xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp (revision a7dea1671b87c07d2d266f836bfa8b58efc7c134)
1 //===-- AMDGPUISelDAGToDAG.cpp - A dag to dag inst selector for AMDGPU ----===//
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 /// \file
10 /// Defines an instruction selector for the AMDGPU target.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "AMDGPU.h"
15 #include "AMDGPUArgumentUsageInfo.h"
16 #include "AMDGPUISelLowering.h" // For AMDGPUISD
17 #include "AMDGPUInstrInfo.h"
18 #include "AMDGPUPerfHintAnalysis.h"
19 #include "AMDGPURegisterInfo.h"
20 #include "AMDGPUSubtarget.h"
21 #include "AMDGPUTargetMachine.h"
22 #include "SIDefines.h"
23 #include "SIISelLowering.h"
24 #include "SIInstrInfo.h"
25 #include "SIMachineFunctionInfo.h"
26 #include "SIRegisterInfo.h"
27 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
28 #include "llvm/ADT/APInt.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/StringRef.h"
31 #include "llvm/Analysis/LegacyDivergenceAnalysis.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/CodeGen/FunctionLoweringInfo.h"
34 #include "llvm/CodeGen/ISDOpcodes.h"
35 #include "llvm/CodeGen/MachineFunction.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/SelectionDAG.h"
38 #include "llvm/CodeGen/SelectionDAGISel.h"
39 #include "llvm/CodeGen/SelectionDAGNodes.h"
40 #include "llvm/CodeGen/ValueTypes.h"
41 #include "llvm/IR/BasicBlock.h"
42 #ifdef EXPENSIVE_CHECKS
43 #include "llvm/IR/Dominators.h"
44 #endif
45 #include "llvm/IR/Instruction.h"
46 #include "llvm/MC/MCInstrDesc.h"
47 #include "llvm/Support/Casting.h"
48 #include "llvm/Support/CodeGen.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MachineValueType.h"
51 #include "llvm/Support/MathExtras.h"
52 #include <cassert>
53 #include <cstdint>
54 #include <new>
55 #include <vector>
56 
57 #define DEBUG_TYPE "isel"
58 
59 using namespace llvm;
60 
61 namespace llvm {
62 
63 class R600InstrInfo;
64 
65 } // end namespace llvm
66 
67 //===----------------------------------------------------------------------===//
68 // Instruction Selector Implementation
69 //===----------------------------------------------------------------------===//
70 
71 namespace {
72 
73 static bool isNullConstantOrUndef(SDValue V) {
74   if (V.isUndef())
75     return true;
76 
77   ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
78   return Const != nullptr && Const->isNullValue();
79 }
80 
81 static bool getConstantValue(SDValue N, uint32_t &Out) {
82   // This is only used for packed vectors, where ussing 0 for undef should
83   // always be good.
84   if (N.isUndef()) {
85     Out = 0;
86     return true;
87   }
88 
89   if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
90     Out = C->getAPIntValue().getSExtValue();
91     return true;
92   }
93 
94   if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N)) {
95     Out = C->getValueAPF().bitcastToAPInt().getSExtValue();
96     return true;
97   }
98 
99   return false;
100 }
101 
102 // TODO: Handle undef as zero
103 static SDNode *packConstantV2I16(const SDNode *N, SelectionDAG &DAG,
104                                  bool Negate = false) {
105   assert(N->getOpcode() == ISD::BUILD_VECTOR && N->getNumOperands() == 2);
106   uint32_t LHSVal, RHSVal;
107   if (getConstantValue(N->getOperand(0), LHSVal) &&
108       getConstantValue(N->getOperand(1), RHSVal)) {
109     SDLoc SL(N);
110     uint32_t K = Negate ?
111       (-LHSVal & 0xffff) | (-RHSVal << 16) :
112       (LHSVal & 0xffff) | (RHSVal << 16);
113     return DAG.getMachineNode(AMDGPU::S_MOV_B32, SL, N->getValueType(0),
114                               DAG.getTargetConstant(K, SL, MVT::i32));
115   }
116 
117   return nullptr;
118 }
119 
120 static SDNode *packNegConstantV2I16(const SDNode *N, SelectionDAG &DAG) {
121   return packConstantV2I16(N, DAG, true);
122 }
123 
124 /// AMDGPU specific code to select AMDGPU machine instructions for
125 /// SelectionDAG operations.
126 class AMDGPUDAGToDAGISel : public SelectionDAGISel {
127   // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can
128   // make the right decision when generating code for different targets.
129   const GCNSubtarget *Subtarget;
130   bool EnableLateStructurizeCFG;
131 
132 public:
133   explicit AMDGPUDAGToDAGISel(TargetMachine *TM = nullptr,
134                               CodeGenOpt::Level OptLevel = CodeGenOpt::Default)
135     : SelectionDAGISel(*TM, OptLevel) {
136     EnableLateStructurizeCFG = AMDGPUTargetMachine::EnableLateStructurizeCFG;
137   }
138   ~AMDGPUDAGToDAGISel() override = default;
139 
140   void getAnalysisUsage(AnalysisUsage &AU) const override {
141     AU.addRequired<AMDGPUArgumentUsageInfo>();
142     AU.addRequired<LegacyDivergenceAnalysis>();
143 #ifdef EXPENSIVE_CHECKS
144     AU.addRequired<DominatorTreeWrapperPass>();
145     AU.addRequired<LoopInfoWrapperPass>();
146 #endif
147     SelectionDAGISel::getAnalysisUsage(AU);
148   }
149 
150   bool matchLoadD16FromBuildVector(SDNode *N) const;
151 
152   bool runOnMachineFunction(MachineFunction &MF) override;
153   void PreprocessISelDAG() override;
154   void Select(SDNode *N) override;
155   StringRef getPassName() const override;
156   void PostprocessISelDAG() override;
157 
158 protected:
159   void SelectBuildVector(SDNode *N, unsigned RegClassID);
160 
161 private:
162   std::pair<SDValue, SDValue> foldFrameIndex(SDValue N) const;
163   bool isNoNanSrc(SDValue N) const;
164   bool isInlineImmediate(const SDNode *N, bool Negated = false) const;
165   bool isNegInlineImmediate(const SDNode *N) const {
166     return isInlineImmediate(N, true);
167   }
168 
169   bool isVGPRImm(const SDNode *N) const;
170   bool isUniformLoad(const SDNode *N) const;
171   bool isUniformBr(const SDNode *N) const;
172 
173   MachineSDNode *buildSMovImm64(SDLoc &DL, uint64_t Val, EVT VT) const;
174 
175   SDNode *glueCopyToOp(SDNode *N, SDValue NewChain, SDValue Glue) const;
176   SDNode *glueCopyToM0(SDNode *N, SDValue Val) const;
177   SDNode *glueCopyToM0LDSInit(SDNode *N) const;
178 
179   const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const;
180   virtual bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset);
181   virtual bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset);
182   bool isDSOffsetLegal(SDValue Base, unsigned Offset,
183                        unsigned OffsetBits) const;
184   bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const;
185   bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0,
186                                  SDValue &Offset1) const;
187   bool SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
188                    SDValue &SOffset, SDValue &Offset, SDValue &Offen,
189                    SDValue &Idxen, SDValue &Addr64, SDValue &GLC, SDValue &SLC,
190                    SDValue &TFE, SDValue &DLC, SDValue &SWZ) const;
191   bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
192                          SDValue &SOffset, SDValue &Offset, SDValue &GLC,
193                          SDValue &SLC, SDValue &TFE, SDValue &DLC,
194                          SDValue &SWZ) const;
195   bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
196                          SDValue &VAddr, SDValue &SOffset, SDValue &Offset,
197                          SDValue &SLC) const;
198   bool SelectMUBUFScratchOffen(SDNode *Parent,
199                                SDValue Addr, SDValue &RSrc, SDValue &VAddr,
200                                SDValue &SOffset, SDValue &ImmOffset) const;
201   bool SelectMUBUFScratchOffset(SDNode *Parent,
202                                 SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
203                                 SDValue &Offset) const;
204 
205   bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset,
206                          SDValue &Offset, SDValue &GLC, SDValue &SLC,
207                          SDValue &TFE, SDValue &DLC, SDValue &SWZ) const;
208   bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
209                          SDValue &Offset, SDValue &SLC) const;
210   bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
211                          SDValue &Offset) const;
212 
213   template <bool IsSigned>
214   bool SelectFlatOffset(SDNode *N, SDValue Addr, SDValue &VAddr,
215                         SDValue &Offset, SDValue &SLC) const;
216   bool SelectFlatAtomic(SDNode *N, SDValue Addr, SDValue &VAddr,
217                         SDValue &Offset, SDValue &SLC) const;
218   bool SelectFlatAtomicSigned(SDNode *N, SDValue Addr, SDValue &VAddr,
219                               SDValue &Offset, SDValue &SLC) const;
220 
221   bool SelectSMRDOffset(SDValue ByteOffsetNode, SDValue &Offset,
222                         bool &Imm) const;
223   SDValue Expand32BitAddress(SDValue Addr) const;
224   bool SelectSMRD(SDValue Addr, SDValue &SBase, SDValue &Offset,
225                   bool &Imm) const;
226   bool SelectSMRDImm(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
227   bool SelectSMRDImm32(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
228   bool SelectSMRDSgpr(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
229   bool SelectSMRDBufferImm(SDValue Addr, SDValue &Offset) const;
230   bool SelectSMRDBufferImm32(SDValue Addr, SDValue &Offset) const;
231   bool SelectMOVRELOffset(SDValue Index, SDValue &Base, SDValue &Offset) const;
232 
233   bool SelectVOP3Mods_NNaN(SDValue In, SDValue &Src, SDValue &SrcMods) const;
234   bool SelectVOP3Mods_f32(SDValue In, SDValue &Src, SDValue &SrcMods) const;
235   bool SelectVOP3ModsImpl(SDValue In, SDValue &Src, unsigned &SrcMods) const;
236   bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
237   bool SelectVOP3NoMods(SDValue In, SDValue &Src) const;
238   bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
239                        SDValue &Clamp, SDValue &Omod) const;
240   bool SelectVOP3NoMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
241                          SDValue &Clamp, SDValue &Omod) const;
242 
243   bool SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src, SDValue &SrcMods,
244                                  SDValue &Clamp,
245                                  SDValue &Omod) const;
246 
247   bool SelectVOP3OMods(SDValue In, SDValue &Src,
248                        SDValue &Clamp, SDValue &Omod) const;
249 
250   bool SelectVOP3PMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
251   bool SelectVOP3PMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
252                         SDValue &Clamp) const;
253 
254   bool SelectVOP3OpSel(SDValue In, SDValue &Src, SDValue &SrcMods) const;
255   bool SelectVOP3OpSel0(SDValue In, SDValue &Src, SDValue &SrcMods,
256                         SDValue &Clamp) const;
257 
258   bool SelectVOP3OpSelMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
259   bool SelectVOP3OpSelMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
260                             SDValue &Clamp) const;
261   bool SelectVOP3PMadMixModsImpl(SDValue In, SDValue &Src, unsigned &Mods) const;
262   bool SelectVOP3PMadMixMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
263 
264   SDValue getHi16Elt(SDValue In) const;
265 
266   SDValue getMaterializedScalarImm32(int64_t Val, const SDLoc &DL) const;
267 
268   void SelectADD_SUB_I64(SDNode *N);
269   void SelectAddcSubb(SDNode *N);
270   void SelectUADDO_USUBO(SDNode *N);
271   void SelectDIV_SCALE(SDNode *N);
272   void SelectDIV_FMAS(SDNode *N);
273   void SelectMAD_64_32(SDNode *N);
274   void SelectFMA_W_CHAIN(SDNode *N);
275   void SelectFMUL_W_CHAIN(SDNode *N);
276 
277   SDNode *getS_BFE(unsigned Opcode, const SDLoc &DL, SDValue Val,
278                    uint32_t Offset, uint32_t Width);
279   void SelectS_BFEFromShifts(SDNode *N);
280   void SelectS_BFE(SDNode *N);
281   bool isCBranchSCC(const SDNode *N) const;
282   void SelectBRCOND(SDNode *N);
283   void SelectFMAD_FMA(SDNode *N);
284   void SelectATOMIC_CMP_SWAP(SDNode *N);
285   void SelectDSAppendConsume(SDNode *N, unsigned IntrID);
286   void SelectDS_GWS(SDNode *N, unsigned IntrID);
287   void SelectINTRINSIC_W_CHAIN(SDNode *N);
288   void SelectINTRINSIC_WO_CHAIN(SDNode *N);
289   void SelectINTRINSIC_VOID(SDNode *N);
290 
291 protected:
292   // Include the pieces autogenerated from the target description.
293 #include "AMDGPUGenDAGISel.inc"
294 };
295 
296 class R600DAGToDAGISel : public AMDGPUDAGToDAGISel {
297   const R600Subtarget *Subtarget;
298 
299   bool isConstantLoad(const MemSDNode *N, int cbID) const;
300   bool SelectGlobalValueConstantOffset(SDValue Addr, SDValue& IntPtr);
301   bool SelectGlobalValueVariableOffset(SDValue Addr, SDValue &BaseReg,
302                                        SDValue& Offset);
303 public:
304   explicit R600DAGToDAGISel(TargetMachine *TM, CodeGenOpt::Level OptLevel) :
305       AMDGPUDAGToDAGISel(TM, OptLevel) {}
306 
307   void Select(SDNode *N) override;
308 
309   bool SelectADDRIndirect(SDValue Addr, SDValue &Base,
310                           SDValue &Offset) override;
311   bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
312                           SDValue &Offset) override;
313 
314   bool runOnMachineFunction(MachineFunction &MF) override;
315 
316   void PreprocessISelDAG() override {}
317 
318 protected:
319   // Include the pieces autogenerated from the target description.
320 #include "R600GenDAGISel.inc"
321 };
322 
323 static SDValue stripBitcast(SDValue Val) {
324   return Val.getOpcode() == ISD::BITCAST ? Val.getOperand(0) : Val;
325 }
326 
327 // Figure out if this is really an extract of the high 16-bits of a dword.
328 static bool isExtractHiElt(SDValue In, SDValue &Out) {
329   In = stripBitcast(In);
330   if (In.getOpcode() != ISD::TRUNCATE)
331     return false;
332 
333   SDValue Srl = In.getOperand(0);
334   if (Srl.getOpcode() == ISD::SRL) {
335     if (ConstantSDNode *ShiftAmt = dyn_cast<ConstantSDNode>(Srl.getOperand(1))) {
336       if (ShiftAmt->getZExtValue() == 16) {
337         Out = stripBitcast(Srl.getOperand(0));
338         return true;
339       }
340     }
341   }
342 
343   return false;
344 }
345 
346 // Look through operations that obscure just looking at the low 16-bits of the
347 // same register.
348 static SDValue stripExtractLoElt(SDValue In) {
349   if (In.getOpcode() == ISD::TRUNCATE) {
350     SDValue Src = In.getOperand(0);
351     if (Src.getValueType().getSizeInBits() == 32)
352       return stripBitcast(Src);
353   }
354 
355   return In;
356 }
357 
358 }  // end anonymous namespace
359 
360 INITIALIZE_PASS_BEGIN(AMDGPUDAGToDAGISel, "amdgpu-isel",
361                       "AMDGPU DAG->DAG Pattern Instruction Selection", false, false)
362 INITIALIZE_PASS_DEPENDENCY(AMDGPUArgumentUsageInfo)
363 INITIALIZE_PASS_DEPENDENCY(AMDGPUPerfHintAnalysis)
364 INITIALIZE_PASS_DEPENDENCY(LegacyDivergenceAnalysis)
365 #ifdef EXPENSIVE_CHECKS
366 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
367 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
368 #endif
369 INITIALIZE_PASS_END(AMDGPUDAGToDAGISel, "amdgpu-isel",
370                     "AMDGPU DAG->DAG Pattern Instruction Selection", false, false)
371 
372 /// This pass converts a legalized DAG into a AMDGPU-specific
373 // DAG, ready for instruction scheduling.
