xref: /freebsd/contrib/llvm-project/llvm/lib/Target/PowerPC/PPCISelLowering.h (revision 43a5ec4eb41567cc92586503212743d89686d78f)
1 //===-- PPCISelLowering.h - PPC32 DAG Lowering Interface --------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the interfaces that PPC uses to lower LLVM code into a
10 // selection DAG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef LLVM_LIB_TARGET_POWERPC_PPCISELLOWERING_H
15 #define LLVM_LIB_TARGET_POWERPC_PPCISELLOWERING_H
16 
17 #include "PPCInstrInfo.h"
18 #include "llvm/CodeGen/CallingConvLower.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineMemOperand.h"
21 #include "llvm/CodeGen/SelectionDAG.h"
22 #include "llvm/CodeGen/SelectionDAGNodes.h"
23 #include "llvm/CodeGen/TargetLowering.h"
24 #include "llvm/CodeGen/ValueTypes.h"
25 #include "llvm/IR/Attributes.h"
26 #include "llvm/IR/CallingConv.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/InlineAsm.h"
29 #include "llvm/IR/Metadata.h"
30 #include "llvm/IR/Type.h"
31 #include "llvm/Support/MachineValueType.h"
32 #include <utility>
33 
34 namespace llvm {
35 
36   namespace PPCISD {
37 
38     // When adding a NEW PPCISD node please add it to the correct position in
39     // the enum. The order of elements in this enum matters!
40     // Values that are added after this entry:
41     //     STBRX = ISD::FIRST_TARGET_MEMORY_OPCODE
42     // are considered memory opcodes and are treated differently than entries
43     // that come before it. For example, ADD or MUL should be placed before
44     // the ISD::FIRST_TARGET_MEMORY_OPCODE while a LOAD or STORE should come
45     // after it.
46   enum NodeType : unsigned {
47     // Start the numbering where the builtin ops and target ops leave off.
48     FIRST_NUMBER = ISD::BUILTIN_OP_END,
49 
50     /// FSEL - Traditional three-operand fsel node.
51     ///
52     FSEL,
53 
54     /// XSMAXCDP, XSMINCDP - C-type min/max instructions.
55     XSMAXCDP,
56     XSMINCDP,
57 
58     /// FCFID - The FCFID instruction, taking an f64 operand and producing
59     /// and f64 value containing the FP representation of the integer that
60     /// was temporarily in the f64 operand.
61     FCFID,
62 
63     /// Newer FCFID[US] integer-to-floating-point conversion instructions for
64     /// unsigned integers and single-precision outputs.
65     FCFIDU,
66     FCFIDS,
67     FCFIDUS,
68 
69     /// FCTI[D,W]Z - The FCTIDZ and FCTIWZ instructions, taking an f32 or f64
70     /// operand, producing an f64 value containing the integer representation
71     /// of that FP value.
72     FCTIDZ,
73     FCTIWZ,
74 
75     /// Newer FCTI[D,W]UZ floating-point-to-integer conversion instructions for
76     /// unsigned integers with round toward zero.
77     FCTIDUZ,
78     FCTIWUZ,
79 
80     /// Floating-point-to-interger conversion instructions
81     FP_TO_UINT_IN_VSR,
82     FP_TO_SINT_IN_VSR,
83 
84     /// VEXTS, ByteWidth - takes an input in VSFRC and produces an output in
85     /// VSFRC that is sign-extended from ByteWidth to a 64-byte integer.
86     VEXTS,
87 
88     /// Reciprocal estimate instructions (unary FP ops).
89     FRE,
90     FRSQRTE,
91 
92     /// Test instruction for software square root.
93     FTSQRT,
94 
95     /// Square root instruction.
96     FSQRT,
97 
98     /// VPERM - The PPC VPERM Instruction.
99     ///
100     VPERM,
101 
102     /// XXSPLT - The PPC VSX splat instructions
103     ///
104     XXSPLT,
105 
106     /// XXSPLTI_SP_TO_DP - The PPC VSX splat instructions for immediates for
107     /// converting immediate single precision numbers to double precision
108     /// vector or scalar.
109     XXSPLTI_SP_TO_DP,
110 
111     /// XXSPLTI32DX - The PPC XXSPLTI32DX instruction.
112     ///
113     XXSPLTI32DX,
114 
115     /// VECINSERT - The PPC vector insert instruction
116     ///
117     VECINSERT,
118 
119     /// VECSHL - The PPC vector shift left instruction
120     ///
121     VECSHL,
122 
123     /// XXPERMDI - The PPC XXPERMDI instruction
124     ///
125     XXPERMDI,
126 
127     /// The CMPB instruction (takes two operands of i32 or i64).
128     CMPB,
129 
130     /// Hi/Lo - These represent the high and low 16-bit parts of a global
131     /// address respectively.  These nodes have two operands, the first of
132     /// which must be a TargetGlobalAddress, and the second of which must be a
133     /// Constant.  Selected naively, these turn into 'lis G+C' and 'li G+C',
134     /// though these are usually folded into other nodes.
135     Hi,
136     Lo,
137 
138     /// The following two target-specific nodes are used for calls through
139     /// function pointers in the 64-bit SVR4 ABI.
140 
141     /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX)
142     /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
143     /// compute an allocation on the stack.
144     DYNALLOC,
145 
146     /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
147     /// compute an offset from native SP to the address  of the most recent
148     /// dynamic alloca.
149     DYNAREAOFFSET,
150 
151     /// To avoid stack clash, allocation is performed by block and each block is
152     /// probed.
153     PROBED_ALLOCA,
154 
155     /// The result of the mflr at function entry, used for PIC code.
156     GlobalBaseReg,
157 
158     /// These nodes represent PPC shifts.
159     ///
160     /// For scalar types, only the last `n + 1` bits of the shift amounts
161     /// are used, where n is log2(sizeof(element) * 8). See sld/slw, etc.
162     /// for exact behaviors.
163     ///
164     /// For vector types, only the last n bits are used. See vsld.
165     SRL,
166     SRA,
167     SHL,
168 
169     /// FNMSUB - Negated multiply-subtract instruction.
170     FNMSUB,
171 
172     /// EXTSWSLI = The PPC extswsli instruction, which does an extend-sign
173     /// word and shift left immediate.
174     EXTSWSLI,
175 
176     /// The combination of sra[wd]i and addze used to implemented signed
177     /// integer division by a power of 2. The first operand is the dividend,
178     /// and the second is the constant shift amount (representing the
179     /// divisor).
180     SRA_ADDZE,
181 
182     /// CALL - A direct function call.
183     /// CALL_NOP is a call with the special NOP which follows 64-bit
184     /// CALL_NOTOC the caller does not use the TOC.
185     /// SVR4 calls and 32-bit/64-bit AIX calls.
186     CALL,
187     CALL_NOP,
188     CALL_NOTOC,
189 
190     /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a
191     /// MTCTR instruction.
192     MTCTR,
193 
194     /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a
195     /// BCTRL instruction.
196     BCTRL,
197 
198     /// CHAIN,FLAG = BCTRL(CHAIN, ADDR, INFLAG) - The combination of a bctrl
199     /// instruction and the TOC reload required on 64-bit ELF, 32-bit AIX
200     /// and 64-bit AIX.
201     BCTRL_LOAD_TOC,
202 
203     /// Return with a flag operand, matched by 'blr'
204     RET_FLAG,
205 
206     /// R32 = MFOCRF(CRREG, INFLAG) - Represents the MFOCRF instruction.
207     /// This copies the bits corresponding to the specified CRREG into the
208     /// resultant GPR.  Bits corresponding to other CR regs are undefined.
209     MFOCRF,
210 
211     /// Direct move from a VSX register to a GPR
212     MFVSR,
213 
214     /// Direct move from a GPR to a VSX register (algebraic)
215     MTVSRA,
216 
217     /// Direct move from a GPR to a VSX register (zero)
218     MTVSRZ,
219 
220     /// Direct move of 2 consecutive GPR to a VSX register.
