 &&
8565         !Subtarget.hasFPCVT() &&
8566         !DAG.getTarget().Options.UnsafeFPMath) {
8567 
8568       // Twiddle input to make sure the low 11 bits are zero.  (If this
8569       // is the case, we are guaranteed the value will fit into the 53 bit
8570       // mantissa of an IEEE double-precision value without rounding.)
8571       // If any of those low 11 bits were not zero originally, make sure
8572       // bit 12 (value 2048) is set instead, so that the final rounding
8573       // to single-precision gets the correct result.
8574       SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
8575                                   SINT, DAG.getConstant(2047, dl, MVT::i64));
8576       Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
8577                           Round, DAG.getConstant(2047, dl, MVT::i64));
8578       Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
8579       Round = DAG.getNode(ISD::AND, dl, MVT::i64,
8580                           Round, DAG.getConstant(-2048, dl, MVT::i64));
8581 
8582       // However, we cannot use that value unconditionally: if the magnitude
8583       // of the input value is small, the bit-twiddling we did above might
8584       // end up visibly changing the output.  Fortunately, in that case, we
8585       // don't need to twiddle bits since the original input will convert
8586       // exactly to double-precision floating-point already.  Therefore,
8587       // construct a conditional to use the original value if the top 11
8588       // bits are all sign-bit copies, and use the rounded value computed
8589       // above otherwise.
8590       SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
8591                                  SINT, DAG.getConstant(53, dl, MVT::i32));
8592       Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
8593                          Cond, DAG.getConstant(1, dl, MVT::i64));
8594       Cond = DAG.getSetCC(
8595           dl,
8596           getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
8597           Cond, DAG.getConstant(1, dl, MVT::i64), ISD::SETUGT);
8598 
8599       SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
8600     }
8601 
8602     ReuseLoadInfo RLI;
8603     SDValue Bits;
8604 
8605     MachineFunction &MF = DAG.getMachineFunction();
8606     if (canReuseLoadAddress(SINT, MVT::i64, RLI, DAG)) {
8607       Bits = DAG.getLoad(MVT::f64, dl, RLI.Chain, RLI.Ptr, RLI.MPI,
8608                          RLI.Alignment, RLI.MMOFlags(), RLI.AAInfo, RLI.Ranges);
8609       spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG);
8610     } else if (Subtarget.hasLFIWAX() &&
8611                canReuseLoadAddress(SINT, MVT::i32, RLI, DAG, ISD::SEXTLOAD)) {
8612       MachineMemOperand *MMO =
8613         MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
8614                                 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
8615       SDValue Ops[] = { RLI.Chain, RLI.Ptr };
8616       Bits = DAG.getMemIntrinsicNode(PPCISD::LFIWAX, dl,
8617                                      DAG.getVTList(MVT::f64, MVT::Other),
8618                                      Ops, MVT::i32, MMO);
8619       spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG);
8620     } else if (Subtarget.hasFPCVT() &&
8621                canReuseLoadAddress(SINT, MVT::i32, RLI, DAG, ISD::ZEXTLOAD)) {
8622       MachineMemOperand *MMO =
8623         MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
8624                                 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
8625       SDValue Ops[] = { RLI.Chain, RLI.Ptr };
8626       Bits = DAG.getMemIntrinsicNode(PPCISD::LFIWZX, dl,
8627                                      DAG.getVTList(MVT::f64, MVT::Other),
8628                                      Ops, MVT::i32, MMO);
8629       spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG);
8630     } else if (((Subtarget.hasLFIWAX() &&
8631                  SINT.getOpcode() == ISD::SIGN_EXTEND) ||
8632                 (Subtarget.hasFPCVT() &&
8633                  SINT.getOpcode() == ISD::ZERO_EXTEND)) &&
8634                SINT.getOperand(0).getValueType() == MVT::i32) {
8635       MachineFrameInfo &MFI = MF.getFrameInfo();
8636       EVT PtrVT = getPointerTy(DAG.getDataLayout());
8637 
8638       int FrameIdx = MFI.CreateStackObject(4, Align(4), false);
8639       SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
8640 
8641       SDValue Store = DAG.getStore(Chain, dl, SINT.getOperand(0), FIdx,
8642                                    MachinePointerInfo::getFixedStack(
8643                                        DAG.getMachineFunction(), FrameIdx));
8644       Chain = Store;
8645 
8646       assert(cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 &&
8647              "Expected an i32 store");
8648 
8649       RLI.Ptr = FIdx;
8650       RLI.Chain = Chain;
8651       RLI.MPI =
8652           MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx);
8653       RLI.Alignment = Align(4);
8654 
8655       MachineMemOperand *MMO =
8656         MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
8657                                 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
8658       SDValue Ops[] = { RLI.Chain, RLI.Ptr };
8659       Bits = DAG.getMemIntrinsicNode(SINT.getOpcode() == ISD::ZERO_EXTEND ?
8660                                      PPCISD::LFIWZX : PPCISD::LFIWAX,
8661                                      dl, DAG.getVTList(MVT::f64, MVT::Other),
8662                                      Ops, MVT::i32, MMO);
8663       Chain = Bits.getValue(1);
8664     } else
8665       Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
8666 
8667     SDValue FP = convertIntToFP(Op, Bits, DAG, Subtarget, Chain);
8668     if (IsStrict)
8669       Chain = FP.getValue(1);
8670 
8671     if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT()) {
8672       if (IsStrict)
8673         FP = DAG.getNode(ISD::STRICT_FP_ROUND, dl,
8674                          DAG.getVTList(MVT::f32, MVT::Other),
8675                          {Chain, FP, DAG.getIntPtrConstant(0, dl)}, Flags);
8676       else
8677         FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP,
8678                          DAG.getIntPtrConstant(0, dl, /*isTarget=*/true));
8679     }
8680     return FP;
8681   }
8682 
8683   assert(Src.getValueType() == MVT::i32 &&
8684          "Unhandled INT_TO_FP type in custom expander!");
8685   // Since we only generate this in 64-bit mode, we can take advantage of
8686   // 64-bit registers.  In particular, sign extend the input value into the
8687   // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
8688   // then lfd it and fcfid it.
8689   MachineFunction &MF = DAG.getMachineFunction();
8690   MachineFrameInfo &MFI = MF.getFrameInfo();
8691   EVT PtrVT = getPointerTy(MF.getDataLayout());
8692 
8693   SDValue Ld;
8694   if (Subtarget.hasLFIWAX() || Subtarget.hasFPCVT()) {
8695     ReuseLoadInfo RLI;
8696     bool ReusingLoad;
8697     if (!(ReusingLoad = canReuseLoadAddress(Src, MVT::i32, RLI, DAG))) {
8698       int FrameIdx = MFI.CreateStackObject(4, Align(4), false);
8699       SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
8700 
8701       SDValue Store = DAG.getStore(Chain, dl, Src, FIdx,
8702                                    MachinePointerInfo::getFixedStack(
8703                                        DAG.getMachineFunction(), FrameIdx));
8704       Chain = Store;
8705 
8706       assert(cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 &&
8707              "Expected an i32 store");
8708 
8709       RLI.Ptr = FIdx;
8710       RLI.Chain = Chain;
8711       RLI.MPI =
8712           MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx);
8713       RLI.Alignment = Align(4);
8714     }
8715 
8716     MachineMemOperand *MMO =
8717       MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
8718                               RLI.Alignment, RLI.AAInfo, RLI.Ranges);
8719     SDValue Ops[] = { RLI.Chain, RLI.Ptr };
8720     Ld = DAG.getMemIntrinsicNode(IsSigned ? PPCISD::LFIWAX : PPCISD::LFIWZX, dl,
8721                                  DAG.getVTList(MVT::f64, MVT::Other), Ops,
8722                                  MVT::i32, MMO);
8723     Chain = Ld.getValue(1);
8724     if (ReusingLoad)
8725       spliceIntoChain(RLI.ResChain, Ld.getValue(1), DAG);
8726   } else {
8727     assert(Subtarget.isPPC64() &&
8728            "i32->FP without LFIWAX supported only on PPC64");
8729 
8730     int FrameIdx = MFI.CreateStackObject(8, Align(8), false);
8731     SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
8732 
8733     SDValue Ext64 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i64, Src);
8734 
8735     // STD the extended value into the stack slot.
8736     SDValue Store = DAG.getStore(
8737         Chain, dl, Ext64, FIdx,
8738         MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx));
8739     Chain = Store;
8740 
8741     // Load the value as a double.
8742     Ld = DAG.getLoad(
8743         MVT::f64, dl, Chain, FIdx,
8744         MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx));
8745     Chain = Ld.getValue(1);
8746   }
8747 
8748   // FCFID it and return it.
8749   SDValue FP = convertIntToFP(Op, Ld, DAG, Subtarget, Chain);
8750   if (IsStrict)
8751     Chain = FP.getValue(1);
8752   if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT()) {
8753     if (IsStrict)
8754       FP = DAG.getNode(ISD::STRICT_FP_ROUND, dl,
8755                        DAG.getVTList(MVT::f32, MVT::Other),
8756                        {Chain, FP, DAG.getIntPtrConstant(0, dl)}, Flags);
8757     else
8758       FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP,
8759                        DAG.getIntPtrConstant(0, dl, /*isTarget=*/true));
8760   }
8761   return FP;
8762 }
8763 
8764 SDValue PPCTargetLowering::LowerGET_ROUNDING(SDValue Op,
8765                                              SelectionDAG &DAG) const {
8766   SDLoc dl(Op);
8767   /*
8768    The rounding mode is in bits 30:31 of FPSR, and has the following
8769    settings:
8770      00 Round to nearest
8771      01 Round to 0
8772      10 Round to +inf
8773      11 Round to -inf
8774 
8775   GET_ROUNDING, on the other hand, expects the following:
8776     -1 Undefined
8777      0 Round to 0
8778      1 Round to nearest
8779      2 Round to +inf
8780      3 Round to -inf
8781 
8782   To perform the conversion, we do:
8783     ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
8784   */
8785 
8786   MachineFunction &MF = DAG.getMachineFunction();
8787   EVT VT = Op.getValueType();
8788   EVT PtrVT = getPointerTy(MF.getDataLayout());
8789 
8790   // Save FP Control Word to register
8791   SDValue Chain = Op.getOperand(0);
8792   SDValue MFFS = DAG.getNode(PPCISD::MFFS, dl, {MVT::f64, MVT::Other}, Chain);
8793   Chain = MFFS.getValue(1);
8794 
8795   SDValue CWD;
8796   if (isTypeLegal(MVT::i64)) {
8797     CWD = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
8798                       DAG.getNode(ISD::BITCAST, dl, MVT::i64, MFFS));
8799   } else {
8800     // Save FP register to stack slot
8801     int SSFI = MF.getFrameInfo().CreateStackObject(8, Align(8), false);
8802     SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
8803     Chain = DAG.getStore(Chain, dl, MFFS, StackSlot, MachinePointerInfo());
8804 
8805     // Load FP Control Word from low 32 bits of stack slot.
8806     assert(hasBigEndianPartOrdering(MVT::i64, MF.getDataLayout()) &&
8807            "Stack slot adjustment is valid only on big endian subtargets!");
8808     SDValue Four = DAG.getConstant(4, dl, PtrVT);
8809     SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
8810     CWD = DAG.getLoad(MVT::i32, dl, Chain, Addr, MachinePointerInfo());
8811     Chain = CWD.getValue(1);
8812   }
8813 
8814   // Transform as necessary
8815   SDValue CWD1 =
8816     DAG.getNode(ISD::AND, dl, MVT::i32,
8817                 CWD, DAG.getConstant(3, dl, MVT::i32));
8818   SDValue CWD2 =
8819     DAG.getNode(ISD::SRL, dl, MVT::i32,
8820                 DAG.getNode(ISD::AND, dl, MVT::i32,
8821                             DAG.getNode(ISD::XOR, dl, MVT::i32,
8822                                         CWD, DAG.getConstant(3, dl, MVT::i32)),
8823                             DAG.getConstant(3, dl, MVT::i32)),
8824                 DAG.getConstant(1, dl, MVT::i32));
8825 
8826   SDValue RetVal =
8827     DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
8828 
8829   RetVal =
8830       DAG.getNode((VT.getSizeInBits() < 16 ? ISD::TRUNCATE : ISD::ZERO_EXTEND),
8831                   dl, VT, RetVal);
8832 
8833   return DAG.getMergeValues({RetVal, Chain}, dl);
8834 }
8835 
8836 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
8837   EVT VT = Op.getValueType();
8838   unsigned BitWidth = VT.getSizeInBits();
8839   SDLoc dl(Op);
8840   assert(Op.getNumOperands() == 3 &&
8841          VT == Op.getOperand(1).getValueType() &&
8842          "Unexpected SHL!");
8843 
8844   // Expand into a bunch of logical ops.  Note that these ops
8845   // depend on the PPC behavior for oversized shift amounts.
8846   SDValue Lo = Op.getOperand(0);
8847   SDValue Hi = Op.getOperand(1);
8848   SDValue Amt = Op.getOperand(2);
8849   EVT AmtVT = Amt.getValueType();
8850 
8851   SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
8852                              DAG.getConstant(BitWidth, dl, AmtVT), Amt);
8853   SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
8854   SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
8855   SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
8856   SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
8857                              DAG.getConstant(-BitWidth, dl, AmtVT));
8858   SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
8859   SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
8860   SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
8861   SDValue OutOps[] = { OutLo, OutHi };
8862   return DAG.getMergeValues(OutOps, dl);
8863 }
8864 
8865 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
8866   EVT VT = Op.getValueType();
8867   SDLoc dl(Op);
8868   unsigned BitWidth = VT.getSizeInBits();
8869   assert(Op.getNumOperands() == 3 &&
8870          VT == Op.getOperand(1).getValueType() &&
8871          "Unexpected SRL!");
8872 
8873   // Expand into a bunch of logical ops.  Note that these ops
8874   // depend on the PPC behavior for oversized shift amounts.
8875   SDValue Lo = Op.getOperand(0);
8876   SDValue Hi = Op.getOperand(1);
8877   SDValue Amt = Op.getOperand(2);
8878   EVT AmtVT = Amt.getValueType();
8879 
8880   SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
8881                              DAG.getConstant(BitWidth, dl, AmtVT), Amt);
8882   SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
8883   SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
8884   SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
8885   SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
8886                              DAG.getConstant(-BitWidth, dl, AmtVT));
8887   SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
8888   SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
8889   SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
8890   SDValue OutOps[] = { OutLo, OutHi };
8891   return DAG.getMergeValues(OutOps, dl);
8892 }
8893 
8894 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
8895   SDLoc dl(Op);
8896   EVT VT = Op.getValueType();
8897   unsigned BitWidth = VT.getSizeInBits();
8898   assert(Op.getNumOperands() == 3 &&
8899          VT == Op.getOperand(1).getValueType() &&
8900          "Unexpected SRA!");
8901 
8902   // Expand into a bunch of logical ops, followed by a select_cc.
8903   SDValue Lo = Op.getOperand(0);
8904   SDValue Hi = Op.getOperand(1);
8905   SDValue Amt = Op.getOperand(2);
8906   EVT AmtVT = Amt.getValueType();
8907 
8908   SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
8909                              DAG.getConstant(BitWidth, dl, AmtVT), Amt);
8910   SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
8911   SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
8912   SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
8913   SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
8914                              DAG.getConstant(-BitWidth, dl, AmtVT));
8915   SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
8916   SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
8917   SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, dl, AmtVT),
8918                                   Tmp4, Tmp6, ISD::SETLE);
8919   SDValue OutOps[] = { OutLo, OutHi };
8920   return DAG.getMergeValues(OutOps, dl);
8921 }
8922 
8923 SDValue PPCTargetLowering::LowerFunnelShift(SDValue Op,
8924                                             SelectionDAG &DAG) const {
8925   SDLoc dl(Op);
8926   EVT VT = Op.getValueType();
8927   unsigned BitWidth = VT.getSizeInBits();
8928 
8929   bool IsFSHL = Op.getOpcode() == ISD::FSHL;
8930   SDValue X = Op.getOperand(0);
8931   SDValue Y = Op.getOperand(1);
8932   SDValue Z = Op.getOperand(2);
8933   EVT AmtVT = Z.getValueType();
8934 
8935   // fshl: (X << (Z % BW)) | (Y >> (BW - (Z % BW)))
8936   // fshr: (X << (BW - (Z % BW))) | (Y >> (Z % BW))
8937   // This is simpler than TargetLowering::expandFunnelShift because we can rely
8938   // on PowerPC shift by BW being well defined.
8939   Z = DAG.getNode(ISD::AND, dl, AmtVT, Z,
8940                   DAG.getConstant(BitWidth - 1, dl, AmtVT));
8941   SDValue SubZ =
8942       DAG.getNode(ISD::SUB, dl, AmtVT, DAG.getConstant(BitWidth, dl, AmtVT), Z);
8943   X = DAG.getNode(PPCISD::SHL, dl, VT, X, IsFSHL ? Z : SubZ);
8944   Y = DAG.getNode(PPCISD::SRL, dl, VT, Y, IsFSHL ? SubZ : Z);
8945   return DAG.getNode(ISD::OR, dl, VT, X, Y);
8946 }
8947 
8948 //===----------------------------------------------------------------------===//
8949 // Vector related lowering.
8950 //
8951 
8952 /// getCanonicalConstSplat - Build a canonical splat immediate of Val with an
8953 /// element size of SplatSize. Cast the result to VT.
8954 static SDValue getCanonicalConstSplat(uint64_t Val, unsigned SplatSize, EVT VT,
8955                                       SelectionDAG &DAG, const SDLoc &dl) {
8956   static const MVT VTys[] = { // canonical VT to use for each size.
8957     MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
8958   };
8959 
8960   EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
8961 
8962   // For a splat with all ones, turn it to vspltisb 0xFF to canonicalize.
8963   if (Val == ((1LLU << (SplatSize * 8)) - 1)) {
8964     SplatSize = 1;
8965     Val = 0xFF;
8966   }
8967 
8968   EVT CanonicalVT = VTys[SplatSize-1];
8969 
8970   // Build a canonical splat for this value.
8971   return DAG.getBitcast(ReqVT, DAG.getConstant(Val, dl, CanonicalVT));
8972 }
8973 
8974 /// BuildIntrinsicOp - Return a unary operator intrinsic node with the
8975 /// specified intrinsic ID.
8976 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op, SelectionDAG &DAG,
8977                                 const SDLoc &dl, EVT DestVT = MVT::Other) {
8978   if (DestVT == MVT::Other) DestVT = Op.getValueType();
8979   return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
8980                      DAG.getConstant(IID, dl, MVT::i32), Op);
8981 }
8982 
8983 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
8984 /// specified intrinsic ID.
8985 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
8986                                 SelectionDAG &DAG, const SDLoc &dl,
8987                                 EVT DestVT = MVT::Other) {
8988   if (DestVT == MVT::Other) DestVT = LHS.getValueType();
8989   return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
8990                      DAG.getConstant(IID, dl, MVT::i32), LHS, RHS);
8991 }
8992 
8993 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
8994 /// specified intrinsic ID.
8995 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
8996                                 SDValue Op2, SelectionDAG &DAG, const SDLoc &dl,
8997                                 EVT DestVT = MVT::Other) {
8998   if (DestVT == MVT::Other) DestVT = Op0.getValueType();
8999   return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
9000                      DAG.getConstant(IID, dl, MVT::i32), Op0, Op1, Op2);
9001 }
9002 
9003 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
9004 /// amount.  The result has the specified value type.
9005 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt, EVT VT,
9006                            SelectionDAG &DAG, const SDLoc &dl) {
9007   // Force LHS/RHS to be the right type.
9008   LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
9009   RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
9010 
9011   int Ops[16];
9012   for (unsigned i = 0; i != 16; ++i)
9013     Ops[i] = i + Amt;
9014   SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
9015   return DAG.getNode(ISD::BITCAST, dl, VT, T);
9016 }
9017 
9018 /// Do we have an efficient pattern in a .td file for this node?
9019 ///
9020 /// \param V - pointer to the BuildVectorSDNode being matched
9021 /// \param HasDirectMove - does this subtarget have VSR <-> GPR direct moves?
9022 ///
9023 /// There are some patterns where it is beneficial to keep a BUILD_VECTOR
9024 /// node as a BUILD_VECTOR node rather than expanding it. The patterns where
9025 /// the opposite is true (expansion is beneficial) are:
9026 /// - The node builds a vector out of integers that are not 32 or 64-bits
9027 /// - The node builds a vector out of constants
9028 /// - The node is a "load-and-splat"
9029 /// In all other cases, we will choose to keep the BUILD_VECTOR.
9030 static bool haveEfficientBuildVectorPattern(BuildVectorSDNode *V,
9031                                             bool HasDirectMove,
9032                                             bool HasP8Vector) {
9033   EVT VecVT = V->getValueType(0);
9034   bool RightType = VecVT == MVT::v2f64 ||
9035     (HasP8Vector && VecVT == MVT::v4f32) ||
9036     (HasDirectMove && (VecVT == MVT::v2i64 || VecVT == MVT::v4i32));
9037   if (!RightType)
9038     return false;
9039 
9040   bool IsSplat = true;
9041   bool IsLoad = false;
9042   SDValue Op0 = V->getOperand(0);
9043 
9044   // This function is called in a block that confirms the node is not a constant
9045   // splat. So a constant BUILD_VECTOR here means the vector is built out of
9046   // different constants.
9047   if (V->isConstant())
9048     return false;
9049   for (int i = 0, e = V->getNumOperands(); i < e; ++i) {
9050     if (V->getOperand(i).isUndef())
9051       return false;
9052     // We want to expand nodes that represent load-and-splat even if the
9053     // loaded value is a floating point truncation or conversion to int.
9054     if (V->getOperand(i).getOpcode() == ISD::LOAD ||
9055         (V->getOperand(i).getOpcode() == ISD::FP_ROUND &&
9056          V->getOperand(i).getOperand(0).getOpcode() == ISD::LOAD) ||
9057         (V->getOperand(i).getOpcode() == ISD::FP_TO_SINT &&
9058          V->getOperand(i).getOperand(0).getOpcode() == ISD::LOAD) ||
9059         (V->getOperand(i).getOpcode() == ISD::FP_TO_UINT &&
9060          V->getOperand(i).getOperand(0).getOpcode() == ISD::LOAD))
9061       IsLoad = true;
9062     // If the operands are different or the input is not a load and has more
9063     // uses than just this BV node, then it isn't a splat.
9064     if (V->getOperand(i) != Op0 ||
9065         (!IsLoad && !V->isOnlyUserOf(V->getOperand(i).getNode())))
9066       IsSplat = false;
9067   }
9068   return !(IsSplat && IsLoad);
9069 }
9070 
9071 // Lower BITCAST(f128, (build_pair i64, i64)) to BUILD_FP128.
9072 SDValue PPCTargetLowering::LowerBITCAST(SDValue Op, SelectionDAG &DAG) const {
9073 
9074   SDLoc dl(Op);
9075   SDValue Op0 = Op->getOperand(0);
9076 
9077   if ((Op.getValueType() != MVT::f128) ||
9078       (Op0.getOpcode() != ISD::BUILD_PAIR) ||
9079       (Op0.getOperand(0).getValueType() != MVT::i64) ||
9080       (Op0.getOperand(1).getValueType() != MVT::i64))
9081     return SDValue();
9082 
9083   return DAG.getNode(PPCISD::BUILD_FP128, dl, MVT::f128, Op0.getOperand(0),
9084                      Op0.getOperand(1));
9085 }
9086 
9087 static const SDValue *getNormalLoadInput(const SDValue &Op, bool &IsPermuted) {
9088   const SDValue *InputLoad = &Op;
9089   while (InputLoad->getOpcode() == ISD::BITCAST)
9090     InputLoad = &InputLoad->getOperand(0);
9091   if (InputLoad->getOpcode() == ISD::SCALAR_TO_VECTOR ||
9092       InputLoad->getOpcode() == PPCISD::SCALAR_TO_VECTOR_PERMUTED) {
9093     IsPermuted = InputLoad->getOpcode() == PPCISD::SCALAR_TO_VECTOR_PERMUTED;
9094     InputLoad = &InputLoad->getOperand(0);
9095   }
9096   if (InputLoad->getOpcode() != ISD::LOAD)
9097     return nullptr;
9098   LoadSDNode *LD = cast<LoadSDNode>(*InputLoad);
9099   return ISD::isNormalLoad(LD) ? InputLoad : nullptr;
9100 }
9101 
9102 // Convert the argument APFloat to a single precision APFloat if there is no
9103 // loss in information during the conversion to single precision APFloat and the
9104 // resulting number is not a denormal number. Return true if successful.
9105 bool llvm::convertToNonDenormSingle(APFloat &ArgAPFloat) {
9106   APFloat APFloatToConvert = ArgAPFloat;
9107   bool LosesInfo = true;
9108   APFloatToConvert.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
9109                            &LosesInfo);
9110   bool Success = (!LosesInfo && !APFloatToConvert.isDenormal());
9111   if (Success)
9112     ArgAPFloat = APFloatToConvert;
9113   return Success;
9114 }
9115 
9116 // Bitcast the argument APInt to a double and convert it to a single precision
9117 // APFloat, bitcast the APFloat to an APInt and assign it to the original
9118 // argument if there is no loss in information during the conversion from
9119 // double to single precision APFloat and the resulting number is not a denormal
9120 // number. Return true if successful.
9121 bool llvm::convertToNonDenormSingle(APInt &ArgAPInt) {
9122   double DpValue = ArgAPInt.bitsToDouble();
9123   APFloat APFloatDp(DpValue);
9124   bool Success = convertToNonDenormSingle(APFloatDp);
9125   if (Success)
9126     ArgAPInt = APFloatDp.bitcastToAPInt();
9127   return Success;
9128 }
9129 
9130 // Nondestructive check for convertTonNonDenormSingle.
9131 bool llvm::checkConvertToNonDenormSingle(APFloat &ArgAPFloat) {
9132   // Only convert if it loses info, since XXSPLTIDP should
9133   // handle the other case.
9134   APFloat APFloatToConvert = ArgAPFloat;
9135   bool LosesInfo = true;
9136   APFloatToConvert.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
9137                            &LosesInfo);
9138 
9139   return (!LosesInfo && !APFloatToConvert.isDenormal());
9140 }
9141 
9142 static bool isValidSplatLoad(const PPCSubtarget &Subtarget, const SDValue &Op,
9143                              unsigned &Opcode) {
9144   LoadSDNode *InputNode = dyn_cast<LoadSDNode>(Op.getOperand(0));
9145   if (!InputNode || !Subtarget.hasVSX() || !ISD::isUNINDEXEDLoad(InputNode))
9146     return false;
9147 
9148   EVT Ty = Op->getValueType(0);
9149   // For v2f64, v4f32 and v4i32 types, we require the load to be non-extending
9150   // as we cannot handle extending loads for these types.
9151   if ((Ty == MVT::v2f64 || Ty == MVT::v4f32 || Ty == MVT::v4i32) &&
9152       ISD::isNON_EXTLoad(InputNode))
9153     return true;
9154 
9155   EVT MemVT = InputNode->getMemoryVT();
9156   // For v8i16 and v16i8 types, extending loads can be handled as long as the
9157   // memory VT is the same vector element VT type.
9158   // The loads feeding into the v8i16 and v16i8 types will be extending because
9159   // scalar i8/i16 are not legal types.
9160   if ((Ty == MVT::v8i16 || Ty == MVT::v16i8) && ISD::isEXTLoad(InputNode) &&
9161       (MemVT == Ty.getVectorElementType()))
9162     return true;
9163 
9164   if (Ty == MVT::v2i64) {
9165     // Check the extend type, when the input type is i32, and the output vector
9166     // type is v2i64.
9167     if (MemVT == MVT::i32) {
9168       if (ISD::isZEXTLoad(InputNode))
9169         Opcode = PPCISD::ZEXT_LD_SPLAT;
9170       if (ISD::isSEXTLoad(InputNode))
9171         Opcode = PPCISD::SEXT_LD_SPLAT;
9172     }
9173     return true;
9174   }
9175   return false;
9176 }
9177 
9178 // If this is a case we can't handle, return null and let the default
9179 // expansion code take care of it.  If we CAN select this case, and if it
9180 // selects to a single instruction, return Op.  Otherwise, if we can codegen
9181 // this case more efficiently than a constant pool load, lower it to the
9182 // sequence of ops that should be used.
9183 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
9184                                              SelectionDAG &DAG) const {
9185   SDLoc dl(Op);
9186   BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9187   assert(BVN && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
9188 
9189   // Check if this is a splat of a constant value.
9190   APInt APSplatBits, APSplatUndef;
9191   unsigned SplatBitSize;
9192   bool HasAnyUndefs;
9193   bool BVNIsConstantSplat =
9194       BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
9195                            HasAnyUndefs, 0, !Subtarget.isLittleEndian());
9196 
9197   // If it is a splat of a double, check if we can shrink it to a 32 bit
9198   // non-denormal float which when converted back to double gives us the same
9199   // double. This is to exploit the XXSPLTIDP instruction.
9200   // If we lose precision, we use XXSPLTI32DX.
9201   if (BVNIsConstantSplat && (SplatBitSize == 64) &&
9202       Subtarget.hasPrefixInstrs()) {
9203     // Check the type first to short-circuit so we don't modify APSplatBits if
9204     // this block isn't executed.
9205     if ((Op->getValueType(0) == MVT::v2f64) &&
9206         convertToNonDenormSingle(APSplatBits)) {
9207       SDValue SplatNode = DAG.getNode(
9208           PPCISD::XXSPLTI_SP_TO_DP, dl, MVT::v2f64,
9209           DAG.getTargetConstant(APSplatBits.getZExtValue(), dl, MVT::i32));
9210       return DAG.getBitcast(Op.getValueType(), SplatNode);
9211     } else {
9212       // We may lose precision, so we have to use XXSPLTI32DX.
9213 
9214       uint32_t Hi =
9215           (uint32_t)((APSplatBits.getZExtValue() & 0xFFFFFFFF00000000LL) >> 32);
9216       uint32_t Lo =
9217           (uint32_t)(APSplatBits.getZExtValue() & 0xFFFFFFFF);
9218       SDValue SplatNode = DAG.getUNDEF(MVT::v2i64);
9219 
9220       if (!Hi || !Lo)
9221         // If either load is 0, then we should generate XXLXOR to set to 0.
9222         SplatNode = DAG.getTargetConstant(0, dl, MVT::v2i64);
9223 
9224       if (Hi)
9225         SplatNode = DAG.getNode(
9226             PPCISD::XXSPLTI32DX, dl, MVT::v2i64, SplatNode,
9227             DAG.getTargetConstant(0, dl, MVT::i32),
9228             DAG.getTargetConstant(Hi, dl, MVT::i32));
9229 
9230       if (Lo)
9231         SplatNode =
9232             DAG.getNode(PPCISD::XXSPLTI32DX, dl, MVT::v2i64, SplatNode,
9233                         DAG.getTargetConstant(1, dl, MVT::i32),
9234                         DAG.getTargetConstant(Lo, dl, MVT::i32));
9235 
9236       return DAG.getBitcast(Op.getValueType(), SplatNode);
9237     }
9238   }
9239 
9240   if (!BVNIsConstantSplat || SplatBitSize > 32) {
9241     unsigned NewOpcode = PPCISD::LD_SPLAT;
9242 
9243     // Handle load-and-splat patterns as we have instructions that will do this
9244     // in one go.
9245     if (DAG.isSplatValue(Op, true) &&
9246         isValidSplatLoad(Subtarget, Op, NewOpcode)) {
9247       const SDValue *InputLoad = &Op.getOperand(0);
9248       LoadSDNode *LD = cast<LoadSDNode>(*InputLoad);
9249 
9250       // If the input load is an extending load, it will be an i32 -> i64
9251       // extending load and isValidSplatLoad() will update NewOpcode.
9252       unsigned MemorySize = LD->getMemoryVT().getScalarSizeInBits();
9253       unsigned ElementSize =
9254           MemorySize * ((NewOpcode == PPCISD::LD_SPLAT) ? 1 : 2);
9255 
9256       assert(((ElementSize == 2 * MemorySize)
9257                   ? (NewOpcode == PPCISD::ZEXT_LD_SPLAT ||
9258                      NewOpcode == PPCISD::SEXT_LD_SPLAT)
9259                   : (NewOpcode == PPCISD::LD_SPLAT)) &&
9260              "Unmatched element size and opcode!\n");
9261 
9262       // Checking for a single use of this load, we have to check for vector
9263       // width (128 bits) / ElementSize uses (since each operand of the
9264       // BUILD_VECTOR is a separate use of the value.
9265       unsigned NumUsesOfInputLD = 128 / ElementSize;
9266       for (SDValue BVInOp : Op->ops())
9267         if (BVInOp.isUndef())
9268           NumUsesOfInputLD--;
9269 
9270       // Exclude somes case where LD_SPLAT is worse than scalar_to_vector:
9271       // Below cases should also happen for "lfiwzx/lfiwax + LE target + index
9272       // 1" and "lxvrhx + BE target + index 7" and "lxvrbx + BE target + index
9273       // 15", but funciton IsValidSplatLoad() now will only return true when
9274       // the data at index 0 is not nullptr. So we will not get into trouble for
9275       // these cases.
9276       //
9277       // case 1 - lfiwzx/lfiwax
9278       // 1.1: load result is i32 and is sign/zero extend to i64;
9279       // 1.2: build a v2i64 vector type with above loaded value;
9280       // 1.3: the vector has only one value at index 0, others are all undef;
9281       // 1.4: on BE target, so that lfiwzx/lfiwax does not need any permute.
9282       if (NumUsesOfInputLD == 1 &&
9283           (Op->getValueType(0) == MVT::v2i64 && NewOpcode != PPCISD::LD_SPLAT &&
9284            !Subtarget.isLittleEndian() && Subtarget.hasVSX() &&
9285            Subtarget.hasLFIWAX()))
9286         return SDValue();
9287 
9288       // case 2 - lxvr[hb]x
9289       // 2.1: load result is at most i16;
9290       // 2.2: build a vector with above loaded value;
9291       // 2.3: the vector has only one value at index 0, others are all undef;
9292       // 2.4: on LE target, so that lxvr[hb]x does not need any permute.
9293       if (NumUsesOfInputLD == 1 && Subtarget.isLittleEndian() &&
9294           Subtarget.isISA3_1() && ElementSize <= 16)
9295         return SDValue();
9296 
9297       assert(NumUsesOfInputLD > 0 && "No uses of input LD of a build_vector?");
9298       if (InputLoad->getNode()->hasNUsesOfValue(NumUsesOfInputLD, 0) &&
9299           Subtarget.hasVSX()) {
9300         SDValue Ops[] = {
9301           LD->getChain(),    // Chain
9302           LD->getBasePtr(),  // Ptr
9303           DAG.getValueType(Op.getValueType()) // VT
9304         };
9305         SDValue LdSplt = DAG.getMemIntrinsicNode(
9306             NewOpcode, dl, DAG.getVTList(Op.getValueType(), MVT::Other), Ops,
9307             LD->getMemoryVT(), LD->getMemOperand());
9308         // Replace all uses of the output chain of the original load with the
9309         // output chain of the new load.
9310         DAG.ReplaceAllUsesOfValueWith(InputLoad->getValue(1),
9311                                       LdSplt.getValue(1));
9312         return LdSplt;
9313       }
9314     }
9315 
9316     // In 64BIT mode BUILD_VECTOR nodes that are not constant splats of up to
9317     // 32-bits can be lowered to VSX instructions under certain conditions.
9318     // Without VSX, there is no pattern more efficient than expanding the node.
9319     if (Subtarget.hasVSX() && Subtarget.isPPC64() &&
9320         haveEfficientBuildVectorPattern(BVN, Subtarget.hasDirectMove(),
9321                                         Subtarget.hasP8Vector()))
9322       return Op;
9323     return SDValue();
9324   }
9325 
9326   uint64_t SplatBits = APSplatBits.getZExtValue();
9327   uint64_t SplatUndef = APSplatUndef.getZExtValue();
9328   unsigned SplatSize = SplatBitSize / 8;
9329 
9330   // First, handle single instruction cases.
9331 
9332   // All zeros?
9333   if (SplatBits == 0) {
9334     // Canonicalize all zero vectors to be v4i32.
9335     if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
9336       SDValue Z = DAG.getConstant(0, dl, MVT::v4i32);
9337       Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
9338     }
9339     return Op;
9340   }
9341 
9342   // We have XXSPLTIW for constant splats four bytes wide.
9343   // Given vector length is a multiple of 4, 2-byte splats can be replaced
9344   // with 4-byte splats. We replicate the SplatBits in case of 2-byte splat to
9345   // make a 4-byte splat element. For example: 2-byte splat of 0xABAB can be
9346   // turned into a 4-byte splat of 0xABABABAB.
9347   if (Subtarget.hasPrefixInstrs() && SplatSize == 2)
9348     return getCanonicalConstSplat(SplatBits | (SplatBits << 16), SplatSize * 2,
9349                                   Op.getValueType(), DAG, dl);
9350 
9351   if (Subtarget.hasPrefixInstrs() && SplatSize == 4)
9352     return getCanonicalConstSplat(SplatBits, SplatSize, Op.getValueType(), DAG,
9353                                   dl);
9354 
9355   // We have XXSPLTIB for constant splats one byte wide.
9356   if (Subtarget.hasP9Vector() && SplatSize == 1)
9357     return getCanonicalConstSplat(SplatBits, SplatSize, Op.getValueType(), DAG,
9358                                   dl);
9359 
9360   // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
9361   int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
9362                     (32-SplatBitSize));
9363   if (SextVal >= -16 && SextVal <= 15)
9364     return getCanonicalConstSplat(SextVal, SplatSize, Op.getValueType(), DAG,
9365                                   dl);
9366 
9367   // Two instruction sequences.
9368 
9369   // If this value is in the range [-32,30] and is even, use:
9370   //     VSPLTI[bhw](val/2) + VSPLTI[bhw](val/2)
9371   // If this value is in the range [17,31] and is odd, use:
9372   //     VSPLTI[bhw](val-16) - VSPLTI[bhw](-16)
9373   // If this value is in the range [-31,-17] and is odd, use:
9374   //     VSPLTI[bhw](val+16) + VSPLTI[bhw](-16)
9375   // Note the last two are three-instruction sequences.
9376   if (SextVal >= -32 && SextVal <= 31) {
9377     // To avoid having these optimizations undone by constant folding,
9378     // we convert to a pseudo that will be expanded later into one of
9379     // the above forms.
9380     SDValue Elt = DAG.getConstant(SextVal, dl, MVT::i32);
9381     EVT VT = (SplatSize == 1 ? MVT::v16i8 :
9382               (SplatSize == 2 ? MVT::v8i16 : MVT::v4i32));
9383     SDValue EltSize = DAG.getConstant(SplatSize, dl, MVT::i32);
9384     SDValue RetVal = DAG.getNode(PPCISD::VADD_SPLAT, dl, VT, Elt, EltSize);
9385     if (VT == Op.getValueType())
9386       return RetVal;
9387     else
9388       return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), RetVal);
9389   }
9390 
9391   // If this is 0x8000_0000 x 4, turn into vspltisw + vslw.  If it is
9392   // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000).  This is important
9393   // for fneg/fabs.
9394   if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
9395     // Make -1 and vspltisw -1:
9396     SDValue OnesV = getCanonicalConstSplat(-1, 4, MVT::v4i32, DAG, dl);
9397 
9398     // Make the VSLW intrinsic, computing 0x8000_0000.
9399     SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
9400                                    OnesV, DAG, dl);
9401 
9402     // xor by OnesV to invert it.
9403     Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
9404     return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
9405   }
9406 
9407   // Check to see if this is a wide variety of vsplti*, binop self cases.
9408   static const signed char SplatCsts[] = {
9409     -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
9410     -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
9411   };
9412 
9413   for (unsigned idx = 0; idx < std::size(SplatCsts); ++idx) {
9414     // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
9415     // cases which are ambiguous (e.g. formation of 0x8000_0000).  'vsplti -1'
9416     int i = SplatCsts[idx];
9417 
9418     // Figure out what shift amount will be used by altivec if shifted by i in
9419     // this splat size.
9420     unsigned TypeShiftAmt = i & (SplatBitSize-1);
9421 
9422     // vsplti + shl self.
9423     if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
9424       SDValue Res = getCanonicalConstSplat(i, SplatSize, MVT::Other, DAG, dl);
9425       static const unsigned IIDs[] = { // Intrinsic to use for each size.
9426         Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
9427         Intrinsic::ppc_altivec_vslw
9428       };
9429       Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
9430       return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
9431     }
9432 
9433     // vsplti + srl self.
9434     if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
9435       SDValue Res = getCanonicalConstSplat(i, SplatSize, MVT::Other, DAG, dl);
9436       static const unsigned IIDs[] = { // Intrinsic to use for each size.
9437         Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
9438         Intrinsic::ppc_altivec_vsrw
9439       };
9440       Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
9441       return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
9442     }
9443 
9444     // vsplti + rol self.
9445     if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
9446                          ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
9447       SDValue Res = getCanonicalConstSplat(i, SplatSize, MVT::Other, DAG, dl);
9448       static const unsigned IIDs[] = { // Intrinsic to use for each size.
9449         Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
9450         Intrinsic::ppc_altivec_vrlw
9451       };
9452       Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
9453       return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
9454     }
9455 
9456     // t = vsplti c, result = vsldoi t, t, 1
9457     if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
9458       SDValue T = getCanonicalConstSplat(i, SplatSize, MVT::v16i8, DAG, dl);
9459       unsigned Amt = Subtarget.isLittleEndian() ? 15 : 1;
9460       return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl);
9461     }
9462     // t = vsplti c, result = vsldoi t, t, 2
9463     if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
9464       SDValue T = getCanonicalConstSplat(i, SplatSize, MVT::v16i8, DAG, dl);
9465       unsigned Amt = Subtarget.isLittleEndian() ? 14 : 2;
9466       return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl);
9467     }
9468     // t = vsplti c, result = vsldoi t, t, 3
9469     if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
9470       SDValue T = getCanonicalConstSplat(i, SplatSize, MVT::v16i8, DAG, dl);
9471       unsigned Amt = Subtarget.isLittleEndian() ? 13 : 3;
9472       return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl);
9473     }
9474   }
9475 
9476   return SDValue();
9477 }
9478 
9479 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
9480 /// the specified operations to build the shuffle.
9481 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
9482                                       SDValue RHS, SelectionDAG &DAG,
9483                                       const SDLoc &dl) {
9484   unsigned OpNum = (PFEntry >> 26) & 0x0F;
9485   unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
9486   unsigned RHSID = (PFEntry >>  0) & ((1 << 13)-1);
9487 
9488   enum {
9489     OP_COPY = 0,  // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
9490     OP_VMRGHW,
9491     OP_VMRGLW,
9492     OP_VSPLTISW0,
9493     OP_VSPLTISW1,
9494     OP_VSPLTISW2,
9495     OP_VSPLTISW3,
9496     OP_VSLDOI4,
9497     OP_VSLDOI8,
9498     OP_VSLDOI12
9499   };
9500 
9501   if (OpNum == OP_COPY) {
9502     if (LHSID == (1*9+2)*9+3) return LHS;
9503     assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
9504     return RHS;
9505   }
9506 
9507   SDValue OpLHS, OpRHS;
9508   OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
9509   OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
9510 
9511   int ShufIdxs[16];
9512   switch (OpNum) {
9513   default: llvm_unreachable("Unknown i32 permute!");
9514   case OP_VMRGHW:
9515     ShufIdxs[ 0] =  0; ShufIdxs[ 1] =  1; ShufIdxs[ 2] =  2; ShufIdxs[ 3] =  3;
9516     ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
9517     ShufIdxs[ 8] =  4; ShufIdxs[ 9] =  5; ShufIdxs[10] =  6; ShufIdxs[11] =  7;
9518     ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
9519     break;
9520   case OP_VMRGLW:
9521     ShufIdxs[ 0] =  8; ShufIdxs[ 1] =  9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
9522     ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
9523     ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
9524     ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
9525     break;
9526   case OP_VSPLTISW0:
9527     for (unsigned i = 0; i != 16; ++i)
9528       ShufIdxs[i] = (i&3)+0;
9529     break;
9530   case OP_VSPLTISW1:
9531     for (unsigned i = 0; i != 16; ++i)
9532       ShufIdxs[i] = (i&3)+4;
9533     break;
9534   case OP_VSPLTISW2:
9535     for (unsigned i = 0; i != 16; ++i)
9536       ShufIdxs[i] = (i&3)+8;
9537     break;
9538   case OP_VSPLTISW3:
9539     for (unsigned i = 0; i != 16; ++i)
9540       ShufIdxs[i] = (i&3)+12;
9541     break;
9542   case OP_VSLDOI4:
9543     return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
9544   case OP_VSLDOI8:
9545     return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
9546   case OP_VSLDOI12:
9547     return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
9548   }
9549   EVT VT = OpLHS.getValueType();
9550   OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
9551   OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
9552   SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
9553   return DAG.getNode(ISD::BITCAST, dl, VT, T);
9554 }
9555 
9556 /// lowerToVINSERTB - Return the SDValue if this VECTOR_SHUFFLE can be handled
9557 /// by the VINSERTB instruction introduced in ISA 3.0, else just return default
9558 /// SDValue.
9559 SDValue PPCTargetLowering::lowerToVINSERTB(ShuffleVectorSDNode *N,
9560                                            SelectionDAG &DAG) const {
9561   const unsigned BytesInVector = 16;
9562   bool IsLE = Subtarget.isLittleEndian();
9563   SDLoc dl(N);
9564   SDValue V1 = N->getOperand(0);
9565   SDValue V2 = N->getOperand(1);
9566   unsigned ShiftElts = 0, InsertAtByte = 0;
9567   bool Swap = false;
9568 
9569   // Shifts required to get the byte we want at element 7.
9570   unsigned LittleEndianShifts[] = {8, 7,  6,  5,  4,  3,  2,  1,
9571                                    0, 15, 14, 13, 12, 11, 10, 9};
9572   unsigned BigEndianShifts[] = {9, 10, 11, 12, 13, 14, 15, 0,
9573                                 1, 2,  3,  4,  5,  6,  7,  8};
9574 
9575   ArrayRef<int> Mask = N->getMask();
9576   int OriginalOrder[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
9577 
9578   // For each mask element, find out if we're just inserting something
9579   // from V2 into V1 or vice versa.
9580   // Possible permutations inserting an element from V2 into V1:
9581   //   X, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
9582   //   0, X, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
9583   //   ...
9584   //   0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, X
9585   // Inserting from V1 into V2 will be similar, except mask range will be
9586   // [16,31].
9587 
9588   bool FoundCandidate = false;
9589   // If both vector operands for the shuffle are the same vector, the mask
9590   // will contain only elements from the first one and the second one will be
9591   // undef.
9592   unsigned VINSERTBSrcElem = IsLE ? 8 : 7;
9593   // Go through the mask of half-words to find an element that's being moved
9594   // from one vector to the other.
9595   for (unsigned i = 0; i < BytesInVector; ++i) {
9596     unsigned CurrentElement = Mask[i];
9597     // If 2nd operand is undefined, we should only look for element 7 in the
9598     // Mask.
9599     if (V2.isUndef() && CurrentElement != VINSERTBSrcElem)
9600       continue;
9601 
9602     bool OtherElementsInOrder = true;
9603     // Examine the other elements in the Mask to see if they're in original
9604     // order.
9605     for (unsigned j = 0; j < BytesInVector; ++j) {
9606       if (j == i)
9607         continue;
9608       // If CurrentElement is from V1 [0,15], then we the rest of the Mask to be
9609       // from V2 [16,31] and vice versa.  Unless the 2nd operand is undefined,
9610       // in which we always assume we're always picking from the 1st operand.
9611       int MaskOffset =
9612           (!V2.isUndef() && CurrentElement < BytesInVector) ? BytesInVector : 0;
9613       if (Mask[j] != OriginalOrder[j] + MaskOffset) {
9614         OtherElementsInOrder = false;
9615         break;
9616       }
9617     }
9618     // If other elements are in original order, we record the number of shifts
9619     // we need to get the element we want into element 7. Also record which byte
9620     // in the vector we should insert into.
9621     if (OtherElementsInOrder) {
9622       // If 2nd operand is undefined, we assume no shifts and no swapping.
9623       if (V2.isUndef()) {
9624         ShiftElts = 0;
9625         Swap = false;
9626       } else {
9627         // Only need the last 4-bits for shifts because operands will be swapped if CurrentElement is >= 2^4.
9628         ShiftElts = IsLE ? LittleEndianShifts[CurrentElement & 0xF]
9629                          : BigEndianShifts[CurrentElement & 0xF];
9630         Swap = CurrentElement < BytesInVector;
9631       }
9632       InsertAtByte = IsLE ? BytesInVector - (i + 1) : i;
9633       FoundCandidate = true;
9634       break;
9635     }
9636   }
9637 
9638   if (!FoundCandidate)
9639     return SDValue();
9640 
9641   // Candidate found, construct the proper SDAG sequence with VINSERTB,
9642   // optionally with VECSHL if shift is required.
9643   if (Swap)
9644     std::swap(V1, V2);
9645   if (V2.isUndef())
9646     V2 = V1;
9647   if (ShiftElts) {
9648     SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v16i8, V2, V2,
9649                               DAG.getConstant(ShiftElts, dl, MVT::i32));
9650     return DAG.getNode(PPCISD::VECINSERT, dl, MVT::v16i8, V1, Shl,
9651                        DAG.getConstant(InsertAtByte, dl, MVT::i32));
9652   }
9653   return DAG.getNode(PPCISD::VECINSERT, dl, MVT::v16i8, V1, V2,
9654                      DAG.getConstant(InsertAtByte, dl, MVT::i32));
9655 }
9656 
9657 /// lowerToVINSERTH - Return the SDValue if this VECTOR_SHUFFLE can be handled
9658 /// by the VINSERTH instruction introduced in ISA 3.0, else just return default
9659 /// SDValue.
9660 SDValue PPCTargetLowering::lowerToVINSERTH(ShuffleVectorSDNode *N,
9661                                            SelectionDAG &DAG) const {
9662   const unsigned NumHalfWords = 8;
9663   const unsigned BytesInVector = NumHalfWords * 2;
9664   // Check that the shuffle is on half-words.
9665   if (!isNByteElemShuffleMask(N, 2, 1))
9666     return SDValue();
9667 
9668   bool IsLE = Subtarget.isLittleEndian();
9669   SDLoc dl(N);
9670   SDValue V1 = N->getOperand(0);
9671   SDValue V2 = N->getOperand(1);
9672   unsigned ShiftElts = 0, InsertAtByte = 0;
9673   bool Swap = false;
9674 
9675   // Shifts required to get the half-word we want at element 3.
9676   unsigned LittleEndianShifts[] = {4, 3, 2, 1, 0, 7, 6, 5};
9677   unsigned BigEndianShifts[] = {5, 6, 7, 0, 1, 2, 3, 4};
9678 
9679   uint32_t Mask = 0;
9680   uint32_t OriginalOrderLow = 0x1234567;
9681   uint32_t OriginalOrderHigh = 0x89ABCDEF;
9682   // Now we look at mask elements 0,2,4,6,8,10,12,14.  Pack the mask into a
9683   // 32-bit space, only need 4-bit nibbles per element.
9684   for (unsigned i = 0; i < NumHalfWords; ++i) {
9685     unsigned MaskShift = (NumHalfWords - 1 - i) * 4;
9686     Mask |= ((uint32_t)(N->getMaskElt(i * 2) / 2) << MaskShift);
9687   }
9688 
9689   // For each mask element, find out if we're just inserting something
9690   // from V2 into V1 or vice versa.  Possible permutations inserting an element
9691   // from V2 into V1:
9692   //   X, 1, 2, 3, 4, 5, 6, 7
9693   //   0, X, 2, 3, 4, 5, 6, 7
9694   //   0, 1, X, 3, 4, 5, 6, 7
9695   //   0, 1, 2, X, 4, 5, 6, 7
9696   //   0, 1, 2, 3, X, 5, 6, 7
9697   //   0, 1, 2, 3, 4, X, 6, 7
9698   //   0, 1, 2, 3, 4, 5, X, 7
9699   //   0, 1, 2, 3, 4, 5, 6, X
9700   // Inserting from V1 into V2 will be similar, except mask range will be [8,15].
9701 
9702   bool FoundCandidate = false;
9703   // Go through the mask of half-words to find an element that's being moved
9704   // from one vector to the other.
9705   for (unsigned i = 0; i < NumHalfWords; ++i) {
9706     unsigned MaskShift = (NumHalfWords - 1 - i) * 4;
9707     uint32_t MaskOneElt = (Mask >> MaskShift) & 0xF;
9708     uint32_t MaskOtherElts = ~(0xF << MaskShift);
9709     uint32_t TargetOrder = 0x0;
9710 
9711     // If both vector operands for the shuffle are the same vector, the mask
9712     // will contain only elements from the first one and the second one will be
9713     // undef.
9714     if (V2.isUndef()) {
9715       ShiftElts = 0;
9716       unsigned VINSERTHSrcElem = IsLE ? 4 : 3;
9717       TargetOrder = OriginalOrderLow;
9718       Swap = false;
9719       // Skip if not the correct element or mask of other elements don't equal
9720       // to our expected order.
9721       if (MaskOneElt == VINSERTHSrcElem &&
9722           (Mask & MaskOtherElts) == (TargetOrder & MaskOtherElts)) {
9723         InsertAtByte = IsLE ? BytesInVector - (i + 1) * 2 : i * 2;
9724         FoundCandidate = true;
9725         break;
9726       }
9727     } else { // If both operands are defined.
9728       // Target order is [8,15] if the current mask is between [0,7].
9729       TargetOrder =
9730           (MaskOneElt < NumHalfWords) ? OriginalOrderHigh : OriginalOrderLow;
9731       // Skip if mask of other elements don't equal our expected order.
9732       if ((Mask & MaskOtherElts) == (TargetOrder & MaskOtherElts)) {
9733         // We only need the last 3 bits for the number of shifts.
9734         ShiftElts = IsLE ? LittleEndianShifts[MaskOneElt & 0x7]
9735                          : BigEndianShifts[MaskOneElt & 0x7];
9736         InsertAtByte = IsLE ? BytesInVector - (i + 1) * 2 : i * 2;
9737         Swap = MaskOneElt < NumHalfWords;
9738         FoundCandidate = true;
9739         break;
9740       }
9741     }
9742   }
9743 
9744   if (!FoundCandidate)
9745     return SDValue();
9746 
9747   // Candidate found, construct the proper SDAG sequence with VINSERTH,
9748   // optionally with VECSHL if shift is required.
9749   if (Swap)
9750     std::swap(V1, V2);
9751   if (V2.isUndef())
9752     V2 = V1;
9753   SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
9754   if (ShiftElts) {
9755     // Double ShiftElts because we're left shifting on v16i8 type.
9756     SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v16i8, V2, V2,
9757                               DAG.getConstant(2 * ShiftElts, dl, MVT::i32));
9758     SDValue Conv2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, Shl);
9759     SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v8i16, Conv1, Conv2,
9760                               DAG.getConstant(InsertAtByte, dl, MVT::i32));
9761     return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins);
9762   }
9763   SDValue Conv2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
9764   SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v8i16, Conv1, Conv2,
9765                             DAG.getConstant(InsertAtByte, dl, MVT::i32));
9766   return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins);
9767 }
9768 
9769 /// lowerToXXSPLTI32DX - Return the SDValue if this VECTOR_SHUFFLE can be
9770 /// handled by the XXSPLTI32DX instruction introduced in ISA 3.1, otherwise
9771 /// return the default SDValue.
9772 SDValue PPCTargetLowering::lowerToXXSPLTI32DX(ShuffleVectorSDNode *SVN,
9773                                               SelectionDAG &DAG) const {
9774   // The LHS and RHS may be bitcasts to v16i8 as we canonicalize shuffles
9775   // to v16i8. Peek through the bitcasts to get the actual operands.
9776   SDValue LHS = peekThroughBitcasts(SVN->getOperand(0));
9777   SDValue RHS = peekThroughBitcasts(SVN->getOperand(1));
9778 
9779   auto ShuffleMask = SVN->getMask();
9780   SDValue VecShuffle(SVN, 0);
9781   SDLoc DL(SVN);
9782 
9783   // Check that we have a four byte shuffle.
9784   if (!isNByteElemShuffleMask(SVN, 4, 1))
9785     return SDValue();
9786 
9787   // Canonicalize the RHS being a BUILD_VECTOR when lowering to xxsplti32dx.
9788   if (RHS->getOpcode() != ISD::BUILD_VECTOR) {
9789     std::swap(LHS, RHS);
9790     VecShuffle = peekThroughBitcasts(DAG.getCommutedVectorShuffle(*SVN));
9791     ShuffleVectorSDNode *CommutedSV = dyn_cast<ShuffleVectorSDNode>(VecShuffle);
9792     if (!CommutedSV)
9793       return SDValue();
9794     ShuffleMask = CommutedSV->getMask();
9795   }
9796 
9797   // Ensure that the RHS is a vector of constants.
9798   BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
9799   if (!BVN)
9800     return SDValue();
9801 
9802   // Check if RHS is a splat of 4-bytes (or smaller).
