1 //===-- SIFoldOperands.cpp - Fold operands --- ----------------------------===//
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 /// \file
8 //===----------------------------------------------------------------------===//
9 //
10
11 #include "AMDGPU.h"
12 #include "GCNSubtarget.h"
13 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
14 #include "SIMachineFunctionInfo.h"
15 #include "llvm/ADT/DepthFirstIterator.h"
16 #include "llvm/CodeGen/MachineFunctionPass.h"
17 #include "llvm/CodeGen/MachineOperand.h"
18
19 #define DEBUG_TYPE "si-fold-operands"
20 using namespace llvm;
21
22 namespace {
23
24 struct FoldCandidate {
25 MachineInstr *UseMI;
26 union {
27 MachineOperand *OpToFold;
28 uint64_t ImmToFold;
29 int FrameIndexToFold;
30 };
31 int ShrinkOpcode;
32 unsigned UseOpNo;
33 MachineOperand::MachineOperandType Kind;
34 bool Commuted;
35
FoldCandidate__anon62acd4700111::FoldCandidate36 FoldCandidate(MachineInstr *MI, unsigned OpNo, MachineOperand *FoldOp,
37 bool Commuted_ = false,
38 int ShrinkOp = -1) :
39 UseMI(MI), OpToFold(nullptr), ShrinkOpcode(ShrinkOp), UseOpNo(OpNo),
40 Kind(FoldOp->getType()),
41 Commuted(Commuted_) {
42 if (FoldOp->isImm()) {
43 ImmToFold = FoldOp->getImm();
44 } else if (FoldOp->isFI()) {
45 FrameIndexToFold = FoldOp->getIndex();
46 } else {
47 assert(FoldOp->isReg() || FoldOp->isGlobal());
48 OpToFold = FoldOp;
49 }
50 }
51
isFI__anon62acd4700111::FoldCandidate52 bool isFI() const {
53 return Kind == MachineOperand::MO_FrameIndex;
54 }
55
isImm__anon62acd4700111::FoldCandidate56 bool isImm() const {
57 return Kind == MachineOperand::MO_Immediate;
58 }
59
isReg__anon62acd4700111::FoldCandidate60 bool isReg() const {
61 return Kind == MachineOperand::MO_Register;
62 }
63
isGlobal__anon62acd4700111::FoldCandidate64 bool isGlobal() const { return Kind == MachineOperand::MO_GlobalAddress; }
65
needsShrink__anon62acd4700111::FoldCandidate66 bool needsShrink() const { return ShrinkOpcode != -1; }
67 };
68
69 class SIFoldOperands : public MachineFunctionPass {
70 public:
71 static char ID;
72 MachineRegisterInfo *MRI;
73 const SIInstrInfo *TII;
74 const SIRegisterInfo *TRI;
75 const GCNSubtarget *ST;
76 const SIMachineFunctionInfo *MFI;
77
78 bool frameIndexMayFold(const MachineInstr &UseMI, int OpNo,
79 const MachineOperand &OpToFold) const;
80
81 bool updateOperand(FoldCandidate &Fold) const;
82
83 bool canUseImmWithOpSel(FoldCandidate &Fold) const;
84
85 bool tryFoldImmWithOpSel(FoldCandidate &Fold) const;
86
87 bool tryAddToFoldList(SmallVectorImpl<FoldCandidate> &FoldList,
88 MachineInstr *MI, unsigned OpNo,
89 MachineOperand *OpToFold) const;
90 bool isUseSafeToFold(const MachineInstr &MI,
91 const MachineOperand &UseMO) const;
92 bool
93 getRegSeqInit(SmallVectorImpl<std::pair<MachineOperand *, unsigned>> &Defs,
94 Register UseReg, uint8_t OpTy) const;
95 bool tryToFoldACImm(const MachineOperand &OpToFold, MachineInstr *UseMI,
96 unsigned UseOpIdx,
97 SmallVectorImpl<FoldCandidate> &FoldList) const;
98 void foldOperand(MachineOperand &OpToFold,
99 MachineInstr *UseMI,
100 int UseOpIdx,
101 SmallVectorImpl<FoldCandidate> &FoldList,
102 SmallVectorImpl<MachineInstr *> &CopiesToReplace) const;
103
104 MachineOperand *getImmOrMaterializedImm(MachineOperand &Op) const;
105 bool tryConstantFoldOp(MachineInstr *MI) const;
106 bool tryFoldCndMask(MachineInstr &MI) const;
107 bool tryFoldZeroHighBits(MachineInstr &MI) const;
108 bool foldInstOperand(MachineInstr &MI, MachineOperand &OpToFold) const;
109 bool tryFoldFoldableCopy(MachineInstr &MI,
110 MachineOperand *&CurrentKnownM0Val) const;
111
112 const MachineOperand *isClamp(const MachineInstr &MI) const;
113 bool tryFoldClamp(MachineInstr &MI);
114
115 std::pair<const MachineOperand *, int> isOMod(const MachineInstr &MI) const;
116 bool tryFoldOMod(MachineInstr &MI);
117 bool tryFoldRegSequence(MachineInstr &MI);
118 bool tryFoldPhiAGPR(MachineInstr &MI);
119 bool tryFoldLoad(MachineInstr &MI);
120
121 bool tryOptimizeAGPRPhis(MachineBasicBlock &MBB);
122
123 public:
SIFoldOperands()124 SIFoldOperands() : MachineFunctionPass(ID) {
125 initializeSIFoldOperandsPass(*PassRegistry::getPassRegistry());
126 }
127
128 bool runOnMachineFunction(MachineFunction &MF) override;
129
getPassName() const130 StringRef getPassName() const override { return "SI Fold Operands"; }
131
getAnalysisUsage(AnalysisUsage & AU) const132 void getAnalysisUsage(AnalysisUsage &AU) const override {
133 AU.setPreservesCFG();
134 MachineFunctionPass::getAnalysisUsage(AU);
135 }
136 };
137
138 } // End anonymous namespace.
139
140 INITIALIZE_PASS(SIFoldOperands, DEBUG_TYPE,
141 "SI Fold Operands", false, false)
142
143 char SIFoldOperands::ID = 0;
144
145 char &llvm::SIFoldOperandsID = SIFoldOperands::ID;
146
getRegOpRC(const MachineRegisterInfo & MRI,const TargetRegisterInfo & TRI,const MachineOperand & MO)147 static const TargetRegisterClass *getRegOpRC(const MachineRegisterInfo &MRI,
148 const TargetRegisterInfo &TRI,
149 const MachineOperand &MO) {
150 const TargetRegisterClass *RC = MRI.getRegClass(MO.getReg());
151 if (const TargetRegisterClass *SubRC =
152 TRI.getSubRegisterClass(RC, MO.getSubReg()))
153 RC = SubRC;
154 return RC;
155 }
156
157 // Map multiply-accumulate opcode to corresponding multiply-add opcode if any.
macToMad(unsigned Opc)158 static unsigned macToMad(unsigned Opc) {
159 switch (Opc) {
160 case AMDGPU::V_MAC_F32_e64:
161 return AMDGPU::V_MAD_F32_e64;
162 case AMDGPU::V_MAC_F16_e64:
163 return AMDGPU::V_MAD_F16_e64;
164 case AMDGPU::V_FMAC_F32_e64:
165 return AMDGPU::V_FMA_F32_e64;
166 case AMDGPU::V_FMAC_F16_e64:
167 return AMDGPU::V_FMA_F16_gfx9_e64;
168 case AMDGPU::V_FMAC_F16_t16_e64:
169 return AMDGPU::V_FMA_F16_gfx9_e64;
170 case AMDGPU::V_FMAC_LEGACY_F32_e64:
171 return AMDGPU::V_FMA_LEGACY_F32_e64;
172 case AMDGPU::V_FMAC_F64_e64:
173 return AMDGPU::V_FMA_F64_e64;
174 }
175 return AMDGPU::INSTRUCTION_LIST_END;
176 }
177
178 // TODO: Add heuristic that the frame index might not fit in the addressing mode
179 // immediate offset to avoid materializing in loops.
frameIndexMayFold(const MachineInstr & UseMI,int OpNo,const MachineOperand & OpToFold) const180 bool SIFoldOperands::frameIndexMayFold(const MachineInstr &UseMI, int OpNo,
181 const MachineOperand &OpToFold) const {
182 if (!OpToFold.isFI())
183 return false;
184
185 const unsigned Opc = UseMI.getOpcode();
186 if (TII->isMUBUF(UseMI))
187 return OpNo == AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vaddr);
188 if (!TII->isFLATScratch(UseMI))
189 return false;
190
191 int SIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::saddr);
192 if (OpNo == SIdx)
193 return true;
194
195 int VIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vaddr);
196 return OpNo == VIdx && SIdx == -1;
197 }
198
createSIFoldOperandsPass()199 FunctionPass *llvm::createSIFoldOperandsPass() {
200 return new SIFoldOperands();
201 }
202
canUseImmWithOpSel(FoldCandidate & Fold) const203 bool SIFoldOperands::canUseImmWithOpSel(FoldCandidate &Fold) const {
204 MachineInstr *MI = Fold.UseMI;
205 MachineOperand &Old = MI->getOperand(Fold.UseOpNo);
206 const uint64_t TSFlags = MI->getDesc().TSFlags;
207
208 assert(Old.isReg() && Fold.isImm());
209
210 if (!(TSFlags & SIInstrFlags::IsPacked) || (TSFlags & SIInstrFlags::IsMAI) ||
211 (TSFlags & SIInstrFlags::IsWMMA) || (TSFlags & SIInstrFlags::IsSWMMAC) ||
212 (ST->hasDOTOpSelHazard() && (TSFlags & SIInstrFlags::IsDOT)))
213 return false;
214
215 unsigned Opcode = MI->getOpcode();
216 int OpNo = MI->getOperandNo(&Old);
217 uint8_t OpType = TII->get(Opcode).operands()[OpNo].OperandType;
218 switch (OpType) {
219 default:
220 return false;
221 case AMDGPU::OPERAND_REG_IMM_V2FP16:
222 case AMDGPU::OPERAND_REG_IMM_V2BF16:
223 case AMDGPU::OPERAND_REG_IMM_V2INT16:
224 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
225 case AMDGPU::OPERAND_REG_INLINE_C_V2BF16:
226 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
227 break;
228 }
229
230 return true;
231 }
232
tryFoldImmWithOpSel(FoldCandidate & Fold) const233 bool SIFoldOperands::tryFoldImmWithOpSel(FoldCandidate &Fold) const {
234 MachineInstr *MI = Fold.UseMI;
235 MachineOperand &Old = MI->getOperand(Fold.UseOpNo);
236 unsigned Opcode = MI->getOpcode();
237 int OpNo = MI->getOperandNo(&Old);
238 uint8_t OpType = TII->get(Opcode).operands()[OpNo].OperandType;
239
240 // If the literal can be inlined as-is, apply it and short-circuit the
241 // tests below. The main motivation for this is to avoid unintuitive
242 // uses of opsel.
243 if (AMDGPU::isInlinableLiteralV216(Fold.ImmToFold, OpType)) {
244 Old.ChangeToImmediate(Fold.ImmToFold);
245 return true;
246 }
247
248 // Refer to op_sel/op_sel_hi and check if we can change the immediate and
249 // op_sel in a way that allows an inline constant.
250 int ModIdx = -1;
251 unsigned SrcIdx = ~0;
252 if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0)) {
253 ModIdx = AMDGPU::OpName::src0_modifiers;
254 SrcIdx = 0;
255 } else if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1)) {
256 ModIdx = AMDGPU::OpName::src1_modifiers;
257 SrcIdx = 1;
258 } else if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2)) {
259 ModIdx = AMDGPU::OpName::src2_modifiers;
260 SrcIdx = 2;
261 }
262 assert(ModIdx != -1);
263 ModIdx = AMDGPU::getNamedOperandIdx(Opcode, ModIdx);
264 MachineOperand &Mod = MI->getOperand(ModIdx);
265 unsigned ModVal = Mod.getImm();
266
267 uint16_t ImmLo = static_cast<uint16_t>(
268 Fold.ImmToFold >> (ModVal & SISrcMods::OP_SEL_0 ? 16 : 0));
269 uint16_t ImmHi = static_cast<uint16_t>(
270 Fold.ImmToFold >> (ModVal & SISrcMods::OP_SEL_1 ? 16 : 0));
271 uint32_t Imm = (static_cast<uint32_t>(ImmHi) << 16) | ImmLo;
272 unsigned NewModVal = ModVal & ~(SISrcMods::OP_SEL_0 | SISrcMods::OP_SEL_1);
273
274 // Helper function that attempts to inline the given value with a newly
275 // chosen opsel pattern.
276 auto tryFoldToInline = [&](uint32_t Imm) -> bool {
277 if (AMDGPU::isInlinableLiteralV216(Imm, OpType)) {
278 Mod.setImm(NewModVal | SISrcMods::OP_SEL_1);
279 Old.ChangeToImmediate(Imm);
280 return true;
281 }
282
283 // Try to shuffle the halves around and leverage opsel to get an inline
284 // constant.