221     BUILD_FP128,
222 
223     /// BUILD_SPE64 and EXTRACT_SPE are analogous to BUILD_PAIR and
224     /// EXTRACT_ELEMENT but take f64 arguments instead of i64, as i64 is
225     /// unsupported for this target.
226     /// Merge 2 GPRs to a single SPE register.
227     BUILD_SPE64,
228 
229     /// Extract SPE register component, second argument is high or low.
230     EXTRACT_SPE,
231 
232     /// Extract a subvector from signed integer vector and convert to FP.
233     /// It is primarily used to convert a (widened) illegal integer vector
234     /// type to a legal floating point vector type.
235     /// For example v2i32 -> widened to v4i32 -> v2f64
236     SINT_VEC_TO_FP,
237 
238     /// Extract a subvector from unsigned integer vector and convert to FP.
239     /// As with SINT_VEC_TO_FP, used for converting illegal types.
240     UINT_VEC_TO_FP,
241 
242     /// PowerPC instructions that have SCALAR_TO_VECTOR semantics tend to
243     /// place the value into the least significant element of the most
244     /// significant doubleword in the vector. This is not element zero for
245     /// anything smaller than a doubleword on either endianness. This node has
246     /// the same semantics as SCALAR_TO_VECTOR except that the value remains in
247     /// the aforementioned location in the vector register.
248     SCALAR_TO_VECTOR_PERMUTED,
249 
250     // FIXME: Remove these once the ANDI glue bug is fixed:
251     /// i1 = ANDI_rec_1_[EQ|GT]_BIT(i32 or i64 x) - Represents the result of the
252     /// eq or gt bit of CR0 after executing andi. x, 1. This is used to
253     /// implement truncation of i32 or i64 to i1.
254     ANDI_rec_1_EQ_BIT,
255     ANDI_rec_1_GT_BIT,
256 
257     // READ_TIME_BASE - A read of the 64-bit time-base register on a 32-bit
258     // target (returns (Lo, Hi)). It takes a chain operand.
259     READ_TIME_BASE,
260 
261     // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
262     EH_SJLJ_SETJMP,
263 
264     // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
265     EH_SJLJ_LONGJMP,
266 
267     /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP*
268     /// instructions.  For lack of better number, we use the opcode number
269     /// encoding for the OPC field to identify the compare.  For example, 838
270     /// is VCMPGTSH.
271     VCMP,
272 
273     /// RESVEC, OUTFLAG = VCMP_rec(LHS, RHS, OPC) - Represents one of the
274     /// altivec VCMP*_rec instructions.  For lack of better number, we use the
275     /// opcode number encoding for the OPC field to identify the compare.  For
276     /// example, 838 is VCMPGTSH.
277     VCMP_rec,
278 
279     /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This
280     /// corresponds to the COND_BRANCH pseudo instruction.  CRRC is the
281     /// condition register to branch on, OPC is the branch opcode to use (e.g.
282     /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is
283     /// an optional input flag argument.
284     COND_BRANCH,
285 
286     /// CHAIN = BDNZ CHAIN, DESTBB - These are used to create counter-based
287     /// loops.
288     BDNZ,
289     BDZ,
290 
291     /// F8RC = FADDRTZ F8RC, F8RC - This is an FADD done with rounding
292     /// towards zero.  Used only as part of the long double-to-int
293     /// conversion sequence.
294     FADDRTZ,
295 
296     /// F8RC = MFFS - This moves the FPSCR (not modeled) into the register.
297     MFFS,
298 
299     /// TC_RETURN - A tail call return.
300     ///   operand #0 chain
301     ///   operand #1 callee (register or absolute)
302     ///   operand #2 stack adjustment
303     ///   operand #3 optional in flag
304     TC_RETURN,
305 
306     /// ch, gl = CR6[UN]SET ch, inglue - Toggle CR bit 6 for SVR4 vararg calls
307     CR6SET,
308     CR6UNSET,
309 
310     /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by initial-exec TLS
311     /// for non-position independent code on PPC32.
312     PPC32_GOT,
313 
314     /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by general dynamic and
315     /// local dynamic TLS and position indendepent code on PPC32.
316     PPC32_PICGOT,
317 
318     /// G8RC = ADDIS_GOT_TPREL_HA %x2, Symbol - Used by the initial-exec
319     /// TLS model, produces an ADDIS8 instruction that adds the GOT
320     /// base to sym\@got\@tprel\@ha.
321     ADDIS_GOT_TPREL_HA,
322 
323     /// G8RC = LD_GOT_TPREL_L Symbol, G8RReg - Used by the initial-exec
324     /// TLS model, produces a LD instruction with base register G8RReg
325     /// and offset sym\@got\@tprel\@l.  This completes the addition that
326     /// finds the offset of "sym" relative to the thread pointer.
327     LD_GOT_TPREL_L,
328 
329     /// G8RC = ADD_TLS G8RReg, Symbol - Used by the initial-exec TLS
330     /// model, produces an ADD instruction that adds the contents of
331     /// G8RReg to the thread pointer.  Symbol contains a relocation
332     /// sym\@tls which is to be replaced by the thread pointer and
333     /// identifies to the linker that the instruction is part of a
334     /// TLS sequence.
335     ADD_TLS,
336 
337     /// G8RC = ADDIS_TLSGD_HA %x2, Symbol - For the general-dynamic TLS
338     /// model, produces an ADDIS8 instruction that adds the GOT base
339     /// register to sym\@got\@tlsgd\@ha.
340     ADDIS_TLSGD_HA,
341 
342     /// %x3 = ADDI_TLSGD_L G8RReg, Symbol - For the general-dynamic TLS
343     /// model, produces an ADDI8 instruction that adds G8RReg to
344     /// sym\@got\@tlsgd\@l and stores the result in X3.  Hidden by
345     /// ADDIS_TLSGD_L_ADDR until after register assignment.
346     ADDI_TLSGD_L,
347 
348     /// %x3 = GET_TLS_ADDR %x3, Symbol - For the general-dynamic TLS
349     /// model, produces a call to __tls_get_addr(sym\@tlsgd).  Hidden by
350     /// ADDIS_TLSGD_L_ADDR until after register assignment.
351     GET_TLS_ADDR,
352 
353     /// G8RC = ADDI_TLSGD_L_ADDR G8RReg, Symbol, Symbol - Op that
354     /// combines ADDI_TLSGD_L and GET_TLS_ADDR until expansion following
355     /// register assignment.
356     ADDI_TLSGD_L_ADDR,
357 
358     /// GPRC = TLSGD_AIX, TOC_ENTRY, TOC_ENTRY
359     /// G8RC = TLSGD_AIX, TOC_ENTRY, TOC_ENTRY
360     /// Op that combines two register copies of TOC entries
361     /// (region handle into R3 and variable offset into R4) followed by a
362     /// GET_TLS_ADDR node which will be expanded to a call to __get_tls_addr.
363     /// This node is used in 64-bit mode as well (in which case the result is
364     /// G8RC and inputs are X3/X4).
365     TLSGD_AIX,
366 
367     /// G8RC = ADDIS_TLSLD_HA %x2, Symbol - For the local-dynamic TLS
368     /// model, produces an ADDIS8 instruction that adds the GOT base
369     /// register to sym\@got\@tlsld\@ha.
370     ADDIS_TLSLD_HA,
371 
372     /// %x3 = ADDI_TLSLD_L G8RReg, Symbol - For the local-dynamic TLS
373     /// model, produces an ADDI8 instruction that adds G8RReg to
374     /// sym\@got\@tlsld\@l and stores the result in X3.  Hidden by
375     /// ADDIS_TLSLD_L_ADDR until after register assignment.
376     ADDI_TLSLD_L,
377 
378     /// %x3 = GET_TLSLD_ADDR %x3, Symbol - For the local-dynamic TLS
379     /// model, produces a call to __tls_get_addr(sym\@tlsld).  Hidden by
380     /// ADDIS_TLSLD_L_ADDR until after register assignment.