374 FunctionPass *llvm::createAMDGPUISelDag(TargetMachine *TM,
375                                         CodeGenOpt::Level OptLevel) {
376   return new AMDGPUDAGToDAGISel(TM, OptLevel);
377 }
378 
379 /// This pass converts a legalized DAG into a R600-specific
380 // DAG, ready for instruction scheduling.
381 FunctionPass *llvm::createR600ISelDag(TargetMachine *TM,
382                                       CodeGenOpt::Level OptLevel) {
383   return new R600DAGToDAGISel(TM, OptLevel);
384 }
385 
386 bool AMDGPUDAGToDAGISel::runOnMachineFunction(MachineFunction &MF) {
387 #ifdef EXPENSIVE_CHECKS
388   DominatorTree & DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
389   LoopInfo * LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
390   for (auto &L : LI->getLoopsInPreorder()) {
391     assert(L->isLCSSAForm(DT));
392   }
393 #endif
394   Subtarget = &MF.getSubtarget<GCNSubtarget>();
395   return SelectionDAGISel::runOnMachineFunction(MF);
396 }
397 
398 bool AMDGPUDAGToDAGISel::matchLoadD16FromBuildVector(SDNode *N) const {
399   assert(Subtarget->d16PreservesUnusedBits());
400   MVT VT = N->getValueType(0).getSimpleVT();
401   if (VT != MVT::v2i16 && VT != MVT::v2f16)
402     return false;
403 
404   SDValue Lo = N->getOperand(0);
405   SDValue Hi = N->getOperand(1);
406 
407   LoadSDNode *LdHi = dyn_cast<LoadSDNode>(stripBitcast(Hi));
408 
409   // build_vector lo, (load ptr) -> load_d16_hi ptr, lo
410   // build_vector lo, (zextload ptr from i8) -> load_d16_hi_u8 ptr, lo
411   // build_vector lo, (sextload ptr from i8) -> load_d16_hi_i8 ptr, lo
412 
413   // Need to check for possible indirect dependencies on the other half of the
414   // vector to avoid introducing a cycle.
415   if (LdHi && Hi.hasOneUse() && !LdHi->isPredecessorOf(Lo.getNode())) {
416     SDVTList VTList = CurDAG->getVTList(VT, MVT::Other);
417 
418     SDValue TiedIn = CurDAG->getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), VT, Lo);
419     SDValue Ops[] = {
420       LdHi->getChain(), LdHi->getBasePtr(), TiedIn
421     };
422 
423     unsigned LoadOp = AMDGPUISD::LOAD_D16_HI;
424     if (LdHi->getMemoryVT() == MVT::i8) {
425       LoadOp = LdHi->getExtensionType() == ISD::SEXTLOAD ?
426         AMDGPUISD::LOAD_D16_HI_I8 : AMDGPUISD::LOAD_D16_HI_U8;
427     } else {
428       assert(LdHi->getMemoryVT() == MVT::i16);
429     }
430 
431     SDValue NewLoadHi =
432       CurDAG->getMemIntrinsicNode(LoadOp, SDLoc(LdHi), VTList,
433                                   Ops, LdHi->getMemoryVT(),
434                                   LdHi->getMemOperand());
435 
436     CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), NewLoadHi);
437     CurDAG->ReplaceAllUsesOfValueWith(SDValue(LdHi, 1), NewLoadHi.getValue(1));
438     return true;
439   }
440 
441   // build_vector (load ptr), hi -> load_d16_lo ptr, hi
442   // build_vector (zextload ptr from i8), hi -> load_d16_lo_u8 ptr, hi
443   // build_vector (sextload ptr from i8), hi -> load_d16_lo_i8 ptr, hi
444   LoadSDNode *LdLo = dyn_cast<LoadSDNode>(stripBitcast(Lo));
445   if (LdLo && Lo.hasOneUse()) {
446     SDValue TiedIn = getHi16Elt(Hi);
447     if (!TiedIn || LdLo->isPredecessorOf(TiedIn.getNode()))
448       return false;
449 
450     SDVTList VTList = CurDAG->getVTList(VT, MVT::Other);
451     unsigned LoadOp = AMDGPUISD::LOAD_D16_LO;
452     if (LdLo->getMemoryVT() == MVT::i8) {
453       LoadOp = LdLo->getExtensionType() == ISD::SEXTLOAD ?
454         AMDGPUISD::LOAD_D16_LO_I8 : AMDGPUISD::LOAD_D16_LO_U8;
455     } else {
456       assert(LdLo->getMemoryVT() == MVT::i16);
457     }
458 
459     TiedIn = CurDAG->getNode(ISD::BITCAST, SDLoc(N), VT, TiedIn);
460 
461     SDValue Ops[] = {
462       LdLo->getChain(), LdLo->getBasePtr(), TiedIn
463     };
464 
465     SDValue NewLoadLo =
466       CurDAG->getMemIntrinsicNode(LoadOp, SDLoc(LdLo), VTList,
467                                   Ops, LdLo->getMemoryVT(),
468                                   LdLo->getMemOperand());
469 
470     CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), NewLoadLo);
471     CurDAG->ReplaceAllUsesOfValueWith(SDValue(LdLo, 1), NewLoadLo.getValue(1));
472     return true;
473   }
474 
475   return false;
476 }
477 
478 void AMDGPUDAGToDAGISel::PreprocessISelDAG() {
479   if (!Subtarget->d16PreservesUnusedBits())
480     return;
481 
482   SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end();
483 
484   bool MadeChange = false;
485   while (Position != CurDAG->allnodes_begin()) {
486     SDNode *N = &*--Position;
487     if (N->use_empty())
488       continue;
489 
490     switch (N->getOpcode()) {
491     case ISD::BUILD_VECTOR:
492       MadeChange |= matchLoadD16FromBuildVector(N);
493       break;
494     default:
495       break;
496     }
497   }
498 
499   if (MadeChange) {
500     CurDAG->RemoveDeadNodes();
501     LLVM_DEBUG(dbgs() << "After PreProcess:\n";
502                CurDAG->dump(););
503   }
504 }
505 
506 bool AMDGPUDAGToDAGISel::isNoNanSrc(SDValue N) const {
507   if (TM.Options.NoNaNsFPMath)
508     return true;
509 
510   // TODO: Move into isKnownNeverNaN
511   if (N->getFlags().isDefined())
512     return N->getFlags().hasNoNaNs();
513 
514   return CurDAG->isKnownNeverNaN(N);
515 }
516 
517 bool AMDGPUDAGToDAGISel::isInlineImmediate(const SDNode *N,
518                                            bool Negated) const {
519   if (N->isUndef())
520     return true;
521 
522   const SIInstrInfo *TII = Subtarget->getInstrInfo();
523   if (Negated) {
524     if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N))
525       return TII->isInlineConstant(-C->getAPIntValue());
526 
527     if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N))
528       return TII->isInlineConstant(-C->getValueAPF().bitcastToAPInt());
529 
530   } else {
531     if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N))
532       return TII->isInlineConstant(C->getAPIntValue());
533 
534     if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N))
535       return TII->isInlineConstant(C->getValueAPF().bitcastToAPInt());
536   }
537 
538   return false;
539 }
540 
541 /// Determine the register class for \p OpNo
542 /// \returns The register class of the virtual register that will be used for
543 /// the given operand number \OpNo or NULL if the register class cannot be
544 /// determined.
545 const TargetRegisterClass *AMDGPUDAGToDAGISel::getOperandRegClass(SDNode *N,
546                                                           unsigned OpNo) const {
547   if (!N->isMachineOpcode()) {
548     if (N->getOpcode() == ISD::CopyToReg) {
549       unsigned Reg = cast<RegisterSDNode>(N->getOperand(1))->getReg();
550       if (Register::isVirtualRegister(Reg)) {
551         MachineRegisterInfo &MRI = CurDAG->getMachineFunction().getRegInfo();
552         return MRI.getRegClass(Reg);
553       }
554 
555       const SIRegisterInfo *TRI
556         = static_cast<const GCNSubtarget *>(Subtarget)->getRegisterInfo();
557       return TRI->getPhysRegClass(Reg);
558     }
559 
560     return nullptr;
561   }
562 
563   switch (N->getMachineOpcode()) {
564   default: {
565     const MCInstrDesc &Desc =
566         Subtarget->getInstrInfo()->get(N->getMachineOpcode());
567     unsigned OpIdx = Desc.getNumDefs() + OpNo;
568     if (OpIdx >= Desc.getNumOperands())
569       return nullptr;
570     int RegClass = Desc.OpInfo[OpIdx].RegClass;
571     if (RegClass == -1)
572       return nullptr;
573 
574     return Subtarget->getRegisterInfo()->getRegClass(RegClass);
575   }
576   case AMDGPU::REG_SEQUENCE: {
577     unsigned RCID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
578     const TargetRegisterClass *SuperRC =
579         Subtarget->getRegisterInfo()->getRegClass(RCID);
580 
581     SDValue SubRegOp = N->getOperand(OpNo + 1);
582     unsigned SubRegIdx = cast<ConstantSDNode>(SubRegOp)->getZExtValue();
583     return Subtarget->getRegisterInfo()->getSubClassWithSubReg(SuperRC,
584                                                               SubRegIdx);
585   }
586   }
587 }
588 
589 SDNode *AMDGPUDAGToDAGISel::glueCopyToOp(SDNode *N, SDValue NewChain,
590                                          SDValue Glue) const {
591   SmallVector <SDValue, 8> Ops;
592   Ops.push_back(NewChain); // Replace the chain.
593   for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
594     Ops.push_back(N->getOperand(i));
595 
596   Ops.push_back(Glue);
597   return CurDAG->MorphNodeTo(N, N->getOpcode(), N->getVTList(), Ops);
598 }
599 
600 SDNode *AMDGPUDAGToDAGISel::glueCopyToM0(SDNode *N, SDValue Val) const {
601   const SITargetLowering& Lowering =
602     *static_cast<const SITargetLowering*>(getTargetLowering());
603 
604   assert(N->getOperand(0).getValueType() == MVT::Other && "Expected chain");
605 
606   SDValue M0 = Lowering.copyToM0(*CurDAG, N->getOperand(0), SDLoc(N), Val);
607   return glueCopyToOp(N, M0, M0.getValue(1));
608 }
609 
610 SDNode *AMDGPUDAGToDAGISel::glueCopyToM0LDSInit(SDNode *N) const {
611   unsigned AS = cast<MemSDNode>(N)->getAddressSpace();
612   if (AS == AMDGPUAS::LOCAL_ADDRESS) {
613     if (Subtarget->ldsRequiresM0Init())
614       return glueCopyToM0(N, CurDAG->getTargetConstant(-1, SDLoc(N), MVT::i32));
615   } else if (AS == AMDGPUAS::REGION_ADDRESS) {
616     MachineFunction &MF = CurDAG->getMachineFunction();
617     unsigned Value = MF.getInfo<SIMachineFunctionInfo>()->getGDSSize();
618     return
619         glueCopyToM0(N, CurDAG->getTargetConstant(Value, SDLoc(N), MVT::i32));
620   }
621   return N;
622 }
623 
624 MachineSDNode *AMDGPUDAGToDAGISel::buildSMovImm64(SDLoc &DL, uint64_t Imm,
625                                                   EVT VT) const {
626   SDNode *Lo = CurDAG->getMachineNode(
627       AMDGPU::S_MOV_B32, DL, MVT::i32,
628       CurDAG->getTargetConstant(Imm & 0xFFFFFFFF, DL, MVT::i32));
629   SDNode *Hi =
630       CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32,
631                              CurDAG->getTargetConstant(Imm >> 32, DL, MVT::i32));
632   const SDValue Ops[] = {
633       CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
634       SDValue(Lo, 0), CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
635       SDValue(Hi, 0), CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32)};
636 
637   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL, VT, Ops);
638 }
639 
640 static unsigned selectSGPRVectorRegClassID(unsigned NumVectorElts) {
641   switch (NumVectorElts) {
642   case 1:
643     return AMDGPU::SReg_32RegClassID;
644   case 2:
645     return AMDGPU::SReg_64RegClassID;
646   case 3:
647     return AMDGPU::SGPR_96RegClassID;
648   case 4:
649     return AMDGPU::SGPR_128RegClassID;
650   case 5:
651     return AMDGPU::SGPR_160RegClassID;
652   case 8:
653     return AMDGPU::SReg_256RegClassID;
654   case 16:
655     return AMDGPU::SReg_512RegClassID;
656   case 32:
657     return AMDGPU::SReg_1024RegClassID;
658   }
659 
660   llvm_unreachable("invalid vector size");
661 }
662 
663 void AMDGPUDAGToDAGISel::SelectBuildVector(SDNode *N, unsigned RegClassID) {
664   EVT VT = N->getValueType(0);
665   unsigned NumVectorElts = VT.getVectorNumElements();
666   EVT EltVT = VT.getVectorElementType();
667   SDLoc DL(N);
668   SDValue RegClass = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32);
669 
670   if (NumVectorElts == 1) {
671     CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS, EltVT, N->getOperand(0),
672                          RegClass);
673     return;
674   }
675 
676   assert(NumVectorElts <= 32 && "Vectors with more than 32 elements not "
677                                   "supported yet");
678   // 32 = Max Num Vector Elements
679   // 2 = 2 REG_SEQUENCE operands per element (value, subreg index)
680   // 1 = Vector Register Class
681   SmallVector<SDValue, 32 * 2 + 1> RegSeqArgs(NumVectorElts * 2 + 1);
682 
683   RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, DL, MVT::i32);
684   bool IsRegSeq = true;
685   unsigned NOps = N->getNumOperands();
686   for (unsigned i = 0; i < NOps; i++) {
687     // XXX: Why is this here?
688     if (isa<RegisterSDNode>(N->getOperand(i))) {
689       IsRegSeq = false;
690       break;
691     }
692     unsigned Sub = AMDGPURegisterInfo::getSubRegFromChannel(i);
693     RegSeqArgs[1 + (2 * i)] = N->getOperand(i);
694     RegSeqArgs[1 + (2 * i) + 1] = CurDAG->getTargetConstant(Sub, DL, MVT::i32);
695   }
696   if (NOps != NumVectorElts) {
697     // Fill in the missing undef elements if this was a scalar_to_vector.
698     assert(N->getOpcode() == ISD::SCALAR_TO_VECTOR && NOps < NumVectorElts);
699     MachineSDNode *ImpDef = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
700                                                    DL, EltVT);
701     for (unsigned i = NOps; i < NumVectorElts; ++i) {
702       unsigned Sub = AMDGPURegisterInfo::getSubRegFromChannel(i);
703       RegSeqArgs[1 + (2 * i)] = SDValue(ImpDef, 0);
704       RegSeqArgs[1 + (2 * i) + 1] =
705           CurDAG->getTargetConstant(Sub, DL, MVT::i32);
706     }
707   }
708 
709   if (!IsRegSeq)
710     SelectCode(N);
711   CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(), RegSeqArgs);
712 }
713 
714 void AMDGPUDAGToDAGISel::Select(SDNode *N) {
715   unsigned int Opc = N->getOpcode();
716   if (N->isMachineOpcode()) {
717     N->setNodeId(-1);
718     return;   // Already selected.