9803   APInt APSplatValue, APSplatUndef;
9804   unsigned SplatBitSize;
9805   bool HasAnyUndefs;
9806   if (!BVN->isConstantSplat(APSplatValue, APSplatUndef, SplatBitSize,
9807                             HasAnyUndefs, 0, !Subtarget.isLittleEndian()) ||
9808       SplatBitSize > 32)
9809     return SDValue();
9810 
9811   // Check that the shuffle mask matches the semantics of XXSPLTI32DX.
9812   // The instruction splats a constant C into two words of the source vector
9813   // producing { C, Unchanged, C, Unchanged } or { Unchanged, C, Unchanged, C }.
9814   // Thus we check that the shuffle mask is the equivalent  of
9815   // <0, [4-7], 2, [4-7]> or <[4-7], 1, [4-7], 3> respectively.
9816   // Note: the check above of isNByteElemShuffleMask() ensures that the bytes
9817   // within each word are consecutive, so we only need to check the first byte.
9818   SDValue Index;
9819   bool IsLE = Subtarget.isLittleEndian();
9820   if ((ShuffleMask[0] == 0 && ShuffleMask[8] == 8) &&
9821       (ShuffleMask[4] % 4 == 0 && ShuffleMask[12] % 4 == 0 &&
9822        ShuffleMask[4] > 15 && ShuffleMask[12] > 15))
9823     Index = DAG.getTargetConstant(IsLE ? 0 : 1, DL, MVT::i32);
9824   else if ((ShuffleMask[4] == 4 && ShuffleMask[12] == 12) &&
9825            (ShuffleMask[0] % 4 == 0 && ShuffleMask[8] % 4 == 0 &&
9826             ShuffleMask[0] > 15 && ShuffleMask[8] > 15))
9827     Index = DAG.getTargetConstant(IsLE ? 1 : 0, DL, MVT::i32);
9828   else
9829     return SDValue();
9830 
9831   // If the splat is narrower than 32-bits, we need to get the 32-bit value
9832   // for XXSPLTI32DX.
9833   unsigned SplatVal = APSplatValue.getZExtValue();
9834   for (; SplatBitSize < 32; SplatBitSize <<= 1)
9835     SplatVal |= (SplatVal << SplatBitSize);
9836 
9837   SDValue SplatNode = DAG.getNode(
9838       PPCISD::XXSPLTI32DX, DL, MVT::v2i64, DAG.getBitcast(MVT::v2i64, LHS),
9839       Index, DAG.getTargetConstant(SplatVal, DL, MVT::i32));
9840   return DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, SplatNode);
9841 }
9842 
9843 /// LowerROTL - Custom lowering for ROTL(v1i128) to vector_shuffle(v16i8).
9844 /// We lower ROTL(v1i128) to vector_shuffle(v16i8) only if shift amount is
9845 /// a multiple of 8. Otherwise convert it to a scalar rotation(i128)
9846 /// i.e (or (shl x, C1), (srl x, 128-C1)).
9847 SDValue PPCTargetLowering::LowerROTL(SDValue Op, SelectionDAG &DAG) const {
9848   assert(Op.getOpcode() == ISD::ROTL && "Should only be called for ISD::ROTL");
9849   assert(Op.getValueType() == MVT::v1i128 &&
9850          "Only set v1i128 as custom, other type shouldn't reach here!");
9851   SDLoc dl(Op);
9852   SDValue N0 = peekThroughBitcasts(Op.getOperand(0));
9853   SDValue N1 = peekThroughBitcasts(Op.getOperand(1));
9854   unsigned SHLAmt = N1.getConstantOperandVal(0);
9855   if (SHLAmt % 8 == 0) {
9856     std::array<int, 16> Mask;
9857     std::iota(Mask.begin(), Mask.end(), 0);
9858     std::rotate(Mask.begin(), Mask.begin() + SHLAmt / 8, Mask.end());
9859     if (SDValue Shuffle =
9860             DAG.getVectorShuffle(MVT::v16i8, dl, DAG.getBitcast(MVT::v16i8, N0),
9861                                  DAG.getUNDEF(MVT::v16i8), Mask))
9862       return DAG.getNode(ISD::BITCAST, dl, MVT::v1i128, Shuffle);
9863   }
9864   SDValue ArgVal = DAG.getBitcast(MVT::i128, N0);
9865   SDValue SHLOp = DAG.getNode(ISD::SHL, dl, MVT::i128, ArgVal,
9866                               DAG.getConstant(SHLAmt, dl, MVT::i32));
9867   SDValue SRLOp = DAG.getNode(ISD::SRL, dl, MVT::i128, ArgVal,
9868                               DAG.getConstant(128 - SHLAmt, dl, MVT::i32));
9869   SDValue OROp = DAG.getNode(ISD::OR, dl, MVT::i128, SHLOp, SRLOp);
9870   return DAG.getNode(ISD::BITCAST, dl, MVT::v1i128, OROp);
9871 }
9872 
9873 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE.  If this
9874 /// is a shuffle we can handle in a single instruction, return it.  Otherwise,
9875 /// return the code it can be lowered into.  Worst case, it can always be
9876 /// lowered into a vperm.
9877 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
9878                                                SelectionDAG &DAG) const {
9879   SDLoc dl(Op);
9880   SDValue V1 = Op.getOperand(0);
9881   SDValue V2 = Op.getOperand(1);
9882   ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9883 
9884   // Any nodes that were combined in the target-independent combiner prior
9885   // to vector legalization will not be sent to the target combine. Try to
9886   // combine it here.
9887   if (SDValue NewShuffle = combineVectorShuffle(SVOp, DAG)) {
9888     if (!isa<ShuffleVectorSDNode>(NewShuffle))
9889       return NewShuffle;
9890     Op = NewShuffle;
9891     SVOp = cast<ShuffleVectorSDNode>(Op);
9892     V1 = Op.getOperand(0);
9893     V2 = Op.getOperand(1);
9894   }
9895   EVT VT = Op.getValueType();
9896   bool isLittleEndian = Subtarget.isLittleEndian();
9897 
9898   unsigned ShiftElts, InsertAtByte;
9899   bool Swap = false;
9900 
9901   // If this is a load-and-splat, we can do that with a single instruction
9902   // in some cases. However if the load has multiple uses, we don't want to
9903   // combine it because that will just produce multiple loads.
9904   bool IsPermutedLoad = false;
9905   const SDValue *InputLoad = getNormalLoadInput(V1, IsPermutedLoad);
9906   if (InputLoad && Subtarget.hasVSX() && V2.isUndef() &&
9907       (PPC::isSplatShuffleMask(SVOp, 4) || PPC::isSplatShuffleMask(SVOp, 8)) &&
9908       InputLoad->hasOneUse()) {
9909     bool IsFourByte = PPC::isSplatShuffleMask(SVOp, 4);
9910     int SplatIdx =
9911       PPC::getSplatIdxForPPCMnemonics(SVOp, IsFourByte ? 4 : 8, DAG);
9912 
9913     // The splat index for permuted loads will be in the left half of the vector
9914     // which is strictly wider than the loaded value by 8 bytes. So we need to
9915     // adjust the splat index to point to the correct address in memory.
9916     if (IsPermutedLoad) {
9917       assert((isLittleEndian || IsFourByte) &&
9918              "Unexpected size for permuted load on big endian target");
9919       SplatIdx += IsFourByte ? 2 : 1;
9920       assert((SplatIdx < (IsFourByte ? 4 : 2)) &&
9921              "Splat of a value outside of the loaded memory");
9922     }
9923 
9924     LoadSDNode *LD = cast<LoadSDNode>(*InputLoad);
9925     // For 4-byte load-and-splat, we need Power9.
9926     if ((IsFourByte && Subtarget.hasP9Vector()) || !IsFourByte) {
9927       uint64_t Offset = 0;
9928       if (IsFourByte)
9929         Offset = isLittleEndian ? (3 - SplatIdx) * 4 : SplatIdx * 4;
9930       else
9931         Offset = isLittleEndian ? (1 - SplatIdx) * 8 : SplatIdx * 8;
9932 
9933       // If the width of the load is the same as the width of the splat,
9934       // loading with an offset would load the wrong memory.
9935       if (LD->getValueType(0).getSizeInBits() == (IsFourByte ? 32 : 64))
9936         Offset = 0;
9937 
9938       SDValue BasePtr = LD->getBasePtr();
9939       if (Offset != 0)
9940         BasePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
9941                               BasePtr, DAG.getIntPtrConstant(Offset, dl));
9942       SDValue Ops[] = {
9943         LD->getChain(),    // Chain
9944         BasePtr,           // BasePtr
9945         DAG.getValueType(Op.getValueType()) // VT
9946       };
9947       SDVTList VTL =
9948         DAG.getVTList(IsFourByte ? MVT::v4i32 : MVT::v2i64, MVT::Other);
9949       SDValue LdSplt =
9950         DAG.getMemIntrinsicNode(PPCISD::LD_SPLAT, dl, VTL,
9951                                 Ops, LD->getMemoryVT(), LD->getMemOperand());
9952       DAG.ReplaceAllUsesOfValueWith(InputLoad->getValue(1), LdSplt.getValue(1));
9953       if (LdSplt.getValueType() != SVOp->getValueType(0))
9954         LdSplt = DAG.getBitcast(SVOp->getValueType(0), LdSplt);
9955       return LdSplt;
9956     }
9957   }
9958 
9959   // All v2i64 and v2f64 shuffles are legal
9960   if (VT == MVT::v2i64 || VT == MVT::v2f64)
9961     return Op;
9962 
9963   if (Subtarget.hasP9Vector() &&
9964       PPC::isXXINSERTWMask(SVOp, ShiftElts, InsertAtByte, Swap,
9965                            isLittleEndian)) {
9966     if (Swap)
9967       std::swap(V1, V2);
9968     SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
9969     SDValue Conv2 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V2);
9970     if (ShiftElts) {
9971       SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v4i32, Conv2, Conv2,
9972                                 DAG.getConstant(ShiftElts, dl, MVT::i32));
9973       SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v4i32, Conv1, Shl,
9974                                 DAG.getConstant(InsertAtByte, dl, MVT::i32));
9975       return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins);
9976     }
9977     SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v4i32, Conv1, Conv2,
9978                               DAG.getConstant(InsertAtByte, dl, MVT::i32));
9979     return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins);
9980   }
9981 
9982   if (Subtarget.hasPrefixInstrs()) {
9983     SDValue SplatInsertNode;
9984     if ((SplatInsertNode = lowerToXXSPLTI32DX(SVOp, DAG)))
9985       return SplatInsertNode;
9986   }
9987 
9988   if (Subtarget.hasP9Altivec()) {
9989     SDValue NewISDNode;
9990     if ((NewISDNode = lowerToVINSERTH(SVOp, DAG)))
9991       return NewISDNode;
9992 
9993     if ((NewISDNode = lowerToVINSERTB(SVOp, DAG)))
9994       return NewISDNode;
9995   }
9996 
9997   if (Subtarget.hasVSX() &&
9998       PPC::isXXSLDWIShuffleMask(SVOp, ShiftElts, Swap, isLittleEndian)) {
9999     if (Swap)
10000       std::swap(V1, V2);
10001     SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
10002     SDValue Conv2 =
10003         DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V2.isUndef() ? V1 : V2);
10004 
10005     SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v4i32, Conv1, Conv2,
10006                               DAG.getConstant(ShiftElts, dl, MVT::i32));
10007     return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Shl);
10008   }
10009 
10010   if (Subtarget.hasVSX() &&
10011     PPC::isXXPERMDIShuffleMask(SVOp, ShiftElts, Swap, isLittleEndian)) {
10012     if (Swap)
10013       std::swap(V1, V2);
10014     SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1);
10015     SDValue Conv2 =
10016         DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2.isUndef() ? V1 : V2);
10017 
10018     SDValue PermDI = DAG.getNode(PPCISD::XXPERMDI, dl, MVT::v2i64, Conv1, Conv2,
10019                               DAG.getConstant(ShiftElts, dl, MVT::i32));
10020     return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, PermDI);
10021   }
10022 
10023   if (Subtarget.hasP9Vector()) {
10024      if (PPC::isXXBRHShuffleMask(SVOp)) {
10025       SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
10026       SDValue ReveHWord = DAG.getNode(ISD::BSWAP, dl, MVT::v8i16, Conv);
10027       return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveHWord);
10028     } else if (PPC::isXXBRWShuffleMask(SVOp)) {
10029       SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
10030       SDValue ReveWord = DAG.getNode(ISD::BSWAP, dl, MVT::v4i32, Conv);
10031       return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveWord);
10032     } else if (PPC::isXXBRDShuffleMask(SVOp)) {
10033       SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1);
10034       SDValue ReveDWord = DAG.getNode(ISD::BSWAP, dl, MVT::v2i64, Conv);
10035       return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveDWord);
10036     } else if (PPC::isXXBRQShuffleMask(SVOp)) {
10037       SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v1i128, V1);
10038       SDValue ReveQWord = DAG.getNode(ISD::BSWAP, dl, MVT::v1i128, Conv);
10039       return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveQWord);
10040     }
10041   }
10042 
10043   if (Subtarget.hasVSX()) {
10044     if (V2.isUndef() && PPC::isSplatShuffleMask(SVOp, 4)) {
10045       int SplatIdx = PPC::getSplatIdxForPPCMnemonics(SVOp, 4, DAG);
10046 
10047       SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
10048       SDValue Splat = DAG.getNode(PPCISD::XXSPLT, dl, MVT::v4i32, Conv,
10049                                   DAG.getConstant(SplatIdx, dl, MVT::i32));
10050       return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Splat);
10051     }
10052 
10053     // Left shifts of 8 bytes are actually swaps. Convert accordingly.
10054     if (V2.isUndef() && PPC::isVSLDOIShuffleMask(SVOp, 1, DAG) == 8) {
10055       SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
10056       SDValue Swap = DAG.getNode(PPCISD::SWAP_NO_CHAIN, dl, MVT::v2f64, Conv);
10057       return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Swap);
10058     }
10059   }
10060 
10061   // Cases that are handled by instructions that take permute immediates
10062   // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
10063   // selected by the instruction selector.
10064   if (V2.isUndef()) {
10065     if (PPC::isSplatShuffleMask(SVOp, 1) ||
10066         PPC::isSplatShuffleMask(SVOp, 2) ||
10067         PPC::isSplatShuffleMask(SVOp, 4) ||
10068         PPC::isVPKUWUMShuffleMask(SVOp, 1, DAG) ||
10069         PPC::isVPKUHUMShuffleMask(SVOp, 1, DAG) ||
10070         PPC::isVSLDOIShuffleMask(SVOp, 1, DAG) != -1 ||
10071         PPC::isVMRGLShuffleMask(SVOp, 1, 1, DAG) ||
10072         PPC::isVMRGLShuffleMask(SVOp, 2, 1, DAG) ||
10073         PPC::isVMRGLShuffleMask(SVOp, 4, 1, DAG) ||
10074         PPC::isVMRGHShuffleMask(SVOp, 1, 1, DAG) ||
10075         PPC::isVMRGHShuffleMask(SVOp, 2, 1, DAG) ||
10076         PPC::isVMRGHShuffleMask(SVOp, 4, 1, DAG) ||
10077         (Subtarget.hasP8Altivec() && (
10078          PPC::isVPKUDUMShuffleMask(SVOp, 1, DAG) ||
10079          PPC::isVMRGEOShuffleMask(SVOp, true, 1, DAG) ||
10080          PPC::isVMRGEOShuffleMask(SVOp, false, 1, DAG)))) {
10081       return Op;
10082     }
10083   }
10084 
10085   // Altivec has a variety of "shuffle immediates" that take two vector inputs
10086   // and produce a fixed permutation.  If any of these match, do not lower to
10087   // VPERM.
10088   unsigned int ShuffleKind = isLittleEndian ? 2 : 0;
10089   if (PPC::isVPKUWUMShuffleMask(SVOp, ShuffleKind, DAG) ||
10090       PPC::isVPKUHUMShuffleMask(SVOp, ShuffleKind, DAG) ||
10091       PPC::isVSLDOIShuffleMask(SVOp, ShuffleKind, DAG) != -1 ||
10092       PPC::isVMRGLShuffleMask(SVOp, 1, ShuffleKind, DAG) ||
10093       PPC::isVMRGLShuffleMask(SVOp, 2, ShuffleKind, DAG) ||
10094       PPC::isVMRGLShuffleMask(SVOp, 4, ShuffleKind, DAG) ||
10095       PPC::isVMRGHShuffleMask(SVOp, 1, ShuffleKind, DAG) ||
10096       PPC::isVMRGHShuffleMask(SVOp, 2, ShuffleKind, DAG) ||
10097       PPC::isVMRGHShuffleMask(SVOp, 4, ShuffleKind, DAG) ||
10098       (Subtarget.hasP8Altivec() && (
10099        PPC::isVPKUDUMShuffleMask(SVOp, ShuffleKind, DAG) ||
10100        PPC::isVMRGEOShuffleMask(SVOp, true, ShuffleKind, DAG) ||
10101        PPC::isVMRGEOShuffleMask(SVOp, false, ShuffleKind, DAG))))
10102     return Op;
10103 
10104   // Check to see if this is a shuffle of 4-byte values.  If so, we can use our
10105   // perfect shuffle table to emit an optimal matching sequence.
10106   ArrayRef<int> PermMask = SVOp->getMask();
10107 
10108   if (!DisablePerfectShuffle && !isLittleEndian) {
10109     unsigned PFIndexes[4];
10110     bool isFourElementShuffle = true;
10111     for (unsigned i = 0; i != 4 && isFourElementShuffle;
10112          ++i) {                           // Element number
10113       unsigned EltNo = 8;                 // Start out undef.
10114       for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
10115         if (PermMask[i * 4 + j] < 0)
10116           continue; // Undef, ignore it.
10117 
10118         unsigned ByteSource = PermMask[i * 4 + j];
10119         if ((ByteSource & 3) != j) {
10120           isFourElementShuffle = false;
10121           break;
10122         }
10123 
10124         if (EltNo == 8) {
10125           EltNo = ByteSource / 4;
10126         } else if (EltNo != ByteSource / 4) {
10127           isFourElementShuffle = false;
10128           break;
10129         }
10130       }
10131       PFIndexes[i] = EltNo;
10132     }
10133 
10134     // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
10135     // perfect shuffle vector to determine if it is cost effective to do this as
10136     // discrete instructions, or whether we should use a vperm.
10137     // For now, we skip this for little endian until such time as we have a
10138     // little-endian perfect shuffle table.
10139     if (isFourElementShuffle) {
10140       // Compute the index in the perfect shuffle table.
10141       unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
10142                               PFIndexes[2] * 9 + PFIndexes[3];
10143 
10144       unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
10145       unsigned Cost = (PFEntry >> 30);
10146 
10147       // Determining when to avoid vperm is tricky.  Many things affect the cost
10148       // of vperm, particularly how many times the perm mask needs to be
10149       // computed. For example, if the perm mask can be hoisted out of a loop or
10150       // is already used (perhaps because there are multiple permutes with the
10151       // same shuffle mask?) the vperm has a cost of 1.  OTOH, hoisting the
10152       // permute mask out of the loop requires an extra register.
10153       //
10154       // As a compromise, we only emit discrete instructions if the shuffle can
10155       // be generated in 3 or fewer operations.  When we have loop information
10156       // available, if this block is within a loop, we should avoid using vperm
10157       // for 3-operation perms and use a constant pool load instead.
10158       if (Cost < 3)
10159         return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
10160     }
10161   }
10162 
10163   // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
10164   // vector that will get spilled to the constant pool.
10165   if (V2.isUndef()) V2 = V1;
10166 
10167   return LowerVPERM(Op, DAG, PermMask, VT, V1, V2);
10168 }
10169 
10170 SDValue PPCTargetLowering::LowerVPERM(SDValue Op, SelectionDAG &DAG,
10171                                       ArrayRef<int> PermMask, EVT VT,
10172                                       SDValue V1, SDValue V2) const {
10173   unsigned Opcode = PPCISD::VPERM;
10174   EVT ValType = V1.getValueType();
10175   SDLoc dl(Op);
10176   bool NeedSwap = false;
10177   bool isLittleEndian = Subtarget.isLittleEndian();
10178   bool isPPC64 = Subtarget.isPPC64();
10179 
10180   // Only need to place items backwards in LE,
10181   // the mask will be properly calculated.
10182   if (isLittleEndian)
10183     std::swap(V1, V2);
10184 
10185   if (Subtarget.isISA3_0() && (V1->hasOneUse() || V2->hasOneUse())) {
10186     LLVM_DEBUG(dbgs() << "At least one of two input vectors are dead - using "
10187                          "XXPERM instead\n");
10188     Opcode = PPCISD::XXPERM;
10189 
10190     // if V2 is dead, then we swap V1 and V2 so we can
10191     // use V2 as the destination instead.
10192     if (!V1->hasOneUse() && V2->hasOneUse()) {
10193       std::swap(V1, V2);
10194       NeedSwap = !NeedSwap;
10195     }
10196   }
10197 
10198   // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
10199   // that it is in input element units, not in bytes.  Convert now.
10200 
10201   // For little endian, the order of the input vectors is reversed, and
10202   // the permutation mask is complemented with respect to 31.  This is
10203   // necessary to produce proper semantics with the big-endian-based vperm
10204   // instruction.
10205   EVT EltVT = V1.getValueType().getVectorElementType();
10206   unsigned BytesPerElement = EltVT.getSizeInBits() / 8;
10207 
10208   bool V1HasXXSWAPD = V1->getOperand(0)->getOpcode() == PPCISD::XXSWAPD;
10209   bool V2HasXXSWAPD = V2->getOperand(0)->getOpcode() == PPCISD::XXSWAPD;
10210 
10211   /*
10212   Vectors will be appended like so: [ V1 | v2 ]
10213   XXSWAPD on V1:
10214   [   A   |   B   |   C   |   D   ] -> [   C   |   D   |   A   |   B   ]
10215      0-3     4-7     8-11   12-15         0-3     4-7     8-11   12-15
10216   i.e.  index of A, B += 8, and index of C, D -= 8.
10217   XXSWAPD on V2:
10218   [   E   |   F   |   G   |   H   ] -> [   G   |   H   |   E   |   F   ]
10219     16-19   20-23   24-27   28-31        16-19   20-23   24-27   28-31
10220   i.e.  index of E, F += 8, index of G, H -= 8
10221   Swap V1 and V2:
10222   [   V1   |   V2  ] -> [   V2   |   V1   ]
10223      0-15     16-31        0-15     16-31
10224   i.e.  index of V1 += 16, index of V2 -= 16
10225   */
10226 
10227   SmallVector<SDValue, 16> ResultMask;
10228   for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
10229     unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
10230 
10231     if (Opcode == PPCISD::XXPERM) {
10232       if (V1HasXXSWAPD) {
10233         if (SrcElt < 8)
10234           SrcElt += 8;
10235         else if (SrcElt < 16)
10236           SrcElt -= 8;
10237       }
10238       if (V2HasXXSWAPD) {
10239         if (SrcElt > 23)
10240           SrcElt -= 8;
10241         else if (SrcElt > 15)
10242           SrcElt += 8;
10243       }
10244       if (NeedSwap) {
10245         if (SrcElt < 16)
10246           SrcElt += 16;
10247         else
10248           SrcElt -= 16;
10249       }
10250     }
10251 
10252     for (unsigned j = 0; j != BytesPerElement; ++j)
10253       if (isLittleEndian)
10254         ResultMask.push_back(
10255             DAG.getConstant(31 - (SrcElt * BytesPerElement + j), dl, MVT::i32));
10256       else
10257         ResultMask.push_back(
10258             DAG.getConstant(SrcElt * BytesPerElement + j, dl, MVT::i32));
10259   }
10260 
10261   if (Opcode == PPCISD::XXPERM && (V1HasXXSWAPD || V2HasXXSWAPD)) {
10262     if (V1HasXXSWAPD) {
10263       dl = SDLoc(V1->getOperand(0));
10264       V1 = V1->getOperand(0)->getOperand(1);
10265     }
10266     if (V2HasXXSWAPD) {
10267       dl = SDLoc(V2->getOperand(0));
10268       V2 = V2->getOperand(0)->getOperand(1);
10269     }
10270     if (isPPC64 && ValType != MVT::v2f64)
10271       V1 = DAG.getBitcast(MVT::v2f64, V1);
10272     if (isPPC64 && V2.getValueType() != MVT::v2f64)
10273       V2 = DAG.getBitcast(MVT::v2f64, V2);
10274   }
10275 
10276   ShufflesHandledWithVPERM++;
10277   SDValue VPermMask = DAG.getBuildVector(MVT::v16i8, dl, ResultMask);
10278   LLVM_DEBUG({
10279     ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10280     if (Opcode == PPCISD::XXPERM) {
10281       dbgs() << "Emitting a XXPERM for the following shuffle:\n";
10282     } else {
10283       dbgs() << "Emitting a VPERM for the following shuffle:\n";
10284     }
10285     SVOp->dump();
10286     dbgs() << "With the following permute control vector:\n";
10287     VPermMask.dump();
10288   });
10289 
10290   if (Opcode == PPCISD::XXPERM)
10291     VPermMask = DAG.getBitcast(MVT::v4i32, VPermMask);
10292 
10293   SDValue VPERMNode =
10294       DAG.getNode(Opcode, dl, V1.getValueType(), V1, V2, VPermMask);
10295 
10296   VPERMNode = DAG.getBitcast(ValType, VPERMNode);
10297   return VPERMNode;
10298 }
10299 
10300 /// getVectorCompareInfo - Given an intrinsic, return false if it is not a
10301 /// vector comparison.  If it is, return true and fill in Opc/isDot with
10302 /// information about the intrinsic.
10303 static bool getVectorCompareInfo(SDValue Intrin, int &CompareOpc,
10304                                  bool &isDot, const PPCSubtarget &Subtarget) {
10305   unsigned IntrinsicID =
10306       cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
10307   CompareOpc = -1;
10308   isDot = false;
10309   switch (IntrinsicID) {
10310   default:
10311     return false;
10312   // Comparison predicates.
10313   case Intrinsic::ppc_altivec_vcmpbfp_p:
10314     CompareOpc = 966;
10315     isDot = true;
10316     break;
10317   case Intrinsic::ppc_altivec_vcmpeqfp_p:
10318     CompareOpc = 198;
10319     isDot = true;
10320     break;
10321   case Intrinsic::ppc_altivec_vcmpequb_p:
10322     CompareOpc = 6;
10323     isDot = true;
10324     break;
10325   case Intrinsic::ppc_altivec_vcmpequh_p:
10326     CompareOpc = 70;
10327     isDot = true;
10328     break;
10329   case Intrinsic::ppc_altivec_vcmpequw_p:
10330     CompareOpc = 134;
10331     isDot = true;
10332     break;
10333   case Intrinsic::ppc_altivec_vcmpequd_p:
10334     if (Subtarget.hasVSX() || Subtarget.hasP8Altivec()) {
10335       CompareOpc = 199;
10336       isDot = true;
10337     } else
10338       return false;
10339     break;
10340   case Intrinsic::ppc_altivec_vcmpneb_p:
10341   case Intrinsic::ppc_altivec_vcmpneh_p:
10342   case Intrinsic::ppc_altivec_vcmpnew_p:
10343   case Intrinsic::ppc_altivec_vcmpnezb_p:
10344   case Intrinsic::ppc_altivec_vcmpnezh_p:
10345   case Intrinsic::ppc_altivec_vcmpnezw_p:
10346     if (Subtarget.hasP9Altivec()) {
10347       switch (IntrinsicID) {
10348       default:
10349         llvm_unreachable("Unknown comparison intrinsic.");
10350       case Intrinsic::ppc_altivec_vcmpneb_p:
10351         CompareOpc = 7;
10352         break;
10353       case Intrinsic::ppc_altivec_vcmpneh_p:
10354         CompareOpc = 71;
10355         break;
10356       case Intrinsic::ppc_altivec_vcmpnew_p:
10357         CompareOpc = 135;
10358         break;
10359       case Intrinsic::ppc_altivec_vcmpnezb_p:
10360         CompareOpc = 263;
10361         break;
10362       case Intrinsic::ppc_altivec_vcmpnezh_p:
10363         CompareOpc = 327;
10364         break;
10365       case Intrinsic::ppc_altivec_vcmpnezw_p:
10366         CompareOpc = 391;
10367         break;
10368       }
10369       isDot = true;
10370     } else
10371       return false;
10372     break;
10373   case Intrinsic::ppc_altivec_vcmpgefp_p:
10374     CompareOpc = 454;
10375     isDot = true;
10376     break;
10377   case Intrinsic::ppc_altivec_vcmpgtfp_p:
10378     CompareOpc = 710;
10379     isDot = true;
10380     break;
10381   case Intrinsic::ppc_altivec_vcmpgtsb_p:
10382     CompareOpc = 774;
10383     isDot = true;
10384     break;
10385   case Intrinsic::ppc_altivec_vcmpgtsh_p:
10386     CompareOpc = 838;
10387     isDot = true;
10388     break;
10389   case Intrinsic::ppc_altivec_vcmpgtsw_p:
10390     CompareOpc = 902;
10391     isDot = true;
10392     break;
10393   case Intrinsic::ppc_altivec_vcmpgtsd_p:
10394     if (Subtarget.hasVSX() || Subtarget.hasP8Altivec()) {
10395       CompareOpc = 967;
10396       isDot = true;
10397     } else
10398       return false;
10399     break;
10400   case Intrinsic::ppc_altivec_vcmpgtub_p:
10401     CompareOpc = 518;
10402     isDot = true;
10403     break;
10404   case Intrinsic::ppc_altivec_vcmpgtuh_p:
10405     CompareOpc = 582;
10406     isDot = true;
10407     break;
10408   case Intrinsic::ppc_altivec_vcmpgtuw_p:
10409     CompareOpc = 646;
10410     isDot = true;
10411     break;
10412   case Intrinsic::ppc_altivec_vcmpgtud_p:
10413     if (Subtarget.hasVSX() || Subtarget.hasP8Altivec()) {
10414       CompareOpc = 711;
10415       isDot = true;
10416     } else
10417       return false;
10418     break;
10419 
10420   case Intrinsic::ppc_altivec_vcmpequq:
10421   case Intrinsic::ppc_altivec_vcmpgtsq:
10422   case Intrinsic::ppc_altivec_vcmpgtuq:
10423     if (!Subtarget.isISA3_1())
10424       return false;
10425     switch (IntrinsicID) {
10426     default:
10427       llvm_unreachable("Unknown comparison intrinsic.");
10428     case Intrinsic::ppc_altivec_vcmpequq:
10429       CompareOpc = 455;
10430       break;
10431     case Intrinsic::ppc_altivec_vcmpgtsq:
10432       CompareOpc = 903;
10433       break;
10434     case Intrinsic::ppc_altivec_vcmpgtuq:
10435       CompareOpc = 647;
10436       break;
10437     }
10438     break;
10439 
10440   // VSX predicate comparisons use the same infrastructure
10441   case Intrinsic::ppc_vsx_xvcmpeqdp_p:
10442   case Intrinsic::ppc_vsx_xvcmpgedp_p:
10443   case Intrinsic::ppc_vsx_xvcmpgtdp_p:
10444   case Intrinsic::ppc_vsx_xvcmpeqsp_p:
10445   case Intrinsic::ppc_vsx_xvcmpgesp_p:
10446   case Intrinsic::ppc_vsx_xvcmpgtsp_p:
10447     if (Subtarget.hasVSX()) {
10448       switch (IntrinsicID) {
10449       case Intrinsic::ppc_vsx_xvcmpeqdp_p:
10450         CompareOpc = 99;
10451         break;
10452       case Intrinsic::ppc_vsx_xvcmpgedp_p:
10453         CompareOpc = 115;
10454         break;
10455       case Intrinsic::ppc_vsx_xvcmpgtdp_p:
10456         CompareOpc = 107;
10457         break;
10458       case Intrinsic::ppc_vsx_xvcmpeqsp_p:
10459         CompareOpc = 67;
10460         break;
10461       case Intrinsic::ppc_vsx_xvcmpgesp_p:
10462         CompareOpc = 83;
10463         break;
10464       case Intrinsic::ppc_vsx_xvcmpgtsp_p:
10465         CompareOpc = 75;
10466         break;
10467       }
10468       isDot = true;
10469     } else
10470       return false;
10471     break;
10472 
10473   // Normal Comparisons.
10474   case Intrinsic::ppc_altivec_vcmpbfp:
10475     CompareOpc = 966;
10476     break;
10477   case Intrinsic::ppc_altivec_vcmpeqfp:
10478     CompareOpc = 198;
10479     break;
10480   case Intrinsic::ppc_altivec_vcmpequb:
10481     CompareOpc = 6;
10482     break;
10483   case Intrinsic::ppc_altivec_vcmpequh:
10484     CompareOpc = 70;
10485     break;
10486   case Intrinsic::ppc_altivec_vcmpequw:
10487     CompareOpc = 134;
10488     break;
10489   case Intrinsic::ppc_altivec_vcmpequd:
10490     if (Subtarget.hasP8Altivec())
10491       CompareOpc = 199;
10492     else
10493       return false;
10494     break;
10495   case Intrinsic::ppc_altivec_vcmpneb:
10496   case Intrinsic::ppc_altivec_vcmpneh:
10497   case Intrinsic::ppc_altivec_vcmpnew:
10498   case Intrinsic::ppc_altivec_vcmpnezb:
10499   case Intrinsic::ppc_altivec_vcmpnezh:
10500   case Intrinsic::ppc_altivec_vcmpnezw:
10501     if (Subtarget.hasP9Altivec())
10502       switch (IntrinsicID) {
10503       default:
10504         llvm_unreachable("Unknown comparison intrinsic.");
10505       case Intrinsic::ppc_altivec_vcmpneb:
10506         CompareOpc = 7;
10507         break;
10508       case Intrinsic::ppc_altivec_vcmpneh:
10509         CompareOpc = 71;
10510         break;
10511       case Intrinsic::ppc_altivec_vcmpnew:
10512         CompareOpc = 135;
10513         break;
10514       case Intrinsic::ppc_altivec_vcmpnezb:
10515         CompareOpc = 263;
10516         break;
10517       case Intrinsic::ppc_altivec_vcmpnezh:
10518         CompareOpc = 327;
10519         break;
10520       case Intrinsic::ppc_altivec_vcmpnezw:
10521         CompareOpc = 391;
10522         break;
10523       }
10524     else
10525       return false;
10526     break;
10527   case Intrinsic::ppc_altivec_vcmpgefp:
10528     CompareOpc = 454;
10529     break;
10530   case Intrinsic::ppc_altivec_vcmpgtfp:
10531     CompareOpc = 710;
10532     break;
10533   case Intrinsic::ppc_altivec_vcmpgtsb:
10534     CompareOpc = 774;
10535     break;
10536   case Intrinsic::ppc_altivec_vcmpgtsh:
10537     CompareOpc = 838;
10538     break;
10539   case Intrinsic::ppc_altivec_vcmpgtsw:
10540     CompareOpc = 902;
10541     break;
10542   case Intrinsic::ppc_altivec_vcmpgtsd:
10543     if (Subtarget.hasP8Altivec())
10544       CompareOpc = 967;
10545     else
10546       return false;
10547     break;
10548   case Intrinsic::ppc_altivec_vcmpgtub:
10549     CompareOpc = 518;
10550     break;
10551   case Intrinsic::ppc_altivec_vcmpgtuh:
10552     CompareOpc = 582;
10553     break;
10554   case Intrinsic::ppc_altivec_vcmpgtuw:
10555     CompareOpc = 646;
10556     break;
10557   case Intrinsic::ppc_altivec_vcmpgtud:
10558     if (Subtarget.hasP8Altivec())
10559       CompareOpc = 711;
10560     else
10561       return false;
10562     break;
10563   case Intrinsic::ppc_altivec_vcmpequq_p:
10564   case Intrinsic::ppc_altivec_vcmpgtsq_p:
10565   case Intrinsic::ppc_altivec_vcmpgtuq_p:
10566     if (!Subtarget.isISA3_1())
10567       return false;
10568     switch (IntrinsicID) {
10569     default:
10570       llvm_unreachable("Unknown comparison intrinsic.");
10571     case Intrinsic::ppc_altivec_vcmpequq_p:
10572       CompareOpc = 455;
10573       break;
10574     case Intrinsic::ppc_altivec_vcmpgtsq_p:
10575       CompareOpc = 903;
10576       break;
10577     case Intrinsic::ppc_altivec_vcmpgtuq_p:
10578       CompareOpc = 647;
10579       break;
10580     }
10581     isDot = true;
10582     break;
10583   }
10584   return true;
10585 }
10586 
10587 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
10588 /// lower, do it, otherwise return null.
10589 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
10590                                                    SelectionDAG &DAG) const {
10591   unsigned IntrinsicID =
10592     cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
10593 
10594   SDLoc dl(Op);
10595 
10596   switch (IntrinsicID) {
10597   case Intrinsic::thread_pointer:
10598     // Reads the thread pointer register, used for __builtin_thread_pointer.
10599     if (Subtarget.isPPC64())
10600       return DAG.getRegister(PPC::X13, MVT::i64);
10601     return DAG.getRegister(PPC::R2, MVT::i32);
10602 
10603   case Intrinsic::ppc_mma_disassemble_acc: {
10604     if (Subtarget.isISAFuture()) {
10605       EVT ReturnTypes[] = {MVT::v256i1, MVT::v256i1};
10606       SDValue WideVec = SDValue(DAG.getMachineNode(PPC::DMXXEXTFDMR512, dl,
10607                                                    ArrayRef(ReturnTypes, 2),
10608                                                    Op.getOperand(1)),
10609                                 0);
10610       SmallVector<SDValue, 4> RetOps;
10611       SDValue Value = SDValue(WideVec.getNode(), 0);
10612       SDValue Value2 = SDValue(WideVec.getNode(), 1);
10613 
10614       SDValue Extract;
10615       Extract = DAG.getNode(
10616           PPCISD::EXTRACT_VSX_REG, dl, MVT::v16i8,
10617           Subtarget.isLittleEndian() ? Value2 : Value,
10618           DAG.getConstant(Subtarget.isLittleEndian() ? 1 : 0,
10619                           dl, getPointerTy(DAG.getDataLayout())));
10620       RetOps.push_back(Extract);
10621       Extract = DAG.getNode(
10622           PPCISD::EXTRACT_VSX_REG, dl, MVT::v16i8,
10623           Subtarget.isLittleEndian() ? Value2 : Value,
10624           DAG.getConstant(Subtarget.isLittleEndian() ? 0 : 1,
10625                           dl, getPointerTy(DAG.getDataLayout())));
10626       RetOps.push_back(Extract);
10627       Extract = DAG.getNode(
10628           PPCISD::EXTRACT_VSX_REG, dl, MVT::v16i8,
10629           Subtarget.isLittleEndian() ? Value : Value2,
10630           DAG.getConstant(Subtarget.isLittleEndian() ? 1 : 0,
10631                           dl, getPointerTy(DAG.getDataLayout())));
10632       RetOps.push_back(Extract);
10633       Extract = DAG.getNode(
10634           PPCISD::EXTRACT_VSX_REG, dl, MVT::v16i8,
10635           Subtarget.isLittleEndian() ? Value : Value2,
10636           DAG.getConstant(Subtarget.isLittleEndian() ? 0 : 1,
10637                           dl, getPointerTy(DAG.getDataLayout())));
10638       RetOps.push_back(Extract);
10639       return DAG.getMergeValues(RetOps, dl);
10640     }
10641     LLVM_FALLTHROUGH;
10642   }
10643   case Intrinsic::ppc_vsx_disassemble_pair: {
10644     int NumVecs = 2;
10645     SDValue WideVec = Op.getOperand(1);
10646     if (IntrinsicID == Intrinsic::ppc_mma_disassemble_acc) {
10647       NumVecs = 4;
10648       WideVec = DAG.getNode(PPCISD::XXMFACC, dl, MVT::v512i1, WideVec);
10649     }
10650     SmallVector<SDValue, 4> RetOps;
10651     for (int VecNo = 0; VecNo < NumVecs; VecNo++) {
10652       SDValue Extract = DAG.getNode(
10653           PPCISD::EXTRACT_VSX_REG, dl, MVT::v16i8, WideVec,
10654           DAG.getConstant(Subtarget.isLittleEndian() ? NumVecs - 1 - VecNo
10655                                                      : VecNo,
10656                           dl, getPointerTy(DAG.getDataLayout())));
10657       RetOps.push_back(Extract);
10658     }
10659     return DAG.getMergeValues(RetOps, dl);
10660   }
10661 
10662   case Intrinsic::ppc_unpack_longdouble: {
10663     auto *Idx = dyn_cast<ConstantSDNode>(Op.getOperand(2));
10664     assert(Idx && (Idx->getSExtValue() == 0 || Idx->getSExtValue() == 1) &&
10665            "Argument of long double unpack must be 0 or 1!");
10666     return DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::f64, Op.getOperand(1),
10667                        DAG.getConstant(!!(Idx->getSExtValue()), dl,
10668                                        Idx->getValueType(0)));
10669   }
10670 
10671   case Intrinsic::ppc_compare_exp_lt:
10672   case Intrinsic::ppc_compare_exp_gt:
10673   case Intrinsic::ppc_compare_exp_eq:
10674   case Intrinsic::ppc_compare_exp_uo: {
10675     unsigned Pred;
10676     switch (IntrinsicID) {
10677     case Intrinsic::ppc_compare_exp_lt:
10678       Pred = PPC::PRED_LT;
10679       break;
10680     case Intrinsic::ppc_compare_exp_gt:
10681       Pred = PPC::PRED_GT;
10682       break;
10683     case Intrinsic::ppc_compare_exp_eq:
10684       Pred = PPC::PRED_EQ;
10685       break;
10686     case Intrinsic::ppc_compare_exp_uo:
10687       Pred = PPC::PRED_UN;
10688       break;
10689     }
10690     return SDValue(
10691         DAG.getMachineNode(
10692             PPC::SELECT_CC_I4, dl, MVT::i32,
10693             {SDValue(DAG.getMachineNode(PPC::XSCMPEXPDP, dl, MVT::i32,
10694                                         Op.getOperand(1), Op.getOperand(2)),
10695                      0),
10696              DAG.getConstant(1, dl, MVT::i32), DAG.getConstant(0, dl, MVT::i32),
10697              DAG.getTargetConstant(Pred, dl, MVT::i32)}),
10698         0);
10699   }
10700   case Intrinsic::ppc_test_data_class: {
10701     EVT OpVT = Op.getOperand(1).getValueType();
10702     unsigned CmprOpc = OpVT == MVT::f128 ? PPC::XSTSTDCQP
10703                                          : (OpVT == MVT::f64 ? PPC::XSTSTDCDP
10704                                                              : PPC::XSTSTDCSP);
10705     return SDValue(
10706         DAG.getMachineNode(
10707             PPC::SELECT_CC_I4, dl, MVT::i32,
10708             {SDValue(DAG.getMachineNode(CmprOpc, dl, MVT::i32, Op.getOperand(2),
10709                                         Op.getOperand(1)),
10710                      0),
10711              DAG.getConstant(1, dl, MVT::i32), DAG.getConstant(0, dl, MVT::i32),
10712              DAG.getTargetConstant(PPC::PRED_EQ, dl, MVT::i32)}),
10713         0);
10714   }
10715   case Intrinsic::ppc_fnmsub: {
10716     EVT VT = Op.getOperand(1).getValueType();
10717     if (!Subtarget.hasVSX() || (!Subtarget.hasFloat128() && VT == MVT::f128))
10718       return DAG.getNode(
10719           ISD::FNEG, dl, VT,
10720           DAG.getNode(ISD::FMA, dl, VT, Op.getOperand(1), Op.getOperand(2),
10721                       DAG.getNode(ISD::FNEG, dl, VT, Op.getOperand(3))));
10722     return DAG.getNode(PPCISD::FNMSUB, dl, VT, Op.getOperand(1),
10723                        Op.getOperand(2), Op.getOperand(3));
10724   }
10725   case Intrinsic::ppc_convert_f128_to_ppcf128:
10726   case Intrinsic::ppc_convert_ppcf128_to_f128: {
10727     RTLIB::Libcall LC = IntrinsicID == Intrinsic::ppc_convert_ppcf128_to_f128
10728                             ? RTLIB::CONVERT_PPCF128_F128
10729                             : RTLIB::CONVERT_F128_PPCF128;
10730     MakeLibCallOptions CallOptions;
10731     std::pair<SDValue, SDValue> Result =
10732         makeLibCall(DAG, LC, Op.getValueType(), Op.getOperand(1), CallOptions,
10733                     dl, SDValue());
10734     return Result.first;
10735   }
10736   case Intrinsic::ppc_maxfe:
10737   case Intrinsic::ppc_maxfl:
10738   case Intrinsic::ppc_maxfs:
10739   case Intrinsic::ppc_minfe:
10740   case Intrinsic::ppc_minfl:
10741   case Intrinsic::ppc_minfs: {
10742     EVT VT = Op.getValueType();
10743     assert(
10744         all_of(Op->ops().drop_front(4),
10745                [VT](const SDUse &Use) { return Use.getValueType() == VT; }) &&
10746         "ppc_[max|min]f[e|l|s] must have uniform type arguments");
10747     (void)VT;
10748     ISD::CondCode CC = ISD::SETGT;
10749     if (IntrinsicID == Intrinsic::ppc_minfe ||
10750         IntrinsicID == Intrinsic::ppc_minfl ||
10751         IntrinsicID == Intrinsic::ppc_minfs)
10752       CC = ISD::SETLT;
10753     unsigned I = Op.getNumOperands() - 2, Cnt = I;
10754     SDValue Res = Op.getOperand(I);
10755     for (--I; Cnt != 0; --Cnt, I = (--I == 0 ? (Op.getNumOperands() - 1) : I)) {
10756       Res =
10757           DAG.getSelectCC(dl, Res, Op.getOperand(I), Res, Op.getOperand(I), CC);
10758     }
10759     return Res;
10760   }
10761   }
10762 
10763   // If this is a lowered altivec predicate compare, CompareOpc is set to the
10764   // opcode number of the comparison.
10765   int CompareOpc;
10766   bool isDot;
10767   if (!getVectorCompareInfo(Op, CompareOpc, isDot, Subtarget))
10768     return SDValue();    // Don't custom lower most intrinsics.
10769 
10770   // If this is a non-dot comparison, make the VCMP node and we are done.
10771   if (!isDot) {
10772     SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
10773                               Op.getOperand(1), Op.getOperand(2),
10774                               DAG.getConstant(CompareOpc, dl, MVT::i32));
10775     return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
10776   }
10777 
10778   // Create the PPCISD altivec 'dot' comparison node.
10779   SDValue Ops[] = {
10780     Op.getOperand(2),  // LHS
10781     Op.getOperand(3),  // RHS
10782     DAG.getConstant(CompareOpc, dl, MVT::i32)
10783   };
10784   EVT VTs[] = { Op.getOperand(2).getValueType(), MVT::Glue };
10785   SDValue CompNode = DAG.getNode(PPCISD::VCMP_rec, dl, VTs, Ops);
10786 
10787   // Now that we have the comparison, emit a copy from the CR to a GPR.
10788   // This is flagged to the above dot comparison.
10789   SDValue Flags = DAG.getNode(PPCISD::MFOCRF, dl, MVT::i32,
10790                                 DAG.getRegister(PPC::CR6, MVT::i32),
10791                                 CompNode.getValue(1));
10792 
10793   // Unpack the result based on how the target uses it.
10794   unsigned BitNo;   // Bit # of CR6.
10795   bool InvertBit;   // Invert result?
10796   switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
10797   default:  // Can't happen, don't crash on invalid number though.
10798   case 0:   // Return the value of the EQ bit of CR6.
10799     BitNo = 0; InvertBit = false;
10800     break;
10801   case 1:   // Return the inverted value of the EQ bit of CR6.
10802     BitNo = 0; InvertBit = true;
10803     break;
10804   case 2:   // Return the value of the LT bit of CR6.
10805     BitNo = 2; InvertBit = false;
10806     break;
10807   case 3:   // Return the inverted value of the LT bit of CR6.
10808     BitNo = 2; InvertBit = true;
10809     break;
10810   }
10811 
10812   // Shift the bit into the low position.
10813   Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
10814                       DAG.getConstant(8 - (3 - BitNo), dl, MVT::i32));
10815   // Isolate the bit.
10816   Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
10817                       DAG.getConstant(1, dl, MVT::i32));
10818 
10819   // If we are supposed to, toggle the bit.
10820   if (InvertBit)
10821     Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
10822                         DAG.getConstant(1, dl, MVT::i32));
10823   return Flags;
10824 }
10825 
10826 SDValue PPCTargetLowering::LowerINTRINSIC_VOID(SDValue Op,
10827                                                SelectionDAG &DAG) const {
10828   // SelectionDAGBuilder::visitTargetIntrinsic may insert one extra chain to
10829   // the beginning of the argument list.
10830   int ArgStart = isa<ConstantSDNode>(Op.getOperand(0)) ? 0 : 1;
10831   SDLoc DL(Op);
10832   switch (cast<ConstantSDNode>(Op.getOperand(ArgStart))->getZExtValue()) {
10833   case Intrinsic::ppc_cfence: {
10834     assert(ArgStart == 1 && "llvm.ppc.cfence must carry a chain argument.");
10835     assert(Subtarget.isPPC64() && "Only 64-bit is supported for now.");
10836     SDValue Val = Op.getOperand(ArgStart + 1);
10837     EVT Ty = Val.getValueType();
10838     if (Ty == MVT::i128) {
10839       // FIXME: Testing one of two paired registers is sufficient to guarantee
10840       // ordering?
10841       Val = DAG.getNode(ISD::TRUNCATE, DL, MVT::i64, Val);
10842     }
10843     return SDValue(
10844         DAG.getMachineNode(PPC::CFENCE8, DL, MVT::Other,
10845                            DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Val),
10846                            Op.getOperand(0)),
10847         0);
10848   }
10849   default:
10850     break;
10851   }
10852   return SDValue();
10853 }
10854 
10855 // Lower scalar BSWAP64 to xxbrd.
10856 SDValue PPCTargetLowering::LowerBSWAP(SDValue Op, SelectionDAG &DAG) const {
10857   SDLoc dl(Op);
10858   if (!Subtarget.isPPC64())
10859     return Op;
10860   // MTVSRDD
10861   Op = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64, Op.getOperand(0),
10862                    Op.getOperand(0));
10863   // XXBRD
10864   Op = DAG.getNode(ISD::BSWAP, dl, MVT::v2i64, Op);
10865   // MFVSRD
10866   int VectorIndex = 0;
10867   if (Subtarget.isLittleEndian())
10868     VectorIndex = 1;
10869   Op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Op,
10870                    DAG.getTargetConstant(VectorIndex, dl, MVT::i32));
10871   return Op;
10872 }
10873 
10874 // ATOMIC_CMP_SWAP for i8/i16 needs to zero-extend its input since it will be
10875 // compared to a value that is atomically loaded (atomic loads zero-extend).
10876 SDValue PPCTargetLowering::LowerATOMIC_CMP_SWAP(SDValue Op,
10877                                                 SelectionDAG &DAG) const {
10878   assert(Op.getOpcode() == ISD::ATOMIC_CMP_SWAP &&
10879          "Expecting an atomic compare-and-swap here.");
10880   SDLoc dl(Op);
10881   auto *AtomicNode = cast<AtomicSDNode>(Op.getNode());
10882   EVT MemVT = AtomicNode->getMemoryVT();
10883   if (MemVT.getSizeInBits() >= 32)
10884     return Op;
10885 
10886   SDValue CmpOp = Op.getOperand(2);
10887   // If this is already correctly zero-extended, leave it alone.
10888   auto HighBits = APInt::getHighBitsSet(32, 32 - MemVT.getSizeInBits());
10889   if (DAG.MaskedValueIsZero(CmpOp, HighBits))
10890     return Op;
10891 
10892   // Clear the high bits of the compare operand.
10893   unsigned MaskVal = (1 << MemVT.getSizeInBits()) - 1;
10894   SDValue NewCmpOp =
10895     DAG.getNode(ISD::AND, dl, MVT::i32, CmpOp,
10896                 DAG.getConstant(MaskVal, dl, MVT::i32));
10897 
10898   // Replace the existing compare operand with the properly zero-extended one.
10899   SmallVector<SDValue, 4> Ops;
10900   for (int i = 0, e = AtomicNode->getNumOperands(); i < e; i++)
10901     Ops.push_back(AtomicNode->getOperand(i));
10902   Ops[2] = NewCmpOp;
10903   MachineMemOperand *MMO = AtomicNode->getMemOperand();
10904   SDVTList Tys = DAG.getVTList(MVT::i32, MVT::Other);
10905   auto NodeTy =
10906     (MemVT == MVT::i8) ? PPCISD::ATOMIC_CMP_SWAP_8 : PPCISD::ATOMIC_CMP_SWAP_16;
10907   return DAG.getMemIntrinsicNode(NodeTy, dl, Tys, Ops, MemVT, MMO);
10908 }
10909 
10910 SDValue PPCTargetLowering::LowerATOMIC_LOAD_STORE(SDValue Op,
10911                                                   SelectionDAG &DAG) const {
10912   AtomicSDNode *N = cast<AtomicSDNode>(Op.getNode());
10913   EVT MemVT = N->getMemoryVT();
10914   assert(MemVT.getSimpleVT() == MVT::i128 &&
10915          "Expect quadword atomic operations");
10916   SDLoc dl(N);
10917   unsigned Opc = N->getOpcode();
10918   switch (Opc) {
10919   case ISD::ATOMIC_LOAD: {
10920     // Lower quadword atomic load to int_ppc_atomic_load_i128 which will be
10921     // lowered to ppc instructions by pattern matching instruction selector.
10922     SDVTList Tys = DAG.getVTList(MVT::i64, MVT::i64, MVT::Other);
10923     SmallVector<SDValue, 4> Ops{
10924         N->getOperand(0),
10925         DAG.getConstant(Intrinsic::ppc_atomic_load_i128, dl, MVT::i32)};
10926     for (int I = 1, E = N->getNumOperands(); I < E; ++I)
10927       Ops.push_back(N->getOperand(I));
10928     SDValue LoadedVal = DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl, Tys,
10929                                                 Ops, MemVT, N->getMemOperand());
10930     SDValue ValLo = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i128, LoadedVal);
10931     SDValue ValHi =
10932         DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i128, LoadedVal.getValue(1));
10933     ValHi = DAG.getNode(ISD::SHL, dl, MVT::i128, ValHi,
10934                         DAG.getConstant(64, dl, MVT::i32));
10935     SDValue Val =
10936         DAG.getNode(ISD::OR, dl, {MVT::i128, MVT::Other}, {ValLo, ValHi});
10937     return DAG.getNode(ISD::MERGE_VALUES, dl, {MVT::i128, MVT::Other},
10938                        {Val, LoadedVal.getValue(2)});
10939   }
10940   case ISD::ATOMIC_STORE: {
10941     // Lower quadword atomic store to int_ppc_atomic_store_i128 which will be
10942     // lowered to ppc instructions by pattern matching instruction selector.
10943     SDVTList Tys = DAG.getVTList(MVT::Other);
10944     SmallVector<SDValue, 4> Ops{
10945         N->getOperand(0),
10946         DAG.getConstant(Intrinsic::ppc_atomic_store_i128, dl, MVT::i32)};
10947     SDValue Val = N->getOperand(2);
10948     SDValue ValLo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i64, Val);
10949     SDValue ValHi = DAG.getNode(ISD::SRL, dl, MVT::i128, Val,
10950                                 DAG.getConstant(64, dl, MVT::i32));
10951     ValHi = DAG.getNode(ISD::TRUNCATE, dl, MVT::i64, ValHi);
10952     Ops.push_back(ValLo);
10953     Ops.push_back(ValHi);
10954     Ops.push_back(N->getOperand(1));
10955     return DAG.getMemIntrinsicNode(ISD::INTRINSIC_VOID, dl, Tys, Ops, MemVT,
10956                                    N->getMemOperand());
10957   }
10958   default:
10959     llvm_unreachable("Unexpected atomic opcode");
10960   }
10961 }
10962 
10963 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
10964                                                  SelectionDAG &DAG) const {
10965   SDLoc dl(Op);
10966   // Create a stack slot that is 16-byte aligned.
10967   MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
10968   int FrameIdx = MFI.CreateStackObject(16, Align(16), false);
10969   EVT PtrVT = getPointerTy(DAG.getDataLayout());
10970   SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
10971 
10972   // Store the input value into Value#0 of the stack slot.
10973   SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0), FIdx,
10974                                MachinePointerInfo());
10975   // Load it out.
10976   return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo());
10977 }
10978 
10979 SDValue PPCTargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
10980                                                   SelectionDAG &DAG) const {
10981   assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT &&
10982          "Should only be called for ISD::INSERT_VECTOR_ELT");
10983 
10984   ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(2));
10985 
10986   EVT VT = Op.getValueType();
10987   SDLoc dl(Op);
10988   SDValue V1 = Op.getOperand(0);
10989   SDValue V2 = Op.getOperand(1);
10990 
10991   if (VT == MVT::v2f64 && C)
10992     return Op;
10993 
10994   if (Subtarget.hasP9Vector()) {
10995     // A f32 load feeding into a v4f32 insert_vector_elt is handled in this way
10996     // because on P10, it allows this specific insert_vector_elt load pattern to
10997     // utilize the refactored load and store infrastructure in order to exploit
10998     // prefixed loads.