381     GET_TLSLD_ADDR,
382 
383     /// G8RC = ADDI_TLSLD_L_ADDR G8RReg, Symbol, Symbol - Op that
384     /// combines ADDI_TLSLD_L and GET_TLSLD_ADDR until expansion
385     /// following register assignment.
386     ADDI_TLSLD_L_ADDR,
387 
388     /// G8RC = ADDIS_DTPREL_HA %x3, Symbol - For the local-dynamic TLS
389     /// model, produces an ADDIS8 instruction that adds X3 to
390     /// sym\@dtprel\@ha.
391     ADDIS_DTPREL_HA,
392 
393     /// G8RC = ADDI_DTPREL_L G8RReg, Symbol - For the local-dynamic TLS
394     /// model, produces an ADDI8 instruction that adds G8RReg to
395     /// sym\@got\@dtprel\@l.
396     ADDI_DTPREL_L,
397 
398     /// G8RC = PADDI_DTPREL %x3, Symbol - For the pc-rel based local-dynamic TLS
399     /// model, produces a PADDI8 instruction that adds X3 to sym\@dtprel.
400     PADDI_DTPREL,
401 
402     /// VRRC = VADD_SPLAT Elt, EltSize - Temporary node to be expanded
403     /// during instruction selection to optimize a BUILD_VECTOR into
404     /// operations on splats.  This is necessary to avoid losing these
405     /// optimizations due to constant folding.
406     VADD_SPLAT,
407 
408     /// CHAIN = SC CHAIN, Imm128 - System call.  The 7-bit unsigned
409     /// operand identifies the operating system entry point.
410     SC,
411 
412     /// CHAIN = CLRBHRB CHAIN - Clear branch history rolling buffer.
413     CLRBHRB,
414 
415     /// GPRC, CHAIN = MFBHRBE CHAIN, Entry, Dummy - Move from branch
416     /// history rolling buffer entry.
417     MFBHRBE,
418 
419     /// CHAIN = RFEBB CHAIN, State - Return from event-based branch.
420     RFEBB,
421 
422     /// VSRC, CHAIN = XXSWAPD CHAIN, VSRC - Occurs only for little
423     /// endian.  Maps to an xxswapd instruction that corrects an lxvd2x
424     /// or stxvd2x instruction.  The chain is necessary because the
425     /// sequence replaces a load and needs to provide the same number
426     /// of outputs.
427     XXSWAPD,
428 
429     /// An SDNode for swaps that are not associated with any loads/stores
430     /// and thereby have no chain.
431     SWAP_NO_CHAIN,
432 
433     /// An SDNode for Power9 vector absolute value difference.
434     /// operand #0 vector
435     /// operand #1 vector
436     /// operand #2 constant i32 0 or 1, to indicate whether needs to patch
437     /// the most significant bit for signed i32
438     ///
439     /// Power9 VABSD* instructions are designed to support unsigned integer
440     /// vectors (byte/halfword/word), if we want to make use of them for signed
441     /// integer vectors, we have to flip their sign bits first. To flip sign bit
442     /// for byte/halfword integer vector would become inefficient, but for word
443     /// integer vector, we can leverage XVNEGSP to make it efficiently. eg:
444     /// abs(sub(a,b)) => VABSDUW(a+0x80000000, b+0x80000000)
445     ///               => VABSDUW((XVNEGSP a), (XVNEGSP b))
446     VABSD,
447 
448     /// FP_EXTEND_HALF(VECTOR, IDX) - Custom extend upper (IDX=0) half or
449     /// lower (IDX=1) half of v4f32 to v2f64.
450     FP_EXTEND_HALF,
451 
452     /// MAT_PCREL_ADDR = Materialize a PC Relative address. This can be done
453     /// either through an add like PADDI or through a PC Relative load like
454     /// PLD.
455     MAT_PCREL_ADDR,
456 
457     /// TLS_DYNAMIC_MAT_PCREL_ADDR = Materialize a PC Relative address for
458     /// TLS global address when using dynamic access models. This can be done
459     /// through an add like PADDI.
460     TLS_DYNAMIC_MAT_PCREL_ADDR,
461 
462     /// TLS_LOCAL_EXEC_MAT_ADDR = Materialize an address for TLS global address
463     /// when using local exec access models, and when prefixed instructions are
464     /// available. This is used with ADD_TLS to produce an add like PADDI.
465     TLS_LOCAL_EXEC_MAT_ADDR,
466 
467     /// ACC_BUILD = Build an accumulator register from 4 VSX registers.
468     ACC_BUILD,
469 
470     /// PAIR_BUILD = Build a vector pair register from 2 VSX registers.
471     PAIR_BUILD,
472 
473     /// EXTRACT_VSX_REG = Extract one of the underlying vsx registers of
474     /// an accumulator or pair register. This node is needed because
475     /// EXTRACT_SUBVECTOR expects the input and output vectors to have the same
476     /// element type.
477     EXTRACT_VSX_REG,
478 
479     /// XXMFACC = This corresponds to the xxmfacc instruction.
480     XXMFACC,
481 
482     // Constrained conversion from floating point to int
483     STRICT_FCTIDZ = ISD::FIRST_TARGET_STRICTFP_OPCODE,
484     STRICT_FCTIWZ,
485     STRICT_FCTIDUZ,
486     STRICT_FCTIWUZ,
487 
488     /// Constrained integer-to-floating-point conversion instructions.
489     STRICT_FCFID,
490     STRICT_FCFIDU,
491     STRICT_FCFIDS,
492     STRICT_FCFIDUS,
493 
494     /// Constrained floating point add in round-to-zero mode.
495     STRICT_FADDRTZ,
496 
497     /// CHAIN = STBRX CHAIN, GPRC, Ptr, Type - This is a
498     /// byte-swapping store instruction.  It byte-swaps the low "Type" bits of
499     /// the GPRC input, then stores it through Ptr.  Type can be either i16 or
500     /// i32.
501     STBRX = ISD::FIRST_TARGET_MEMORY_OPCODE,
502 
503     /// GPRC, CHAIN = LBRX CHAIN, Ptr, Type - This is a
504     /// byte-swapping load instruction.  It loads "Type" bits, byte swaps it,
505     /// then puts it in the bottom bits of the GPRC.  TYPE can be either i16
506     /// or i32.
507     LBRX,
508 
509     /// STFIWX - The STFIWX instruction.  The first operand is an input token
510     /// chain, then an f64 value to store, then an address to store it to.
511     STFIWX,
512 
513     /// GPRC, CHAIN = LFIWAX CHAIN, Ptr - This is a floating-point
514     /// load which sign-extends from a 32-bit integer value into the
515     /// destination 64-bit register.
516     LFIWAX,
517 
518     /// GPRC, CHAIN = LFIWZX CHAIN, Ptr - This is a floating-point
519     /// load which zero-extends from a 32-bit integer value into the
520     /// destination 64-bit register.
521     LFIWZX,
522 
523     /// GPRC, CHAIN = LXSIZX, CHAIN, Ptr, ByteWidth - This is a load of an
524     /// integer smaller than 64 bits into a VSR. The integer is zero-extended.
525     /// This can be used for converting loaded integers to floating point.
526     LXSIZX,
527 
528     /// STXSIX - The STXSI[bh]X instruction. The first operand is an input
529     /// chain, then an f64 value to store, then an address to store it to,
530     /// followed by a byte-width for the store.
531     STXSIX,
532 
533     /// VSRC, CHAIN = LXVD2X_LE CHAIN, Ptr - Occurs only for little endian.
534     /// Maps directly to an lxvd2x instruction that will be followed by
535     /// an xxswapd.
536     LXVD2X,
537 
538     /// LXVRZX - Load VSX Vector Rightmost and Zero Extend
539     /// This node represents v1i128 BUILD_VECTOR of a zero extending load
540     /// instruction from <byte, halfword, word, or doubleword> to i128.
541     /// Allows utilization of the Load VSX Vector Rightmost Instructions.