719   }
720 
721   // isa<MemSDNode> almost works but is slightly too permissive for some DS
722   // intrinsics.
723   if (Opc == ISD::LOAD || Opc == ISD::STORE || isa<AtomicSDNode>(N) ||
724       (Opc == AMDGPUISD::ATOMIC_INC || Opc == AMDGPUISD::ATOMIC_DEC ||
725        Opc == ISD::ATOMIC_LOAD_FADD ||
726        Opc == AMDGPUISD::ATOMIC_LOAD_FMIN ||
727        Opc == AMDGPUISD::ATOMIC_LOAD_FMAX)) {
728     N = glueCopyToM0LDSInit(N);
729     SelectCode(N);
730     return;
731   }
732 
733   switch (Opc) {
734   default:
735     break;
736   // We are selecting i64 ADD here instead of custom lower it during
737   // DAG legalization, so we can fold some i64 ADDs used for address
738   // calculation into the LOAD and STORE instructions.
739   case ISD::ADDC:
740   case ISD::ADDE:
741   case ISD::SUBC:
742   case ISD::SUBE: {
743     if (N->getValueType(0) != MVT::i64)
744       break;
745 
746     SelectADD_SUB_I64(N);
747     return;
748   }
749   case ISD::ADDCARRY:
750   case ISD::SUBCARRY:
751     if (N->getValueType(0) != MVT::i32)
752       break;
753 
754     SelectAddcSubb(N);
755     return;
756   case ISD::UADDO:
757   case ISD::USUBO: {
758     SelectUADDO_USUBO(N);
759     return;
760   }
761   case AMDGPUISD::FMUL_W_CHAIN: {
762     SelectFMUL_W_CHAIN(N);
763     return;
764   }
765   case AMDGPUISD::FMA_W_CHAIN: {
766     SelectFMA_W_CHAIN(N);
767     return;
768   }
769 
770   case ISD::SCALAR_TO_VECTOR:
771   case ISD::BUILD_VECTOR: {
772     EVT VT = N->getValueType(0);
773     unsigned NumVectorElts = VT.getVectorNumElements();
774     if (VT.getScalarSizeInBits() == 16) {
775       if (Opc == ISD::BUILD_VECTOR && NumVectorElts == 2) {
776         if (SDNode *Packed = packConstantV2I16(N, *CurDAG)) {
777           ReplaceNode(N, Packed);
778           return;
779         }
780       }
781 
782       break;
783     }
784 
785     assert(VT.getVectorElementType().bitsEq(MVT::i32));
786     unsigned RegClassID = selectSGPRVectorRegClassID(NumVectorElts);
787     SelectBuildVector(N, RegClassID);
788     return;
789   }
790   case ISD::BUILD_PAIR: {
791     SDValue RC, SubReg0, SubReg1;
792     SDLoc DL(N);
793     if (N->getValueType(0) == MVT::i128) {
794       RC = CurDAG->getTargetConstant(AMDGPU::SGPR_128RegClassID, DL, MVT::i32);
795       SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0_sub1, DL, MVT::i32);
796       SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub2_sub3, DL, MVT::i32);
797     } else if (N->getValueType(0) == MVT::i64) {
798       RC = CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32);
799       SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
800       SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
801     } else {
802       llvm_unreachable("Unhandled value type for BUILD_PAIR");
803     }
804     const SDValue Ops[] = { RC, N->getOperand(0), SubReg0,
805                             N->getOperand(1), SubReg1 };
806     ReplaceNode(N, CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL,
807                                           N->getValueType(0), Ops));
808     return;
809   }
810 
811   case ISD::Constant:
812   case ISD::ConstantFP: {
813     if (N->getValueType(0).getSizeInBits() != 64 || isInlineImmediate(N))
814       break;
815 
816     uint64_t Imm;
817     if (ConstantFPSDNode *FP = dyn_cast<ConstantFPSDNode>(N))
818       Imm = FP->getValueAPF().bitcastToAPInt().getZExtValue();
819     else {
820       ConstantSDNode *C = cast<ConstantSDNode>(N);
821       Imm = C->getZExtValue();
822     }
823 
824     SDLoc DL(N);
825     ReplaceNode(N, buildSMovImm64(DL, Imm, N->getValueType(0)));
826     return;
827   }
828   case AMDGPUISD::BFE_I32:
829   case AMDGPUISD::BFE_U32: {
830     // There is a scalar version available, but unlike the vector version which
831     // has a separate operand for the offset and width, the scalar version packs
832     // the width and offset into a single operand. Try to move to the scalar
833     // version if the offsets are constant, so that we can try to keep extended
834     // loads of kernel arguments in SGPRs.
835 
836     // TODO: Technically we could try to pattern match scalar bitshifts of
837     // dynamic values, but it's probably not useful.
838     ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
839     if (!Offset)
840       break;
841 
842     ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
843     if (!Width)
844       break;
845 
846     bool Signed = Opc == AMDGPUISD::BFE_I32;
847 
848     uint32_t OffsetVal = Offset->getZExtValue();
849     uint32_t WidthVal = Width->getZExtValue();
850 
851     ReplaceNode(N, getS_BFE(Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32,
852                             SDLoc(N), N->getOperand(0), OffsetVal, WidthVal));
853     return;
854   }
855   case AMDGPUISD::DIV_SCALE: {
856     SelectDIV_SCALE(N);
857     return;
858   }
859   case AMDGPUISD::DIV_FMAS: {
860     SelectDIV_FMAS(N);
861     return;
862   }
863   case AMDGPUISD::MAD_I64_I32:
864   case AMDGPUISD::MAD_U64_U32: {
865     SelectMAD_64_32(N);
866     return;
867   }
868   case ISD::CopyToReg: {
869     const SITargetLowering& Lowering =
870       *static_cast<const SITargetLowering*>(getTargetLowering());
871     N = Lowering.legalizeTargetIndependentNode(N, *CurDAG);
872     break;
873   }
874   case ISD::AND:
875   case ISD::SRL:
876   case ISD::SRA:
877   case ISD::SIGN_EXTEND_INREG:
878     if (N->getValueType(0) != MVT::i32)
879       break;
880 
881     SelectS_BFE(N);
882     return;
883   case ISD::BRCOND:
884     SelectBRCOND(N);
885     return;
886   case ISD::FMAD:
887   case ISD::FMA:
888     SelectFMAD_FMA(N);
889     return;
890   case AMDGPUISD::ATOMIC_CMP_SWAP:
891     SelectATOMIC_CMP_SWAP(N);
892     return;
893   case AMDGPUISD::CVT_PKRTZ_F16_F32:
894   case AMDGPUISD::CVT_PKNORM_I16_F32:
895   case AMDGPUISD::CVT_PKNORM_U16_F32:
896   case AMDGPUISD::CVT_PK_U16_U32:
897   case AMDGPUISD::CVT_PK_I16_I32: {
898     // Hack around using a legal type if f16 is illegal.
899     if (N->getValueType(0) == MVT::i32) {
900       MVT NewVT = Opc == AMDGPUISD::CVT_PKRTZ_F16_F32 ? MVT::v2f16 : MVT::v2i16;
901       N = CurDAG->MorphNodeTo(N, N->getOpcode(), CurDAG->getVTList(NewVT),
902                               { N->getOperand(0), N->getOperand(1) });
903       SelectCode(N);
904       return;
905     }
906 
907     break;
908   }
909   case ISD::INTRINSIC_W_CHAIN: {
910     SelectINTRINSIC_W_CHAIN(N);
911     return;
912   }
913   case ISD::INTRINSIC_WO_CHAIN: {
914     SelectINTRINSIC_WO_CHAIN(N);
915     return;
916   }
917   case ISD::INTRINSIC_VOID: {
918     SelectINTRINSIC_VOID(N);
919     return;
920   }
921   }
922 
923   SelectCode(N);
924 }
925 
926 bool AMDGPUDAGToDAGISel::isUniformBr(const SDNode *N) const {
927   const BasicBlock *BB = FuncInfo->MBB->getBasicBlock();
928   const Instruction *Term = BB->getTerminator();
929   return Term->getMetadata("amdgpu.uniform") ||
930          Term->getMetadata("structurizecfg.uniform");
931 }
932 
933 StringRef AMDGPUDAGToDAGISel::getPassName() const {
934   return "AMDGPU DAG->DAG Pattern Instruction Selection";
935 }
936 
937 //===----------------------------------------------------------------------===//
938 // Complex Patterns
939 //===----------------------------------------------------------------------===//
940 
941 bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
942                                             SDValue &Offset) {
943   return false;
944 }
945 
946 bool AMDGPUDAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base,
947                                             SDValue &Offset) {
948   ConstantSDNode *C;
949   SDLoc DL(Addr);
950 
951   if ((C = dyn_cast<ConstantSDNode>(Addr))) {
952     Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
953     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
954   } else if ((Addr.getOpcode() == AMDGPUISD::DWORDADDR) &&
955              (C = dyn_cast<ConstantSDNode>(Addr.getOperand(0)))) {
956     Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
957     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
958   } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
959             (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) {
960     Base = Addr.getOperand(0);
961     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
962   } else {
963     Base = Addr;
964     Offset = CurDAG->getTargetConstant(0, DL, MVT::i32);
965   }
966 
967   return true;
968 }
969 
970 SDValue AMDGPUDAGToDAGISel::getMaterializedScalarImm32(int64_t Val,
971                                                        const SDLoc &DL) const {
972   SDNode *Mov = CurDAG->getMachineNode(
973     AMDGPU::S_MOV_B32, DL, MVT::i32,
974     CurDAG->getTargetConstant(Val, DL, MVT::i32));
975   return SDValue(Mov, 0);
976 }
977 
978 // FIXME: Should only handle addcarry/subcarry
979 void AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) {
980   SDLoc DL(N);
981   SDValue LHS = N->getOperand(0);
982   SDValue RHS = N->getOperand(1);
983 
984   unsigned Opcode = N->getOpcode();
985   bool ConsumeCarry = (Opcode == ISD::ADDE || Opcode == ISD::SUBE);
986   bool ProduceCarry =
987       ConsumeCarry || Opcode == ISD::ADDC || Opcode == ISD::SUBC;
988   bool IsAdd = Opcode == ISD::ADD || Opcode == ISD::ADDC || Opcode == ISD::ADDE;
989 
990   SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
991   SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
992 
993   SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
994                                        DL, MVT::i32, LHS, Sub0);
995   SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
996                                        DL, MVT::i32, LHS, Sub1);
997 
998   SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
999                                        DL, MVT::i32, RHS, Sub0);
1000   SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
1001                                        DL, MVT::i32, RHS, Sub1);
1002 
1003   SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue);
1004 
1005   unsigned Opc = IsAdd ? AMDGPU::S_ADD_U32 : AMDGPU::S_SUB_U32;
1006   unsigned CarryOpc = IsAdd ? AMDGPU::S_ADDC_U32 : AMDGPU::S_SUBB_U32;
1007 
1008   SDNode *AddLo;
1009   if (!ConsumeCarry) {
1010     SDValue Args[] = { SDValue(Lo0, 0), SDValue(Lo1, 0) };
1011     AddLo = CurDAG->getMachineNode(Opc, DL, VTList, Args);
1012   } else {
1013     SDValue Args[] = { SDValue(Lo0, 0), SDValue(Lo1, 0), N->getOperand(2) };
1014     AddLo = CurDAG->getMachineNode(CarryOpc, DL, VTList, Args);
1015   }
1016   SDValue AddHiArgs[] = {
1017     SDValue(Hi0, 0),
1018     SDValue(Hi1, 0),
1019     SDValue(AddLo, 1)
1020   };
1021   SDNode *AddHi = CurDAG->getMachineNode(CarryOpc, DL, VTList, AddHiArgs);
1022 
1023   SDValue RegSequenceArgs[] = {
1024     CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
1025     SDValue(AddLo,0),
1026     Sub0,
1027     SDValue(AddHi,0),
1028     Sub1,
1029   };
1030   SDNode *RegSequence = CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
1031                                                MVT::i64, RegSequenceArgs);
1032 
1033   if (ProduceCarry) {
1034     // Replace the carry-use
1035     ReplaceUses(SDValue(N, 1), SDValue(AddHi, 1));
1036   }
1037 
1038   // Replace the remaining uses.
1039   ReplaceNode(N, RegSequence);
1040 }
1041 
1042 void AMDGPUDAGToDAGISel::SelectAddcSubb(SDNode *N) {
1043   SDLoc DL(N);
1044   SDValue LHS = N->getOperand(0);
1045   SDValue RHS = N->getOperand(1);
1046   SDValue CI = N->getOperand(2);
1047 
1048   unsigned Opc = N->getOpcode() == ISD::ADDCARRY ? AMDGPU::V_ADDC_U32_e64
1049                                                  : AMDGPU::V_SUBB_U32_e64;
1050   CurDAG->SelectNodeTo(
1051       N, Opc, N->getVTList(),
1052       {LHS, RHS, CI, CurDAG->getTargetConstant(0, {}, MVT::i1) /*clamp bit*/});
1053 }
1054 
1055 void AMDGPUDAGToDAGISel::SelectUADDO_USUBO(SDNode *N) {
1056   // The name of the opcodes are misleading. v_add_i32/v_sub_i32 have unsigned
1057   // carry out despite the _i32 name. These were renamed in VI to _U32.
1058   // FIXME: We should probably rename the opcodes here.
1059   unsigned Opc = N->getOpcode() == ISD::UADDO ?
1060     AMDGPU::V_ADD_I32_e64 : AMDGPU::V_SUB_I32_e64;
1061 
1062   CurDAG->SelectNodeTo(
1063       N, Opc, N->getVTList(),
1064       {N->getOperand(0), N->getOperand(1),
1065        CurDAG->getTargetConstant(0, {}, MVT::i1) /*clamp bit*/});
1066 }
1067 
1068 void AMDGPUDAGToDAGISel::SelectFMA_W_CHAIN(SDNode *N) {
1069   SDLoc SL(N);
1070   //  src0_modifiers, src0,  src1_modifiers, src1, src2_modifiers, src2, clamp, omod
1071   SDValue Ops[10];
1072 
1073   SelectVOP3Mods0(N->getOperand(1), Ops[1], Ops[0], Ops[6], Ops[7]);
1074   SelectVOP3Mods(N->getOperand(2), Ops[3], Ops[2]);
1075   SelectVOP3Mods(N->getOperand(3), Ops[5], Ops[4]);
1076   Ops[8] = N->getOperand(0);
1077   Ops[9] = N->getOperand(4);
1078 
1079   CurDAG->SelectNodeTo(N, AMDGPU::V_FMA_F32, N->getVTList(), Ops);
1080 }
1081 
1082 void AMDGPUDAGToDAGISel::SelectFMUL_W_CHAIN(SDNode *N) {
1083   SDLoc SL(N);
1084   //    src0_modifiers, src0,  src1_modifiers, src1, clamp, omod
1085   SDValue Ops[8];
1086 
1087   SelectVOP3Mods0(N->getOperand(1), Ops[1], Ops[0], Ops[4], Ops[5]);
1088   SelectVOP3Mods(N->getOperand(2), Ops[3], Ops[2]);
1089   Ops[6] = N->getOperand(0);
1090   Ops[7] = N->getOperand(3);
1091 
1092   CurDAG->SelectNodeTo(N, AMDGPU::V_MUL_F32_e64, N->getVTList(), Ops);
1093 }
1094 
1095 // We need to handle this here because tablegen doesn't support matching
1096 // instructions with multiple outputs.