10999     // On targets with inexpensive direct moves (Power9 and up), a
11000     // (insert_vector_elt v4f32:$vec, (f32 load)) is always better as an integer
11001     // load since a single precision load will involve conversion to double
11002     // precision on the load followed by another conversion to single precision.
11003     if ((VT == MVT::v4f32) && (V2.getValueType() == MVT::f32) &&
11004         (isa<LoadSDNode>(V2))) {
11005       SDValue BitcastVector = DAG.getBitcast(MVT::v4i32, V1);
11006       SDValue BitcastLoad = DAG.getBitcast(MVT::i32, V2);
11007       SDValue InsVecElt =
11008           DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v4i32, BitcastVector,
11009                       BitcastLoad, Op.getOperand(2));
11010       return DAG.getBitcast(MVT::v4f32, InsVecElt);
11011     }
11012   }
11013 
11014   if (Subtarget.isISA3_1()) {
11015     if ((VT == MVT::v2i64 || VT == MVT::v2f64) && !Subtarget.isPPC64())
11016       return SDValue();
11017     // On P10, we have legal lowering for constant and variable indices for
11018     // all vectors.
11019     if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
11020         VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
11021       return Op;
11022   }
11023 
11024   // Before P10, we have legal lowering for constant indices but not for
11025   // variable ones.
11026   if (!C)
11027     return SDValue();
11028 
11029   // We can use MTVSRZ + VECINSERT for v8i16 and v16i8 types.
11030   if (VT == MVT::v8i16 || VT == MVT::v16i8) {
11031     SDValue Mtvsrz = DAG.getNode(PPCISD::MTVSRZ, dl, VT, V2);
11032     unsigned BytesInEachElement = VT.getVectorElementType().getSizeInBits() / 8;
11033     unsigned InsertAtElement = C->getZExtValue();
11034     unsigned InsertAtByte = InsertAtElement * BytesInEachElement;
11035     if (Subtarget.isLittleEndian()) {
11036       InsertAtByte = (16 - BytesInEachElement) - InsertAtByte;
11037     }
11038     return DAG.getNode(PPCISD::VECINSERT, dl, VT, V1, Mtvsrz,
11039                        DAG.getConstant(InsertAtByte, dl, MVT::i32));
11040   }
11041   return Op;
11042 }
11043 
11044 SDValue PPCTargetLowering::LowerVectorLoad(SDValue Op,
11045                                            SelectionDAG &DAG) const {
11046   SDLoc dl(Op);
11047   LoadSDNode *LN = cast<LoadSDNode>(Op.getNode());
11048   SDValue LoadChain = LN->getChain();
11049   SDValue BasePtr = LN->getBasePtr();
11050   EVT VT = Op.getValueType();
11051 
11052   if (VT != MVT::v256i1 && VT != MVT::v512i1)
11053     return Op;
11054 
11055   // Type v256i1 is used for pairs and v512i1 is used for accumulators.
11056   // Here we create 2 or 4 v16i8 loads to load the pair or accumulator value in
11057   // 2 or 4 vsx registers.
11058   assert((VT != MVT::v512i1 || Subtarget.hasMMA()) &&
11059          "Type unsupported without MMA");
11060   assert((VT != MVT::v256i1 || Subtarget.pairedVectorMemops()) &&
11061          "Type unsupported without paired vector support");
11062   Align Alignment = LN->getAlign();
11063   SmallVector<SDValue, 4> Loads;
11064   SmallVector<SDValue, 4> LoadChains;
11065   unsigned NumVecs = VT.getSizeInBits() / 128;
11066   for (unsigned Idx = 0; Idx < NumVecs; ++Idx) {
11067     SDValue Load =
11068         DAG.getLoad(MVT::v16i8, dl, LoadChain, BasePtr,
11069                     LN->getPointerInfo().getWithOffset(Idx * 16),
11070                     commonAlignment(Alignment, Idx * 16),
11071                     LN->getMemOperand()->getFlags(), LN->getAAInfo());
11072     BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
11073                           DAG.getConstant(16, dl, BasePtr.getValueType()));
11074     Loads.push_back(Load);
11075     LoadChains.push_back(Load.getValue(1));
11076   }
11077   if (Subtarget.isLittleEndian()) {
11078     std::reverse(Loads.begin(), Loads.end());
11079     std::reverse(LoadChains.begin(), LoadChains.end());
11080   }
11081   SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
11082   SDValue Value =
11083       DAG.getNode(VT == MVT::v512i1 ? PPCISD::ACC_BUILD : PPCISD::PAIR_BUILD,
11084                   dl, VT, Loads);
11085   SDValue RetOps[] = {Value, TF};
11086   return DAG.getMergeValues(RetOps, dl);
11087 }
11088 
11089 SDValue PPCTargetLowering::LowerVectorStore(SDValue Op,
11090                                             SelectionDAG &DAG) const {
11091   SDLoc dl(Op);
11092   StoreSDNode *SN = cast<StoreSDNode>(Op.getNode());
11093   SDValue StoreChain = SN->getChain();
11094   SDValue BasePtr = SN->getBasePtr();
11095   SDValue Value = SN->getValue();
11096   SDValue Value2 = SN->getValue();
11097   EVT StoreVT = Value.getValueType();
11098 
11099   if (StoreVT != MVT::v256i1 && StoreVT != MVT::v512i1)
11100     return Op;
11101 
11102   // Type v256i1 is used for pairs and v512i1 is used for accumulators.
11103   // Here we create 2 or 4 v16i8 stores to store the pair or accumulator
11104   // underlying registers individually.
11105   assert((StoreVT != MVT::v512i1 || Subtarget.hasMMA()) &&
11106          "Type unsupported without MMA");
11107   assert((StoreVT != MVT::v256i1 || Subtarget.pairedVectorMemops()) &&
11108          "Type unsupported without paired vector support");
11109   Align Alignment = SN->getAlign();
11110   SmallVector<SDValue, 4> Stores;
11111   unsigned NumVecs = 2;
11112   if (StoreVT == MVT::v512i1) {
11113     if (Subtarget.isISAFuture()) {
11114       EVT ReturnTypes[] = {MVT::v256i1, MVT::v256i1};
11115       MachineSDNode *ExtNode = DAG.getMachineNode(
11116           PPC::DMXXEXTFDMR512, dl, ArrayRef(ReturnTypes, 2), Op.getOperand(1));
11117 
11118       Value = SDValue(ExtNode, 0);
11119       Value2 = SDValue(ExtNode, 1);
11120     } else
11121       Value = DAG.getNode(PPCISD::XXMFACC, dl, MVT::v512i1, Value);
11122     NumVecs = 4;
11123   }
11124   for (unsigned Idx = 0; Idx < NumVecs; ++Idx) {
11125     unsigned VecNum = Subtarget.isLittleEndian() ? NumVecs - 1 - Idx : Idx;
11126     SDValue Elt;
11127     if (Subtarget.isISAFuture()) {
11128       VecNum = Subtarget.isLittleEndian() ? 1 - (Idx % 2) : (Idx % 2);
11129       Elt = DAG.getNode(PPCISD::EXTRACT_VSX_REG, dl, MVT::v16i8,
11130                         Idx > 1 ? Value2 : Value,
11131                         DAG.getConstant(VecNum, dl, getPointerTy(DAG.getDataLayout())));
11132     } else
11133       Elt = DAG.getNode(PPCISD::EXTRACT_VSX_REG, dl, MVT::v16i8, Value,
11134                         DAG.getConstant(VecNum, dl, getPointerTy(DAG.getDataLayout())));
11135 
11136     SDValue Store =
11137         DAG.getStore(StoreChain, dl, Elt, BasePtr,
11138                      SN->getPointerInfo().getWithOffset(Idx * 16),
11139                      commonAlignment(Alignment, Idx * 16),
11140                      SN->getMemOperand()->getFlags(), SN->getAAInfo());
11141     BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
11142                           DAG.getConstant(16, dl, BasePtr.getValueType()));
11143     Stores.push_back(Store);
11144   }
11145   SDValue TF = DAG.getTokenFactor(dl, Stores);
11146   return TF;
11147 }
11148 
11149 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
11150   SDLoc dl(Op);
11151   if (Op.getValueType() == MVT::v4i32) {
11152     SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
11153 
11154     SDValue Zero = getCanonicalConstSplat(0, 1, MVT::v4i32, DAG, dl);
11155     // +16 as shift amt.
11156     SDValue Neg16 = getCanonicalConstSplat(-16, 4, MVT::v4i32, DAG, dl);
11157     SDValue RHSSwap =   // = vrlw RHS, 16
11158       BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
11159 
11160     // Shrinkify inputs to v8i16.
11161     LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
11162     RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
11163     RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
11164 
11165     // Low parts multiplied together, generating 32-bit results (we ignore the
11166     // top parts).
11167     SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
11168                                         LHS, RHS, DAG, dl, MVT::v4i32);
11169 
11170     SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
11171                                       LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
11172     // Shift the high parts up 16 bits.
11173     HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
11174                               Neg16, DAG, dl);
11175     return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
11176   } else if (Op.getValueType() == MVT::v16i8) {
11177     SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
11178     bool isLittleEndian = Subtarget.isLittleEndian();
11179 
11180     // Multiply the even 8-bit parts, producing 16-bit sums.
11181     SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
11182                                            LHS, RHS, DAG, dl, MVT::v8i16);
11183     EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
11184 
11185     // Multiply the odd 8-bit parts, producing 16-bit sums.
11186     SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
11187                                           LHS, RHS, DAG, dl, MVT::v8i16);
11188     OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
11189 
11190     // Merge the results together.  Because vmuleub and vmuloub are
11191     // instructions with a big-endian bias, we must reverse the
11192     // element numbering and reverse the meaning of "odd" and "even"
11193     // when generating little endian code.
11194     int Ops[16];
11195     for (unsigned i = 0; i != 8; ++i) {
11196       if (isLittleEndian) {
11197         Ops[i*2  ] = 2*i;
11198         Ops[i*2+1] = 2*i+16;
11199       } else {
11200         Ops[i*2  ] = 2*i+1;
11201         Ops[i*2+1] = 2*i+1+16;
11202       }
11203     }
11204     if (isLittleEndian)
11205       return DAG.getVectorShuffle(MVT::v16i8, dl, OddParts, EvenParts, Ops);
11206     else
11207       return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
11208   } else {
11209     llvm_unreachable("Unknown mul to lower!");
11210   }
11211 }
11212 
11213 SDValue PPCTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
11214   bool IsStrict = Op->isStrictFPOpcode();
11215   if (Op.getOperand(IsStrict ? 1 : 0).getValueType() == MVT::f128 &&
11216       !Subtarget.hasP9Vector())
11217     return SDValue();
11218 
11219   return Op;
11220 }
11221 
11222 // Custom lowering for fpext vf32 to v2f64
11223 SDValue PPCTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
11224 
11225   assert(Op.getOpcode() == ISD::FP_EXTEND &&
11226          "Should only be called for ISD::FP_EXTEND");
11227 
11228   // FIXME: handle extends from half precision float vectors on P9.
11229   // We only want to custom lower an extend from v2f32 to v2f64.
11230   if (Op.getValueType() != MVT::v2f64 ||
11231       Op.getOperand(0).getValueType() != MVT::v2f32)
11232     return SDValue();
11233 
11234   SDLoc dl(Op);
11235   SDValue Op0 = Op.getOperand(0);
11236 
11237   switch (Op0.getOpcode()) {
11238   default:
11239     return SDValue();
11240   case ISD::EXTRACT_SUBVECTOR: {
11241     assert(Op0.getNumOperands() == 2 &&
11242            isa<ConstantSDNode>(Op0->getOperand(1)) &&
11243            "Node should have 2 operands with second one being a constant!");
11244 
11245     if (Op0.getOperand(0).getValueType() != MVT::v4f32)
11246       return SDValue();
11247 
11248     // Custom lower is only done for high or low doubleword.
11249     int Idx = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
11250     if (Idx % 2 != 0)
11251       return SDValue();
11252 
11253     // Since input is v4f32, at this point Idx is either 0 or 2.
11254     // Shift to get the doubleword position we want.
11255     int DWord = Idx >> 1;
11256 
11257     // High and low word positions are different on little endian.
11258     if (Subtarget.isLittleEndian())
11259       DWord ^= 0x1;
11260 
11261     return DAG.getNode(PPCISD::FP_EXTEND_HALF, dl, MVT::v2f64,
11262                        Op0.getOperand(0), DAG.getConstant(DWord, dl, MVT::i32));
11263   }
11264   case ISD::FADD:
11265   case ISD::FMUL:
11266   case ISD::FSUB: {
11267     SDValue NewLoad[2];
11268     for (unsigned i = 0, ie = Op0.getNumOperands(); i != ie; ++i) {
11269       // Ensure both input are loads.
11270       SDValue LdOp = Op0.getOperand(i);
11271       if (LdOp.getOpcode() != ISD::LOAD)
11272         return SDValue();
11273       // Generate new load node.
11274       LoadSDNode *LD = cast<LoadSDNode>(LdOp);
11275       SDValue LoadOps[] = {LD->getChain(), LD->getBasePtr()};
11276       NewLoad[i] = DAG.getMemIntrinsicNode(
11277           PPCISD::LD_VSX_LH, dl, DAG.getVTList(MVT::v4f32, MVT::Other), LoadOps,
11278           LD->getMemoryVT(), LD->getMemOperand());
11279     }
11280     SDValue NewOp =
11281         DAG.getNode(Op0.getOpcode(), SDLoc(Op0), MVT::v4f32, NewLoad[0],
11282                     NewLoad[1], Op0.getNode()->getFlags());
11283     return DAG.getNode(PPCISD::FP_EXTEND_HALF, dl, MVT::v2f64, NewOp,
11284                        DAG.getConstant(0, dl, MVT::i32));
11285   }
11286   case ISD::LOAD: {
11287     LoadSDNode *LD = cast<LoadSDNode>(Op0);
11288     SDValue LoadOps[] = {LD->getChain(), LD->getBasePtr()};
11289     SDValue NewLd = DAG.getMemIntrinsicNode(
11290         PPCISD::LD_VSX_LH, dl, DAG.getVTList(MVT::v4f32, MVT::Other), LoadOps,
11291         LD->getMemoryVT(), LD->getMemOperand());
11292     return DAG.getNode(PPCISD::FP_EXTEND_HALF, dl, MVT::v2f64, NewLd,
11293                        DAG.getConstant(0, dl, MVT::i32));
11294   }
11295   }
11296   llvm_unreachable("ERROR:Should return for all cases within swtich.");
11297 }
11298 
11299 /// LowerOperation - Provide custom lowering hooks for some operations.
11300 ///
11301 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
11302   switch (Op.getOpcode()) {
11303   default: llvm_unreachable("Wasn't expecting to be able to lower this!");
11304   case ISD::FPOW:               return lowerPow(Op, DAG);
11305   case ISD::FSIN:               return lowerSin(Op, DAG);
11306   case ISD::FCOS:               return lowerCos(Op, DAG);
11307   case ISD::FLOG:               return lowerLog(Op, DAG);
11308   case ISD::FLOG10:             return lowerLog10(Op, DAG);
11309   case ISD::FEXP:               return lowerExp(Op, DAG);
11310   case ISD::ConstantPool:       return LowerConstantPool(Op, DAG);
11311   case ISD::BlockAddress:       return LowerBlockAddress(Op, DAG);
11312   case ISD::GlobalAddress:      return LowerGlobalAddress(Op, DAG);
11313   case ISD::GlobalTLSAddress:   return LowerGlobalTLSAddress(Op, DAG);
11314   case ISD::JumpTable:          return LowerJumpTable(Op, DAG);
11315   case ISD::STRICT_FSETCC:
11316   case ISD::STRICT_FSETCCS:
11317   case ISD::SETCC:              return LowerSETCC(Op, DAG);
11318   case ISD::INIT_TRAMPOLINE:    return LowerINIT_TRAMPOLINE(Op, DAG);
11319   case ISD::ADJUST_TRAMPOLINE:  return LowerADJUST_TRAMPOLINE(Op, DAG);
11320 
11321   case ISD::INLINEASM:
11322   case ISD::INLINEASM_BR:       return LowerINLINEASM(Op, DAG);
11323   // Variable argument lowering.
11324   case ISD::VASTART:            return LowerVASTART(Op, DAG);
11325   case ISD::VAARG:              return LowerVAARG(Op, DAG);
11326   case ISD::VACOPY:             return LowerVACOPY(Op, DAG);
11327 
11328   case ISD::STACKRESTORE:       return LowerSTACKRESTORE(Op, DAG);
11329   case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
11330   case ISD::GET_DYNAMIC_AREA_OFFSET:
11331     return LowerGET_DYNAMIC_AREA_OFFSET(Op, DAG);
11332 
11333   // Exception handling lowering.
11334   case ISD::EH_DWARF_CFA:       return LowerEH_DWARF_CFA(Op, DAG);
11335   case ISD::EH_SJLJ_SETJMP:     return lowerEH_SJLJ_SETJMP(Op, DAG);
11336   case ISD::EH_SJLJ_LONGJMP:    return lowerEH_SJLJ_LONGJMP(Op, DAG);
11337 
11338   case ISD::LOAD:               return LowerLOAD(Op, DAG);
11339   case ISD::STORE:              return LowerSTORE(Op, DAG);
11340   case ISD::TRUNCATE:           return LowerTRUNCATE(Op, DAG);
11341   case ISD::SELECT_CC:          return LowerSELECT_CC(Op, DAG);
11342   case ISD::STRICT_FP_TO_UINT:
11343   case ISD::STRICT_FP_TO_SINT:
11344   case ISD::FP_TO_UINT:
11345   case ISD::FP_TO_SINT:         return LowerFP_TO_INT(Op, DAG, SDLoc(Op));
11346   case ISD::STRICT_UINT_TO_FP:
11347   case ISD::STRICT_SINT_TO_FP:
11348   case ISD::UINT_TO_FP:
11349   case ISD::SINT_TO_FP:         return LowerINT_TO_FP(Op, DAG);
11350   case ISD::GET_ROUNDING:       return LowerGET_ROUNDING(Op, DAG);
11351 
11352   // Lower 64-bit shifts.
11353   case ISD::SHL_PARTS:          return LowerSHL_PARTS(Op, DAG);
11354   case ISD::SRL_PARTS:          return LowerSRL_PARTS(Op, DAG);
11355   case ISD::SRA_PARTS:          return LowerSRA_PARTS(Op, DAG);
11356 
11357   case ISD::FSHL:               return LowerFunnelShift(Op, DAG);
11358   case ISD::FSHR:               return LowerFunnelShift(Op, DAG);
11359 
11360   // Vector-related lowering.
11361   case ISD::BUILD_VECTOR:       return LowerBUILD_VECTOR(Op, DAG);
11362   case ISD::VECTOR_SHUFFLE:     return LowerVECTOR_SHUFFLE(Op, DAG);
11363   case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
11364   case ISD::SCALAR_TO_VECTOR:   return LowerSCALAR_TO_VECTOR(Op, DAG);
11365   case ISD::INSERT_VECTOR_ELT:  return LowerINSERT_VECTOR_ELT(Op, DAG);
11366   case ISD::MUL:                return LowerMUL(Op, DAG);
11367   case ISD::FP_EXTEND:          return LowerFP_EXTEND(Op, DAG);
11368   case ISD::STRICT_FP_ROUND:
11369   case ISD::FP_ROUND:
11370     return LowerFP_ROUND(Op, DAG);
11371   case ISD::ROTL:               return LowerROTL(Op, DAG);
11372 
11373   // For counter-based loop handling.
11374   case ISD::INTRINSIC_W_CHAIN:  return SDValue();
11375 
11376   case ISD::BITCAST:            return LowerBITCAST(Op, DAG);
11377 
11378   // Frame & Return address.
11379   case ISD::RETURNADDR:         return LowerRETURNADDR(Op, DAG);
11380   case ISD::FRAMEADDR:          return LowerFRAMEADDR(Op, DAG);
11381 
11382   case ISD::INTRINSIC_VOID:
11383     return LowerINTRINSIC_VOID(Op, DAG);
11384   case ISD::BSWAP:
11385     return LowerBSWAP(Op, DAG);
11386   case ISD::ATOMIC_CMP_SWAP:
11387     return LowerATOMIC_CMP_SWAP(Op, DAG);
11388   case ISD::ATOMIC_STORE:
11389     return LowerATOMIC_LOAD_STORE(Op, DAG);
11390   }
11391 }
11392 
11393 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
11394                                            SmallVectorImpl<SDValue>&Results,
11395                                            SelectionDAG &DAG) const {
11396   SDLoc dl(N);
11397   switch (N->getOpcode()) {
11398   default:
11399     llvm_unreachable("Do not know how to custom type legalize this operation!");
11400   case ISD::ATOMIC_LOAD: {
11401     SDValue Res = LowerATOMIC_LOAD_STORE(SDValue(N, 0), DAG);
11402     Results.push_back(Res);
11403     Results.push_back(Res.getValue(1));
11404     break;
11405   }
11406   case ISD::READCYCLECOUNTER: {
11407     SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
11408     SDValue RTB = DAG.getNode(PPCISD::READ_TIME_BASE, dl, VTs, N->getOperand(0));
11409 
11410     Results.push_back(
11411         DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, RTB, RTB.getValue(1)));
11412     Results.push_back(RTB.getValue(2));
11413     break;
11414   }
11415   case ISD::INTRINSIC_W_CHAIN: {
11416     if (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() !=
11417         Intrinsic::loop_decrement)
11418       break;
11419 
11420     assert(N->getValueType(0) == MVT::i1 &&
11421            "Unexpected result type for CTR decrement intrinsic");
11422     EVT SVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(),
11423                                  N->getValueType(0));
11424     SDVTList VTs = DAG.getVTList(SVT, MVT::Other);
11425     SDValue NewInt = DAG.getNode(N->getOpcode(), dl, VTs, N->getOperand(0),
11426                                  N->getOperand(1));
11427 
11428     Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewInt));
11429     Results.push_back(NewInt.getValue(1));
11430     break;
11431   }
11432   case ISD::INTRINSIC_WO_CHAIN: {
11433     switch (cast<ConstantSDNode>(N->getOperand(0))->getZExtValue()) {
11434     case Intrinsic::ppc_pack_longdouble:
11435       Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
11436                                     N->getOperand(2), N->getOperand(1)));
11437       break;
11438     case Intrinsic::ppc_maxfe:
11439     case Intrinsic::ppc_minfe:
11440     case Intrinsic::ppc_fnmsub:
11441     case Intrinsic::ppc_convert_f128_to_ppcf128:
11442       Results.push_back(LowerINTRINSIC_WO_CHAIN(SDValue(N, 0), DAG));
11443       break;
11444     }
11445     break;
11446   }
11447   case ISD::VAARG: {
11448     if (!Subtarget.isSVR4ABI() || Subtarget.isPPC64())
11449       return;
11450 
11451     EVT VT = N->getValueType(0);
11452 
11453     if (VT == MVT::i64) {
11454       SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG);
11455 
11456       Results.push_back(NewNode);
11457       Results.push_back(NewNode.getValue(1));
11458     }
11459     return;
11460   }
11461   case ISD::STRICT_FP_TO_SINT:
11462   case ISD::STRICT_FP_TO_UINT:
11463   case ISD::FP_TO_SINT:
11464   case ISD::FP_TO_UINT: {
11465     // LowerFP_TO_INT() can only handle f32 and f64.
11466     if (N->getOperand(N->isStrictFPOpcode() ? 1 : 0).getValueType() ==
11467         MVT::ppcf128)
11468       return;
11469     SDValue LoweredValue = LowerFP_TO_INT(SDValue(N, 0), DAG, dl);
11470     Results.push_back(LoweredValue);
11471     if (N->isStrictFPOpcode())
11472       Results.push_back(LoweredValue.getValue(1));
11473     return;
11474   }
11475   case ISD::TRUNCATE: {
11476     if (!N->getValueType(0).isVector())
11477       return;
11478     SDValue Lowered = LowerTRUNCATEVector(SDValue(N, 0), DAG);
11479     if (Lowered)
11480       Results.push_back(Lowered);
11481     return;
11482   }
11483   case ISD::FSHL:
11484   case ISD::FSHR:
11485     // Don't handle funnel shifts here.
11486     return;
11487   case ISD::BITCAST:
11488     // Don't handle bitcast here.
11489     return;
11490   case ISD::FP_EXTEND:
11491     SDValue Lowered = LowerFP_EXTEND(SDValue(N, 0), DAG);
11492     if (Lowered)
11493       Results.push_back(Lowered);
11494     return;
11495   }
11496 }
11497 
11498 //===----------------------------------------------------------------------===//
11499 //  Other Lowering Code
11500 //===----------------------------------------------------------------------===//
11501 
11502 static Instruction *callIntrinsic(IRBuilderBase &Builder, Intrinsic::ID Id) {
11503   Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11504   Function *Func = Intrinsic::getDeclaration(M, Id);
11505   return Builder.CreateCall(Func, {});
11506 }
11507 
11508 // The mappings for emitLeading/TrailingFence is taken from
11509 // http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11510 Instruction *PPCTargetLowering::emitLeadingFence(IRBuilderBase &Builder,
11511                                                  Instruction *Inst,
11512                                                  AtomicOrdering Ord) const {
11513   if (Ord == AtomicOrdering::SequentiallyConsistent)
11514     return callIntrinsic(Builder, Intrinsic::ppc_sync);
11515   if (isReleaseOrStronger(Ord))
11516     return callIntrinsic(Builder, Intrinsic::ppc_lwsync);
11517   return nullptr;
11518 }
11519 
11520 Instruction *PPCTargetLowering::emitTrailingFence(IRBuilderBase &Builder,
11521                                                   Instruction *Inst,
11522                                                   AtomicOrdering Ord) const {
11523   if (Inst->hasAtomicLoad() && isAcquireOrStronger(Ord)) {
11524     // See http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html and
11525     // http://www.rdrop.com/users/paulmck/scalability/paper/N2745r.2011.03.04a.html
11526     // and http://www.cl.cam.ac.uk/~pes20/cppppc/ for justification.
11527     if (isa<LoadInst>(Inst) && Subtarget.isPPC64())
11528       return Builder.CreateCall(
11529           Intrinsic::getDeclaration(
11530               Builder.GetInsertBlock()->getParent()->getParent(),
11531               Intrinsic::ppc_cfence, {Inst->getType()}),
11532           {Inst});
11533     // FIXME: Can use isync for rmw operation.
11534     return callIntrinsic(Builder, Intrinsic::ppc_lwsync);
11535   }
11536   return nullptr;
11537 }
11538 
11539 MachineBasicBlock *
11540 PPCTargetLowering::EmitAtomicBinary(MachineInstr &MI, MachineBasicBlock *BB,
11541                                     unsigned AtomicSize,
11542                                     unsigned BinOpcode,
11543                                     unsigned CmpOpcode,
11544                                     unsigned CmpPred) const {
11545   // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
11546   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
11547 
11548   auto LoadMnemonic = PPC::LDARX;
11549   auto StoreMnemonic = PPC::STDCX;
11550   switch (AtomicSize) {
11551   default:
11552     llvm_unreachable("Unexpected size of atomic entity");
11553   case 1:
11554     LoadMnemonic = PPC::LBARX;
11555     StoreMnemonic = PPC::STBCX;
11556     assert(Subtarget.hasPartwordAtomics() && "Call this only with size >=4");
11557     break;
11558   case 2:
11559     LoadMnemonic = PPC::LHARX;
11560     StoreMnemonic = PPC::STHCX;
11561     assert(Subtarget.hasPartwordAtomics() && "Call this only with size >=4");
11562     break;
11563   case 4:
11564     LoadMnemonic = PPC::LWARX;
11565     StoreMnemonic = PPC::STWCX;
11566     break;
11567   case 8:
11568     LoadMnemonic = PPC::LDARX;
11569     StoreMnemonic = PPC::STDCX;
11570     break;
11571   }
11572 
11573   const BasicBlock *LLVM_BB = BB->getBasicBlock();
11574   MachineFunction *F = BB->getParent();
11575   MachineFunction::iterator It = ++BB->getIterator();
11576 
11577   Register dest = MI.getOperand(0).getReg();
11578   Register ptrA = MI.getOperand(1).getReg();
11579   Register ptrB = MI.getOperand(2).getReg();
11580   Register incr = MI.getOperand(3).getReg();
11581   DebugLoc dl = MI.getDebugLoc();
11582 
11583   MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
11584   MachineBasicBlock *loop2MBB =
11585     CmpOpcode ? F->CreateMachineBasicBlock(LLVM_BB) : nullptr;
11586   MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
11587   F->insert(It, loopMBB);
11588   if (CmpOpcode)
11589     F->insert(It, loop2MBB);
11590   F->insert(It, exitMBB);
11591   exitMBB->splice(exitMBB->begin(), BB,
11592                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
11593   exitMBB->transferSuccessorsAndUpdatePHIs(BB);
11594 
11595   MachineRegisterInfo &RegInfo = F->getRegInfo();
11596   Register TmpReg = (!BinOpcode) ? incr :
11597     RegInfo.createVirtualRegister( AtomicSize == 8 ? &PPC::G8RCRegClass
11598                                            : &PPC::GPRCRegClass);
11599 
11600   //  thisMBB:
11601   //   ...
11602   //   fallthrough --> loopMBB
11603   BB->addSuccessor(loopMBB);
11604 
11605   //  loopMBB:
11606   //   l[wd]arx dest, ptr
11607   //   add r0, dest, incr
11608   //   st[wd]cx. r0, ptr
11609   //   bne- loopMBB
11610   //   fallthrough --> exitMBB
11611 
11612   // For max/min...
11613   //  loopMBB:
11614   //   l[wd]arx dest, ptr
11615   //   cmpl?[wd] dest, incr
11616   //   bgt exitMBB
11617   //  loop2MBB:
11618   //   st[wd]cx. dest, ptr
11619   //   bne- loopMBB
11620   //   fallthrough --> exitMBB
11621 
11622   BB = loopMBB;
11623   BuildMI(BB, dl, TII->get(LoadMnemonic), dest)
11624     .addReg(ptrA).addReg(ptrB);
11625   if (BinOpcode)
11626     BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
11627   if (CmpOpcode) {
11628     Register CrReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass);
11629     // Signed comparisons of byte or halfword values must be sign-extended.
11630     if (CmpOpcode == PPC::CMPW && AtomicSize < 4) {
11631       Register ExtReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
11632       BuildMI(BB, dl, TII->get(AtomicSize == 1 ? PPC::EXTSB : PPC::EXTSH),
11633               ExtReg).addReg(dest);
11634       BuildMI(BB, dl, TII->get(CmpOpcode), CrReg).addReg(ExtReg).addReg(incr);
11635     } else
11636       BuildMI(BB, dl, TII->get(CmpOpcode), CrReg).addReg(dest).addReg(incr);
11637 
11638     BuildMI(BB, dl, TII->get(PPC::BCC))
11639         .addImm(CmpPred)
11640         .addReg(CrReg)
11641         .addMBB(exitMBB);
11642     BB->addSuccessor(loop2MBB);
11643     BB->addSuccessor(exitMBB);
11644     BB = loop2MBB;
11645   }
11646   BuildMI(BB, dl, TII->get(StoreMnemonic))
11647     .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
11648   BuildMI(BB, dl, TII->get(PPC::BCC))
11649     .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
11650   BB->addSuccessor(loopMBB);
11651   BB->addSuccessor(exitMBB);
11652 
11653   //  exitMBB:
11654   //   ...
11655   BB = exitMBB;
11656   return BB;
11657 }
11658 
11659 static bool isSignExtended(MachineInstr &MI, const PPCInstrInfo *TII) {
11660   switch(MI.getOpcode()) {
11661   default:
11662     return false;
11663   case PPC::COPY:
11664     return TII->isSignExtended(MI.getOperand(1).getReg(),
11665                                &MI.getMF()->getRegInfo());
11666   case PPC::LHA:
11667   case PPC::LHA8:
11668   case PPC::LHAU:
11669   case PPC::LHAU8:
11670   case PPC::LHAUX:
11671   case PPC::LHAUX8:
11672   case PPC::LHAX:
11673   case PPC::LHAX8:
11674   case PPC::LWA:
11675   case PPC::LWAUX:
11676   case PPC::LWAX:
11677   case PPC::LWAX_32:
11678   case PPC::LWA_32:
11679   case PPC::PLHA:
11680   case PPC::PLHA8:
11681   case PPC::PLHA8pc:
11682   case PPC::PLHApc:
11683   case PPC::PLWA:
11684   case PPC::PLWA8:
11685   case PPC::PLWA8pc:
11686   case PPC::PLWApc:
11687   case PPC::EXTSB:
11688   case PPC::EXTSB8:
11689   case PPC::EXTSB8_32_64:
11690   case PPC::EXTSB8_rec:
11691   case PPC::EXTSB_rec:
11692   case PPC::EXTSH:
11693   case PPC::EXTSH8:
11694   case PPC::EXTSH8_32_64:
11695   case PPC::EXTSH8_rec:
11696   case PPC::EXTSH_rec:
11697   case PPC::EXTSW:
11698   case PPC::EXTSWSLI:
11699   case PPC::EXTSWSLI_32_64:
11700   case PPC::EXTSWSLI_32_64_rec:
11701   case PPC::EXTSWSLI_rec:
11702   case PPC::EXTSW_32:
11703   case PPC::EXTSW_32_64:
11704   case PPC::EXTSW_32_64_rec:
11705   case PPC::EXTSW_rec:
11706   case PPC::SRAW:
11707   case PPC::SRAWI:
11708   case PPC::SRAWI_rec:
11709   case PPC::SRAW_rec:
11710     return true;
11711   }
11712   return false;
11713 }
11714 
11715 MachineBasicBlock *PPCTargetLowering::EmitPartwordAtomicBinary(
11716     MachineInstr &MI, MachineBasicBlock *BB,
11717     bool is8bit, // operation
11718     unsigned BinOpcode, unsigned CmpOpcode, unsigned CmpPred) const {
11719   // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
11720   const PPCInstrInfo *TII = Subtarget.getInstrInfo();
11721 
11722   // If this is a signed comparison and the value being compared is not known
11723   // to be sign extended, sign extend it here.
11724   DebugLoc dl = MI.getDebugLoc();
11725   MachineFunction *F = BB->getParent();
11726   MachineRegisterInfo &RegInfo = F->getRegInfo();
11727   Register incr = MI.getOperand(3).getReg();
11728   bool IsSignExtended =
11729       incr.isVirtual() && isSignExtended(*RegInfo.getVRegDef(incr), TII);
11730 
11731   if (CmpOpcode == PPC::CMPW && !IsSignExtended) {
11732     Register ValueReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
11733     BuildMI(*BB, MI, dl, TII->get(is8bit ? PPC::EXTSB : PPC::EXTSH), ValueReg)
11734         .addReg(MI.getOperand(3).getReg());
11735     MI.getOperand(3).setReg(ValueReg);
11736   }
11737   // If we support part-word atomic mnemonics, just use them
11738   if (Subtarget.hasPartwordAtomics())
11739     return EmitAtomicBinary(MI, BB, is8bit ? 1 : 2, BinOpcode, CmpOpcode,
11740                             CmpPred);
11741 
11742   // In 64 bit mode we have to use 64 bits for addresses, even though the
11743   // lwarx/stwcx are 32 bits.  With the 32-bit atomics we can use address
11744   // registers without caring whether they're 32 or 64, but here we're
11745   // doing actual arithmetic on the addresses.
11746   bool is64bit = Subtarget.isPPC64();
11747   bool isLittleEndian = Subtarget.isLittleEndian();
11748   unsigned ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO;
11749 
11750   const BasicBlock *LLVM_BB = BB->getBasicBlock();
11751   MachineFunction::iterator It = ++BB->getIterator();
11752 
11753   Register dest = MI.getOperand(0).getReg();
11754   Register ptrA = MI.getOperand(1).getReg();
11755   Register ptrB = MI.getOperand(2).getReg();
11756 
11757   MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
11758   MachineBasicBlock *loop2MBB =
11759       CmpOpcode ? F->CreateMachineBasicBlock(LLVM_BB) : nullptr;
11760   MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
11761   F->insert(It, loopMBB);
11762   if (CmpOpcode)
11763     F->insert(It, loop2MBB);
11764   F->insert(It, exitMBB);
11765   exitMBB->splice(exitMBB->begin(), BB,
11766                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
11767   exitMBB->transferSuccessorsAndUpdatePHIs(BB);
11768 
11769   const TargetRegisterClass *RC =
11770       is64bit ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
11771   const TargetRegisterClass *GPRC = &PPC::GPRCRegClass;
11772 
11773   Register PtrReg = RegInfo.createVirtualRegister(RC);
11774   Register Shift1Reg = RegInfo.createVirtualRegister(GPRC);
11775   Register ShiftReg =
11776       isLittleEndian ? Shift1Reg : RegInfo.createVirtualRegister(GPRC);
11777   Register Incr2Reg = RegInfo.createVirtualRegister(GPRC);
11778   Register MaskReg = RegInfo.createVirtualRegister(GPRC);
11779   Register Mask2Reg = RegInfo.createVirtualRegister(GPRC);
11780   Register Mask3Reg = RegInfo.createVirtualRegister(GPRC);
11781   Register Tmp2Reg = RegInfo.createVirtualRegister(GPRC);
11782   Register Tmp3Reg = RegInfo.createVirtualRegister(GPRC);
11783   Register Tmp4Reg = RegInfo.createVirtualRegister(GPRC);
11784   Register TmpDestReg = RegInfo.createVirtualRegister(GPRC);
11785   Register SrwDestReg = RegInfo.createVirtualRegister(GPRC);
11786   Register Ptr1Reg;
11787   Register TmpReg =
11788       (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(GPRC);
11789 
11790   //  thisMBB:
11791   //   ...
11792   //   fallthrough --> loopMBB
11793   BB->addSuccessor(loopMBB);
11794 
11795   // The 4-byte load must be aligned, while a char or short may be
11796   // anywhere in the word.  Hence all this nasty bookkeeping code.
11797   //   add ptr1, ptrA, ptrB [copy if ptrA==0]
11798   //   rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
11799   //   xori shift, shift1, 24 [16]
11800   //   rlwinm ptr, ptr1, 0, 0, 29
11801   //   slw incr2, incr, shift
11802   //   li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
11803   //   slw mask, mask2, shift
11804   //  loopMBB:
11805   //   lwarx tmpDest, ptr
11806   //   add tmp, tmpDest, incr2
11807   //   andc tmp2, tmpDest, mask
11808   //   and tmp3, tmp, mask
11809   //   or tmp4, tmp3, tmp2
11810   //   stwcx. tmp4, ptr
11811   //   bne- loopMBB
11812   //   fallthrough --> exitMBB
11813   //   srw SrwDest, tmpDest, shift
11814   //   rlwinm SrwDest, SrwDest, 0, 24 [16], 31
11815   if (ptrA != ZeroReg) {
11816     Ptr1Reg = RegInfo.createVirtualRegister(RC);
11817     BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
11818         .addReg(ptrA)
11819         .addReg(ptrB);
11820   } else {
11821     Ptr1Reg = ptrB;
11822   }
11823   // We need use 32-bit subregister to avoid mismatch register class in 64-bit
11824   // mode.
11825   BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg)
11826       .addReg(Ptr1Reg, 0, is64bit ? PPC::sub_32 : 0)
11827       .addImm(3)
11828       .addImm(27)
11829       .addImm(is8bit ? 28 : 27);
11830   if (!isLittleEndian)
11831     BuildMI(BB, dl, TII->get(PPC::XORI), ShiftReg)
11832         .addReg(Shift1Reg)
11833         .addImm(is8bit ? 24 : 16);
11834   if (is64bit)
11835     BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
11836         .addReg(Ptr1Reg)
11837         .addImm(0)
11838         .addImm(61);
11839   else
11840     BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
11841         .addReg(Ptr1Reg)
11842         .addImm(0)
11843         .addImm(0)
11844         .addImm(29);
11845   BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg).addReg(incr).addReg(ShiftReg);
11846   if (is8bit)
11847     BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
11848   else {
11849     BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
11850     BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
11851         .addReg(Mask3Reg)
11852         .addImm(65535);
11853   }
11854   BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
11855       .addReg(Mask2Reg)
11856       .addReg(ShiftReg);
11857 
11858   BB = loopMBB;
11859   BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
11860       .addReg(ZeroReg)
11861       .addReg(PtrReg);
11862   if (BinOpcode)
11863     BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
11864         .addReg(Incr2Reg)
11865         .addReg(TmpDestReg);
11866   BuildMI(BB, dl, TII->get(PPC::ANDC), Tmp2Reg)
11867       .addReg(TmpDestReg)
11868       .addReg(MaskReg);
11869   BuildMI(BB, dl, TII->get(PPC::AND), Tmp3Reg).addReg(TmpReg).addReg(MaskReg);
11870   if (CmpOpcode) {
11871     // For unsigned comparisons, we can directly compare the shifted values.
11872     // For signed comparisons we shift and sign extend.
11873     Register SReg = RegInfo.createVirtualRegister(GPRC);
11874     Register CrReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass);
11875     BuildMI(BB, dl, TII->get(PPC::AND), SReg)
11876         .addReg(TmpDestReg)
11877         .addReg(MaskReg);
11878     unsigned ValueReg = SReg;
11879     unsigned CmpReg = Incr2Reg;
11880     if (CmpOpcode == PPC::CMPW) {
11881       ValueReg = RegInfo.createVirtualRegister(GPRC);
11882       BuildMI(BB, dl, TII->get(PPC::SRW), ValueReg)
11883           .addReg(SReg)
11884           .addReg(ShiftReg);
11885       Register ValueSReg = RegInfo.createVirtualRegister(GPRC);
11886       BuildMI(BB, dl, TII->get(is8bit ? PPC::EXTSB : PPC::EXTSH), ValueSReg)
11887           .addReg(ValueReg);
11888       ValueReg = ValueSReg;
11889       CmpReg = incr;
11890     }
11891     BuildMI(BB, dl, TII->get(CmpOpcode), CrReg).addReg(ValueReg).addReg(CmpReg);
11892     BuildMI(BB, dl, TII->get(PPC::BCC))
11893         .addImm(CmpPred)
11894         .addReg(CrReg)
11895         .addMBB(exitMBB);
11896     BB->addSuccessor(loop2MBB);
11897     BB->addSuccessor(exitMBB);
11898     BB = loop2MBB;
11899   }
11900   BuildMI(BB, dl, TII->get(PPC::OR), Tmp4Reg).addReg(Tmp3Reg).addReg(Tmp2Reg);
11901   BuildMI(BB, dl, TII->get(PPC::STWCX))
11902       .addReg(Tmp4Reg)
11903       .addReg(ZeroReg)
11904       .addReg(PtrReg);
11905   BuildMI(BB, dl, TII->get(PPC::BCC))
11906       .addImm(PPC::PRED_NE)
11907       .addReg(PPC::CR0)
11908       .addMBB(loopMBB);
11909   BB->addSuccessor(loopMBB);
11910   BB->addSuccessor(exitMBB);
11911 
11912   //  exitMBB:
11913   //   ...
11914   BB = exitMBB;
11915   // Since the shift amount is not a constant, we need to clear
11916   // the upper bits with a separate RLWINM.
11917   BuildMI(*BB, BB->begin(), dl, TII->get(PPC::RLWINM), dest)
11918       .addReg(SrwDestReg)
11919       .addImm(0)
11920       .addImm(is8bit ? 24 : 16)
11921       .addImm(31);
11922   BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), SrwDestReg)
11923       .addReg(TmpDestReg)
11924       .addReg(ShiftReg);
11925   return BB;
11926 }
11927 
11928 llvm::MachineBasicBlock *
11929 PPCTargetLowering::emitEHSjLjSetJmp(MachineInstr &MI,
11930                                     MachineBasicBlock *MBB) const {
11931   DebugLoc DL = MI.getDebugLoc();
11932   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
11933   const PPCRegisterInfo *TRI = Subtarget.getRegisterInfo();
11934 
11935   MachineFunction *MF = MBB->getParent();
11936   MachineRegisterInfo &MRI = MF->getRegInfo();
11937 
11938   const BasicBlock *BB = MBB->getBasicBlock();
11939   MachineFunction::iterator I = ++MBB->getIterator();
11940 
11941   Register DstReg = MI.getOperand(0).getReg();
11942   const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
11943   assert(TRI->isTypeLegalForClass(*RC, MVT::i32) && "Invalid destination!");
11944   Register mainDstReg = MRI.createVirtualRegister(RC);
11945   Register restoreDstReg = MRI.createVirtualRegister(RC);
11946 
11947   MVT PVT = getPointerTy(MF->getDataLayout());
11948   assert((PVT == MVT::i64 || PVT == MVT::i32) &&
11949          "Invalid Pointer Size!");
11950   // For v = setjmp(buf), we generate
11951   //
11952   // thisMBB:
11953   //  SjLjSetup mainMBB
11954   //  bl mainMBB
11955   //  v_restore = 1
11956   //  b sinkMBB
11957   //
11958   // mainMBB:
11959   //  buf[LabelOffset] = LR
11960   //  v_main = 0
11961   //
11962   // sinkMBB:
11963   //  v = phi(main, restore)
11964   //
11965 
11966   MachineBasicBlock *thisMBB = MBB;
11967   MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
11968   MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
11969   MF->insert(I, mainMBB);
11970   MF->insert(I, sinkMBB);
11971 
11972   MachineInstrBuilder MIB;
11973 
11974   // Transfer the remainder of BB and its successor edges to sinkMBB.
11975   sinkMBB->splice(sinkMBB->begin(), MBB,
11976                   std::next(MachineBasicBlock::iterator(MI)), MBB->end());
11977   sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
11978 
11979   // Note that the structure of the jmp_buf used here is not compatible
11980   // with that used by libc, and is not designed to be. Specifically, it
11981   // stores only those 'reserved' registers that LLVM does not otherwise
11982   // understand how to spill. Also, by convention, by the time this
11983   // intrinsic is called, Clang has already stored the frame address in the
11984   // first slot of the buffer and stack address in the third. Following the
11985   // X86 target code, we'll store the jump address in the second slot. We also
11986   // need to save the TOC pointer (R2) to handle jumps between shared
11987   // libraries, and that will be stored in the fourth slot. The thread
11988   // identifier (R13) is not affected.
11989 
11990   // thisMBB:
11991   const int64_t LabelOffset = 1 * PVT.getStoreSize();
11992   const int64_t TOCOffset   = 3 * PVT.getStoreSize();
11993   const int64_t BPOffset    = 4 * PVT.getStoreSize();
11994 
11995   // Prepare IP either in reg.
11996   const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
11997   Register LabelReg = MRI.createVirtualRegister(PtrRC);
11998   Register BufReg = MI.getOperand(1).getReg();
11999 
12000   if (Subtarget.is64BitELFABI()) {
12001     setUsesTOCBasePtr(*MBB->getParent());
12002     MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::STD))
12003               .addReg(PPC::X2)
12004               .addImm(TOCOffset)
12005               .addReg(BufReg)
12006               .cloneMemRefs(MI);
12007   }
12008 
12009   // Naked functions never have a base pointer, and so we use r1. For all
12010   // other functions, this decision must be delayed until during PEI.
12011   unsigned BaseReg;
12012   if (MF->getFunction().hasFnAttribute(Attribute::Naked))
12013     BaseReg = Subtarget.isPPC64() ? PPC::X1 : PPC::R1;
12014   else
12015     BaseReg = Subtarget.isPPC64() ? PPC::BP8 : PPC::BP;
12016 
12017   MIB = BuildMI(*thisMBB, MI, DL,
12018                 TII->get(Subtarget.isPPC64() ? PPC::STD : PPC::STW))
12019             .addReg(BaseReg)
12020             .addImm(BPOffset)
12021             .addReg(BufReg)
12022             .cloneMemRefs(MI);
12023 
12024   // Setup
12025   MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::BCLalways)).addMBB(mainMBB);
12026   MIB.addRegMask(TRI->getNoPreservedMask());
12027 
12028   BuildMI(*thisMBB, MI, DL, TII->get(PPC::LI), restoreDstReg).addImm(1);
12029 
12030   MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::EH_SjLj_Setup))
12031           .addMBB(mainMBB);
12032   MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::B)).addMBB(sinkMBB);
12033 
12034   thisMBB->addSuccessor(mainMBB, BranchProbability::getZero());
12035   thisMBB->addSuccessor(sinkMBB, BranchProbability::getOne());
12036 
12037   // mainMBB:
12038   //  mainDstReg = 0
12039   MIB =
12040       BuildMI(mainMBB, DL,
12041               TII->get(Subtarget.isPPC64() ? PPC::MFLR8 : PPC::MFLR), LabelReg);
12042 
12043   // Store IP
12044   if (Subtarget.isPPC64()) {
12045     MIB = BuildMI(mainMBB, DL, TII->get(PPC::STD))
12046             .addReg(LabelReg)
12047             .addImm(LabelOffset)
12048             .addReg(BufReg);
12049   } else {
12050     MIB = BuildMI(mainMBB, DL, TII->get(PPC::STW))
12051             .addReg(LabelReg)
12052             .addImm(LabelOffset)
12053             .addReg(BufReg);
12054   }
12055   MIB.cloneMemRefs(MI);
12056 
12057   BuildMI(mainMBB, DL, TII->get(PPC::LI), mainDstReg).addImm(0);
12058   mainMBB->addSuccessor(sinkMBB);
12059 
12060   // sinkMBB:
12061   BuildMI(*sinkMBB, sinkMBB->begin(), DL,
12062           TII->get(PPC::PHI), DstReg)
12063     .addReg(mainDstReg).addMBB(mainMBB)
12064     .addReg(restoreDstReg).addMBB(thisMBB);
12065 
12066   MI.eraseFromParent();
12067   return sinkMBB;
12068 }
12069 
12070 MachineBasicBlock *
12071 PPCTargetLowering::emitEHSjLjLongJmp(MachineInstr &MI,
12072                                      MachineBasicBlock *MBB) const {
12073   DebugLoc DL = MI.getDebugLoc();
12074   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
12075 
12076   MachineFunction *MF = MBB->getParent();
12077   MachineRegisterInfo &MRI = MF->getRegInfo();
12078 
12079   MVT PVT = getPointerTy(MF->getDataLayout());
12080   assert((PVT == MVT::i64 || PVT == MVT::i32) &&
12081          "Invalid Pointer Size!");
12082 
12083   const TargetRegisterClass *RC =
12084     (PVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
12085   Register Tmp = MRI.createVirtualRegister(RC);
12086   // Since FP is only updated here but NOT referenced, it's treated as GPR.
12087   unsigned FP  = (PVT == MVT::i64) ? PPC::X31 : PPC::R31;
12088   unsigned SP  = (PVT == MVT::i64) ? PPC::X1 : PPC::R1;
12089   unsigned BP =
12090       (PVT == MVT::i64)
12091           ? PPC::X30
12092           : (Subtarget.isSVR4ABI() && isPositionIndependent() ? PPC::R29
12093                                                               : PPC::R30);
12094 
12095   MachineInstrBuilder MIB;
12096 
12097   const int64_t LabelOffset = 1 * PVT.getStoreSize();
12098   const int64_t SPOffset    = 2 * PVT.getStoreSize();
12099   const int64_t TOCOffset   = 3 * PVT.getStoreSize();
12100   const int64_t BPOffset    = 4 * PVT.getStoreSize();
12101 
12102   Register BufReg = MI.getOperand(0).getReg();
12103 
12104   // Reload FP (the jumped-to function may not have had a
12105   // frame pointer, and if so, then its r31 will be restored
12106   // as necessary).
12107   if (PVT == MVT::i64) {
12108     MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), FP)
12109             .addImm(0)
12110             .addReg(BufReg);
12111   } else {
12112     MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), FP)
12113             .addImm(0)
12114             .addReg(BufReg);
12115   }
12116   MIB.cloneMemRefs(MI);
12117 
12118   // Reload IP
12119   if (PVT == MVT::i64) {
12120     MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), Tmp)
12121             .addImm(LabelOffset)
12122             .addReg(BufReg);
12123   } else {
12124     MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), Tmp)
12125             .addImm(LabelOffset)
12126             .addReg(BufReg);
12127   }
12128   MIB.cloneMemRefs(MI);
12129 
12130   // Reload SP
12131   if (PVT == MVT::i64) {
12132     MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), SP)
12133             .addImm(SPOffset)
12134             .addReg(BufReg);
12135   } else {
12136     MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), SP)
12137             .addImm(SPOffset)
12138             .addReg(BufReg);
12139   }
12140   MIB.cloneMemRefs(MI);
12141 
12142   // Reload BP
12143   if (PVT == MVT::i64) {
12144     MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), BP)
12145             .addImm(BPOffset)
12146             .addReg(BufReg);
12147   } else {
12148     MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), BP)
12149             .addImm(BPOffset)
12150             .addReg(BufReg);
12151   }
12152   MIB.cloneMemRefs(MI);
12153 
12154   // Reload TOC
12155   if (PVT == MVT::i64 && Subtarget.isSVR4ABI()) {
12156     setUsesTOCBasePtr(*MBB->getParent());
12157     MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), PPC::X2)
12158               .addImm(TOCOffset)
12159               .addReg(BufReg)
12160               .cloneMemRefs(MI);
12161   }
12162 
12163   // Jump
12164   BuildMI(*MBB, MI, DL,
12165           TII->get(PVT == MVT::i64 ? PPC::MTCTR8 : PPC::MTCTR)).addReg(Tmp);
12166   BuildMI(*MBB, MI, DL, TII->get(PVT == MVT::i64 ? PPC::BCTR8 : PPC::BCTR));
12167 
12168   MI.eraseFromParent();
12169   return MBB;
12170 }
12171 
12172 bool PPCTargetLowering::hasInlineStackProbe(const MachineFunction &MF) const {
12173   // If the function specifically requests inline stack probes, emit them.
12174   if (MF.getFunction().hasFnAttribute("probe-stack"))
12175     return MF.getFunction().getFnAttribute("probe-stack").getValueAsString() ==
12176            "inline-asm";
12177   return false;
12178 }
12179 
12180 unsigned PPCTargetLowering::getStackProbeSize(const MachineFunction &MF) const {
12181   const TargetFrameLowering *TFI = Subtarget.getFrameLowering();
12182   unsigned StackAlign = TFI->getStackAlignment();
12183   assert(StackAlign >= 1 && isPowerOf2_32(StackAlign) &&
12184          "Unexpected stack alignment");
12185   // The default stack probe size is 4096 if the function has no
12186   // stack-probe-size attribute.
12187   const Function &Fn = MF.getFunction();
12188   unsigned StackProbeSize =
12189       Fn.getFnAttributeAsParsedInteger("stack-probe-size", 4096);
12190   // Round down to the stack alignment.
12191   StackProbeSize &= ~(StackAlign - 1);
12192   return StackProbeSize ? StackProbeSize : StackAlign;
12193 }
12194 
12195 // Lower dynamic stack allocation with probing. `emitProbedAlloca` is splitted
12196 // into three phases. In the first phase, it uses pseudo instruction
12197 // PREPARE_PROBED_ALLOCA to get the future result of actual FramePointer and
12198 // FinalStackPtr. In the second phase, it generates a loop for probing blocks.
12199 // At last, it uses pseudo instruction DYNAREAOFFSET to get the future result of
12200 // MaxCallFrameSize so that it can calculate correct data area pointer.
12201 MachineBasicBlock *
12202 PPCTargetLowering::emitProbedAlloca(MachineInstr &MI,
12203                                     MachineBasicBlock *MBB) const {
12204   const bool isPPC64 = Subtarget.isPPC64();
12205   MachineFunction *MF = MBB->getParent();
12206   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
12207   DebugLoc DL = MI.getDebugLoc();
12208   const unsigned ProbeSize = getStackProbeSize(*MF);
12209   const BasicBlock *ProbedBB = MBB->getBasicBlock();
12210   MachineRegisterInfo &MRI = MF->getRegInfo();
12211   // The CFG of probing stack looks as
12212   //         +-----+
12213   //         | MBB |
12214   //         +--+--+
12215   //            |
12216   //       +----v----+
12217   //  +--->+ TestMBB +---+
12218   //  |    +----+----+   |
12219   //  |         |        |
12220   //  |   +-----v----+   |
12221   //  +---+ BlockMBB |   |
12222   //      +----------+   |
12223   //                     |
12224   //       +---------+   |
12225   //       | TailMBB +<--+
12226   //       +---------+
12227   // In MBB, calculate previous frame pointer and final stack pointer.
12228   // In TestMBB, test if sp is equal to final stack pointer, if so, jump to
12229   // TailMBB. In BlockMBB, update the sp atomically and jump back to TestMBB.
12230   // TailMBB is spliced via \p MI.
12231   MachineBasicBlock *TestMBB = MF->CreateMachineBasicBlock(ProbedBB);
12232   MachineBasicBlock *TailMBB = MF->CreateMachineBasicBlock(ProbedBB);
12233   MachineBasicBlock *BlockMBB = MF->CreateMachineBasicBlock(ProbedBB);
12234 
12235   MachineFunction::iterator MBBIter = ++MBB->getIterator();
12236   MF->insert(MBBIter, TestMBB);
12237   MF->insert(MBBIter, BlockMBB);
12238   MF->insert(MBBIter, TailMBB);
12239 
12240   const TargetRegisterClass *G8RC = &PPC::G8RCRegClass;
12241   const TargetRegisterClass *GPRC = &PPC::GPRCRegClass;
12242 
12243   Register DstReg = MI.getOperand(0).getReg();
12244   Register NegSizeReg = MI.getOperand(1).getReg();
12245   Register SPReg = isPPC64 ? PPC::X1 : PPC::R1;
12246   Register FinalStackPtr = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC);
12247   Register FramePointer = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC);
12248   Register ActualNegSizeReg = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC);
12249 
12250   // Since value of NegSizeReg might be realigned in prologepilog, insert a
12251   // PREPARE_PROBED_ALLOCA pseudo instruction to get actual FramePointer and
12252   // NegSize.