542     LXVRZX,
543 
544     /// VSRC, CHAIN = LOAD_VEC_BE CHAIN, Ptr - Occurs only for little endian.
545     /// Maps directly to one of lxvd2x/lxvw4x/lxvh8x/lxvb16x depending on
546     /// the vector type to load vector in big-endian element order.
547     LOAD_VEC_BE,
548 
549     /// VSRC, CHAIN = LD_VSX_LH CHAIN, Ptr - This is a floating-point load of a
550     /// v2f32 value into the lower half of a VSR register.
551     LD_VSX_LH,
552 
553     /// VSRC, CHAIN = LD_SPLAT, CHAIN, Ptr - a splatting load memory
554     /// instructions such as LXVDSX, LXVWSX.
555     LD_SPLAT,
556 
557     /// CHAIN = STXVD2X CHAIN, VSRC, Ptr - Occurs only for little endian.
558     /// Maps directly to an stxvd2x instruction that will be preceded by
559     /// an xxswapd.
560     STXVD2X,
561 
562     /// CHAIN = STORE_VEC_BE CHAIN, VSRC, Ptr - Occurs only for little endian.
563     /// Maps directly to one of stxvd2x/stxvw4x/stxvh8x/stxvb16x depending on
564     /// the vector type to store vector in big-endian element order.
565     STORE_VEC_BE,
566 
567     /// Store scalar integers from VSR.
568     ST_VSR_SCAL_INT,
569 
570     /// ATOMIC_CMP_SWAP - the exact same as the target-independent nodes
571     /// except they ensure that the compare input is zero-extended for
572     /// sub-word versions because the atomic loads zero-extend.
573     ATOMIC_CMP_SWAP_8,
574     ATOMIC_CMP_SWAP_16,
575 
576     /// GPRC = TOC_ENTRY GA, TOC
577     /// Loads the entry for GA from the TOC, where the TOC base is given by
578     /// the last operand.
579     TOC_ENTRY
580   };
581 
582   } // end namespace PPCISD
583 
584   /// Define some predicates that are used for node matching.
585   namespace PPC {
586 
587     /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
588     /// VPKUHUM instruction.
589     bool isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
590                               SelectionDAG &DAG);
591 
592     /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
593     /// VPKUWUM instruction.
594     bool isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
595                               SelectionDAG &DAG);
596 
597     /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a
598     /// VPKUDUM instruction.
599     bool isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
600                               SelectionDAG &DAG);
601 
602     /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
603     /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
604     bool isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
605                             unsigned ShuffleKind, SelectionDAG &DAG);
606 
607     /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
608     /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
609     bool isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
610                             unsigned ShuffleKind, SelectionDAG &DAG);
611 
612     /// isVMRGEOShuffleMask - Return true if this is a shuffle mask suitable for
613     /// a VMRGEW or VMRGOW instruction
614     bool isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven,
615                              unsigned ShuffleKind, SelectionDAG &DAG);
616     /// isXXSLDWIShuffleMask - Return true if this is a shuffle mask suitable
617     /// for a XXSLDWI instruction.
618     bool isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
619                               bool &Swap, bool IsLE);
620 
621     /// isXXBRHShuffleMask - Return true if this is a shuffle mask suitable
622     /// for a XXBRH instruction.
623     bool isXXBRHShuffleMask(ShuffleVectorSDNode *N);
624 
625     /// isXXBRWShuffleMask - Return true if this is a shuffle mask suitable
626     /// for a XXBRW instruction.
627     bool isXXBRWShuffleMask(ShuffleVectorSDNode *N);
628 
629     /// isXXBRDShuffleMask - Return true if this is a shuffle mask suitable
630     /// for a XXBRD instruction.
631     bool isXXBRDShuffleMask(ShuffleVectorSDNode *N);
632 
633     /// isXXBRQShuffleMask - Return true if this is a shuffle mask suitable
634     /// for a XXBRQ instruction.
635     bool isXXBRQShuffleMask(ShuffleVectorSDNode *N);
636 
637     /// isXXPERMDIShuffleMask - Return true if this is a shuffle mask suitable
638     /// for a XXPERMDI instruction.
639     bool isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
640                               bool &Swap, bool IsLE);
641 
642     /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the
643     /// shift amount, otherwise return -1.
644     int isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
645                             SelectionDAG &DAG);
646 
647     /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
648     /// specifies a splat of a single element that is suitable for input to
649     /// VSPLTB/VSPLTH/VSPLTW.
650     bool isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize);
651 
652     /// isXXINSERTWMask - Return true if this VECTOR_SHUFFLE can be handled by
653     /// the XXINSERTW instruction introduced in ISA 3.0. This is essentially any
654     /// shuffle of v4f32/v4i32 vectors that just inserts one element from one
655     /// vector into the other. This function will also set a couple of
656     /// output parameters for how much the source vector needs to be shifted and
657     /// what byte number needs to be specified for the instruction to put the
658     /// element in the desired location of the target vector.
659     bool isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
660                          unsigned &InsertAtByte, bool &Swap, bool IsLE);
661 
662     /// getSplatIdxForPPCMnemonics - Return the splat index as a value that is
663     /// appropriate for PPC mnemonics (which have a big endian bias - namely
664     /// elements are counted from the left of the vector register).
665     unsigned getSplatIdxForPPCMnemonics(SDNode *N, unsigned EltSize,
666                                         SelectionDAG &DAG);
667 
668     /// get_VSPLTI_elt - If this is a build_vector of constants which can be
669     /// formed by using a vspltis[bhw] instruction of the specified element
670     /// size, return the constant being splatted.  The ByteSize field indicates
671     /// the number of bytes of each element [124] -> [bhw].
672     SDValue get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
673 
674     // Flags for computing the optimal addressing mode for loads and stores.
675     enum MemOpFlags {
676       MOF_None = 0,
677 
678       // Extension mode for integer loads.
679       MOF_SExt = 1,
680       MOF_ZExt = 1 << 1,
681       MOF_NoExt = 1 << 2,
682 
683       // Address computation flags.
684       MOF_NotAddNorCst = 1 << 5,      // Not const. or sum of ptr and scalar.
685       MOF_RPlusSImm16 = 1 << 6,       // Reg plus signed 16-bit constant.
686       MOF_RPlusLo = 1 << 7,           // Reg plus signed 16-bit relocation
687       MOF_RPlusSImm16Mult4 = 1 << 8,  // Reg plus 16-bit signed multiple of 4.
688       MOF_RPlusSImm16Mult16 = 1 << 9, // Reg plus 16-bit signed multiple of 16.
689       MOF_RPlusSImm34 = 1 << 10,      // Reg plus 34-bit signed constant.
690       MOF_RPlusR = 1 << 11,           // Sum of two variables.
691       MOF_PCRel = 1 << 12,            // PC-Relative relocation.
692       MOF_AddrIsSImm32 = 1 << 13,     // A simple 32-bit constant.
693 
694       // The in-memory type.
695       MOF_SubWordInt = 1 << 15,
696       MOF_WordInt = 1 << 16,
697       MOF_DoubleWordInt = 1 << 17,
698       MOF_ScalarFloat = 1 << 18, // Scalar single or double precision.
699       MOF_Vector = 1 << 19,      // Vector types and quad precision scalars.
700       MOF_Vector256 = 1 << 20,
701 
702       // Subtarget features.
703       MOF_SubtargetBeforeP9 = 1 << 22,
704       MOF_SubtargetP9 = 1 << 23,
705       MOF_SubtargetP10 = 1 << 24,
706       MOF_SubtargetSPE = 1 << 25
707     };
708 
709     // The addressing modes for loads and stores.
710     enum AddrMode {
711       AM_None,
712       AM_DForm,
713       AM_DSForm,
714       AM_DQForm,
715       AM_XForm,
716     };
717   } // end namespace PPC
718 
719   class PPCTargetLowering : public TargetLowering {
720     const PPCSubtarget &Subtarget;
721 
722   public:
723     explicit PPCTargetLowering(const PPCTargetMachine &TM,
724                                const PPCSubtarget &STI);
725 
726     /// getTargetNodeName() - This method returns the name of a target specific
727     /// DAG node.