1097 void AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) {
1098   SDLoc SL(N);
1099   EVT VT = N->getValueType(0);
1100 
1101   assert(VT == MVT::f32 || VT == MVT::f64);
1102 
1103   unsigned Opc
1104     = (VT == MVT::f64) ? AMDGPU::V_DIV_SCALE_F64 : AMDGPU::V_DIV_SCALE_F32;
1105 
1106   SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2) };
1107   CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
1108 }
1109 
1110 void AMDGPUDAGToDAGISel::SelectDIV_FMAS(SDNode *N) {
1111   const GCNSubtarget *ST = static_cast<const GCNSubtarget *>(Subtarget);
1112   const SIRegisterInfo *TRI = ST->getRegisterInfo();
1113 
1114   SDLoc SL(N);
1115   EVT VT = N->getValueType(0);
1116 
1117   assert(VT == MVT::f32 || VT == MVT::f64);
1118 
1119   unsigned Opc
1120     = (VT == MVT::f64) ? AMDGPU::V_DIV_FMAS_F64 : AMDGPU::V_DIV_FMAS_F32;
1121 
1122   SDValue CarryIn = N->getOperand(3);
1123   // V_DIV_FMAS implicitly reads VCC.
1124   SDValue VCC = CurDAG->getCopyToReg(CurDAG->getEntryNode(), SL,
1125                                      TRI->getVCC(), CarryIn, SDValue());
1126 
1127   SDValue Ops[10];
1128 
1129   SelectVOP3Mods0(N->getOperand(0), Ops[1], Ops[0], Ops[6], Ops[7]);
1130   SelectVOP3Mods(N->getOperand(1), Ops[3], Ops[2]);
1131   SelectVOP3Mods(N->getOperand(2), Ops[5], Ops[4]);
1132 
1133   Ops[8] = VCC;
1134   Ops[9] = VCC.getValue(1);
1135 
1136   CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
1137 }
1138 
1139 // We need to handle this here because tablegen doesn't support matching
1140 // instructions with multiple outputs.
1141 void AMDGPUDAGToDAGISel::SelectMAD_64_32(SDNode *N) {
1142   SDLoc SL(N);
1143   bool Signed = N->getOpcode() == AMDGPUISD::MAD_I64_I32;
1144   unsigned Opc = Signed ? AMDGPU::V_MAD_I64_I32 : AMDGPU::V_MAD_U64_U32;
1145 
1146   SDValue Clamp = CurDAG->getTargetConstant(0, SL, MVT::i1);
1147   SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
1148                     Clamp };
1149   CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
1150 }
1151 
1152 bool AMDGPUDAGToDAGISel::isDSOffsetLegal(SDValue Base, unsigned Offset,
1153                                          unsigned OffsetBits) const {
1154   if ((OffsetBits == 16 && !isUInt<16>(Offset)) ||
1155       (OffsetBits == 8 && !isUInt<8>(Offset)))
1156     return false;
1157 
1158   if (Subtarget->hasUsableDSOffset() ||
1159       Subtarget->unsafeDSOffsetFoldingEnabled())
1160     return true;
1161 
1162   // On Southern Islands instruction with a negative base value and an offset
1163   // don't seem to work.
1164   return CurDAG->SignBitIsZero(Base);
1165 }
1166 
1167 bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base,
1168                                               SDValue &Offset) const {
1169   SDLoc DL(Addr);
1170   if (CurDAG->isBaseWithConstantOffset(Addr)) {
1171     SDValue N0 = Addr.getOperand(0);
1172     SDValue N1 = Addr.getOperand(1);
1173     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
1174     if (isDSOffsetLegal(N0, C1->getSExtValue(), 16)) {
1175       // (add n0, c0)
1176       Base = N0;
1177       Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16);
1178       return true;
1179     }
1180   } else if (Addr.getOpcode() == ISD::SUB) {
1181     // sub C, x -> add (sub 0, x), C
1182     if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr.getOperand(0))) {
1183       int64_t ByteOffset = C->getSExtValue();
1184       if (isUInt<16>(ByteOffset)) {
1185         SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
1186 
1187         // XXX - This is kind of hacky. Create a dummy sub node so we can check
1188         // the known bits in isDSOffsetLegal. We need to emit the selected node
1189         // here, so this is thrown away.
1190         SDValue Sub = CurDAG->getNode(ISD::SUB, DL, MVT::i32,
1191                                       Zero, Addr.getOperand(1));
1192 
1193         if (isDSOffsetLegal(Sub, ByteOffset, 16)) {
1194           SmallVector<SDValue, 3> Opnds;
1195           Opnds.push_back(Zero);
1196           Opnds.push_back(Addr.getOperand(1));
1197 
1198           // FIXME: Select to VOP3 version for with-carry.
1199           unsigned SubOp = AMDGPU::V_SUB_I32_e32;
1200           if (Subtarget->hasAddNoCarry()) {
1201             SubOp = AMDGPU::V_SUB_U32_e64;
1202             Opnds.push_back(
1203                 CurDAG->getTargetConstant(0, {}, MVT::i1)); // clamp bit
1204           }
1205 
1206           MachineSDNode *MachineSub =
1207               CurDAG->getMachineNode(SubOp, DL, MVT::i32, Opnds);
1208 
1209           Base = SDValue(MachineSub, 0);
1210           Offset = CurDAG->getTargetConstant(ByteOffset, DL, MVT::i16);
1211           return true;
1212         }
1213       }
1214     }
1215   } else if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
1216     // If we have a constant address, prefer to put the constant into the
1217     // offset. This can save moves to load the constant address since multiple
1218     // operations can share the zero base address register, and enables merging
1219     // into read2 / write2 instructions.
1220 
1221     SDLoc DL(Addr);
1222 
1223     if (isUInt<16>(CAddr->getZExtValue())) {
1224       SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
1225       MachineSDNode *MovZero = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
1226                                  DL, MVT::i32, Zero);
1227       Base = SDValue(MovZero, 0);
1228       Offset = CurDAG->getTargetConstant(CAddr->getZExtValue(), DL, MVT::i16);
1229       return true;
1230     }
1231   }
1232 
1233   // default case
1234   Base = Addr;
1235   Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i16);
1236   return true;
1237 }
1238 
1239 // TODO: If offset is too big, put low 16-bit into offset.
1240 bool AMDGPUDAGToDAGISel::SelectDS64Bit4ByteAligned(SDValue Addr, SDValue &Base,
1241                                                    SDValue &Offset0,
1242                                                    SDValue &Offset1) const {
1243   SDLoc DL(Addr);
1244 
1245   if (CurDAG->isBaseWithConstantOffset(Addr)) {
1246     SDValue N0 = Addr.getOperand(0);
1247     SDValue N1 = Addr.getOperand(1);
1248     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
1249     unsigned DWordOffset0 = C1->getZExtValue() / 4;
1250     unsigned DWordOffset1 = DWordOffset0 + 1;
1251     // (add n0, c0)
1252     if (isDSOffsetLegal(N0, DWordOffset1, 8)) {
1253       Base = N0;
1254       Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8);
1255       Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8);
1256       return true;
1257     }
1258   } else if (Addr.getOpcode() == ISD::SUB) {
1259     // sub C, x -> add (sub 0, x), C
1260     if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr.getOperand(0))) {
1261       unsigned DWordOffset0 = C->getZExtValue() / 4;
1262       unsigned DWordOffset1 = DWordOffset0 + 1;
1263 
1264       if (isUInt<8>(DWordOffset0)) {
1265         SDLoc DL(Addr);
1266         SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
1267 
1268         // XXX - This is kind of hacky. Create a dummy sub node so we can check
1269         // the known bits in isDSOffsetLegal. We need to emit the selected node
1270         // here, so this is thrown away.
1271         SDValue Sub = CurDAG->getNode(ISD::SUB, DL, MVT::i32,
1272                                       Zero, Addr.getOperand(1));
1273 
1274         if (isDSOffsetLegal(Sub, DWordOffset1, 8)) {
1275           SmallVector<SDValue, 3> Opnds;
1276           Opnds.push_back(Zero);
1277           Opnds.push_back(Addr.getOperand(1));
1278           unsigned SubOp = AMDGPU::V_SUB_I32_e32;
1279           if (Subtarget->hasAddNoCarry()) {
1280             SubOp = AMDGPU::V_SUB_U32_e64;
1281             Opnds.push_back(
1282                 CurDAG->getTargetConstant(0, {}, MVT::i1)); // clamp bit
1283           }
1284 
1285           MachineSDNode *MachineSub
1286             = CurDAG->getMachineNode(SubOp, DL, MVT::i32, Opnds);
1287 
1288           Base = SDValue(MachineSub, 0);
1289           Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8);
1290           Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8);
1291           return true;
1292         }
1293       }
1294     }
1295   } else if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
1296     unsigned DWordOffset0 = CAddr->getZExtValue() / 4;
1297     unsigned DWordOffset1 = DWordOffset0 + 1;
1298     assert(4 * DWordOffset0 == CAddr->getZExtValue());
1299 
1300     if (isUInt<8>(DWordOffset0) && isUInt<8>(DWordOffset1)) {
1301       SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
1302       MachineSDNode *MovZero
1303         = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
1304                                  DL, MVT::i32, Zero);
1305       Base = SDValue(MovZero, 0);
1306       Offset0 = CurDAG->getTargetConstant(DWordOffset0, DL, MVT::i8);
1307       Offset1 = CurDAG->getTargetConstant(DWordOffset1, DL, MVT::i8);
1308       return true;
1309     }
1310   }
1311 
1312   // default case
1313 
1314   Base = Addr;
1315   Offset0 = CurDAG->getTargetConstant(0, DL, MVT::i8);
1316   Offset1 = CurDAG->getTargetConstant(1, DL, MVT::i8);
1317   return true;
1318 }
1319 
1320 bool AMDGPUDAGToDAGISel::SelectMUBUF(SDValue Addr, SDValue &Ptr,
1321                                      SDValue &VAddr, SDValue &SOffset,
1322                                      SDValue &Offset, SDValue &Offen,
1323                                      SDValue &Idxen, SDValue &Addr64,
1324                                      SDValue &GLC, SDValue &SLC,
1325                                      SDValue &TFE, SDValue &DLC,
1326                                      SDValue &SWZ) const {
1327   // Subtarget prefers to use flat instruction
1328   if (Subtarget->useFlatForGlobal())
1329     return false;
1330 
1331   SDLoc DL(Addr);
1332 
1333   if (!GLC.getNode())
1334     GLC = CurDAG->getTargetConstant(0, DL, MVT::i1);
1335   if (!SLC.getNode())
1336     SLC = CurDAG->getTargetConstant(0, DL, MVT::i1);
1337   TFE = CurDAG->getTargetConstant(0, DL, MVT::i1);
1338   DLC = CurDAG->getTargetConstant(0, DL, MVT::i1);
1339   SWZ = CurDAG->getTargetConstant(0, DL, MVT::i1);
1340 
1341   Idxen = CurDAG->getTargetConstant(0, DL, MVT::i1);
1342   Offen = CurDAG->getTargetConstant(0, DL, MVT::i1);
1343   Addr64 = CurDAG->getTargetConstant(0, DL, MVT::i1);
1344   SOffset = CurDAG->getTargetConstant(0, DL, MVT::i32);
1345 
1346   ConstantSDNode *C1 = nullptr;
1347   SDValue N0 = Addr;
1348   if (CurDAG->isBaseWithConstantOffset(Addr)) {
1349     C1 = cast<ConstantSDNode>(Addr.getOperand(1));
1350     if (isUInt<32>(C1->getZExtValue()))
1351       N0 = Addr.getOperand(0);
1352     else
1353       C1 = nullptr;
1354   }
1355 
1356   if (N0.getOpcode() == ISD::ADD) {
1357     // (add N2, N3) -> addr64, or
1358     // (add (add N2, N3), C1) -> addr64
1359     SDValue N2 = N0.getOperand(0);
1360     SDValue N3 = N0.getOperand(1);
1361     Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1);
1362 
1363     if (N2->isDivergent()) {
1364       if (N3->isDivergent()) {
1365         // Both N2 and N3 are divergent. Use N0 (the result of the add) as the
1366         // addr64, and construct the resource from a 0 address.
1367         Ptr = SDValue(buildSMovImm64(DL, 0, MVT::v2i32), 0);
1368         VAddr = N0;
1369       } else {
1370         // N2 is divergent, N3 is not.
1371         Ptr = N3;
1372         VAddr = N2;
1373       }
1374     } else {
1375       // N2 is not divergent.
1376       Ptr = N2;
1377       VAddr = N3;
1378     }
1379     Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1380   } else if (N0->isDivergent()) {
1381     // N0 is divergent. Use it as the addr64, and construct the resource from a
1382     // 0 address.
1383     Ptr = SDValue(buildSMovImm64(DL, 0, MVT::v2i32), 0);
1384     VAddr = N0;
1385     Addr64 = CurDAG->getTargetConstant(1, DL, MVT::i1);
1386   } else {
1387     // N0 -> offset, or
1388     // (N0 + C1) -> offset
1389     VAddr = CurDAG->getTargetConstant(0, DL, MVT::i32);
1390     Ptr = N0;
1391   }
1392 
1393   if (!C1) {
1394     // No offset.
1395     Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1396     return true;
1397   }
1398 
1399   if (SIInstrInfo::isLegalMUBUFImmOffset(C1->getZExtValue())) {
1400     // Legal offset for instruction.
1401     Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16);
1402     return true;
1403   }
1404 
1405   // Illegal offset, store it in soffset.
1406   Offset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1407   SOffset =
1408       SDValue(CurDAG->getMachineNode(
1409                   AMDGPU::S_MOV_B32, DL, MVT::i32,
1410                   CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i32)),
1411               0);
1412   return true;
1413 }
1414 
1415 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
1416                                            SDValue &VAddr, SDValue &SOffset,
1417                                            SDValue &Offset, SDValue &GLC,
1418                                            SDValue &SLC, SDValue &TFE,
1419                                            SDValue &DLC, SDValue &SWZ) const {
1420   SDValue Ptr, Offen, Idxen, Addr64;
1421 
1422   // addr64 bit was removed for volcanic islands.
1423   if (!Subtarget->hasAddr64())
1424     return false;
1425 
1426   if (!SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
1427               GLC, SLC, TFE, DLC, SWZ))
1428     return false;
1429 
1430   ConstantSDNode *C = cast<ConstantSDNode>(Addr64);
1431   if (C->getSExtValue()) {
1432     SDLoc DL(Addr);
1433 
1434     const SITargetLowering& Lowering =
1435       *static_cast<const SITargetLowering*>(getTargetLowering());
1436 
1437     SRsrc = SDValue(Lowering.wrapAddr64Rsrc(*CurDAG, DL, Ptr), 0);
1438     return true;
1439   }
1440 
1441   return false;
1442 }
1443 
1444 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
1445                                            SDValue &VAddr, SDValue &SOffset,
1446                                            SDValue &Offset,
1447                                            SDValue &SLC) const {
1448   SLC = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i1);
1449   SDValue GLC, TFE, DLC, SWZ;
1450 
1451   return SelectMUBUFAddr64(Addr, SRsrc, VAddr, SOffset, Offset, GLC, SLC, TFE, DLC, SWZ);
1452 }
1453 
1454 static bool isStackPtrRelative(const MachinePointerInfo &PtrInfo) {
1455   auto PSV = PtrInfo.V.dyn_cast<const PseudoSourceValue *>();
1456   return PSV && PSV->isStack();
1457 }
1458 
1459 std::pair<SDValue, SDValue> AMDGPUDAGToDAGISel::foldFrameIndex(SDValue N) const {
1460   const MachineFunction &MF = CurDAG->getMachineFunction();
1461   const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1462 
1463   if (auto FI = dyn_cast<FrameIndexSDNode>(N)) {
1464     SDValue TFI = CurDAG->getTargetFrameIndex(FI->getIndex(),
1465                                               FI->getValueType(0));
1466 
1467     // If we can resolve this to a frame index access, this will be relative to
1468     // either the stack or frame pointer SGPR.