285 uint16_t Lo = static_cast<uint16_t>(Imm);
286 uint16_t Hi = static_cast<uint16_t>(Imm >> 16);
287 if (Lo == Hi) {
288 if (AMDGPU::isInlinableLiteralV216(Lo, OpType)) {
289 Mod.setImm(NewModVal);
290 Old.ChangeToImmediate(Lo);
291 return true;
292 }
293
294 if (static_cast<int16_t>(Lo) < 0) {
295 int32_t SExt = static_cast<int16_t>(Lo);
296 if (AMDGPU::isInlinableLiteralV216(SExt, OpType)) {
297 Mod.setImm(NewModVal);
298 Old.ChangeToImmediate(SExt);
299 return true;
300 }
301 }
302
303 // This check is only useful for integer instructions
304 if (OpType == AMDGPU::OPERAND_REG_IMM_V2INT16 ||
305 OpType == AMDGPU::OPERAND_REG_INLINE_AC_V2INT16) {
306 if (AMDGPU::isInlinableLiteralV216(Lo << 16, OpType)) {
307 Mod.setImm(NewModVal | SISrcMods::OP_SEL_0 | SISrcMods::OP_SEL_1);
308 Old.ChangeToImmediate(static_cast<uint32_t>(Lo) << 16);
309 return true;
310 }
311 }
312 } else {
313 uint32_t Swapped = (static_cast<uint32_t>(Lo) << 16) | Hi;
314 if (AMDGPU::isInlinableLiteralV216(Swapped, OpType)) {
315 Mod.setImm(NewModVal | SISrcMods::OP_SEL_0);
316 Old.ChangeToImmediate(Swapped);
317 return true;
318 }
319 }
320
321 return false;
322 };
323
324 if (tryFoldToInline(Imm))
325 return true;
326
327 // Replace integer addition by subtraction and vice versa if it allows
328 // folding the immediate to an inline constant.
329 //
330 // We should only ever get here for SrcIdx == 1 due to canonicalization
331 // earlier in the pipeline, but we double-check here to be safe / fully
332 // general.
333 bool IsUAdd = Opcode == AMDGPU::V_PK_ADD_U16;
334 bool IsUSub = Opcode == AMDGPU::V_PK_SUB_U16;
335 if (SrcIdx == 1 && (IsUAdd || IsUSub)) {
336 unsigned ClampIdx =
337 AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::clamp);
338 bool Clamp = MI->getOperand(ClampIdx).getImm() != 0;
339
340 if (!Clamp) {
341 uint16_t NegLo = -static_cast<uint16_t>(Imm);
342 uint16_t NegHi = -static_cast<uint16_t>(Imm >> 16);
343 uint32_t NegImm = (static_cast<uint32_t>(NegHi) << 16) | NegLo;
344
345 if (tryFoldToInline(NegImm)) {
346 unsigned NegOpcode =
347 IsUAdd ? AMDGPU::V_PK_SUB_U16 : AMDGPU::V_PK_ADD_U16;
348 MI->setDesc(TII->get(NegOpcode));
349 return true;
350 }
351 }
352 }
353
354 return false;
355 }
356
updateOperand(FoldCandidate & Fold) const357 bool SIFoldOperands::updateOperand(FoldCandidate &Fold) const {
358 MachineInstr *MI = Fold.UseMI;
359 MachineOperand &Old = MI->getOperand(Fold.UseOpNo);
360 assert(Old.isReg());
361
362 if (Fold.isImm() && canUseImmWithOpSel(Fold)) {
363 if (tryFoldImmWithOpSel(Fold))
364 return true;
365
366 // We can't represent the candidate as an inline constant. Try as a literal
367 // with the original opsel, checking constant bus limitations.
368 MachineOperand New = MachineOperand::CreateImm(Fold.ImmToFold);
369 int OpNo = MI->getOperandNo(&Old);
370 if (!TII->isOperandLegal(*MI, OpNo, &New))
371 return false;
372 Old.ChangeToImmediate(Fold.ImmToFold);
373 return true;
374 }
375
376 if ((Fold.isImm() || Fold.isFI() || Fold.isGlobal()) && Fold.needsShrink()) {
377 MachineBasicBlock *MBB = MI->getParent();
378 auto Liveness = MBB->computeRegisterLiveness(TRI, AMDGPU::VCC, MI, 16);
379 if (Liveness != MachineBasicBlock::LQR_Dead) {
380 LLVM_DEBUG(dbgs() << "Not shrinking " << MI << " due to vcc liveness\n");
381 return false;
382 }
383
384 int Op32 = Fold.ShrinkOpcode;
385 MachineOperand &Dst0 = MI->getOperand(0);
386 MachineOperand &Dst1 = MI->getOperand(1);
387 assert(Dst0.isDef() && Dst1.isDef());
388
389 bool HaveNonDbgCarryUse = !MRI->use_nodbg_empty(Dst1.getReg());
390
391 const TargetRegisterClass *Dst0RC = MRI->getRegClass(Dst0.getReg());
392 Register NewReg0 = MRI->createVirtualRegister(Dst0RC);
393
394 MachineInstr *Inst32 = TII->buildShrunkInst(*MI, Op32);
395
396 if (HaveNonDbgCarryUse) {
397 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(AMDGPU::COPY),
398 Dst1.getReg())
399 .addReg(AMDGPU::VCC, RegState::Kill);
400 }
401
402 // Keep the old instruction around to avoid breaking iterators, but
403 // replace it with a dummy instruction to remove uses.
404 //
405 // FIXME: We should not invert how this pass looks at operands to avoid
406 // this. Should track set of foldable movs instead of looking for uses
407 // when looking at a use.
408 Dst0.setReg(NewReg0);
409 for (unsigned I = MI->getNumOperands() - 1; I > 0; --I)
410 MI->removeOperand(I);
411 MI->setDesc(TII->get(AMDGPU::IMPLICIT_DEF));
412
413 if (Fold.Commuted)
414 TII->commuteInstruction(*Inst32, false);
415 return true;
416 }
417
418 assert(!Fold.needsShrink() && "not handled");
419
420 if (Fold.isImm()) {
421 if (Old.isTied()) {
422 int NewMFMAOpc = AMDGPU::getMFMAEarlyClobberOp(MI->getOpcode());
423 if (NewMFMAOpc == -1)
424 return false;
425 MI->setDesc(TII->get(NewMFMAOpc));
426 MI->untieRegOperand(0);
427 }
428 Old.ChangeToImmediate(Fold.ImmToFold);
429 return true;
430 }
431
432 if (Fold.isGlobal()) {
433 Old.ChangeToGA(Fold.OpToFold->getGlobal(), Fold.OpToFold->getOffset(),
434 Fold.OpToFold->getTargetFlags());
435 return true;
436 }
437
438 if (Fold.isFI()) {
439 Old.ChangeToFrameIndex(Fold.FrameIndexToFold);
440 return true;
441 }
442
443 MachineOperand *New = Fold.OpToFold;
444 Old.substVirtReg(New->getReg(), New->getSubReg(), *TRI);
445 Old.setIsUndef(New->isUndef());
446 return true;
447 }
448
isUseMIInFoldList(ArrayRef<FoldCandidate> FoldList,const MachineInstr * MI)449 static bool isUseMIInFoldList(ArrayRef<FoldCandidate> FoldList,
450 const MachineInstr *MI) {
451 return any_of(FoldList, [&](const auto &C) { return C.UseMI == MI; });
452 }
453
appendFoldCandidate(SmallVectorImpl<FoldCandidate> & FoldList,MachineInstr * MI,unsigned OpNo,MachineOperand * FoldOp,bool Commuted=false,int ShrinkOp=-1)454 static void appendFoldCandidate(SmallVectorImpl<FoldCandidate> &FoldList,
455 MachineInstr *MI, unsigned OpNo,
456 MachineOperand *FoldOp, bool Commuted = false,
457 int ShrinkOp = -1) {
458 // Skip additional folding on the same operand.
459 for (FoldCandidate &Fold : FoldList)
460 if (Fold.UseMI == MI && Fold.UseOpNo == OpNo)
461 return;
462 LLVM_DEBUG(dbgs() << "Append " << (Commuted ? "commuted" : "normal")
463 << " operand " << OpNo << "\n " << *MI);
464 FoldList.emplace_back(MI, OpNo, FoldOp, Commuted, ShrinkOp);
465 }
466
tryAddToFoldList(SmallVectorImpl<FoldCandidate> & FoldList,MachineInstr * MI,unsigned OpNo,MachineOperand * OpToFold) const467 bool SIFoldOperands::tryAddToFoldList(SmallVectorImpl<FoldCandidate> &FoldList,
468 MachineInstr *MI, unsigned OpNo,
469 MachineOperand *OpToFold) const {
470 const unsigned Opc = MI->getOpcode();
471
472 auto tryToFoldAsFMAAKorMK = [&]() {
473 if (!OpToFold->isImm())
474 return false;
475
476 const bool TryAK = OpNo == 3;
477 const unsigned NewOpc = TryAK ? AMDGPU::S_FMAAK_F32 : AMDGPU::S_FMAMK_F32;
478 MI->setDesc(TII->get(NewOpc));
479
480 // We have to fold into operand which would be Imm not into OpNo.
481 bool FoldAsFMAAKorMK =
482 tryAddToFoldList(FoldList, MI, TryAK ? 3 : 2, OpToFold);
483 if (FoldAsFMAAKorMK) {
484 // Untie Src2 of fmac.
485 MI->untieRegOperand(3);
486 // For fmamk swap operands 1 and 2 if OpToFold was meant for operand 1.
487 if (OpNo == 1) {
488 MachineOperand &Op1 = MI->getOperand(1);
489 MachineOperand &Op2 = MI->getOperand(2);
490 Register OldReg = Op1.getReg();
491 // Operand 2 might be an inlinable constant
492 if (Op2.isImm()) {
493 Op1.ChangeToImmediate(Op2.getImm());
494 Op2.ChangeToRegister(OldReg, false);
495 } else {
496 Op1.setReg(Op2.getReg());
497 Op2.setReg(OldReg);
498 }
499 }
500 return true;
501 }
502 MI->setDesc(TII->get(Opc));
503 return false;
504 };
505
506 bool IsLegal = TII->isOperandLegal(*MI, OpNo, OpToFold);
507 if (!IsLegal && OpToFold->isImm()) {
508 FoldCandidate Fold(MI, OpNo, OpToFold);
509 IsLegal = canUseImmWithOpSel(Fold);
510 }
511
512 if (!IsLegal) {
513 // Special case for v_mac_{f16, f32}_e64 if we are trying to fold into src2
514 unsigned NewOpc = macToMad(Opc);
515 if (NewOpc != AMDGPU::INSTRUCTION_LIST_END) {
516 // Check if changing this to a v_mad_{f16, f32} instruction will allow us
517 // to fold the operand.
518 MI->setDesc(TII->get(NewOpc));
519 bool AddOpSel = !AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::op_sel) &&
520 AMDGPU::hasNamedOperand(NewOpc, AMDGPU::OpName::op_sel);
521 if (AddOpSel)
522 MI->addOperand(MachineOperand::CreateImm(0));
523 bool FoldAsMAD = tryAddToFoldList(FoldList, MI, OpNo, OpToFold);
524 if (FoldAsMAD) {
525 MI->untieRegOperand(OpNo);
526 return true;
527 }
528 if (AddOpSel)
529 MI->removeOperand(MI->getNumExplicitOperands() - 1);
530 MI->setDesc(TII->get(Opc));
531 }
532
533 // Special case for s_fmac_f32 if we are trying to fold into Src2.
534 // By transforming into fmaak we can untie Src2 and make folding legal.
535 if (Opc == AMDGPU::S_FMAC_F32 && OpNo == 3) {
536 if (tryToFoldAsFMAAKorMK())
537 return true;
538 }
539
540 // Special case for s_setreg_b32
541 if (OpToFold->isImm()) {
542 unsigned ImmOpc = 0;
543 if (Opc == AMDGPU::S_SETREG_B32)
544 ImmOpc = AMDGPU::S_SETREG_IMM32_B32;
545 else if (Opc == AMDGPU::S_SETREG_B32_mode)
546 ImmOpc = AMDGPU::S_SETREG_IMM32_B32_mode;
547 if (ImmOpc) {
548 MI->setDesc(TII->get(ImmOpc));
549 appendFoldCandidate(FoldList, MI, OpNo, OpToFold);
550 return true;
551 }
552 }
553
554 // If we are already folding into another operand of MI, then
555 // we can't commute the instruction, otherwise we risk making the
556 // other fold illegal.
557 if (isUseMIInFoldList(FoldList, MI))
558 return false;
559
560 // Operand is not legal, so try to commute the instruction to
561 // see if this makes it possible to fold.
562 unsigned CommuteOpNo = TargetInstrInfo::CommuteAnyOperandIndex;
563 bool CanCommute = TII->findCommutedOpIndices(*MI, OpNo, CommuteOpNo);
564 if (!CanCommute)
565 return false;
566
567 // One of operands might be an Imm operand, and OpNo may refer to it after
568 // the call of commuteInstruction() below. Such situations are avoided
569 // here explicitly as OpNo must be a register operand to be a candidate
570 // for memory folding.