728     const char *getTargetNodeName(unsigned Opcode) const override;
729 
730     bool isSelectSupported(SelectSupportKind Kind) const override {
731       // PowerPC does not support scalar condition selects on vectors.
732       return (Kind != SelectSupportKind::ScalarCondVectorVal);
733     }
734 
735     /// getPreferredVectorAction - The code we generate when vector types are
736     /// legalized by promoting the integer element type is often much worse
737     /// than code we generate if we widen the type for applicable vector types.
738     /// The issue with promoting is that the vector is scalaraized, individual
739     /// elements promoted and then the vector is rebuilt. So say we load a pair
740     /// of v4i8's and shuffle them. This will turn into a mess of 8 extending
741     /// loads, moves back into VSR's (or memory ops if we don't have moves) and
742     /// then the VPERM for the shuffle. All in all a very slow sequence.
743     TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(MVT VT)
744       const override {
745       if (!VT.isScalableVector() && VT.getVectorNumElements() != 1 &&
746           VT.getScalarSizeInBits() % 8 == 0)
747         return TypeWidenVector;
748       return TargetLoweringBase::getPreferredVectorAction(VT);
749     }
750 
751     bool useSoftFloat() const override;
752 
753     bool hasSPE() const;
754 
755     MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
756       return MVT::i32;
757     }
758 
759     bool isCheapToSpeculateCttz() const override {
760       return true;
761     }
762 
763     bool isCheapToSpeculateCtlz() const override {
764       return true;
765     }
766 
767     bool isCtlzFast() const override {
768       return true;
769     }
770 
771     bool isEqualityCmpFoldedWithSignedCmp() const override {
772       return false;
773     }
774 
775     bool hasAndNotCompare(SDValue) const override {
776       return true;
777     }
778 
779     bool preferIncOfAddToSubOfNot(EVT VT) const override;
780 
781     bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
782       return VT.isScalarInteger();
783     }
784 
785     SDValue getNegatedExpression(SDValue Op, SelectionDAG &DAG, bool LegalOps,
786                                  bool OptForSize, NegatibleCost &Cost,
787                                  unsigned Depth = 0) const override;
788 
789     /// getSetCCResultType - Return the ISD::SETCC ValueType
790     EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
791                            EVT VT) const override;
792 
793     /// Return true if target always benefits from combining into FMA for a
794     /// given value type. This must typically return false on targets where FMA
795     /// takes more cycles to execute than FADD.
796     bool enableAggressiveFMAFusion(EVT VT) const override;
797 
798     /// getPreIndexedAddressParts - returns true by value, base pointer and
799     /// offset pointer and addressing mode by reference if the node's address
800     /// can be legally represented as pre-indexed load / store address.
801     bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
802                                    SDValue &Offset,
803                                    ISD::MemIndexedMode &AM,
804                                    SelectionDAG &DAG) const override;
805 
806     /// SelectAddressEVXRegReg - Given the specified addressed, check to see if
807     /// it can be more efficiently represented as [r+imm].
808     bool SelectAddressEVXRegReg(SDValue N, SDValue &Base, SDValue &Index,
809                                 SelectionDAG &DAG) const;
810 
811     /// SelectAddressRegReg - Given the specified addressed, check to see if it
812     /// can be more efficiently represented as [r+imm]. If \p EncodingAlignment
813     /// is non-zero, only accept displacement which is not suitable for [r+imm].
814     /// Returns false if it can be represented by [r+imm], which are preferred.
815     bool SelectAddressRegReg(SDValue N, SDValue &Base, SDValue &Index,
816                              SelectionDAG &DAG,
817                              MaybeAlign EncodingAlignment = None) const;
818 
819     /// SelectAddressRegImm - Returns true if the address N can be represented
820     /// by a base register plus a signed 16-bit displacement [r+imm], and if it
821     /// is not better represented as reg+reg. If \p EncodingAlignment is
822     /// non-zero, only accept displacements suitable for instruction encoding
823     /// requirement, i.e. multiples of 4 for DS form.
824     bool SelectAddressRegImm(SDValue N, SDValue &Disp, SDValue &Base,
825                              SelectionDAG &DAG,
826                              MaybeAlign EncodingAlignment) const;
827     bool SelectAddressRegImm34(SDValue N, SDValue &Disp, SDValue &Base,
828                                SelectionDAG &DAG) const;
829 
830     /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
831     /// represented as an indexed [r+r] operation.
832     bool SelectAddressRegRegOnly(SDValue N, SDValue &Base, SDValue &Index,
833                                  SelectionDAG &DAG) const;
834 
835     /// SelectAddressPCRel - Represent the specified address as pc relative to
836     /// be represented as [pc+imm]
837     bool SelectAddressPCRel(SDValue N, SDValue &Base) const;
838 
839     Sched::Preference getSchedulingPreference(SDNode *N) const override;
840 
841     /// LowerOperation - Provide custom lowering hooks for some operations.
842     ///
843     SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
844 
845     /// ReplaceNodeResults - Replace the results of node with an illegal result
846     /// type with new values built out of custom code.
847     ///
848     void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
849                             SelectionDAG &DAG) const override;
850 
851     SDValue expandVSXLoadForLE(SDNode *N, DAGCombinerInfo &DCI) const;
852     SDValue expandVSXStoreForLE(SDNode *N, DAGCombinerInfo &DCI) const;
853 
854     SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
855 
856     SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
857                           SmallVectorImpl<SDNode *> &Created) const override;
858 
859     Register getRegisterByName(const char* RegName, LLT VT,
860                                const MachineFunction &MF) const override;
861 
862     void computeKnownBitsForTargetNode(const SDValue Op,
863                                        KnownBits &Known,
864                                        const APInt &DemandedElts,
865                                        const SelectionDAG &DAG,
866                                        unsigned Depth = 0) const override;
867 
868     Align getPrefLoopAlignment(MachineLoop *ML) const override;
869 
870     bool shouldInsertFencesForAtomic(const Instruction *I) const override {
871       return true;
872     }
873 
874     Instruction *emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst,
875                                   AtomicOrdering Ord) const override;
876     Instruction *emitTrailingFence(IRBuilderBase &Builder, Instruction *Inst,
877                                    AtomicOrdering Ord) const override;
878 
879     TargetLowering::AtomicExpansionKind
880     shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
881 
882     TargetLowering::AtomicExpansionKind
883     shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override;
884 
885     Value *emitMaskedAtomicRMWIntrinsic(IRBuilderBase &Builder,
886                                         AtomicRMWInst *AI, Value *AlignedAddr,
887                                         Value *Incr, Value *Mask,
888                                         Value *ShiftAmt,
889                                         AtomicOrdering Ord) const override;
890     Value *emitMaskedAtomicCmpXchgIntrinsic(IRBuilderBase &Builder,
891                                             AtomicCmpXchgInst *CI,
892                                             Value *AlignedAddr, Value *CmpVal,
893                                             Value *NewVal, Value *Mask,
894                                             AtomicOrdering Ord) const override;
895 
896     MachineBasicBlock *
897     EmitInstrWithCustomInserter(MachineInstr &MI,
898                                 MachineBasicBlock *MBB) const override;
899     MachineBasicBlock *EmitAtomicBinary(MachineInstr &MI,
900                                         MachineBasicBlock *MBB,
901                                         unsigned AtomicSize,
902                                         unsigned BinOpcode,
903                                         unsigned CmpOpcode = 0,
904                                         unsigned CmpPred = 0) const;
905     MachineBasicBlock *EmitPartwordAtomicBinary(MachineInstr &MI,
906                                                 MachineBasicBlock *MBB,
907                                                 bool is8bit,
908                                                 unsigned Opcode,
909                                                 unsigned CmpOpcode = 0,
910                                                 unsigned CmpPred = 0) const;
911 
912     MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
913                                         MachineBasicBlock *MBB) const;
914 
915     MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
916                                          MachineBasicBlock *MBB) const;
917 
918     MachineBasicBlock *emitProbedAlloca(MachineInstr &MI,
919                                         MachineBasicBlock *MBB) const;
920 
921     bool hasInlineStackProbe(MachineFunction &MF) const override;
922 
923     unsigned getStackProbeSize(MachineFunction &MF) const;
924 
925     ConstraintType getConstraintType(StringRef Constraint) const override;
926 
927     /// Examine constraint string and operand type and determine a weight value.