1469     return std::make_pair(
1470         TFI, CurDAG->getRegister(Info->getStackPtrOffsetReg(), MVT::i32));
1471   }
1472 
1473   // If we don't know this private access is a local stack object, it needs to
1474   // be relative to the entry point's scratch wave offset register.
1475   return std::make_pair(N, CurDAG->getRegister(Info->getScratchWaveOffsetReg(),
1476                                                MVT::i32));
1477 }
1478 
1479 bool AMDGPUDAGToDAGISel::SelectMUBUFScratchOffen(SDNode *Parent,
1480                                                  SDValue Addr, SDValue &Rsrc,
1481                                                  SDValue &VAddr, SDValue &SOffset,
1482                                                  SDValue &ImmOffset) const {
1483 
1484   SDLoc DL(Addr);
1485   MachineFunction &MF = CurDAG->getMachineFunction();
1486   const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1487 
1488   Rsrc = CurDAG->getRegister(Info->getScratchRSrcReg(), MVT::v4i32);
1489 
1490   if (ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
1491     unsigned Imm = CAddr->getZExtValue();
1492 
1493     SDValue HighBits = CurDAG->getTargetConstant(Imm & ~4095, DL, MVT::i32);
1494     MachineSDNode *MovHighBits = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
1495                                                         DL, MVT::i32, HighBits);
1496     VAddr = SDValue(MovHighBits, 0);
1497 
1498     // In a call sequence, stores to the argument stack area are relative to the
1499     // stack pointer.
1500     const MachinePointerInfo &PtrInfo = cast<MemSDNode>(Parent)->getPointerInfo();
1501     unsigned SOffsetReg = isStackPtrRelative(PtrInfo) ?
1502       Info->getStackPtrOffsetReg() : Info->getScratchWaveOffsetReg();
1503 
1504     SOffset = CurDAG->getRegister(SOffsetReg, MVT::i32);
1505     ImmOffset = CurDAG->getTargetConstant(Imm & 4095, DL, MVT::i16);
1506     return true;
1507   }
1508 
1509   if (CurDAG->isBaseWithConstantOffset(Addr)) {
1510     // (add n0, c1)
1511 
1512     SDValue N0 = Addr.getOperand(0);
1513     SDValue N1 = Addr.getOperand(1);
1514 
1515     // Offsets in vaddr must be positive if range checking is enabled.
1516     //
1517     // The total computation of vaddr + soffset + offset must not overflow.  If
1518     // vaddr is negative, even if offset is 0 the sgpr offset add will end up
1519     // overflowing.
1520     //
1521     // Prior to gfx9, MUBUF instructions with the vaddr offset enabled would
1522     // always perform a range check. If a negative vaddr base index was used,
1523     // this would fail the range check. The overall address computation would
1524     // compute a valid address, but this doesn't happen due to the range
1525     // check. For out-of-bounds MUBUF loads, a 0 is returned.
1526     //
1527     // Therefore it should be safe to fold any VGPR offset on gfx9 into the
1528     // MUBUF vaddr, but not on older subtargets which can only do this if the
1529     // sign bit is known 0.
1530     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
1531     if (SIInstrInfo::isLegalMUBUFImmOffset(C1->getZExtValue()) &&
1532         (!Subtarget->privateMemoryResourceIsRangeChecked() ||
1533          CurDAG->SignBitIsZero(N0))) {
1534       std::tie(VAddr, SOffset) = foldFrameIndex(N0);
1535       ImmOffset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i16);
1536       return true;
1537     }
1538   }
1539 
1540   // (node)
1541   std::tie(VAddr, SOffset) = foldFrameIndex(Addr);
1542   ImmOffset = CurDAG->getTargetConstant(0, DL, MVT::i16);
1543   return true;
1544 }
1545 
1546 bool AMDGPUDAGToDAGISel::SelectMUBUFScratchOffset(SDNode *Parent,
1547                                                   SDValue Addr,
1548                                                   SDValue &SRsrc,
1549                                                   SDValue &SOffset,
1550                                                   SDValue &Offset) const {
1551   ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr);
1552   if (!CAddr || !SIInstrInfo::isLegalMUBUFImmOffset(CAddr->getZExtValue()))
1553     return false;
1554 
1555   SDLoc DL(Addr);
1556   MachineFunction &MF = CurDAG->getMachineFunction();
1557   const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1558 
1559   SRsrc = CurDAG->getRegister(Info->getScratchRSrcReg(), MVT::v4i32);
1560 
1561   const MachinePointerInfo &PtrInfo = cast<MemSDNode>(Parent)->getPointerInfo();
1562   unsigned SOffsetReg = isStackPtrRelative(PtrInfo) ?
1563     Info->getStackPtrOffsetReg() : Info->getScratchWaveOffsetReg();
1564 
1565   // FIXME: Get from MachinePointerInfo? We should only be using the frame
1566   // offset if we know this is in a call sequence.
1567   SOffset = CurDAG->getRegister(SOffsetReg, MVT::i32);
1568 
1569   Offset = CurDAG->getTargetConstant(CAddr->getZExtValue(), DL, MVT::i16);
1570   return true;
1571 }
1572 
1573 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1574                                            SDValue &SOffset, SDValue &Offset,
1575                                            SDValue &GLC, SDValue &SLC,
1576                                            SDValue &TFE, SDValue &DLC,
1577                                            SDValue &SWZ) const {
1578   SDValue Ptr, VAddr, Offen, Idxen, Addr64;
1579   const SIInstrInfo *TII =
1580     static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
1581 
1582   if (!SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
1583               GLC, SLC, TFE, DLC, SWZ))
1584     return false;
1585 
1586   if (!cast<ConstantSDNode>(Offen)->getSExtValue() &&
1587       !cast<ConstantSDNode>(Idxen)->getSExtValue() &&
1588       !cast<ConstantSDNode>(Addr64)->getSExtValue()) {
1589     uint64_t Rsrc = TII->getDefaultRsrcDataFormat() |
1590                     APInt::getAllOnesValue(32).getZExtValue(); // Size
1591     SDLoc DL(Addr);
1592 
1593     const SITargetLowering& Lowering =
1594       *static_cast<const SITargetLowering*>(getTargetLowering());
1595 
1596     SRsrc = SDValue(Lowering.buildRSRC(*CurDAG, DL, Ptr, 0, Rsrc), 0);
1597     return true;
1598   }
1599   return false;
1600 }
1601 
1602 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1603                                            SDValue &Soffset, SDValue &Offset
1604                                            ) const {
1605   SDValue GLC, SLC, TFE, DLC, SWZ;
1606 
1607   return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE, DLC, SWZ);
1608 }
1609 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1610                                            SDValue &Soffset, SDValue &Offset,
1611                                            SDValue &SLC) const {
1612   SDValue GLC, TFE, DLC, SWZ;
1613 
1614   return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE, DLC, SWZ);
1615 }
1616 
1617 // Find a load or store from corresponding pattern root.
1618 // Roots may be build_vector, bitconvert or their combinations.
1619 static MemSDNode* findMemSDNode(SDNode *N) {
1620   N = AMDGPUTargetLowering::stripBitcast(SDValue(N,0)).getNode();
1621   if (MemSDNode *MN = dyn_cast<MemSDNode>(N))
1622     return MN;
1623   assert(isa<BuildVectorSDNode>(N));
1624   for (SDValue V : N->op_values())
1625     if (MemSDNode *MN =
1626           dyn_cast<MemSDNode>(AMDGPUTargetLowering::stripBitcast(V)))
1627       return MN;
1628   llvm_unreachable("cannot find MemSDNode in the pattern!");
1629 }
1630 
1631 template <bool IsSigned>
1632 bool AMDGPUDAGToDAGISel::SelectFlatOffset(SDNode *N,
1633                                           SDValue Addr,
1634                                           SDValue &VAddr,
1635                                           SDValue &Offset,
1636                                           SDValue &SLC) const {
1637   int64_t OffsetVal = 0;
1638 
1639   if (Subtarget->hasFlatInstOffsets() &&
1640       (!Subtarget->hasFlatSegmentOffsetBug() ||
1641        findMemSDNode(N)->getAddressSpace() != AMDGPUAS::FLAT_ADDRESS) &&
1642       CurDAG->isBaseWithConstantOffset(Addr)) {
1643     SDValue N0 = Addr.getOperand(0);
1644     SDValue N1 = Addr.getOperand(1);
1645     uint64_t COffsetVal = cast<ConstantSDNode>(N1)->getSExtValue();
1646 
1647     const SIInstrInfo *TII = Subtarget->getInstrInfo();
1648     unsigned AS = findMemSDNode(N)->getAddressSpace();
1649     if (TII->isLegalFLATOffset(COffsetVal, AS, IsSigned)) {
1650       Addr = N0;
1651       OffsetVal = COffsetVal;
1652     } else {
1653       // If the offset doesn't fit, put the low bits into the offset field and
1654       // add the rest.
1655 
1656       SDLoc DL(N);
1657       uint64_t ImmField;
1658       const unsigned NumBits = TII->getNumFlatOffsetBits(AS, IsSigned);
1659       if (IsSigned) {
1660         ImmField = SignExtend64(COffsetVal, NumBits);
1661 
1662         // Don't use a negative offset field if the base offset is positive.
1663         // Since the scheduler currently relies on the offset field, doing so
1664         // could result in strange scheduling decisions.
1665 
1666         // TODO: Should we not do this in the opposite direction as well?
1667         if (static_cast<int64_t>(COffsetVal) > 0) {
1668           if (static_cast<int64_t>(ImmField) < 0) {
1669             const uint64_t OffsetMask = maskTrailingOnes<uint64_t>(NumBits - 1);
1670             ImmField = COffsetVal & OffsetMask;
1671           }
1672         }
1673       } else {
1674         // TODO: Should we do this for a negative offset?
1675         const uint64_t OffsetMask = maskTrailingOnes<uint64_t>(NumBits);
1676         ImmField = COffsetVal & OffsetMask;
1677       }
1678 
1679       uint64_t RemainderOffset = COffsetVal - ImmField;
1680 
1681       assert(TII->isLegalFLATOffset(ImmField, AS, IsSigned));
1682       assert(RemainderOffset + ImmField == COffsetVal);
1683 
1684       OffsetVal = ImmField;
1685 
1686       // TODO: Should this try to use a scalar add pseudo if the base address is
1687       // uniform and saddr is usable?
1688       SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
1689       SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
1690 
1691       SDNode *N0Lo = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
1692                                             DL, MVT::i32, N0, Sub0);
1693       SDNode *N0Hi = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
1694                                             DL, MVT::i32, N0, Sub1);
1695 
1696       SDValue AddOffsetLo
1697         = getMaterializedScalarImm32(Lo_32(RemainderOffset), DL);
1698       SDValue AddOffsetHi
1699         = getMaterializedScalarImm32(Hi_32(RemainderOffset), DL);
1700 
1701       SDVTList VTs = CurDAG->getVTList(MVT::i32, MVT::i1);
1702       SDValue Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1);
1703 
1704       SDNode *Add = CurDAG->getMachineNode(
1705         AMDGPU::V_ADD_I32_e64, DL, VTs,
1706         {AddOffsetLo, SDValue(N0Lo, 0), Clamp});
1707 
1708       SDNode *Addc = CurDAG->getMachineNode(
1709         AMDGPU::V_ADDC_U32_e64, DL, VTs,
1710         {AddOffsetHi, SDValue(N0Hi, 0), SDValue(Add, 1), Clamp});
1711 
1712       SDValue RegSequenceArgs[] = {
1713         CurDAG->getTargetConstant(AMDGPU::VReg_64RegClassID, DL, MVT::i32),
1714         SDValue(Add, 0), Sub0, SDValue(Addc, 0), Sub1
1715       };
1716 
1717       Addr = SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
1718                                             MVT::i64, RegSequenceArgs), 0);
1719     }
1720   }
1721 
1722   VAddr = Addr;
1723   Offset = CurDAG->getTargetConstant(OffsetVal, SDLoc(), MVT::i16);
1724   SLC = CurDAG->getTargetConstant(0, SDLoc(), MVT::i1);
1725   return true;
1726 }
1727 
1728 bool AMDGPUDAGToDAGISel::SelectFlatAtomic(SDNode *N,
1729                                           SDValue Addr,
1730                                           SDValue &VAddr,
1731                                           SDValue &Offset,
1732                                           SDValue &SLC) const {
1733   return SelectFlatOffset<false>(N, Addr, VAddr, Offset, SLC);
1734 }
1735 
1736 bool AMDGPUDAGToDAGISel::SelectFlatAtomicSigned(SDNode *N,
1737                                                 SDValue Addr,
1738                                                 SDValue &VAddr,
1739                                                 SDValue &Offset,
1740                                                 SDValue &SLC) const {
1741   return SelectFlatOffset<true>(N, Addr, VAddr, Offset, SLC);
1742 }
1743 
1744 bool AMDGPUDAGToDAGISel::SelectSMRDOffset(SDValue ByteOffsetNode,
1745                                           SDValue &Offset, bool &Imm) const {
1746 
1747   // FIXME: Handle non-constant offsets.
1748   ConstantSDNode *C = dyn_cast<ConstantSDNode>(ByteOffsetNode);
1749   if (!C)
1750     return false;
1751 
1752   SDLoc SL(ByteOffsetNode);
1753   GCNSubtarget::Generation Gen = Subtarget->getGeneration();
1754   int64_t ByteOffset = C->getSExtValue();
1755   int64_t EncodedOffset = AMDGPU::getSMRDEncodedOffset(*Subtarget, ByteOffset);
1756 
1757   if (AMDGPU::isLegalSMRDImmOffset(*Subtarget, ByteOffset)) {
1758     Offset = CurDAG->getTargetConstant(EncodedOffset, SL, MVT::i32);
1759     Imm = true;
1760     return true;
1761   }
1762 
1763   if (!isUInt<32>(EncodedOffset) || !isUInt<32>(ByteOffset))
1764     return false;
1765 
1766   if (Gen == AMDGPUSubtarget::SEA_ISLANDS && isUInt<32>(EncodedOffset)) {
1767     // 32-bit Immediates are supported on Sea Islands.
1768     Offset = CurDAG->getTargetConstant(EncodedOffset, SL, MVT::i32);
1769   } else {
1770     SDValue C32Bit = CurDAG->getTargetConstant(ByteOffset, SL, MVT::i32);
1771     Offset = SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SL, MVT::i32,
1772                                             C32Bit), 0);
1773   }
1774   Imm = false;
1775   return true;
1776 }
1777 
1778 SDValue AMDGPUDAGToDAGISel::Expand32BitAddress(SDValue Addr) const {
1779   if (Addr.getValueType() != MVT::i32)
1780     return Addr;
1781 
1782   // Zero-extend a 32-bit address.