571 if (!MI->getOperand(OpNo).isReg() || !MI->getOperand(CommuteOpNo).isReg())
572 return false;
573
574 if (!TII->commuteInstruction(*MI, false, OpNo, CommuteOpNo))
575 return false;
576
577 int Op32 = -1;
578 if (!TII->isOperandLegal(*MI, CommuteOpNo, OpToFold)) {
579 if ((Opc != AMDGPU::V_ADD_CO_U32_e64 && Opc != AMDGPU::V_SUB_CO_U32_e64 &&
580 Opc != AMDGPU::V_SUBREV_CO_U32_e64) || // FIXME
581 (!OpToFold->isImm() && !OpToFold->isFI() && !OpToFold->isGlobal())) {
582 TII->commuteInstruction(*MI, false, OpNo, CommuteOpNo);
583 return false;
584 }
585
586 // Verify the other operand is a VGPR, otherwise we would violate the
587 // constant bus restriction.
588 MachineOperand &OtherOp = MI->getOperand(OpNo);
589 if (!OtherOp.isReg() ||
590 !TII->getRegisterInfo().isVGPR(*MRI, OtherOp.getReg()))
591 return false;
592
593 assert(MI->getOperand(1).isDef());
594
595 // Make sure to get the 32-bit version of the commuted opcode.
596 unsigned MaybeCommutedOpc = MI->getOpcode();
597 Op32 = AMDGPU::getVOPe32(MaybeCommutedOpc);
598 }
599
600 appendFoldCandidate(FoldList, MI, CommuteOpNo, OpToFold, true, Op32);
601 return true;
602 }
603
604 // Inlineable constant might have been folded into Imm operand of fmaak or
605 // fmamk and we are trying to fold a non-inlinable constant.
606 if ((Opc == AMDGPU::S_FMAAK_F32 || Opc == AMDGPU::S_FMAMK_F32) &&
607 !OpToFold->isReg() && !TII->isInlineConstant(*OpToFold)) {
608 unsigned ImmIdx = Opc == AMDGPU::S_FMAAK_F32 ? 3 : 2;
609 MachineOperand &OpImm = MI->getOperand(ImmIdx);
610 if (!OpImm.isReg() &&
611 TII->isInlineConstant(*MI, MI->getOperand(OpNo), OpImm))
612 return tryToFoldAsFMAAKorMK();
613 }
614
615 // Special case for s_fmac_f32 if we are trying to fold into Src0 or Src1.
616 // By changing into fmamk we can untie Src2.
617 // If folding for Src0 happens first and it is identical operand to Src1 we
618 // should avoid transforming into fmamk which requires commuting as it would
619 // cause folding into Src1 to fail later on due to wrong OpNo used.
620 if (Opc == AMDGPU::S_FMAC_F32 &&
621 (OpNo != 1 || !MI->getOperand(1).isIdenticalTo(MI->getOperand(2)))) {
622 if (tryToFoldAsFMAAKorMK())
623 return true;
624 }
625
626 // Check the case where we might introduce a second constant operand to a
627 // scalar instruction
628 if (TII->isSALU(MI->getOpcode())) {
629 const MCInstrDesc &InstDesc = MI->getDesc();
630 const MCOperandInfo &OpInfo = InstDesc.operands()[OpNo];
631
632 // Fine if the operand can be encoded as an inline constant
633 if (!OpToFold->isReg() && !TII->isInlineConstant(*OpToFold, OpInfo)) {
634 // Otherwise check for another constant
635 for (unsigned i = 0, e = InstDesc.getNumOperands(); i != e; ++i) {
636 auto &Op = MI->getOperand(i);
637 if (OpNo != i && !Op.isReg() &&
638 !TII->isInlineConstant(Op, InstDesc.operands()[i]))
639 return false;
640 }
641 }
642 }
643
644 appendFoldCandidate(FoldList, MI, OpNo, OpToFold);
645 return true;
646 }
647
isUseSafeToFold(const MachineInstr & MI,const MachineOperand & UseMO) const648 bool SIFoldOperands::isUseSafeToFold(const MachineInstr &MI,
649 const MachineOperand &UseMO) const {
650 // Operands of SDWA instructions must be registers.
651 return !TII->isSDWA(MI);
652 }
653
654 // Find a def of the UseReg, check if it is a reg_sequence and find initializers
655 // for each subreg, tracking it to foldable inline immediate if possible.
656 // Returns true on success.
getRegSeqInit(SmallVectorImpl<std::pair<MachineOperand *,unsigned>> & Defs,Register UseReg,uint8_t OpTy) const657 bool SIFoldOperands::getRegSeqInit(
658 SmallVectorImpl<std::pair<MachineOperand *, unsigned>> &Defs,
659 Register UseReg, uint8_t OpTy) const {
660 MachineInstr *Def = MRI->getVRegDef(UseReg);
661 if (!Def || !Def->isRegSequence())
662 return false;
663
664 for (unsigned I = 1, E = Def->getNumExplicitOperands(); I < E; I += 2) {
665 MachineOperand *Sub = &Def->getOperand(I);
666 assert(Sub->isReg());
667
668 for (MachineInstr *SubDef = MRI->getVRegDef(Sub->getReg());
669 SubDef && Sub->isReg() && Sub->getReg().isVirtual() &&
670 !Sub->getSubReg() && TII->isFoldableCopy(*SubDef);
671 SubDef = MRI->getVRegDef(Sub->getReg())) {
672 MachineOperand *Op = &SubDef->getOperand(1);
673 if (Op->isImm()) {
674 if (TII->isInlineConstant(*Op, OpTy))
675 Sub = Op;
676 break;
677 }
678 if (!Op->isReg() || Op->getReg().isPhysical())
679 break;
680 Sub = Op;
681 }
682
683 Defs.emplace_back(Sub, Def->getOperand(I + 1).getImm());
684 }
685
686 return true;
687 }
688
tryToFoldACImm(const MachineOperand & OpToFold,MachineInstr * UseMI,unsigned UseOpIdx,SmallVectorImpl<FoldCandidate> & FoldList) const689 bool SIFoldOperands::tryToFoldACImm(
690 const MachineOperand &OpToFold, MachineInstr *UseMI, unsigned UseOpIdx,
691 SmallVectorImpl<FoldCandidate> &FoldList) const {
692 const MCInstrDesc &Desc = UseMI->getDesc();
693 if (UseOpIdx >= Desc.getNumOperands())
694 return false;
695
696 if (!AMDGPU::isSISrcInlinableOperand(Desc, UseOpIdx))
697 return false;
698
699 uint8_t OpTy = Desc.operands()[UseOpIdx].OperandType;
700 if (OpToFold.isImm() && TII->isInlineConstant(OpToFold, OpTy) &&
701 TII->isOperandLegal(*UseMI, UseOpIdx, &OpToFold)) {
702 UseMI->getOperand(UseOpIdx).ChangeToImmediate(OpToFold.getImm());
703 return true;
704 }
705
706 if (!OpToFold.isReg())
707 return false;
708
709 Register UseReg = OpToFold.getReg();
710 if (!UseReg.isVirtual())
711 return false;
712
713 if (isUseMIInFoldList(FoldList, UseMI))
714 return false;
715
716 // Maybe it is just a COPY of an immediate itself.
717 MachineInstr *Def = MRI->getVRegDef(UseReg);
718 MachineOperand &UseOp = UseMI->getOperand(UseOpIdx);
719 if (!UseOp.getSubReg() && Def && TII->isFoldableCopy(*Def)) {
720 MachineOperand &DefOp = Def->getOperand(1);
721 if (DefOp.isImm() && TII->isInlineConstant(DefOp, OpTy) &&
722 TII->isOperandLegal(*UseMI, UseOpIdx, &DefOp)) {
723 UseMI->getOperand(UseOpIdx).ChangeToImmediate(DefOp.getImm());
724 return true;
725 }
726 }
727
728 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs;
729 if (!getRegSeqInit(Defs, UseReg, OpTy))
730 return false;
731
732 int32_t Imm;
733 for (unsigned I = 0, E = Defs.size(); I != E; ++I) {
734 const MachineOperand *Op = Defs[I].first;
735 if (!Op->isImm())
736 return false;
737
738 auto SubImm = Op->getImm();
739 if (!I) {
740 Imm = SubImm;
741 if (!TII->isInlineConstant(*Op, OpTy) ||
742 !TII->isOperandLegal(*UseMI, UseOpIdx, Op))
743 return false;
744
745 continue;
746 }
747 if (Imm != SubImm)
748 return false; // Can only fold splat constants
749 }
750
751 appendFoldCandidate(FoldList, UseMI, UseOpIdx, Defs[0].first);
752 return true;
753 }
754
foldOperand(MachineOperand & OpToFold,MachineInstr * UseMI,int UseOpIdx,SmallVectorImpl<FoldCandidate> & FoldList,SmallVectorImpl<MachineInstr * > & CopiesToReplace) const755 void SIFoldOperands::foldOperand(
756 MachineOperand &OpToFold,
757 MachineInstr *UseMI,
758 int UseOpIdx,
759 SmallVectorImpl<FoldCandidate> &FoldList,
760 SmallVectorImpl<MachineInstr *> &CopiesToReplace) const {
761 const MachineOperand *UseOp = &UseMI->getOperand(UseOpIdx);
762
763 if (!isUseSafeToFold(*UseMI, *UseOp))
764 return;
765
766 // FIXME: Fold operands with subregs.
767 if (UseOp->isReg() && OpToFold.isReg() &&
768 (UseOp->isImplicit() || UseOp->getSubReg() != AMDGPU::NoSubRegister))
769 return;
770
771 // Special case for REG_SEQUENCE: We can't fold literals into
772 // REG_SEQUENCE instructions, so we have to fold them into the
773 // uses of REG_SEQUENCE.
774 if (UseMI->isRegSequence()) {
775 Register RegSeqDstReg = UseMI->getOperand(0).getReg();
776 unsigned RegSeqDstSubReg = UseMI->getOperand(UseOpIdx + 1).getImm();
777
778 // Grab the use operands first
779 SmallVector<MachineOperand *, 4> UsesToProcess;
780 for (auto &Use : MRI->use_nodbg_operands(RegSeqDstReg))
781 UsesToProcess.push_back(&Use);
782 for (auto *RSUse : UsesToProcess) {
783 MachineInstr *RSUseMI = RSUse->getParent();
784
785 if (tryToFoldACImm(UseMI->getOperand(0), RSUseMI,
786 RSUseMI->getOperandNo(RSUse), FoldList))
787 continue;
788
789 if (RSUse->getSubReg() != RegSeqDstSubReg)
790 continue;
791
792 foldOperand(OpToFold, RSUseMI, RSUseMI->getOperandNo(RSUse), FoldList,
793 CopiesToReplace);
794 }
795 return;
796 }
797
798 if (tryToFoldACImm(OpToFold, UseMI, UseOpIdx, FoldList))
799 return;
800
801 if (frameIndexMayFold(*UseMI, UseOpIdx, OpToFold)) {
802 // Verify that this is a stack access.
803 // FIXME: Should probably use stack pseudos before frame lowering.
804
805 if (TII->isMUBUF(*UseMI)) {
806 if (TII->getNamedOperand(*UseMI, AMDGPU::OpName::srsrc)->getReg() !=
807 MFI->getScratchRSrcReg())
808 return;
809
810 // Ensure this is either relative to the current frame or the current
811 // wave.
812 MachineOperand &SOff =
813 *TII->getNamedOperand(*UseMI, AMDGPU::OpName::soffset);
814 if (!SOff.isImm() || SOff.getImm() != 0)
815 return;
816 }
817
818 // A frame index will resolve to a positive constant, so it should always be
819 // safe to fold the addressing mode, even pre-GFX9.
820 UseMI->getOperand(UseOpIdx).ChangeToFrameIndex(OpToFold.getIndex());
821
822 const unsigned Opc = UseMI->getOpcode();
823 if (TII->isFLATScratch(*UseMI) &&
824 AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::vaddr) &&
825 !AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::saddr)) {
826 unsigned NewOpc = AMDGPU::getFlatScratchInstSSfromSV(Opc);
827 UseMI->setDesc(TII->get(NewOpc));
828 }
829
830 return;
831 }
832
833 bool FoldingImmLike =
834 OpToFold.isImm() || OpToFold.isFI() || OpToFold.isGlobal();
835
836 if (FoldingImmLike && UseMI->isCopy()) {
837 Register DestReg = UseMI->getOperand(0).getReg();
838 Register SrcReg = UseMI->getOperand(1).getReg();
839 assert(SrcReg.isVirtual());
840
841 const TargetRegisterClass *SrcRC = MRI->getRegClass(SrcReg);
842
843 // Don't fold into a copy to a physical register with the same class. Doing
844 // so would interfere with the register coalescer's logic which would avoid
845 // redundant initializations.