12253   unsigned ProbeOpc;
12254   if (!MRI.hasOneNonDBGUse(NegSizeReg))
12255     ProbeOpc =
12256         isPPC64 ? PPC::PREPARE_PROBED_ALLOCA_64 : PPC::PREPARE_PROBED_ALLOCA_32;
12257   else
12258     // By introducing PREPARE_PROBED_ALLOCA_NEGSIZE_OPT, ActualNegSizeReg
12259     // and NegSizeReg will be allocated in the same phyreg to avoid
12260     // redundant copy when NegSizeReg has only one use which is current MI and
12261     // will be replaced by PREPARE_PROBED_ALLOCA then.
12262     ProbeOpc = isPPC64 ? PPC::PREPARE_PROBED_ALLOCA_NEGSIZE_SAME_REG_64
12263                        : PPC::PREPARE_PROBED_ALLOCA_NEGSIZE_SAME_REG_32;
12264   BuildMI(*MBB, {MI}, DL, TII->get(ProbeOpc), FramePointer)
12265       .addDef(ActualNegSizeReg)
12266       .addReg(NegSizeReg)
12267       .add(MI.getOperand(2))
12268       .add(MI.getOperand(3));
12269 
12270   // Calculate final stack pointer, which equals to SP + ActualNegSize.
12271   BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::ADD8 : PPC::ADD4),
12272           FinalStackPtr)
12273       .addReg(SPReg)
12274       .addReg(ActualNegSizeReg);
12275 
12276   // Materialize a scratch register for update.
12277   int64_t NegProbeSize = -(int64_t)ProbeSize;
12278   assert(isInt<32>(NegProbeSize) && "Unhandled probe size!");
12279   Register ScratchReg = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC);
12280   if (!isInt<16>(NegProbeSize)) {
12281     Register TempReg = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC);
12282     BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::LIS8 : PPC::LIS), TempReg)
12283         .addImm(NegProbeSize >> 16);
12284     BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::ORI8 : PPC::ORI),
12285             ScratchReg)
12286         .addReg(TempReg)
12287         .addImm(NegProbeSize & 0xFFFF);
12288   } else
12289     BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::LI8 : PPC::LI), ScratchReg)
12290         .addImm(NegProbeSize);
12291 
12292   {
12293     // Probing leading residual part.
12294     Register Div = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC);
12295     BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::DIVD : PPC::DIVW), Div)
12296         .addReg(ActualNegSizeReg)
12297         .addReg(ScratchReg);
12298     Register Mul = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC);
12299     BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::MULLD : PPC::MULLW), Mul)
12300         .addReg(Div)
12301         .addReg(ScratchReg);
12302     Register NegMod = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC);
12303     BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::SUBF8 : PPC::SUBF), NegMod)
12304         .addReg(Mul)
12305         .addReg(ActualNegSizeReg);
12306     BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::STDUX : PPC::STWUX), SPReg)
12307         .addReg(FramePointer)
12308         .addReg(SPReg)
12309         .addReg(NegMod);
12310   }
12311 
12312   {
12313     // Remaining part should be multiple of ProbeSize.
12314     Register CmpResult = MRI.createVirtualRegister(&PPC::CRRCRegClass);
12315     BuildMI(TestMBB, DL, TII->get(isPPC64 ? PPC::CMPD : PPC::CMPW), CmpResult)
12316         .addReg(SPReg)
12317         .addReg(FinalStackPtr);
12318     BuildMI(TestMBB, DL, TII->get(PPC::BCC))
12319         .addImm(PPC::PRED_EQ)
12320         .addReg(CmpResult)
12321         .addMBB(TailMBB);
12322     TestMBB->addSuccessor(BlockMBB);
12323     TestMBB->addSuccessor(TailMBB);
12324   }
12325 
12326   {
12327     // Touch the block.
12328     // |P...|P...|P...
12329     BuildMI(BlockMBB, DL, TII->get(isPPC64 ? PPC::STDUX : PPC::STWUX), SPReg)
12330         .addReg(FramePointer)
12331         .addReg(SPReg)
12332         .addReg(ScratchReg);
12333     BuildMI(BlockMBB, DL, TII->get(PPC::B)).addMBB(TestMBB);
12334     BlockMBB->addSuccessor(TestMBB);
12335   }
12336 
12337   // Calculation of MaxCallFrameSize is deferred to prologepilog, use
12338   // DYNAREAOFFSET pseudo instruction to get the future result.
12339   Register MaxCallFrameSizeReg =
12340       MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC);
12341   BuildMI(TailMBB, DL,
12342           TII->get(isPPC64 ? PPC::DYNAREAOFFSET8 : PPC::DYNAREAOFFSET),
12343           MaxCallFrameSizeReg)
12344       .add(MI.getOperand(2))
12345       .add(MI.getOperand(3));
12346   BuildMI(TailMBB, DL, TII->get(isPPC64 ? PPC::ADD8 : PPC::ADD4), DstReg)
12347       .addReg(SPReg)
12348       .addReg(MaxCallFrameSizeReg);
12349 
12350   // Splice instructions after MI to TailMBB.
12351   TailMBB->splice(TailMBB->end(), MBB,
12352                   std::next(MachineBasicBlock::iterator(MI)), MBB->end());
12353   TailMBB->transferSuccessorsAndUpdatePHIs(MBB);
12354   MBB->addSuccessor(TestMBB);
12355 
12356   // Delete the pseudo instruction.
12357   MI.eraseFromParent();
12358 
12359   ++NumDynamicAllocaProbed;
12360   return TailMBB;
12361 }
12362 
12363 MachineBasicBlock *
12364 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
12365                                                MachineBasicBlock *BB) const {
12366   if (MI.getOpcode() == TargetOpcode::STACKMAP ||
12367       MI.getOpcode() == TargetOpcode::PATCHPOINT) {
12368     if (Subtarget.is64BitELFABI() &&
12369         MI.getOpcode() == TargetOpcode::PATCHPOINT &&
12370         !Subtarget.isUsingPCRelativeCalls()) {
12371       // Call lowering should have added an r2 operand to indicate a dependence
12372       // on the TOC base pointer value. It can't however, because there is no
12373       // way to mark the dependence as implicit there, and so the stackmap code
12374       // will confuse it with a regular operand. Instead, add the dependence
12375       // here.
12376       MI.addOperand(MachineOperand::CreateReg(PPC::X2, false, true));
12377     }
12378 
12379     return emitPatchPoint(MI, BB);
12380   }
12381 
12382   if (MI.getOpcode() == PPC::EH_SjLj_SetJmp32 ||
12383       MI.getOpcode() == PPC::EH_SjLj_SetJmp64) {
12384     return emitEHSjLjSetJmp(MI, BB);
12385   } else if (MI.getOpcode() == PPC::EH_SjLj_LongJmp32 ||
12386              MI.getOpcode() == PPC::EH_SjLj_LongJmp64) {
12387     return emitEHSjLjLongJmp(MI, BB);
12388   }
12389 
12390   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
12391 
12392   // To "insert" these instructions we actually have to insert their
12393   // control-flow patterns.
12394   const BasicBlock *LLVM_BB = BB->getBasicBlock();
12395   MachineFunction::iterator It = ++BB->getIterator();
12396 
12397   MachineFunction *F = BB->getParent();
12398   MachineRegisterInfo &MRI = F->getRegInfo();
12399 
12400   if (MI.getOpcode() == PPC::SELECT_CC_I4 ||
12401       MI.getOpcode() == PPC::SELECT_CC_I8 || MI.getOpcode() == PPC::SELECT_I4 ||
12402       MI.getOpcode() == PPC::SELECT_I8) {
12403     SmallVector<MachineOperand, 2> Cond;
12404     if (MI.getOpcode() == PPC::SELECT_CC_I4 ||
12405         MI.getOpcode() == PPC::SELECT_CC_I8)
12406       Cond.push_back(MI.getOperand(4));
12407     else
12408       Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET));
12409     Cond.push_back(MI.getOperand(1));
12410 
12411     DebugLoc dl = MI.getDebugLoc();
12412     TII->insertSelect(*BB, MI, dl, MI.getOperand(0).getReg(), Cond,
12413                       MI.getOperand(2).getReg(), MI.getOperand(3).getReg());
12414   } else if (MI.getOpcode() == PPC::SELECT_CC_F4 ||
12415              MI.getOpcode() == PPC::SELECT_CC_F8 ||
12416              MI.getOpcode() == PPC::SELECT_CC_F16 ||
12417              MI.getOpcode() == PPC::SELECT_CC_VRRC ||
12418              MI.getOpcode() == PPC::SELECT_CC_VSFRC ||
12419              MI.getOpcode() == PPC::SELECT_CC_VSSRC ||
12420              MI.getOpcode() == PPC::SELECT_CC_VSRC ||
12421              MI.getOpcode() == PPC::SELECT_CC_SPE4 ||
12422              MI.getOpcode() == PPC::SELECT_CC_SPE ||
12423              MI.getOpcode() == PPC::SELECT_F4 ||
12424              MI.getOpcode() == PPC::SELECT_F8 ||
12425              MI.getOpcode() == PPC::SELECT_F16 ||
12426              MI.getOpcode() == PPC::SELECT_SPE ||
12427              MI.getOpcode() == PPC::SELECT_SPE4 ||
12428              MI.getOpcode() == PPC::SELECT_VRRC ||
12429              MI.getOpcode() == PPC::SELECT_VSFRC ||
12430              MI.getOpcode() == PPC::SELECT_VSSRC ||
12431              MI.getOpcode() == PPC::SELECT_VSRC) {
12432     // The incoming instruction knows the destination vreg to set, the
12433     // condition code register to branch on, the true/false values to
12434     // select between, and a branch opcode to use.
12435 
12436     //  thisMBB:
12437     //  ...
12438     //   TrueVal = ...
12439     //   cmpTY ccX, r1, r2
12440     //   bCC copy1MBB
12441     //   fallthrough --> copy0MBB
12442     MachineBasicBlock *thisMBB = BB;
12443     MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
12444     MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
12445     DebugLoc dl = MI.getDebugLoc();
12446     F->insert(It, copy0MBB);
12447     F->insert(It, sinkMBB);
12448 
12449     // Transfer the remainder of BB and its successor edges to sinkMBB.
12450     sinkMBB->splice(sinkMBB->begin(), BB,
12451                     std::next(MachineBasicBlock::iterator(MI)), BB->end());
12452     sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
12453 
12454     // Next, add the true and fallthrough blocks as its successors.
12455     BB->addSuccessor(copy0MBB);
12456     BB->addSuccessor(sinkMBB);
12457 
12458     if (MI.getOpcode() == PPC::SELECT_I4 || MI.getOpcode() == PPC::SELECT_I8 ||
12459         MI.getOpcode() == PPC::SELECT_F4 || MI.getOpcode() == PPC::SELECT_F8 ||
12460         MI.getOpcode() == PPC::SELECT_F16 ||
12461         MI.getOpcode() == PPC::SELECT_SPE4 ||
12462         MI.getOpcode() == PPC::SELECT_SPE ||
12463         MI.getOpcode() == PPC::SELECT_VRRC ||
12464         MI.getOpcode() == PPC::SELECT_VSFRC ||
12465         MI.getOpcode() == PPC::SELECT_VSSRC ||
12466         MI.getOpcode() == PPC::SELECT_VSRC) {
12467       BuildMI(BB, dl, TII->get(PPC::BC))
12468           .addReg(MI.getOperand(1).getReg())
12469           .addMBB(sinkMBB);
12470     } else {
12471       unsigned SelectPred = MI.getOperand(4).getImm();
12472       BuildMI(BB, dl, TII->get(PPC::BCC))
12473           .addImm(SelectPred)
12474           .addReg(MI.getOperand(1).getReg())
12475           .addMBB(sinkMBB);
12476     }
12477 
12478     //  copy0MBB:
12479     //   %FalseValue = ...
12480     //   # fallthrough to sinkMBB
12481     BB = copy0MBB;
12482 
12483     // Update machine-CFG edges
12484     BB->addSuccessor(sinkMBB);
12485 
12486     //  sinkMBB:
12487     //   %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
12488     //  ...
12489     BB = sinkMBB;
12490     BuildMI(*BB, BB->begin(), dl, TII->get(PPC::PHI), MI.getOperand(0).getReg())
12491         .addReg(MI.getOperand(3).getReg())
12492         .addMBB(copy0MBB)
12493         .addReg(MI.getOperand(2).getReg())
12494         .addMBB(thisMBB);
12495   } else if (MI.getOpcode() == PPC::ReadTB) {
12496     // To read the 64-bit time-base register on a 32-bit target, we read the
12497     // two halves. Should the counter have wrapped while it was being read, we
12498     // need to try again.
12499     // ...
12500     // readLoop:
12501     // mfspr Rx,TBU # load from TBU
12502     // mfspr Ry,TB  # load from TB
12503     // mfspr Rz,TBU # load from TBU
12504     // cmpw crX,Rx,Rz # check if 'old'='new'
12505     // bne readLoop   # branch if they're not equal
12506     // ...
12507 
12508     MachineBasicBlock *readMBB = F->CreateMachineBasicBlock(LLVM_BB);
12509     MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
12510     DebugLoc dl = MI.getDebugLoc();
12511     F->insert(It, readMBB);
12512     F->insert(It, sinkMBB);
12513 
12514     // Transfer the remainder of BB and its successor edges to sinkMBB.
12515     sinkMBB->splice(sinkMBB->begin(), BB,
12516                     std::next(MachineBasicBlock::iterator(MI)), BB->end());
12517     sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
12518 
12519     BB->addSuccessor(readMBB);
12520     BB = readMBB;
12521 
12522     MachineRegisterInfo &RegInfo = F->getRegInfo();
12523     Register ReadAgainReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
12524     Register LoReg = MI.getOperand(0).getReg();
12525     Register HiReg = MI.getOperand(1).getReg();
12526 
12527     BuildMI(BB, dl, TII->get(PPC::MFSPR), HiReg).addImm(269);
12528     BuildMI(BB, dl, TII->get(PPC::MFSPR), LoReg).addImm(268);
12529     BuildMI(BB, dl, TII->get(PPC::MFSPR), ReadAgainReg).addImm(269);
12530 
12531     Register CmpReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass);
12532 
12533     BuildMI(BB, dl, TII->get(PPC::CMPW), CmpReg)
12534         .addReg(HiReg)
12535         .addReg(ReadAgainReg);
12536     BuildMI(BB, dl, TII->get(PPC::BCC))
12537         .addImm(PPC::PRED_NE)
12538         .addReg(CmpReg)
12539         .addMBB(readMBB);
12540 
12541     BB->addSuccessor(readMBB);
12542     BB->addSuccessor(sinkMBB);
12543   } else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
12544     BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
12545   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
12546     BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
12547   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
12548     BB = EmitAtomicBinary(MI, BB, 4, PPC::ADD4);
12549   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
12550     BB = EmitAtomicBinary(MI, BB, 8, PPC::ADD8);
12551 
12552   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
12553     BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
12554   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
12555     BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
12556   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
12557     BB = EmitAtomicBinary(MI, BB, 4, PPC::AND);
12558   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
12559     BB = EmitAtomicBinary(MI, BB, 8, PPC::AND8);
12560 
12561   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
12562     BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
12563   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
12564     BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
12565   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
12566     BB = EmitAtomicBinary(MI, BB, 4, PPC::OR);
12567   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
12568     BB = EmitAtomicBinary(MI, BB, 8, PPC::OR8);
12569 
12570   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
12571     BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
12572   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
12573     BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
12574   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
12575     BB = EmitAtomicBinary(MI, BB, 4, PPC::XOR);
12576   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
12577     BB = EmitAtomicBinary(MI, BB, 8, PPC::XOR8);
12578 
12579   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
12580     BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::NAND);
12581   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
12582     BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::NAND);
12583   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
12584     BB = EmitAtomicBinary(MI, BB, 4, PPC::NAND);
12585   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
12586     BB = EmitAtomicBinary(MI, BB, 8, PPC::NAND8);
12587 
12588   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
12589     BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
12590   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
12591     BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
12592   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
12593     BB = EmitAtomicBinary(MI, BB, 4, PPC::SUBF);
12594   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
12595     BB = EmitAtomicBinary(MI, BB, 8, PPC::SUBF8);
12596 
12597   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I8)
12598     BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPW, PPC::PRED_LT);
12599   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I16)
12600     BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPW, PPC::PRED_LT);
12601   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I32)
12602     BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPW, PPC::PRED_LT);
12603   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I64)
12604     BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPD, PPC::PRED_LT);
12605 
12606   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I8)
12607     BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPW, PPC::PRED_GT);
12608   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I16)
12609     BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPW, PPC::PRED_GT);
12610   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I32)
12611     BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPW, PPC::PRED_GT);
12612   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I64)
12613     BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPD, PPC::PRED_GT);
12614 
12615   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I8)
12616     BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPLW, PPC::PRED_LT);
12617   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I16)
12618     BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPLW, PPC::PRED_LT);
12619   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I32)
12620     BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPLW, PPC::PRED_LT);
12621   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I64)
12622     BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPLD, PPC::PRED_LT);
12623 
12624   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I8)
12625     BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPLW, PPC::PRED_GT);
12626   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I16)
12627     BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPLW, PPC::PRED_GT);
12628   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I32)
12629     BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPLW, PPC::PRED_GT);
12630   else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I64)
12631     BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPLD, PPC::PRED_GT);
12632 
12633   else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I8)
12634     BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
12635   else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I16)
12636     BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
12637   else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I32)
12638     BB = EmitAtomicBinary(MI, BB, 4, 0);
12639   else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I64)
12640     BB = EmitAtomicBinary(MI, BB, 8, 0);
12641   else if (MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
12642            MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I64 ||
12643            (Subtarget.hasPartwordAtomics() &&
12644             MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8) ||
12645            (Subtarget.hasPartwordAtomics() &&
12646             MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I16)) {
12647     bool is64bit = MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
12648 
12649     auto LoadMnemonic = PPC::LDARX;
12650     auto StoreMnemonic = PPC::STDCX;
12651     switch (MI.getOpcode()) {
12652     default:
12653       llvm_unreachable("Compare and swap of unknown size");
12654     case PPC::ATOMIC_CMP_SWAP_I8:
12655       LoadMnemonic = PPC::LBARX;
12656       StoreMnemonic = PPC::STBCX;
12657       assert(Subtarget.hasPartwordAtomics() && "No support partword atomics.");
12658       break;
12659     case PPC::ATOMIC_CMP_SWAP_I16:
12660       LoadMnemonic = PPC::LHARX;
12661       StoreMnemonic = PPC::STHCX;
12662       assert(Subtarget.hasPartwordAtomics() && "No support partword atomics.");
12663       break;
12664     case PPC::ATOMIC_CMP_SWAP_I32:
12665       LoadMnemonic = PPC::LWARX;
12666       StoreMnemonic = PPC::STWCX;
12667       break;
12668     case PPC::ATOMIC_CMP_SWAP_I64:
12669       LoadMnemonic = PPC::LDARX;
12670       StoreMnemonic = PPC::STDCX;
12671       break;
12672     }
12673     MachineRegisterInfo &RegInfo = F->getRegInfo();
12674     Register dest = MI.getOperand(0).getReg();
12675     Register ptrA = MI.getOperand(1).getReg();
12676     Register ptrB = MI.getOperand(2).getReg();
12677     Register CrReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass);
12678     Register oldval = MI.getOperand(3).getReg();
12679     Register newval = MI.getOperand(4).getReg();
12680     DebugLoc dl = MI.getDebugLoc();
12681 
12682     MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
12683     MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
12684     MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
12685     F->insert(It, loop1MBB);
12686     F->insert(It, loop2MBB);
12687     F->insert(It, exitMBB);
12688     exitMBB->splice(exitMBB->begin(), BB,
12689                     std::next(MachineBasicBlock::iterator(MI)), BB->end());
12690     exitMBB->transferSuccessorsAndUpdatePHIs(BB);
12691 
12692     //  thisMBB:
12693     //   ...
12694     //   fallthrough --> loopMBB
12695     BB->addSuccessor(loop1MBB);
12696 
12697     // loop1MBB:
12698     //   l[bhwd]arx dest, ptr
12699     //   cmp[wd] dest, oldval
12700     //   bne- exitBB
12701     // loop2MBB:
12702     //   st[bhwd]cx. newval, ptr
12703     //   bne- loopMBB
12704     //   b exitBB
12705     // exitBB:
12706     BB = loop1MBB;
12707     BuildMI(BB, dl, TII->get(LoadMnemonic), dest).addReg(ptrA).addReg(ptrB);
12708     BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), CrReg)
12709         .addReg(dest)
12710         .addReg(oldval);
12711     BuildMI(BB, dl, TII->get(PPC::BCC))
12712         .addImm(PPC::PRED_NE)
12713         .addReg(CrReg)
12714         .addMBB(exitMBB);
12715     BB->addSuccessor(loop2MBB);
12716     BB->addSuccessor(exitMBB);
12717 
12718     BB = loop2MBB;
12719     BuildMI(BB, dl, TII->get(StoreMnemonic))
12720         .addReg(newval)
12721         .addReg(ptrA)
12722         .addReg(ptrB);
12723     BuildMI(BB, dl, TII->get(PPC::BCC))
12724         .addImm(PPC::PRED_NE)
12725         .addReg(PPC::CR0)
12726         .addMBB(loop1MBB);
12727     BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
12728     BB->addSuccessor(loop1MBB);
12729     BB->addSuccessor(exitMBB);
12730 
12731     //  exitMBB:
12732     //   ...
12733     BB = exitMBB;
12734   } else if (MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
12735              MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
12736     // We must use 64-bit registers for addresses when targeting 64-bit,
12737     // since we're actually doing arithmetic on them.  Other registers
12738     // can be 32-bit.
12739     bool is64bit = Subtarget.isPPC64();
12740     bool isLittleEndian = Subtarget.isLittleEndian();
12741     bool is8bit = MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
12742 
12743     Register dest = MI.getOperand(0).getReg();
12744     Register ptrA = MI.getOperand(1).getReg();
12745     Register ptrB = MI.getOperand(2).getReg();
12746     Register oldval = MI.getOperand(3).getReg();
12747     Register newval = MI.getOperand(4).getReg();
12748     DebugLoc dl = MI.getDebugLoc();
12749 
12750     MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
12751     MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
12752     MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
12753     F->insert(It, loop1MBB);
12754     F->insert(It, loop2MBB);
12755     F->insert(It, exitMBB);
12756     exitMBB->splice(exitMBB->begin(), BB,
12757                     std::next(MachineBasicBlock::iterator(MI)), BB->end());
12758     exitMBB->transferSuccessorsAndUpdatePHIs(BB);
12759 
12760     MachineRegisterInfo &RegInfo = F->getRegInfo();
12761     const TargetRegisterClass *RC =
12762         is64bit ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
12763     const TargetRegisterClass *GPRC = &PPC::GPRCRegClass;
12764 
12765     Register PtrReg = RegInfo.createVirtualRegister(RC);
12766     Register Shift1Reg = RegInfo.createVirtualRegister(GPRC);
12767     Register ShiftReg =
12768         isLittleEndian ? Shift1Reg : RegInfo.createVirtualRegister(GPRC);
12769     Register NewVal2Reg = RegInfo.createVirtualRegister(GPRC);
12770     Register NewVal3Reg = RegInfo.createVirtualRegister(GPRC);
12771     Register OldVal2Reg = RegInfo.createVirtualRegister(GPRC);
12772     Register OldVal3Reg = RegInfo.createVirtualRegister(GPRC);
12773     Register MaskReg = RegInfo.createVirtualRegister(GPRC);
12774     Register Mask2Reg = RegInfo.createVirtualRegister(GPRC);
12775     Register Mask3Reg = RegInfo.createVirtualRegister(GPRC);
12776     Register Tmp2Reg = RegInfo.createVirtualRegister(GPRC);
12777     Register Tmp4Reg = RegInfo.createVirtualRegister(GPRC);
12778     Register TmpDestReg = RegInfo.createVirtualRegister(GPRC);
12779     Register Ptr1Reg;
12780     Register TmpReg = RegInfo.createVirtualRegister(GPRC);
12781     Register ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO;
12782     Register CrReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass);
12783     //  thisMBB:
12784     //   ...
12785     //   fallthrough --> loopMBB
12786     BB->addSuccessor(loop1MBB);
12787 
12788     // The 4-byte load must be aligned, while a char or short may be
12789     // anywhere in the word.  Hence all this nasty bookkeeping code.
12790     //   add ptr1, ptrA, ptrB [copy if ptrA==0]
12791     //   rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
12792     //   xori shift, shift1, 24 [16]
12793     //   rlwinm ptr, ptr1, 0, 0, 29
12794     //   slw newval2, newval, shift
12795     //   slw oldval2, oldval,shift
12796     //   li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
12797     //   slw mask, mask2, shift
12798     //   and newval3, newval2, mask
12799     //   and oldval3, oldval2, mask
12800     // loop1MBB:
12801     //   lwarx tmpDest, ptr
12802     //   and tmp, tmpDest, mask
12803     //   cmpw tmp, oldval3
12804     //   bne- exitBB
12805     // loop2MBB:
12806     //   andc tmp2, tmpDest, mask
12807     //   or tmp4, tmp2, newval3
12808     //   stwcx. tmp4, ptr
12809     //   bne- loop1MBB
12810     //   b exitBB
12811     // exitBB:
12812     //   srw dest, tmpDest, shift
12813     if (ptrA != ZeroReg) {
12814       Ptr1Reg = RegInfo.createVirtualRegister(RC);
12815       BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
12816           .addReg(ptrA)
12817           .addReg(ptrB);
12818     } else {
12819       Ptr1Reg = ptrB;
12820     }
12821 
12822     // We need use 32-bit subregister to avoid mismatch register class in 64-bit
12823     // mode.
12824     BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg)
12825         .addReg(Ptr1Reg, 0, is64bit ? PPC::sub_32 : 0)
12826         .addImm(3)
12827         .addImm(27)
12828         .addImm(is8bit ? 28 : 27);
12829     if (!isLittleEndian)
12830       BuildMI(BB, dl, TII->get(PPC::XORI), ShiftReg)
12831           .addReg(Shift1Reg)
12832           .addImm(is8bit ? 24 : 16);
12833     if (is64bit)
12834       BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
12835           .addReg(Ptr1Reg)
12836           .addImm(0)
12837           .addImm(61);
12838     else
12839       BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
12840           .addReg(Ptr1Reg)
12841           .addImm(0)
12842           .addImm(0)
12843           .addImm(29);
12844     BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
12845         .addReg(newval)
12846         .addReg(ShiftReg);
12847     BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
12848         .addReg(oldval)
12849         .addReg(ShiftReg);
12850     if (is8bit)
12851       BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
12852     else {
12853       BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
12854       BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
12855           .addReg(Mask3Reg)
12856           .addImm(65535);
12857     }
12858     BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
12859         .addReg(Mask2Reg)
12860         .addReg(ShiftReg);
12861     BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
12862         .addReg(NewVal2Reg)
12863         .addReg(MaskReg);
12864     BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
12865         .addReg(OldVal2Reg)
12866         .addReg(MaskReg);
12867 
12868     BB = loop1MBB;
12869     BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
12870         .addReg(ZeroReg)
12871         .addReg(PtrReg);
12872     BuildMI(BB, dl, TII->get(PPC::AND), TmpReg)
12873         .addReg(TmpDestReg)
12874         .addReg(MaskReg);
12875     BuildMI(BB, dl, TII->get(PPC::CMPW), CrReg)
12876         .addReg(TmpReg)
12877         .addReg(OldVal3Reg);
12878     BuildMI(BB, dl, TII->get(PPC::BCC))
12879         .addImm(PPC::PRED_NE)
12880         .addReg(CrReg)
12881         .addMBB(exitMBB);
12882     BB->addSuccessor(loop2MBB);
12883     BB->addSuccessor(exitMBB);
12884 
12885     BB = loop2MBB;
12886     BuildMI(BB, dl, TII->get(PPC::ANDC), Tmp2Reg)
12887         .addReg(TmpDestReg)
12888         .addReg(MaskReg);
12889     BuildMI(BB, dl, TII->get(PPC::OR), Tmp4Reg)
12890         .addReg(Tmp2Reg)
12891         .addReg(NewVal3Reg);
12892     BuildMI(BB, dl, TII->get(PPC::STWCX))
12893         .addReg(Tmp4Reg)
12894         .addReg(ZeroReg)
12895         .addReg(PtrReg);
12896     BuildMI(BB, dl, TII->get(PPC::BCC))
12897         .addImm(PPC::PRED_NE)
12898         .addReg(PPC::CR0)
12899         .addMBB(loop1MBB);
12900     BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
12901     BB->addSuccessor(loop1MBB);
12902     BB->addSuccessor(exitMBB);
12903 
12904     //  exitMBB:
12905     //   ...
12906     BB = exitMBB;
12907     BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest)
12908         .addReg(TmpReg)
12909         .addReg(ShiftReg);
12910   } else if (MI.getOpcode() == PPC::FADDrtz) {
12911     // This pseudo performs an FADD with rounding mode temporarily forced
12912     // to round-to-zero.  We emit this via custom inserter since the FPSCR
12913     // is not modeled at the SelectionDAG level.
12914     Register Dest = MI.getOperand(0).getReg();
12915     Register Src1 = MI.getOperand(1).getReg();
12916     Register Src2 = MI.getOperand(2).getReg();
12917     DebugLoc dl = MI.getDebugLoc();
12918 
12919     MachineRegisterInfo &RegInfo = F->getRegInfo();
12920     Register MFFSReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
12921 
12922     // Save FPSCR value.
12923     BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), MFFSReg);
12924 
12925     // Set rounding mode to round-to-zero.
12926     BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB1))
12927         .addImm(31)
12928         .addReg(PPC::RM, RegState::ImplicitDefine);
12929 
12930     BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB0))
12931         .addImm(30)
12932         .addReg(PPC::RM, RegState::ImplicitDefine);
12933 
12934     // Perform addition.
12935     auto MIB = BuildMI(*BB, MI, dl, TII->get(PPC::FADD), Dest)
12936                    .addReg(Src1)
12937                    .addReg(Src2);
12938     if (MI.getFlag(MachineInstr::NoFPExcept))
12939       MIB.setMIFlag(MachineInstr::NoFPExcept);
12940 
12941     // Restore FPSCR value.
12942     BuildMI(*BB, MI, dl, TII->get(PPC::MTFSFb)).addImm(1).addReg(MFFSReg);
12943   } else if (MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT ||
12944              MI.getOpcode() == PPC::ANDI_rec_1_GT_BIT ||
12945              MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT8 ||
12946              MI.getOpcode() == PPC::ANDI_rec_1_GT_BIT8) {
12947     unsigned Opcode = (MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT8 ||
12948                        MI.getOpcode() == PPC::ANDI_rec_1_GT_BIT8)
12949                           ? PPC::ANDI8_rec
12950                           : PPC::ANDI_rec;
12951     bool IsEQ = (MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT ||
12952                  MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT8);
12953 
12954     MachineRegisterInfo &RegInfo = F->getRegInfo();
12955     Register Dest = RegInfo.createVirtualRegister(
12956         Opcode == PPC::ANDI_rec ? &PPC::GPRCRegClass : &PPC::G8RCRegClass);
12957 
12958     DebugLoc Dl = MI.getDebugLoc();
12959     BuildMI(*BB, MI, Dl, TII->get(Opcode), Dest)
12960         .addReg(MI.getOperand(1).getReg())
12961         .addImm(1);
12962     BuildMI(*BB, MI, Dl, TII->get(TargetOpcode::COPY),
12963             MI.getOperand(0).getReg())
12964         .addReg(IsEQ ? PPC::CR0EQ : PPC::CR0GT);
12965   } else if (MI.getOpcode() == PPC::TCHECK_RET) {
12966     DebugLoc Dl = MI.getDebugLoc();
12967     MachineRegisterInfo &RegInfo = F->getRegInfo();
12968     Register CRReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass);
12969     BuildMI(*BB, MI, Dl, TII->get(PPC::TCHECK), CRReg);
12970     BuildMI(*BB, MI, Dl, TII->get(TargetOpcode::COPY),
12971             MI.getOperand(0).getReg())
12972         .addReg(CRReg);
12973   } else if (MI.getOpcode() == PPC::TBEGIN_RET) {
12974     DebugLoc Dl = MI.getDebugLoc();
12975     unsigned Imm = MI.getOperand(1).getImm();
12976     BuildMI(*BB, MI, Dl, TII->get(PPC::TBEGIN)).addImm(Imm);
12977     BuildMI(*BB, MI, Dl, TII->get(TargetOpcode::COPY),
12978             MI.getOperand(0).getReg())
12979         .addReg(PPC::CR0EQ);
12980   } else if (MI.getOpcode() == PPC::SETRNDi) {
12981     DebugLoc dl = MI.getDebugLoc();
12982     Register OldFPSCRReg = MI.getOperand(0).getReg();
12983 
12984     // Save FPSCR value.
12985     if (MRI.use_empty(OldFPSCRReg))
12986       BuildMI(*BB, MI, dl, TII->get(TargetOpcode::IMPLICIT_DEF), OldFPSCRReg);
12987     else
12988       BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), OldFPSCRReg);
12989 
12990     // The floating point rounding mode is in the bits 62:63 of FPCSR, and has
12991     // the following settings:
12992     //   00 Round to nearest
12993     //   01 Round to 0
12994     //   10 Round to +inf
12995     //   11 Round to -inf
12996 
12997     // When the operand is immediate, using the two least significant bits of
12998     // the immediate to set the bits 62:63 of FPSCR.
12999     unsigned Mode = MI.getOperand(1).getImm();
13000     BuildMI(*BB, MI, dl, TII->get((Mode & 1) ? PPC::MTFSB1 : PPC::MTFSB0))
13001         .addImm(31)
13002         .addReg(PPC::RM, RegState::ImplicitDefine);
13003 
13004     BuildMI(*BB, MI, dl, TII->get((Mode & 2) ? PPC::MTFSB1 : PPC::MTFSB0))
13005         .addImm(30)
13006         .addReg(PPC::RM, RegState::ImplicitDefine);
13007   } else if (MI.getOpcode() == PPC::SETRND) {
13008     DebugLoc dl = MI.getDebugLoc();
13009 
13010     // Copy register from F8RCRegClass::SrcReg to G8RCRegClass::DestReg
13011     // or copy register from G8RCRegClass::SrcReg to F8RCRegClass::DestReg.
13012     // If the target doesn't have DirectMove, we should use stack to do the
13013     // conversion, because the target doesn't have the instructions like mtvsrd
13014     // or mfvsrd to do this conversion directly.
13015     auto copyRegFromG8RCOrF8RC = [&] (unsigned DestReg, unsigned SrcReg) {
13016       if (Subtarget.hasDirectMove()) {
13017         BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), DestReg)
13018           .addReg(SrcReg);
13019       } else {
13020         // Use stack to do the register copy.
13021         unsigned StoreOp = PPC::STD, LoadOp = PPC::LFD;
13022         MachineRegisterInfo &RegInfo = F->getRegInfo();
13023         const TargetRegisterClass *RC = RegInfo.getRegClass(SrcReg);
13024         if (RC == &PPC::F8RCRegClass) {
13025           // Copy register from F8RCRegClass to G8RCRegclass.
13026           assert((RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) &&
13027                  "Unsupported RegClass.");
13028 
13029           StoreOp = PPC::STFD;
13030           LoadOp = PPC::LD;
13031         } else {
13032           // Copy register from G8RCRegClass to F8RCRegclass.
13033           assert((RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) &&
13034                  (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) &&
13035                  "Unsupported RegClass.");
13036         }
13037 
13038         MachineFrameInfo &MFI = F->getFrameInfo();
13039         int FrameIdx = MFI.CreateStackObject(8, Align(8), false);
13040 
13041         MachineMemOperand *MMOStore = F->getMachineMemOperand(
13042             MachinePointerInfo::getFixedStack(*F, FrameIdx, 0),
13043             MachineMemOperand::MOStore, MFI.getObjectSize(FrameIdx),
13044             MFI.getObjectAlign(FrameIdx));
13045 
13046         // Store the SrcReg into the stack.
13047         BuildMI(*BB, MI, dl, TII->get(StoreOp))
13048           .addReg(SrcReg)
13049           .addImm(0)
13050           .addFrameIndex(FrameIdx)
13051           .addMemOperand(MMOStore);
13052 
13053         MachineMemOperand *MMOLoad = F->getMachineMemOperand(
13054             MachinePointerInfo::getFixedStack(*F, FrameIdx, 0),
13055             MachineMemOperand::MOLoad, MFI.getObjectSize(FrameIdx),
13056             MFI.getObjectAlign(FrameIdx));
13057 
13058         // Load from the stack where SrcReg is stored, and save to DestReg,
13059         // so we have done the RegClass conversion from RegClass::SrcReg to
13060         // RegClass::DestReg.
13061         BuildMI(*BB, MI, dl, TII->get(LoadOp), DestReg)
13062           .addImm(0)
13063           .addFrameIndex(FrameIdx)
13064           .addMemOperand(MMOLoad);
13065       }
13066     };
13067 
13068     Register OldFPSCRReg = MI.getOperand(0).getReg();
13069 
13070     // Save FPSCR value.
13071     BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), OldFPSCRReg);
13072 
13073     // When the operand is gprc register, use two least significant bits of the
13074     // register and mtfsf instruction to set the bits 62:63 of FPSCR.
13075     //
13076     // copy OldFPSCRTmpReg, OldFPSCRReg
13077     // (INSERT_SUBREG ExtSrcReg, (IMPLICIT_DEF ImDefReg), SrcOp, 1)
13078     // rldimi NewFPSCRTmpReg, ExtSrcReg, OldFPSCRReg, 0, 62
13079     // copy NewFPSCRReg, NewFPSCRTmpReg
13080     // mtfsf 255, NewFPSCRReg
13081     MachineOperand SrcOp = MI.getOperand(1);
13082     MachineRegisterInfo &RegInfo = F->getRegInfo();
13083     Register OldFPSCRTmpReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
13084 
13085     copyRegFromG8RCOrF8RC(OldFPSCRTmpReg, OldFPSCRReg);
13086 
13087     Register ImDefReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
13088     Register ExtSrcReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
13089 
13090     // The first operand of INSERT_SUBREG should be a register which has
13091     // subregisters, we only care about its RegClass, so we should use an
13092     // IMPLICIT_DEF register.
13093     BuildMI(*BB, MI, dl, TII->get(TargetOpcode::IMPLICIT_DEF), ImDefReg);
13094     BuildMI(*BB, MI, dl, TII->get(PPC::INSERT_SUBREG), ExtSrcReg)
13095       .addReg(ImDefReg)
13096       .add(SrcOp)
13097       .addImm(1);
13098 
13099     Register NewFPSCRTmpReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
13100     BuildMI(*BB, MI, dl, TII->get(PPC::RLDIMI), NewFPSCRTmpReg)
13101       .addReg(OldFPSCRTmpReg)
13102       .addReg(ExtSrcReg)
13103       .addImm(0)
13104       .addImm(62);
13105 
13106     Register NewFPSCRReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
13107     copyRegFromG8RCOrF8RC(NewFPSCRReg, NewFPSCRTmpReg);
13108 
13109     // The mask 255 means that put the 32:63 bits of NewFPSCRReg to the 32:63
13110     // bits of FPSCR.
13111     BuildMI(*BB, MI, dl, TII->get(PPC::MTFSF))
13112       .addImm(255)
13113       .addReg(NewFPSCRReg)
13114       .addImm(0)
13115       .addImm(0);
13116   } else if (MI.getOpcode() == PPC::SETFLM) {
13117     DebugLoc Dl = MI.getDebugLoc();
13118 
13119     // Result of setflm is previous FPSCR content, so we need to save it first.
13120     Register OldFPSCRReg = MI.getOperand(0).getReg();
13121     if (MRI.use_empty(OldFPSCRReg))
13122       BuildMI(*BB, MI, Dl, TII->get(TargetOpcode::IMPLICIT_DEF), OldFPSCRReg);
13123     else
13124       BuildMI(*BB, MI, Dl, TII->get(PPC::MFFS), OldFPSCRReg);
13125 
13126     // Put bits in 32:63 to FPSCR.
13127     Register NewFPSCRReg = MI.getOperand(1).getReg();
13128     BuildMI(*BB, MI, Dl, TII->get(PPC::MTFSF))
13129         .addImm(255)
13130         .addReg(NewFPSCRReg)
13131         .addImm(0)
13132         .addImm(0);
13133   } else if (MI.getOpcode() == PPC::PROBED_ALLOCA_32 ||
13134              MI.getOpcode() == PPC::PROBED_ALLOCA_64) {
13135     return emitProbedAlloca(MI, BB);
13136   } else if (MI.getOpcode() == PPC::SPLIT_QUADWORD) {
13137     DebugLoc DL = MI.getDebugLoc();
13138     Register Src = MI.getOperand(2).getReg();
13139     Register Lo = MI.getOperand(0).getReg();
13140     Register Hi = MI.getOperand(1).getReg();
13141     BuildMI(*BB, MI, DL, TII->get(TargetOpcode::COPY))
13142         .addDef(Lo)
13143         .addUse(Src, 0, PPC::sub_gp8_x1);
13144     BuildMI(*BB, MI, DL, TII->get(TargetOpcode::COPY))
13145         .addDef(Hi)
13146         .addUse(Src, 0, PPC::sub_gp8_x0);
13147   } else if (MI.getOpcode() == PPC::LQX_PSEUDO ||
13148              MI.getOpcode() == PPC::STQX_PSEUDO) {
13149     DebugLoc DL = MI.getDebugLoc();
13150     // Ptr is used as the ptr_rc_no_r0 part
13151     // of LQ/STQ's memory operand and adding result of RA and RB,
13152     // so it has to be g8rc_and_g8rc_nox0.
13153     Register Ptr =
13154         F->getRegInfo().createVirtualRegister(&PPC::G8RC_and_G8RC_NOX0RegClass);
13155     Register Val = MI.getOperand(0).getReg();
13156     Register RA = MI.getOperand(1).getReg();
13157     Register RB = MI.getOperand(2).getReg();
13158     BuildMI(*BB, MI, DL, TII->get(PPC::ADD8), Ptr).addReg(RA).addReg(RB);
13159     BuildMI(*BB, MI, DL,
13160             MI.getOpcode() == PPC::LQX_PSEUDO ? TII->get(PPC::LQ)
13161                                               : TII->get(PPC::STQ))
13162         .addReg(Val, MI.getOpcode() == PPC::LQX_PSEUDO ? RegState::Define : 0)
13163         .addImm(0)
13164         .addReg(Ptr);
13165   } else {
13166     llvm_unreachable("Unexpected instr type to insert");
13167   }
13168 
13169   MI.eraseFromParent(); // The pseudo instruction is gone now.
13170   return BB;
13171 }
13172 
13173 //===----------------------------------------------------------------------===//
13174 // Target Optimization Hooks
13175 //===----------------------------------------------------------------------===//
13176 
13177 static int getEstimateRefinementSteps(EVT VT, const PPCSubtarget &Subtarget) {
13178   // For the estimates, convergence is quadratic, so we essentially double the
13179   // number of digits correct after every iteration. For both FRE and FRSQRTE,
13180   // the minimum architected relative accuracy is 2^-5. When hasRecipPrec(),
13181   // this is 2^-14. IEEE float has 23 digits and double has 52 digits.
13182   int RefinementSteps = Subtarget.hasRecipPrec() ? 1 : 3;
13183   if (VT.getScalarType() == MVT::f64)
13184     RefinementSteps++;
13185   return RefinementSteps;
13186 }
13187 
13188 SDValue PPCTargetLowering::getSqrtInputTest(SDValue Op, SelectionDAG &DAG,
13189                                             const DenormalMode &Mode) const {
13190   // We only have VSX Vector Test for software Square Root.
13191   EVT VT = Op.getValueType();
13192   if (!isTypeLegal(MVT::i1) ||
13193       (VT != MVT::f64 &&
13194        ((VT != MVT::v2f64 && VT != MVT::v4f32) || !Subtarget.hasVSX())))
13195     return TargetLowering::getSqrtInputTest(Op, DAG, Mode);
13196 
13197   SDLoc DL(Op);
13198   // The output register of FTSQRT is CR field.
13199   SDValue FTSQRT = DAG.getNode(PPCISD::FTSQRT, DL, MVT::i32, Op);
13200   // ftsqrt BF,FRB
13201   // Let e_b be the unbiased exponent of the double-precision
13202   // floating-point operand in register FRB.
13203   // fe_flag is set to 1 if either of the following conditions occurs.
13204   //   - The double-precision floating-point operand in register FRB is a zero,
13205   //     a NaN, or an infinity, or a negative value.
13206   //   - e_b is less than or equal to -970.
13207   // Otherwise fe_flag is set to 0.
13208   // Both VSX and non-VSX versions would set EQ bit in the CR if the number is
13209   // not eligible for iteration. (zero/negative/infinity/nan or unbiased
13210   // exponent is less than -970)
13211   SDValue SRIdxVal = DAG.getTargetConstant(PPC::sub_eq, DL, MVT::i32);
13212   return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::i1,
13213                                     FTSQRT, SRIdxVal),
13214                  0);
13215 }
13216 
13217 SDValue
13218 PPCTargetLowering::getSqrtResultForDenormInput(SDValue Op,
13219                                                SelectionDAG &DAG) const {
13220   // We only have VSX Vector Square Root.
13221   EVT VT = Op.getValueType();
13222   if (VT != MVT::f64 &&
13223       ((VT != MVT::v2f64 && VT != MVT::v4f32) || !Subtarget.hasVSX()))
13224     return TargetLowering::getSqrtResultForDenormInput(Op, DAG);
13225 
13226   return DAG.getNode(PPCISD::FSQRT, SDLoc(Op), VT, Op);
13227 }
13228 
13229 SDValue PPCTargetLowering::getSqrtEstimate(SDValue Operand, SelectionDAG &DAG,
13230                                            int Enabled, int &RefinementSteps,
13231                                            bool &UseOneConstNR,
13232                                            bool Reciprocal) const {
13233   EVT VT = Operand.getValueType();
13234   if ((VT == MVT::f32 && Subtarget.hasFRSQRTES()) ||
13235       (VT == MVT::f64 && Subtarget.hasFRSQRTE()) ||
13236       (VT == MVT::v4f32 && Subtarget.hasAltivec()) ||
13237       (VT == MVT::v2f64 && Subtarget.hasVSX())) {
13238     if (RefinementSteps == ReciprocalEstimate::Unspecified)
13239       RefinementSteps = getEstimateRefinementSteps(VT, Subtarget);
13240 
13241     // The Newton-Raphson computation with a single constant does not provide
13242     // enough accuracy on some CPUs.
13243     UseOneConstNR = !Subtarget.needsTwoConstNR();
13244     return DAG.getNode(PPCISD::FRSQRTE, SDLoc(Operand), VT, Operand);
13245   }
13246   return SDValue();
13247 }
13248 
13249 SDValue PPCTargetLowering::getRecipEstimate(SDValue Operand, SelectionDAG &DAG,
13250                                             int Enabled,
13251                                             int &RefinementSteps) const {
13252   EVT VT = Operand.getValueType();
13253   if ((VT == MVT::f32 && Subtarget.hasFRES()) ||
13254       (VT == MVT::f64 && Subtarget.hasFRE()) ||
13255       (VT == MVT::v4f32 && Subtarget.hasAltivec()) ||
13256       (VT == MVT::v2f64 && Subtarget.hasVSX())) {
13257     if (RefinementSteps == ReciprocalEstimate::Unspecified)
13258       RefinementSteps = getEstimateRefinementSteps(VT, Subtarget);
13259     return DAG.getNode(PPCISD::FRE, SDLoc(Operand), VT, Operand);
13260   }
13261   return SDValue();
13262 }
13263 
13264 unsigned PPCTargetLowering::combineRepeatedFPDivisors() const {
13265   // Note: This functionality is used only when unsafe-fp-math is enabled, and
13266   // on cores with reciprocal estimates (which are used when unsafe-fp-math is
13267   // enabled for division), this functionality is redundant with the default
13268   // combiner logic (once the division -> reciprocal/multiply transformation
13269   // has taken place). As a result, this matters more for older cores than for
13270   // newer ones.
13271 
13272   // Combine multiple FDIVs with the same divisor into multiple FMULs by the
13273   // reciprocal if there are two or more FDIVs (for embedded cores with only
13274   // one FP pipeline) for three or more FDIVs (for generic OOO cores).
13275   switch (Subtarget.getCPUDirective()) {
13276   default:
13277     return 3;
13278   case PPC::DIR_440:
13279   case PPC::DIR_A2:
13280   case PPC::DIR_E500:
13281   case PPC::DIR_E500mc:
13282   case PPC::DIR_E5500:
13283     return 2;
13284   }
13285 }
13286 
13287 // isConsecutiveLSLoc needs to work even if all adds have not yet been
13288 // collapsed, and so we need to look through chains of them.
13289 static void getBaseWithConstantOffset(SDValue Loc, SDValue &Base,
13290                                      int64_t& Offset, SelectionDAG &DAG) {
13291   if (DAG.isBaseWithConstantOffset(Loc)) {
13292     Base = Loc.getOperand(0);
13293     Offset += cast<ConstantSDNode>(Loc.getOperand(1))->getSExtValue();
13294 
13295     // The base might itself be a base plus an offset, and if so, accumulate
13296     // that as well.
13297     getBaseWithConstantOffset(Loc.getOperand(0), Base, Offset, DAG);
13298   }
13299 }
13300 
13301 static bool isConsecutiveLSLoc(SDValue Loc, EVT VT, LSBaseSDNode *Base,
13302                             unsigned Bytes, int Dist,
13303                             SelectionDAG &DAG) {
13304   if (VT.getSizeInBits() / 8 != Bytes)
13305     return false;
13306 
13307   SDValue BaseLoc = Base->getBasePtr();
13308   if (Loc.getOpcode() == ISD::FrameIndex) {
13309     if (BaseLoc.getOpcode() != ISD::FrameIndex)
13310       return false;
13311     const MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
13312     int FI  = cast<FrameIndexSDNode>(Loc)->getIndex();
13313     int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
13314     int FS  = MFI.getObjectSize(FI);
13315     int BFS = MFI.getObjectSize(BFI);
13316     if (FS != BFS || FS != (int)Bytes) return false;
13317     return MFI.getObjectOffset(FI) == (MFI.getObjectOffset(BFI) + Dist*Bytes);
13318   }
13319 
13320   SDValue Base1 = Loc, Base2 = BaseLoc;
13321   int64_t Offset1 = 0, Offset2 = 0;
13322   getBaseWithConstantOffset(Loc, Base1, Offset1, DAG);
13323   getBaseWithConstantOffset(BaseLoc, Base2, Offset2, DAG);
13324   if (Base1 == Base2 && Offset1 == (Offset2 + Dist * Bytes))
13325     return true;
13326 
13327   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13328   const GlobalValue *GV1 = nullptr;
13329   const GlobalValue *GV2 = nullptr;
13330   Offset1 = 0;
13331   Offset2 = 0;
13332   bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1);
13333   bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
13334   if (isGA1 && isGA2 && GV1 == GV2)
13335     return Offset1 == (Offset2 + Dist*Bytes);
13336   return false;
13337 }
13338 
13339 // Like SelectionDAG::isConsecutiveLoad, but also works for stores, and does
13340 // not enforce equality of the chain operands.
13341 static bool isConsecutiveLS(SDNode *N, LSBaseSDNode *Base,
13342                             unsigned Bytes, int Dist,
13343                             SelectionDAG &DAG) {
13344   if (LSBaseSDNode *LS = dyn_cast<LSBaseSDNode>(N)) {
13345     EVT VT = LS->getMemoryVT();
13346     SDValue Loc = LS->getBasePtr();
13347     return isConsecutiveLSLoc(Loc, VT, Base, Bytes, Dist, DAG);
13348   }
13349 
13350   if (N->getOpcode() == ISD::INTRINSIC_W_CHAIN) {
13351     EVT VT;
13352     switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
13353     default: return false;
13354     case Intrinsic::ppc_altivec_lvx:
13355     case Intrinsic::ppc_altivec_lvxl:
13356     case Intrinsic::ppc_vsx_lxvw4x:
13357     case Intrinsic::ppc_vsx_lxvw4x_be:
13358       VT = MVT::v4i32;
13359       break;
13360     case Intrinsic::ppc_vsx_lxvd2x:
13361     case Intrinsic::ppc_vsx_lxvd2x_be:
13362       VT = MVT::v2f64;
13363       break;
13364     case Intrinsic::ppc_altivec_lvebx:
13365       VT = MVT::i8;
13366       break;
13367     case Intrinsic::ppc_altivec_lvehx:
13368       VT = MVT::i16;
13369       break;
13370     case Intrinsic::ppc_altivec_lvewx:
13371       VT = MVT::i32;
13372       break;
13373     }
13374 
13375     return isConsecutiveLSLoc(N->getOperand(2), VT, Base, Bytes, Dist, DAG);
13376   }
13377 
13378   if (N->getOpcode() == ISD::INTRINSIC_VOID) {
13379     EVT VT;
13380     switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
13381     default: return false;
13382     case Intrinsic::ppc_altivec_stvx:
13383     case Intrinsic::ppc_altivec_stvxl:
13384     case Intrinsic::ppc_vsx_stxvw4x:
13385       VT = MVT::v4i32;
13386       break;
13387     case Intrinsic::ppc_vsx_stxvd2x:
13388       VT = MVT::v2f64;
13389       break;
13390     case Intrinsic::ppc_vsx_stxvw4x_be:
13391       VT = MVT::v4i32;
13392       break;
13393     case Intrinsic::ppc_vsx_stxvd2x_be:
13394       VT = MVT::v2f64;
13395       break;
13396     case Intrinsic::ppc_altivec_stvebx:
13397       VT = MVT::i8;
13398       break;
13399     case Intrinsic::ppc_altivec_stvehx:
13400       VT = MVT::i16;
13401       break;
13402     case Intrinsic::ppc_altivec_stvewx:
13403       VT = MVT::i32;
13404       break;
13405     }
13406 
13407     return isConsecutiveLSLoc(N->getOperand(3), VT, Base, Bytes, Dist, DAG);
13408   }
13409 
13410   return false;
13411 }
13412 
13413 // Return true is there is a nearyby consecutive load to the one provided
13414 // (regardless of alignment). We search up and down the chain, looking though
13415 // token factors and other loads (but nothing else). As a result, a true result
13416 // indicates that it is safe to create a new consecutive load adjacent to the
13417 // load provided.
13418 static bool findConsecutiveLoad(LoadSDNode *LD, SelectionDAG &DAG) {
13419   SDValue Chain = LD->getChain();
13420   EVT VT = LD->getMemoryVT();
13421 
13422   SmallSet<SDNode *, 16> LoadRoots;
13423   SmallVector<SDNode *, 8> Queue(1, Chain.getNode());
13424   SmallSet<SDNode *, 16> Visited;
13425 
13426   // First, search up the chain, branching to follow all token-factor operands.
13427   // If we find a consecutive load, then we're done, otherwise, record all
13428   // nodes just above the top-level loads and token factors.
13429   while (!Queue.empty()) {
13430     SDNode *ChainNext = Queue.pop_back_val();
13431     if (!Visited.insert(ChainNext).second)
13432       continue;
13433 
13434     if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(ChainNext)) {
13435       if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
13436         return true;
13437 
13438       if (!Visited.count(ChainLD->getChain().getNode()))
13439         Queue.push_back(ChainLD->getChain().getNode());
13440     } else if (ChainNext->getOpcode() == ISD::TokenFactor) {
13441       for (const SDUse &O : ChainNext->ops())
13442         if (!Visited.count(O.getNode()))
13443           Queue.push_back(O.getNode());
13444     } else
13445       LoadRoots.insert(ChainNext);
13446   }
13447 
13448   // Second, search down the chain, starting from the top-level nodes recorded
13449   // in the first phase. These top-level nodes are the nodes just above all
13450   // loads and token factors. Starting with their uses, recursively look though
13451   // all loads (just the chain uses) and token factors to find a consecutive
13452   // load.
13453   Visited.clear();
13454   Queue.clear();
13455 
13456   for (SDNode *I : LoadRoots) {
13457     Queue.push_back(I);
13458 
13459     while (!Queue.empty()) {
13460       SDNode *LoadRoot = Queue.pop_back_val();
13461       if (!Visited.insert(LoadRoot).second)
13462         continue;
13463 
13464       if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(LoadRoot))
13465         if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
13466           return true;
13467 
13468       for (SDNode *U : LoadRoot->uses())
13469         if (((isa<MemSDNode>(U) &&
13470               cast<MemSDNode>(U)->getChain().getNode() == LoadRoot) ||
13471              U->getOpcode() == ISD::TokenFactor) &&
13472             !Visited.count(U))
13473           Queue.push_back(U);
13474     }
13475   }
13476 
13477   return false;
13478 }
13479 
13480 /// This function is called when we have proved that a SETCC node can be replaced
13481 /// by subtraction (and other supporting instructions) so that the result of
13482 /// comparison is kept in a GPR instead of CR. This function is purely for
13483 /// codegen purposes and has some flags to guide the codegen process.