1783   SDLoc SL(Addr);
1784 
1785   const MachineFunction &MF = CurDAG->getMachineFunction();
1786   const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
1787   unsigned AddrHiVal = Info->get32BitAddressHighBits();
1788   SDValue AddrHi = CurDAG->getTargetConstant(AddrHiVal, SL, MVT::i32);
1789 
1790   const SDValue Ops[] = {
1791     CurDAG->getTargetConstant(AMDGPU::SReg_64_XEXECRegClassID, SL, MVT::i32),
1792     Addr,
1793     CurDAG->getTargetConstant(AMDGPU::sub0, SL, MVT::i32),
1794     SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SL, MVT::i32, AddrHi),
1795             0),
1796     CurDAG->getTargetConstant(AMDGPU::sub1, SL, MVT::i32),
1797   };
1798 
1799   return SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, SL, MVT::i64,
1800                                         Ops), 0);
1801 }
1802 
1803 bool AMDGPUDAGToDAGISel::SelectSMRD(SDValue Addr, SDValue &SBase,
1804                                      SDValue &Offset, bool &Imm) const {
1805   SDLoc SL(Addr);
1806 
1807   // A 32-bit (address + offset) should not cause unsigned 32-bit integer
1808   // wraparound, because s_load instructions perform the addition in 64 bits.
1809   if ((Addr.getValueType() != MVT::i32 ||
1810        Addr->getFlags().hasNoUnsignedWrap()) &&
1811       CurDAG->isBaseWithConstantOffset(Addr)) {
1812     SDValue N0 = Addr.getOperand(0);
1813     SDValue N1 = Addr.getOperand(1);
1814 
1815     if (SelectSMRDOffset(N1, Offset, Imm)) {
1816       SBase = Expand32BitAddress(N0);
1817       return true;
1818     }
1819   }
1820   SBase = Expand32BitAddress(Addr);
1821   Offset = CurDAG->getTargetConstant(0, SL, MVT::i32);
1822   Imm = true;
1823   return true;
1824 }
1825 
1826 bool AMDGPUDAGToDAGISel::SelectSMRDImm(SDValue Addr, SDValue &SBase,
1827                                        SDValue &Offset) const {
1828   bool Imm;
1829   return SelectSMRD(Addr, SBase, Offset, Imm) && Imm;
1830 }
1831 
1832 bool AMDGPUDAGToDAGISel::SelectSMRDImm32(SDValue Addr, SDValue &SBase,
1833                                          SDValue &Offset) const {
1834 
1835   if (Subtarget->getGeneration() != AMDGPUSubtarget::SEA_ISLANDS)
1836     return false;
1837 
1838   bool Imm;
1839   if (!SelectSMRD(Addr, SBase, Offset, Imm))
1840     return false;
1841 
1842   return !Imm && isa<ConstantSDNode>(Offset);
1843 }
1844 
1845 bool AMDGPUDAGToDAGISel::SelectSMRDSgpr(SDValue Addr, SDValue &SBase,
1846                                         SDValue &Offset) const {
1847   bool Imm;
1848   return SelectSMRD(Addr, SBase, Offset, Imm) && !Imm &&
1849          !isa<ConstantSDNode>(Offset);
1850 }
1851 
1852 bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm(SDValue Addr,
1853                                              SDValue &Offset) const {
1854   bool Imm;
1855   return SelectSMRDOffset(Addr, Offset, Imm) && Imm;
1856 }
1857 
1858 bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm32(SDValue Addr,
1859                                                SDValue &Offset) const {
1860   if (Subtarget->getGeneration() != AMDGPUSubtarget::SEA_ISLANDS)
1861     return false;
1862 
1863   bool Imm;
1864   if (!SelectSMRDOffset(Addr, Offset, Imm))
1865     return false;
1866 
1867   return !Imm && isa<ConstantSDNode>(Offset);
1868 }
1869 
1870 bool AMDGPUDAGToDAGISel::SelectMOVRELOffset(SDValue Index,
1871                                             SDValue &Base,
1872                                             SDValue &Offset) const {
1873   SDLoc DL(Index);
1874 
1875   if (CurDAG->isBaseWithConstantOffset(Index)) {
1876     SDValue N0 = Index.getOperand(0);
1877     SDValue N1 = Index.getOperand(1);
1878     ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
1879 
1880     // (add n0, c0)
1881     // Don't peel off the offset (c0) if doing so could possibly lead
1882     // the base (n0) to be negative.
1883     if (C1->getSExtValue() <= 0 || CurDAG->SignBitIsZero(N0)) {
1884       Base = N0;
1885       Offset = CurDAG->getTargetConstant(C1->getZExtValue(), DL, MVT::i32);
1886       return true;
1887     }
1888   }
1889 
1890   if (isa<ConstantSDNode>(Index))
1891     return false;
1892 
1893   Base = Index;
1894   Offset = CurDAG->getTargetConstant(0, DL, MVT::i32);
1895   return true;
1896 }
1897 
1898 SDNode *AMDGPUDAGToDAGISel::getS_BFE(unsigned Opcode, const SDLoc &DL,
1899                                      SDValue Val, uint32_t Offset,
1900                                      uint32_t Width) {
1901   // Transformation function, pack the offset and width of a BFE into
1902   // the format expected by the S_BFE_I32 / S_BFE_U32. In the second
1903   // source, bits [5:0] contain the offset and bits [22:16] the width.
1904   uint32_t PackedVal = Offset | (Width << 16);
1905   SDValue PackedConst = CurDAG->getTargetConstant(PackedVal, DL, MVT::i32);
1906 
1907   return CurDAG->getMachineNode(Opcode, DL, MVT::i32, Val, PackedConst);
1908 }
1909 
1910 void AMDGPUDAGToDAGISel::SelectS_BFEFromShifts(SDNode *N) {
1911   // "(a << b) srl c)" ---> "BFE_U32 a, (c-b), (32-c)
1912   // "(a << b) sra c)" ---> "BFE_I32 a, (c-b), (32-c)
1913   // Predicate: 0 < b <= c < 32
1914 
1915   const SDValue &Shl = N->getOperand(0);
1916   ConstantSDNode *B = dyn_cast<ConstantSDNode>(Shl->getOperand(1));
1917   ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
1918 
1919   if (B && C) {
1920     uint32_t BVal = B->getZExtValue();
1921     uint32_t CVal = C->getZExtValue();
1922 
1923     if (0 < BVal && BVal <= CVal && CVal < 32) {
1924       bool Signed = N->getOpcode() == ISD::SRA;
1925       unsigned Opcode = Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32;
1926 
1927       ReplaceNode(N, getS_BFE(Opcode, SDLoc(N), Shl.getOperand(0), CVal - BVal,
1928                               32 - CVal));
1929       return;
1930     }
1931   }
1932   SelectCode(N);
1933 }
1934 
1935 void AMDGPUDAGToDAGISel::SelectS_BFE(SDNode *N) {
1936   switch (N->getOpcode()) {
1937   case ISD::AND:
1938     if (N->getOperand(0).getOpcode() == ISD::SRL) {
1939       // "(a srl b) & mask" ---> "BFE_U32 a, b, popcount(mask)"
1940       // Predicate: isMask(mask)
1941       const SDValue &Srl = N->getOperand(0);
1942       ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(Srl.getOperand(1));
1943       ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(N->getOperand(1));
1944 
1945       if (Shift && Mask) {
1946         uint32_t ShiftVal = Shift->getZExtValue();
1947         uint32_t MaskVal = Mask->getZExtValue();
1948 
1949         if (isMask_32(MaskVal)) {
1950           uint32_t WidthVal = countPopulation(MaskVal);
1951 
1952           ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N),
1953                                   Srl.getOperand(0), ShiftVal, WidthVal));
1954           return;
1955         }
1956       }
1957     }
1958     break;
1959   case ISD::SRL:
1960     if (N->getOperand(0).getOpcode() == ISD::AND) {
1961       // "(a & mask) srl b)" ---> "BFE_U32 a, b, popcount(mask >> b)"
1962       // Predicate: isMask(mask >> b)
1963       const SDValue &And = N->getOperand(0);
1964       ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(N->getOperand(1));
1965       ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(And->getOperand(1));
1966 
1967       if (Shift && Mask) {
1968         uint32_t ShiftVal = Shift->getZExtValue();
1969         uint32_t MaskVal = Mask->getZExtValue() >> ShiftVal;
1970 
1971         if (isMask_32(MaskVal)) {
1972           uint32_t WidthVal = countPopulation(MaskVal);
1973 
1974           ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_U32, SDLoc(N),
1975                                   And.getOperand(0), ShiftVal, WidthVal));
1976           return;
1977         }
1978       }
1979     } else if (N->getOperand(0).getOpcode() == ISD::SHL) {
1980       SelectS_BFEFromShifts(N);
1981       return;
1982     }
1983     break;
1984   case ISD::SRA:
1985     if (N->getOperand(0).getOpcode() == ISD::SHL) {
1986       SelectS_BFEFromShifts(N);
1987       return;
1988     }
1989     break;
1990 
1991   case ISD::SIGN_EXTEND_INREG: {
1992     // sext_inreg (srl x, 16), i8 -> bfe_i32 x, 16, 8
1993     SDValue Src = N->getOperand(0);
1994     if (Src.getOpcode() != ISD::SRL)
1995       break;
1996 
1997     const ConstantSDNode *Amt = dyn_cast<ConstantSDNode>(Src.getOperand(1));
1998     if (!Amt)
1999       break;
2000 
2001     unsigned Width = cast<VTSDNode>(N->getOperand(1))->getVT().getSizeInBits();
2002     ReplaceNode(N, getS_BFE(AMDGPU::S_BFE_I32, SDLoc(N), Src.getOperand(0),
2003                             Amt->getZExtValue(), Width));
2004     return;
2005   }
2006   }
2007 
2008   SelectCode(N);
2009 }
2010 
2011 bool AMDGPUDAGToDAGISel::isCBranchSCC(const SDNode *N) const {
2012   assert(N->getOpcode() == ISD::BRCOND);
2013   if (!N->hasOneUse())
2014     return false;
2015 
2016   SDValue Cond = N->getOperand(1);
2017   if (Cond.getOpcode() == ISD::CopyToReg)
2018     Cond = Cond.getOperand(2);
2019 
2020   if (Cond.getOpcode() != ISD::SETCC || !Cond.hasOneUse())
2021     return false;
2022 
2023   MVT VT = Cond.getOperand(0).getSimpleValueType();
2024   if (VT == MVT::i32)
2025     return true;
2026 
2027   if (VT == MVT::i64) {
2028     auto ST = static_cast<const GCNSubtarget *>(Subtarget);
2029 
2030     ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
2031     return (CC == ISD::SETEQ || CC == ISD::SETNE) && ST->hasScalarCompareEq64();
2032   }
2033 
2034   return false;
2035 }
2036 
2037 void AMDGPUDAGToDAGISel::SelectBRCOND(SDNode *N) {
2038   SDValue Cond = N->getOperand(1);
2039 
2040   if (Cond.isUndef()) {
2041     CurDAG->SelectNodeTo(N, AMDGPU::SI_BR_UNDEF, MVT::Other,
2042                          N->getOperand(2), N->getOperand(0));
2043     return;
2044   }
2045 
2046   const GCNSubtarget *ST = static_cast<const GCNSubtarget *>(Subtarget);
2047   const SIRegisterInfo *TRI = ST->getRegisterInfo();
2048 
2049   bool UseSCCBr = isCBranchSCC(N) && isUniformBr(N);
2050   unsigned BrOp = UseSCCBr ? AMDGPU::S_CBRANCH_SCC1 : AMDGPU::S_CBRANCH_VCCNZ;
2051   unsigned CondReg = UseSCCBr ? (unsigned)AMDGPU::SCC : TRI->getVCC();
2052   SDLoc SL(N);
2053 
2054   if (!UseSCCBr) {
2055     // This is the case that we are selecting to S_CBRANCH_VCCNZ.  We have not
2056     // analyzed what generates the vcc value, so we do not know whether vcc
2057     // bits for disabled lanes are 0.  Thus we need to mask out bits for
2058     // disabled lanes.
2059     //
2060     // For the case that we select S_CBRANCH_SCC1 and it gets
2061     // changed to S_CBRANCH_VCCNZ in SIFixSGPRCopies, SIFixSGPRCopies calls
2062     // SIInstrInfo::moveToVALU which inserts the S_AND).
2063     //
2064     // We could add an analysis of what generates the vcc value here and omit
2065     // the S_AND when is unnecessary. But it would be better to add a separate
2066     // pass after SIFixSGPRCopies to do the unnecessary S_AND removal, so it
2067     // catches both cases.
2068     Cond = SDValue(CurDAG->getMachineNode(ST->isWave32() ? AMDGPU::S_AND_B32
2069                                                          : AMDGPU::S_AND_B64,
2070                      SL, MVT::i1,
2071                      CurDAG->getRegister(ST->isWave32() ? AMDGPU::EXEC_LO
2072                                                         : AMDGPU::EXEC,
2073                                          MVT::i1),
2074                     Cond),
2075                    0);
2076   }
2077 
2078   SDValue VCC = CurDAG->getCopyToReg(N->getOperand(0), SL, CondReg, Cond);
2079   CurDAG->SelectNodeTo(N, BrOp, MVT::Other,
2080                        N->getOperand(2), // Basic Block
2081                        VCC.getValue(0));
2082 }
2083 
2084 void AMDGPUDAGToDAGISel::SelectFMAD_FMA(SDNode *N) {
2085   MVT VT = N->getSimpleValueType(0);
2086   bool IsFMA = N->getOpcode() == ISD::FMA;
2087   if (VT != MVT::f32 || (!Subtarget->hasMadMixInsts() &&
2088                          !Subtarget->hasFmaMixInsts()) ||
2089       ((IsFMA && Subtarget->hasMadMixInsts()) ||
2090        (!IsFMA && Subtarget->hasFmaMixInsts()))) {
2091     SelectCode(N);
2092     return;
2093   }
2094 
2095   SDValue Src0 = N->getOperand(0);
2096   SDValue Src1 = N->getOperand(1);
2097   SDValue Src2 = N->getOperand(2);
2098   unsigned Src0Mods, Src1Mods, Src2Mods;
2099 
2100   // Avoid using v_mad_mix_f32/v_fma_mix_f32 unless there is actually an operand
2101   // using the conversion from f16.
2102   bool Sel0 = SelectVOP3PMadMixModsImpl(Src0, Src0, Src0Mods);
2103   bool Sel1 = SelectVOP3PMadMixModsImpl(Src1, Src1, Src1Mods);
2104   bool Sel2 = SelectVOP3PMadMixModsImpl(Src2, Src2, Src2Mods);
2105 
2106   assert((IsFMA || !Subtarget->hasFP32Denormals()) &&
2107          "fmad selected with denormals enabled");
2108   // TODO: We can select this with f32 denormals enabled if all the sources are
2109   // converted from f16 (in which case fmad isn't legal).
2110 
2111   if (Sel0 || Sel1 || Sel2) {
2112     // For dummy operands.