846 if (DestReg.isPhysical() && SrcRC->contains(DestReg))
847 return;
848
849 const TargetRegisterClass *DestRC = TRI->getRegClassForReg(*MRI, DestReg);
850 if (!DestReg.isPhysical()) {
851 if (DestRC == &AMDGPU::AGPR_32RegClass &&
852 TII->isInlineConstant(OpToFold, AMDGPU::OPERAND_REG_INLINE_C_INT32)) {
853 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64));
854 UseMI->getOperand(1).ChangeToImmediate(OpToFold.getImm());
855 CopiesToReplace.push_back(UseMI);
856 return;
857 }
858 }
859
860 // In order to fold immediates into copies, we need to change the
861 // copy to a MOV.
862
863 unsigned MovOp = TII->getMovOpcode(DestRC);
864 if (MovOp == AMDGPU::COPY)
865 return;
866
867 MachineInstr::mop_iterator ImpOpI = UseMI->implicit_operands().begin();
868 MachineInstr::mop_iterator ImpOpE = UseMI->implicit_operands().end();
869 while (ImpOpI != ImpOpE) {
870 MachineInstr::mop_iterator Tmp = ImpOpI;
871 ImpOpI++;
872 UseMI->removeOperand(UseMI->getOperandNo(Tmp));
873 }
874 UseMI->setDesc(TII->get(MovOp));
875
876 if (MovOp == AMDGPU::V_MOV_B16_t16_e64) {
877 const auto &SrcOp = UseMI->getOperand(UseOpIdx);
878 MachineOperand NewSrcOp(SrcOp);
879 MachineFunction *MF = UseMI->getParent()->getParent();
880 UseMI->removeOperand(1);
881 UseMI->addOperand(*MF, MachineOperand::CreateImm(0)); // src0_modifiers
882 UseMI->addOperand(NewSrcOp); // src0
883 UseMI->addOperand(*MF, MachineOperand::CreateImm(0)); // op_sel
884 UseOpIdx = 2;
885 UseOp = &UseMI->getOperand(UseOpIdx);
886 }
887 CopiesToReplace.push_back(UseMI);
888 } else {
889 if (UseMI->isCopy() && OpToFold.isReg() &&
890 UseMI->getOperand(0).getReg().isVirtual() &&
891 !UseMI->getOperand(1).getSubReg()) {
892 LLVM_DEBUG(dbgs() << "Folding " << OpToFold << "\n into " << *UseMI);
893 unsigned Size = TII->getOpSize(*UseMI, 1);
894 Register UseReg = OpToFold.getReg();
895 UseMI->getOperand(1).setReg(UseReg);
896 UseMI->getOperand(1).setSubReg(OpToFold.getSubReg());
897 UseMI->getOperand(1).setIsKill(false);
898 CopiesToReplace.push_back(UseMI);
899 OpToFold.setIsKill(false);
900
901 // Remove kill flags as kills may now be out of order with uses.
902 MRI->clearKillFlags(OpToFold.getReg());
903
904 // That is very tricky to store a value into an AGPR. v_accvgpr_write_b32
905 // can only accept VGPR or inline immediate. Recreate a reg_sequence with
906 // its initializers right here, so we will rematerialize immediates and
907 // avoid copies via different reg classes.
908 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs;
909 if (Size > 4 && TRI->isAGPR(*MRI, UseMI->getOperand(0).getReg()) &&
910 getRegSeqInit(Defs, UseReg, AMDGPU::OPERAND_REG_INLINE_C_INT32)) {
911 const DebugLoc &DL = UseMI->getDebugLoc();
912 MachineBasicBlock &MBB = *UseMI->getParent();
913
914 UseMI->setDesc(TII->get(AMDGPU::REG_SEQUENCE));
915 for (unsigned I = UseMI->getNumOperands() - 1; I > 0; --I)
916 UseMI->removeOperand(I);
917
918 MachineInstrBuilder B(*MBB.getParent(), UseMI);
919 DenseMap<TargetInstrInfo::RegSubRegPair, Register> VGPRCopies;
920 SmallSetVector<TargetInstrInfo::RegSubRegPair, 32> SeenAGPRs;
921 for (unsigned I = 0; I < Size / 4; ++I) {
922 MachineOperand *Def = Defs[I].first;
923 TargetInstrInfo::RegSubRegPair CopyToVGPR;
924 if (Def->isImm() &&
925 TII->isInlineConstant(*Def, AMDGPU::OPERAND_REG_INLINE_C_INT32)) {
926 int64_t Imm = Def->getImm();
927
928 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass);
929 BuildMI(MBB, UseMI, DL,
930 TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), Tmp).addImm(Imm);
931 B.addReg(Tmp);
932 } else if (Def->isReg() && TRI->isAGPR(*MRI, Def->getReg())) {
933 auto Src = getRegSubRegPair(*Def);
934 Def->setIsKill(false);
935 if (!SeenAGPRs.insert(Src)) {
936 // We cannot build a reg_sequence out of the same registers, they
937 // must be copied. Better do it here before copyPhysReg() created
938 // several reads to do the AGPR->VGPR->AGPR copy.
939 CopyToVGPR = Src;
940 } else {
941 B.addReg(Src.Reg, Def->isUndef() ? RegState::Undef : 0,
942 Src.SubReg);
943 }
944 } else {
945 assert(Def->isReg());
946 Def->setIsKill(false);
947 auto Src = getRegSubRegPair(*Def);
948
949 // Direct copy from SGPR to AGPR is not possible. To avoid creation
950 // of exploded copies SGPR->VGPR->AGPR in the copyPhysReg() later,
951 // create a copy here and track if we already have such a copy.
952 if (TRI->isSGPRReg(*MRI, Src.Reg)) {
953 CopyToVGPR = Src;
954 } else {
955 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass);
956 BuildMI(MBB, UseMI, DL, TII->get(AMDGPU::COPY), Tmp).add(*Def);
957 B.addReg(Tmp);
958 }
959 }
960
961 if (CopyToVGPR.Reg) {
962 Register Vgpr;
963 if (VGPRCopies.count(CopyToVGPR)) {
964 Vgpr = VGPRCopies[CopyToVGPR];
965 } else {
966 Vgpr = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
967 BuildMI(MBB, UseMI, DL, TII->get(AMDGPU::COPY), Vgpr).add(*Def);
968 VGPRCopies[CopyToVGPR] = Vgpr;
969 }
970 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass);
971 BuildMI(MBB, UseMI, DL,
972 TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), Tmp).addReg(Vgpr);
973 B.addReg(Tmp);
974 }
975
976 B.addImm(Defs[I].second);
977 }
978 LLVM_DEBUG(dbgs() << "Folded " << *UseMI);
979 return;
980 }
981
982 if (Size != 4)
983 return;
984
985 Register Reg0 = UseMI->getOperand(0).getReg();
986 Register Reg1 = UseMI->getOperand(1).getReg();
987 if (TRI->isAGPR(*MRI, Reg0) && TRI->isVGPR(*MRI, Reg1))
988 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64));
989 else if (TRI->isVGPR(*MRI, Reg0) && TRI->isAGPR(*MRI, Reg1))
990 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_READ_B32_e64));
991 else if (ST->hasGFX90AInsts() && TRI->isAGPR(*MRI, Reg0) &&
992 TRI->isAGPR(*MRI, Reg1))
993 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_MOV_B32));
994 return;
995 }
996
997 unsigned UseOpc = UseMI->getOpcode();
998 if (UseOpc == AMDGPU::V_READFIRSTLANE_B32 ||
999 (UseOpc == AMDGPU::V_READLANE_B32 &&
1000 (int)UseOpIdx ==
1001 AMDGPU::getNamedOperandIdx(UseOpc, AMDGPU::OpName::src0))) {
1002 // %vgpr = V_MOV_B32 imm
1003 // %sgpr = V_READFIRSTLANE_B32 %vgpr
1004 // =>
1005 // %sgpr = S_MOV_B32 imm
1006 if (FoldingImmLike) {
1007 if (execMayBeModifiedBeforeUse(*MRI,
1008 UseMI->getOperand(UseOpIdx).getReg(),
1009 *OpToFold.getParent(),
1010 *UseMI))
1011 return;
1012
1013 UseMI->setDesc(TII->get(AMDGPU::S_MOV_B32));
1014
1015 if (OpToFold.isImm())
1016 UseMI->getOperand(1).ChangeToImmediate(OpToFold.getImm());
1017 else
1018 UseMI->getOperand(1).ChangeToFrameIndex(OpToFold.getIndex());
1019 UseMI->removeOperand(2); // Remove exec read (or src1 for readlane)
1020 return;
1021 }
1022
1023 if (OpToFold.isReg() && TRI->isSGPRReg(*MRI, OpToFold.getReg())) {
1024 if (execMayBeModifiedBeforeUse(*MRI,
1025 UseMI->getOperand(UseOpIdx).getReg(),
1026 *OpToFold.getParent(),
1027 *UseMI))
1028 return;
1029
1030 // %vgpr = COPY %sgpr0
1031 // %sgpr1 = V_READFIRSTLANE_B32 %vgpr
1032 // =>
1033 // %sgpr1 = COPY %sgpr0
1034 UseMI->setDesc(TII->get(AMDGPU::COPY));
1035 UseMI->getOperand(1).setReg(OpToFold.getReg());
1036 UseMI->getOperand(1).setSubReg(OpToFold.getSubReg());
1037 UseMI->getOperand(1).setIsKill(false);
1038 UseMI->removeOperand(2); // Remove exec read (or src1 for readlane)
1039 return;
1040 }
1041 }
1042
1043 const MCInstrDesc &UseDesc = UseMI->getDesc();
1044
1045 // Don't fold into target independent nodes. Target independent opcodes
1046 // don't have defined register classes.
1047 if (UseDesc.isVariadic() || UseOp->isImplicit() ||
1048 UseDesc.operands()[UseOpIdx].RegClass == -1)
1049 return;
1050 }
1051
1052 if (!FoldingImmLike) {
1053 if (OpToFold.isReg() && ST->needsAlignedVGPRs()) {
1054 // Don't fold if OpToFold doesn't hold an aligned register.
1055 const TargetRegisterClass *RC =
1056 TRI->getRegClassForReg(*MRI, OpToFold.getReg());
1057 assert(RC);
1058 if (TRI->hasVectorRegisters(RC) && OpToFold.getSubReg()) {
1059 unsigned SubReg = OpToFold.getSubReg();
1060 if (const TargetRegisterClass *SubRC =
1061 TRI->getSubRegisterClass(RC, SubReg))
1062 RC = SubRC;
1063 }
1064
1065 if (!RC || !TRI->isProperlyAlignedRC(*RC))
1066 return;
1067 }
1068
1069 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold);
1070
1071 // FIXME: We could try to change the instruction from 64-bit to 32-bit
1072 // to enable more folding opportunities. The shrink operands pass
1073 // already does this.
1074 return;
1075 }
1076
1077
1078 const MCInstrDesc &FoldDesc = OpToFold.getParent()->getDesc();
1079 const TargetRegisterClass *FoldRC =
1080 TRI->getRegClass(FoldDesc.operands()[0].RegClass);
1081
1082 // Split 64-bit constants into 32-bits for folding.
1083 if (UseOp->getSubReg() && AMDGPU::getRegBitWidth(*FoldRC) == 64) {
1084 Register UseReg = UseOp->getReg();
1085 const TargetRegisterClass *UseRC = MRI->getRegClass(UseReg);
1086 if (AMDGPU::getRegBitWidth(*UseRC) != 64)
1087 return;
1088
1089 APInt Imm(64, OpToFold.getImm());
1090 if (UseOp->getSubReg() == AMDGPU::sub0) {
1091 Imm = Imm.getLoBits(32);
1092 } else {
1093 assert(UseOp->getSubReg() == AMDGPU::sub1);
1094 Imm = Imm.getHiBits(32);
1095 }
1096
1097 MachineOperand ImmOp = MachineOperand::CreateImm(Imm.getSExtValue());
1098 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &ImmOp);
1099 return;
1100 }
1101
1102 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold);
1103 }
1104
evalBinaryInstruction(unsigned Opcode,int32_t & Result,uint32_t LHS,uint32_t RHS)1105 static bool evalBinaryInstruction(unsigned Opcode, int32_t &Result,
1106 uint32_t LHS, uint32_t RHS) {
1107 switch (Opcode) {
1108 case AMDGPU::V_AND_B32_e64:
1109 case AMDGPU::V_AND_B32_e32:
1110 case AMDGPU::S_AND_B32:
1111 Result = LHS & RHS;
1112 return true;
1113 case AMDGPU::V_OR_B32_e64:
1114 case AMDGPU::V_OR_B32_e32:
1115 case AMDGPU::S_OR_B32:
1116 Result = LHS | RHS;
1117 return true;
1118 case AMDGPU::V_XOR_B32_e64:
1119 case AMDGPU::V_XOR_B32_e32:
1120 case AMDGPU::S_XOR_B32:
1121 Result = LHS ^ RHS;
1122 return true;
1123 case AMDGPU::S_XNOR_B32:
1124 Result = ~(LHS ^ RHS);
1125 return true;
1126 case AMDGPU::S_NAND_B32:
1127 Result = ~(LHS & RHS);
1128 return true;
1129 case AMDGPU::S_NOR_B32:
1130 Result = ~(LHS | RHS);
1131 return true;
1132 case AMDGPU::S_ANDN2_B32:
1133 Result = LHS & ~RHS;
1134 return true;
1135 case AMDGPU::S_ORN2_B32:
1136 Result = LHS | ~RHS;
1137 return true;
1138 case AMDGPU::V_LSHL_B32_e64:
1139 case AMDGPU::V_LSHL_B32_e32:
1140 case AMDGPU::S_LSHL_B32:
1141 // The instruction ignores the high bits for out of bounds shifts.