928     /// The operand object must already have been set up with the operand type.
929     ConstraintWeight getSingleConstraintMatchWeight(
930       AsmOperandInfo &info, const char *constraint) const override;
931 
932     std::pair<unsigned, const TargetRegisterClass *>
933     getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
934                                  StringRef Constraint, MVT VT) const override;
935 
936     /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
937     /// function arguments in the caller parameter area.  This is the actual
938     /// alignment, not its logarithm.
939     unsigned getByValTypeAlignment(Type *Ty,
940                                    const DataLayout &DL) const override;
941 
942     /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
943     /// vector.  If it is invalid, don't add anything to Ops.
944     void LowerAsmOperandForConstraint(SDValue Op,
945                                       std::string &Constraint,
946                                       std::vector<SDValue> &Ops,
947                                       SelectionDAG &DAG) const override;
948 
949     unsigned
950     getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
951       if (ConstraintCode == "es")
952         return InlineAsm::Constraint_es;
953       else if (ConstraintCode == "Q")
954         return InlineAsm::Constraint_Q;
955       else if (ConstraintCode == "Z")
956         return InlineAsm::Constraint_Z;
957       else if (ConstraintCode == "Zy")
958         return InlineAsm::Constraint_Zy;
959       return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
960     }
961 
962     /// isLegalAddressingMode - Return true if the addressing mode represented
963     /// by AM is legal for this target, for a load/store of the specified type.
964     bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
965                                Type *Ty, unsigned AS,
966                                Instruction *I = nullptr) const override;
967 
968     /// isLegalICmpImmediate - Return true if the specified immediate is legal
969     /// icmp immediate, that is the target has icmp instructions which can
970     /// compare a register against the immediate without having to materialize
971     /// the immediate into a register.
972     bool isLegalICmpImmediate(int64_t Imm) const override;
973 
974     /// isLegalAddImmediate - Return true if the specified immediate is legal
975     /// add immediate, that is the target has add instructions which can
976     /// add a register and the immediate without having to materialize
977     /// the immediate into a register.
978     bool isLegalAddImmediate(int64_t Imm) const override;
979 
980     /// isTruncateFree - Return true if it's free to truncate a value of
981     /// type Ty1 to type Ty2. e.g. On PPC it's free to truncate a i64 value in
982     /// register X1 to i32 by referencing its sub-register R1.
983     bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
984     bool isTruncateFree(EVT VT1, EVT VT2) const override;
985 
986     bool isZExtFree(SDValue Val, EVT VT2) const override;
987 
988     bool isFPExtFree(EVT DestVT, EVT SrcVT) const override;
989 
990     /// Returns true if it is beneficial to convert a load of a constant
991     /// to just the constant itself.
992     bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
993                                            Type *Ty) const override;
994 
995     bool convertSelectOfConstantsToMath(EVT VT) const override {
996       return true;
997     }
998 
999     bool decomposeMulByConstant(LLVMContext &Context, EVT VT,
1000                                 SDValue C) const override;
1001 
1002     bool isDesirableToTransformToIntegerOp(unsigned Opc,
1003                                            EVT VT) const override {
1004       // Only handle float load/store pair because float(fpr) load/store
1005       // instruction has more cycles than integer(gpr) load/store in PPC.
1006       if (Opc != ISD::LOAD && Opc != ISD::STORE)
1007         return false;
1008       if (VT != MVT::f32 && VT != MVT::f64)
1009         return false;
1010 
1011       return true;
1012     }
1013 
1014     // Returns true if the address of the global is stored in TOC entry.
1015     bool isAccessedAsGotIndirect(SDValue N) const;
1016 
1017     bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
1018 
1019     bool getTgtMemIntrinsic(IntrinsicInfo &Info,
1020                             const CallInst &I,
1021                             MachineFunction &MF,
1022                             unsigned Intrinsic) const override;
1023 
1024     /// It returns EVT::Other if the type should be determined using generic
1025     /// target-independent logic.
1026     EVT getOptimalMemOpType(const MemOp &Op,
1027                             const AttributeList &FuncAttributes) const override;
1028 
1029     /// Is unaligned memory access allowed for the given type, and is it fast
1030     /// relative to software emulation.
1031     bool allowsMisalignedMemoryAccesses(
1032         EVT VT, unsigned AddrSpace, Align Alignment = Align(1),
1033         MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
1034         bool *Fast = nullptr) const override;
1035 
1036     /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
1037     /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
1038     /// expanded to FMAs when this method returns true, otherwise fmuladd is
1039     /// expanded to fmul + fadd.
1040     bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
1041                                     EVT VT) const override;
1042 
1043     bool isFMAFasterThanFMulAndFAdd(const Function &F, Type *Ty) const override;
1044 
1045     /// isProfitableToHoist - Check if it is profitable to hoist instruction
1046     /// \p I to its dominator block.
1047     /// For example, it is not profitable if \p I and it's only user can form a
1048     /// FMA instruction, because Powerpc prefers FMADD.
1049     bool isProfitableToHoist(Instruction *I) const override;
1050 
1051     const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
1052 
1053     // Should we expand the build vector with shuffles?
1054     bool
1055     shouldExpandBuildVectorWithShuffles(EVT VT,
1056                                         unsigned DefinedValues) const override;
1057 
1058     // Keep the zero-extensions for arguments to libcalls.
1059     bool shouldKeepZExtForFP16Conv() const override { return true; }
1060 
1061     /// createFastISel - This method returns a target-specific FastISel object,
1062     /// or null if the target does not support "fast" instruction selection.
1063     FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
1064                              const TargetLibraryInfo *LibInfo) const override;
1065 
1066     /// Returns true if an argument of type Ty needs to be passed in a
1067     /// contiguous block of registers in calling convention CallConv.
1068     bool functionArgumentNeedsConsecutiveRegisters(
1069         Type *Ty, CallingConv::ID CallConv, bool isVarArg,
1070         const DataLayout &DL) const override {
1071       // We support any array type as "consecutive" block in the parameter
1072       // save area.  The element type defines the alignment requirement and
1073       // whether the argument should go in GPRs, FPRs, or VRs if available.
1074       //
1075       // Note that clang uses this capability both to implement the ELFv2
1076       // homogeneous float/vector aggregate ABI, and to avoid having to use
1077       // "byval" when passing aggregates that might fully fit in registers.
1078       return Ty->isArrayTy();
1079     }
1080 
1081     /// If a physical register, this returns the register that receives the
1082     /// exception address on entry to an EH pad.
1083     Register
1084     getExceptionPointerRegister(const Constant *PersonalityFn) const override;
1085 
1086     /// If a physical register, this returns the register that receives the
1087     /// exception typeid on entry to a landing pad.
1088     Register
1089     getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
1090 
1091     /// Override to support customized stack guard loading.
1092     bool useLoadStackGuardNode() const override;
1093     void insertSSPDeclarations(Module &M) const override;
1094     Value *getSDagStackGuard(const Module &M) const override;
1095 
1096     bool isFPImmLegal(const APFloat &Imm, EVT VT,
1097                       bool ForCodeSize) const override;
1098 
1099     unsigned getJumpTableEncoding() const override;
1100     bool isJumpTableRelative() const override;
1101     SDValue getPICJumpTableRelocBase(SDValue Table,
1102                                      SelectionDAG &DAG) const override;
1103     const MCExpr *getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
1104                                                unsigned JTI,
1105                                                MCContext &Ctx) const override;
1106 
1107     /// SelectOptimalAddrMode - Based on a node N and it's Parent (a MemSDNode),
1108     /// compute the address flags of the node, get the optimal address mode
1109     /// based on the flags, and set the Base and Disp based on the address mode.