2113     SDValue Zero = CurDAG->getTargetConstant(0, SDLoc(), MVT::i32);
2114     SDValue Ops[] = {
2115       CurDAG->getTargetConstant(Src0Mods, SDLoc(), MVT::i32), Src0,
2116       CurDAG->getTargetConstant(Src1Mods, SDLoc(), MVT::i32), Src1,
2117       CurDAG->getTargetConstant(Src2Mods, SDLoc(), MVT::i32), Src2,
2118       CurDAG->getTargetConstant(0, SDLoc(), MVT::i1),
2119       Zero, Zero
2120     };
2121 
2122     CurDAG->SelectNodeTo(N,
2123                          IsFMA ? AMDGPU::V_FMA_MIX_F32 : AMDGPU::V_MAD_MIX_F32,
2124                          MVT::f32, Ops);
2125   } else {
2126     SelectCode(N);
2127   }
2128 }
2129 
2130 // This is here because there isn't a way to use the generated sub0_sub1 as the
2131 // subreg index to EXTRACT_SUBREG in tablegen.
2132 void AMDGPUDAGToDAGISel::SelectATOMIC_CMP_SWAP(SDNode *N) {
2133   MemSDNode *Mem = cast<MemSDNode>(N);
2134   unsigned AS = Mem->getAddressSpace();
2135   if (AS == AMDGPUAS::FLAT_ADDRESS) {
2136     SelectCode(N);
2137     return;
2138   }
2139 
2140   MVT VT = N->getSimpleValueType(0);
2141   bool Is32 = (VT == MVT::i32);
2142   SDLoc SL(N);
2143 
2144   MachineSDNode *CmpSwap = nullptr;
2145   if (Subtarget->hasAddr64()) {
2146     SDValue SRsrc, VAddr, SOffset, Offset, SLC;
2147 
2148     if (SelectMUBUFAddr64(Mem->getBasePtr(), SRsrc, VAddr, SOffset, Offset, SLC)) {
2149       unsigned Opcode = Is32 ? AMDGPU::BUFFER_ATOMIC_CMPSWAP_ADDR64_RTN :
2150         AMDGPU::BUFFER_ATOMIC_CMPSWAP_X2_ADDR64_RTN;
2151       SDValue CmpVal = Mem->getOperand(2);
2152 
2153       // XXX - Do we care about glue operands?
2154 
2155       SDValue Ops[] = {
2156         CmpVal, VAddr, SRsrc, SOffset, Offset, SLC, Mem->getChain()
2157       };
2158 
2159       CmpSwap = CurDAG->getMachineNode(Opcode, SL, Mem->getVTList(), Ops);
2160     }
2161   }
2162 
2163   if (!CmpSwap) {
2164     SDValue SRsrc, SOffset, Offset, SLC;
2165     if (SelectMUBUFOffset(Mem->getBasePtr(), SRsrc, SOffset, Offset, SLC)) {
2166       unsigned Opcode = Is32 ? AMDGPU::BUFFER_ATOMIC_CMPSWAP_OFFSET_RTN :
2167         AMDGPU::BUFFER_ATOMIC_CMPSWAP_X2_OFFSET_RTN;
2168 
2169       SDValue CmpVal = Mem->getOperand(2);
2170       SDValue Ops[] = {
2171         CmpVal, SRsrc, SOffset, Offset, SLC, Mem->getChain()
2172       };
2173 
2174       CmpSwap = CurDAG->getMachineNode(Opcode, SL, Mem->getVTList(), Ops);
2175     }
2176   }
2177 
2178   if (!CmpSwap) {
2179     SelectCode(N);
2180     return;
2181   }
2182 
2183   MachineMemOperand *MMO = Mem->getMemOperand();
2184   CurDAG->setNodeMemRefs(CmpSwap, {MMO});
2185 
2186   unsigned SubReg = Is32 ? AMDGPU::sub0 : AMDGPU::sub0_sub1;
2187   SDValue Extract
2188     = CurDAG->getTargetExtractSubreg(SubReg, SL, VT, SDValue(CmpSwap, 0));
2189 
2190   ReplaceUses(SDValue(N, 0), Extract);
2191   ReplaceUses(SDValue(N, 1), SDValue(CmpSwap, 1));
2192   CurDAG->RemoveDeadNode(N);
2193 }
2194 
2195 void AMDGPUDAGToDAGISel::SelectDSAppendConsume(SDNode *N, unsigned IntrID) {
2196   // The address is assumed to be uniform, so if it ends up in a VGPR, it will
2197   // be copied to an SGPR with readfirstlane.
2198   unsigned Opc = IntrID == Intrinsic::amdgcn_ds_append ?
2199     AMDGPU::DS_APPEND : AMDGPU::DS_CONSUME;
2200 
2201   SDValue Chain = N->getOperand(0);
2202   SDValue Ptr = N->getOperand(2);
2203   MemIntrinsicSDNode *M = cast<MemIntrinsicSDNode>(N);
2204   MachineMemOperand *MMO = M->getMemOperand();
2205   bool IsGDS = M->getAddressSpace() == AMDGPUAS::REGION_ADDRESS;
2206 
2207   SDValue Offset;
2208   if (CurDAG->isBaseWithConstantOffset(Ptr)) {
2209     SDValue PtrBase = Ptr.getOperand(0);
2210     SDValue PtrOffset = Ptr.getOperand(1);
2211 
2212     const APInt &OffsetVal = cast<ConstantSDNode>(PtrOffset)->getAPIntValue();
2213     if (isDSOffsetLegal(PtrBase, OffsetVal.getZExtValue(), 16)) {
2214       N = glueCopyToM0(N, PtrBase);
2215       Offset = CurDAG->getTargetConstant(OffsetVal, SDLoc(), MVT::i32);
2216     }
2217   }
2218 
2219   if (!Offset) {
2220     N = glueCopyToM0(N, Ptr);
2221     Offset = CurDAG->getTargetConstant(0, SDLoc(), MVT::i32);
2222   }
2223 
2224   SDValue Ops[] = {
2225     Offset,
2226     CurDAG->getTargetConstant(IsGDS, SDLoc(), MVT::i32),
2227     Chain,
2228     N->getOperand(N->getNumOperands() - 1) // New glue
2229   };
2230 
2231   SDNode *Selected = CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
2232   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Selected), {MMO});
2233 }
2234 
2235 static unsigned gwsIntrinToOpcode(unsigned IntrID) {
2236   switch (IntrID) {
2237   case Intrinsic::amdgcn_ds_gws_init:
2238     return AMDGPU::DS_GWS_INIT;
2239   case Intrinsic::amdgcn_ds_gws_barrier:
2240     return AMDGPU::DS_GWS_BARRIER;
2241   case Intrinsic::amdgcn_ds_gws_sema_v:
2242     return AMDGPU::DS_GWS_SEMA_V;
2243   case Intrinsic::amdgcn_ds_gws_sema_br:
2244     return AMDGPU::DS_GWS_SEMA_BR;
2245   case Intrinsic::amdgcn_ds_gws_sema_p:
2246     return AMDGPU::DS_GWS_SEMA_P;
2247   case Intrinsic::amdgcn_ds_gws_sema_release_all:
2248     return AMDGPU::DS_GWS_SEMA_RELEASE_ALL;
2249   default:
2250     llvm_unreachable("not a gws intrinsic");
2251   }
2252 }
2253 
2254 void AMDGPUDAGToDAGISel::SelectDS_GWS(SDNode *N, unsigned IntrID) {
2255   if (IntrID == Intrinsic::amdgcn_ds_gws_sema_release_all &&
2256       !Subtarget->hasGWSSemaReleaseAll()) {
2257     // Let this error.
2258     SelectCode(N);
2259     return;
2260   }
2261 
2262   // Chain, intrinsic ID, vsrc, offset
2263   const bool HasVSrc = N->getNumOperands() == 4;
2264   assert(HasVSrc || N->getNumOperands() == 3);
2265 
2266   SDLoc SL(N);
2267   SDValue BaseOffset = N->getOperand(HasVSrc ? 3 : 2);
2268   int ImmOffset = 0;
2269   MemIntrinsicSDNode *M = cast<MemIntrinsicSDNode>(N);
2270   MachineMemOperand *MMO = M->getMemOperand();
2271 
2272   // Don't worry if the offset ends up in a VGPR. Only one lane will have
2273   // effect, so SIFixSGPRCopies will validly insert readfirstlane.
2274 
2275   // The resource id offset is computed as (<isa opaque base> + M0[21:16] +
2276   // offset field) % 64. Some versions of the programming guide omit the m0
2277   // part, or claim it's from offset 0.
2278   if (ConstantSDNode *ConstOffset = dyn_cast<ConstantSDNode>(BaseOffset)) {
2279     // If we have a constant offset, try to use the 0 in m0 as the base.
2280     // TODO: Look into changing the default m0 initialization value. If the
2281     // default -1 only set the low 16-bits, we could leave it as-is and add 1 to
2282     // the immediate offset.
2283     glueCopyToM0(N, CurDAG->getTargetConstant(0, SL, MVT::i32));
2284     ImmOffset = ConstOffset->getZExtValue();
2285   } else {
2286     if (CurDAG->isBaseWithConstantOffset(BaseOffset)) {
2287       ImmOffset = BaseOffset.getConstantOperandVal(1);
2288       BaseOffset = BaseOffset.getOperand(0);
2289     }
2290 
2291     // Prefer to do the shift in an SGPR since it should be possible to use m0
2292     // as the result directly. If it's already an SGPR, it will be eliminated
2293     // later.
2294     SDNode *SGPROffset
2295       = CurDAG->getMachineNode(AMDGPU::V_READFIRSTLANE_B32, SL, MVT::i32,
2296                                BaseOffset);
2297     // Shift to offset in m0
2298     SDNode *M0Base
2299       = CurDAG->getMachineNode(AMDGPU::S_LSHL_B32, SL, MVT::i32,
2300                                SDValue(SGPROffset, 0),
2301                                CurDAG->getTargetConstant(16, SL, MVT::i32));
2302     glueCopyToM0(N, SDValue(M0Base, 0));
2303   }
2304 
2305   SDValue Chain = N->getOperand(0);
2306   SDValue OffsetField = CurDAG->getTargetConstant(ImmOffset, SL, MVT::i32);
2307 
2308   // TODO: Can this just be removed from the instruction?
2309   SDValue GDS = CurDAG->getTargetConstant(1, SL, MVT::i1);
2310 
2311   const unsigned Opc = gwsIntrinToOpcode(IntrID);
2312   SmallVector<SDValue, 5> Ops;
2313   if (HasVSrc)
2314     Ops.push_back(N->getOperand(2));
2315   Ops.push_back(OffsetField);
2316   Ops.push_back(GDS);
2317   Ops.push_back(Chain);
2318 
2319   SDNode *Selected = CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
2320   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Selected), {MMO});
2321 }
2322 
2323 void AMDGPUDAGToDAGISel::SelectINTRINSIC_W_CHAIN(SDNode *N) {
2324   unsigned IntrID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2325   switch (IntrID) {
2326   case Intrinsic::amdgcn_ds_append:
2327   case Intrinsic::amdgcn_ds_consume: {
2328     if (N->getValueType(0) != MVT::i32)
2329       break;
2330     SelectDSAppendConsume(N, IntrID);
2331     return;
2332   }
2333   }
2334 
2335   SelectCode(N);
2336 }
2337 
2338 void AMDGPUDAGToDAGISel::SelectINTRINSIC_WO_CHAIN(SDNode *N) {
2339   unsigned IntrID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2340   unsigned Opcode;
2341   switch (IntrID) {
2342   case Intrinsic::amdgcn_wqm:
2343     Opcode = AMDGPU::WQM;
2344     break;
2345   case Intrinsic::amdgcn_softwqm:
2346     Opcode = AMDGPU::SOFT_WQM;
2347     break;
2348   case Intrinsic::amdgcn_wwm:
2349     Opcode = AMDGPU::WWM;
2350     break;
2351   default:
2352     SelectCode(N);
2353     return;
2354   }
2355 
2356   SDValue Src = N->getOperand(1);
2357   CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), {Src});
2358 }
2359 
2360 void AMDGPUDAGToDAGISel::SelectINTRINSIC_VOID(SDNode *N) {
2361   unsigned IntrID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2362   switch (IntrID) {
2363   case Intrinsic::amdgcn_ds_gws_init:
2364   case Intrinsic::amdgcn_ds_gws_barrier:
2365   case Intrinsic::amdgcn_ds_gws_sema_v:
2366   case Intrinsic::amdgcn_ds_gws_sema_br:
2367   case Intrinsic::amdgcn_ds_gws_sema_p:
2368   case Intrinsic::amdgcn_ds_gws_sema_release_all:
2369     SelectDS_GWS(N, IntrID);
2370     return;
2371   default:
2372     break;
2373   }
2374 
2375   SelectCode(N);
2376 }
2377 
2378 bool AMDGPUDAGToDAGISel::SelectVOP3ModsImpl(SDValue In, SDValue &Src,
2379                                             unsigned &Mods) const {
2380   Mods = 0;
2381   Src = In;
2382 
2383   if (Src.getOpcode() == ISD::FNEG) {
2384     Mods |= SISrcMods::NEG;
2385     Src = Src.getOperand(0);
2386   }
2387 
2388   if (Src.getOpcode() == ISD::FABS) {
2389     Mods |= SISrcMods::ABS;
2390     Src = Src.getOperand(0);
2391   }
2392 
2393   return true;
2394 }
2395 
2396 bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src,
2397                                         SDValue &SrcMods) const {
2398   unsigned Mods;
2399   if (SelectVOP3ModsImpl(In, Src, Mods)) {
2400     SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
2401     return true;
2402   }
2403 
2404   return false;
2405 }
2406 
2407 bool AMDGPUDAGToDAGISel::SelectVOP3Mods_NNaN(SDValue In, SDValue &Src,
2408                                              SDValue &SrcMods) const {
2409   SelectVOP3Mods(In, Src, SrcMods);
2410   return isNoNanSrc(Src);
2411 }
2412 
2413 bool AMDGPUDAGToDAGISel::SelectVOP3Mods_f32(SDValue In, SDValue &Src,
2414                                             SDValue &SrcMods) const {
2415   if (In.getValueType() == MVT::f32)
2416     return SelectVOP3Mods(In, Src, SrcMods);
2417   Src = In;
2418   SrcMods = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32);;
2419   return true;
2420 }
2421 
2422 bool AMDGPUDAGToDAGISel::SelectVOP3NoMods(SDValue In, SDValue &Src) const {
2423   if (In.getOpcode() == ISD::FABS || In.getOpcode() == ISD::FNEG)
2424     return false;
2425 
2426   Src = In;
2427   return true;
2428 }
2429 
2430 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0(SDValue In, SDValue &Src,
2431                                          SDValue &SrcMods, SDValue &Clamp,
2432                                          SDValue &Omod) const {
2433   SDLoc DL(In);
2434   Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1);
2435   Omod = CurDAG->getTargetConstant(0, DL, MVT::i1);
2436 
2437   return SelectVOP3Mods(In, Src, SrcMods);
2438 }
2439 
2440 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src,
2441                                                    SDValue &SrcMods,
2442                                                    SDValue &Clamp,
2443                                                    SDValue &Omod) const {
2444   Clamp = Omod = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32);
2445   return SelectVOP3Mods(In, Src, SrcMods);
2446 }
2447 
2448 bool AMDGPUDAGToDAGISel::SelectVOP3OMods(SDValue In, SDValue &Src,
2449                                          SDValue &Clamp, SDValue &Omod) const {
2450   Src = In;
2451 
2452   SDLoc DL(In);
2453   Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1);
2454   Omod = CurDAG->getTargetConstant(0, DL, MVT::i1);
2455 
2456   return true;
2457 }
2458 
2459 bool AMDGPUDAGToDAGISel::SelectVOP3PMods(SDValue In, SDValue &Src,
2460                                          SDValue &SrcMods) const {
2461   unsigned Mods = 0;
2462   Src = In;
2463 
2464   if (Src.getOpcode() == ISD::FNEG) {
2465     Mods ^= (SISrcMods::NEG | SISrcMods::NEG_HI);
2466     Src = Src.getOperand(0);
2467   }
2468 
2469   if (Src.getOpcode() == ISD::BUILD_VECTOR) {
2470     unsigned VecMods = Mods;
2471 
2472     SDValue Lo = stripBitcast(Src.getOperand(0));
2473     SDValue Hi = stripBitcast(Src.getOperand(1));
2474 
2475     if (Lo.getOpcode() == ISD::FNEG) {
2476       Lo = stripBitcast(Lo.getOperand(0));
2477       Mods ^= SISrcMods::NEG;
2478     }
2479 
2480     if (Hi.getOpcode() == ISD::FNEG) {
2481       Hi = stripBitcast(Hi.getOperand(0));
2482       Mods ^= SISrcMods::NEG_HI;
2483     }
2484 
2485     if (isExtractHiElt(Lo, Lo))
2486       Mods |= SISrcMods::OP_SEL_0;
2487 
2488     if (isExtractHiElt(Hi, Hi))
2489       Mods |= SISrcMods::OP_SEL_1;
2490 
2491     Lo = stripExtractLoElt(Lo);
2492     Hi = stripExtractLoElt(Hi);
2493 
2494     if (Lo == Hi && !isInlineImmediate(Lo.getNode())) {
2495       // Really a scalar input. Just select from the low half of the register to
2496       // avoid packing.