1142 Result = LHS << (RHS & 31);
1143 return true;
1144 case AMDGPU::V_LSHLREV_B32_e64:
1145 case AMDGPU::V_LSHLREV_B32_e32:
1146 Result = RHS << (LHS & 31);
1147 return true;
1148 case AMDGPU::V_LSHR_B32_e64:
1149 case AMDGPU::V_LSHR_B32_e32:
1150 case AMDGPU::S_LSHR_B32:
1151 Result = LHS >> (RHS & 31);
1152 return true;
1153 case AMDGPU::V_LSHRREV_B32_e64:
1154 case AMDGPU::V_LSHRREV_B32_e32:
1155 Result = RHS >> (LHS & 31);
1156 return true;
1157 case AMDGPU::V_ASHR_I32_e64:
1158 case AMDGPU::V_ASHR_I32_e32:
1159 case AMDGPU::S_ASHR_I32:
1160 Result = static_cast<int32_t>(LHS) >> (RHS & 31);
1161 return true;
1162 case AMDGPU::V_ASHRREV_I32_e64:
1163 case AMDGPU::V_ASHRREV_I32_e32:
1164 Result = static_cast<int32_t>(RHS) >> (LHS & 31);
1165 return true;
1166 default:
1167 return false;
1168 }
1169 }
1170
getMovOpc(bool IsScalar)1171 static unsigned getMovOpc(bool IsScalar) {
1172 return IsScalar ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32;
1173 }
1174
mutateCopyOp(MachineInstr & MI,const MCInstrDesc & NewDesc)1175 static void mutateCopyOp(MachineInstr &MI, const MCInstrDesc &NewDesc) {
1176 MI.setDesc(NewDesc);
1177
1178 // Remove any leftover implicit operands from mutating the instruction. e.g.
1179 // if we replace an s_and_b32 with a copy, we don't need the implicit scc def
1180 // anymore.
1181 const MCInstrDesc &Desc = MI.getDesc();
1182 unsigned NumOps = Desc.getNumOperands() + Desc.implicit_uses().size() +
1183 Desc.implicit_defs().size();
1184
1185 for (unsigned I = MI.getNumOperands() - 1; I >= NumOps; --I)
1186 MI.removeOperand(I);
1187 }
1188
1189 MachineOperand *
getImmOrMaterializedImm(MachineOperand & Op) const1190 SIFoldOperands::getImmOrMaterializedImm(MachineOperand &Op) const {
1191 // If this has a subregister, it obviously is a register source.
1192 if (!Op.isReg() || Op.getSubReg() != AMDGPU::NoSubRegister ||
1193 !Op.getReg().isVirtual())
1194 return &Op;
1195
1196 MachineInstr *Def = MRI->getVRegDef(Op.getReg());
1197 if (Def && Def->isMoveImmediate()) {
1198 MachineOperand &ImmSrc = Def->getOperand(1);
1199 if (ImmSrc.isImm())
1200 return &ImmSrc;
1201 }
1202
1203 return &Op;
1204 }
1205
1206 // Try to simplify operations with a constant that may appear after instruction
1207 // selection.
1208 // TODO: See if a frame index with a fixed offset can fold.
tryConstantFoldOp(MachineInstr * MI) const1209 bool SIFoldOperands::tryConstantFoldOp(MachineInstr *MI) const {
1210 if (!MI->allImplicitDefsAreDead())
1211 return false;
1212
1213 unsigned Opc = MI->getOpcode();
1214
1215 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
1216 if (Src0Idx == -1)
1217 return false;
1218 MachineOperand *Src0 = getImmOrMaterializedImm(MI->getOperand(Src0Idx));
1219
1220 if ((Opc == AMDGPU::V_NOT_B32_e64 || Opc == AMDGPU::V_NOT_B32_e32 ||
1221 Opc == AMDGPU::S_NOT_B32) &&
1222 Src0->isImm()) {
1223 MI->getOperand(1).ChangeToImmediate(~Src0->getImm());
1224 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_NOT_B32)));
1225 return true;
1226 }
1227
1228 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
1229 if (Src1Idx == -1)
1230 return false;
1231 MachineOperand *Src1 = getImmOrMaterializedImm(MI->getOperand(Src1Idx));
1232
1233 if (!Src0->isImm() && !Src1->isImm())
1234 return false;
1235
1236 // and k0, k1 -> v_mov_b32 (k0 & k1)
1237 // or k0, k1 -> v_mov_b32 (k0 | k1)
1238 // xor k0, k1 -> v_mov_b32 (k0 ^ k1)
1239 if (Src0->isImm() && Src1->isImm()) {
1240 int32_t NewImm;
1241 if (!evalBinaryInstruction(Opc, NewImm, Src0->getImm(), Src1->getImm()))
1242 return false;
1243
1244 bool IsSGPR = TRI->isSGPRReg(*MRI, MI->getOperand(0).getReg());
1245
1246 // Be careful to change the right operand, src0 may belong to a different
1247 // instruction.
1248 MI->getOperand(Src0Idx).ChangeToImmediate(NewImm);
1249 MI->removeOperand(Src1Idx);
1250 mutateCopyOp(*MI, TII->get(getMovOpc(IsSGPR)));
1251 return true;
1252 }
1253
1254 if (!MI->isCommutable())
1255 return false;
1256
1257 if (Src0->isImm() && !Src1->isImm()) {
1258 std::swap(Src0, Src1);
1259 std::swap(Src0Idx, Src1Idx);
1260 }
1261
1262 int32_t Src1Val = static_cast<int32_t>(Src1->getImm());
1263 if (Opc == AMDGPU::V_OR_B32_e64 ||
1264 Opc == AMDGPU::V_OR_B32_e32 ||
1265 Opc == AMDGPU::S_OR_B32) {
1266 if (Src1Val == 0) {
1267 // y = or x, 0 => y = copy x
1268 MI->removeOperand(Src1Idx);
1269 mutateCopyOp(*MI, TII->get(AMDGPU::COPY));
1270 } else if (Src1Val == -1) {
1271 // y = or x, -1 => y = v_mov_b32 -1
1272 MI->removeOperand(Src1Idx);
1273 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_OR_B32)));
1274 } else
1275 return false;
1276
1277 return true;
1278 }
1279
1280 if (Opc == AMDGPU::V_AND_B32_e64 || Opc == AMDGPU::V_AND_B32_e32 ||
1281 Opc == AMDGPU::S_AND_B32) {
1282 if (Src1Val == 0) {
1283 // y = and x, 0 => y = v_mov_b32 0
1284 MI->removeOperand(Src0Idx);
1285 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_AND_B32)));
1286 } else if (Src1Val == -1) {
1287 // y = and x, -1 => y = copy x
1288 MI->removeOperand(Src1Idx);
1289 mutateCopyOp(*MI, TII->get(AMDGPU::COPY));
1290 } else
1291 return false;
1292
1293 return true;
1294 }
1295
1296 if (Opc == AMDGPU::V_XOR_B32_e64 || Opc == AMDGPU::V_XOR_B32_e32 ||
1297 Opc == AMDGPU::S_XOR_B32) {
1298 if (Src1Val == 0) {
1299 // y = xor x, 0 => y = copy x
1300 MI->removeOperand(Src1Idx);
1301 mutateCopyOp(*MI, TII->get(AMDGPU::COPY));
1302 return true;
1303 }
1304 }
1305
1306 return false;
1307 }
1308
1309 // Try to fold an instruction into a simpler one
tryFoldCndMask(MachineInstr & MI) const1310 bool SIFoldOperands::tryFoldCndMask(MachineInstr &MI) const {
1311 unsigned Opc = MI.getOpcode();
1312 if (Opc != AMDGPU::V_CNDMASK_B32_e32 && Opc != AMDGPU::V_CNDMASK_B32_e64 &&
1313 Opc != AMDGPU::V_CNDMASK_B64_PSEUDO)
1314 return false;
1315
1316 MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
1317 MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
1318 if (!Src1->isIdenticalTo(*Src0)) {
1319 auto *Src0Imm = getImmOrMaterializedImm(*Src0);
1320 auto *Src1Imm = getImmOrMaterializedImm(*Src1);
1321 if (!Src1Imm->isIdenticalTo(*Src0Imm))
1322 return false;
1323 }
1324
1325 int Src1ModIdx =
1326 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1_modifiers);
1327 int Src0ModIdx =
1328 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0_modifiers);
1329 if ((Src1ModIdx != -1 && MI.getOperand(Src1ModIdx).getImm() != 0) ||
1330 (Src0ModIdx != -1 && MI.getOperand(Src0ModIdx).getImm() != 0))
1331 return false;
1332
1333 LLVM_DEBUG(dbgs() << "Folded " << MI << " into ");
1334 auto &NewDesc =
1335 TII->get(Src0->isReg() ? (unsigned)AMDGPU::COPY : getMovOpc(false));
1336 int Src2Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2);
1337 if (Src2Idx != -1)
1338 MI.removeOperand(Src2Idx);
1339 MI.removeOperand(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1));
1340 if (Src1ModIdx != -1)
1341 MI.removeOperand(Src1ModIdx);
1342 if (Src0ModIdx != -1)
1343 MI.removeOperand(Src0ModIdx);
1344 mutateCopyOp(MI, NewDesc);
1345 LLVM_DEBUG(dbgs() << MI);
1346 return true;
1347 }
1348
tryFoldZeroHighBits(MachineInstr & MI) const1349 bool SIFoldOperands::tryFoldZeroHighBits(MachineInstr &MI) const {
1350 if (MI.getOpcode() != AMDGPU::V_AND_B32_e64 &&
1351 MI.getOpcode() != AMDGPU::V_AND_B32_e32)
1352 return false;
1353
1354 MachineOperand *Src0 = getImmOrMaterializedImm(MI.getOperand(1));
1355 if (!Src0->isImm() || Src0->getImm() != 0xffff)
1356 return false;
1357
1358 Register Src1 = MI.getOperand(2).getReg();
1359 MachineInstr *SrcDef = MRI->getVRegDef(Src1);
1360 if (!ST->zeroesHigh16BitsOfDest(SrcDef->getOpcode()))
1361 return false;
1362
1363 Register Dst = MI.getOperand(0).getReg();
1364 MRI->replaceRegWith(Dst, Src1);
1365 if (!MI.getOperand(2).isKill())
1366 MRI->clearKillFlags(Src1);
1367 MI.eraseFromParent();
1368 return true;
1369 }
1370
foldInstOperand(MachineInstr & MI,MachineOperand & OpToFold) const1371 bool SIFoldOperands::foldInstOperand(MachineInstr &MI,
1372 MachineOperand &OpToFold) const {
1373 // We need mutate the operands of new mov instructions to add implicit
1374 // uses of EXEC, but adding them invalidates the use_iterator, so defer
1375 // this.
1376 SmallVector<MachineInstr *, 4> CopiesToReplace;
1377 SmallVector<FoldCandidate, 4> FoldList;
1378 MachineOperand &Dst = MI.getOperand(0);
1379 bool Changed = false;
1380
1381 if (OpToFold.isImm()) {
1382 for (auto &UseMI :
1383 make_early_inc_range(MRI->use_nodbg_instructions(Dst.getReg()))) {
1384 // Folding the immediate may reveal operations that can be constant
1385 // folded or replaced with a copy. This can happen for example after
1386 // frame indices are lowered to constants or from splitting 64-bit
1387 // constants.
1388 //
1389 // We may also encounter cases where one or both operands are
1390 // immediates materialized into a register, which would ordinarily not
1391 // be folded due to multiple uses or operand constraints.
1392 if (tryConstantFoldOp(&UseMI)) {
1393 LLVM_DEBUG(dbgs() << "Constant folded " << UseMI);
1394 Changed = true;
1395 }
1396 }
1397 }
1398
1399 SmallVector<MachineOperand *, 4> UsesToProcess;
1400 for (auto &Use : MRI->use_nodbg_operands(Dst.getReg()))
1401 UsesToProcess.push_back(&Use);
1402 for (auto *U : UsesToProcess) {
1403 MachineInstr *UseMI = U->getParent();
1404 foldOperand(OpToFold, UseMI, UseMI->getOperandNo(U), FoldList,
1405 CopiesToReplace);
1406 }
1407
1408 if (CopiesToReplace.empty() && FoldList.empty())
1409 return Changed;
1410
1411 MachineFunction *MF = MI.getParent()->getParent();
1412 // Make sure we add EXEC uses to any new v_mov instructions created.