13484 static SDValue generateEquivalentSub(SDNode *N, int Size, bool Complement,
13485                                      bool Swap, SDLoc &DL, SelectionDAG &DAG) {
13486   assert(N->getOpcode() == ISD::SETCC && "ISD::SETCC Expected.");
13487 
13488   // Zero extend the operands to the largest legal integer. Originally, they
13489   // must be of a strictly smaller size.
13490   auto Op0 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N->getOperand(0),
13491                          DAG.getConstant(Size, DL, MVT::i32));
13492   auto Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N->getOperand(1),
13493                          DAG.getConstant(Size, DL, MVT::i32));
13494 
13495   // Swap if needed. Depends on the condition code.
13496   if (Swap)
13497     std::swap(Op0, Op1);
13498 
13499   // Subtract extended integers.
13500   auto SubNode = DAG.getNode(ISD::SUB, DL, MVT::i64, Op0, Op1);
13501 
13502   // Move the sign bit to the least significant position and zero out the rest.
13503   // Now the least significant bit carries the result of original comparison.
13504   auto Shifted = DAG.getNode(ISD::SRL, DL, MVT::i64, SubNode,
13505                              DAG.getConstant(Size - 1, DL, MVT::i32));
13506   auto Final = Shifted;
13507 
13508   // Complement the result if needed. Based on the condition code.
13509   if (Complement)
13510     Final = DAG.getNode(ISD::XOR, DL, MVT::i64, Shifted,
13511                         DAG.getConstant(1, DL, MVT::i64));
13512 
13513   return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Final);
13514 }
13515 
13516 SDValue PPCTargetLowering::ConvertSETCCToSubtract(SDNode *N,
13517                                                   DAGCombinerInfo &DCI) const {
13518   assert(N->getOpcode() == ISD::SETCC && "ISD::SETCC Expected.");
13519 
13520   SelectionDAG &DAG = DCI.DAG;
13521   SDLoc DL(N);
13522 
13523   // Size of integers being compared has a critical role in the following
13524   // analysis, so we prefer to do this when all types are legal.
13525   if (!DCI.isAfterLegalizeDAG())
13526     return SDValue();
13527 
13528   // If all users of SETCC extend its value to a legal integer type
13529   // then we replace SETCC with a subtraction
13530   for (const SDNode *U : N->uses())
13531     if (U->getOpcode() != ISD::ZERO_EXTEND)
13532       return SDValue();
13533 
13534   ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
13535   auto OpSize = N->getOperand(0).getValueSizeInBits();
13536 
13537   unsigned Size = DAG.getDataLayout().getLargestLegalIntTypeSizeInBits();
13538 
13539   if (OpSize < Size) {
13540     switch (CC) {
13541     default: break;
13542     case ISD::SETULT:
13543       return generateEquivalentSub(N, Size, false, false, DL, DAG);
13544     case ISD::SETULE:
13545       return generateEquivalentSub(N, Size, true, true, DL, DAG);
13546     case ISD::SETUGT:
13547       return generateEquivalentSub(N, Size, false, true, DL, DAG);
13548     case ISD::SETUGE:
13549       return generateEquivalentSub(N, Size, true, false, DL, DAG);
13550     }
13551   }
13552 
13553   return SDValue();
13554 }
13555 
13556 SDValue PPCTargetLowering::DAGCombineTruncBoolExt(SDNode *N,
13557                                                   DAGCombinerInfo &DCI) const {
13558   SelectionDAG &DAG = DCI.DAG;
13559   SDLoc dl(N);
13560 
13561   assert(Subtarget.useCRBits() && "Expecting to be tracking CR bits");
13562   // If we're tracking CR bits, we need to be careful that we don't have:
13563   //   trunc(binary-ops(zext(x), zext(y)))
13564   // or
13565   //   trunc(binary-ops(binary-ops(zext(x), zext(y)), ...)
13566   // such that we're unnecessarily moving things into GPRs when it would be
13567   // better to keep them in CR bits.
13568 
13569   // Note that trunc here can be an actual i1 trunc, or can be the effective
13570   // truncation that comes from a setcc or select_cc.
13571   if (N->getOpcode() == ISD::TRUNCATE &&
13572       N->getValueType(0) != MVT::i1)
13573     return SDValue();
13574 
13575   if (N->getOperand(0).getValueType() != MVT::i32 &&
13576       N->getOperand(0).getValueType() != MVT::i64)
13577     return SDValue();
13578 
13579   if (N->getOpcode() == ISD::SETCC ||
13580       N->getOpcode() == ISD::SELECT_CC) {
13581     // If we're looking at a comparison, then we need to make sure that the
13582     // high bits (all except for the first) don't matter the result.
13583     ISD::CondCode CC =
13584       cast<CondCodeSDNode>(N->getOperand(
13585         N->getOpcode() == ISD::SETCC ? 2 : 4))->get();
13586     unsigned OpBits = N->getOperand(0).getValueSizeInBits();
13587 
13588     if (ISD::isSignedIntSetCC(CC)) {
13589       if (DAG.ComputeNumSignBits(N->getOperand(0)) != OpBits ||
13590           DAG.ComputeNumSignBits(N->getOperand(1)) != OpBits)
13591         return SDValue();
13592     } else if (ISD::isUnsignedIntSetCC(CC)) {
13593       if (!DAG.MaskedValueIsZero(N->getOperand(0),
13594                                  APInt::getHighBitsSet(OpBits, OpBits-1)) ||
13595           !DAG.MaskedValueIsZero(N->getOperand(1),
13596                                  APInt::getHighBitsSet(OpBits, OpBits-1)))
13597         return (N->getOpcode() == ISD::SETCC ? ConvertSETCCToSubtract(N, DCI)
13598                                              : SDValue());
13599     } else {
13600       // This is neither a signed nor an unsigned comparison, just make sure
13601       // that the high bits are equal.
13602       KnownBits Op1Known = DAG.computeKnownBits(N->getOperand(0));
13603       KnownBits Op2Known = DAG.computeKnownBits(N->getOperand(1));
13604 
13605       // We don't really care about what is known about the first bit (if
13606       // anything), so pretend that it is known zero for both to ensure they can
13607       // be compared as constants.
13608       Op1Known.Zero.setBit(0); Op1Known.One.clearBit(0);
13609       Op2Known.Zero.setBit(0); Op2Known.One.clearBit(0);
13610 
13611       if (!Op1Known.isConstant() || !Op2Known.isConstant() ||
13612           Op1Known.getConstant() != Op2Known.getConstant())
13613         return SDValue();
13614     }
13615   }
13616 
13617   // We now know that the higher-order bits are irrelevant, we just need to
13618   // make sure that all of the intermediate operations are bit operations, and
13619   // all inputs are extensions.
13620   if (N->getOperand(0).getOpcode() != ISD::AND &&
13621       N->getOperand(0).getOpcode() != ISD::OR  &&
13622       N->getOperand(0).getOpcode() != ISD::XOR &&
13623       N->getOperand(0).getOpcode() != ISD::SELECT &&
13624       N->getOperand(0).getOpcode() != ISD::SELECT_CC &&
13625       N->getOperand(0).getOpcode() != ISD::TRUNCATE &&
13626       N->getOperand(0).getOpcode() != ISD::SIGN_EXTEND &&
13627       N->getOperand(0).getOpcode() != ISD::ZERO_EXTEND &&
13628       N->getOperand(0).getOpcode() != ISD::ANY_EXTEND)
13629     return SDValue();
13630 
13631   if ((N->getOpcode() == ISD::SETCC || N->getOpcode() == ISD::SELECT_CC) &&
13632       N->getOperand(1).getOpcode() != ISD::AND &&
13633       N->getOperand(1).getOpcode() != ISD::OR  &&
13634       N->getOperand(1).getOpcode() != ISD::XOR &&
13635       N->getOperand(1).getOpcode() != ISD::SELECT &&
13636       N->getOperand(1).getOpcode() != ISD::SELECT_CC &&
13637       N->getOperand(1).getOpcode() != ISD::TRUNCATE &&
13638       N->getOperand(1).getOpcode() != ISD::SIGN_EXTEND &&
13639       N->getOperand(1).getOpcode() != ISD::ZERO_EXTEND &&
13640       N->getOperand(1).getOpcode() != ISD::ANY_EXTEND)
13641     return SDValue();
13642 
13643   SmallVector<SDValue, 4> Inputs;
13644   SmallVector<SDValue, 8> BinOps, PromOps;
13645   SmallPtrSet<SDNode *, 16> Visited;
13646 
13647   for (unsigned i = 0; i < 2; ++i) {
13648     if (((N->getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
13649           N->getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
13650           N->getOperand(i).getOpcode() == ISD::ANY_EXTEND) &&
13651           N->getOperand(i).getOperand(0).getValueType() == MVT::i1) ||
13652         isa<ConstantSDNode>(N->getOperand(i)))
13653       Inputs.push_back(N->getOperand(i));
13654     else
13655       BinOps.push_back(N->getOperand(i));
13656 
13657     if (N->getOpcode() == ISD::TRUNCATE)
13658       break;
13659   }
13660 
13661   // Visit all inputs, collect all binary operations (and, or, xor and
13662   // select) that are all fed by extensions.
13663   while (!BinOps.empty()) {
13664     SDValue BinOp = BinOps.pop_back_val();
13665 
13666     if (!Visited.insert(BinOp.getNode()).second)
13667       continue;
13668 
13669     PromOps.push_back(BinOp);
13670 
13671     for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) {
13672       // The condition of the select is not promoted.
13673       if (BinOp.getOpcode() == ISD::SELECT && i == 0)
13674         continue;
13675       if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3)
13676         continue;
13677 
13678       if (((BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
13679             BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
13680             BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) &&
13681            BinOp.getOperand(i).getOperand(0).getValueType() == MVT::i1) ||
13682           isa<ConstantSDNode>(BinOp.getOperand(i))) {
13683         Inputs.push_back(BinOp.getOperand(i));
13684       } else if (BinOp.getOperand(i).getOpcode() == ISD::AND ||
13685                  BinOp.getOperand(i).getOpcode() == ISD::OR  ||
13686                  BinOp.getOperand(i).getOpcode() == ISD::XOR ||
13687                  BinOp.getOperand(i).getOpcode() == ISD::SELECT ||
13688                  BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC ||
13689                  BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE ||
13690                  BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
13691                  BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
13692                  BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) {
13693         BinOps.push_back(BinOp.getOperand(i));
13694       } else {
13695         // We have an input that is not an extension or another binary
13696         // operation; we'll abort this transformation.
13697         return SDValue();
13698       }
13699     }
13700   }
13701 
13702   // Make sure that this is a self-contained cluster of operations (which
13703   // is not quite the same thing as saying that everything has only one
13704   // use).
13705   for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
13706     if (isa<ConstantSDNode>(Inputs[i]))
13707       continue;
13708 
13709     for (const SDNode *User : Inputs[i].getNode()->uses()) {
13710       if (User != N && !Visited.count(User))
13711         return SDValue();
13712 
13713       // Make sure that we're not going to promote the non-output-value
13714       // operand(s) or SELECT or SELECT_CC.
13715       // FIXME: Although we could sometimes handle this, and it does occur in
13716       // practice that one of the condition inputs to the select is also one of
13717       // the outputs, we currently can't deal with this.
13718       if (User->getOpcode() == ISD::SELECT) {
13719         if (User->getOperand(0) == Inputs[i])
13720           return SDValue();
13721       } else if (User->getOpcode() == ISD::SELECT_CC) {
13722         if (User->getOperand(0) == Inputs[i] ||
13723             User->getOperand(1) == Inputs[i])
13724           return SDValue();
13725       }
13726     }
13727   }
13728 
13729   for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) {
13730     for (const SDNode *User : PromOps[i].getNode()->uses()) {
13731       if (User != N && !Visited.count(User))
13732         return SDValue();
13733 
13734       // Make sure that we're not going to promote the non-output-value
13735       // operand(s) or SELECT or SELECT_CC.
13736       // FIXME: Although we could sometimes handle this, and it does occur in
13737       // practice that one of the condition inputs to the select is also one of
13738       // the outputs, we currently can't deal with this.
13739       if (User->getOpcode() == ISD::SELECT) {
13740         if (User->getOperand(0) == PromOps[i])
13741           return SDValue();
13742       } else if (User->getOpcode() == ISD::SELECT_CC) {
13743         if (User->getOperand(0) == PromOps[i] ||
13744             User->getOperand(1) == PromOps[i])
13745           return SDValue();
13746       }
13747     }
13748   }
13749 
13750   // Replace all inputs with the extension operand.
13751   for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
13752     // Constants may have users outside the cluster of to-be-promoted nodes,
13753     // and so we need to replace those as we do the promotions.
13754     if (isa<ConstantSDNode>(Inputs[i]))
13755       continue;
13756     else
13757       DAG.ReplaceAllUsesOfValueWith(Inputs[i], Inputs[i].getOperand(0));
13758   }
13759 
13760   std::list<HandleSDNode> PromOpHandles;
13761   for (auto &PromOp : PromOps)
13762     PromOpHandles.emplace_back(PromOp);
13763 
13764   // Replace all operations (these are all the same, but have a different
13765   // (i1) return type). DAG.getNode will validate that the types of
13766   // a binary operator match, so go through the list in reverse so that
13767   // we've likely promoted both operands first. Any intermediate truncations or
13768   // extensions disappear.
13769   while (!PromOpHandles.empty()) {
13770     SDValue PromOp = PromOpHandles.back().getValue();
13771     PromOpHandles.pop_back();
13772 
13773     if (PromOp.getOpcode() == ISD::TRUNCATE ||
13774         PromOp.getOpcode() == ISD::SIGN_EXTEND ||
13775         PromOp.getOpcode() == ISD::ZERO_EXTEND ||
13776         PromOp.getOpcode() == ISD::ANY_EXTEND) {
13777       if (!isa<ConstantSDNode>(PromOp.getOperand(0)) &&
13778           PromOp.getOperand(0).getValueType() != MVT::i1) {
13779         // The operand is not yet ready (see comment below).
13780         PromOpHandles.emplace_front(PromOp);
13781         continue;
13782       }
13783 
13784       SDValue RepValue = PromOp.getOperand(0);
13785       if (isa<ConstantSDNode>(RepValue))
13786         RepValue = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, RepValue);
13787 
13788       DAG.ReplaceAllUsesOfValueWith(PromOp, RepValue);
13789       continue;
13790     }
13791 
13792     unsigned C;
13793     switch (PromOp.getOpcode()) {
13794     default:             C = 0; break;
13795     case ISD::SELECT:    C = 1; break;
13796     case ISD::SELECT_CC: C = 2; break;
13797     }
13798 
13799     if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) &&
13800          PromOp.getOperand(C).getValueType() != MVT::i1) ||
13801         (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) &&
13802          PromOp.getOperand(C+1).getValueType() != MVT::i1)) {
13803       // The to-be-promoted operands of this node have not yet been
13804       // promoted (this should be rare because we're going through the
13805       // list backward, but if one of the operands has several users in
13806       // this cluster of to-be-promoted nodes, it is possible).
13807       PromOpHandles.emplace_front(PromOp);
13808       continue;
13809     }
13810 
13811     SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(),
13812                                 PromOp.getNode()->op_end());
13813 
13814     // If there are any constant inputs, make sure they're replaced now.
13815     for (unsigned i = 0; i < 2; ++i)
13816       if (isa<ConstantSDNode>(Ops[C+i]))
13817         Ops[C+i] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ops[C+i]);
13818 
13819     DAG.ReplaceAllUsesOfValueWith(PromOp,
13820       DAG.getNode(PromOp.getOpcode(), dl, MVT::i1, Ops));
13821   }
13822 
13823   // Now we're left with the initial truncation itself.
13824   if (N->getOpcode() == ISD::TRUNCATE)
13825     return N->getOperand(0);
13826 
13827   // Otherwise, this is a comparison. The operands to be compared have just
13828   // changed type (to i1), but everything else is the same.
13829   return SDValue(N, 0);
13830 }
13831 
13832 SDValue PPCTargetLowering::DAGCombineExtBoolTrunc(SDNode *N,
13833                                                   DAGCombinerInfo &DCI) const {
13834   SelectionDAG &DAG = DCI.DAG;
13835   SDLoc dl(N);
13836 
13837   // If we're tracking CR bits, we need to be careful that we don't have:
13838   //   zext(binary-ops(trunc(x), trunc(y)))
13839   // or
13840   //   zext(binary-ops(binary-ops(trunc(x), trunc(y)), ...)
13841   // such that we're unnecessarily moving things into CR bits that can more
13842   // efficiently stay in GPRs. Note that if we're not certain that the high
13843   // bits are set as required by the final extension, we still may need to do
13844   // some masking to get the proper behavior.
13845 
13846   // This same functionality is important on PPC64 when dealing with
13847   // 32-to-64-bit extensions; these occur often when 32-bit values are used as
13848   // the return values of functions. Because it is so similar, it is handled
13849   // here as well.
13850 
13851   if (N->getValueType(0) != MVT::i32 &&
13852       N->getValueType(0) != MVT::i64)
13853     return SDValue();
13854 
13855   if (!((N->getOperand(0).getValueType() == MVT::i1 && Subtarget.useCRBits()) ||
13856         (N->getOperand(0).getValueType() == MVT::i32 && Subtarget.isPPC64())))
13857     return SDValue();
13858 
13859   if (N->getOperand(0).getOpcode() != ISD::AND &&
13860       N->getOperand(0).getOpcode() != ISD::OR  &&
13861       N->getOperand(0).getOpcode() != ISD::XOR &&
13862       N->getOperand(0).getOpcode() != ISD::SELECT &&
13863       N->getOperand(0).getOpcode() != ISD::SELECT_CC)
13864     return SDValue();
13865 
13866   SmallVector<SDValue, 4> Inputs;
13867   SmallVector<SDValue, 8> BinOps(1, N->getOperand(0)), PromOps;
13868   SmallPtrSet<SDNode *, 16> Visited;
13869 
13870   // Visit all inputs, collect all binary operations (and, or, xor and
13871   // select) that are all fed by truncations.
13872   while (!BinOps.empty()) {
13873     SDValue BinOp = BinOps.pop_back_val();
13874 
13875     if (!Visited.insert(BinOp.getNode()).second)
13876       continue;
13877 
13878     PromOps.push_back(BinOp);
13879 
13880     for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) {
13881       // The condition of the select is not promoted.
13882       if (BinOp.getOpcode() == ISD::SELECT && i == 0)
13883         continue;
13884       if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3)
13885         continue;
13886 
13887       if (BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE ||
13888           isa<ConstantSDNode>(BinOp.getOperand(i))) {
13889         Inputs.push_back(BinOp.getOperand(i));
13890       } else if (BinOp.getOperand(i).getOpcode() == ISD::AND ||
13891                  BinOp.getOperand(i).getOpcode() == ISD::OR  ||
13892                  BinOp.getOperand(i).getOpcode() == ISD::XOR ||
13893                  BinOp.getOperand(i).getOpcode() == ISD::SELECT ||
13894                  BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC) {
13895         BinOps.push_back(BinOp.getOperand(i));
13896       } else {
13897         // We have an input that is not a truncation or another binary
13898         // operation; we'll abort this transformation.
13899         return SDValue();
13900       }
13901     }
13902   }
13903 
13904   // The operands of a select that must be truncated when the select is
13905   // promoted because the operand is actually part of the to-be-promoted set.
13906   DenseMap<SDNode *, EVT> SelectTruncOp[2];
13907 
13908   // Make sure that this is a self-contained cluster of operations (which
13909   // is not quite the same thing as saying that everything has only one
13910   // use).
13911   for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
13912     if (isa<ConstantSDNode>(Inputs[i]))
13913       continue;
13914 
13915     for (SDNode *User : Inputs[i].getNode()->uses()) {
13916       if (User != N && !Visited.count(User))
13917         return SDValue();
13918 
13919       // If we're going to promote the non-output-value operand(s) or SELECT or
13920       // SELECT_CC, record them for truncation.
13921       if (User->getOpcode() == ISD::SELECT) {
13922         if (User->getOperand(0) == Inputs[i])
13923           SelectTruncOp[0].insert(std::make_pair(User,
13924                                     User->getOperand(0).getValueType()));
13925       } else if (User->getOpcode() == ISD::SELECT_CC) {
13926         if (User->getOperand(0) == Inputs[i])
13927           SelectTruncOp[0].insert(std::make_pair(User,
13928                                     User->getOperand(0).getValueType()));
13929         if (User->getOperand(1) == Inputs[i])
13930           SelectTruncOp[1].insert(std::make_pair(User,
13931                                     User->getOperand(1).getValueType()));
13932       }
13933     }
13934   }
13935 
13936   for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) {
13937     for (SDNode *User : PromOps[i].getNode()->uses()) {
13938       if (User != N && !Visited.count(User))
13939         return SDValue();
13940 
13941       // If we're going to promote the non-output-value operand(s) or SELECT or
13942       // SELECT_CC, record them for truncation.
13943       if (User->getOpcode() == ISD::SELECT) {
13944         if (User->getOperand(0) == PromOps[i])
13945           SelectTruncOp[0].insert(std::make_pair(User,
13946                                     User->getOperand(0).getValueType()));
13947       } else if (User->getOpcode() == ISD::SELECT_CC) {
13948         if (User->getOperand(0) == PromOps[i])
13949           SelectTruncOp[0].insert(std::make_pair(User,
13950                                     User->getOperand(0).getValueType()));
13951         if (User->getOperand(1) == PromOps[i])
13952           SelectTruncOp[1].insert(std::make_pair(User,
13953                                     User->getOperand(1).getValueType()));
13954       }
13955     }
13956   }
13957 
13958   unsigned PromBits = N->getOperand(0).getValueSizeInBits();
13959   bool ReallyNeedsExt = false;
13960   if (N->getOpcode() != ISD::ANY_EXTEND) {
13961     // If all of the inputs are not already sign/zero extended, then
13962     // we'll still need to do that at the end.
13963     for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
13964       if (isa<ConstantSDNode>(Inputs[i]))
13965         continue;
13966 
13967       unsigned OpBits =
13968         Inputs[i].getOperand(0).getValueSizeInBits();
13969       assert(PromBits < OpBits && "Truncation not to a smaller bit count?");
13970 
13971       if ((N->getOpcode() == ISD::ZERO_EXTEND &&
13972            !DAG.MaskedValueIsZero(Inputs[i].getOperand(0),
13973                                   APInt::getHighBitsSet(OpBits,
13974                                                         OpBits-PromBits))) ||
13975           (N->getOpcode() == ISD::SIGN_EXTEND &&
13976            DAG.ComputeNumSignBits(Inputs[i].getOperand(0)) <
13977              (OpBits-(PromBits-1)))) {
13978         ReallyNeedsExt = true;
13979         break;
13980       }
13981     }
13982   }
13983 
13984   // Replace all inputs, either with the truncation operand, or a
13985   // truncation or extension to the final output type.
13986   for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
13987     // Constant inputs need to be replaced with the to-be-promoted nodes that
13988     // use them because they might have users outside of the cluster of
13989     // promoted nodes.
13990     if (isa<ConstantSDNode>(Inputs[i]))
13991       continue;
13992 
13993     SDValue InSrc = Inputs[i].getOperand(0);
13994     if (Inputs[i].getValueType() == N->getValueType(0))
13995       DAG.ReplaceAllUsesOfValueWith(Inputs[i], InSrc);
13996     else if (N->getOpcode() == ISD::SIGN_EXTEND)
13997       DAG.ReplaceAllUsesOfValueWith(Inputs[i],
13998         DAG.getSExtOrTrunc(InSrc, dl, N->getValueType(0)));
13999     else if (N->getOpcode() == ISD::ZERO_EXTEND)
14000       DAG.ReplaceAllUsesOfValueWith(Inputs[i],
14001         DAG.getZExtOrTrunc(InSrc, dl, N->getValueType(0)));
14002     else
14003       DAG.ReplaceAllUsesOfValueWith(Inputs[i],
14004         DAG.getAnyExtOrTrunc(InSrc, dl, N->getValueType(0)));
14005   }
14006 
14007   std::list<HandleSDNode> PromOpHandles;
14008   for (auto &PromOp : PromOps)
14009     PromOpHandles.emplace_back(PromOp);
14010 
14011   // Replace all operations (these are all the same, but have a different
14012   // (promoted) return type). DAG.getNode will validate that the types of
14013   // a binary operator match, so go through the list in reverse so that
14014   // we've likely promoted both operands first.
14015   while (!PromOpHandles.empty()) {
14016     SDValue PromOp = PromOpHandles.back().getValue();
14017     PromOpHandles.pop_back();
14018 
14019     unsigned C;
14020     switch (PromOp.getOpcode()) {
14021     default:             C = 0; break;
14022     case ISD::SELECT:    C = 1; break;
14023     case ISD::SELECT_CC: C = 2; break;
14024     }
14025 
14026     if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) &&
14027          PromOp.getOperand(C).getValueType() != N->getValueType(0)) ||
14028         (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) &&
14029          PromOp.getOperand(C+1).getValueType() != N->getValueType(0))) {
14030       // The to-be-promoted operands of this node have not yet been
14031       // promoted (this should be rare because we're going through the
14032       // list backward, but if one of the operands has several users in
14033       // this cluster of to-be-promoted nodes, it is possible).
14034       PromOpHandles.emplace_front(PromOp);
14035       continue;
14036     }
14037 
14038     // For SELECT and SELECT_CC nodes, we do a similar check for any
14039     // to-be-promoted comparison inputs.
14040     if (PromOp.getOpcode() == ISD::SELECT ||
14041         PromOp.getOpcode() == ISD::SELECT_CC) {
14042       if ((SelectTruncOp[0].count(PromOp.getNode()) &&
14043            PromOp.getOperand(0).getValueType() != N->getValueType(0)) ||
14044           (SelectTruncOp[1].count(PromOp.getNode()) &&
14045            PromOp.getOperand(1).getValueType() != N->getValueType(0))) {
14046         PromOpHandles.emplace_front(PromOp);
14047         continue;
14048       }
14049     }
14050 
14051     SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(),
14052                                 PromOp.getNode()->op_end());
14053 
14054     // If this node has constant inputs, then they'll need to be promoted here.
14055     for (unsigned i = 0; i < 2; ++i) {
14056       if (!isa<ConstantSDNode>(Ops[C+i]))
14057         continue;
14058       if (Ops[C+i].getValueType() == N->getValueType(0))
14059         continue;
14060 
14061       if (N->getOpcode() == ISD::SIGN_EXTEND)
14062         Ops[C+i] = DAG.getSExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
14063       else if (N->getOpcode() == ISD::ZERO_EXTEND)
14064         Ops[C+i] = DAG.getZExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
14065       else
14066         Ops[C+i] = DAG.getAnyExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
14067     }
14068 
14069     // If we've promoted the comparison inputs of a SELECT or SELECT_CC,
14070     // truncate them again to the original value type.
14071     if (PromOp.getOpcode() == ISD::SELECT ||
14072         PromOp.getOpcode() == ISD::SELECT_CC) {
14073       auto SI0 = SelectTruncOp[0].find(PromOp.getNode());
14074       if (SI0 != SelectTruncOp[0].end())
14075         Ops[0] = DAG.getNode(ISD::TRUNCATE, dl, SI0->second, Ops[0]);
14076       auto SI1 = SelectTruncOp[1].find(PromOp.getNode());
14077       if (SI1 != SelectTruncOp[1].end())
14078         Ops[1] = DAG.getNode(ISD::TRUNCATE, dl, SI1->second, Ops[1]);
14079     }
14080 
14081     DAG.ReplaceAllUsesOfValueWith(PromOp,
14082       DAG.getNode(PromOp.getOpcode(), dl, N->getValueType(0), Ops));
14083   }
14084 
14085   // Now we're left with the initial extension itself.
14086   if (!ReallyNeedsExt)
14087     return N->getOperand(0);
14088 
14089   // To zero extend, just mask off everything except for the first bit (in the
14090   // i1 case).
14091   if (N->getOpcode() == ISD::ZERO_EXTEND)
14092     return DAG.getNode(ISD::AND, dl, N->getValueType(0), N->getOperand(0),
14093                        DAG.getConstant(APInt::getLowBitsSet(
14094                                          N->getValueSizeInBits(0), PromBits),
14095                                        dl, N->getValueType(0)));
14096 
14097   assert(N->getOpcode() == ISD::SIGN_EXTEND &&
14098          "Invalid extension type");
14099   EVT ShiftAmountTy = getShiftAmountTy(N->getValueType(0), DAG.getDataLayout());
14100   SDValue ShiftCst =
14101       DAG.getConstant(N->getValueSizeInBits(0) - PromBits, dl, ShiftAmountTy);
14102   return DAG.getNode(
14103       ISD::SRA, dl, N->getValueType(0),
14104       DAG.getNode(ISD::SHL, dl, N->getValueType(0), N->getOperand(0), ShiftCst),
14105       ShiftCst);
14106 }
14107 
14108 SDValue PPCTargetLowering::combineSetCC(SDNode *N,
14109                                         DAGCombinerInfo &DCI) const {
14110   assert(N->getOpcode() == ISD::SETCC &&
14111          "Should be called with a SETCC node");
14112 
14113   ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
14114   if (CC == ISD::SETNE || CC == ISD::SETEQ) {
14115     SDValue LHS = N->getOperand(0);
14116     SDValue RHS = N->getOperand(1);
14117 
14118     // If there is a '0 - y' pattern, canonicalize the pattern to the RHS.
14119     if (LHS.getOpcode() == ISD::SUB && isNullConstant(LHS.getOperand(0)) &&
14120         LHS.hasOneUse())
14121       std::swap(LHS, RHS);
14122 
14123     // x == 0-y --> x+y == 0
14124     // x != 0-y --> x+y != 0
14125     if (RHS.getOpcode() == ISD::SUB && isNullConstant(RHS.getOperand(0)) &&
14126         RHS.hasOneUse()) {
14127       SDLoc DL(N);
14128       SelectionDAG &DAG = DCI.DAG;
14129       EVT VT = N->getValueType(0);
14130       EVT OpVT = LHS.getValueType();
14131       SDValue Add = DAG.getNode(ISD::ADD, DL, OpVT, LHS, RHS.getOperand(1));
14132       return DAG.getSetCC(DL, VT, Add, DAG.getConstant(0, DL, OpVT), CC);
14133     }
14134   }
14135 
14136   return DAGCombineTruncBoolExt(N, DCI);
14137 }
14138 
14139 // Is this an extending load from an f32 to an f64?
14140 static bool isFPExtLoad(SDValue Op) {
14141   if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op.getNode()))
14142     return LD->getExtensionType() == ISD::EXTLOAD &&
14143       Op.getValueType() == MVT::f64;
14144   return false;
14145 }
14146 
14147 /// Reduces the number of fp-to-int conversion when building a vector.
14148 ///
14149 /// If this vector is built out of floating to integer conversions,
14150 /// transform it to a vector built out of floating point values followed by a
14151 /// single floating to integer conversion of the vector.
14152 /// Namely  (build_vector (fptosi $A), (fptosi $B), ...)
14153 /// becomes (fptosi (build_vector ($A, $B, ...)))
14154 SDValue PPCTargetLowering::
14155 combineElementTruncationToVectorTruncation(SDNode *N,
14156                                            DAGCombinerInfo &DCI) const {
14157   assert(N->getOpcode() == ISD::BUILD_VECTOR &&
14158          "Should be called with a BUILD_VECTOR node");
14159 
14160   SelectionDAG &DAG = DCI.DAG;
14161   SDLoc dl(N);
14162 
14163   SDValue FirstInput = N->getOperand(0);
14164   assert(FirstInput.getOpcode() == PPCISD::MFVSR &&
14165          "The input operand must be an fp-to-int conversion.");
14166 
14167   // This combine happens after legalization so the fp_to_[su]i nodes are
14168   // already converted to PPCSISD nodes.
14169   unsigned FirstConversion = FirstInput.getOperand(0).getOpcode();
14170   if (FirstConversion == PPCISD::FCTIDZ ||
14171       FirstConversion == PPCISD::FCTIDUZ ||
14172       FirstConversion == PPCISD::FCTIWZ ||
14173       FirstConversion == PPCISD::FCTIWUZ) {
14174     bool IsSplat = true;
14175     bool Is32Bit = FirstConversion == PPCISD::FCTIWZ ||
14176       FirstConversion == PPCISD::FCTIWUZ;
14177     EVT SrcVT = FirstInput.getOperand(0).getValueType();
14178     SmallVector<SDValue, 4> Ops;
14179     EVT TargetVT = N->getValueType(0);
14180     for (int i = 0, e = N->getNumOperands(); i < e; ++i) {
14181       SDValue NextOp = N->getOperand(i);
14182       if (NextOp.getOpcode() != PPCISD::MFVSR)
14183         return SDValue();
14184       unsigned NextConversion = NextOp.getOperand(0).getOpcode();
14185       if (NextConversion != FirstConversion)
14186         return SDValue();
14187       // If we are converting to 32-bit integers, we need to add an FP_ROUND.
14188       // This is not valid if the input was originally double precision. It is
14189       // also not profitable to do unless this is an extending load in which
14190       // case doing this combine will allow us to combine consecutive loads.
14191       if (Is32Bit && !isFPExtLoad(NextOp.getOperand(0).getOperand(0)))
14192         return SDValue();
14193       if (N->getOperand(i) != FirstInput)
14194         IsSplat = false;
14195     }
14196 
14197     // If this is a splat, we leave it as-is since there will be only a single
14198     // fp-to-int conversion followed by a splat of the integer. This is better
14199     // for 32-bit and smaller ints and neutral for 64-bit ints.
14200     if (IsSplat)
14201       return SDValue();
14202 
14203     // Now that we know we have the right type of node, get its operands
14204     for (int i = 0, e = N->getNumOperands(); i < e; ++i) {
14205       SDValue In = N->getOperand(i).getOperand(0);
14206       if (Is32Bit) {
14207         // For 32-bit values, we need to add an FP_ROUND node (if we made it
14208         // here, we know that all inputs are extending loads so this is safe).
14209         if (In.isUndef())
14210           Ops.push_back(DAG.getUNDEF(SrcVT));
14211         else {
14212           SDValue Trunc =
14213               DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, In.getOperand(0),
14214                           DAG.getIntPtrConstant(1, dl, /*isTarget=*/true));
14215           Ops.push_back(Trunc);
14216         }
14217       } else
14218         Ops.push_back(In.isUndef() ? DAG.getUNDEF(SrcVT) : In.getOperand(0));
14219     }
14220 
14221     unsigned Opcode;
14222     if (FirstConversion == PPCISD::FCTIDZ ||
14223         FirstConversion == PPCISD::FCTIWZ)
14224       Opcode = ISD::FP_TO_SINT;
14225     else
14226       Opcode = ISD::FP_TO_UINT;
14227 
14228     EVT NewVT = TargetVT == MVT::v2i64 ? MVT::v2f64 : MVT::v4f32;
14229     SDValue BV = DAG.getBuildVector(NewVT, dl, Ops);
14230     return DAG.getNode(Opcode, dl, TargetVT, BV);
14231   }
14232   return SDValue();
14233 }
14234 
14235 /// Reduce the number of loads when building a vector.
14236 ///
14237 /// Building a vector out of multiple loads can be converted to a load
14238 /// of the vector type if the loads are consecutive. If the loads are
14239 /// consecutive but in descending order, a shuffle is added at the end
14240 /// to reorder the vector.
14241 static SDValue combineBVOfConsecutiveLoads(SDNode *N, SelectionDAG &DAG) {
14242   assert(N->getOpcode() == ISD::BUILD_VECTOR &&
14243          "Should be called with a BUILD_VECTOR node");
14244 
14245   SDLoc dl(N);
14246 
14247   // Return early for non byte-sized type, as they can't be consecutive.
14248   if (!N->getValueType(0).getVectorElementType().isByteSized())
14249     return SDValue();
14250 
14251   bool InputsAreConsecutiveLoads = true;
14252   bool InputsAreReverseConsecutive = true;
14253   unsigned ElemSize = N->getValueType(0).getScalarType().getStoreSize();
14254   SDValue FirstInput = N->getOperand(0);
14255   bool IsRoundOfExtLoad = false;
14256   LoadSDNode *FirstLoad = nullptr;
14257 
14258   if (FirstInput.getOpcode() == ISD::FP_ROUND &&
14259       FirstInput.getOperand(0).getOpcode() == ISD::LOAD) {
14260     FirstLoad = cast<LoadSDNode>(FirstInput.getOperand(0));
14261     IsRoundOfExtLoad = FirstLoad->getExtensionType() == ISD::EXTLOAD;
14262   }
14263   // Not a build vector of (possibly fp_rounded) loads.
14264   if ((!IsRoundOfExtLoad && FirstInput.getOpcode() != ISD::LOAD) ||
14265       N->getNumOperands() == 1)
14266     return SDValue();
14267 
14268   if (!IsRoundOfExtLoad)
14269     FirstLoad = cast<LoadSDNode>(FirstInput);
14270 
14271   SmallVector<LoadSDNode *, 4> InputLoads;
14272   InputLoads.push_back(FirstLoad);
14273   for (int i = 1, e = N->getNumOperands(); i < e; ++i) {
14274     // If any inputs are fp_round(extload), they all must be.
14275     if (IsRoundOfExtLoad && N->getOperand(i).getOpcode() != ISD::FP_ROUND)
14276       return SDValue();
14277 
14278     SDValue NextInput = IsRoundOfExtLoad ? N->getOperand(i).getOperand(0) :
14279       N->getOperand(i);
14280     if (NextInput.getOpcode() != ISD::LOAD)
14281       return SDValue();
14282 
14283     SDValue PreviousInput =
14284       IsRoundOfExtLoad ? N->getOperand(i-1).getOperand(0) : N->getOperand(i-1);
14285     LoadSDNode *LD1 = cast<LoadSDNode>(PreviousInput);
14286     LoadSDNode *LD2 = cast<LoadSDNode>(NextInput);
14287 
14288     // If any inputs are fp_round(extload), they all must be.
14289     if (IsRoundOfExtLoad && LD2->getExtensionType() != ISD::EXTLOAD)
14290       return SDValue();
14291 
14292     // We only care about regular loads. The PPC-specific load intrinsics
14293     // will not lead to a merge opportunity.
14294     if (!DAG.areNonVolatileConsecutiveLoads(LD2, LD1, ElemSize, 1))
14295       InputsAreConsecutiveLoads = false;
14296     if (!DAG.areNonVolatileConsecutiveLoads(LD1, LD2, ElemSize, 1))
14297       InputsAreReverseConsecutive = false;
14298 
14299     // Exit early if the loads are neither consecutive nor reverse consecutive.
14300     if (!InputsAreConsecutiveLoads && !InputsAreReverseConsecutive)
14301       return SDValue();
14302     InputLoads.push_back(LD2);
14303   }
14304 
14305   assert(!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive) &&
14306          "The loads cannot be both consecutive and reverse consecutive.");
14307 
14308   SDValue WideLoad;
14309   SDValue ReturnSDVal;
14310   if (InputsAreConsecutiveLoads) {
14311     assert(FirstLoad && "Input needs to be a LoadSDNode.");
14312     WideLoad = DAG.getLoad(N->getValueType(0), dl, FirstLoad->getChain(),
14313                            FirstLoad->getBasePtr(), FirstLoad->getPointerInfo(),
14314                            FirstLoad->getAlign());
14315     ReturnSDVal = WideLoad;
14316   } else if (InputsAreReverseConsecutive) {
14317     LoadSDNode *LastLoad = InputLoads.back();
14318     assert(LastLoad && "Input needs to be a LoadSDNode.");
14319     WideLoad = DAG.getLoad(N->getValueType(0), dl, LastLoad->getChain(),
14320                            LastLoad->getBasePtr(), LastLoad->getPointerInfo(),
14321                            LastLoad->getAlign());
14322     SmallVector<int, 16> Ops;
14323     for (int i = N->getNumOperands() - 1; i >= 0; i--)
14324       Ops.push_back(i);
14325 
14326     ReturnSDVal = DAG.getVectorShuffle(N->getValueType(0), dl, WideLoad,
14327                                        DAG.getUNDEF(N->getValueType(0)), Ops);
14328   } else
14329     return SDValue();
14330 
14331   for (auto *LD : InputLoads)
14332     DAG.makeEquivalentMemoryOrdering(LD, WideLoad);
14333   return ReturnSDVal;
14334 }
14335 
14336 // This function adds the required vector_shuffle needed to get
14337 // the elements of the vector extract in the correct position
14338 // as specified by the CorrectElems encoding.
14339 static SDValue addShuffleForVecExtend(SDNode *N, SelectionDAG &DAG,
14340                                       SDValue Input, uint64_t Elems,
14341                                       uint64_t CorrectElems) {
14342   SDLoc dl(N);
14343 
14344   unsigned NumElems = Input.getValueType().getVectorNumElements();
14345   SmallVector<int, 16> ShuffleMask(NumElems, -1);
14346 
14347   // Knowing the element indices being extracted from the original
14348   // vector and the order in which they're being inserted, just put
14349   // them at element indices required for the instruction.
14350   for (unsigned i = 0; i < N->getNumOperands(); i++) {
14351     if (DAG.getDataLayout().isLittleEndian())
14352       ShuffleMask[CorrectElems & 0xF] = Elems & 0xF;
14353     else
14354       ShuffleMask[(CorrectElems & 0xF0) >> 4] = (Elems & 0xF0) >> 4;
14355     CorrectElems = CorrectElems >> 8;
14356     Elems = Elems >> 8;
14357   }
14358 
14359   SDValue Shuffle =
14360       DAG.getVectorShuffle(Input.getValueType(), dl, Input,
14361                            DAG.getUNDEF(Input.getValueType()), ShuffleMask);
14362 
14363   EVT VT = N->getValueType(0);
14364   SDValue Conv = DAG.getBitcast(VT, Shuffle);
14365 
14366   EVT ExtVT = EVT::getVectorVT(*DAG.getContext(),
14367                                Input.getValueType().getVectorElementType(),
14368                                VT.getVectorNumElements());
14369   return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, VT, Conv,
14370                      DAG.getValueType(ExtVT));
14371 }
14372 
14373 // Look for build vector patterns where input operands come from sign
14374 // extended vector_extract elements of specific indices. If the correct indices
14375 // aren't used, add a vector shuffle to fix up the indices and create
14376 // SIGN_EXTEND_INREG node which selects the vector sign extend instructions
14377 // during instruction selection.
14378 static SDValue combineBVOfVecSExt(SDNode *N, SelectionDAG &DAG) {
14379   // This array encodes the indices that the vector sign extend instructions
14380   // extract from when extending from one type to another for both BE and LE.
14381   // The right nibble of each byte corresponds to the LE incides.
14382   // and the left nibble of each byte corresponds to the BE incides.
14383   // For example: 0x3074B8FC  byte->word
14384   // For LE: the allowed indices are: 0x0,0x4,0x8,0xC
14385   // For BE: the allowed indices are: 0x3,0x7,0xB,0xF
14386   // For example: 0x000070F8  byte->double word
14387   // For LE: the allowed indices are: 0x0,0x8
14388   // For BE: the allowed indices are: 0x7,0xF
14389   uint64_t TargetElems[] = {
14390       0x3074B8FC, // b->w
14391       0x000070F8, // b->d
14392       0x10325476, // h->w
14393       0x00003074, // h->d
14394       0x00001032, // w->d
14395   };
14396 
14397   uint64_t Elems = 0;
14398   int Index;
14399   SDValue Input;
14400 
14401   auto isSExtOfVecExtract = [&](SDValue Op) -> bool {
14402     if (!Op)
14403       return false;
14404     if (Op.getOpcode() != ISD::SIGN_EXTEND &&
14405         Op.getOpcode() != ISD::SIGN_EXTEND_INREG)
14406       return false;
14407 
14408     // A SIGN_EXTEND_INREG might be fed by an ANY_EXTEND to produce a value
14409     // of the right width.
14410     SDValue Extract = Op.getOperand(0);
14411     if (Extract.getOpcode() == ISD::ANY_EXTEND)
14412       Extract = Extract.getOperand(0);
14413     if (Extract.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
14414       return false;
14415 
14416     ConstantSDNode *ExtOp = dyn_cast<ConstantSDNode>(Extract.getOperand(1));
14417     if (!ExtOp)
14418       return false;
14419 
14420     Index = ExtOp->getZExtValue();
14421     if (Input && Input != Extract.getOperand(0))
14422       return false;
14423 
14424     if (!Input)
14425       Input = Extract.getOperand(0);
14426 
14427     Elems = Elems << 8;
14428     Index = DAG.getDataLayout().isLittleEndian() ? Index : Index << 4;
14429     Elems |= Index;
14430 
14431     return true;
14432   };
14433 
14434   // If the build vector operands aren't sign extended vector extracts,
14435   // of the same input vector, then return.
14436   for (unsigned i = 0; i < N->getNumOperands(); i++) {
14437     if (!isSExtOfVecExtract(N->getOperand(i))) {
14438       return SDValue();
14439     }
14440   }
14441 
14442   // If the vector extract indicies are not correct, add the appropriate
14443   // vector_shuffle.
14444   int TgtElemArrayIdx;
14445   int InputSize = Input.getValueType().getScalarSizeInBits();
14446   int OutputSize = N->getValueType(0).getScalarSizeInBits();
14447   if (InputSize + OutputSize == 40)
14448     TgtElemArrayIdx = 0;
14449   else if (InputSize + OutputSize == 72)
14450     TgtElemArrayIdx = 1;
14451   else if (InputSize + OutputSize == 48)
14452     TgtElemArrayIdx = 2;
14453   else if (InputSize + OutputSize == 80)
14454     TgtElemArrayIdx = 3;
14455   else if (InputSize + OutputSize == 96)
14456     TgtElemArrayIdx = 4;
14457   else
14458     return SDValue();
14459 
14460   uint64_t CorrectElems = TargetElems[TgtElemArrayIdx];
14461   CorrectElems = DAG.getDataLayout().isLittleEndian()
14462                      ? CorrectElems & 0x0F0F0F0F0F0F0F0F
14463                      : CorrectElems & 0xF0F0F0F0F0F0F0F0;
14464   if (Elems != CorrectElems) {
14465     return addShuffleForVecExtend(N, DAG, Input, Elems, CorrectElems);
14466   }
14467 
14468   // Regular lowering will catch cases where a shuffle is not needed.
14469   return SDValue();
14470 }
14471 
14472 // Look for the pattern of a load from a narrow width to i128, feeding
14473 // into a BUILD_VECTOR of v1i128. Replace this sequence with a PPCISD node
14474 // (LXVRZX). This node represents a zero extending load that will be matched
14475 // to the Load VSX Vector Rightmost instructions.
14476 static SDValue combineBVZEXTLOAD(SDNode *N, SelectionDAG &DAG) {
14477   SDLoc DL(N);
14478 
14479   // This combine is only eligible for a BUILD_VECTOR of v1i128.
14480   if (N->getValueType(0) != MVT::v1i128)
14481     return SDValue();
14482 
14483   SDValue Operand = N->getOperand(0);
14484   // Proceed with the transformation if the operand to the BUILD_VECTOR
14485   // is a load instruction.
14486   if (Operand.getOpcode() != ISD::LOAD)
14487     return SDValue();
14488 
14489   auto *LD = cast<LoadSDNode>(Operand);
14490   EVT MemoryType = LD->getMemoryVT();
14491 
14492   // This transformation is only valid if the we are loading either a byte,
14493   // halfword, word, or doubleword.
14494   bool ValidLDType = MemoryType == MVT::i8 || MemoryType == MVT::i16 ||
14495                      MemoryType == MVT::i32 || MemoryType == MVT::i64;
14496 
14497   // Ensure that the load from the narrow width is being zero extended to i128.
14498   if (!ValidLDType ||
14499       (LD->getExtensionType() != ISD::ZEXTLOAD &&
14500        LD->getExtensionType() != ISD::EXTLOAD))
14501     return SDValue();
14502 
14503   SDValue LoadOps[] = {
14504       LD->getChain(), LD->getBasePtr(),
14505       DAG.getIntPtrConstant(MemoryType.getScalarSizeInBits(), DL)};
14506 
14507   return DAG.getMemIntrinsicNode(PPCISD::LXVRZX, DL,
14508                                  DAG.getVTList(MVT::v1i128, MVT::Other),
14509                                  LoadOps, MemoryType, LD->getMemOperand());
14510 }
14511 
14512 SDValue PPCTargetLowering::DAGCombineBuildVector(SDNode *N,
14513                                                  DAGCombinerInfo &DCI) const {
14514   assert(N->getOpcode() == ISD::BUILD_VECTOR &&
14515          "Should be called with a BUILD_VECTOR node");
14516 
14517   SelectionDAG &DAG = DCI.DAG;
14518   SDLoc dl(N);
14519 
14520   if (!Subtarget.hasVSX())
14521     return SDValue();
14522 
14523   // The target independent DAG combiner will leave a build_vector of
14524   // float-to-int conversions intact. We can generate MUCH better code for
14525   // a float-to-int conversion of a vector of floats.
14526   SDValue FirstInput = N->getOperand(0);
14527   if (FirstInput.getOpcode() == PPCISD::MFVSR) {
14528     SDValue Reduced = combineElementTruncationToVectorTruncation(N, DCI);
14529     if (Reduced)
14530       return Reduced;
14531   }
14532 
14533   // If we're building a vector out of consecutive loads, just load that
14534   // vector type.
14535   SDValue Reduced = combineBVOfConsecutiveLoads(N, DAG);
14536   if (Reduced)
14537     return Reduced;
14538 
14539   // If we're building a vector out of extended elements from another vector
14540   // we have P9 vector integer extend instructions. The code assumes legal
14541   // input types (i.e. it can't handle things like v4i16) so do not run before
14542   // legalization.
14543   if (Subtarget.hasP9Altivec() && !DCI.isBeforeLegalize()) {
14544     Reduced = combineBVOfVecSExt(N, DAG);
14545     if (Reduced)
14546       return Reduced;
14547   }
14548 
14549   // On Power10, the Load VSX Vector Rightmost instructions can be utilized
14550   // if this is a BUILD_VECTOR of v1i128, and if the operand to the BUILD_VECTOR
14551   // is a load from <valid narrow width> to i128.
14552   if (Subtarget.isISA3_1()) {
14553     SDValue BVOfZLoad = combineBVZEXTLOAD(N, DAG);
14554     if (BVOfZLoad)
14555       return BVOfZLoad;
14556   }
14557 
14558   if (N->getValueType(0) != MVT::v2f64)
14559     return SDValue();
14560 
14561   // Looking for:
14562   // (build_vector ([su]int_to_fp (extractelt 0)), [su]int_to_fp (extractelt 1))
14563   if (FirstInput.getOpcode() != ISD::SINT_TO_FP &&
14564       FirstInput.getOpcode() != ISD::UINT_TO_FP)
14565     return SDValue();
14566   if (N->getOperand(1).getOpcode() != ISD::SINT_TO_FP &&
14567       N->getOperand(1).getOpcode() != ISD::UINT_TO_FP)
14568     return SDValue();
14569   if (FirstInput.getOpcode() != N->getOperand(1).getOpcode())
14570     return SDValue();
14571 
14572   SDValue Ext1 = FirstInput.getOperand(0);
14573   SDValue Ext2 = N->getOperand(1).getOperand(0);
14574   if(Ext1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
14575      Ext2.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
14576     return SDValue();
14577 
14578   ConstantSDNode *Ext1Op = dyn_cast<ConstantSDNode>(Ext1.getOperand(1));
14579   ConstantSDNode *Ext2Op = dyn_cast<ConstantSDNode>(Ext2.getOperand(1));
14580   if (!Ext1Op || !Ext2Op)
14581     return SDValue();
14582   if (Ext1.getOperand(0).getValueType() != MVT::v4i32 ||
14583       Ext1.getOperand(0) != Ext2.getOperand(0))
14584     return SDValue();
14585 
14586   int FirstElem = Ext1Op->getZExtValue();
14587   int SecondElem = Ext2Op->getZExtValue();
14588   int SubvecIdx;
14589   if (FirstElem == 0 && SecondElem == 1)
14590     SubvecIdx = Subtarget.isLittleEndian() ? 1 : 0;
14591   else if (FirstElem == 2 && SecondElem == 3)
14592     SubvecIdx = Subtarget.isLittleEndian() ? 0 : 1;
14593   else
14594     return SDValue();
14595 
14596   SDValue SrcVec = Ext1.getOperand(0);
14597   auto NodeType = (N->getOperand(1).getOpcode() == ISD::SINT_TO_FP) ?
14598     PPCISD::SINT_VEC_TO_FP : PPCISD::UINT_VEC_TO_FP;
14599   return DAG.getNode(NodeType, dl, MVT::v2f64,
14600                      SrcVec, DAG.getIntPtrConstant(SubvecIdx, dl));
14601 }
14602 
14603 SDValue PPCTargetLowering::combineFPToIntToFP(SDNode *N,
14604                                               DAGCombinerInfo &DCI) const {
14605   assert((N->getOpcode() == ISD::SINT_TO_FP ||
14606           N->getOpcode() == ISD::UINT_TO_FP) &&
14607          "Need an int -> FP conversion node here");
14608 
14609   if (useSoftFloat() || !Subtarget.has64BitSupport())
14610     return SDValue();
14611 
14612   SelectionDAG &DAG = DCI.DAG;
14613   SDLoc dl(N);
14614   SDValue Op(N, 0);
14615 
14616   // Don't handle ppc_fp128 here or conversions that are out-of-range capable
14617   // from the hardware.
14618   if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
14619     return SDValue();
14620   if (!Op.getOperand(0).getValueType().isSimple())
14621     return SDValue();
14622   if (Op.getOperand(0).getValueType().getSimpleVT() <= MVT(MVT::i1) ||
14623       Op.getOperand(0).getValueType().getSimpleVT() > MVT(MVT::i64))
14624     return SDValue();
14625 
14626   SDValue FirstOperand(Op.getOperand(0));
14627   bool SubWordLoad = FirstOperand.getOpcode() == ISD::LOAD &&
14628     (FirstOperand.getValueType() == MVT::i8 ||
14629      FirstOperand.getValueType() == MVT::i16);
14630   if (Subtarget.hasP9Vector() && Subtarget.hasP9Altivec() && SubWordLoad) {
14631     bool Signed = N->getOpcode() == ISD::SINT_TO_FP;
14632     bool DstDouble = Op.getValueType() == MVT::f64;
14633     unsigned ConvOp = Signed ?
14634       (DstDouble ? PPCISD::FCFID  : PPCISD::FCFIDS) :
14635       (DstDouble ? PPCISD::FCFIDU : PPCISD::FCFIDUS);
14636     SDValue WidthConst =
14637       DAG.getIntPtrConstant(FirstOperand.getValueType() == MVT::i8 ? 1 : 2,
14638                             dl, false);
14639     LoadSDNode *LDN = cast<LoadSDNode>(FirstOperand.getNode());
14640     SDValue Ops[] = { LDN->getChain(), LDN->getBasePtr(), WidthConst };
14641     SDValue Ld = DAG.getMemIntrinsicNode(PPCISD::LXSIZX, dl,
14642                                          DAG.getVTList(MVT::f64, MVT::Other),
14643                                          Ops, MVT::i8, LDN->getMemOperand());
14644 
14645     // For signed conversion, we need to sign-extend the value in the VSR
14646     if (Signed) {
14647       SDValue ExtOps[] = { Ld, WidthConst };
14648       SDValue Ext = DAG.getNode(PPCISD::VEXTS, dl, MVT::f64, ExtOps);
14649       return DAG.getNode(ConvOp, dl, DstDouble ? MVT::f64 : MVT::f32, Ext);
14650     } else
14651       return DAG.getNode(ConvOp, dl, DstDouble ? MVT::f64 : MVT::f32, Ld);
14652   }
14653 
14654 
14655   // For i32 intermediate values, unfortunately, the conversion functions
14656   // leave the upper 32 bits of the value are undefined. Within the set of
14657   // scalar instructions, we have no method for zero- or sign-extending the
14658   // value. Thus, we cannot handle i32 intermediate values here.
14659   if (Op.getOperand(0).getValueType() == MVT::i32)
14660     return SDValue();
14661 
14662   assert((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) &&
14663          "UINT_TO_FP is supported only with FPCVT");
14664 
14665   // If we have FCFIDS, then use it when converting to single-precision.
14666   // Otherwise, convert to double-precision and then round.
14667   unsigned FCFOp = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32)
14668                        ? (Op.getOpcode() == ISD::UINT_TO_FP ? PPCISD::FCFIDUS
14669                                                             : PPCISD::FCFIDS)
14670                        : (Op.getOpcode() == ISD::UINT_TO_FP ? PPCISD::FCFIDU
14671                                                             : PPCISD::FCFID);
14672   MVT FCFTy = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32)
14673                   ? MVT::f32
14674                   : MVT::f64;
14675 
14676   // If we're converting from a float, to an int, and back to a float again,
14677   // then we don't need the store/load pair at all.