2497 
2498       Src = Lo;
2499       SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
2500       return true;
2501     }
2502 
2503     Mods = VecMods;
2504   }
2505 
2506   // Packed instructions do not have abs modifiers.
2507   Mods |= SISrcMods::OP_SEL_1;
2508 
2509   SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
2510   return true;
2511 }
2512 
2513 bool AMDGPUDAGToDAGISel::SelectVOP3PMods0(SDValue In, SDValue &Src,
2514                                           SDValue &SrcMods,
2515                                           SDValue &Clamp) const {
2516   SDLoc SL(In);
2517 
2518   // FIXME: Handle clamp and op_sel
2519   Clamp = CurDAG->getTargetConstant(0, SL, MVT::i32);
2520 
2521   return SelectVOP3PMods(In, Src, SrcMods);
2522 }
2523 
2524 bool AMDGPUDAGToDAGISel::SelectVOP3OpSel(SDValue In, SDValue &Src,
2525                                          SDValue &SrcMods) const {
2526   Src = In;
2527   // FIXME: Handle op_sel
2528   SrcMods = CurDAG->getTargetConstant(0, SDLoc(In), MVT::i32);
2529   return true;
2530 }
2531 
2532 bool AMDGPUDAGToDAGISel::SelectVOP3OpSel0(SDValue In, SDValue &Src,
2533                                           SDValue &SrcMods,
2534                                           SDValue &Clamp) const {
2535   SDLoc SL(In);
2536 
2537   // FIXME: Handle clamp
2538   Clamp = CurDAG->getTargetConstant(0, SL, MVT::i32);
2539 
2540   return SelectVOP3OpSel(In, Src, SrcMods);
2541 }
2542 
2543 bool AMDGPUDAGToDAGISel::SelectVOP3OpSelMods(SDValue In, SDValue &Src,
2544                                              SDValue &SrcMods) const {
2545   // FIXME: Handle op_sel
2546   return SelectVOP3Mods(In, Src, SrcMods);
2547 }
2548 
2549 bool AMDGPUDAGToDAGISel::SelectVOP3OpSelMods0(SDValue In, SDValue &Src,
2550                                               SDValue &SrcMods,
2551                                               SDValue &Clamp) const {
2552   SDLoc SL(In);
2553 
2554   // FIXME: Handle clamp
2555   Clamp = CurDAG->getTargetConstant(0, SL, MVT::i32);
2556 
2557   return SelectVOP3OpSelMods(In, Src, SrcMods);
2558 }
2559 
2560 // The return value is not whether the match is possible (which it always is),
2561 // but whether or not it a conversion is really used.
2562 bool AMDGPUDAGToDAGISel::SelectVOP3PMadMixModsImpl(SDValue In, SDValue &Src,
2563                                                    unsigned &Mods) const {
2564   Mods = 0;
2565   SelectVOP3ModsImpl(In, Src, Mods);
2566 
2567   if (Src.getOpcode() == ISD::FP_EXTEND) {
2568     Src = Src.getOperand(0);
2569     assert(Src.getValueType() == MVT::f16);
2570     Src = stripBitcast(Src);
2571 
2572     // Be careful about folding modifiers if we already have an abs. fneg is
2573     // applied last, so we don't want to apply an earlier fneg.
2574     if ((Mods & SISrcMods::ABS) == 0) {
2575       unsigned ModsTmp;
2576       SelectVOP3ModsImpl(Src, Src, ModsTmp);
2577 
2578       if ((ModsTmp & SISrcMods::NEG) != 0)
2579         Mods ^= SISrcMods::NEG;
2580 
2581       if ((ModsTmp & SISrcMods::ABS) != 0)
2582         Mods |= SISrcMods::ABS;
2583     }
2584 
2585     // op_sel/op_sel_hi decide the source type and source.
2586     // If the source's op_sel_hi is set, it indicates to do a conversion from fp16.
2587     // If the sources's op_sel is set, it picks the high half of the source
2588     // register.
2589 
2590     Mods |= SISrcMods::OP_SEL_1;
2591     if (isExtractHiElt(Src, Src)) {
2592       Mods |= SISrcMods::OP_SEL_0;
2593 
2594       // TODO: Should we try to look for neg/abs here?
2595     }
2596 
2597     return true;
2598   }
2599 
2600   return false;
2601 }
2602 
2603 bool AMDGPUDAGToDAGISel::SelectVOP3PMadMixMods(SDValue In, SDValue &Src,
2604                                                SDValue &SrcMods) const {
2605   unsigned Mods = 0;
2606   SelectVOP3PMadMixModsImpl(In, Src, Mods);
2607   SrcMods = CurDAG->getTargetConstant(Mods, SDLoc(In), MVT::i32);
2608   return true;
2609 }
2610 
2611 SDValue AMDGPUDAGToDAGISel::getHi16Elt(SDValue In) const {
2612   if (In.isUndef())
2613     return CurDAG->getUNDEF(MVT::i32);
2614 
2615   if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(In)) {
2616     SDLoc SL(In);
2617     return CurDAG->getConstant(C->getZExtValue() << 16, SL, MVT::i32);
2618   }
2619 
2620   if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(In)) {
2621     SDLoc SL(In);
2622     return CurDAG->getConstant(
2623       C->getValueAPF().bitcastToAPInt().getZExtValue() << 16, SL, MVT::i32);
2624   }
2625 
2626   SDValue Src;
2627   if (isExtractHiElt(In, Src))
2628     return Src;
2629 
2630   return SDValue();
2631 }
2632 
2633 bool AMDGPUDAGToDAGISel::isVGPRImm(const SDNode * N) const {
2634   assert(CurDAG->getTarget().getTargetTriple().getArch() == Triple::amdgcn);
2635 
2636   const SIRegisterInfo *SIRI =
2637     static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
2638   const SIInstrInfo * SII =
2639     static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
2640 
2641   unsigned Limit = 0;
2642   bool AllUsesAcceptSReg = true;
2643   for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
2644     Limit < 10 && U != E; ++U, ++Limit) {
2645     const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
2646 
2647     // If the register class is unknown, it could be an unknown
2648     // register class that needs to be an SGPR, e.g. an inline asm
2649     // constraint
2650     if (!RC || SIRI->isSGPRClass(RC))
2651       return false;
2652 
2653     if (RC != &AMDGPU::VS_32RegClass) {
2654       AllUsesAcceptSReg = false;
2655       SDNode * User = *U;
2656       if (User->isMachineOpcode()) {
2657         unsigned Opc = User->getMachineOpcode();
2658         MCInstrDesc Desc = SII->get(Opc);
2659         if (Desc.isCommutable()) {
2660           unsigned OpIdx = Desc.getNumDefs() + U.getOperandNo();
2661           unsigned CommuteIdx1 = TargetInstrInfo::CommuteAnyOperandIndex;
2662           if (SII->findCommutedOpIndices(Desc, OpIdx, CommuteIdx1)) {
2663             unsigned CommutedOpNo = CommuteIdx1 - Desc.getNumDefs();
2664             const TargetRegisterClass *CommutedRC = getOperandRegClass(*U, CommutedOpNo);
2665             if (CommutedRC == &AMDGPU::VS_32RegClass)
2666               AllUsesAcceptSReg = true;
2667           }
2668         }
2669       }
2670       // If "AllUsesAcceptSReg == false" so far we haven't suceeded
2671       // commuting current user. This means have at least one use
2672       // that strictly require VGPR. Thus, we will not attempt to commute
2673       // other user instructions.
2674       if (!AllUsesAcceptSReg)
2675         break;
2676     }
2677   }
2678   return !AllUsesAcceptSReg && (Limit < 10);
2679 }
2680 
2681 bool AMDGPUDAGToDAGISel::isUniformLoad(const SDNode * N) const {
2682   auto Ld = cast<LoadSDNode>(N);
2683 
2684   return Ld->getAlignment() >= 4 &&
2685         (
2686           (
2687             (
2688               Ld->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS       ||
2689               Ld->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT
2690             )
2691             &&
2692             !N->isDivergent()
2693           )
2694           ||
2695           (
2696             Subtarget->getScalarizeGlobalBehavior() &&
2697             Ld->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
2698             !Ld->isVolatile() &&
2699             !N->isDivergent() &&
2700             static_cast<const SITargetLowering *>(
2701               getTargetLowering())->isMemOpHasNoClobberedMemOperand(N)
2702           )
2703         );
2704 }
2705 
2706 void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
2707   const AMDGPUTargetLowering& Lowering =
2708     *static_cast<const AMDGPUTargetLowering*>(getTargetLowering());
2709   bool IsModified = false;
2710   do {
2711     IsModified = false;
2712 
2713     // Go over all selected nodes and try to fold them a bit more
2714     SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_begin();
2715     while (Position != CurDAG->allnodes_end()) {
2716       SDNode *Node = &*Position++;
2717       MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(Node);
2718       if (!MachineNode)
2719         continue;
2720 
2721       SDNode *ResNode = Lowering.PostISelFolding(MachineNode, *CurDAG);
2722       if (ResNode != Node) {
2723         if (ResNode)
2724           ReplaceUses(Node, ResNode);
2725         IsModified = true;
2726       }
2727     }
2728     CurDAG->RemoveDeadNodes();
2729   } while (IsModified);
2730 }
2731 
2732 bool R600DAGToDAGISel::runOnMachineFunction(MachineFunction &MF) {
2733   Subtarget = &MF.getSubtarget<R600Subtarget>();
2734   return SelectionDAGISel::runOnMachineFunction(MF);
2735 }
2736 
2737 bool R600DAGToDAGISel::isConstantLoad(const MemSDNode *N, int CbId) const {
2738   if (!N->readMem())
2739     return false;
2740   if (CbId == -1)
2741     return N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS ||
2742            N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT;
2743 
2744   return N->getAddressSpace() == AMDGPUAS::CONSTANT_BUFFER_0 + CbId;
2745 }
2746 
2747 bool R600DAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr,
2748                                                          SDValue& IntPtr) {
2749   if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Addr)) {
2750     IntPtr = CurDAG->getIntPtrConstant(Cst->getZExtValue() / 4, SDLoc(Addr),
2751                                        true);
2752     return true;
2753   }
2754   return false;
2755 }
2756 
2757 bool R600DAGToDAGISel::SelectGlobalValueVariableOffset(SDValue Addr,
2758     SDValue& BaseReg, SDValue &Offset) {
2759   if (!isa<ConstantSDNode>(Addr)) {
2760     BaseReg = Addr;
2761     Offset = CurDAG->getIntPtrConstant(0, SDLoc(Addr), true);
2762     return true;
2763   }
2764   return false;
2765 }
2766 
2767 void R600DAGToDAGISel::Select(SDNode *N) {
2768   unsigned int Opc = N->getOpcode();
2769   if (N->isMachineOpcode()) {
2770     N->setNodeId(-1);
2771     return;   // Already selected.
2772   }
2773 
2774   switch (Opc) {
2775   default: break;
2776   case AMDGPUISD::BUILD_VERTICAL_VECTOR:
2777   case ISD::SCALAR_TO_VECTOR:
2778   case ISD::BUILD_VECTOR: {
2779     EVT VT = N->getValueType(0);
2780     unsigned NumVectorElts = VT.getVectorNumElements();
2781     unsigned RegClassID;
2782     // BUILD_VECTOR was lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
2783     // that adds a 128 bits reg copy when going through TwoAddressInstructions
2784     // pass. We want to avoid 128 bits copies as much as possible because they
2785     // can't be bundled by our scheduler.
2786     switch(NumVectorElts) {
2787     case 2: RegClassID = R600::R600_Reg64RegClassID; break;
2788     case 4:
2789       if (Opc == AMDGPUISD::BUILD_VERTICAL_VECTOR)
2790         RegClassID = R600::R600_Reg128VerticalRegClassID;
2791       else
2792         RegClassID = R600::R600_Reg128RegClassID;
2793       break;
2794     default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
2795     }
2796     SelectBuildVector(N, RegClassID);
2797     return;
2798   }
2799   }
2800 
2801   SelectCode(N);
2802 }
2803 
2804 bool R600DAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base,
2805                                           SDValue &Offset) {
2806   ConstantSDNode *C;
2807   SDLoc DL(Addr);
2808 
2809   if ((C = dyn_cast<ConstantSDNode>(Addr))) {
2810     Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
2811     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
2812   } else if ((Addr.getOpcode() == AMDGPUISD::DWORDADDR) &&
2813              (C = dyn_cast<ConstantSDNode>(Addr.getOperand(0)))) {
2814     Base = CurDAG->getRegister(R600::INDIRECT_BASE_ADDR, MVT::i32);
2815     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
2816   } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
2817             (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) {
2818     Base = Addr.getOperand(0);
2819     Offset = CurDAG->getTargetConstant(C->getZExtValue(), DL, MVT::i32);
2820   } else {
2821     Base = Addr;
2822     Offset = CurDAG->getTargetConstant(0, DL, MVT::i32);
2823   }
2824 
2825   return true;
2826 }
2827 
2828 bool R600DAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
2829                                           SDValue &Offset) {
2830   ConstantSDNode *IMMOffset;
2831 
2832   if (Addr.getOpcode() == ISD::ADD
2833       && (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))
2834       && isInt<16>(IMMOffset->getZExtValue())) {
2835 
2836       Base = Addr.getOperand(0);
2837       Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr),
2838                                          MVT::i32);
2839       return true;
2840   // If the pointer address is constant, we can move it to the offset field.
2841   } else if ((IMMOffset = dyn_cast<ConstantSDNode>(Addr))
2842              && isInt<16>(IMMOffset->getZExtValue())) {
2843     Base = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
2844                                   SDLoc(CurDAG->getEntryNode()),
2845                                   R600::ZERO, MVT::i32);
2846     Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), SDLoc(Addr),
2847                                        MVT::i32);
2848     return true;
2849   }
2850 
2851   // Default case, no offset
2852   Base = Addr;
2853   Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), MVT::i32);
2854   return true;
2855 }
2856