1413 for (MachineInstr *Copy : CopiesToReplace)
1414 Copy->addImplicitDefUseOperands(*MF);
1415
1416 for (FoldCandidate &Fold : FoldList) {
1417 assert(!Fold.isReg() || Fold.OpToFold);
1418 if (Fold.isReg() && Fold.OpToFold->getReg().isVirtual()) {
1419 Register Reg = Fold.OpToFold->getReg();
1420 MachineInstr *DefMI = Fold.OpToFold->getParent();
1421 if (DefMI->readsRegister(AMDGPU::EXEC, TRI) &&
1422 execMayBeModifiedBeforeUse(*MRI, Reg, *DefMI, *Fold.UseMI))
1423 continue;
1424 }
1425 if (updateOperand(Fold)) {
1426 // Clear kill flags.
1427 if (Fold.isReg()) {
1428 assert(Fold.OpToFold && Fold.OpToFold->isReg());
1429 // FIXME: Probably shouldn't bother trying to fold if not an
1430 // SGPR. PeepholeOptimizer can eliminate redundant VGPR->VGPR
1431 // copies.
1432 MRI->clearKillFlags(Fold.OpToFold->getReg());
1433 }
1434 LLVM_DEBUG(dbgs() << "Folded source from " << MI << " into OpNo "
1435 << static_cast<int>(Fold.UseOpNo) << " of "
1436 << *Fold.UseMI);
1437 } else if (Fold.Commuted) {
1438 // Restoring instruction's original operand order if fold has failed.
1439 TII->commuteInstruction(*Fold.UseMI, false);
1440 }
1441 }
1442 return true;
1443 }
1444
tryFoldFoldableCopy(MachineInstr & MI,MachineOperand * & CurrentKnownM0Val) const1445 bool SIFoldOperands::tryFoldFoldableCopy(
1446 MachineInstr &MI, MachineOperand *&CurrentKnownM0Val) const {
1447 // Specially track simple redefs of m0 to the same value in a block, so we
1448 // can erase the later ones.
1449 if (MI.getOperand(0).getReg() == AMDGPU::M0) {
1450 MachineOperand &NewM0Val = MI.getOperand(1);
1451 if (CurrentKnownM0Val && CurrentKnownM0Val->isIdenticalTo(NewM0Val)) {
1452 MI.eraseFromParent();
1453 return true;
1454 }
1455
1456 // We aren't tracking other physical registers
1457 CurrentKnownM0Val = (NewM0Val.isReg() && NewM0Val.getReg().isPhysical())
1458 ? nullptr
1459 : &NewM0Val;
1460 return false;
1461 }
1462
1463 MachineOperand &OpToFold = MI.getOperand(1);
1464 bool FoldingImm = OpToFold.isImm() || OpToFold.isFI() || OpToFold.isGlobal();
1465
1466 // FIXME: We could also be folding things like TargetIndexes.
1467 if (!FoldingImm && !OpToFold.isReg())
1468 return false;
1469
1470 if (OpToFold.isReg() && !OpToFold.getReg().isVirtual())
1471 return false;
1472
1473 // Prevent folding operands backwards in the function. For example,
1474 // the COPY opcode must not be replaced by 1 in this example:
1475 //
1476 // %3 = COPY %vgpr0; VGPR_32:%3
1477 // ...
1478 // %vgpr0 = V_MOV_B32_e32 1, implicit %exec
1479 if (!MI.getOperand(0).getReg().isVirtual())
1480 return false;
1481
1482 bool Changed = foldInstOperand(MI, OpToFold);
1483
1484 // If we managed to fold all uses of this copy then we might as well
1485 // delete it now.
1486 // The only reason we need to follow chains of copies here is that
1487 // tryFoldRegSequence looks forward through copies before folding a
1488 // REG_SEQUENCE into its eventual users.
1489 auto *InstToErase = &MI;
1490 while (MRI->use_nodbg_empty(InstToErase->getOperand(0).getReg())) {
1491 auto &SrcOp = InstToErase->getOperand(1);
1492 auto SrcReg = SrcOp.isReg() ? SrcOp.getReg() : Register();
1493 InstToErase->eraseFromParent();
1494 Changed = true;
1495 InstToErase = nullptr;
1496 if (!SrcReg || SrcReg.isPhysical())
1497 break;
1498 InstToErase = MRI->getVRegDef(SrcReg);
1499 if (!InstToErase || !TII->isFoldableCopy(*InstToErase))
1500 break;
1501 }
1502
1503 if (InstToErase && InstToErase->isRegSequence() &&
1504 MRI->use_nodbg_empty(InstToErase->getOperand(0).getReg())) {
1505 InstToErase->eraseFromParent();
1506 Changed = true;
1507 }
1508
1509 return Changed;
1510 }
1511
1512 // Clamp patterns are canonically selected to v_max_* instructions, so only
1513 // handle them.
isClamp(const MachineInstr & MI) const1514 const MachineOperand *SIFoldOperands::isClamp(const MachineInstr &MI) const {
1515 unsigned Op = MI.getOpcode();
1516 switch (Op) {
1517 case AMDGPU::V_MAX_F32_e64:
1518 case AMDGPU::V_MAX_F16_e64:
1519 case AMDGPU::V_MAX_F16_t16_e64:
1520 case AMDGPU::V_MAX_F16_fake16_e64:
1521 case AMDGPU::V_MAX_F64_e64:
1522 case AMDGPU::V_MAX_NUM_F64_e64:
1523 case AMDGPU::V_PK_MAX_F16: {
1524 if (MI.mayRaiseFPException())
1525 return nullptr;
1526
1527 if (!TII->getNamedOperand(MI, AMDGPU::OpName::clamp)->getImm())
1528 return nullptr;
1529
1530 // Make sure sources are identical.
1531 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
1532 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
1533 if (!Src0->isReg() || !Src1->isReg() ||
1534 Src0->getReg() != Src1->getReg() ||
1535 Src0->getSubReg() != Src1->getSubReg() ||
1536 Src0->getSubReg() != AMDGPU::NoSubRegister)
1537 return nullptr;
1538
1539 // Can't fold up if we have modifiers.
1540 if (TII->hasModifiersSet(MI, AMDGPU::OpName::omod))
1541 return nullptr;
1542
1543 unsigned Src0Mods
1544 = TII->getNamedOperand(MI, AMDGPU::OpName::src0_modifiers)->getImm();
1545 unsigned Src1Mods
1546 = TII->getNamedOperand(MI, AMDGPU::OpName::src1_modifiers)->getImm();
1547
1548 // Having a 0 op_sel_hi would require swizzling the output in the source
1549 // instruction, which we can't do.
1550 unsigned UnsetMods = (Op == AMDGPU::V_PK_MAX_F16) ? SISrcMods::OP_SEL_1
1551 : 0u;
1552 if (Src0Mods != UnsetMods && Src1Mods != UnsetMods)
1553 return nullptr;
1554 return Src0;
1555 }
1556 default:
1557 return nullptr;
1558 }
1559 }
1560
1561 // FIXME: Clamp for v_mad_mixhi_f16 handled during isel.
tryFoldClamp(MachineInstr & MI)1562 bool SIFoldOperands::tryFoldClamp(MachineInstr &MI) {
1563 const MachineOperand *ClampSrc = isClamp(MI);
1564 if (!ClampSrc || !MRI->hasOneNonDBGUser(ClampSrc->getReg()))
1565 return false;
1566
1567 MachineInstr *Def = MRI->getVRegDef(ClampSrc->getReg());
1568
1569 // The type of clamp must be compatible.
1570 if (TII->getClampMask(*Def) != TII->getClampMask(MI))
1571 return false;
1572
1573 if (Def->mayRaiseFPException())
1574 return false;
1575
1576 MachineOperand *DefClamp = TII->getNamedOperand(*Def, AMDGPU::OpName::clamp);
1577 if (!DefClamp)
1578 return false;
1579
1580 LLVM_DEBUG(dbgs() << "Folding clamp " << *DefClamp << " into " << *Def);
1581
1582 // Clamp is applied after omod, so it is OK if omod is set.
1583 DefClamp->setImm(1);
1584
1585 Register DefReg = Def->getOperand(0).getReg();
1586 Register MIDstReg = MI.getOperand(0).getReg();
1587 if (TRI->isSGPRReg(*MRI, DefReg)) {
1588 // Pseudo scalar instructions have a SGPR for dst and clamp is a v_max*
1589 // instruction with a VGPR dst.
1590 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), TII->get(AMDGPU::COPY),
1591 MIDstReg)
1592 .addReg(DefReg);
1593 } else {
1594 MRI->replaceRegWith(MIDstReg, DefReg);
1595 }
1596 MI.eraseFromParent();
1597
1598 // Use of output modifiers forces VOP3 encoding for a VOP2 mac/fmac
1599 // instruction, so we might as well convert it to the more flexible VOP3-only
1600 // mad/fma form.
1601 if (TII->convertToThreeAddress(*Def, nullptr, nullptr))
1602 Def->eraseFromParent();
1603
1604 return true;
1605 }
1606
getOModValue(unsigned Opc,int64_t Val)1607 static int getOModValue(unsigned Opc, int64_t Val) {
1608 switch (Opc) {
1609 case AMDGPU::V_MUL_F64_e64:
1610 case AMDGPU::V_MUL_F64_pseudo_e64: {
1611 switch (Val) {
1612 case 0x3fe0000000000000: // 0.5
1613 return SIOutMods::DIV2;
1614 case 0x4000000000000000: // 2.0
1615 return SIOutMods::MUL2;
1616 case 0x4010000000000000: // 4.0
1617 return SIOutMods::MUL4;
1618 default:
1619 return SIOutMods::NONE;
1620 }
1621 }
1622 case AMDGPU::V_MUL_F32_e64: {
1623 switch (static_cast<uint32_t>(Val)) {
1624 case 0x3f000000: // 0.5
1625 return SIOutMods::DIV2;
1626 case 0x40000000: // 2.0
1627 return SIOutMods::MUL2;
1628 case 0x40800000: // 4.0
1629 return SIOutMods::MUL4;
1630 default:
1631 return SIOutMods::NONE;
1632 }
1633 }
1634 case AMDGPU::V_MUL_F16_e64:
1635 case AMDGPU::V_MUL_F16_t16_e64:
1636 case AMDGPU::V_MUL_F16_fake16_e64: {
1637 switch (static_cast<uint16_t>(Val)) {
1638 case 0x3800: // 0.5
1639 return SIOutMods::DIV2;
1640 case 0x4000: // 2.0
1641 return SIOutMods::MUL2;
1642 case 0x4400: // 4.0
1643 return SIOutMods::MUL4;
1644 default:
1645 return SIOutMods::NONE;
1646 }
1647 }
1648 default:
1649 llvm_unreachable("invalid mul opcode");
1650 }
1651 }
1652
1653 // FIXME: Does this really not support denormals with f16?
1654 // FIXME: Does this need to check IEEE mode bit? SNaNs are generally not
1655 // handled, so will anything other than that break?
1656 std::pair<const MachineOperand *, int>
isOMod(const MachineInstr & MI) const1657 SIFoldOperands::isOMod(const MachineInstr &MI) const {
1658 unsigned Op = MI.getOpcode();
1659 switch (Op) {
1660 case AMDGPU::V_MUL_F64_e64:
1661 case AMDGPU::V_MUL_F64_pseudo_e64:
1662 case AMDGPU::V_MUL_F32_e64:
1663 case AMDGPU::V_MUL_F16_t16_e64:
1664 case AMDGPU::V_MUL_F16_fake16_e64:
1665 case AMDGPU::V_MUL_F16_e64: {
1666 // If output denormals are enabled, omod is ignored.
1667 if ((Op == AMDGPU::V_MUL_F32_e64 &&
1668 MFI->getMode().FP32Denormals.Output != DenormalMode::PreserveSign) ||
1669 ((Op == AMDGPU::V_MUL_F64_e64 || Op == AMDGPU::V_MUL_F64_pseudo_e64 ||
1670 Op == AMDGPU::V_MUL_F16_e64 || Op == AMDGPU::V_MUL_F16_t16_e64 ||
1671 Op == AMDGPU::V_MUL_F16_fake16_e64) &&
1672 MFI->getMode().FP64FP16Denormals.Output !=
1673 DenormalMode::PreserveSign) ||
1674 MI.mayRaiseFPException())
1675 return std::pair(nullptr, SIOutMods::NONE);
1676
1677 const MachineOperand *RegOp = nullptr;
1678 const MachineOperand *ImmOp = nullptr;
1679 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
1680 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
1681 if (Src0->isImm()) {
1682 ImmOp = Src0;
1683 RegOp = Src1;
1684 } else if (Src1->isImm()) {
1685 ImmOp = Src1;
1686 RegOp = Src0;
1687 } else
1688 return std::pair(nullptr, SIOutMods::NONE);
1689
1690 int OMod = getOModValue(Op, ImmOp->getImm());
1691 if (OMod == SIOutMods::NONE ||
1692 TII->hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers) ||
1693 TII->hasModifiersSet(MI, AMDGPU::OpName::src1_modifiers) ||
1694 TII->hasModifiersSet(MI, AMDGPU::OpName::omod) ||
1695 TII->hasModifiersSet(MI, AMDGPU::OpName::clamp))
1696 return std::pair(nullptr, SIOutMods::NONE);
1697
1698 return std::pair(RegOp, OMod);
1699 }
1700 case AMDGPU::V_ADD_F64_e64:
1701 case AMDGPU::V_ADD_F64_pseudo_e64:
1702 case AMDGPU::V_ADD_F32_e64:
1703 case AMDGPU::V_ADD_F16_e64:
1704 case AMDGPU::V_ADD_F16_t16_e64:
1705 case AMDGPU::V_ADD_F16_fake16_e64: {
1706 // If output denormals are enabled, omod is ignored.