1110     PPC::AddrMode SelectOptimalAddrMode(const SDNode *Parent, SDValue N,
1111                                         SDValue &Disp, SDValue &Base,
1112                                         SelectionDAG &DAG,
1113                                         MaybeAlign Align) const;
1114     /// SelectForceXFormMode - Given the specified address, force it to be
1115     /// represented as an indexed [r+r] operation (an XForm instruction).
1116     PPC::AddrMode SelectForceXFormMode(SDValue N, SDValue &Disp, SDValue &Base,
1117                                        SelectionDAG &DAG) const;
1118 
1119     /// Structure that collects some common arguments that get passed around
1120     /// between the functions for call lowering.
1121     struct CallFlags {
1122       const CallingConv::ID CallConv;
1123       const bool IsTailCall : 1;
1124       const bool IsVarArg : 1;
1125       const bool IsPatchPoint : 1;
1126       const bool IsIndirect : 1;
1127       const bool HasNest : 1;
1128       const bool NoMerge : 1;
1129 
1130       CallFlags(CallingConv::ID CC, bool IsTailCall, bool IsVarArg,
1131                 bool IsPatchPoint, bool IsIndirect, bool HasNest, bool NoMerge)
1132           : CallConv(CC), IsTailCall(IsTailCall), IsVarArg(IsVarArg),
1133             IsPatchPoint(IsPatchPoint), IsIndirect(IsIndirect),
1134             HasNest(HasNest), NoMerge(NoMerge) {}
1135     };
1136 
1137     CCAssignFn *ccAssignFnForCall(CallingConv::ID CC, bool Return,
1138                                   bool IsVarArg) const;
1139 
1140   private:
1141     struct ReuseLoadInfo {
1142       SDValue Ptr;
1143       SDValue Chain;
1144       SDValue ResChain;
1145       MachinePointerInfo MPI;
1146       bool IsDereferenceable = false;
1147       bool IsInvariant = false;
1148       Align Alignment;
1149       AAMDNodes AAInfo;
1150       const MDNode *Ranges = nullptr;
1151 
1152       ReuseLoadInfo() = default;
1153 
1154       MachineMemOperand::Flags MMOFlags() const {
1155         MachineMemOperand::Flags F = MachineMemOperand::MONone;
1156         if (IsDereferenceable)
1157           F |= MachineMemOperand::MODereferenceable;
1158         if (IsInvariant)
1159           F |= MachineMemOperand::MOInvariant;
1160         return F;
1161       }
1162     };
1163 
1164     // Map that relates a set of common address flags to PPC addressing modes.
1165     std::map<PPC::AddrMode, SmallVector<unsigned, 16>> AddrModesMap;
1166     void initializeAddrModeMap();
1167 
1168     bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI,
1169                              SelectionDAG &DAG,
1170                              ISD::LoadExtType ET = ISD::NON_EXTLOAD) const;
1171     void spliceIntoChain(SDValue ResChain, SDValue NewResChain,
1172                          SelectionDAG &DAG) const;
1173 
1174     void LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
1175                                 SelectionDAG &DAG, const SDLoc &dl) const;
1176     SDValue LowerFP_TO_INTDirectMove(SDValue Op, SelectionDAG &DAG,
1177                                      const SDLoc &dl) const;
1178 
1179     bool directMoveIsProfitable(const SDValue &Op) const;
1180     SDValue LowerINT_TO_FPDirectMove(SDValue Op, SelectionDAG &DAG,
1181                                      const SDLoc &dl) const;
1182 
1183     SDValue LowerINT_TO_FPVector(SDValue Op, SelectionDAG &DAG,
1184                                  const SDLoc &dl) const;
1185 
1186     SDValue LowerTRUNCATEVector(SDValue Op, SelectionDAG &DAG) const;
1187 
1188     SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
1189     SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
1190 
1191     bool
1192     IsEligibleForTailCallOptimization(SDValue Callee,
1193                                       CallingConv::ID CalleeCC,
1194                                       bool isVarArg,
1195                                       const SmallVectorImpl<ISD::InputArg> &Ins,
1196                                       SelectionDAG& DAG) const;
1197 
1198     bool IsEligibleForTailCallOptimization_64SVR4(
1199         SDValue Callee, CallingConv::ID CalleeCC, const CallBase *CB,
1200         bool isVarArg, const SmallVectorImpl<ISD::OutputArg> &Outs,
1201         const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const;
1202 
1203     SDValue EmitTailCallLoadFPAndRetAddr(SelectionDAG &DAG, int SPDiff,
1204                                          SDValue Chain, SDValue &LROpOut,
1205                                          SDValue &FPOpOut,
1206                                          const SDLoc &dl) const;
1207 
1208     SDValue getTOCEntry(SelectionDAG &DAG, const SDLoc &dl, SDValue GA) const;
1209 
1210     SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1211     SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
1212     SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
1213     SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
1214     SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
1215     SDValue LowerGlobalTLSAddressAIX(SDValue Op, SelectionDAG &DAG) const;
1216     SDValue LowerGlobalTLSAddressLinux(SDValue Op, SelectionDAG &DAG) const;
1217     SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
1218     SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
1219     SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
1220     SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1221     SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1222     SDValue LowerINLINEASM(SDValue Op, SelectionDAG &DAG) const;
1223     SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
1224     SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
1225     SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
1226     SDValue LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG) const;
1227     SDValue LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, SelectionDAG &DAG) const;
1228     SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
1229     SDValue LowerEH_DWARF_CFA(SDValue Op, SelectionDAG &DAG) const;
1230     SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
1231     SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
1232     SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
1233     SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
1234     SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
1235                            const SDLoc &dl) const;
1236     SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1237     SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
1238     SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const;
1239     SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const;
1240     SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const;
1241     SDValue LowerFunnelShift(SDValue Op, SelectionDAG &DAG) const;
1242     SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
1243     SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
1244     SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1245     SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
1246     SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
1247     SDValue LowerBSWAP(SDValue Op, SelectionDAG &DAG) const;
1248     SDValue LowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const;
1249     SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
1250     SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
1251     SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
1252     SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
1253     SDValue LowerROTL(SDValue Op, SelectionDAG &DAG) const;
1254 
1255     SDValue LowerVectorLoad(SDValue Op, SelectionDAG &DAG) const;
1256     SDValue LowerVectorStore(SDValue Op, SelectionDAG &DAG) const;
1257 
1258     SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
1259                             CallingConv::ID CallConv, bool isVarArg,
1260                             const SmallVectorImpl<ISD::InputArg> &Ins,
1261                             const SDLoc &dl, SelectionDAG &DAG,
1262                             SmallVectorImpl<SDValue> &InVals) const;
1263 
1264     SDValue FinishCall(CallFlags CFlags, const SDLoc &dl, SelectionDAG &DAG,
1265                        SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
1266                        SDValue InFlag, SDValue Chain, SDValue CallSeqStart,
1267                        SDValue &Callee, int SPDiff, unsigned NumBytes,
1268                        const SmallVectorImpl<ISD::InputArg> &Ins,
1269                        SmallVectorImpl<SDValue> &InVals,
1270                        const CallBase *CB) const;
1271 
1272     SDValue
1273     LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1274                          const SmallVectorImpl<ISD::InputArg> &Ins,
1275                          const SDLoc &dl, SelectionDAG &DAG,
1276                          SmallVectorImpl<SDValue> &InVals) const override;
1277 