14678   if ((Op.getOperand(0).getOpcode() == ISD::FP_TO_UINT &&
14679        Subtarget.hasFPCVT()) ||
14680       (Op.getOperand(0).getOpcode() == ISD::FP_TO_SINT)) {
14681     SDValue Src = Op.getOperand(0).getOperand(0);
14682     if (Src.getValueType() == MVT::f32) {
14683       Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
14684       DCI.AddToWorklist(Src.getNode());
14685     } else if (Src.getValueType() != MVT::f64) {
14686       // Make sure that we don't pick up a ppc_fp128 source value.
14687       return SDValue();
14688     }
14689 
14690     unsigned FCTOp =
14691       Op.getOperand(0).getOpcode() == ISD::FP_TO_SINT ? PPCISD::FCTIDZ :
14692                                                         PPCISD::FCTIDUZ;
14693 
14694     SDValue Tmp = DAG.getNode(FCTOp, dl, MVT::f64, Src);
14695     SDValue FP = DAG.getNode(FCFOp, dl, FCFTy, Tmp);
14696 
14697     if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT()) {
14698       FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP,
14699                        DAG.getIntPtrConstant(0, dl, /*isTarget=*/true));
14700       DCI.AddToWorklist(FP.getNode());
14701     }
14702 
14703     return FP;
14704   }
14705 
14706   return SDValue();
14707 }
14708 
14709 // expandVSXLoadForLE - Convert VSX loads (which may be intrinsics for
14710 // builtins) into loads with swaps.
14711 SDValue PPCTargetLowering::expandVSXLoadForLE(SDNode *N,
14712                                               DAGCombinerInfo &DCI) const {
14713   // Delay VSX load for LE combine until after LegalizeOps to prioritize other
14714   // load combines.
14715   if (DCI.isBeforeLegalizeOps())
14716     return SDValue();
14717 
14718   SelectionDAG &DAG = DCI.DAG;
14719   SDLoc dl(N);
14720   SDValue Chain;
14721   SDValue Base;
14722   MachineMemOperand *MMO;
14723 
14724   switch (N->getOpcode()) {
14725   default:
14726     llvm_unreachable("Unexpected opcode for little endian VSX load");
14727   case ISD::LOAD: {
14728     LoadSDNode *LD = cast<LoadSDNode>(N);
14729     Chain = LD->getChain();
14730     Base = LD->getBasePtr();
14731     MMO = LD->getMemOperand();
14732     // If the MMO suggests this isn't a load of a full vector, leave
14733     // things alone.  For a built-in, we have to make the change for
14734     // correctness, so if there is a size problem that will be a bug.
14735     if (MMO->getSize() < 16)
14736       return SDValue();
14737     break;
14738   }
14739   case ISD::INTRINSIC_W_CHAIN: {
14740     MemIntrinsicSDNode *Intrin = cast<MemIntrinsicSDNode>(N);
14741     Chain = Intrin->getChain();
14742     // Similarly to the store case below, Intrin->getBasePtr() doesn't get
14743     // us what we want. Get operand 2 instead.
14744     Base = Intrin->getOperand(2);
14745     MMO = Intrin->getMemOperand();
14746     break;
14747   }
14748   }
14749 
14750   MVT VecTy = N->getValueType(0).getSimpleVT();
14751 
14752   SDValue LoadOps[] = { Chain, Base };
14753   SDValue Load = DAG.getMemIntrinsicNode(PPCISD::LXVD2X, dl,
14754                                          DAG.getVTList(MVT::v2f64, MVT::Other),
14755                                          LoadOps, MVT::v2f64, MMO);
14756 
14757   DCI.AddToWorklist(Load.getNode());
14758   Chain = Load.getValue(1);
14759   SDValue Swap = DAG.getNode(
14760       PPCISD::XXSWAPD, dl, DAG.getVTList(MVT::v2f64, MVT::Other), Chain, Load);
14761   DCI.AddToWorklist(Swap.getNode());
14762 
14763   // Add a bitcast if the resulting load type doesn't match v2f64.
14764   if (VecTy != MVT::v2f64) {
14765     SDValue N = DAG.getNode(ISD::BITCAST, dl, VecTy, Swap);
14766     DCI.AddToWorklist(N.getNode());
14767     // Package {bitcast value, swap's chain} to match Load's shape.
14768     return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(VecTy, MVT::Other),
14769                        N, Swap.getValue(1));
14770   }
14771 
14772   return Swap;
14773 }
14774 
14775 // expandVSXStoreForLE - Convert VSX stores (which may be intrinsics for
14776 // builtins) into stores with swaps.
14777 SDValue PPCTargetLowering::expandVSXStoreForLE(SDNode *N,
14778                                                DAGCombinerInfo &DCI) const {
14779   // Delay VSX store for LE combine until after LegalizeOps to prioritize other
14780   // store combines.
14781   if (DCI.isBeforeLegalizeOps())
14782     return SDValue();
14783 
14784   SelectionDAG &DAG = DCI.DAG;
14785   SDLoc dl(N);
14786   SDValue Chain;
14787   SDValue Base;
14788   unsigned SrcOpnd;
14789   MachineMemOperand *MMO;
14790 
14791   switch (N->getOpcode()) {
14792   default:
14793     llvm_unreachable("Unexpected opcode for little endian VSX store");
14794   case ISD::STORE: {
14795     StoreSDNode *ST = cast<StoreSDNode>(N);
14796     Chain = ST->getChain();
14797     Base = ST->getBasePtr();
14798     MMO = ST->getMemOperand();
14799     SrcOpnd = 1;
14800     // If the MMO suggests this isn't a store of a full vector, leave
14801     // things alone.  For a built-in, we have to make the change for
14802     // correctness, so if there is a size problem that will be a bug.
14803     if (MMO->getSize() < 16)
14804       return SDValue();
14805     break;
14806   }
14807   case ISD::INTRINSIC_VOID: {
14808     MemIntrinsicSDNode *Intrin = cast<MemIntrinsicSDNode>(N);
14809     Chain = Intrin->getChain();
14810     // Intrin->getBasePtr() oddly does not get what we want.
14811     Base = Intrin->getOperand(3);
14812     MMO = Intrin->getMemOperand();
14813     SrcOpnd = 2;
14814     break;
14815   }
14816   }
14817 
14818   SDValue Src = N->getOperand(SrcOpnd);
14819   MVT VecTy = Src.getValueType().getSimpleVT();
14820 
14821   // All stores are done as v2f64 and possible bit cast.
14822   if (VecTy != MVT::v2f64) {
14823     Src = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Src);
14824     DCI.AddToWorklist(Src.getNode());
14825   }
14826 
14827   SDValue Swap = DAG.getNode(PPCISD::XXSWAPD, dl,
14828                              DAG.getVTList(MVT::v2f64, MVT::Other), Chain, Src);
14829   DCI.AddToWorklist(Swap.getNode());
14830   Chain = Swap.getValue(1);
14831   SDValue StoreOps[] = { Chain, Swap, Base };
14832   SDValue Store = DAG.getMemIntrinsicNode(PPCISD::STXVD2X, dl,
14833                                           DAG.getVTList(MVT::Other),
14834                                           StoreOps, VecTy, MMO);
14835   DCI.AddToWorklist(Store.getNode());
14836   return Store;
14837 }
14838 
14839 // Handle DAG combine for STORE (FP_TO_INT F).
14840 SDValue PPCTargetLowering::combineStoreFPToInt(SDNode *N,
14841                                                DAGCombinerInfo &DCI) const {
14842 
14843   SelectionDAG &DAG = DCI.DAG;
14844   SDLoc dl(N);
14845   unsigned Opcode = N->getOperand(1).getOpcode();
14846 
14847   assert((Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT)
14848          && "Not a FP_TO_INT Instruction!");
14849 
14850   SDValue Val = N->getOperand(1).getOperand(0);
14851   EVT Op1VT = N->getOperand(1).getValueType();
14852   EVT ResVT = Val.getValueType();
14853 
14854   if (!isTypeLegal(ResVT))
14855     return SDValue();
14856 
14857   // Only perform combine for conversion to i64/i32 or power9 i16/i8.
14858   bool ValidTypeForStoreFltAsInt =
14859         (Op1VT == MVT::i32 || Op1VT == MVT::i64 ||
14860          (Subtarget.hasP9Vector() && (Op1VT == MVT::i16 || Op1VT == MVT::i8)));
14861 
14862   if (ResVT == MVT::f128 && !Subtarget.hasP9Vector())
14863     return SDValue();
14864 
14865   if (ResVT == MVT::ppcf128 || !Subtarget.hasP8Vector() ||
14866       cast<StoreSDNode>(N)->isTruncatingStore() || !ValidTypeForStoreFltAsInt)
14867     return SDValue();
14868 
14869   // Extend f32 values to f64
14870   if (ResVT.getScalarSizeInBits() == 32) {
14871     Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
14872     DCI.AddToWorklist(Val.getNode());
14873   }
14874 
14875   // Set signed or unsigned conversion opcode.
14876   unsigned ConvOpcode = (Opcode == ISD::FP_TO_SINT) ?
14877                           PPCISD::FP_TO_SINT_IN_VSR :
14878                           PPCISD::FP_TO_UINT_IN_VSR;
14879 
14880   Val = DAG.getNode(ConvOpcode,
14881                     dl, ResVT == MVT::f128 ? MVT::f128 : MVT::f64, Val);
14882   DCI.AddToWorklist(Val.getNode());
14883 
14884   // Set number of bytes being converted.
14885   unsigned ByteSize = Op1VT.getScalarSizeInBits() / 8;
14886   SDValue Ops[] = { N->getOperand(0), Val, N->getOperand(2),
14887                     DAG.getIntPtrConstant(ByteSize, dl, false),
14888                     DAG.getValueType(Op1VT) };
14889 
14890   Val = DAG.getMemIntrinsicNode(PPCISD::ST_VSR_SCAL_INT, dl,
14891           DAG.getVTList(MVT::Other), Ops,
14892           cast<StoreSDNode>(N)->getMemoryVT(),
14893           cast<StoreSDNode>(N)->getMemOperand());
14894 
14895   DCI.AddToWorklist(Val.getNode());
14896   return Val;
14897 }
14898 
14899 static bool isAlternatingShuffMask(const ArrayRef<int> &Mask, int NumElts) {
14900   // Check that the source of the element keeps flipping
14901   // (i.e. Mask[i] < NumElts -> Mask[i+i] >= NumElts).
14902   bool PrevElemFromFirstVec = Mask[0] < NumElts;
14903   for (int i = 1, e = Mask.size(); i < e; i++) {
14904     if (PrevElemFromFirstVec && Mask[i] < NumElts)
14905       return false;
14906     if (!PrevElemFromFirstVec && Mask[i] >= NumElts)
14907       return false;
14908     PrevElemFromFirstVec = !PrevElemFromFirstVec;
14909   }
14910   return true;
14911 }
14912 
14913 static bool isSplatBV(SDValue Op) {
14914   if (Op.getOpcode() != ISD::BUILD_VECTOR)
14915     return false;
14916   SDValue FirstOp;
14917 
14918   // Find first non-undef input.
14919   for (int i = 0, e = Op.getNumOperands(); i < e; i++) {
14920     FirstOp = Op.getOperand(i);
14921     if (!FirstOp.isUndef())
14922       break;
14923   }
14924 
14925   // All inputs are undef or the same as the first non-undef input.
14926   for (int i = 1, e = Op.getNumOperands(); i < e; i++)
14927     if (Op.getOperand(i) != FirstOp && !Op.getOperand(i).isUndef())
14928       return false;
14929   return true;
14930 }
14931 
14932 static SDValue isScalarToVec(SDValue Op) {
14933   if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR)
14934     return Op;
14935   if (Op.getOpcode() != ISD::BITCAST)
14936     return SDValue();
14937   Op = Op.getOperand(0);
14938   if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR)
14939     return Op;
14940   return SDValue();
14941 }
14942 
14943 // Fix up the shuffle mask to account for the fact that the result of
14944 // scalar_to_vector is not in lane zero. This just takes all values in
14945 // the ranges specified by the min/max indices and adds the number of
14946 // elements required to ensure each element comes from the respective
14947 // position in the valid lane.
14948 // On little endian, that's just the corresponding element in the other
14949 // half of the vector. On big endian, it is in the same half but right
14950 // justified rather than left justified in that half.
14951 static void fixupShuffleMaskForPermutedSToV(SmallVectorImpl<int> &ShuffV,
14952                                             int LHSMaxIdx, int RHSMinIdx,
14953                                             int RHSMaxIdx, int HalfVec,
14954                                             unsigned ValidLaneWidth,
14955                                             const PPCSubtarget &Subtarget) {
14956   for (int i = 0, e = ShuffV.size(); i < e; i++) {
14957     int Idx = ShuffV[i];
14958     if ((Idx >= 0 && Idx < LHSMaxIdx) || (Idx >= RHSMinIdx && Idx < RHSMaxIdx))
14959       ShuffV[i] +=
14960           Subtarget.isLittleEndian() ? HalfVec : HalfVec - ValidLaneWidth;
14961   }
14962 }
14963 
14964 // Replace a SCALAR_TO_VECTOR with a SCALAR_TO_VECTOR_PERMUTED except if
14965 // the original is:
14966 // (<n x Ty> (scalar_to_vector (Ty (extract_elt <n x Ty> %a, C))))
14967 // In such a case, just change the shuffle mask to extract the element
14968 // from the permuted index.
14969 static SDValue getSToVPermuted(SDValue OrigSToV, SelectionDAG &DAG,
14970                                const PPCSubtarget &Subtarget) {
14971   SDLoc dl(OrigSToV);
14972   EVT VT = OrigSToV.getValueType();
14973   assert(OrigSToV.getOpcode() == ISD::SCALAR_TO_VECTOR &&
14974          "Expecting a SCALAR_TO_VECTOR here");
14975   SDValue Input = OrigSToV.getOperand(0);
14976 
14977   if (Input.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
14978     ConstantSDNode *Idx = dyn_cast<ConstantSDNode>(Input.getOperand(1));
14979     SDValue OrigVector = Input.getOperand(0);
14980 
14981     // Can't handle non-const element indices or different vector types
14982     // for the input to the extract and the output of the scalar_to_vector.
14983     if (Idx && VT == OrigVector.getValueType()) {
14984       unsigned NumElts = VT.getVectorNumElements();
14985       assert(
14986           NumElts > 1 &&
14987           "Cannot produce a permuted scalar_to_vector for one element vector");
14988       SmallVector<int, 16> NewMask(NumElts, -1);
14989       unsigned ResultInElt = NumElts / 2;
14990       ResultInElt -= Subtarget.isLittleEndian() ? 0 : 1;
14991       NewMask[ResultInElt] = Idx->getZExtValue();
14992       return DAG.getVectorShuffle(VT, dl, OrigVector, OrigVector, NewMask);
14993     }
14994   }
14995   return DAG.getNode(PPCISD::SCALAR_TO_VECTOR_PERMUTED, dl, VT,
14996                      OrigSToV.getOperand(0));
14997 }
14998 
14999 // On little endian subtargets, combine shuffles such as:
15000 // vector_shuffle<16,1,17,3,18,5,19,7,20,9,21,11,22,13,23,15>, <zero>, %b
15001 // into:
15002 // vector_shuffle<16,0,17,1,18,2,19,3,20,4,21,5,22,6,23,7>, <zero>, %b
15003 // because the latter can be matched to a single instruction merge.
15004 // Furthermore, SCALAR_TO_VECTOR on little endian always involves a permute
15005 // to put the value into element zero. Adjust the shuffle mask so that the
15006 // vector can remain in permuted form (to prevent a swap prior to a shuffle).
15007 // On big endian targets, this is still useful for SCALAR_TO_VECTOR
15008 // nodes with elements smaller than doubleword because all the ways
15009 // of getting scalar data into a vector register put the value in the
15010 // rightmost element of the left half of the vector.
15011 SDValue PPCTargetLowering::combineVectorShuffle(ShuffleVectorSDNode *SVN,
15012                                                 SelectionDAG &DAG) const {
15013   SDValue LHS = SVN->getOperand(0);
15014   SDValue RHS = SVN->getOperand(1);
15015   auto Mask = SVN->getMask();
15016   int NumElts = LHS.getValueType().getVectorNumElements();
15017   SDValue Res(SVN, 0);
15018   SDLoc dl(SVN);
15019   bool IsLittleEndian = Subtarget.isLittleEndian();
15020 
15021   // On big endian targets this is only useful for subtargets with direct moves.
15022   // On little endian targets it would be useful for all subtargets with VSX.
15023   // However adding special handling for LE subtargets without direct moves
15024   // would be wasted effort since the minimum arch for LE is ISA 2.07 (Power8)
15025   // which includes direct moves.
15026   if (!Subtarget.hasDirectMove())
15027     return Res;
15028 
15029   // If this is not a shuffle of a shuffle and the first element comes from
15030   // the second vector, canonicalize to the commuted form. This will make it
15031   // more likely to match one of the single instruction patterns.
15032   if (Mask[0] >= NumElts && LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
15033       RHS.getOpcode() != ISD::VECTOR_SHUFFLE) {
15034     std::swap(LHS, RHS);
15035     Res = DAG.getCommutedVectorShuffle(*SVN);
15036     Mask = cast<ShuffleVectorSDNode>(Res)->getMask();
15037   }
15038 
15039   // Adjust the shuffle mask if either input vector comes from a
15040   // SCALAR_TO_VECTOR and keep the respective input vector in permuted
15041   // form (to prevent the need for a swap).
15042   SmallVector<int, 16> ShuffV(Mask);
15043   SDValue SToVLHS = isScalarToVec(LHS);
15044   SDValue SToVRHS = isScalarToVec(RHS);
15045   if (SToVLHS || SToVRHS) {
15046     // FIXME: If both LHS and RHS are SCALAR_TO_VECTOR, but are not the
15047     // same type and have differing element sizes, then do not perform
15048     // the following transformation. The current transformation for
15049     // SCALAR_TO_VECTOR assumes that both input vectors have the same
15050     // element size. This will be updated in the future to account for
15051     // differing sizes of the LHS and RHS.
15052     if (SToVLHS && SToVRHS &&
15053         (SToVLHS.getValueType().getScalarSizeInBits() !=
15054          SToVRHS.getValueType().getScalarSizeInBits()))
15055       return Res;
15056 
15057     int NumEltsIn = SToVLHS ? SToVLHS.getValueType().getVectorNumElements()
15058                             : SToVRHS.getValueType().getVectorNumElements();
15059     int NumEltsOut = ShuffV.size();
15060     // The width of the "valid lane" (i.e. the lane that contains the value that
15061     // is vectorized) needs to be expressed in terms of the number of elements
15062     // of the shuffle. It is thereby the ratio of the values before and after
15063     // any bitcast.
15064     unsigned ValidLaneWidth =
15065         SToVLHS ? SToVLHS.getValueType().getScalarSizeInBits() /
15066                       LHS.getValueType().getScalarSizeInBits()
15067                 : SToVRHS.getValueType().getScalarSizeInBits() /
15068                       RHS.getValueType().getScalarSizeInBits();
15069 
15070     // Initially assume that neither input is permuted. These will be adjusted
15071     // accordingly if either input is.
15072     int LHSMaxIdx = -1;
15073     int RHSMinIdx = -1;
15074     int RHSMaxIdx = -1;
15075     int HalfVec = LHS.getValueType().getVectorNumElements() / 2;
15076 
15077     // Get the permuted scalar to vector nodes for the source(s) that come from
15078     // ISD::SCALAR_TO_VECTOR.
15079     // On big endian systems, this only makes sense for element sizes smaller
15080     // than 64 bits since for 64-bit elements, all instructions already put
15081     // the value into element zero. Since scalar size of LHS and RHS may differ
15082     // after isScalarToVec, this should be checked using their own sizes.
15083     if (SToVLHS) {
15084       if (!IsLittleEndian && SToVLHS.getValueType().getScalarSizeInBits() >= 64)
15085         return Res;
15086       // Set up the values for the shuffle vector fixup.
15087       LHSMaxIdx = NumEltsOut / NumEltsIn;
15088       SToVLHS = getSToVPermuted(SToVLHS, DAG, Subtarget);
15089       if (SToVLHS.getValueType() != LHS.getValueType())
15090         SToVLHS = DAG.getBitcast(LHS.getValueType(), SToVLHS);
15091       LHS = SToVLHS;
15092     }
15093     if (SToVRHS) {
15094       if (!IsLittleEndian && SToVRHS.getValueType().getScalarSizeInBits() >= 64)
15095         return Res;
15096       RHSMinIdx = NumEltsOut;
15097       RHSMaxIdx = NumEltsOut / NumEltsIn + RHSMinIdx;
15098       SToVRHS = getSToVPermuted(SToVRHS, DAG, Subtarget);
15099       if (SToVRHS.getValueType() != RHS.getValueType())
15100         SToVRHS = DAG.getBitcast(RHS.getValueType(), SToVRHS);
15101       RHS = SToVRHS;
15102     }
15103 
15104     // Fix up the shuffle mask to reflect where the desired element actually is.
15105     // The minimum and maximum indices that correspond to element zero for both
15106     // the LHS and RHS are computed and will control which shuffle mask entries
15107     // are to be changed. For example, if the RHS is permuted, any shuffle mask
15108     // entries in the range [RHSMinIdx,RHSMaxIdx) will be adjusted.
15109     fixupShuffleMaskForPermutedSToV(ShuffV, LHSMaxIdx, RHSMinIdx, RHSMaxIdx,
15110                                     HalfVec, ValidLaneWidth, Subtarget);
15111     Res = DAG.getVectorShuffle(SVN->getValueType(0), dl, LHS, RHS, ShuffV);
15112 
15113     // We may have simplified away the shuffle. We won't be able to do anything
15114     // further with it here.
15115     if (!isa<ShuffleVectorSDNode>(Res))
15116       return Res;
15117     Mask = cast<ShuffleVectorSDNode>(Res)->getMask();
15118   }
15119 
15120   SDValue TheSplat = IsLittleEndian ? RHS : LHS;
15121   // The common case after we commuted the shuffle is that the RHS is a splat
15122   // and we have elements coming in from the splat at indices that are not
15123   // conducive to using a merge.
15124   // Example:
15125   // vector_shuffle<0,17,1,19,2,21,3,23,4,25,5,27,6,29,7,31> t1, <zero>
15126   if (!isSplatBV(TheSplat))
15127     return Res;
15128 
15129   // We are looking for a mask such that all even elements are from
15130   // one vector and all odd elements from the other.
15131   if (!isAlternatingShuffMask(Mask, NumElts))
15132     return Res;
15133 
15134   // Adjust the mask so we are pulling in the same index from the splat
15135   // as the index from the interesting vector in consecutive elements.
15136   if (IsLittleEndian) {
15137     // Example (even elements from first vector):
15138     // vector_shuffle<0,16,1,17,2,18,3,19,4,20,5,21,6,22,7,23> t1, <zero>
15139     if (Mask[0] < NumElts)
15140       for (int i = 1, e = Mask.size(); i < e; i += 2) {
15141         if (ShuffV[i] < 0)
15142           continue;
15143         ShuffV[i] = (ShuffV[i - 1] + NumElts);
15144       }
15145     // Example (odd elements from first vector):
15146     // vector_shuffle<16,0,17,1,18,2,19,3,20,4,21,5,22,6,23,7> t1, <zero>
15147     else
15148       for (int i = 0, e = Mask.size(); i < e; i += 2) {
15149         if (ShuffV[i] < 0)
15150           continue;
15151         ShuffV[i] = (ShuffV[i + 1] + NumElts);
15152       }
15153   } else {
15154     // Example (even elements from first vector):
15155     // vector_shuffle<0,16,1,17,2,18,3,19,4,20,5,21,6,22,7,23> <zero>, t1
15156     if (Mask[0] < NumElts)
15157       for (int i = 0, e = Mask.size(); i < e; i += 2) {
15158         if (ShuffV[i] < 0)
15159           continue;
15160         ShuffV[i] = ShuffV[i + 1] - NumElts;
15161       }
15162     // Example (odd elements from first vector):
15163     // vector_shuffle<16,0,17,1,18,2,19,3,20,4,21,5,22,6,23,7> <zero>, t1
15164     else
15165       for (int i = 1, e = Mask.size(); i < e; i += 2) {
15166         if (ShuffV[i] < 0)
15167           continue;
15168         ShuffV[i] = ShuffV[i - 1] - NumElts;
15169       }
15170   }
15171 
15172   // If the RHS has undefs, we need to remove them since we may have created
15173   // a shuffle that adds those instead of the splat value.
15174   SDValue SplatVal =
15175       cast<BuildVectorSDNode>(TheSplat.getNode())->getSplatValue();
15176   TheSplat = DAG.getSplatBuildVector(TheSplat.getValueType(), dl, SplatVal);
15177 
15178   if (IsLittleEndian)
15179     RHS = TheSplat;
15180   else
15181     LHS = TheSplat;
15182   return DAG.getVectorShuffle(SVN->getValueType(0), dl, LHS, RHS, ShuffV);
15183 }
15184 
15185 SDValue PPCTargetLowering::combineVReverseMemOP(ShuffleVectorSDNode *SVN,
15186                                                 LSBaseSDNode *LSBase,
15187                                                 DAGCombinerInfo &DCI) const {
15188   assert((ISD::isNormalLoad(LSBase) || ISD::isNormalStore(LSBase)) &&
15189         "Not a reverse memop pattern!");
15190 
15191   auto IsElementReverse = [](const ShuffleVectorSDNode *SVN) -> bool {
15192     auto Mask = SVN->getMask();
15193     int i = 0;
15194     auto I = Mask.rbegin();
15195     auto E = Mask.rend();
15196 
15197     for (; I != E; ++I) {
15198       if (*I != i)
15199         return false;
15200       i++;
15201     }
15202     return true;
15203   };
15204 
15205   SelectionDAG &DAG = DCI.DAG;
15206   EVT VT = SVN->getValueType(0);
15207 
15208   if (!isTypeLegal(VT) || !Subtarget.isLittleEndian() || !Subtarget.hasVSX())
15209     return SDValue();
15210 
15211   // Before P9, we have PPCVSXSwapRemoval pass to hack the element order.
15212   // See comment in PPCVSXSwapRemoval.cpp.
15213   // It is conflict with PPCVSXSwapRemoval opt. So we don't do it.
15214   if (!Subtarget.hasP9Vector())
15215     return SDValue();
15216 
15217   if(!IsElementReverse(SVN))
15218     return SDValue();
15219 
15220   if (LSBase->getOpcode() == ISD::LOAD) {
15221     // If the load return value 0 has more than one user except the
15222     // shufflevector instruction, it is not profitable to replace the
15223     // shufflevector with a reverse load.
15224     for (SDNode::use_iterator UI = LSBase->use_begin(), UE = LSBase->use_end();
15225          UI != UE; ++UI)
15226       if (UI.getUse().getResNo() == 0 && UI->getOpcode() != ISD::VECTOR_SHUFFLE)
15227         return SDValue();
15228 
15229     SDLoc dl(LSBase);
15230     SDValue LoadOps[] = {LSBase->getChain(), LSBase->getBasePtr()};
15231     return DAG.getMemIntrinsicNode(
15232         PPCISD::LOAD_VEC_BE, dl, DAG.getVTList(VT, MVT::Other), LoadOps,
15233         LSBase->getMemoryVT(), LSBase->getMemOperand());
15234   }
15235 
15236   if (LSBase->getOpcode() == ISD::STORE) {
15237     // If there are other uses of the shuffle, the swap cannot be avoided.
15238     // Forcing the use of an X-Form (since swapped stores only have
15239     // X-Forms) without removing the swap is unprofitable.
15240     if (!SVN->hasOneUse())
15241       return SDValue();
15242 
15243     SDLoc dl(LSBase);
15244     SDValue StoreOps[] = {LSBase->getChain(), SVN->getOperand(0),
15245                           LSBase->getBasePtr()};
15246     return DAG.getMemIntrinsicNode(
15247         PPCISD::STORE_VEC_BE, dl, DAG.getVTList(MVT::Other), StoreOps,
15248         LSBase->getMemoryVT(), LSBase->getMemOperand());
15249   }
15250 
15251   llvm_unreachable("Expected a load or store node here");
15252 }
15253 
15254 static bool isStoreConditional(SDValue Intrin, unsigned &StoreWidth) {
15255   unsigned IntrinsicID =
15256       cast<ConstantSDNode>(Intrin.getOperand(1))->getZExtValue();
15257   if (IntrinsicID == Intrinsic::ppc_stdcx)
15258     StoreWidth = 8;
15259   else if (IntrinsicID == Intrinsic::ppc_stwcx)
15260     StoreWidth = 4;
15261   else if (IntrinsicID == Intrinsic::ppc_sthcx)
15262     StoreWidth = 2;
15263   else if (IntrinsicID == Intrinsic::ppc_stbcx)
15264     StoreWidth = 1;
15265   else
15266     return false;
15267   return true;
15268 }
15269 
15270 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
15271                                              DAGCombinerInfo &DCI) const {
15272   SelectionDAG &DAG = DCI.DAG;
15273   SDLoc dl(N);
15274   switch (N->getOpcode()) {
15275   default: break;
15276   case ISD::ADD:
15277     return combineADD(N, DCI);
15278   case ISD::SHL:
15279     return combineSHL(N, DCI);
15280   case ISD::SRA:
15281     return combineSRA(N, DCI);
15282   case ISD::SRL:
15283     return combineSRL(N, DCI);
15284   case ISD::MUL:
15285     return combineMUL(N, DCI);
15286   case ISD::FMA:
15287   case PPCISD::FNMSUB:
15288     return combineFMALike(N, DCI);
15289   case PPCISD::SHL:
15290     if (isNullConstant(N->getOperand(0))) // 0 << V -> 0.
15291         return N->getOperand(0);
15292     break;
15293   case PPCISD::SRL:
15294     if (isNullConstant(N->getOperand(0))) // 0 >>u V -> 0.
15295         return N->getOperand(0);
15296     break;
15297   case PPCISD::SRA:
15298     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
15299       if (C->isZero() ||  //  0 >>s V -> 0.
15300           C->isAllOnes()) // -1 >>s V -> -1.
15301         return N->getOperand(0);
15302     }
15303     break;
15304   case ISD::SIGN_EXTEND:
15305   case ISD::ZERO_EXTEND:
15306   case ISD::ANY_EXTEND:
15307     return DAGCombineExtBoolTrunc(N, DCI);
15308   case ISD::TRUNCATE:
15309     return combineTRUNCATE(N, DCI);
15310   case ISD::SETCC:
15311     if (SDValue CSCC = combineSetCC(N, DCI))
15312       return CSCC;
15313     [[fallthrough]];
15314   case ISD::SELECT_CC:
15315     return DAGCombineTruncBoolExt(N, DCI);
15316   case ISD::SINT_TO_FP:
15317   case ISD::UINT_TO_FP:
15318     return combineFPToIntToFP(N, DCI);
15319   case ISD::VECTOR_SHUFFLE:
15320     if (ISD::isNormalLoad(N->getOperand(0).getNode())) {
15321       LSBaseSDNode* LSBase = cast<LSBaseSDNode>(N->getOperand(0));
15322       return combineVReverseMemOP(cast<ShuffleVectorSDNode>(N), LSBase, DCI);
15323     }
15324     return combineVectorShuffle(cast<ShuffleVectorSDNode>(N), DCI.DAG);
15325   case ISD::STORE: {
15326 
15327     EVT Op1VT = N->getOperand(1).getValueType();
15328     unsigned Opcode = N->getOperand(1).getOpcode();
15329 
15330     if (Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT) {
15331       SDValue Val= combineStoreFPToInt(N, DCI);
15332       if (Val)
15333         return Val;
15334     }
15335 
15336     if (Opcode == ISD::VECTOR_SHUFFLE && ISD::isNormalStore(N)) {
15337       ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N->getOperand(1));
15338       SDValue Val= combineVReverseMemOP(SVN, cast<LSBaseSDNode>(N), DCI);
15339       if (Val)
15340         return Val;
15341     }
15342 
15343     // Turn STORE (BSWAP) -> sthbrx/stwbrx.
15344     if (cast<StoreSDNode>(N)->isUnindexed() && Opcode == ISD::BSWAP &&
15345         N->getOperand(1).getNode()->hasOneUse() &&
15346         (Op1VT == MVT::i32 || Op1VT == MVT::i16 ||
15347          (Subtarget.hasLDBRX() && Subtarget.isPPC64() && Op1VT == MVT::i64))) {
15348 
15349       // STBRX can only handle simple types and it makes no sense to store less
15350       // two bytes in byte-reversed order.
15351       EVT mVT = cast<StoreSDNode>(N)->getMemoryVT();
15352       if (mVT.isExtended() || mVT.getSizeInBits() < 16)
15353         break;
15354 
15355       SDValue BSwapOp = N->getOperand(1).getOperand(0);
15356       // Do an any-extend to 32-bits if this is a half-word input.
15357       if (BSwapOp.getValueType() == MVT::i16)
15358         BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
15359 
15360       // If the type of BSWAP operand is wider than stored memory width
15361       // it need to be shifted to the right side before STBRX.
15362       if (Op1VT.bitsGT(mVT)) {
15363         int Shift = Op1VT.getSizeInBits() - mVT.getSizeInBits();
15364         BSwapOp = DAG.getNode(ISD::SRL, dl, Op1VT, BSwapOp,
15365                               DAG.getConstant(Shift, dl, MVT::i32));
15366         // Need to truncate if this is a bswap of i64 stored as i32/i16.
15367         if (Op1VT == MVT::i64)
15368           BSwapOp = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BSwapOp);
15369       }
15370 
15371       SDValue Ops[] = {
15372         N->getOperand(0), BSwapOp, N->getOperand(2), DAG.getValueType(mVT)
15373       };
15374       return
15375         DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
15376                                 Ops, cast<StoreSDNode>(N)->getMemoryVT(),
15377                                 cast<StoreSDNode>(N)->getMemOperand());
15378     }
15379 
15380     // STORE Constant:i32<0>  ->  STORE<trunc to i32> Constant:i64<0>
15381     // So it can increase the chance of CSE constant construction.
15382     if (Subtarget.isPPC64() && !DCI.isBeforeLegalize() &&
15383         isa<ConstantSDNode>(N->getOperand(1)) && Op1VT == MVT::i32) {
15384       // Need to sign-extended to 64-bits to handle negative values.
15385       EVT MemVT = cast<StoreSDNode>(N)->getMemoryVT();
15386       uint64_t Val64 = SignExtend64(N->getConstantOperandVal(1),
15387                                     MemVT.getSizeInBits());
15388       SDValue Const64 = DAG.getConstant(Val64, dl, MVT::i64);
15389 
15390       // DAG.getTruncStore() can't be used here because it doesn't accept
15391       // the general (base + offset) addressing mode.
15392       // So we use UpdateNodeOperands and setTruncatingStore instead.
15393       DAG.UpdateNodeOperands(N, N->getOperand(0), Const64, N->getOperand(2),
15394                              N->getOperand(3));
15395       cast<StoreSDNode>(N)->setTruncatingStore(true);
15396       return SDValue(N, 0);
15397     }
15398 
15399     // For little endian, VSX stores require generating xxswapd/lxvd2x.
15400     // Not needed on ISA 3.0 based CPUs since we have a non-permuting store.
15401     if (Op1VT.isSimple()) {
15402       MVT StoreVT = Op1VT.getSimpleVT();
15403       if (Subtarget.needsSwapsForVSXMemOps() &&
15404           (StoreVT == MVT::v2f64 || StoreVT == MVT::v2i64 ||
15405            StoreVT == MVT::v4f32 || StoreVT == MVT::v4i32))
15406         return expandVSXStoreForLE(N, DCI);
15407     }
15408     break;
15409   }
15410   case ISD::LOAD: {
15411     LoadSDNode *LD = cast<LoadSDNode>(N);
15412     EVT VT = LD->getValueType(0);
15413 
15414     // For little endian, VSX loads require generating lxvd2x/xxswapd.
15415     // Not needed on ISA 3.0 based CPUs since we have a non-permuting load.
15416     if (VT.isSimple()) {
15417       MVT LoadVT = VT.getSimpleVT();
15418       if (Subtarget.needsSwapsForVSXMemOps() &&
15419           (LoadVT == MVT::v2f64 || LoadVT == MVT::v2i64 ||
15420            LoadVT == MVT::v4f32 || LoadVT == MVT::v4i32))
15421         return expandVSXLoadForLE(N, DCI);
15422     }
15423 
15424     // We sometimes end up with a 64-bit integer load, from which we extract
15425     // two single-precision floating-point numbers. This happens with
15426     // std::complex<float>, and other similar structures, because of the way we
15427     // canonicalize structure copies. However, if we lack direct moves,
15428     // then the final bitcasts from the extracted integer values to the
15429     // floating-point numbers turn into store/load pairs. Even with direct moves,
15430     // just loading the two floating-point numbers is likely better.
15431     auto ReplaceTwoFloatLoad = [&]() {
15432       if (VT != MVT::i64)
15433         return false;
15434 
15435       if (LD->getExtensionType() != ISD::NON_EXTLOAD ||
15436           LD->isVolatile())
15437         return false;
15438 
15439       //  We're looking for a sequence like this:
15440       //  t13: i64,ch = load<LD8[%ref.tmp]> t0, t6, undef:i64
15441       //      t16: i64 = srl t13, Constant:i32<32>
15442       //    t17: i32 = truncate t16
15443       //  t18: f32 = bitcast t17
15444       //    t19: i32 = truncate t13
15445       //  t20: f32 = bitcast t19
15446 
15447       if (!LD->hasNUsesOfValue(2, 0))
15448         return false;
15449 
15450       auto UI = LD->use_begin();
15451       while (UI.getUse().getResNo() != 0) ++UI;
15452       SDNode *Trunc = *UI++;
15453       while (UI.getUse().getResNo() != 0) ++UI;
15454       SDNode *RightShift = *UI;
15455       if (Trunc->getOpcode() != ISD::TRUNCATE)
15456         std::swap(Trunc, RightShift);
15457 
15458       if (Trunc->getOpcode() != ISD::TRUNCATE ||
15459           Trunc->getValueType(0) != MVT::i32 ||
15460           !Trunc->hasOneUse())
15461         return false;
15462       if (RightShift->getOpcode() != ISD::SRL ||
15463           !isa<ConstantSDNode>(RightShift->getOperand(1)) ||
15464           RightShift->getConstantOperandVal(1) != 32 ||
15465           !RightShift->hasOneUse())
15466         return false;
15467 
15468       SDNode *Trunc2 = *RightShift->use_begin();
15469       if (Trunc2->getOpcode() != ISD::TRUNCATE ||
15470           Trunc2->getValueType(0) != MVT::i32 ||
15471           !Trunc2->hasOneUse())
15472         return false;
15473 
15474       SDNode *Bitcast = *Trunc->use_begin();
15475       SDNode *Bitcast2 = *Trunc2->use_begin();
15476 
15477       if (Bitcast->getOpcode() != ISD::BITCAST ||
15478           Bitcast->getValueType(0) != MVT::f32)
15479         return false;
15480       if (Bitcast2->getOpcode() != ISD::BITCAST ||
15481           Bitcast2->getValueType(0) != MVT::f32)
15482         return false;
15483 
15484       if (Subtarget.isLittleEndian())
15485         std::swap(Bitcast, Bitcast2);
15486 
15487       // Bitcast has the second float (in memory-layout order) and Bitcast2
15488       // has the first one.
15489 
15490       SDValue BasePtr = LD->getBasePtr();
15491       if (LD->isIndexed()) {
15492         assert(LD->getAddressingMode() == ISD::PRE_INC &&
15493                "Non-pre-inc AM on PPC?");
15494         BasePtr =
15495           DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
15496                       LD->getOffset());
15497       }
15498 
15499       auto MMOFlags =
15500           LD->getMemOperand()->getFlags() & ~MachineMemOperand::MOVolatile;
15501       SDValue FloatLoad = DAG.getLoad(MVT::f32, dl, LD->getChain(), BasePtr,
15502                                       LD->getPointerInfo(), LD->getAlign(),
15503                                       MMOFlags, LD->getAAInfo());
15504       SDValue AddPtr =
15505         DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(),
15506                     BasePtr, DAG.getIntPtrConstant(4, dl));
15507       SDValue FloatLoad2 = DAG.getLoad(
15508           MVT::f32, dl, SDValue(FloatLoad.getNode(), 1), AddPtr,
15509           LD->getPointerInfo().getWithOffset(4),
15510           commonAlignment(LD->getAlign(), 4), MMOFlags, LD->getAAInfo());
15511 
15512       if (LD->isIndexed()) {
15513         // Note that DAGCombine should re-form any pre-increment load(s) from
15514         // what is produced here if that makes sense.
15515         DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), BasePtr);
15516       }
15517 
15518       DCI.CombineTo(Bitcast2, FloatLoad);
15519       DCI.CombineTo(Bitcast, FloatLoad2);
15520 
15521       DAG.ReplaceAllUsesOfValueWith(SDValue(LD, LD->isIndexed() ? 2 : 1),
15522                                     SDValue(FloatLoad2.getNode(), 1));
15523       return true;
15524     };
15525 
15526     if (ReplaceTwoFloatLoad())
15527       return SDValue(N, 0);
15528 
15529     EVT MemVT = LD->getMemoryVT();
15530     Type *Ty = MemVT.getTypeForEVT(*DAG.getContext());
15531     Align ABIAlignment = DAG.getDataLayout().getABITypeAlign(Ty);
15532     if (LD->isUnindexed() && VT.isVector() &&
15533         ((Subtarget.hasAltivec() && ISD::isNON_EXTLoad(N) &&
15534           // P8 and later hardware should just use LOAD.
15535           !Subtarget.hasP8Vector() &&
15536           (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
15537            VT == MVT::v4f32))) &&
15538         LD->getAlign() < ABIAlignment) {
15539       // This is a type-legal unaligned Altivec load.
15540       SDValue Chain = LD->getChain();
15541       SDValue Ptr = LD->getBasePtr();
15542       bool isLittleEndian = Subtarget.isLittleEndian();
15543 
15544       // This implements the loading of unaligned vectors as described in
15545       // the venerable Apple Velocity Engine overview. Specifically:
15546       // https://developer.apple.com/hardwaredrivers/ve/alignment.html
15547       // https://developer.apple.com/hardwaredrivers/ve/code_optimization.html
15548       //
15549       // The general idea is to expand a sequence of one or more unaligned
15550       // loads into an alignment-based permutation-control instruction (lvsl
15551       // or lvsr), a series of regular vector loads (which always truncate
15552       // their input address to an aligned address), and a series of
15553       // permutations.  The results of these permutations are the requested
15554       // loaded values.  The trick is that the last "extra" load is not taken
15555       // from the address you might suspect (sizeof(vector) bytes after the
15556       // last requested load), but rather sizeof(vector) - 1 bytes after the
15557       // last requested vector. The point of this is to avoid a page fault if
15558       // the base address happened to be aligned. This works because if the
15559       // base address is aligned, then adding less than a full vector length
15560       // will cause the last vector in the sequence to be (re)loaded.
15561       // Otherwise, the next vector will be fetched as you might suspect was
15562       // necessary.
15563 
15564       // We might be able to reuse the permutation generation from
15565       // a different base address offset from this one by an aligned amount.
15566       // The INTRINSIC_WO_CHAIN DAG combine will attempt to perform this
15567       // optimization later.
15568       Intrinsic::ID Intr, IntrLD, IntrPerm;
15569       MVT PermCntlTy, PermTy, LDTy;
15570       Intr = isLittleEndian ? Intrinsic::ppc_altivec_lvsr
15571                             : Intrinsic::ppc_altivec_lvsl;
15572       IntrLD = Intrinsic::ppc_altivec_lvx;
15573       IntrPerm = Intrinsic::ppc_altivec_vperm;
15574       PermCntlTy = MVT::v16i8;
15575       PermTy = MVT::v4i32;
15576       LDTy = MVT::v4i32;
15577 
15578       SDValue PermCntl = BuildIntrinsicOp(Intr, Ptr, DAG, dl, PermCntlTy);
15579 
15580       // Create the new MMO for the new base load. It is like the original MMO,
15581       // but represents an area in memory almost twice the vector size centered
15582       // on the original address. If the address is unaligned, we might start
15583       // reading up to (sizeof(vector)-1) bytes below the address of the
15584       // original unaligned load.
15585       MachineFunction &MF = DAG.getMachineFunction();
15586       MachineMemOperand *BaseMMO =
15587         MF.getMachineMemOperand(LD->getMemOperand(),
15588                                 -(int64_t)MemVT.getStoreSize()+1,
15589                                 2*MemVT.getStoreSize()-1);
15590 
15591       // Create the new base load.
15592       SDValue LDXIntID =
15593           DAG.getTargetConstant(IntrLD, dl, getPointerTy(MF.getDataLayout()));
15594       SDValue BaseLoadOps[] = { Chain, LDXIntID, Ptr };
15595       SDValue BaseLoad =
15596         DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl,
15597                                 DAG.getVTList(PermTy, MVT::Other),
15598                                 BaseLoadOps, LDTy, BaseMMO);
15599 
15600       // Note that the value of IncOffset (which is provided to the next
15601       // load's pointer info offset value, and thus used to calculate the
15602       // alignment), and the value of IncValue (which is actually used to
15603       // increment the pointer value) are different! This is because we
15604       // require the next load to appear to be aligned, even though it
15605       // is actually offset from the base pointer by a lesser amount.
15606       int IncOffset = VT.getSizeInBits() / 8;
15607       int IncValue = IncOffset;
15608 
15609       // Walk (both up and down) the chain looking for another load at the real
15610       // (aligned) offset (the alignment of the other load does not matter in
15611       // this case). If found, then do not use the offset reduction trick, as
15612       // that will prevent the loads from being later combined (as they would
15613       // otherwise be duplicates).
15614       if (!findConsecutiveLoad(LD, DAG))
15615         --IncValue;
15616 
15617       SDValue Increment =
15618           DAG.getConstant(IncValue, dl, getPointerTy(MF.getDataLayout()));
15619       Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15620 
15621       MachineMemOperand *ExtraMMO =
15622         MF.getMachineMemOperand(LD->getMemOperand(),
15623                                 1, 2*MemVT.getStoreSize()-1);
15624       SDValue ExtraLoadOps[] = { Chain, LDXIntID, Ptr };
15625       SDValue ExtraLoad =
15626         DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl,
15627                                 DAG.getVTList(PermTy, MVT::Other),
15628                                 ExtraLoadOps, LDTy, ExtraMMO);
15629 
15630       SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
15631         BaseLoad.getValue(1), ExtraLoad.getValue(1));
15632 
15633       // Because vperm has a big-endian bias, we must reverse the order
15634       // of the input vectors and complement the permute control vector
15635       // when generating little endian code.  We have already handled the
15636       // latter by using lvsr instead of lvsl, so just reverse BaseLoad
15637       // and ExtraLoad here.
15638       SDValue Perm;
15639       if (isLittleEndian)
15640         Perm = BuildIntrinsicOp(IntrPerm,
15641                                 ExtraLoad, BaseLoad, PermCntl, DAG, dl);
15642       else
15643         Perm = BuildIntrinsicOp(IntrPerm,
15644                                 BaseLoad, ExtraLoad, PermCntl, DAG, dl);
15645 
15646       if (VT != PermTy)
15647         Perm = Subtarget.hasAltivec()
15648                    ? DAG.getNode(ISD::BITCAST, dl, VT, Perm)
15649                    : DAG.getNode(ISD::FP_ROUND, dl, VT, Perm,
15650                                  DAG.getTargetConstant(1, dl, MVT::i64));
15651                                // second argument is 1 because this rounding
15652                                // is always exact.
15653 
15654       // The output of the permutation is our loaded result, the TokenFactor is
15655       // our new chain.
15656       DCI.CombineTo(N, Perm, TF);
15657       return SDValue(N, 0);
15658     }
15659     }
15660     break;
15661     case ISD::INTRINSIC_WO_CHAIN: {
15662       bool isLittleEndian = Subtarget.isLittleEndian();
15663       unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
15664       Intrinsic::ID Intr = (isLittleEndian ? Intrinsic::ppc_altivec_lvsr
15665                                            : Intrinsic::ppc_altivec_lvsl);
15666       if (IID == Intr && N->getOperand(1)->getOpcode() == ISD::ADD) {
15667         SDValue Add = N->getOperand(1);
15668 
15669         int Bits = 4 /* 16 byte alignment */;
15670 
15671         if (DAG.MaskedValueIsZero(Add->getOperand(1),
15672                                   APInt::getAllOnes(Bits /* alignment */)
15673                                       .zext(Add.getScalarValueSizeInBits()))) {
15674           SDNode *BasePtr = Add->getOperand(0).getNode();
15675           for (SDNode *U : BasePtr->uses()) {
15676             if (U->getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
15677                 cast<ConstantSDNode>(U->getOperand(0))->getZExtValue() == IID) {
15678               // We've found another LVSL/LVSR, and this address is an aligned
15679               // multiple of that one. The results will be the same, so use the
15680               // one we've just found instead.
15681 
15682               return SDValue(U, 0);
15683             }
15684           }
15685         }
15686 
15687         if (isa<ConstantSDNode>(Add->getOperand(1))) {
15688           SDNode *BasePtr = Add->getOperand(0).getNode();
15689           for (SDNode *U : BasePtr->uses()) {
15690             if (U->getOpcode() == ISD::ADD &&
15691                 isa<ConstantSDNode>(U->getOperand(1)) &&
15692                 (cast<ConstantSDNode>(Add->getOperand(1))->getZExtValue() -
15693                  cast<ConstantSDNode>(U->getOperand(1))->getZExtValue()) %
15694                         (1ULL << Bits) ==
15695                     0) {
15696               SDNode *OtherAdd = U;
15697               for (SDNode *V : OtherAdd->uses()) {
15698                 if (V->getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
15699                     cast<ConstantSDNode>(V->getOperand(0))->getZExtValue() ==
15700                         IID) {
15701                   return SDValue(V, 0);
15702                 }
15703               }
15704             }
15705           }
15706         }
15707       }
15708 
15709       // Combine vmaxsw/h/b(a, a's negation) to abs(a)
15710       // Expose the vabsduw/h/b opportunity for down stream
15711       if (!DCI.isAfterLegalizeDAG() && Subtarget.hasP9Altivec() &&
15712           (IID == Intrinsic::ppc_altivec_vmaxsw ||
15713            IID == Intrinsic::ppc_altivec_vmaxsh ||
15714            IID == Intrinsic::ppc_altivec_vmaxsb)) {
15715         SDValue V1 = N->getOperand(1);
15716         SDValue V2 = N->getOperand(2);
15717         if ((V1.getSimpleValueType() == MVT::v4i32 ||
15718              V1.getSimpleValueType() == MVT::v8i16 ||
15719              V1.getSimpleValueType() == MVT::v16i8) &&
15720             V1.getSimpleValueType() == V2.getSimpleValueType()) {
15721           // (0-a, a)
15722           if (V1.getOpcode() == ISD::SUB &&
15723               ISD::isBuildVectorAllZeros(V1.getOperand(0).getNode()) &&
15724               V1.getOperand(1) == V2) {
15725             return DAG.getNode(ISD::ABS, dl, V2.getValueType(), V2);
15726           }
15727           // (a, 0-a)
15728           if (V2.getOpcode() == ISD::SUB &&
15729               ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()) &&
15730               V2.getOperand(1) == V1) {
15731             return DAG.getNode(ISD::ABS, dl, V1.getValueType(), V1);
15732           }
15733           // (x-y, y-x)
15734           if (V1.getOpcode() == ISD::SUB && V2.getOpcode() == ISD::SUB &&
15735               V1.getOperand(0) == V2.getOperand(1) &&
15736               V1.getOperand(1) == V2.getOperand(0)) {
15737             return DAG.getNode(ISD::ABS, dl, V1.getValueType(), V1);
15738           }
15739         }
15740       }
15741     }
15742 
15743     break;
15744   case ISD::INTRINSIC_W_CHAIN:
15745     // For little endian, VSX loads require generating lxvd2x/xxswapd.
15746     // Not needed on ISA 3.0 based CPUs since we have a non-permuting load.
15747     if (Subtarget.needsSwapsForVSXMemOps()) {
15748       switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
15749       default:
15750         break;
15751       case Intrinsic::ppc_vsx_lxvw4x:
15752       case Intrinsic::ppc_vsx_lxvd2x:
15753         return expandVSXLoadForLE(N, DCI);
15754       }
15755     }
15756     break;
15757   case ISD::INTRINSIC_VOID:
15758     // For little endian, VSX stores require generating xxswapd/stxvd2x.
15759     // Not needed on ISA 3.0 based CPUs since we have a non-permuting store.
15760     if (Subtarget.needsSwapsForVSXMemOps()) {
15761       switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
15762       default:
15763         break;
15764       case Intrinsic::ppc_vsx_stxvw4x:
15765       case Intrinsic::ppc_vsx_stxvd2x:
15766         return expandVSXStoreForLE(N, DCI);
15767       }
15768     }
15769     break;
15770   case ISD::BSWAP: {
15771     // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
15772     // For subtargets without LDBRX, we can still do better than the default
15773     // expansion even for 64-bit BSWAP (LOAD).
15774     bool Is64BitBswapOn64BitTgt =
15775         Subtarget.isPPC64() && N->getValueType(0) == MVT::i64;
15776     bool IsSingleUseNormalLd = ISD::isNormalLoad(N->getOperand(0).getNode()) &&
15777                                N->getOperand(0).hasOneUse();
15778     if (IsSingleUseNormalLd &&
15779         (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16 ||
15780          (Subtarget.hasLDBRX() && Is64BitBswapOn64BitTgt))) {
15781       SDValue Load = N->getOperand(0);
15782       LoadSDNode *LD = cast<LoadSDNode>(Load);
15783       // Create the byte-swapping load.
15784       SDValue Ops[] = {
15785         LD->getChain(),    // Chain
15786         LD->getBasePtr(),  // Ptr
15787         DAG.getValueType(N->getValueType(0)) // VT
15788       };
15789       SDValue BSLoad =
15790         DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
15791                                 DAG.getVTList(N->getValueType(0) == MVT::i64 ?
15792                                               MVT::i64 : MVT::i32, MVT::Other),
15793                                 Ops, LD->getMemoryVT(), LD->getMemOperand());
15794 
15795       // If this is an i16 load, insert the truncate.
15796       SDValue ResVal = BSLoad;
15797       if (N->getValueType(0) == MVT::i16)
15798         ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
15799 
15800       // First, combine the bswap away.  This makes the value produced by the
15801       // load dead.
15802       DCI.CombineTo(N, ResVal);
15803 
15804       // Next, combine the load away, we give it a bogus result value but a real
15805       // chain result.  The result value is dead because the bswap is dead.
15806       DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
15807 
15808       // Return N so it doesn't get rechecked!
15809       return SDValue(N, 0);
15810     }
15811     // Convert this to two 32-bit bswap loads and a BUILD_PAIR. Do this only
15812     // before legalization so that the BUILD_PAIR is handled correctly.
15813     if (!DCI.isBeforeLegalize() || !Is64BitBswapOn64BitTgt ||
15814         !IsSingleUseNormalLd)
15815       return SDValue();
15816     LoadSDNode *LD = cast<LoadSDNode>(N->getOperand(0));
15817 
15818     // Can't split volatile or atomic loads.
15819     if (!LD->isSimple())
15820       return SDValue();
15821     SDValue BasePtr = LD->getBasePtr();
15822     SDValue Lo = DAG.getLoad(MVT::i32, dl, LD->getChain(), BasePtr,
15823                              LD->getPointerInfo(), LD->getAlign());
15824     Lo = DAG.getNode(ISD::BSWAP, dl, MVT::i32, Lo);
15825     BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
15826                           DAG.getIntPtrConstant(4, dl));
15827     MachineMemOperand *NewMMO = DAG.getMachineFunction().getMachineMemOperand(
15828         LD->getMemOperand(), 4, 4);
15829     SDValue Hi = DAG.getLoad(MVT::i32, dl, LD->getChain(), BasePtr, NewMMO);
15830     Hi = DAG.getNode(ISD::BSWAP, dl, MVT::i32, Hi);
15831     SDValue Res;
15832     if (Subtarget.isLittleEndian())
15833       Res = DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Hi, Lo);
15834     else
15835       Res = DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
15836     SDValue TF =
15837         DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
15838                     Hi.getOperand(0).getValue(1), Lo.getOperand(0).getValue(1));
15839     DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), TF);
15840     return Res;
15841   }
15842   case PPCISD::VCMP:
15843     // If a VCMP_rec node already exists with exactly the same operands as this
15844     // node, use its result instead of this node (VCMP_rec computes both a CR6
15845     // and a normal output).
15846     //
15847     if (!N->getOperand(0).hasOneUse() &&
15848         !N->getOperand(1).hasOneUse() &&
15849         !N->getOperand(2).hasOneUse()) {
15850 
15851       // Scan all of the users of the LHS, looking for VCMP_rec's that match.
15852       SDNode *VCMPrecNode = nullptr;
15853 
15854       SDNode *LHSN = N->getOperand(0).getNode();
15855       for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
15856            UI != E; ++UI)
15857         if (UI->getOpcode() == PPCISD::VCMP_rec &&
15858             UI->getOperand(1) == N->getOperand(1) &&
15859             UI->getOperand(2) == N->getOperand(2) &&
15860             UI->getOperand(0) == N->getOperand(0)) {
15861           VCMPrecNode = *UI;
15862           break;
15863         }
15864 
15865       // If there is no VCMP_rec node, or if the flag value has a single use,
15866       // don't transform this.