1707 if ((Op == AMDGPU::V_ADD_F32_e64 &&
1708 MFI->getMode().FP32Denormals.Output != DenormalMode::PreserveSign) ||
1709 ((Op == AMDGPU::V_ADD_F64_e64 || Op == AMDGPU::V_ADD_F64_pseudo_e64 ||
1710 Op == AMDGPU::V_ADD_F16_e64 || Op == AMDGPU::V_ADD_F16_t16_e64 ||
1711 Op == AMDGPU::V_ADD_F16_fake16_e64) &&
1712 MFI->getMode().FP64FP16Denormals.Output != DenormalMode::PreserveSign))
1713 return std::pair(nullptr, SIOutMods::NONE);
1714
1715 // Look through the DAGCombiner canonicalization fmul x, 2 -> fadd x, x
1716 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
1717 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
1718
1719 if (Src0->isReg() && Src1->isReg() && Src0->getReg() == Src1->getReg() &&
1720 Src0->getSubReg() == Src1->getSubReg() &&
1721 !TII->hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers) &&
1722 !TII->hasModifiersSet(MI, AMDGPU::OpName::src1_modifiers) &&
1723 !TII->hasModifiersSet(MI, AMDGPU::OpName::clamp) &&
1724 !TII->hasModifiersSet(MI, AMDGPU::OpName::omod))
1725 return std::pair(Src0, SIOutMods::MUL2);
1726
1727 return std::pair(nullptr, SIOutMods::NONE);
1728 }
1729 default:
1730 return std::pair(nullptr, SIOutMods::NONE);
1731 }
1732 }
1733
1734 // FIXME: Does this need to check IEEE bit on function?
tryFoldOMod(MachineInstr & MI)1735 bool SIFoldOperands::tryFoldOMod(MachineInstr &MI) {
1736 const MachineOperand *RegOp;
1737 int OMod;
1738 std::tie(RegOp, OMod) = isOMod(MI);
1739 if (OMod == SIOutMods::NONE || !RegOp->isReg() ||
1740 RegOp->getSubReg() != AMDGPU::NoSubRegister ||
1741 !MRI->hasOneNonDBGUser(RegOp->getReg()))
1742 return false;
1743
1744 MachineInstr *Def = MRI->getVRegDef(RegOp->getReg());
1745 MachineOperand *DefOMod = TII->getNamedOperand(*Def, AMDGPU::OpName::omod);
1746 if (!DefOMod || DefOMod->getImm() != SIOutMods::NONE)
1747 return false;
1748
1749 if (Def->mayRaiseFPException())
1750 return false;
1751
1752 // Clamp is applied after omod. If the source already has clamp set, don't
1753 // fold it.
1754 if (TII->hasModifiersSet(*Def, AMDGPU::OpName::clamp))
1755 return false;
1756
1757 LLVM_DEBUG(dbgs() << "Folding omod " << MI << " into " << *Def);
1758
1759 DefOMod->setImm(OMod);
1760 MRI->replaceRegWith(MI.getOperand(0).getReg(), Def->getOperand(0).getReg());
1761 MI.eraseFromParent();
1762
1763 // Use of output modifiers forces VOP3 encoding for a VOP2 mac/fmac
1764 // instruction, so we might as well convert it to the more flexible VOP3-only
1765 // mad/fma form.
1766 if (TII->convertToThreeAddress(*Def, nullptr, nullptr))
1767 Def->eraseFromParent();
1768
1769 return true;
1770 }
1771
1772 // Try to fold a reg_sequence with vgpr output and agpr inputs into an
1773 // instruction which can take an agpr. So far that means a store.
tryFoldRegSequence(MachineInstr & MI)1774 bool SIFoldOperands::tryFoldRegSequence(MachineInstr &MI) {
1775 assert(MI.isRegSequence());
1776 auto Reg = MI.getOperand(0).getReg();
1777
1778 if (!ST->hasGFX90AInsts() || !TRI->isVGPR(*MRI, Reg) ||
1779 !MRI->hasOneNonDBGUse(Reg))
1780 return false;
1781
1782 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs;
1783 if (!getRegSeqInit(Defs, Reg, MCOI::OPERAND_REGISTER))
1784 return false;
1785
1786 for (auto &[Op, SubIdx] : Defs) {
1787 if (!Op->isReg())
1788 return false;
1789 if (TRI->isAGPR(*MRI, Op->getReg()))
1790 continue;
1791 // Maybe this is a COPY from AREG
1792 const MachineInstr *SubDef = MRI->getVRegDef(Op->getReg());
1793 if (!SubDef || !SubDef->isCopy() || SubDef->getOperand(1).getSubReg())
1794 return false;
1795 if (!TRI->isAGPR(*MRI, SubDef->getOperand(1).getReg()))
1796 return false;
1797 }
1798
1799 MachineOperand *Op = &*MRI->use_nodbg_begin(Reg);
1800 MachineInstr *UseMI = Op->getParent();
1801 while (UseMI->isCopy() && !Op->getSubReg()) {
1802 Reg = UseMI->getOperand(0).getReg();
1803 if (!TRI->isVGPR(*MRI, Reg) || !MRI->hasOneNonDBGUse(Reg))
1804 return false;
1805 Op = &*MRI->use_nodbg_begin(Reg);
1806 UseMI = Op->getParent();
1807 }
1808
1809 if (Op->getSubReg())
1810 return false;
1811
1812 unsigned OpIdx = Op - &UseMI->getOperand(0);
1813 const MCInstrDesc &InstDesc = UseMI->getDesc();
1814 const TargetRegisterClass *OpRC =
1815 TII->getRegClass(InstDesc, OpIdx, TRI, *MI.getMF());
1816 if (!OpRC || !TRI->isVectorSuperClass(OpRC))
1817 return false;
1818
1819 const auto *NewDstRC = TRI->getEquivalentAGPRClass(MRI->getRegClass(Reg));
1820 auto Dst = MRI->createVirtualRegister(NewDstRC);
1821 auto RS = BuildMI(*MI.getParent(), MI, MI.getDebugLoc(),
1822 TII->get(AMDGPU::REG_SEQUENCE), Dst);
1823
1824 for (auto &[Def, SubIdx] : Defs) {
1825 Def->setIsKill(false);
1826 if (TRI->isAGPR(*MRI, Def->getReg())) {
1827 RS.add(*Def);
1828 } else { // This is a copy
1829 MachineInstr *SubDef = MRI->getVRegDef(Def->getReg());
1830 SubDef->getOperand(1).setIsKill(false);
1831 RS.addReg(SubDef->getOperand(1).getReg(), 0, Def->getSubReg());
1832 }
1833 RS.addImm(SubIdx);
1834 }
1835
1836 Op->setReg(Dst);
1837 if (!TII->isOperandLegal(*UseMI, OpIdx, Op)) {
1838 Op->setReg(Reg);
1839 RS->eraseFromParent();
1840 return false;
1841 }
1842
1843 LLVM_DEBUG(dbgs() << "Folded " << *RS << " into " << *UseMI);
1844
1845 // Erase the REG_SEQUENCE eagerly, unless we followed a chain of COPY users,
1846 // in which case we can erase them all later in runOnMachineFunction.
1847 if (MRI->use_nodbg_empty(MI.getOperand(0).getReg()))
1848 MI.eraseFromParent();
1849 return true;
1850 }
1851
1852 /// Checks whether \p Copy is a AGPR -> VGPR copy. Returns `true` on success and
1853 /// stores the AGPR register in \p OutReg and the subreg in \p OutSubReg
isAGPRCopy(const SIRegisterInfo & TRI,const MachineRegisterInfo & MRI,const MachineInstr & Copy,Register & OutReg,unsigned & OutSubReg)1854 static bool isAGPRCopy(const SIRegisterInfo &TRI,
1855 const MachineRegisterInfo &MRI, const MachineInstr &Copy,
1856 Register &OutReg, unsigned &OutSubReg) {
1857 assert(Copy.isCopy());
1858
1859 const MachineOperand &CopySrc = Copy.getOperand(1);
1860 Register CopySrcReg = CopySrc.getReg();
1861 if (!CopySrcReg.isVirtual())
1862 return false;
1863
1864 // Common case: copy from AGPR directly, e.g.
1865 // %1:vgpr_32 = COPY %0:agpr_32
1866 if (TRI.isAGPR(MRI, CopySrcReg)) {
1867 OutReg = CopySrcReg;
1868 OutSubReg = CopySrc.getSubReg();
1869 return true;
1870 }
1871
1872 // Sometimes it can also involve two copies, e.g.
1873 // %1:vgpr_256 = COPY %0:agpr_256
1874 // %2:vgpr_32 = COPY %1:vgpr_256.sub0
1875 const MachineInstr *CopySrcDef = MRI.getVRegDef(CopySrcReg);
1876 if (!CopySrcDef || !CopySrcDef->isCopy())
1877 return false;
1878
1879 const MachineOperand &OtherCopySrc = CopySrcDef->getOperand(1);
1880 Register OtherCopySrcReg = OtherCopySrc.getReg();
1881 if (!OtherCopySrcReg.isVirtual() ||
1882 CopySrcDef->getOperand(0).getSubReg() != AMDGPU::NoSubRegister ||
1883 OtherCopySrc.getSubReg() != AMDGPU::NoSubRegister ||
1884 !TRI.isAGPR(MRI, OtherCopySrcReg))
1885 return false;
1886
1887 OutReg = OtherCopySrcReg;
1888 OutSubReg = CopySrc.getSubReg();
1889 return true;
1890 }
1891
1892 // Try to hoist an AGPR to VGPR copy across a PHI.
1893 // This should allow folding of an AGPR into a consumer which may support it.
1894 //
1895 // Example 1: LCSSA PHI
1896 // loop:
1897 // %1:vreg = COPY %0:areg
1898 // exit:
1899 // %2:vreg = PHI %1:vreg, %loop
1900 // =>
1901 // loop:
1902 // exit:
1903 // %1:areg = PHI %0:areg, %loop
1904 // %2:vreg = COPY %1:areg
1905 //
1906 // Example 2: PHI with multiple incoming values:
1907 // entry:
1908 // %1:vreg = GLOBAL_LOAD(..)
1909 // loop:
1910 // %2:vreg = PHI %1:vreg, %entry, %5:vreg, %loop
1911 // %3:areg = COPY %2:vreg
1912 // %4:areg = (instr using %3:areg)
1913 // %5:vreg = COPY %4:areg
1914 // =>
1915 // entry:
1916 // %1:vreg = GLOBAL_LOAD(..)
1917 // %2:areg = COPY %1:vreg
1918 // loop:
1919 // %3:areg = PHI %2:areg, %entry, %X:areg,
1920 // %4:areg = (instr using %3:areg)
tryFoldPhiAGPR(MachineInstr & PHI)1921 bool SIFoldOperands::tryFoldPhiAGPR(MachineInstr &PHI) {
1922 assert(PHI.isPHI());
1923
1924 Register PhiOut = PHI.getOperand(0).getReg();
1925 if (!TRI->isVGPR(*MRI, PhiOut))
1926 return false;
1927
1928 // Iterate once over all incoming values of the PHI to check if this PHI is
1929 // eligible, and determine the exact AGPR RC we'll target.
1930 const TargetRegisterClass *ARC = nullptr;
1931 for (unsigned K = 1; K < PHI.getNumExplicitOperands(); K += 2) {
1932 MachineOperand &MO = PHI.getOperand(K);
1933 MachineInstr *Copy = MRI->getVRegDef(MO.getReg());
1934 if (!Copy || !Copy->isCopy())
1935 continue;
1936
1937 Register AGPRSrc;
1938 unsigned AGPRRegMask = AMDGPU::NoSubRegister;
1939 if (!isAGPRCopy(*TRI, *MRI, *Copy, AGPRSrc, AGPRRegMask))
1940 continue;
1941
1942 const TargetRegisterClass *CopyInRC = MRI->getRegClass(AGPRSrc);
1943 if (const auto *SubRC = TRI->getSubRegisterClass(CopyInRC, AGPRRegMask))
1944 CopyInRC = SubRC;
1945
1946 if (ARC && !ARC->hasSubClassEq(CopyInRC))
1947 return false;
1948 ARC = CopyInRC;
1949 }
1950
1951 if (!ARC)
1952 return false;
1953
1954 bool IsAGPR32 = (ARC == &AMDGPU::AGPR_32RegClass);
1955
1956 // Rewrite the PHI's incoming values to ARC.