1278     SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
1279                       SmallVectorImpl<SDValue> &InVals) const override;
1280 
1281     bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1282                         bool isVarArg,
1283                         const SmallVectorImpl<ISD::OutputArg> &Outs,
1284                         LLVMContext &Context) const override;
1285 
1286     SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1287                         const SmallVectorImpl<ISD::OutputArg> &Outs,
1288                         const SmallVectorImpl<SDValue> &OutVals,
1289                         const SDLoc &dl, SelectionDAG &DAG) const override;
1290 
1291     SDValue extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
1292                               SelectionDAG &DAG, SDValue ArgVal,
1293                               const SDLoc &dl) const;
1294 
1295     SDValue LowerFormalArguments_AIX(
1296         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1297         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1298         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1299     SDValue LowerFormalArguments_64SVR4(
1300         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1301         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1302         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1303     SDValue LowerFormalArguments_32SVR4(
1304         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1305         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1306         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1307 
1308     SDValue createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
1309                                        SDValue CallSeqStart,
1310                                        ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1311                                        const SDLoc &dl) const;
1312 
1313     SDValue LowerCall_64SVR4(SDValue Chain, SDValue Callee, CallFlags CFlags,
1314                              const SmallVectorImpl<ISD::OutputArg> &Outs,
1315                              const SmallVectorImpl<SDValue> &OutVals,
1316                              const SmallVectorImpl<ISD::InputArg> &Ins,
1317                              const SDLoc &dl, SelectionDAG &DAG,
1318                              SmallVectorImpl<SDValue> &InVals,
1319                              const CallBase *CB) const;
1320     SDValue LowerCall_32SVR4(SDValue Chain, SDValue Callee, CallFlags CFlags,
1321                              const SmallVectorImpl<ISD::OutputArg> &Outs,
1322                              const SmallVectorImpl<SDValue> &OutVals,
1323                              const SmallVectorImpl<ISD::InputArg> &Ins,
1324                              const SDLoc &dl, SelectionDAG &DAG,
1325                              SmallVectorImpl<SDValue> &InVals,
1326                              const CallBase *CB) const;
1327     SDValue LowerCall_AIX(SDValue Chain, SDValue Callee, CallFlags CFlags,
1328                           const SmallVectorImpl<ISD::OutputArg> &Outs,
1329                           const SmallVectorImpl<SDValue> &OutVals,
1330                           const SmallVectorImpl<ISD::InputArg> &Ins,
1331                           const SDLoc &dl, SelectionDAG &DAG,
1332                           SmallVectorImpl<SDValue> &InVals,
1333                           const CallBase *CB) const;
1334 
1335     SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1336     SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1337     SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) const;
1338 
1339     SDValue DAGCombineExtBoolTrunc(SDNode *N, DAGCombinerInfo &DCI) const;
1340     SDValue DAGCombineBuildVector(SDNode *N, DAGCombinerInfo &DCI) const;
1341     SDValue DAGCombineTruncBoolExt(SDNode *N, DAGCombinerInfo &DCI) const;
1342     SDValue combineStoreFPToInt(SDNode *N, DAGCombinerInfo &DCI) const;
1343     SDValue combineFPToIntToFP(SDNode *N, DAGCombinerInfo &DCI) const;
1344     SDValue combineSHL(SDNode *N, DAGCombinerInfo &DCI) const;
1345     SDValue combineSRA(SDNode *N, DAGCombinerInfo &DCI) const;
1346     SDValue combineSRL(SDNode *N, DAGCombinerInfo &DCI) const;
1347     SDValue combineMUL(SDNode *N, DAGCombinerInfo &DCI) const;
1348     SDValue combineADD(SDNode *N, DAGCombinerInfo &DCI) const;
1349     SDValue combineFMALike(SDNode *N, DAGCombinerInfo &DCI) const;
1350     SDValue combineTRUNCATE(SDNode *N, DAGCombinerInfo &DCI) const;
1351     SDValue combineSetCC(SDNode *N, DAGCombinerInfo &DCI) const;
1352     SDValue combineABS(SDNode *N, DAGCombinerInfo &DCI) const;
1353     SDValue combineVSelect(SDNode *N, DAGCombinerInfo &DCI) const;
1354     SDValue combineVectorShuffle(ShuffleVectorSDNode *SVN,
1355                                  SelectionDAG &DAG) const;
1356     SDValue combineVReverseMemOP(ShuffleVectorSDNode *SVN, LSBaseSDNode *LSBase,
1357                                  DAGCombinerInfo &DCI) const;
1358 
1359     /// ConvertSETCCToSubtract - looks at SETCC that compares ints. It replaces
1360     /// SETCC with integer subtraction when (1) there is a legal way of doing it
1361     /// (2) keeping the result of comparison in GPR has performance benefit.
1362     SDValue ConvertSETCCToSubtract(SDNode *N, DAGCombinerInfo &DCI) const;
1363 
1364     SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1365                             int &RefinementSteps, bool &UseOneConstNR,
1366                             bool Reciprocal) const override;
1367     SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1368                              int &RefinementSteps) const override;
1369     SDValue getSqrtInputTest(SDValue Operand, SelectionDAG &DAG,
1370                              const DenormalMode &Mode) const override;
1371     SDValue getSqrtResultForDenormInput(SDValue Operand,
1372                                         SelectionDAG &DAG) const override;
1373     unsigned combineRepeatedFPDivisors() const override;
1374 
1375     SDValue
1376     combineElementTruncationToVectorTruncation(SDNode *N,
1377                                                DAGCombinerInfo &DCI) const;
1378 
1379     /// lowerToVINSERTH - Return the SDValue if this VECTOR_SHUFFLE can be
1380     /// handled by the VINSERTH instruction introduced in ISA 3.0. This is
1381     /// essentially any shuffle of v8i16 vectors that just inserts one element
1382     /// from one vector into the other.
1383     SDValue lowerToVINSERTH(ShuffleVectorSDNode *N, SelectionDAG &DAG) const;
1384 
1385     /// lowerToVINSERTB - Return the SDValue if this VECTOR_SHUFFLE can be
1386     /// handled by the VINSERTB instruction introduced in ISA 3.0. This is
1387     /// essentially v16i8 vector version of VINSERTH.
1388     SDValue lowerToVINSERTB(ShuffleVectorSDNode *N, SelectionDAG &DAG) const;
1389 
1390     /// lowerToXXSPLTI32DX - Return the SDValue if this VECTOR_SHUFFLE can be
1391     /// handled by the XXSPLTI32DX instruction introduced in ISA 3.1.
1392     SDValue lowerToXXSPLTI32DX(ShuffleVectorSDNode *N, SelectionDAG &DAG) const;
1393 
1394     // Return whether the call instruction can potentially be optimized to a
1395     // tail call. This will cause the optimizers to attempt to move, or
1396     // duplicate return instructions to help enable tail call optimizations.
1397     bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
1398     bool hasBitPreservingFPLogic(EVT VT) const override;
1399     bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override;
1400 
1401     /// getAddrModeForFlags - Based on the set of address flags, select the most
1402     /// optimal instruction format to match by.
1403     PPC::AddrMode getAddrModeForFlags(unsigned Flags) const;
1404 
1405     /// computeMOFlags - Given a node N and it's Parent (a MemSDNode), compute
1406     /// the address flags of the load/store instruction that is to be matched.
1407     /// The address flags are stored in a map, which is then searched
1408     /// through to determine the optimal load/store instruction format.
1409     unsigned computeMOFlags(const SDNode *Parent, SDValue N,
1410                             SelectionDAG &DAG) const;
1411   }; // end class PPCTargetLowering
1412 
1413   namespace PPC {
1414 
1415     FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
1416                              const TargetLibraryInfo *LibInfo);
1417 
1418   } // end namespace PPC
1419 
1420   bool isIntS16Immediate(SDNode *N, int16_t &Imm);
1421   bool isIntS16Immediate(SDValue Op, int16_t &Imm);
1422   bool isIntS34Immediate(SDNode *N, int64_t &Imm);
1423   bool isIntS34Immediate(SDValue Op, int64_t &Imm);
1424 
1425   bool convertToNonDenormSingle(APInt &ArgAPInt);
1426   bool convertToNonDenormSingle(APFloat &ArgAPFloat);
1427   bool checkConvertToNonDenormSingle(APFloat &ArgAPFloat);
1428 
1429 } // end namespace llvm
1430 
1431 #endif // LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H
1432