15867       if (!VCMPrecNode || VCMPrecNode->hasNUsesOfValue(0, 1))
15868         break;
15869 
15870       // Look at the (necessarily single) use of the flag value.  If it has a
15871       // chain, this transformation is more complex.  Note that multiple things
15872       // could use the value result, which we should ignore.
15873       SDNode *FlagUser = nullptr;
15874       for (SDNode::use_iterator UI = VCMPrecNode->use_begin();
15875            FlagUser == nullptr; ++UI) {
15876         assert(UI != VCMPrecNode->use_end() && "Didn't find user!");
15877         SDNode *User = *UI;
15878         for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
15879           if (User->getOperand(i) == SDValue(VCMPrecNode, 1)) {
15880             FlagUser = User;
15881             break;
15882           }
15883         }
15884       }
15885 
15886       // If the user is a MFOCRF instruction, we know this is safe.
15887       // Otherwise we give up for right now.
15888       if (FlagUser->getOpcode() == PPCISD::MFOCRF)
15889         return SDValue(VCMPrecNode, 0);
15890     }
15891     break;
15892   case ISD::BR_CC: {
15893     // If this is a branch on an altivec predicate comparison, lower this so
15894     // that we don't have to do a MFOCRF: instead, branch directly on CR6.  This
15895     // lowering is done pre-legalize, because the legalizer lowers the predicate
15896     // compare down to code that is difficult to reassemble.
15897     // This code also handles branches that depend on the result of a store
15898     // conditional.
15899     ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
15900     SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
15901 
15902     int CompareOpc;
15903     bool isDot;
15904 
15905     if (!isa<ConstantSDNode>(RHS) || (CC != ISD::SETEQ && CC != ISD::SETNE))
15906       break;
15907 
15908     // Since we are doing this pre-legalize, the RHS can be a constant of
15909     // arbitrary bitwidth which may cause issues when trying to get the value
15910     // from the underlying APInt.
15911     auto RHSAPInt = cast<ConstantSDNode>(RHS)->getAPIntValue();
15912     if (!RHSAPInt.isIntN(64))
15913       break;
15914 
15915     unsigned Val = RHSAPInt.getZExtValue();
15916     auto isImpossibleCompare = [&]() {
15917       // If this is a comparison against something other than 0/1, then we know
15918       // that the condition is never/always true.
15919       if (Val != 0 && Val != 1) {
15920         if (CC == ISD::SETEQ)      // Cond never true, remove branch.
15921           return N->getOperand(0);
15922         // Always !=, turn it into an unconditional branch.
15923         return DAG.getNode(ISD::BR, dl, MVT::Other,
15924                            N->getOperand(0), N->getOperand(4));
15925       }
15926       return SDValue();
15927     };
15928     // Combine branches fed by store conditional instructions (st[bhwd]cx).
15929     unsigned StoreWidth = 0;
15930     if (LHS.getOpcode() == ISD::INTRINSIC_W_CHAIN &&
15931         isStoreConditional(LHS, StoreWidth)) {
15932       if (SDValue Impossible = isImpossibleCompare())
15933         return Impossible;
15934       PPC::Predicate CompOpc;
15935       // eq 0 => ne
15936       // ne 0 => eq
15937       // eq 1 => eq
15938       // ne 1 => ne
15939       if (Val == 0)
15940         CompOpc = CC == ISD::SETEQ ? PPC::PRED_NE : PPC::PRED_EQ;
15941       else
15942         CompOpc = CC == ISD::SETEQ ? PPC::PRED_EQ : PPC::PRED_NE;
15943 
15944       SDValue Ops[] = {LHS.getOperand(0), LHS.getOperand(2), LHS.getOperand(3),
15945                        DAG.getConstant(StoreWidth, dl, MVT::i32)};
15946       auto *MemNode = cast<MemSDNode>(LHS);
15947       SDValue ConstSt = DAG.getMemIntrinsicNode(
15948           PPCISD::STORE_COND, dl,
15949           DAG.getVTList(MVT::i32, MVT::Other, MVT::Glue), Ops,
15950           MemNode->getMemoryVT(), MemNode->getMemOperand());
15951 
15952       SDValue InChain;
15953       // Unchain the branch from the original store conditional.
15954       if (N->getOperand(0) == LHS.getValue(1))
15955         InChain = LHS.getOperand(0);
15956       else if (N->getOperand(0).getOpcode() == ISD::TokenFactor) {
15957         SmallVector<SDValue, 4> InChains;
15958         SDValue InTF = N->getOperand(0);
15959         for (int i = 0, e = InTF.getNumOperands(); i < e; i++)
15960           if (InTF.getOperand(i) != LHS.getValue(1))
15961             InChains.push_back(InTF.getOperand(i));
15962         InChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, InChains);
15963       }
15964 
15965       return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, InChain,
15966                          DAG.getConstant(CompOpc, dl, MVT::i32),
15967                          DAG.getRegister(PPC::CR0, MVT::i32), N->getOperand(4),
15968                          ConstSt.getValue(2));
15969     }
15970 
15971     if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
15972         getVectorCompareInfo(LHS, CompareOpc, isDot, Subtarget)) {
15973       assert(isDot && "Can't compare against a vector result!");
15974 
15975       if (SDValue Impossible = isImpossibleCompare())
15976         return Impossible;
15977 
15978       bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
15979       // Create the PPCISD altivec 'dot' comparison node.
15980       SDValue Ops[] = {
15981         LHS.getOperand(2),  // LHS of compare
15982         LHS.getOperand(3),  // RHS of compare
15983         DAG.getConstant(CompareOpc, dl, MVT::i32)
15984       };
15985       EVT VTs[] = { LHS.getOperand(2).getValueType(), MVT::Glue };
15986       SDValue CompNode = DAG.getNode(PPCISD::VCMP_rec, dl, VTs, Ops);
15987 
15988       // Unpack the result based on how the target uses it.
15989       PPC::Predicate CompOpc;
15990       switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
15991       default:  // Can't happen, don't crash on invalid number though.
15992       case 0:   // Branch on the value of the EQ bit of CR6.
15993         CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
15994         break;
15995       case 1:   // Branch on the inverted value of the EQ bit of CR6.
15996         CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
15997         break;
15998       case 2:   // Branch on the value of the LT bit of CR6.
15999         CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
16000         break;
16001       case 3:   // Branch on the inverted value of the LT bit of CR6.
16002         CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
16003         break;
16004       }
16005 
16006       return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
16007                          DAG.getConstant(CompOpc, dl, MVT::i32),
16008                          DAG.getRegister(PPC::CR6, MVT::i32),
16009                          N->getOperand(4), CompNode.getValue(1));
16010     }
16011     break;
16012   }
16013   case ISD::BUILD_VECTOR:
16014     return DAGCombineBuildVector(N, DCI);
16015   case ISD::VSELECT:
16016     return combineVSelect(N, DCI);
16017   }
16018 
16019   return SDValue();
16020 }
16021 
16022 SDValue
16023 PPCTargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
16024                                  SelectionDAG &DAG,
16025                                  SmallVectorImpl<SDNode *> &Created) const {
16026   // fold (sdiv X, pow2)
16027   EVT VT = N->getValueType(0);
16028   if (VT == MVT::i64 && !Subtarget.isPPC64())
16029     return SDValue();
16030   if ((VT != MVT::i32 && VT != MVT::i64) ||
16031       !(Divisor.isPowerOf2() || Divisor.isNegatedPowerOf2()))
16032     return SDValue();
16033 
16034   SDLoc DL(N);
16035   SDValue N0 = N->getOperand(0);
16036 
16037   bool IsNegPow2 = Divisor.isNegatedPowerOf2();
16038   unsigned Lg2 = (IsNegPow2 ? -Divisor : Divisor).countTrailingZeros();
16039   SDValue ShiftAmt = DAG.getConstant(Lg2, DL, VT);
16040 
16041   SDValue Op = DAG.getNode(PPCISD::SRA_ADDZE, DL, VT, N0, ShiftAmt);
16042   Created.push_back(Op.getNode());
16043 
16044   if (IsNegPow2) {
16045     Op = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Op);
16046     Created.push_back(Op.getNode());
16047   }
16048 
16049   return Op;
16050 }
16051 
16052 //===----------------------------------------------------------------------===//
16053 // Inline Assembly Support
16054 //===----------------------------------------------------------------------===//
16055 
16056 void PPCTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
16057                                                       KnownBits &Known,
16058                                                       const APInt &DemandedElts,
16059                                                       const SelectionDAG &DAG,
16060                                                       unsigned Depth) const {
16061   Known.resetAll();
16062   switch (Op.getOpcode()) {
16063   default: break;
16064   case PPCISD::LBRX: {
16065     // lhbrx is known to have the top bits cleared out.
16066     if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
16067       Known.Zero = 0xFFFF0000;
16068     break;
16069   }
16070   case ISD::INTRINSIC_WO_CHAIN: {
16071     switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
16072     default: break;
16073     case Intrinsic::ppc_altivec_vcmpbfp_p:
16074     case Intrinsic::ppc_altivec_vcmpeqfp_p:
16075     case Intrinsic::ppc_altivec_vcmpequb_p:
16076     case Intrinsic::ppc_altivec_vcmpequh_p:
16077     case Intrinsic::ppc_altivec_vcmpequw_p:
16078     case Intrinsic::ppc_altivec_vcmpequd_p:
16079     case Intrinsic::ppc_altivec_vcmpequq_p:
16080     case Intrinsic::ppc_altivec_vcmpgefp_p:
16081     case Intrinsic::ppc_altivec_vcmpgtfp_p:
16082     case Intrinsic::ppc_altivec_vcmpgtsb_p:
16083     case Intrinsic::ppc_altivec_vcmpgtsh_p:
16084     case Intrinsic::ppc_altivec_vcmpgtsw_p:
16085     case Intrinsic::ppc_altivec_vcmpgtsd_p:
16086     case Intrinsic::ppc_altivec_vcmpgtsq_p:
16087     case Intrinsic::ppc_altivec_vcmpgtub_p:
16088     case Intrinsic::ppc_altivec_vcmpgtuh_p:
16089     case Intrinsic::ppc_altivec_vcmpgtuw_p:
16090     case Intrinsic::ppc_altivec_vcmpgtud_p:
16091     case Intrinsic::ppc_altivec_vcmpgtuq_p:
16092       Known.Zero = ~1U;  // All bits but the low one are known to be zero.
16093       break;
16094     }
16095     break;
16096   }
16097   case ISD::INTRINSIC_W_CHAIN: {
16098     switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
16099     default:
16100       break;
16101     case Intrinsic::ppc_load2r:
16102       // Top bits are cleared for load2r (which is the same as lhbrx).
16103       Known.Zero = 0xFFFF0000;
16104       break;
16105     }
16106     break;
16107   }
16108   }
16109 }
16110 
16111 Align PPCTargetLowering::getPrefLoopAlignment(MachineLoop *ML) const {
16112   switch (Subtarget.getCPUDirective()) {
16113   default: break;
16114   case PPC::DIR_970:
16115   case PPC::DIR_PWR4:
16116   case PPC::DIR_PWR5:
16117   case PPC::DIR_PWR5X:
16118   case PPC::DIR_PWR6:
16119   case PPC::DIR_PWR6X:
16120   case PPC::DIR_PWR7:
16121   case PPC::DIR_PWR8:
16122   case PPC::DIR_PWR9:
16123   case PPC::DIR_PWR10:
16124   case PPC::DIR_PWR_FUTURE: {
16125     if (!ML)
16126       break;
16127 
16128     if (!DisableInnermostLoopAlign32) {
16129       // If the nested loop is an innermost loop, prefer to a 32-byte alignment,
16130       // so that we can decrease cache misses and branch-prediction misses.
16131       // Actual alignment of the loop will depend on the hotness check and other
16132       // logic in alignBlocks.
16133       if (ML->getLoopDepth() > 1 && ML->getSubLoops().empty())
16134         return Align(32);
16135     }
16136 
16137     const PPCInstrInfo *TII = Subtarget.getInstrInfo();
16138 
16139     // For small loops (between 5 and 8 instructions), align to a 32-byte
16140     // boundary so that the entire loop fits in one instruction-cache line.
16141     uint64_t LoopSize = 0;
16142     for (auto I = ML->block_begin(), IE = ML->block_end(); I != IE; ++I)
16143       for (const MachineInstr &J : **I) {
16144         LoopSize += TII->getInstSizeInBytes(J);
16145         if (LoopSize > 32)
16146           break;
16147       }
16148 
16149     if (LoopSize > 16 && LoopSize <= 32)
16150       return Align(32);
16151 
16152     break;
16153   }
16154   }
16155 
16156   return TargetLowering::getPrefLoopAlignment(ML);
16157 }
16158 
16159 /// getConstraintType - Given a constraint, return the type of
16160 /// constraint it is for this target.
16161 PPCTargetLowering::ConstraintType
16162 PPCTargetLowering::getConstraintType(StringRef Constraint) const {
16163   if (Constraint.size() == 1) {
16164     switch (Constraint[0]) {
16165     default: break;
16166     case 'b':
16167     case 'r':
16168     case 'f':
16169     case 'd':
16170     case 'v':
16171     case 'y':
16172       return C_RegisterClass;
16173     case 'Z':
16174       // FIXME: While Z does indicate a memory constraint, it specifically
16175       // indicates an r+r address (used in conjunction with the 'y' modifier
16176       // in the replacement string). Currently, we're forcing the base
16177       // register to be r0 in the asm printer (which is interpreted as zero)
16178       // and forming the complete address in the second register. This is
16179       // suboptimal.
16180       return C_Memory;
16181     }
16182   } else if (Constraint == "wc") { // individual CR bits.
16183     return C_RegisterClass;
16184   } else if (Constraint == "wa" || Constraint == "wd" ||
16185              Constraint == "wf" || Constraint == "ws" ||
16186              Constraint == "wi" || Constraint == "ww") {
16187     return C_RegisterClass; // VSX registers.
16188   }
16189   return TargetLowering::getConstraintType(Constraint);
16190 }
16191 
16192 /// Examine constraint type and operand type and determine a weight value.
16193 /// This object must already have been set up with the operand type
16194 /// and the current alternative constraint selected.
16195 TargetLowering::ConstraintWeight
16196 PPCTargetLowering::getSingleConstraintMatchWeight(
16197     AsmOperandInfo &info, const char *constraint) const {
16198   ConstraintWeight weight = CW_Invalid;
16199   Value *CallOperandVal = info.CallOperandVal;
16200     // If we don't have a value, we can't do a match,
16201     // but allow it at the lowest weight.
16202   if (!CallOperandVal)
16203     return CW_Default;
16204   Type *type = CallOperandVal->getType();
16205 
16206   // Look at the constraint type.
16207   if (StringRef(constraint) == "wc" && type->isIntegerTy(1))
16208     return CW_Register; // an individual CR bit.
16209   else if ((StringRef(constraint) == "wa" ||
16210             StringRef(constraint) == "wd" ||
16211             StringRef(constraint) == "wf") &&
16212            type->isVectorTy())
16213     return CW_Register;
16214   else if (StringRef(constraint) == "wi" && type->isIntegerTy(64))
16215     return CW_Register; // just hold 64-bit integers data.
16216   else if (StringRef(constraint) == "ws" && type->isDoubleTy())
16217     return CW_Register;
16218   else if (StringRef(constraint) == "ww" && type->isFloatTy())
16219     return CW_Register;
16220 
16221   switch (*constraint) {
16222   default:
16223     weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
16224     break;
16225   case 'b':
16226     if (type->isIntegerTy())
16227       weight = CW_Register;
16228     break;
16229   case 'f':
16230     if (type->isFloatTy())
16231       weight = CW_Register;
16232     break;
16233   case 'd':
16234     if (type->isDoubleTy())
16235       weight = CW_Register;
16236     break;
16237   case 'v':
16238     if (type->isVectorTy())
16239       weight = CW_Register;
16240     break;
16241   case 'y':
16242     weight = CW_Register;
16243     break;
16244   case 'Z':
16245     weight = CW_Memory;
16246     break;
16247   }
16248   return weight;
16249 }
16250 
16251 std::pair<unsigned, const TargetRegisterClass *>
16252 PPCTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
16253                                                 StringRef Constraint,
16254                                                 MVT VT) const {
16255   if (Constraint.size() == 1) {
16256     // GCC RS6000 Constraint Letters
16257     switch (Constraint[0]) {
16258     case 'b':   // R1-R31
16259       if (VT == MVT::i64 && Subtarget.isPPC64())
16260         return std::make_pair(0U, &PPC::G8RC_NOX0RegClass);
16261       return std::make_pair(0U, &PPC::GPRC_NOR0RegClass);
16262     case 'r':   // R0-R31
16263       if (VT == MVT::i64 && Subtarget.isPPC64())
16264         return std::make_pair(0U, &PPC::G8RCRegClass);
16265       return std::make_pair(0U, &PPC::GPRCRegClass);
16266     // 'd' and 'f' constraints are both defined to be "the floating point
16267     // registers", where one is for 32-bit and the other for 64-bit. We don't
16268     // really care overly much here so just give them all the same reg classes.
16269     case 'd':
16270     case 'f':
16271       if (Subtarget.hasSPE()) {
16272         if (VT == MVT::f32 || VT == MVT::i32)
16273           return std::make_pair(0U, &PPC::GPRCRegClass);
16274         if (VT == MVT::f64 || VT == MVT::i64)
16275           return std::make_pair(0U, &PPC::SPERCRegClass);
16276       } else {
16277         if (VT == MVT::f32 || VT == MVT::i32)
16278           return std::make_pair(0U, &PPC::F4RCRegClass);
16279         if (VT == MVT::f64 || VT == MVT::i64)
16280           return std::make_pair(0U, &PPC::F8RCRegClass);
16281       }
16282       break;
16283     case 'v':
16284       if (Subtarget.hasAltivec() && VT.isVector())
16285         return std::make_pair(0U, &PPC::VRRCRegClass);
16286       else if (Subtarget.hasVSX())
16287         // Scalars in Altivec registers only make sense with VSX.
16288         return std::make_pair(0U, &PPC::VFRCRegClass);
16289       break;
16290     case 'y':   // crrc
16291       return std::make_pair(0U, &PPC::CRRCRegClass);
16292     }
16293   } else if (Constraint == "wc" && Subtarget.useCRBits()) {
16294     // An individual CR bit.
16295     return std::make_pair(0U, &PPC::CRBITRCRegClass);
16296   } else if ((Constraint == "wa" || Constraint == "wd" ||
16297              Constraint == "wf" || Constraint == "wi") &&
16298              Subtarget.hasVSX()) {
16299     // A VSX register for either a scalar (FP) or vector. There is no
16300     // support for single precision scalars on subtargets prior to Power8.
16301     if (VT.isVector())
16302       return std::make_pair(0U, &PPC::VSRCRegClass);
16303     if (VT == MVT::f32 && Subtarget.hasP8Vector())
16304       return std::make_pair(0U, &PPC::VSSRCRegClass);
16305     return std::make_pair(0U, &PPC::VSFRCRegClass);
16306   } else if ((Constraint == "ws" || Constraint == "ww") && Subtarget.hasVSX()) {
16307     if (VT == MVT::f32 && Subtarget.hasP8Vector())
16308       return std::make_pair(0U, &PPC::VSSRCRegClass);
16309     else
16310       return std::make_pair(0U, &PPC::VSFRCRegClass);
16311   } else if (Constraint == "lr") {
16312     if (VT == MVT::i64)
16313       return std::make_pair(0U, &PPC::LR8RCRegClass);
16314     else
16315       return std::make_pair(0U, &PPC::LRRCRegClass);
16316   }
16317 
16318   // Handle special cases of physical registers that are not properly handled
16319   // by the base class.
16320   if (Constraint[0] == '{' && Constraint[Constraint.size() - 1] == '}') {
16321     // If we name a VSX register, we can't defer to the base class because it
16322     // will not recognize the correct register (their names will be VSL{0-31}
16323     // and V{0-31} so they won't match). So we match them here.
16324     if (Constraint.size() > 3 && Constraint[1] == 'v' && Constraint[2] == 's') {
16325       int VSNum = atoi(Constraint.data() + 3);
16326       assert(VSNum >= 0 && VSNum <= 63 &&
16327              "Attempted to access a vsr out of range");
16328       if (VSNum < 32)
16329         return std::make_pair(PPC::VSL0 + VSNum, &PPC::VSRCRegClass);
16330       return std::make_pair(PPC::V0 + VSNum - 32, &PPC::VSRCRegClass);
16331     }
16332 
16333     // For float registers, we can't defer to the base class as it will match
16334     // the SPILLTOVSRRC class.
16335     if (Constraint.size() > 3 && Constraint[1] == 'f') {
16336       int RegNum = atoi(Constraint.data() + 2);
16337       if (RegNum > 31 || RegNum < 0)
16338         report_fatal_error("Invalid floating point register number");
16339       if (VT == MVT::f32 || VT == MVT::i32)
16340         return Subtarget.hasSPE()
16341                    ? std::make_pair(PPC::R0 + RegNum, &PPC::GPRCRegClass)
16342                    : std::make_pair(PPC::F0 + RegNum, &PPC::F4RCRegClass);
16343       if (VT == MVT::f64 || VT == MVT::i64)
16344         return Subtarget.hasSPE()
16345                    ? std::make_pair(PPC::S0 + RegNum, &PPC::SPERCRegClass)
16346                    : std::make_pair(PPC::F0 + RegNum, &PPC::F8RCRegClass);
16347     }
16348   }
16349 
16350   std::pair<unsigned, const TargetRegisterClass *> R =
16351       TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
16352 
16353   // r[0-9]+ are used, on PPC64, to refer to the corresponding 64-bit registers
16354   // (which we call X[0-9]+). If a 64-bit value has been requested, and a
16355   // 32-bit GPR has been selected, then 'upgrade' it to the 64-bit parent
16356   // register.
16357   // FIXME: If TargetLowering::getRegForInlineAsmConstraint could somehow use
16358   // the AsmName field from *RegisterInfo.td, then this would not be necessary.
16359   if (R.first && VT == MVT::i64 && Subtarget.isPPC64() &&
16360       PPC::GPRCRegClass.contains(R.first))
16361     return std::make_pair(TRI->getMatchingSuperReg(R.first,
16362                             PPC::sub_32, &PPC::G8RCRegClass),
16363                           &PPC::G8RCRegClass);
16364 
16365   // GCC accepts 'cc' as an alias for 'cr0', and we need to do the same.
16366   if (!R.second && StringRef("{cc}").equals_insensitive(Constraint)) {
16367     R.first = PPC::CR0;
16368     R.second = &PPC::CRRCRegClass;
16369   }
16370   // FIXME: This warning should ideally be emitted in the front end.
16371   const auto &TM = getTargetMachine();
16372   if (Subtarget.isAIXABI() && !TM.getAIXExtendedAltivecABI()) {
16373     if (((R.first >= PPC::V20 && R.first <= PPC::V31) ||
16374          (R.first >= PPC::VF20 && R.first <= PPC::VF31)) &&
16375         (R.second == &PPC::VSRCRegClass || R.second == &PPC::VSFRCRegClass))
16376       errs() << "warning: vector registers 20 to 32 are reserved in the "
16377                 "default AIX AltiVec ABI and cannot be used\n";
16378   }
16379 
16380   return R;
16381 }
16382 
16383 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
16384 /// vector.  If it is invalid, don't add anything to Ops.
16385 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
16386                                                      std::string &Constraint,
16387                                                      std::vector<SDValue>&Ops,
16388                                                      SelectionDAG &DAG) const {
16389   SDValue Result;
16390 
16391   // Only support length 1 constraints.
16392   if (Constraint.length() > 1) return;
16393 
16394   char Letter = Constraint[0];
16395   switch (Letter) {
16396   default: break;
16397   case 'I':
16398   case 'J':
16399   case 'K':
16400   case 'L':
16401   case 'M':
16402   case 'N':
16403   case 'O':
16404   case 'P': {
16405     ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
16406     if (!CST) return; // Must be an immediate to match.
16407     SDLoc dl(Op);
16408     int64_t Value = CST->getSExtValue();
16409     EVT TCVT = MVT::i64; // All constants taken to be 64 bits so that negative
16410                          // numbers are printed as such.
16411     switch (Letter) {
16412     default: llvm_unreachable("Unknown constraint letter!");
16413     case 'I':  // "I" is a signed 16-bit constant.
16414       if (isInt<16>(Value))
16415         Result = DAG.getTargetConstant(Value, dl, TCVT);
16416       break;
16417     case 'J':  // "J" is a constant with only the high-order 16 bits nonzero.
16418       if (isShiftedUInt<16, 16>(Value))
16419         Result = DAG.getTargetConstant(Value, dl, TCVT);
16420       break;
16421     case 'L':  // "L" is a signed 16-bit constant shifted left 16 bits.
16422       if (isShiftedInt<16, 16>(Value))
16423         Result = DAG.getTargetConstant(Value, dl, TCVT);
16424       break;
16425     case 'K':  // "K" is a constant with only the low-order 16 bits nonzero.
16426       if (isUInt<16>(Value))
16427         Result = DAG.getTargetConstant(Value, dl, TCVT);
16428       break;
16429     case 'M':  // "M" is a constant that is greater than 31.
16430       if (Value > 31)
16431         Result = DAG.getTargetConstant(Value, dl, TCVT);
16432       break;
16433     case 'N':  // "N" is a positive constant that is an exact power of two.
16434       if (Value > 0 && isPowerOf2_64(Value))
16435         Result = DAG.getTargetConstant(Value, dl, TCVT);
16436       break;
16437     case 'O':  // "O" is the constant zero.
16438       if (Value == 0)
16439         Result = DAG.getTargetConstant(Value, dl, TCVT);
16440       break;
16441     case 'P':  // "P" is a constant whose negation is a signed 16-bit constant.
16442       if (isInt<16>(-Value))
16443         Result = DAG.getTargetConstant(Value, dl, TCVT);
16444       break;
16445     }
16446     break;
16447   }
16448   }
16449 
16450   if (Result.getNode()) {
16451     Ops.push_back(Result);
16452     return;
16453   }
16454 
16455   // Handle standard constraint letters.
16456   TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
16457 }
16458 
16459 void PPCTargetLowering::CollectTargetIntrinsicOperands(const CallInst &I,
16460                                               SmallVectorImpl<SDValue> &Ops,
16461                                               SelectionDAG &DAG) const {
16462   if (I.getNumOperands() <= 1)
16463     return;
16464   if (!isa<ConstantSDNode>(Ops[1].getNode()))
16465     return;
16466   auto IntrinsicID = cast<ConstantSDNode>(Ops[1].getNode())->getZExtValue();
16467   if (IntrinsicID != Intrinsic::ppc_tdw && IntrinsicID != Intrinsic::ppc_tw &&
16468       IntrinsicID != Intrinsic::ppc_trapd && IntrinsicID != Intrinsic::ppc_trap)
16469     return;
16470 
16471   if (I.hasMetadata("annotation")) {
16472     MDNode *MDN = I.getMetadata("annotation");
16473     Ops.push_back(DAG.getMDNode(MDN));
16474   }
16475 }
16476 
16477 // isLegalAddressingMode - Return true if the addressing mode represented
16478 // by AM is legal for this target, for a load/store of the specified type.
16479 bool PPCTargetLowering::isLegalAddressingMode(const DataLayout &DL,
16480                                               const AddrMode &AM, Type *Ty,
16481                                               unsigned AS,
16482                                               Instruction *I) const {
16483   // Vector type r+i form is supported since power9 as DQ form. We don't check
16484   // the offset matching DQ form requirement(off % 16 == 0), because on PowerPC,
16485   // imm form is preferred and the offset can be adjusted to use imm form later
16486   // in pass PPCLoopInstrFormPrep. Also in LSR, for one LSRUse, it uses min and
16487   // max offset to check legal addressing mode, we should be a little aggressive
16488   // to contain other offsets for that LSRUse.
16489   if (Ty->isVectorTy() && AM.BaseOffs != 0 && !Subtarget.hasP9Vector())
16490     return false;
16491 
16492   // PPC allows a sign-extended 16-bit immediate field.
16493   if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
16494     return false;
16495 
16496   // No global is ever allowed as a base.
16497   if (AM.BaseGV)
16498     return false;
16499 
16500   // PPC only support r+r,
16501   switch (AM.Scale) {
16502   case 0:  // "r+i" or just "i", depending on HasBaseReg.
16503     break;
16504   case 1:
16505     if (AM.HasBaseReg && AM.BaseOffs)  // "r+r+i" is not allowed.
16506       return false;
16507     // Otherwise we have r+r or r+i.
16508     break;
16509   case 2:
16510     if (AM.HasBaseReg || AM.BaseOffs)  // 2*r+r  or  2*r+i is not allowed.
16511       return false;
16512     // Allow 2*r as r+r.
16513     break;
16514   default:
16515     // No other scales are supported.
16516     return false;
16517   }
16518 
16519   return true;
16520 }
16521 
16522 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
16523                                            SelectionDAG &DAG) const {
16524   MachineFunction &MF = DAG.getMachineFunction();
16525   MachineFrameInfo &MFI = MF.getFrameInfo();
16526   MFI.setReturnAddressIsTaken(true);
16527 
16528   if (verifyReturnAddressArgumentIsConstant(Op, DAG))
16529     return SDValue();
16530 
16531   SDLoc dl(Op);
16532   unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16533 
16534   // Make sure the function does not optimize away the store of the RA to
16535   // the stack.
16536   PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
16537   FuncInfo->setLRStoreRequired();
16538   bool isPPC64 = Subtarget.isPPC64();
16539   auto PtrVT = getPointerTy(MF.getDataLayout());
16540 
16541   if (Depth > 0) {
16542     // The link register (return address) is saved in the caller's frame
16543     // not the callee's stack frame. So we must get the caller's frame
16544     // address and load the return address at the LR offset from there.
16545     SDValue FrameAddr =
16546         DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
16547                     LowerFRAMEADDR(Op, DAG), MachinePointerInfo());
16548     SDValue Offset =
16549         DAG.getConstant(Subtarget.getFrameLowering()->getReturnSaveOffset(), dl,
16550                         isPPC64 ? MVT::i64 : MVT::i32);
16551     return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16552                        DAG.getNode(ISD::ADD, dl, PtrVT, FrameAddr, Offset),
16553                        MachinePointerInfo());
16554   }
16555 
16556   // Just load the return address off the stack.
16557   SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
16558   return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), RetAddrFI,
16559                      MachinePointerInfo());
16560 }
16561 
16562 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
16563                                           SelectionDAG &DAG) const {
16564   SDLoc dl(Op);
16565   unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16566 
16567   MachineFunction &MF = DAG.getMachineFunction();
16568   MachineFrameInfo &MFI = MF.getFrameInfo();
16569   MFI.setFrameAddressIsTaken(true);
16570 
16571   EVT PtrVT = getPointerTy(MF.getDataLayout());
16572   bool isPPC64 = PtrVT == MVT::i64;
16573 
16574   // Naked functions never have a frame pointer, and so we use r1. For all
16575   // other functions, this decision must be delayed until during PEI.
16576   unsigned FrameReg;
16577   if (MF.getFunction().hasFnAttribute(Attribute::Naked))
16578     FrameReg = isPPC64 ? PPC::X1 : PPC::R1;
16579   else
16580     FrameReg = isPPC64 ? PPC::FP8 : PPC::FP;
16581 
16582   SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
16583                                          PtrVT);
16584   while (Depth--)
16585     FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
16586                             FrameAddr, MachinePointerInfo());
16587   return FrameAddr;
16588 }
16589 
16590 // FIXME? Maybe this could be a TableGen attribute on some registers and
16591 // this table could be generated automatically from RegInfo.
16592 Register PPCTargetLowering::getRegisterByName(const char* RegName, LLT VT,
16593                                               const MachineFunction &MF) const {
16594   bool isPPC64 = Subtarget.isPPC64();
16595 
16596   bool is64Bit = isPPC64 && VT == LLT::scalar(64);
16597   if (!is64Bit && VT != LLT::scalar(32))
16598     report_fatal_error("Invalid register global variable type");
16599 
16600   Register Reg = StringSwitch<Register>(RegName)
16601                      .Case("r1", is64Bit ? PPC::X1 : PPC::R1)
16602                      .Case("r2", isPPC64 ? Register() : PPC::R2)
16603                      .Case("r13", (is64Bit ? PPC::X13 : PPC::R13))
16604                      .Default(Register());
16605 
16606   if (Reg)
16607     return Reg;
16608   report_fatal_error("Invalid register name global variable");
16609 }
16610 
16611 bool PPCTargetLowering::isAccessedAsGotIndirect(SDValue GA) const {
16612   // 32-bit SVR4 ABI access everything as got-indirect.
16613   if (Subtarget.is32BitELFABI())
16614     return true;
16615 
16616   // AIX accesses everything indirectly through the TOC, which is similar to
16617   // the GOT.
16618   if (Subtarget.isAIXABI())
16619     return true;
16620 
16621   CodeModel::Model CModel = getTargetMachine().getCodeModel();
16622   // If it is small or large code model, module locals are accessed
16623   // indirectly by loading their address from .toc/.got.
16624   if (CModel == CodeModel::Small || CModel == CodeModel::Large)
16625     return true;
16626 
16627   // JumpTable and BlockAddress are accessed as got-indirect.
16628   if (isa<JumpTableSDNode>(GA) || isa<BlockAddressSDNode>(GA))
16629     return true;
16630 
16631   if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA))
16632     return Subtarget.isGVIndirectSymbol(G->getGlobal());
16633 
16634   return false;
16635 }
16636 
16637 bool
16638 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
16639   // The PowerPC target isn't yet aware of offsets.
16640   return false;
16641 }
16642 
16643 bool PPCTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
16644                                            const CallInst &I,
16645                                            MachineFunction &MF,
16646                                            unsigned Intrinsic) const {
16647   switch (Intrinsic) {
16648   case Intrinsic::ppc_atomicrmw_xchg_i128:
16649   case Intrinsic::ppc_atomicrmw_add_i128:
16650   case Intrinsic::ppc_atomicrmw_sub_i128:
16651   case Intrinsic::ppc_atomicrmw_nand_i128:
16652   case Intrinsic::ppc_atomicrmw_and_i128:
16653   case Intrinsic::ppc_atomicrmw_or_i128:
16654   case Intrinsic::ppc_atomicrmw_xor_i128:
16655   case Intrinsic::ppc_cmpxchg_i128:
16656     Info.opc = ISD::INTRINSIC_W_CHAIN;
16657     Info.memVT = MVT::i128;
16658     Info.ptrVal = I.getArgOperand(0);
16659     Info.offset = 0;
16660     Info.align = Align(16);
16661     Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore |
16662                  MachineMemOperand::MOVolatile;
16663     return true;
16664   case Intrinsic::ppc_atomic_load_i128:
16665     Info.opc = ISD::INTRINSIC_W_CHAIN;
16666     Info.memVT = MVT::i128;
16667     Info.ptrVal = I.getArgOperand(0);
16668     Info.offset = 0;
16669     Info.align = Align(16);
16670     Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile;
16671     return true;
16672   case Intrinsic::ppc_atomic_store_i128:
16673     Info.opc = ISD::INTRINSIC_VOID;
16674     Info.memVT = MVT::i128;
16675     Info.ptrVal = I.getArgOperand(2);
16676     Info.offset = 0;
16677     Info.align = Align(16);
16678     Info.flags = MachineMemOperand::MOStore | MachineMemOperand::MOVolatile;
16679     return true;
16680   case Intrinsic::ppc_altivec_lvx:
16681   case Intrinsic::ppc_altivec_lvxl:
16682   case Intrinsic::ppc_altivec_lvebx:
16683   case Intrinsic::ppc_altivec_lvehx:
16684   case Intrinsic::ppc_altivec_lvewx:
16685   case Intrinsic::ppc_vsx_lxvd2x:
16686   case Intrinsic::ppc_vsx_lxvw4x:
16687   case Intrinsic::ppc_vsx_lxvd2x_be:
16688   case Intrinsic::ppc_vsx_lxvw4x_be:
16689   case Intrinsic::ppc_vsx_lxvl:
16690   case Intrinsic::ppc_vsx_lxvll: {
16691     EVT VT;
16692     switch (Intrinsic) {
16693     case Intrinsic::ppc_altivec_lvebx:
16694       VT = MVT::i8;
16695       break;
16696     case Intrinsic::ppc_altivec_lvehx:
16697       VT = MVT::i16;
16698       break;
16699     case Intrinsic::ppc_altivec_lvewx:
16700       VT = MVT::i32;
16701       break;
16702     case Intrinsic::ppc_vsx_lxvd2x:
16703     case Intrinsic::ppc_vsx_lxvd2x_be:
16704       VT = MVT::v2f64;
16705       break;
16706     default:
16707       VT = MVT::v4i32;
16708       break;
16709     }
16710 
16711     Info.opc = ISD::INTRINSIC_W_CHAIN;
16712     Info.memVT = VT;
16713     Info.ptrVal = I.getArgOperand(0);
16714     Info.offset = -VT.getStoreSize()+1;
16715     Info.size = 2*VT.getStoreSize()-1;
16716     Info.align = Align(1);
16717     Info.flags = MachineMemOperand::MOLoad;
16718     return true;
16719   }
16720   case Intrinsic::ppc_altivec_stvx:
16721   case Intrinsic::ppc_altivec_stvxl:
16722   case Intrinsic::ppc_altivec_stvebx:
16723   case Intrinsic::ppc_altivec_stvehx:
16724   case Intrinsic::ppc_altivec_stvewx:
16725   case Intrinsic::ppc_vsx_stxvd2x:
16726   case Intrinsic::ppc_vsx_stxvw4x:
16727   case Intrinsic::ppc_vsx_stxvd2x_be:
16728   case Intrinsic::ppc_vsx_stxvw4x_be:
16729   case Intrinsic::ppc_vsx_stxvl:
16730   case Intrinsic::ppc_vsx_stxvll: {
16731     EVT VT;
16732     switch (Intrinsic) {
16733     case Intrinsic::ppc_altivec_stvebx:
16734       VT = MVT::i8;
16735       break;
16736     case Intrinsic::ppc_altivec_stvehx:
16737       VT = MVT::i16;
16738       break;
16739     case Intrinsic::ppc_altivec_stvewx:
16740       VT = MVT::i32;
16741       break;
16742     case Intrinsic::ppc_vsx_stxvd2x:
16743     case Intrinsic::ppc_vsx_stxvd2x_be:
16744       VT = MVT::v2f64;
16745       break;
16746     default:
16747       VT = MVT::v4i32;
16748       break;
16749     }
16750 
16751     Info.opc = ISD::INTRINSIC_VOID;
16752     Info.memVT = VT;
16753     Info.ptrVal = I.getArgOperand(1);
16754     Info.offset = -VT.getStoreSize()+1;
16755     Info.size = 2*VT.getStoreSize()-1;
16756     Info.align = Align(1);
16757     Info.flags = MachineMemOperand::MOStore;
16758     return true;
16759   }
16760   case Intrinsic::ppc_stdcx:
16761   case Intrinsic::ppc_stwcx:
16762   case Intrinsic::ppc_sthcx:
16763   case Intrinsic::ppc_stbcx: {
16764     EVT VT;
16765     auto Alignment = Align(8);
16766     switch (Intrinsic) {
16767     case Intrinsic::ppc_stdcx:
16768       VT = MVT::i64;
16769       break;
16770     case Intrinsic::ppc_stwcx:
16771       VT = MVT::i32;
16772       Alignment = Align(4);
16773       break;
16774     case Intrinsic::ppc_sthcx:
16775       VT = MVT::i16;
16776       Alignment = Align(2);
16777       break;
16778     case Intrinsic::ppc_stbcx:
16779       VT = MVT::i8;
16780       Alignment = Align(1);
16781       break;
16782     }
16783     Info.opc = ISD::INTRINSIC_W_CHAIN;
16784     Info.memVT = VT;
16785     Info.ptrVal = I.getArgOperand(0);
16786     Info.offset = 0;
16787     Info.align = Alignment;
16788     Info.flags = MachineMemOperand::MOStore | MachineMemOperand::MOVolatile;
16789     return true;
16790   }
16791   default:
16792     break;
16793   }
16794 
16795   return false;
16796 }
16797 
16798 /// It returns EVT::Other if the type should be determined using generic
16799 /// target-independent logic.
16800 EVT PPCTargetLowering::getOptimalMemOpType(
16801     const MemOp &Op, const AttributeList &FuncAttributes) const {
16802   if (getTargetMachine().getOptLevel() != CodeGenOpt::None) {
16803     // We should use Altivec/VSX loads and stores when available. For unaligned
16804     // addresses, unaligned VSX loads are only fast starting with the P8.
16805     if (Subtarget.hasAltivec() && Op.size() >= 16 &&
16806         (Op.isAligned(Align(16)) ||
16807          ((Op.isMemset() && Subtarget.hasVSX()) || Subtarget.hasP8Vector())))
16808       return MVT::v4i32;
16809   }
16810 
16811   if (Subtarget.isPPC64()) {
16812     return MVT::i64;
16813   }
16814 
16815   return MVT::i32;
16816 }
16817 
16818 /// Returns true if it is beneficial to convert a load of a constant
16819 /// to just the constant itself.
16820 bool PPCTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
16821                                                           Type *Ty) const {
16822   assert(Ty->isIntegerTy());
16823 
16824   unsigned BitSize = Ty->getPrimitiveSizeInBits();
16825   return !(BitSize == 0 || BitSize > 64);
16826 }
16827 
16828 bool PPCTargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
16829   if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
16830     return false;
16831   unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
16832   unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
16833   return NumBits1 == 64 && NumBits2 == 32;
16834 }
16835 
16836 bool PPCTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
16837   if (!VT1.isInteger() || !VT2.isInteger())
16838     return false;
16839   unsigned NumBits1 = VT1.getSizeInBits();
16840   unsigned NumBits2 = VT2.getSizeInBits();
16841   return NumBits1 == 64 && NumBits2 == 32;
16842 }
16843 
16844 bool PPCTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
16845   // Generally speaking, zexts are not free, but they are free when they can be
16846   // folded with other operations.
16847   if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Val)) {
16848     EVT MemVT = LD->getMemoryVT();
16849     if ((MemVT == MVT::i1 || MemVT == MVT::i8 || MemVT == MVT::i16 ||
16850          (Subtarget.isPPC64() && MemVT == MVT::i32)) &&
16851         (LD->getExtensionType() == ISD::NON_EXTLOAD ||
16852          LD->getExtensionType() == ISD::ZEXTLOAD))
16853       return true;
16854   }
16855 
16856   // FIXME: Add other cases...
16857   //  - 32-bit shifts with a zext to i64
16858   //  - zext after ctlz, bswap, etc.
16859   //  - zext after and by a constant mask
16860 
16861   return TargetLowering::isZExtFree(Val, VT2);
16862 }
16863 
16864 bool PPCTargetLowering::isFPExtFree(EVT DestVT, EVT SrcVT) const {
16865   assert(DestVT.isFloatingPoint() && SrcVT.isFloatingPoint() &&
16866          "invalid fpext types");
16867   // Extending to float128 is not free.
16868   if (DestVT == MVT::f128)
16869     return false;
16870   return true;
16871 }
16872 
16873 bool PPCTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
16874   return isInt<16>(Imm) || isUInt<16>(Imm);
16875 }
16876 
16877 bool PPCTargetLowering::isLegalAddImmediate(int64_t Imm) const {
16878   return isInt<16>(Imm) || isUInt<16>(Imm);
16879 }
16880 
16881 bool PPCTargetLowering::allowsMisalignedMemoryAccesses(EVT VT, unsigned, Align,
16882                                                        MachineMemOperand::Flags,
16883                                                        unsigned *Fast) const {
16884   if (DisablePPCUnaligned)
16885     return false;
16886 
16887   // PowerPC supports unaligned memory access for simple non-vector types.
16888   // Although accessing unaligned addresses is not as efficient as accessing
16889   // aligned addresses, it is generally more efficient than manual expansion,
16890   // and generally only traps for software emulation when crossing page
16891   // boundaries.
16892 
16893   if (!VT.isSimple())
16894     return false;
16895 
16896   if (VT.isFloatingPoint() && !VT.isVector() &&
16897       !Subtarget.allowsUnalignedFPAccess())
16898     return false;
16899 
16900   if (VT.getSimpleVT().isVector()) {
16901     if (Subtarget.hasVSX()) {
16902       if (VT != MVT::v2f64 && VT != MVT::v2i64 &&
16903           VT != MVT::v4f32 && VT != MVT::v4i32)
16904         return false;
16905     } else {
16906       return false;
16907     }
16908   }
16909 
16910   if (VT == MVT::ppcf128)
16911     return false;
16912 
16913   if (Fast)
16914     *Fast = 1;
16915 
16916   return true;
16917 }
16918 
16919 bool PPCTargetLowering::decomposeMulByConstant(LLVMContext &Context, EVT VT,
16920                                                SDValue C) const {
16921   // Check integral scalar types.
16922   if (!VT.isScalarInteger())
16923     return false;
16924   if (auto *ConstNode = dyn_cast<ConstantSDNode>(C.getNode())) {
16925     if (!ConstNode->getAPIntValue().isSignedIntN(64))
16926       return false;
16927     // This transformation will generate >= 2 operations. But the following
16928     // cases will generate <= 2 instructions during ISEL. So exclude them.
16929     // 1. If the constant multiplier fits 16 bits, it can be handled by one
16930     // HW instruction, ie. MULLI
16931     // 2. If the multiplier after shifted fits 16 bits, an extra shift
16932     // instruction is needed than case 1, ie. MULLI and RLDICR
16933     int64_t Imm = ConstNode->getSExtValue();
16934     unsigned Shift = countTrailingZeros<uint64_t>(Imm);
16935     Imm >>= Shift;
16936     if (isInt<16>(Imm))
16937       return false;
16938     uint64_t UImm = static_cast<uint64_t>(Imm);
16939     if (isPowerOf2_64(UImm + 1) || isPowerOf2_64(UImm - 1) ||
16940         isPowerOf2_64(1 - UImm) || isPowerOf2_64(-1 - UImm))
16941       return true;
16942   }
16943   return false;
16944 }
16945 
16946 bool PPCTargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
16947                                                    EVT VT) const {
16948   return isFMAFasterThanFMulAndFAdd(
16949       MF.getFunction(), VT.getTypeForEVT(MF.getFunction().getContext()));
16950 }
16951 
16952 bool PPCTargetLowering::isFMAFasterThanFMulAndFAdd(const Function &F,
16953                                                    Type *Ty) const {
16954   if (Subtarget.hasSPE())
16955     return false;
16956   switch (Ty->getScalarType()->getTypeID()) {
16957   case Type::FloatTyID:
16958   case Type::DoubleTyID:
16959     return true;
16960   case Type::FP128TyID:
16961     return Subtarget.hasP9Vector();
16962   default:
16963     return false;
16964   }
16965 }
16966 
16967 // FIXME: add more patterns which are not profitable to hoist.
16968 bool PPCTargetLowering::isProfitableToHoist(Instruction *I) const {
16969   if (!I->hasOneUse())
16970     return true;
16971 
16972   Instruction *User = I->user_back();
16973   assert(User && "A single use instruction with no uses.");
16974 
16975   switch (I->getOpcode()) {
16976   case Instruction::FMul: {
16977     // Don't break FMA, PowerPC prefers FMA.
16978     if (User->getOpcode() != Instruction::FSub &&
16979         User->getOpcode() != Instruction::FAdd)
16980       return true;
16981 
16982     const TargetOptions &Options = getTargetMachine().Options;
16983     const Function *F = I->getFunction();
16984     const DataLayout &DL = F->getParent()->getDataLayout();
16985     Type *Ty = User->getOperand(0)->getType();
16986 
16987     return !(
16988         isFMAFasterThanFMulAndFAdd(*F, Ty) &&
16989         isOperationLegalOrCustom(ISD::FMA, getValueType(DL, Ty)) &&
16990         (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath));
16991   }
16992   case Instruction::Load: {
16993     // Don't break "store (load float*)" pattern, this pattern will be combined
16994     // to "store (load int32)" in later InstCombine pass. See function
16995     // combineLoadToOperationType. On PowerPC, loading a float point takes more
16996     // cycles than loading a 32 bit integer.
16997     LoadInst *LI = cast<LoadInst>(I);
16998     // For the loads that combineLoadToOperationType does nothing, like
16999     // ordered load, it should be profitable to hoist them.
17000     // For swifterror load, it can only be used for pointer to pointer type, so
17001     // later type check should get rid of this case.
17002     if (!LI->isUnordered())
17003       return true;
17004 
17005     if (User->getOpcode() != Instruction::Store)
17006       return true;
17007 
17008     if (I->getType()->getTypeID() != Type::FloatTyID)
17009       return true;
17010 
17011     return false;
17012   }
17013   default:
17014     return true;
17015   }
17016   return true;
17017 }
17018 
17019 const MCPhysReg *
17020 PPCTargetLowering::getScratchRegisters(CallingConv::ID) const {
17021   // LR is a callee-save register, but we must treat it as clobbered by any call
17022   // site. Hence we include LR in the scratch registers, which are in turn added
17023   // as implicit-defs for stackmaps and patchpoints. The same reasoning applies
17024   // to CTR, which is used by any indirect call.
17025   static const MCPhysReg ScratchRegs[] = {
17026     PPC::X12, PPC::LR8, PPC::CTR8, 0
17027   };
17028 
17029   return ScratchRegs;
17030 }
17031 
17032 Register PPCTargetLowering::getExceptionPointerRegister(
17033     const Constant *PersonalityFn) const {
17034   return Subtarget.isPPC64() ? PPC::X3 : PPC::R3;
17035 }
17036 
17037 Register PPCTargetLowering::getExceptionSelectorRegister(
17038     const Constant *PersonalityFn) const {
17039   return Subtarget.isPPC64() ? PPC::X4 : PPC::R4;
17040 }
17041 
17042 bool
17043 PPCTargetLowering::shouldExpandBuildVectorWithShuffles(
17044                      EVT VT , unsigned DefinedValues) const {
17045   if (VT == MVT::v2i64)
17046     return Subtarget.hasDirectMove(); // Don't need stack ops with direct moves
17047 
17048   if (Subtarget.hasVSX())
17049     return true;
17050 
17051   return TargetLowering::shouldExpandBuildVectorWithShuffles(VT, DefinedValues);
17052 }
17053 
17054 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
17055   if (DisableILPPref || Subtarget.enableMachineScheduler())
17056     return TargetLowering::getSchedulingPreference(N);
17057 
17058   return Sched::ILP;
17059 }
17060 
17061 // Create a fast isel object.
17062 FastISel *
17063 PPCTargetLowering::createFastISel(FunctionLoweringInfo &FuncInfo,
17064                                   const TargetLibraryInfo *LibInfo) const {
17065   return PPC::createFastISel(FuncInfo, LibInfo);
17066 }
17067 
17068 // 'Inverted' means the FMA opcode after negating one multiplicand.
17069 // For example, (fma -a b c) = (fnmsub a b c)
17070 static unsigned invertFMAOpcode(unsigned Opc) {
17071   switch (Opc) {
17072   default:
17073     llvm_unreachable("Invalid FMA opcode for PowerPC!");
17074   case ISD::FMA:
17075     return PPCISD::FNMSUB;
17076   case PPCISD::FNMSUB:
17077     return ISD::FMA;
17078   }
17079 }
17080 
17081 SDValue PPCTargetLowering::getNegatedExpression(SDValue Op, SelectionDAG &DAG,
17082                                                 bool LegalOps, bool OptForSize,
17083                                                 NegatibleCost &Cost,
17084                                                 unsigned Depth) const {
17085   if (Depth > SelectionDAG::MaxRecursionDepth)
17086     return SDValue();
17087 
17088   unsigned Opc = Op.getOpcode();
17089   EVT VT = Op.getValueType();
17090   SDNodeFlags Flags = Op.getNode()->getFlags();
17091 
17092   switch (Opc) {
17093   case PPCISD::FNMSUB:
17094     if (!Op.hasOneUse() || !isTypeLegal(VT))
17095       break;
17096 
17097     const TargetOptions &Options = getTargetMachine().Options;
17098     SDValue N0 = Op.getOperand(0);
17099     SDValue N1 = Op.getOperand(1);
17100     SDValue N2 = Op.getOperand(2);
17101     SDLoc Loc(Op);
17102 
17103     NegatibleCost N2Cost = NegatibleCost::Expensive;
17104     SDValue NegN2 =
17105         getNegatedExpression(N2, DAG, LegalOps, OptForSize, N2Cost, Depth + 1);
17106 
17107     if (!NegN2)
17108       return SDValue();
17109 
17110     // (fneg (fnmsub a b c)) => (fnmsub (fneg a) b (fneg c))
17111     // (fneg (fnmsub a b c)) => (fnmsub a (fneg b) (fneg c))
17112     // These transformations may change sign of zeroes. For example,
17113     // -(-ab-(-c))=-0 while -(-(ab-c))=+0 when a=b=c=1.
17114     if (Flags.hasNoSignedZeros() || Options.NoSignedZerosFPMath) {
17115       // Try and choose the cheaper one to negate.
17116       NegatibleCost N0Cost = NegatibleCost::Expensive;
17117       SDValue NegN0 = getNegatedExpression(N0, DAG, LegalOps, OptForSize,
17118                                            N0Cost, Depth + 1);
17119 
17120       NegatibleCost N1Cost = NegatibleCost::Expensive;
17121       SDValue NegN1 = getNegatedExpression(N1, DAG, LegalOps, OptForSize,
17122                                            N1Cost, Depth + 1);
17123 
17124       if (NegN0 && N0Cost <= N1Cost) {
17125         Cost = std::min(N0Cost, N2Cost);
17126         return DAG.getNode(Opc, Loc, VT, NegN0, N1, NegN2, Flags);
17127       } else if (NegN1) {
17128         Cost = std::min(N1Cost, N2Cost);
17129         return DAG.getNode(Opc, Loc, VT, N0, NegN1, NegN2, Flags);
17130       }
17131     }
17132 
17133     // (fneg (fnmsub a b c)) => (fma a b (fneg c))
17134     if (isOperationLegal(ISD::FMA, VT)) {
17135       Cost = N2Cost;
17136       return DAG.getNode(ISD::FMA, Loc, VT, N0, N1, NegN2, Flags);
17137     }
17138 
17139     break;
17140   }
17141 
17142   return TargetLowering::getNegatedExpression(Op, DAG, LegalOps, OptForSize,
17143                                               Cost, Depth);
17144 }
17145 
17146 // Override to enable LOAD_STACK_GUARD lowering on Linux.
17147 bool PPCTargetLowering::useLoadStackGuardNode() const {
17148   if (!Subtarget.isTargetLinux())
17149     return TargetLowering::useLoadStackGuardNode();
17150   return true;
17151 }
17152 
17153 // Override to disable global variable loading on Linux and insert AIX canary
17154 // word declaration.
17155 void PPCTargetLowering::insertSSPDeclarations(Module &M) const {
17156   if (Subtarget.isAIXABI()) {
17157     M.getOrInsertGlobal(AIXSSPCanaryWordName,
17158                         Type::getInt8PtrTy(M.getContext()));
17159     return;
17160   }
17161   if (!Subtarget.isTargetLinux())
17162     return TargetLowering::insertSSPDeclarations(M);
17163 }
17164 
17165 Value *PPCTargetLowering::getSDagStackGuard(const Module &M) const {
17166   if (Subtarget.isAIXABI())
17167     return M.getGlobalVariable(AIXSSPCanaryWordName);
17168   return TargetLowering::getSDagStackGuard(M);
17169 }
17170 
17171 bool PPCTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
17172                                      bool ForCodeSize) const {
17173   if (!VT.isSimple() || !Subtarget.hasVSX())
17174     return false;
17175 
17176   switch(VT.getSimpleVT().SimpleTy) {
17177   default:
17178     // For FP types that are currently not supported by PPC backend, return
17179     // false. Examples: f16, f80.
17180     return false;
17181   case MVT::f32:
17182   case MVT::f64: {
17183     if (Subtarget.hasPrefixInstrs()) {
17184       // we can materialize all immediatess via XXSPLTI32DX and XXSPLTIDP.
17185       return true;
17186     }
17187     bool IsExact;
17188     APSInt IntResult(16, false);
17189     // The rounding mode doesn't really matter because we only care about floats
17190     // that can be converted to integers exactly.
17191     Imm.convertToInteger(IntResult, APFloat::rmTowardZero, &IsExact);
17192     // For exact values in the range [-16, 15] we can materialize the float.