1957 LLVM_DEBUG(dbgs() << "Folding AGPR copies into: " << PHI);
1958 for (unsigned K = 1; K < PHI.getNumExplicitOperands(); K += 2) {
1959 MachineOperand &MO = PHI.getOperand(K);
1960 Register Reg = MO.getReg();
1961
1962 MachineBasicBlock::iterator InsertPt;
1963 MachineBasicBlock *InsertMBB = nullptr;
1964
1965 // Look at the def of Reg, ignoring all copies.
1966 unsigned CopyOpc = AMDGPU::COPY;
1967 if (MachineInstr *Def = MRI->getVRegDef(Reg)) {
1968
1969 // Look at pre-existing COPY instructions from ARC: Steal the operand. If
1970 // the copy was single-use, it will be removed by DCE later.
1971 if (Def->isCopy()) {
1972 Register AGPRSrc;
1973 unsigned AGPRSubReg = AMDGPU::NoSubRegister;
1974 if (isAGPRCopy(*TRI, *MRI, *Def, AGPRSrc, AGPRSubReg)) {
1975 MO.setReg(AGPRSrc);
1976 MO.setSubReg(AGPRSubReg);
1977 continue;
1978 }
1979
1980 // If this is a multi-use SGPR -> VGPR copy, use V_ACCVGPR_WRITE on
1981 // GFX908 directly instead of a COPY. Otherwise, SIFoldOperand may try
1982 // to fold the sgpr -> vgpr -> agpr copy into a sgpr -> agpr copy which
1983 // is unlikely to be profitable.
1984 //
1985 // Note that V_ACCVGPR_WRITE is only used for AGPR_32.
1986 MachineOperand &CopyIn = Def->getOperand(1);
1987 if (IsAGPR32 && !ST->hasGFX90AInsts() && !MRI->hasOneNonDBGUse(Reg) &&
1988 TRI->isSGPRReg(*MRI, CopyIn.getReg()))
1989 CopyOpc = AMDGPU::V_ACCVGPR_WRITE_B32_e64;
1990 }
1991
1992 InsertMBB = Def->getParent();
1993 InsertPt = InsertMBB->SkipPHIsLabelsAndDebug(++Def->getIterator());
1994 } else {
1995 InsertMBB = PHI.getOperand(MO.getOperandNo() + 1).getMBB();
1996 InsertPt = InsertMBB->getFirstTerminator();
1997 }
1998
1999 Register NewReg = MRI->createVirtualRegister(ARC);
2000 MachineInstr *MI = BuildMI(*InsertMBB, InsertPt, PHI.getDebugLoc(),
2001 TII->get(CopyOpc), NewReg)
2002 .addReg(Reg);
2003 MO.setReg(NewReg);
2004
2005 (void)MI;
2006 LLVM_DEBUG(dbgs() << " Created COPY: " << *MI);
2007 }
2008
2009 // Replace the PHI's result with a new register.
2010 Register NewReg = MRI->createVirtualRegister(ARC);
2011 PHI.getOperand(0).setReg(NewReg);
2012
2013 // COPY that new register back to the original PhiOut register. This COPY will
2014 // usually be folded out later.
2015 MachineBasicBlock *MBB = PHI.getParent();
2016 BuildMI(*MBB, MBB->getFirstNonPHI(), PHI.getDebugLoc(),
2017 TII->get(AMDGPU::COPY), PhiOut)
2018 .addReg(NewReg);
2019
2020 LLVM_DEBUG(dbgs() << " Done: Folded " << PHI);
2021 return true;
2022 }
2023
2024 // Attempt to convert VGPR load to an AGPR load.
tryFoldLoad(MachineInstr & MI)2025 bool SIFoldOperands::tryFoldLoad(MachineInstr &MI) {
2026 assert(MI.mayLoad());
2027 if (!ST->hasGFX90AInsts() || MI.getNumExplicitDefs() != 1)
2028 return false;
2029
2030 MachineOperand &Def = MI.getOperand(0);
2031 if (!Def.isDef())
2032 return false;
2033
2034 Register DefReg = Def.getReg();
2035
2036 if (DefReg.isPhysical() || !TRI->isVGPR(*MRI, DefReg))
2037 return false;
2038
2039 SmallVector<const MachineInstr*, 8> Users;
2040 SmallVector<Register, 8> MoveRegs;
2041 for (const MachineInstr &I : MRI->use_nodbg_instructions(DefReg))
2042 Users.push_back(&I);
2043
2044 if (Users.empty())
2045 return false;
2046
2047 // Check that all uses a copy to an agpr or a reg_sequence producing an agpr.
2048 while (!Users.empty()) {
2049 const MachineInstr *I = Users.pop_back_val();
2050 if (!I->isCopy() && !I->isRegSequence())
2051 return false;
2052 Register DstReg = I->getOperand(0).getReg();
2053 // Physical registers may have more than one instruction definitions
2054 if (DstReg.isPhysical())
2055 return false;
2056 if (TRI->isAGPR(*MRI, DstReg))
2057 continue;
2058 MoveRegs.push_back(DstReg);
2059 for (const MachineInstr &U : MRI->use_nodbg_instructions(DstReg))
2060 Users.push_back(&U);
2061 }
2062
2063 const TargetRegisterClass *RC = MRI->getRegClass(DefReg);
2064 MRI->setRegClass(DefReg, TRI->getEquivalentAGPRClass(RC));
2065 if (!TII->isOperandLegal(MI, 0, &Def)) {
2066 MRI->setRegClass(DefReg, RC);
2067 return false;
2068 }
2069
2070 while (!MoveRegs.empty()) {
2071 Register Reg = MoveRegs.pop_back_val();
2072 MRI->setRegClass(Reg, TRI->getEquivalentAGPRClass(MRI->getRegClass(Reg)));
2073 }
2074
2075 LLVM_DEBUG(dbgs() << "Folded " << MI);
2076
2077 return true;
2078 }
2079
2080 // tryFoldPhiAGPR will aggressively try to create AGPR PHIs.
2081 // For GFX90A and later, this is pretty much always a good thing, but for GFX908
2082 // there's cases where it can create a lot more AGPR-AGPR copies, which are
2083 // expensive on this architecture due to the lack of V_ACCVGPR_MOV.
2084 //
2085 // This function looks at all AGPR PHIs in a basic block and collects their
2086 // operands. Then, it checks for register that are used more than once across
2087 // all PHIs and caches them in a VGPR. This prevents ExpandPostRAPseudo from
2088 // having to create one VGPR temporary per use, which can get very messy if
2089 // these PHIs come from a broken-up large PHI (e.g. 32 AGPR phis, one per vector
2090 // element).
2091 //
2092 // Example
2093 // a:
2094 // %in:agpr_256 = COPY %foo:vgpr_256
2095 // c:
2096 // %x:agpr_32 = ..
2097 // b:
2098 // %0:areg = PHI %in.sub0:agpr_32, %a, %x, %c
2099 // %1:areg = PHI %in.sub0:agpr_32, %a, %y, %c
2100 // %2:areg = PHI %in.sub0:agpr_32, %a, %z, %c
2101 // =>
2102 // a:
2103 // %in:agpr_256 = COPY %foo:vgpr_256
2104 // %tmp:vgpr_32 = V_ACCVGPR_READ_B32_e64 %in.sub0:agpr_32
2105 // %tmp_agpr:agpr_32 = COPY %tmp
2106 // c:
2107 // %x:agpr_32 = ..
2108 // b:
2109 // %0:areg = PHI %tmp_agpr, %a, %x, %c
2110 // %1:areg = PHI %tmp_agpr, %a, %y, %c
2111 // %2:areg = PHI %tmp_agpr, %a, %z, %c
tryOptimizeAGPRPhis(MachineBasicBlock & MBB)2112 bool SIFoldOperands::tryOptimizeAGPRPhis(MachineBasicBlock &MBB) {
2113 // This is only really needed on GFX908 where AGPR-AGPR copies are
2114 // unreasonably difficult.
2115 if (ST->hasGFX90AInsts())
2116 return false;
2117
2118 // Look at all AGPR Phis and collect the register + subregister used.
2119 DenseMap<std::pair<Register, unsigned>, std::vector<MachineOperand *>>
2120 RegToMO;
2121
2122 for (auto &MI : MBB) {
2123 if (!MI.isPHI())
2124 break;
2125
2126 if (!TRI->isAGPR(*MRI, MI.getOperand(0).getReg()))
2127 continue;
2128
2129 for (unsigned K = 1; K < MI.getNumOperands(); K += 2) {
2130 MachineOperand &PhiMO = MI.getOperand(K);
2131 if (!PhiMO.getSubReg())
2132 continue;
2133 RegToMO[{PhiMO.getReg(), PhiMO.getSubReg()}].push_back(&PhiMO);
2134 }
2135 }
2136
2137 // For all (Reg, SubReg) pair that are used more than once, cache the value in
2138 // a VGPR.
2139 bool Changed = false;
2140 for (const auto &[Entry, MOs] : RegToMO) {
2141 if (MOs.size() == 1)
2142 continue;
2143
2144 const auto [Reg, SubReg] = Entry;
2145 MachineInstr *Def = MRI->getVRegDef(Reg);
2146 MachineBasicBlock *DefMBB = Def->getParent();
2147
2148 // Create a copy in a VGPR using V_ACCVGPR_READ_B32_e64 so it's not folded
2149 // out.
2150 const TargetRegisterClass *ARC = getRegOpRC(*MRI, *TRI, *MOs.front());
2151 Register TempVGPR =
2152 MRI->createVirtualRegister(TRI->getEquivalentVGPRClass(ARC));
2153 MachineInstr *VGPRCopy =
2154 BuildMI(*DefMBB, ++Def->getIterator(), Def->getDebugLoc(),
2155 TII->get(AMDGPU::V_ACCVGPR_READ_B32_e64), TempVGPR)
2156 .addReg(Reg, /* flags */ 0, SubReg);
2157
2158 // Copy back to an AGPR and use that instead of the AGPR subreg in all MOs.
2159 Register TempAGPR = MRI->createVirtualRegister(ARC);
2160 BuildMI(*DefMBB, ++VGPRCopy->getIterator(), Def->getDebugLoc(),
2161 TII->get(AMDGPU::COPY), TempAGPR)
2162 .addReg(TempVGPR);
2163
2164 LLVM_DEBUG(dbgs() << "Caching AGPR into VGPR: " << *VGPRCopy);
2165 for (MachineOperand *MO : MOs) {
2166 MO->setReg(TempAGPR);
2167 MO->setSubReg(AMDGPU::NoSubRegister);
2168 LLVM_DEBUG(dbgs() << " Changed PHI Operand: " << *MO << "\n");
2169 }
2170
2171 Changed = true;
2172 }
2173
2174 return Changed;
2175 }
2176
runOnMachineFunction(MachineFunction & MF)2177 bool SIFoldOperands::runOnMachineFunction(MachineFunction &MF) {
2178 if (skipFunction(MF.getFunction()))
2179 return false;
2180
2181 MRI = &MF.getRegInfo();
2182 ST = &MF.getSubtarget<GCNSubtarget>();
2183 TII = ST->getInstrInfo();
2184 TRI = &TII->getRegisterInfo();
2185 MFI = MF.getInfo<SIMachineFunctionInfo>();
2186
2187 // omod is ignored by hardware if IEEE bit is enabled. omod also does not
2188 // correctly handle signed zeros.
2189 //
2190 // FIXME: Also need to check strictfp
2191 bool IsIEEEMode = MFI->getMode().IEEE;
2192 bool HasNSZ = MFI->hasNoSignedZerosFPMath();
2193
2194 bool Changed = false;
2195 for (MachineBasicBlock *MBB : depth_first(&MF)) {
2196 MachineOperand *CurrentKnownM0Val = nullptr;
2197 for (auto &MI : make_early_inc_range(*MBB)) {
2198 Changed |= tryFoldCndMask(MI);
2199
2200 if (tryFoldZeroHighBits(MI)) {
2201 Changed = true;
2202 continue;
2203 }
2204
2205 if (MI.isRegSequence() && tryFoldRegSequence(MI)) {
2206 Changed = true;
2207 continue;
2208 }
2209
2210 if (MI.isPHI() && tryFoldPhiAGPR(MI)) {
2211 Changed = true;
2212 continue;
2213 }
2214
2215 if (MI.mayLoad() && tryFoldLoad(MI)) {
2216 Changed = true;
2217 continue;
2218 }
2219
2220 if (TII->isFoldableCopy(MI)) {
2221 Changed |= tryFoldFoldableCopy(MI, CurrentKnownM0Val);
2222 continue;
2223 }
2224
2225 // Saw an unknown clobber of m0, so we no longer know what it is.
2226 if (CurrentKnownM0Val && MI.modifiesRegister(AMDGPU::M0, TRI))
2227 CurrentKnownM0Val = nullptr;
2228
2229 // TODO: Omod might be OK if there is NSZ only on the source
2230 // instruction, and not the omod multiply.
2231 if (IsIEEEMode || (!HasNSZ && !MI.getFlag(MachineInstr::FmNsz)) ||
2232 !tryFoldOMod(MI))
2233 Changed |= tryFoldClamp(MI);
2234 }
2235
2236 Changed |= tryOptimizeAGPRPhis(*MBB);
2237 }
2238
2239 return Changed;
2240 }
2241