17193     if (IsExact && IntResult <= 15 && IntResult >= -16)
17194       return true;
17195     return Imm.isZero();
17196   }
17197   case MVT::ppcf128:
17198     return Imm.isPosZero();
17199   }
17200 }
17201 
17202 // For vector shift operation op, fold
17203 // (op x, (and y, ((1 << numbits(x)) - 1))) -> (target op x, y)
17204 static SDValue stripModuloOnShift(const TargetLowering &TLI, SDNode *N,
17205                                   SelectionDAG &DAG) {
17206   SDValue N0 = N->getOperand(0);
17207   SDValue N1 = N->getOperand(1);
17208   EVT VT = N0.getValueType();
17209   unsigned OpSizeInBits = VT.getScalarSizeInBits();
17210   unsigned Opcode = N->getOpcode();
17211   unsigned TargetOpcode;
17212 
17213   switch (Opcode) {
17214   default:
17215     llvm_unreachable("Unexpected shift operation");
17216   case ISD::SHL:
17217     TargetOpcode = PPCISD::SHL;
17218     break;
17219   case ISD::SRL:
17220     TargetOpcode = PPCISD::SRL;
17221     break;
17222   case ISD::SRA:
17223     TargetOpcode = PPCISD::SRA;
17224     break;
17225   }
17226 
17227   if (VT.isVector() && TLI.isOperationLegal(Opcode, VT) &&
17228       N1->getOpcode() == ISD::AND)
17229     if (ConstantSDNode *Mask = isConstOrConstSplat(N1->getOperand(1)))
17230       if (Mask->getZExtValue() == OpSizeInBits - 1)
17231         return DAG.getNode(TargetOpcode, SDLoc(N), VT, N0, N1->getOperand(0));
17232 
17233   return SDValue();
17234 }
17235 
17236 SDValue PPCTargetLowering::combineSHL(SDNode *N, DAGCombinerInfo &DCI) const {
17237   if (auto Value = stripModuloOnShift(*this, N, DCI.DAG))
17238     return Value;
17239 
17240   SDValue N0 = N->getOperand(0);
17241   ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N->getOperand(1));
17242   if (!Subtarget.isISA3_0() || !Subtarget.isPPC64() ||
17243       N0.getOpcode() != ISD::SIGN_EXTEND ||
17244       N0.getOperand(0).getValueType() != MVT::i32 || CN1 == nullptr ||
17245       N->getValueType(0) != MVT::i64)
17246     return SDValue();
17247 
17248   // We can't save an operation here if the value is already extended, and
17249   // the existing shift is easier to combine.
17250   SDValue ExtsSrc = N0.getOperand(0);
17251   if (ExtsSrc.getOpcode() == ISD::TRUNCATE &&
17252       ExtsSrc.getOperand(0).getOpcode() == ISD::AssertSext)
17253     return SDValue();
17254 
17255   SDLoc DL(N0);
17256   SDValue ShiftBy = SDValue(CN1, 0);
17257   // We want the shift amount to be i32 on the extswli, but the shift could
17258   // have an i64.
17259   if (ShiftBy.getValueType() == MVT::i64)
17260     ShiftBy = DCI.DAG.getConstant(CN1->getZExtValue(), DL, MVT::i32);
17261 
17262   return DCI.DAG.getNode(PPCISD::EXTSWSLI, DL, MVT::i64, N0->getOperand(0),
17263                          ShiftBy);
17264 }
17265 
17266 SDValue PPCTargetLowering::combineSRA(SDNode *N, DAGCombinerInfo &DCI) const {
17267   if (auto Value = stripModuloOnShift(*this, N, DCI.DAG))
17268     return Value;
17269 
17270   return SDValue();
17271 }
17272 
17273 SDValue PPCTargetLowering::combineSRL(SDNode *N, DAGCombinerInfo &DCI) const {
17274   if (auto Value = stripModuloOnShift(*this, N, DCI.DAG))
17275     return Value;
17276 
17277   return SDValue();
17278 }
17279 
17280 // Transform (add X, (zext(setne Z, C))) -> (addze X, (addic (addi Z, -C), -1))
17281 // Transform (add X, (zext(sete  Z, C))) -> (addze X, (subfic (addi Z, -C), 0))
17282 // When C is zero, the equation (addi Z, -C) can be simplified to Z
17283 // Requirement: -C in [-32768, 32767], X and Z are MVT::i64 types
17284 static SDValue combineADDToADDZE(SDNode *N, SelectionDAG &DAG,
17285                                  const PPCSubtarget &Subtarget) {
17286   if (!Subtarget.isPPC64())
17287     return SDValue();
17288 
17289   SDValue LHS = N->getOperand(0);
17290   SDValue RHS = N->getOperand(1);
17291 
17292   auto isZextOfCompareWithConstant = [](SDValue Op) {
17293     if (Op.getOpcode() != ISD::ZERO_EXTEND || !Op.hasOneUse() ||
17294         Op.getValueType() != MVT::i64)
17295       return false;
17296 
17297     SDValue Cmp = Op.getOperand(0);
17298     if (Cmp.getOpcode() != ISD::SETCC || !Cmp.hasOneUse() ||
17299         Cmp.getOperand(0).getValueType() != MVT::i64)
17300       return false;
17301 
17302     if (auto *Constant = dyn_cast<ConstantSDNode>(Cmp.getOperand(1))) {
17303       int64_t NegConstant = 0 - Constant->getSExtValue();
17304       // Due to the limitations of the addi instruction,
17305       // -C is required to be [-32768, 32767].
17306       return isInt<16>(NegConstant);
17307     }
17308 
17309     return false;
17310   };
17311 
17312   bool LHSHasPattern = isZextOfCompareWithConstant(LHS);
17313   bool RHSHasPattern = isZextOfCompareWithConstant(RHS);
17314 
17315   // If there is a pattern, canonicalize a zext operand to the RHS.
17316   if (LHSHasPattern && !RHSHasPattern)
17317     std::swap(LHS, RHS);
17318   else if (!LHSHasPattern && !RHSHasPattern)
17319     return SDValue();
17320 
17321   SDLoc DL(N);
17322   SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Glue);
17323   SDValue Cmp = RHS.getOperand(0);
17324   SDValue Z = Cmp.getOperand(0);
17325   auto *Constant = cast<ConstantSDNode>(Cmp.getOperand(1));
17326   int64_t NegConstant = 0 - Constant->getSExtValue();
17327 
17328   switch(cast<CondCodeSDNode>(Cmp.getOperand(2))->get()) {
17329   default: break;
17330   case ISD::SETNE: {
17331     //                                 when C == 0
17332     //                             --> addze X, (addic Z, -1).carry
17333     //                            /
17334     // add X, (zext(setne Z, C))--
17335     //                            \    when -32768 <= -C <= 32767 && C != 0
17336     //                             --> addze X, (addic (addi Z, -C), -1).carry
17337     SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Z,
17338                               DAG.getConstant(NegConstant, DL, MVT::i64));
17339     SDValue AddOrZ = NegConstant != 0 ? Add : Z;
17340     SDValue Addc = DAG.getNode(ISD::ADDC, DL, DAG.getVTList(MVT::i64, MVT::Glue),
17341                                AddOrZ, DAG.getConstant(-1ULL, DL, MVT::i64));
17342     return DAG.getNode(ISD::ADDE, DL, VTs, LHS, DAG.getConstant(0, DL, MVT::i64),
17343                        SDValue(Addc.getNode(), 1));
17344     }
17345   case ISD::SETEQ: {
17346     //                                 when C == 0
17347     //                             --> addze X, (subfic Z, 0).carry
17348     //                            /
17349     // add X, (zext(sete  Z, C))--
17350     //                            \    when -32768 <= -C <= 32767 && C != 0
17351     //                             --> addze X, (subfic (addi Z, -C), 0).carry
17352     SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Z,
17353                               DAG.getConstant(NegConstant, DL, MVT::i64));
17354     SDValue AddOrZ = NegConstant != 0 ? Add : Z;
17355     SDValue Subc = DAG.getNode(ISD::SUBC, DL, DAG.getVTList(MVT::i64, MVT::Glue),
17356                                DAG.getConstant(0, DL, MVT::i64), AddOrZ);
17357     return DAG.getNode(ISD::ADDE, DL, VTs, LHS, DAG.getConstant(0, DL, MVT::i64),
17358                        SDValue(Subc.getNode(), 1));
17359     }
17360   }
17361 
17362   return SDValue();
17363 }
17364 
17365 // Transform
17366 // (add C1, (MAT_PCREL_ADDR GlobalAddr+C2)) to
17367 // (MAT_PCREL_ADDR GlobalAddr+(C1+C2))
17368 // In this case both C1 and C2 must be known constants.
17369 // C1+C2 must fit into a 34 bit signed integer.
17370 static SDValue combineADDToMAT_PCREL_ADDR(SDNode *N, SelectionDAG &DAG,
17371                                           const PPCSubtarget &Subtarget) {
17372   if (!Subtarget.isUsingPCRelativeCalls())
17373     return SDValue();
17374 
17375   // Check both Operand 0 and Operand 1 of the ADD node for the PCRel node.
17376   // If we find that node try to cast the Global Address and the Constant.
17377   SDValue LHS = N->getOperand(0);
17378   SDValue RHS = N->getOperand(1);
17379 
17380   if (LHS.getOpcode() != PPCISD::MAT_PCREL_ADDR)
17381     std::swap(LHS, RHS);
17382 
17383   if (LHS.getOpcode() != PPCISD::MAT_PCREL_ADDR)
17384     return SDValue();
17385 
17386   // Operand zero of PPCISD::MAT_PCREL_ADDR is the GA node.
17387   GlobalAddressSDNode *GSDN = dyn_cast<GlobalAddressSDNode>(LHS.getOperand(0));
17388   ConstantSDNode* ConstNode = dyn_cast<ConstantSDNode>(RHS);
17389 
17390   // Check that both casts succeeded.
17391   if (!GSDN || !ConstNode)
17392     return SDValue();
17393 
17394   int64_t NewOffset = GSDN->getOffset() + ConstNode->getSExtValue();
17395   SDLoc DL(GSDN);
17396 
17397   // The signed int offset needs to fit in 34 bits.
17398   if (!isInt<34>(NewOffset))
17399     return SDValue();
17400 
17401   // The new global address is a copy of the old global address except
17402   // that it has the updated Offset.
17403   SDValue GA =
17404       DAG.getTargetGlobalAddress(GSDN->getGlobal(), DL, GSDN->getValueType(0),
17405                                  NewOffset, GSDN->getTargetFlags());
17406   SDValue MatPCRel =
17407       DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, GSDN->getValueType(0), GA);
17408   return MatPCRel;
17409 }
17410 
17411 SDValue PPCTargetLowering::combineADD(SDNode *N, DAGCombinerInfo &DCI) const {
17412   if (auto Value = combineADDToADDZE(N, DCI.DAG, Subtarget))
17413     return Value;
17414 
17415   if (auto Value = combineADDToMAT_PCREL_ADDR(N, DCI.DAG, Subtarget))
17416     return Value;
17417 
17418   return SDValue();
17419 }
17420 
17421 // Detect TRUNCATE operations on bitcasts of float128 values.
17422 // What we are looking for here is the situtation where we extract a subset
17423 // of bits from a 128 bit float.
17424 // This can be of two forms:
17425 // 1) BITCAST of f128 feeding TRUNCATE
17426 // 2) BITCAST of f128 feeding SRL (a shift) feeding TRUNCATE
17427 // The reason this is required is because we do not have a legal i128 type
17428 // and so we want to prevent having to store the f128 and then reload part
17429 // of it.
17430 SDValue PPCTargetLowering::combineTRUNCATE(SDNode *N,
17431                                            DAGCombinerInfo &DCI) const {
17432   // If we are using CRBits then try that first.
17433   if (Subtarget.useCRBits()) {
17434     // Check if CRBits did anything and return that if it did.
17435     if (SDValue CRTruncValue = DAGCombineTruncBoolExt(N, DCI))
17436       return CRTruncValue;
17437   }
17438 
17439   SDLoc dl(N);
17440   SDValue Op0 = N->getOperand(0);
17441 
17442   // Looking for a truncate of i128 to i64.
17443   if (Op0.getValueType() != MVT::i128 || N->getValueType(0) != MVT::i64)
17444     return SDValue();
17445 
17446   int EltToExtract = DCI.DAG.getDataLayout().isBigEndian() ? 1 : 0;
17447 
17448   // SRL feeding TRUNCATE.
17449   if (Op0.getOpcode() == ISD::SRL) {
17450     ConstantSDNode *ConstNode = dyn_cast<ConstantSDNode>(Op0.getOperand(1));
17451     // The right shift has to be by 64 bits.
17452     if (!ConstNode || ConstNode->getZExtValue() != 64)
17453       return SDValue();
17454 
17455     // Switch the element number to extract.
17456     EltToExtract = EltToExtract ? 0 : 1;
17457     // Update Op0 past the SRL.
17458     Op0 = Op0.getOperand(0);
17459   }
17460 
17461   // BITCAST feeding a TRUNCATE possibly via SRL.
17462   if (Op0.getOpcode() == ISD::BITCAST &&
17463       Op0.getValueType() == MVT::i128 &&
17464       Op0.getOperand(0).getValueType() == MVT::f128) {
17465     SDValue Bitcast = DCI.DAG.getBitcast(MVT::v2i64, Op0.getOperand(0));
17466     return DCI.DAG.getNode(
17467         ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Bitcast,
17468         DCI.DAG.getTargetConstant(EltToExtract, dl, MVT::i32));
17469   }
17470   return SDValue();
17471 }
17472 
17473 SDValue PPCTargetLowering::combineMUL(SDNode *N, DAGCombinerInfo &DCI) const {
17474   SelectionDAG &DAG = DCI.DAG;
17475 
17476   ConstantSDNode *ConstOpOrElement = isConstOrConstSplat(N->getOperand(1));
17477   if (!ConstOpOrElement)
17478     return SDValue();
17479 
17480   // An imul is usually smaller than the alternative sequence for legal type.
17481   if (DAG.getMachineFunction().getFunction().hasMinSize() &&
17482       isOperationLegal(ISD::MUL, N->getValueType(0)))
17483     return SDValue();
17484 
17485   auto IsProfitable = [this](bool IsNeg, bool IsAddOne, EVT VT) -> bool {
17486     switch (this->Subtarget.getCPUDirective()) {
17487     default:
17488       // TODO: enhance the condition for subtarget before pwr8
17489       return false;
17490     case PPC::DIR_PWR8:
17491       //  type        mul     add    shl
17492       // scalar        4       1      1
17493       // vector        7       2      2
17494       return true;
17495     case PPC::DIR_PWR9:
17496     case PPC::DIR_PWR10:
17497     case PPC::DIR_PWR_FUTURE:
17498       //  type        mul     add    shl
17499       // scalar        5       2      2
17500       // vector        7       2      2
17501 
17502       // The cycle RATIO of related operations are showed as a table above.
17503       // Because mul is 5(scalar)/7(vector), add/sub/shl are all 2 for both
17504       // scalar and vector type. For 2 instrs patterns, add/sub + shl
17505       // are 4, it is always profitable; but for 3 instrs patterns
17506       // (mul x, -(2^N + 1)) => -(add (shl x, N), x), sub + add + shl are 6.
17507       // So we should only do it for vector type.
17508       return IsAddOne && IsNeg ? VT.isVector() : true;
17509     }
17510   };
17511 
17512   EVT VT = N->getValueType(0);
17513   SDLoc DL(N);
17514 
17515   const APInt &MulAmt = ConstOpOrElement->getAPIntValue();
17516   bool IsNeg = MulAmt.isNegative();
17517   APInt MulAmtAbs = MulAmt.abs();
17518 
17519   if ((MulAmtAbs - 1).isPowerOf2()) {
17520     // (mul x, 2^N + 1) => (add (shl x, N), x)
17521     // (mul x, -(2^N + 1)) => -(add (shl x, N), x)
17522 
17523     if (!IsProfitable(IsNeg, true, VT))
17524       return SDValue();
17525 
17526     SDValue Op0 = N->getOperand(0);
17527     SDValue Op1 =
17528         DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
17529                     DAG.getConstant((MulAmtAbs - 1).logBase2(), DL, VT));
17530     SDValue Res = DAG.getNode(ISD::ADD, DL, VT, Op0, Op1);
17531 
17532     if (!IsNeg)
17533       return Res;
17534 
17535     return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Res);
17536   } else if ((MulAmtAbs + 1).isPowerOf2()) {
17537     // (mul x, 2^N - 1) => (sub (shl x, N), x)
17538     // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
17539 
17540     if (!IsProfitable(IsNeg, false, VT))
17541       return SDValue();
17542 
17543     SDValue Op0 = N->getOperand(0);
17544     SDValue Op1 =
17545         DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
17546                     DAG.getConstant((MulAmtAbs + 1).logBase2(), DL, VT));
17547 
17548     if (!IsNeg)
17549       return DAG.getNode(ISD::SUB, DL, VT, Op1, Op0);
17550     else
17551       return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1);
17552 
17553   } else {
17554     return SDValue();
17555   }
17556 }
17557 
17558 // Combine fma-like op (like fnmsub) with fnegs to appropriate op. Do this
17559 // in combiner since we need to check SD flags and other subtarget features.
17560 SDValue PPCTargetLowering::combineFMALike(SDNode *N,
17561                                           DAGCombinerInfo &DCI) const {
17562   SDValue N0 = N->getOperand(0);
17563   SDValue N1 = N->getOperand(1);
17564   SDValue N2 = N->getOperand(2);
17565   SDNodeFlags Flags = N->getFlags();
17566   EVT VT = N->getValueType(0);
17567   SelectionDAG &DAG = DCI.DAG;
17568   const TargetOptions &Options = getTargetMachine().Options;
17569   unsigned Opc = N->getOpcode();
17570   bool CodeSize = DAG.getMachineFunction().getFunction().hasOptSize();
17571   bool LegalOps = !DCI.isBeforeLegalizeOps();
17572   SDLoc Loc(N);
17573 
17574   if (!isOperationLegal(ISD::FMA, VT))
17575     return SDValue();
17576 
17577   // Allowing transformation to FNMSUB may change sign of zeroes when ab-c=0
17578   // since (fnmsub a b c)=-0 while c-ab=+0.
17579   if (!Flags.hasNoSignedZeros() && !Options.NoSignedZerosFPMath)
17580     return SDValue();
17581 
17582   // (fma (fneg a) b c) => (fnmsub a b c)
17583   // (fnmsub (fneg a) b c) => (fma a b c)
17584   if (SDValue NegN0 = getCheaperNegatedExpression(N0, DAG, LegalOps, CodeSize))
17585     return DAG.getNode(invertFMAOpcode(Opc), Loc, VT, NegN0, N1, N2, Flags);
17586 
17587   // (fma a (fneg b) c) => (fnmsub a b c)
17588   // (fnmsub a (fneg b) c) => (fma a b c)
17589   if (SDValue NegN1 = getCheaperNegatedExpression(N1, DAG, LegalOps, CodeSize))
17590     return DAG.getNode(invertFMAOpcode(Opc), Loc, VT, N0, NegN1, N2, Flags);
17591 
17592   return SDValue();
17593 }
17594 
17595 bool PPCTargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const {
17596   // Only duplicate to increase tail-calls for the 64bit SysV ABIs.
17597   if (!Subtarget.is64BitELFABI())
17598     return false;
17599 
17600   // If not a tail call then no need to proceed.
17601   if (!CI->isTailCall())
17602     return false;
17603 
17604   // If sibling calls have been disabled and tail-calls aren't guaranteed
17605   // there is no reason to duplicate.
17606   auto &TM = getTargetMachine();
17607   if (!TM.Options.GuaranteedTailCallOpt && DisableSCO)
17608     return false;
17609 
17610   // Can't tail call a function called indirectly, or if it has variadic args.
17611   const Function *Callee = CI->getCalledFunction();
17612   if (!Callee || Callee->isVarArg())
17613     return false;
17614 
17615   // Make sure the callee and caller calling conventions are eligible for tco.
17616   const Function *Caller = CI->getParent()->getParent();
17617   if (!areCallingConvEligibleForTCO_64SVR4(Caller->getCallingConv(),
17618                                            CI->getCallingConv()))
17619       return false;
17620 
17621   // If the function is local then we have a good chance at tail-calling it
17622   return getTargetMachine().shouldAssumeDSOLocal(*Caller->getParent(), Callee);
17623 }
17624 
17625 bool PPCTargetLowering::hasBitPreservingFPLogic(EVT VT) const {
17626   if (!Subtarget.hasVSX())
17627     return false;
17628   if (Subtarget.hasP9Vector() && VT == MVT::f128)
17629     return true;
17630   return VT == MVT::f32 || VT == MVT::f64 ||
17631     VT == MVT::v4f32 || VT == MVT::v2f64;
17632 }
17633 
17634 bool PPCTargetLowering::
17635 isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const {
17636   const Value *Mask = AndI.getOperand(1);
17637   // If the mask is suitable for andi. or andis. we should sink the and.
17638   if (const ConstantInt *CI = dyn_cast<ConstantInt>(Mask)) {
17639     // Can't handle constants wider than 64-bits.
17640     if (CI->getBitWidth() > 64)
17641       return false;
17642     int64_t ConstVal = CI->getZExtValue();
17643     return isUInt<16>(ConstVal) ||
17644       (isUInt<16>(ConstVal >> 16) && !(ConstVal & 0xFFFF));
17645   }
17646 
17647   // For non-constant masks, we can always use the record-form and.
17648   return true;
17649 }
17650 
17651 // For type v4i32/v8ii16/v16i8, transform
17652 // from (vselect (setcc a, b, setugt), (sub a, b), (sub b, a)) to (abdu a, b)
17653 // from (vselect (setcc a, b, setuge), (sub a, b), (sub b, a)) to (abdu a, b)
17654 // from (vselect (setcc a, b, setult), (sub b, a), (sub a, b)) to (abdu a, b)
17655 // from (vselect (setcc a, b, setule), (sub b, a), (sub a, b)) to (abdu a, b)
17656 // TODO: Move this to DAGCombiner?
17657 SDValue PPCTargetLowering::combineVSelect(SDNode *N,
17658                                           DAGCombinerInfo &DCI) const {
17659   assert((N->getOpcode() == ISD::VSELECT) && "Need VSELECT node here");
17660   assert(Subtarget.hasP9Altivec() &&
17661          "Only combine this when P9 altivec supported!");
17662 
17663   SelectionDAG &DAG = DCI.DAG;
17664   SDLoc dl(N);
17665   SDValue Cond = N->getOperand(0);
17666   SDValue TrueOpnd = N->getOperand(1);
17667   SDValue FalseOpnd = N->getOperand(2);
17668   EVT VT = N->getOperand(1).getValueType();
17669 
17670   if (Cond.getOpcode() != ISD::SETCC || TrueOpnd.getOpcode() != ISD::SUB ||
17671       FalseOpnd.getOpcode() != ISD::SUB)
17672     return SDValue();
17673 
17674   // ABSD only available for type v4i32/v8i16/v16i8
17675   if (VT != MVT::v4i32 && VT != MVT::v8i16 && VT != MVT::v16i8)
17676     return SDValue();
17677 
17678   // At least to save one more dependent computation
17679   if (!(Cond.hasOneUse() || TrueOpnd.hasOneUse() || FalseOpnd.hasOneUse()))
17680     return SDValue();
17681 
17682   ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
17683 
17684   // Can only handle unsigned comparison here
17685   switch (CC) {
17686   default:
17687     return SDValue();
17688   case ISD::SETUGT:
17689   case ISD::SETUGE:
17690     break;
17691   case ISD::SETULT:
17692   case ISD::SETULE:
17693     std::swap(TrueOpnd, FalseOpnd);
17694     break;
17695   }
17696 
17697   SDValue CmpOpnd1 = Cond.getOperand(0);
17698   SDValue CmpOpnd2 = Cond.getOperand(1);
17699 
17700   // SETCC CmpOpnd1 CmpOpnd2 cond
17701   // TrueOpnd = CmpOpnd1 - CmpOpnd2
17702   // FalseOpnd = CmpOpnd2 - CmpOpnd1
17703   if (TrueOpnd.getOperand(0) == CmpOpnd1 &&
17704       TrueOpnd.getOperand(1) == CmpOpnd2 &&
17705       FalseOpnd.getOperand(0) == CmpOpnd2 &&
17706       FalseOpnd.getOperand(1) == CmpOpnd1) {
17707     return DAG.getNode(ISD::ABDU, dl, N->getOperand(1).getValueType(), CmpOpnd1,
17708                        CmpOpnd2, DAG.getTargetConstant(0, dl, MVT::i32));
17709   }
17710 
17711   return SDValue();
17712 }
17713 
17714 /// getAddrModeForFlags - Based on the set of address flags, select the most
17715 /// optimal instruction format to match by.
17716 PPC::AddrMode PPCTargetLowering::getAddrModeForFlags(unsigned Flags) const {
17717   // This is not a node we should be handling here.
17718   if (Flags == PPC::MOF_None)
17719     return PPC::AM_None;
17720   // Unaligned D-Forms are tried first, followed by the aligned D-Forms.
17721   for (auto FlagSet : AddrModesMap.at(PPC::AM_DForm))
17722     if ((Flags & FlagSet) == FlagSet)
17723       return PPC::AM_DForm;
17724   for (auto FlagSet : AddrModesMap.at(PPC::AM_DSForm))
17725     if ((Flags & FlagSet) == FlagSet)
17726       return PPC::AM_DSForm;
17727   for (auto FlagSet : AddrModesMap.at(PPC::AM_DQForm))
17728     if ((Flags & FlagSet) == FlagSet)
17729       return PPC::AM_DQForm;
17730   for (auto FlagSet : AddrModesMap.at(PPC::AM_PrefixDForm))
17731     if ((Flags & FlagSet) == FlagSet)
17732       return PPC::AM_PrefixDForm;
17733   // If no other forms are selected, return an X-Form as it is the most
17734   // general addressing mode.
17735   return PPC::AM_XForm;
17736 }
17737 
17738 /// Set alignment flags based on whether or not the Frame Index is aligned.
17739 /// Utilized when computing flags for address computation when selecting
17740 /// load and store instructions.
17741 static void setAlignFlagsForFI(SDValue N, unsigned &FlagSet,
17742                                SelectionDAG &DAG) {
17743   bool IsAdd = ((N.getOpcode() == ISD::ADD) || (N.getOpcode() == ISD::OR));
17744   FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(IsAdd ? N.getOperand(0) : N);
17745   if (!FI)
17746     return;
17747   const MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
17748   unsigned FrameIndexAlign = MFI.getObjectAlign(FI->getIndex()).value();
17749   // If this is (add $FI, $S16Imm), the alignment flags are already set
17750   // based on the immediate. We just need to clear the alignment flags
17751   // if the FI alignment is weaker.
17752   if ((FrameIndexAlign % 4) != 0)
17753     FlagSet &= ~PPC::MOF_RPlusSImm16Mult4;
17754   if ((FrameIndexAlign % 16) != 0)
17755     FlagSet &= ~PPC::MOF_RPlusSImm16Mult16;
17756   // If the address is a plain FrameIndex, set alignment flags based on
17757   // FI alignment.
17758   if (!IsAdd) {
17759     if ((FrameIndexAlign % 4) == 0)
17760       FlagSet |= PPC::MOF_RPlusSImm16Mult4;
17761     if ((FrameIndexAlign % 16) == 0)
17762       FlagSet |= PPC::MOF_RPlusSImm16Mult16;
17763   }
17764 }
17765 
17766 /// Given a node, compute flags that are used for address computation when
17767 /// selecting load and store instructions. The flags computed are stored in
17768 /// FlagSet. This function takes into account whether the node is a constant,
17769 /// an ADD, OR, or a constant, and computes the address flags accordingly.
17770 static void computeFlagsForAddressComputation(SDValue N, unsigned &FlagSet,
17771                                               SelectionDAG &DAG) {
17772   // Set the alignment flags for the node depending on if the node is
17773   // 4-byte or 16-byte aligned.
17774   auto SetAlignFlagsForImm = [&](uint64_t Imm) {
17775     if ((Imm & 0x3) == 0)
17776       FlagSet |= PPC::MOF_RPlusSImm16Mult4;
17777     if ((Imm & 0xf) == 0)
17778       FlagSet |= PPC::MOF_RPlusSImm16Mult16;
17779   };
17780 
17781   if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
17782     // All 32-bit constants can be computed as LIS + Disp.
17783     const APInt &ConstImm = CN->getAPIntValue();
17784     if (ConstImm.isSignedIntN(32)) { // Flag to handle 32-bit constants.
17785       FlagSet |= PPC::MOF_AddrIsSImm32;
17786       SetAlignFlagsForImm(ConstImm.getZExtValue());
17787       setAlignFlagsForFI(N, FlagSet, DAG);
17788     }
17789     if (ConstImm.isSignedIntN(34)) // Flag to handle 34-bit constants.
17790       FlagSet |= PPC::MOF_RPlusSImm34;
17791     else // Let constant materialization handle large constants.
17792       FlagSet |= PPC::MOF_NotAddNorCst;
17793   } else if (N.getOpcode() == ISD::ADD || provablyDisjointOr(DAG, N)) {
17794     // This address can be represented as an addition of:
17795     // - Register + Imm16 (possibly a multiple of 4/16)
17796     // - Register + Imm34
17797     // - Register + PPCISD::Lo
17798     // - Register + Register
17799     // In any case, we won't have to match this as Base + Zero.
17800     SDValue RHS = N.getOperand(1);
17801     if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(RHS)) {
17802       const APInt &ConstImm = CN->getAPIntValue();
17803       if (ConstImm.isSignedIntN(16)) {
17804         FlagSet |= PPC::MOF_RPlusSImm16; // Signed 16-bit immediates.
17805         SetAlignFlagsForImm(ConstImm.getZExtValue());
17806         setAlignFlagsForFI(N, FlagSet, DAG);
17807       }
17808       if (ConstImm.isSignedIntN(34))
17809         FlagSet |= PPC::MOF_RPlusSImm34; // Signed 34-bit immediates.
17810       else
17811         FlagSet |= PPC::MOF_RPlusR; // Register.
17812     } else if (RHS.getOpcode() == PPCISD::Lo &&
17813                !cast<ConstantSDNode>(RHS.getOperand(1))->getZExtValue())
17814       FlagSet |= PPC::MOF_RPlusLo; // PPCISD::Lo.
17815     else
17816       FlagSet |= PPC::MOF_RPlusR;
17817   } else { // The address computation is not a constant or an addition.
17818     setAlignFlagsForFI(N, FlagSet, DAG);
17819     FlagSet |= PPC::MOF_NotAddNorCst;
17820   }
17821 }
17822 
17823 static bool isPCRelNode(SDValue N) {
17824   return (N.getOpcode() == PPCISD::MAT_PCREL_ADDR ||
17825       isValidPCRelNode<ConstantPoolSDNode>(N) ||
17826       isValidPCRelNode<GlobalAddressSDNode>(N) ||
17827       isValidPCRelNode<JumpTableSDNode>(N) ||
17828       isValidPCRelNode<BlockAddressSDNode>(N));
17829 }
17830 
17831 /// computeMOFlags - Given a node N and it's Parent (a MemSDNode), compute
17832 /// the address flags of the load/store instruction that is to be matched.
17833 unsigned PPCTargetLowering::computeMOFlags(const SDNode *Parent, SDValue N,
17834                                            SelectionDAG &DAG) const {
17835   unsigned FlagSet = PPC::MOF_None;
17836 
17837   // Compute subtarget flags.
17838   if (!Subtarget.hasP9Vector())
17839     FlagSet |= PPC::MOF_SubtargetBeforeP9;
17840   else {
17841     FlagSet |= PPC::MOF_SubtargetP9;
17842     if (Subtarget.hasPrefixInstrs())
17843       FlagSet |= PPC::MOF_SubtargetP10;
17844   }
17845   if (Subtarget.hasSPE())
17846     FlagSet |= PPC::MOF_SubtargetSPE;
17847 
17848   // Check if we have a PCRel node and return early.
17849   if ((FlagSet & PPC::MOF_SubtargetP10) && isPCRelNode(N))
17850     return FlagSet;
17851 
17852   // If the node is the paired load/store intrinsics, compute flags for
17853   // address computation and return early.
17854   unsigned ParentOp = Parent->getOpcode();
17855   if (Subtarget.isISA3_1() && ((ParentOp == ISD::INTRINSIC_W_CHAIN) ||
17856                                (ParentOp == ISD::INTRINSIC_VOID))) {
17857     unsigned ID = cast<ConstantSDNode>(Parent->getOperand(1))->getZExtValue();
17858     if ((ID == Intrinsic::ppc_vsx_lxvp) || (ID == Intrinsic::ppc_vsx_stxvp)) {
17859       SDValue IntrinOp = (ID == Intrinsic::ppc_vsx_lxvp)
17860                              ? Parent->getOperand(2)
17861                              : Parent->getOperand(3);
17862       computeFlagsForAddressComputation(IntrinOp, FlagSet, DAG);
17863       FlagSet |= PPC::MOF_Vector;
17864       return FlagSet;
17865     }
17866   }
17867 
17868   // Mark this as something we don't want to handle here if it is atomic
17869   // or pre-increment instruction.
17870   if (const LSBaseSDNode *LSB = dyn_cast<LSBaseSDNode>(Parent))
17871     if (LSB->isIndexed())
17872       return PPC::MOF_None;
17873 
17874   // Compute in-memory type flags. This is based on if there are scalars,
17875   // floats or vectors.
17876   const MemSDNode *MN = dyn_cast<MemSDNode>(Parent);
17877   assert(MN && "Parent should be a MemSDNode!");
17878   EVT MemVT = MN->getMemoryVT();
17879   unsigned Size = MemVT.getSizeInBits();
17880   if (MemVT.isScalarInteger()) {
17881     assert(Size <= 128 &&
17882            "Not expecting scalar integers larger than 16 bytes!");
17883     if (Size < 32)
17884       FlagSet |= PPC::MOF_SubWordInt;
17885     else if (Size == 32)
17886       FlagSet |= PPC::MOF_WordInt;
17887     else
17888       FlagSet |= PPC::MOF_DoubleWordInt;
17889   } else if (MemVT.isVector() && !MemVT.isFloatingPoint()) { // Integer vectors.
17890     if (Size == 128)
17891       FlagSet |= PPC::MOF_Vector;
17892     else if (Size == 256) {
17893       assert(Subtarget.pairedVectorMemops() &&
17894              "256-bit vectors are only available when paired vector memops is "
17895              "enabled!");
17896       FlagSet |= PPC::MOF_Vector;
17897     } else
17898       llvm_unreachable("Not expecting illegal vectors!");
17899   } else { // Floating point type: can be scalar, f128 or vector types.
17900     if (Size == 32 || Size == 64)
17901       FlagSet |= PPC::MOF_ScalarFloat;
17902     else if (MemVT == MVT::f128 || MemVT.isVector())
17903       FlagSet |= PPC::MOF_Vector;
17904     else
17905       llvm_unreachable("Not expecting illegal scalar floats!");
17906   }
17907 
17908   // Compute flags for address computation.
17909   computeFlagsForAddressComputation(N, FlagSet, DAG);
17910 
17911   // Compute type extension flags.
17912   if (const LoadSDNode *LN = dyn_cast<LoadSDNode>(Parent)) {
17913     switch (LN->getExtensionType()) {
17914     case ISD::SEXTLOAD:
17915       FlagSet |= PPC::MOF_SExt;
17916       break;
17917     case ISD::EXTLOAD:
17918     case ISD::ZEXTLOAD:
17919       FlagSet |= PPC::MOF_ZExt;
17920       break;
17921     case ISD::NON_EXTLOAD:
17922       FlagSet |= PPC::MOF_NoExt;
17923       break;
17924     }
17925   } else
17926     FlagSet |= PPC::MOF_NoExt;
17927 
17928   // For integers, no extension is the same as zero extension.
17929   // We set the extension mode to zero extension so we don't have
17930   // to add separate entries in AddrModesMap for loads and stores.
17931   if (MemVT.isScalarInteger() && (FlagSet & PPC::MOF_NoExt)) {
17932     FlagSet |= PPC::MOF_ZExt;
17933     FlagSet &= ~PPC::MOF_NoExt;
17934   }
17935 
17936   // If we don't have prefixed instructions, 34-bit constants should be
17937   // treated as PPC::MOF_NotAddNorCst so they can match D-Forms.
17938   bool IsNonP1034BitConst =
17939       ((PPC::MOF_RPlusSImm34 | PPC::MOF_AddrIsSImm32 | PPC::MOF_SubtargetP10) &
17940        FlagSet) == PPC::MOF_RPlusSImm34;
17941   if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::OR &&
17942       IsNonP1034BitConst)
17943     FlagSet |= PPC::MOF_NotAddNorCst;
17944 
17945   return FlagSet;
17946 }
17947 
17948 /// SelectForceXFormMode - Given the specified address, force it to be
17949 /// represented as an indexed [r+r] operation (an XForm instruction).
17950 PPC::AddrMode PPCTargetLowering::SelectForceXFormMode(SDValue N, SDValue &Disp,
17951                                                       SDValue &Base,
17952                                                       SelectionDAG &DAG) const {
17953 
17954   PPC::AddrMode Mode = PPC::AM_XForm;
17955   int16_t ForceXFormImm = 0;
17956   if (provablyDisjointOr(DAG, N) &&
17957       !isIntS16Immediate(N.getOperand(1), ForceXFormImm)) {
17958     Disp = N.getOperand(0);
17959     Base = N.getOperand(1);
17960     return Mode;
17961   }
17962 
17963   // If the address is the result of an add, we will utilize the fact that the
17964   // address calculation includes an implicit add.  However, we can reduce
17965   // register pressure if we do not materialize a constant just for use as the
17966   // index register.  We only get rid of the add if it is not an add of a
17967   // value and a 16-bit signed constant and both have a single use.
17968   if (N.getOpcode() == ISD::ADD &&
17969       (!isIntS16Immediate(N.getOperand(1), ForceXFormImm) ||
17970        !N.getOperand(1).hasOneUse() || !N.getOperand(0).hasOneUse())) {
17971     Disp = N.getOperand(0);
17972     Base = N.getOperand(1);
17973     return Mode;
17974   }
17975 
17976   // Otherwise, use R0 as the base register.
17977   Disp = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO,
17978                          N.getValueType());
17979   Base = N;
17980 
17981   return Mode;
17982 }
17983 
17984 bool PPCTargetLowering::splitValueIntoRegisterParts(
17985     SelectionDAG &DAG, const SDLoc &DL, SDValue Val, SDValue *Parts,
17986     unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC) const {
17987   EVT ValVT = Val.getValueType();
17988   // If we are splitting a scalar integer into f64 parts (i.e. so they
17989   // can be placed into VFRC registers), we need to zero extend and
17990   // bitcast the values. This will ensure the value is placed into a
17991   // VSR using direct moves or stack operations as needed.
17992   if (PartVT == MVT::f64 &&
17993       (ValVT == MVT::i32 || ValVT == MVT::i16 || ValVT == MVT::i8)) {
17994     Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val);
17995     Val = DAG.getNode(ISD::BITCAST, DL, MVT::f64, Val);
17996     Parts[0] = Val;
17997     return true;
17998   }
17999   return false;
18000 }
18001 
18002 SDValue PPCTargetLowering::lowerToLibCall(const char *LibCallName, SDValue Op,
18003                                           SelectionDAG &DAG) const {
18004   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18005   TargetLowering::CallLoweringInfo CLI(DAG);
18006   EVT RetVT = Op.getValueType();
18007   Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
18008   SDValue Callee =
18009       DAG.getExternalSymbol(LibCallName, TLI.getPointerTy(DAG.getDataLayout()));
18010   bool SignExtend = TLI.shouldSignExtendTypeInLibCall(RetVT, false);
18011   TargetLowering::ArgListTy Args;
18012   TargetLowering::ArgListEntry Entry;
18013   for (const SDValue &N : Op->op_values()) {
18014     EVT ArgVT = N.getValueType();
18015     Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18016     Entry.Node = N;
18017     Entry.Ty = ArgTy;
18018     Entry.IsSExt = TLI.shouldSignExtendTypeInLibCall(ArgVT, SignExtend);
18019     Entry.IsZExt = !Entry.IsSExt;
18020     Args.push_back(Entry);
18021   }
18022 
18023   SDValue InChain = DAG.getEntryNode();
18024   SDValue TCChain = InChain;
18025   const Function &F = DAG.getMachineFunction().getFunction();
18026   bool isTailCall =
18027       TLI.isInTailCallPosition(DAG, Op.getNode(), TCChain) &&
18028       (RetTy == F.getReturnType() || F.getReturnType()->isVoidTy());
18029   if (isTailCall)
18030     InChain = TCChain;
18031   CLI.setDebugLoc(SDLoc(Op))
18032       .setChain(InChain)
18033       .setLibCallee(CallingConv::C, RetTy, Callee, std::move(Args))
18034       .setTailCall(isTailCall)
18035       .setSExtResult(SignExtend)
18036       .setZExtResult(!SignExtend)
18037       .setIsPostTypeLegalization(true);
18038   return TLI.LowerCallTo(CLI).first;
18039 }
18040 
18041 SDValue PPCTargetLowering::lowerLibCallBasedOnType(
18042     const char *LibCallFloatName, const char *LibCallDoubleName, SDValue Op,
18043     SelectionDAG &DAG) const {
18044   if (Op.getValueType() == MVT::f32)
18045     return lowerToLibCall(LibCallFloatName, Op, DAG);
18046 
18047   if (Op.getValueType() == MVT::f64)
18048     return lowerToLibCall(LibCallDoubleName, Op, DAG);
18049 
18050   return SDValue();
18051 }
18052 
18053 bool PPCTargetLowering::isLowringToMASSFiniteSafe(SDValue Op) const {
18054   SDNodeFlags Flags = Op.getNode()->getFlags();
18055   return isLowringToMASSSafe(Op) && Flags.hasNoSignedZeros() &&
18056          Flags.hasNoNaNs() && Flags.hasNoInfs();
18057 }
18058 
18059 bool PPCTargetLowering::isLowringToMASSSafe(SDValue Op) const {
18060   return Op.getNode()->getFlags().hasApproximateFuncs();
18061 }
18062 
18063 bool PPCTargetLowering::isScalarMASSConversionEnabled() const {
18064   return getTargetMachine().Options.PPCGenScalarMASSEntries;
18065 }
18066 
18067 SDValue PPCTargetLowering::lowerLibCallBase(const char *LibCallDoubleName,
18068                                             const char *LibCallFloatName,
18069                                             const char *LibCallDoubleNameFinite,
18070                                             const char *LibCallFloatNameFinite,
18071                                             SDValue Op,
18072                                             SelectionDAG &DAG) const {
18073   if (!isScalarMASSConversionEnabled() || !isLowringToMASSSafe(Op))
18074     return SDValue();
18075 
18076   if (!isLowringToMASSFiniteSafe(Op))
18077     return lowerLibCallBasedOnType(LibCallFloatName, LibCallDoubleName, Op,
18078                                    DAG);
18079 
18080   return lowerLibCallBasedOnType(LibCallFloatNameFinite,
18081                                  LibCallDoubleNameFinite, Op, DAG);
18082 }
18083 
18084 SDValue PPCTargetLowering::lowerPow(SDValue Op, SelectionDAG &DAG) const {
18085   return lowerLibCallBase("__xl_pow", "__xl_powf", "__xl_pow_finite",
18086                           "__xl_powf_finite", Op, DAG);
18087 }
18088 
18089 SDValue PPCTargetLowering::lowerSin(SDValue Op, SelectionDAG &DAG) const {
18090   return lowerLibCallBase("__xl_sin", "__xl_sinf", "__xl_sin_finite",
18091                           "__xl_sinf_finite", Op, DAG);
18092 }
18093 
18094 SDValue PPCTargetLowering::lowerCos(SDValue Op, SelectionDAG &DAG) const {
18095   return lowerLibCallBase("__xl_cos", "__xl_cosf", "__xl_cos_finite",
18096                           "__xl_cosf_finite", Op, DAG);
18097 }
18098 
18099 SDValue PPCTargetLowering::lowerLog(SDValue Op, SelectionDAG &DAG) const {
18100   return lowerLibCallBase("__xl_log", "__xl_logf", "__xl_log_finite",
18101                           "__xl_logf_finite", Op, DAG);
18102 }
18103 
18104 SDValue PPCTargetLowering::lowerLog10(SDValue Op, SelectionDAG &DAG) const {
18105   return lowerLibCallBase("__xl_log10", "__xl_log10f", "__xl_log10_finite",
18106                           "__xl_log10f_finite", Op, DAG);
18107 }
18108 
18109 SDValue PPCTargetLowering::lowerExp(SDValue Op, SelectionDAG &DAG) const {
18110   return lowerLibCallBase("__xl_exp", "__xl_expf", "__xl_exp_finite",
18111                           "__xl_expf_finite", Op, DAG);
18112 }
18113 
18114 // If we happen to match to an aligned D-Form, check if the Frame Index is
18115 // adequately aligned. If it is not, reset the mode to match to X-Form.
18116 static void setXFormForUnalignedFI(SDValue N, unsigned Flags,
18117                                    PPC::AddrMode &Mode) {
18118   if (!isa<FrameIndexSDNode>(N))
18119     return;
18120   if ((Mode == PPC::AM_DSForm && !(Flags & PPC::MOF_RPlusSImm16Mult4)) ||
18121       (Mode == PPC::AM_DQForm && !(Flags & PPC::MOF_RPlusSImm16Mult16)))
18122     Mode = PPC::AM_XForm;
18123 }
18124 
18125 /// SelectOptimalAddrMode - Based on a node N and it's Parent (a MemSDNode),
18126 /// compute the address flags of the node, get the optimal address mode based
18127 /// on the flags, and set the Base and Disp based on the address mode.
18128 PPC::AddrMode PPCTargetLowering::SelectOptimalAddrMode(const SDNode *Parent,
18129                                                        SDValue N, SDValue &Disp,
18130                                                        SDValue &Base,
18131                                                        SelectionDAG &DAG,
18132                                                        MaybeAlign Align) const {
18133   SDLoc DL(Parent);
18134 
18135   // Compute the address flags.
18136   unsigned Flags = computeMOFlags(Parent, N, DAG);
18137 
18138   // Get the optimal address mode based on the Flags.
18139   PPC::AddrMode Mode = getAddrModeForFlags(Flags);
18140 
18141   // If the address mode is DS-Form or DQ-Form, check if the FI is aligned.
18142   // Select an X-Form load if it is not.
18143   setXFormForUnalignedFI(N, Flags, Mode);
18144 
18145   // Set the mode to PC-Relative addressing mode if we have a valid PC-Rel node.
18146   if ((Mode == PPC::AM_XForm) && isPCRelNode(N)) {
18147     assert(Subtarget.isUsingPCRelativeCalls() &&
18148            "Must be using PC-Relative calls when a valid PC-Relative node is "
18149            "present!");
18150     Mode = PPC::AM_PCRel;
18151   }
18152 
18153   // Set Base and Disp accordingly depending on the address mode.
18154   switch (Mode) {
18155   case PPC::AM_DForm:
18156   case PPC::AM_DSForm:
18157   case PPC::AM_DQForm: {
18158     // This is a register plus a 16-bit immediate. The base will be the
18159     // register and the displacement will be the immediate unless it
18160     // isn't sufficiently aligned.
18161     if (Flags & PPC::MOF_RPlusSImm16) {
18162       SDValue Op0 = N.getOperand(0);
18163       SDValue Op1 = N.getOperand(1);
18164       int16_t Imm = cast<ConstantSDNode>(Op1)->getAPIntValue().getZExtValue();
18165       if (!Align || isAligned(*Align, Imm)) {
18166         Disp = DAG.getTargetConstant(Imm, DL, N.getValueType());
18167         Base = Op0;
18168         if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Op0)) {
18169           Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
18170           fixupFuncForFI(DAG, FI->getIndex(), N.getValueType());
18171         }
18172         break;
18173       }
18174     }
18175     // This is a register plus the @lo relocation. The base is the register
18176     // and the displacement is the global address.
18177     else if (Flags & PPC::MOF_RPlusLo) {
18178       Disp = N.getOperand(1).getOperand(0); // The global address.
18179       assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
18180              Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
18181              Disp.getOpcode() == ISD::TargetConstantPool ||
18182              Disp.getOpcode() == ISD::TargetJumpTable);
18183       Base = N.getOperand(0);
18184       break;
18185     }
18186     // This is a constant address at most 32 bits. The base will be
18187     // zero or load-immediate-shifted and the displacement will be
18188     // the low 16 bits of the address.
18189     else if (Flags & PPC::MOF_AddrIsSImm32) {
18190       auto *CN = cast<ConstantSDNode>(N);
18191       EVT CNType = CN->getValueType(0);
18192       uint64_t CNImm = CN->getZExtValue();
18193       // If this address fits entirely in a 16-bit sext immediate field, codegen
18194       // this as "d, 0".
18195       int16_t Imm;
18196       if (isIntS16Immediate(CN, Imm) && (!Align || isAligned(*Align, Imm))) {
18197         Disp = DAG.getTargetConstant(Imm, DL, CNType);
18198         Base = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO,
18199                                CNType);
18200         break;
18201       }
18202       // Handle 32-bit sext immediate with LIS + Addr mode.
18203       if ((CNType == MVT::i32 || isInt<32>(CNImm)) &&
18204           (!Align || isAligned(*Align, CNImm))) {
18205         int32_t Addr = (int32_t)CNImm;
18206         // Otherwise, break this down into LIS + Disp.
18207         Disp = DAG.getTargetConstant((int16_t)Addr, DL, MVT::i32);
18208         Base =
18209             DAG.getTargetConstant((Addr - (int16_t)Addr) >> 16, DL, MVT::i32);
18210         uint32_t LIS = CNType == MVT::i32 ? PPC::LIS : PPC::LIS8;
18211         Base = SDValue(DAG.getMachineNode(LIS, DL, CNType, Base), 0);
18212         break;
18213       }
18214     }
18215     // Otherwise, the PPC:MOF_NotAdd flag is set. Load/Store is Non-foldable.
18216     Disp = DAG.getTargetConstant(0, DL, getPointerTy(DAG.getDataLayout()));
18217     if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N)) {
18218       Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
18219       fixupFuncForFI(DAG, FI->getIndex(), N.getValueType());
18220     } else
18221       Base = N;
18222     break;
18223   }
18224   case PPC::AM_PrefixDForm: {
18225     int64_t Imm34 = 0;
18226     unsigned Opcode = N.getOpcode();
18227     if (((Opcode == ISD::ADD) || (Opcode == ISD::OR)) &&
18228         (isIntS34Immediate(N.getOperand(1), Imm34))) {
18229       // N is an Add/OR Node, and it's operand is a 34-bit signed immediate.
18230       Disp = DAG.getTargetConstant(Imm34, DL, N.getValueType());
18231       if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0)))
18232         Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
18233       else
18234         Base = N.getOperand(0);
18235     } else if (isIntS34Immediate(N, Imm34)) {
18236       // The address is a 34-bit signed immediate.
18237       Disp = DAG.getTargetConstant(Imm34, DL, N.getValueType());
18238       Base = DAG.getRegister(PPC::ZERO8, N.getValueType());
18239     }
18240     break;
18241   }
18242   case PPC::AM_PCRel: {
18243     // When selecting PC-Relative instructions, "Base" is not utilized as
18244     // we select the address as [PC+imm].
18245     Disp = N;
18246     break;
18247   }
18248   case PPC::AM_None:
18249     break;
18250   default: { // By default, X-Form is always available to be selected.
18251     // When a frame index is not aligned, we also match by XForm.
18252     FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N);
18253     Base = FI ? N : N.getOperand(1);
18254     Disp = FI ? DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO,
18255                                 N.getValueType())
18256               : N.getOperand(0);
18257     break;
18258   }
18259   }
18260   return Mode;
18261 }
18262 
18263 CCAssignFn *PPCTargetLowering::ccAssignFnForCall(CallingConv::ID CC,
18264                                                  bool Return,
18265                                                  bool IsVarArg) const {
18266   switch (CC) {
18267   case CallingConv::Cold:
18268     return (Return ? RetCC_PPC_Cold : CC_PPC64_ELF_FIS);
18269   default:
18270     return CC_PPC64_ELF_FIS;
18271   }
18272 }
18273 
18274 bool PPCTargetLowering::shouldInlineQuadwordAtomics() const {
18275   // TODO: 16-byte atomic type support for AIX is in progress; we should be able
18276   // to inline 16-byte atomic ops on AIX too in the future.
18277   return Subtarget.isPPC64() &&
18278          (EnableQuadwordAtomics || !Subtarget.getTargetTriple().isOSAIX()) &&
18279          Subtarget.hasQuadwordAtomics();
18280 }
18281 
18282 TargetLowering::AtomicExpansionKind
18283 PPCTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18284   unsigned Size = AI->getType()->getPrimitiveSizeInBits();
18285   if (shouldInlineQuadwordAtomics() && Size == 128)
18286     return AtomicExpansionKind::MaskedIntrinsic;
18287 
18288   switch (AI->getOperation()) {
18289   case AtomicRMWInst::UIncWrap:
18290   case AtomicRMWInst::UDecWrap:
18291     return AtomicExpansionKind::CmpXChg;
18292   default:
18293     return TargetLowering::shouldExpandAtomicRMWInIR(AI);
18294   }
18295 
18296   llvm_unreachable("unreachable atomicrmw operation");
18297 }
18298 
18299 TargetLowering::AtomicExpansionKind
18300 PPCTargetLowering::shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const {
18301   unsigned Size = AI->getNewValOperand()->getType()->getPrimitiveSizeInBits();
18302   if (shouldInlineQuadwordAtomics() && Size == 128)
18303     return AtomicExpansionKind::MaskedIntrinsic;
18304   return TargetLowering::shouldExpandAtomicCmpXchgInIR(AI);
18305 }
18306 
18307 static Intrinsic::ID
18308 getIntrinsicForAtomicRMWBinOp128(AtomicRMWInst::BinOp BinOp) {
18309   switch (BinOp) {
18310   default:
18311     llvm_unreachable("Unexpected AtomicRMW BinOp");
18312   case AtomicRMWInst::Xchg:
18313     return Intrinsic::ppc_atomicrmw_xchg_i128;
18314   case AtomicRMWInst::Add:
18315     return Intrinsic::ppc_atomicrmw_add_i128;
18316   case AtomicRMWInst::Sub:
18317     return Intrinsic::ppc_atomicrmw_sub_i128;
18318   case AtomicRMWInst::And:
18319     return Intrinsic::ppc_atomicrmw_and_i128;
18320   case AtomicRMWInst::Or:
18321     return Intrinsic::ppc_atomicrmw_or_i128;
18322   case AtomicRMWInst::Xor:
18323     return Intrinsic::ppc_atomicrmw_xor_i128;
18324   case AtomicRMWInst::Nand:
18325     return Intrinsic::ppc_atomicrmw_nand_i128;
18326   }
18327 }
18328 
18329 Value *PPCTargetLowering::emitMaskedAtomicRMWIntrinsic(
18330     IRBuilderBase &Builder, AtomicRMWInst *AI, Value *AlignedAddr, Value *Incr,
18331     Value *Mask, Value *ShiftAmt, AtomicOrdering Ord) const {
18332   assert(shouldInlineQuadwordAtomics() && "Only support quadword now");
18333   Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18334   Type *ValTy = Incr->getType();
18335   assert(ValTy->getPrimitiveSizeInBits() == 128);
18336   Function *RMW = Intrinsic::getDeclaration(
18337       M, getIntrinsicForAtomicRMWBinOp128(AI->getOperation()));
18338   Type *Int64Ty = Type::getInt64Ty(M->getContext());
18339   Value *IncrLo = Builder.CreateTrunc(Incr, Int64Ty, "incr_lo");
18340   Value *IncrHi =
18341       Builder.CreateTrunc(Builder.CreateLShr(Incr, 64), Int64Ty, "incr_hi");
18342   Value *Addr =
18343       Builder.CreateBitCast(AlignedAddr, Type::getInt8PtrTy(M->getContext()));
18344   Value *LoHi = Builder.CreateCall(RMW, {Addr, IncrLo, IncrHi});
18345   Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
18346   Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
18347   Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
18348   Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
18349   return Builder.CreateOr(
18350       Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
18351 }
18352 
18353 Value *PPCTargetLowering::emitMaskedAtomicCmpXchgIntrinsic(
18354     IRBuilderBase &Builder, AtomicCmpXchgInst *CI, Value *AlignedAddr,
18355     Value *CmpVal, Value *NewVal, Value *Mask, AtomicOrdering Ord) const {
18356   assert(shouldInlineQuadwordAtomics() && "Only support quadword now");
18357   Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18358   Type *ValTy = CmpVal->getType();
18359   assert(ValTy->getPrimitiveSizeInBits() == 128);
18360   Function *IntCmpXchg =
18361       Intrinsic::getDeclaration(M, Intrinsic::ppc_cmpxchg_i128);
18362   Type *Int64Ty = Type::getInt64Ty(M->getContext());
18363   Value *CmpLo = Builder.CreateTrunc(CmpVal, Int64Ty, "cmp_lo");
18364   Value *CmpHi =
18365       Builder.CreateTrunc(Builder.CreateLShr(CmpVal, 64), Int64Ty, "cmp_hi");
18366   Value *NewLo = Builder.CreateTrunc(NewVal, Int64Ty, "new_lo");
18367   Value *NewHi =
18368       Builder.CreateTrunc(Builder.CreateLShr(NewVal, 64), Int64Ty, "new_hi");
18369   Value *Addr =
18370       Builder.CreateBitCast(AlignedAddr, Type::getInt8PtrTy(M->getContext()));
18371   emitLeadingFence(Builder, CI, Ord);
18372   Value *LoHi =
18373       Builder.CreateCall(IntCmpXchg, {Addr, CmpLo, CmpHi, NewLo, NewHi});
18374   emitTrailingFence(Builder, CI, Ord);
18375   Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
18376   Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
18377   Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
18378   Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
18379   return Builder.CreateOr(
18380       Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
18381 }
18382