1 //===-- SIShrinkInstructions.cpp - Shrink Instructions --------------------===//
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 /// The pass tries to use the 32-bit encoding for instructions when possible.
8 //===----------------------------------------------------------------------===//
9 //
10
11 #include "AMDGPU.h"
12 #include "GCNSubtarget.h"
13 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
14 #include "Utils/AMDGPUBaseInfo.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/CodeGen/MachineFunctionPass.h"
17
18 #define DEBUG_TYPE "si-shrink-instructions"
19
20 STATISTIC(NumInstructionsShrunk,
21 "Number of 64-bit instruction reduced to 32-bit.");
22 STATISTIC(NumLiteralConstantsFolded,
23 "Number of literal constants folded into 32-bit instructions.");
24
25 using namespace llvm;
26
27 namespace {
28
29 class SIShrinkInstructions : public MachineFunctionPass {
30 MachineFunction *MF;
31 MachineRegisterInfo *MRI;
32 const GCNSubtarget *ST;
33 const SIInstrInfo *TII;
34 const SIRegisterInfo *TRI;
35
36 public:
37 static char ID;
38
39 public:
SIShrinkInstructions()40 SIShrinkInstructions() : MachineFunctionPass(ID) {
41 }
42
43 bool foldImmediates(MachineInstr &MI, bool TryToCommute = true) const;
44 bool shouldShrinkTrue16(MachineInstr &MI) const;
45 bool isKImmOperand(const MachineOperand &Src) const;
46 bool isKUImmOperand(const MachineOperand &Src) const;
47 bool isKImmOrKUImmOperand(const MachineOperand &Src, bool &IsUnsigned) const;
48 void copyExtraImplicitOps(MachineInstr &NewMI, MachineInstr &MI) const;
49 void shrinkScalarCompare(MachineInstr &MI) const;
50 void shrinkMIMG(MachineInstr &MI) const;
51 void shrinkMadFma(MachineInstr &MI) const;
52 bool shrinkScalarLogicOp(MachineInstr &MI) const;
53 bool tryReplaceDeadSDST(MachineInstr &MI) const;
54 bool instAccessReg(iterator_range<MachineInstr::const_mop_iterator> &&R,
55 Register Reg, unsigned SubReg) const;
56 bool instReadsReg(const MachineInstr *MI, unsigned Reg,
57 unsigned SubReg) const;
58 bool instModifiesReg(const MachineInstr *MI, unsigned Reg,
59 unsigned SubReg) const;
60 TargetInstrInfo::RegSubRegPair getSubRegForIndex(Register Reg, unsigned Sub,
61 unsigned I) const;
62 void dropInstructionKeepingImpDefs(MachineInstr &MI) const;
63 MachineInstr *matchSwap(MachineInstr &MovT) const;
64
65 bool runOnMachineFunction(MachineFunction &MF) override;
66
getPassName() const67 StringRef getPassName() const override { return "SI Shrink Instructions"; }
68
getAnalysisUsage(AnalysisUsage & AU) const69 void getAnalysisUsage(AnalysisUsage &AU) const override {
70 AU.setPreservesCFG();
71 MachineFunctionPass::getAnalysisUsage(AU);
72 }
73 };
74
75 } // End anonymous namespace.
76
77 INITIALIZE_PASS(SIShrinkInstructions, DEBUG_TYPE,
78 "SI Shrink Instructions", false, false)
79
80 char SIShrinkInstructions::ID = 0;
81
createSIShrinkInstructionsPass()82 FunctionPass *llvm::createSIShrinkInstructionsPass() {
83 return new SIShrinkInstructions();
84 }
85
86 /// This function checks \p MI for operands defined by a move immediate
87 /// instruction and then folds the literal constant into the instruction if it
88 /// can. This function assumes that \p MI is a VOP1, VOP2, or VOPC instructions.
foldImmediates(MachineInstr & MI,bool TryToCommute) const89 bool SIShrinkInstructions::foldImmediates(MachineInstr &MI,
90 bool TryToCommute) const {
91 assert(TII->isVOP1(MI) || TII->isVOP2(MI) || TII->isVOPC(MI));
92
93 int Src0Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::src0);
94
95 // Try to fold Src0
96 MachineOperand &Src0 = MI.getOperand(Src0Idx);
97 if (Src0.isReg()) {
98 Register Reg = Src0.getReg();
99 if (Reg.isVirtual()) {
100 MachineInstr *Def = MRI->getUniqueVRegDef(Reg);
101 if (Def && Def->isMoveImmediate()) {
102 MachineOperand &MovSrc = Def->getOperand(1);
103 bool ConstantFolded = false;
104
105 if (TII->isOperandLegal(MI, Src0Idx, &MovSrc)) {
106 if (MovSrc.isImm()) {
107 Src0.ChangeToImmediate(MovSrc.getImm());
108 ConstantFolded = true;
109 } else if (MovSrc.isFI()) {
110 Src0.ChangeToFrameIndex(MovSrc.getIndex());
111 ConstantFolded = true;
112 } else if (MovSrc.isGlobal()) {
113 Src0.ChangeToGA(MovSrc.getGlobal(), MovSrc.getOffset(),
114 MovSrc.getTargetFlags());
115 ConstantFolded = true;
116 }
117 }
118
119 if (ConstantFolded) {
120 if (MRI->use_nodbg_empty(Reg))
121 Def->eraseFromParent();
122 ++NumLiteralConstantsFolded;
123 return true;
124 }
125 }
126 }
127 }
128
129 // We have failed to fold src0, so commute the instruction and try again.
130 if (TryToCommute && MI.isCommutable()) {
131 if (TII->commuteInstruction(MI)) {
132 if (foldImmediates(MI, false))
133 return true;
134
135 // Commute back.
136 TII->commuteInstruction(MI);
137 }
138 }
139
140 return false;
141 }
142
143 /// Do not shrink the instruction if its registers are not expressible in the
144 /// shrunk encoding.
shouldShrinkTrue16(MachineInstr & MI) const145 bool SIShrinkInstructions::shouldShrinkTrue16(MachineInstr &MI) const {
146 for (unsigned I = 0, E = MI.getNumExplicitOperands(); I != E; ++I) {
147 const MachineOperand &MO = MI.getOperand(I);
148 if (MO.isReg()) {
149 Register Reg = MO.getReg();
150 assert(!Reg.isVirtual() && "Prior checks should ensure we only shrink "
151 "True16 Instructions post-RA");
152 if (AMDGPU::VGPR_32RegClass.contains(Reg) &&
153 !AMDGPU::VGPR_32_Lo128RegClass.contains(Reg))
154 return false;
155 }
156 }
157 return true;
158 }
159
isKImmOperand(const MachineOperand & Src) const160 bool SIShrinkInstructions::isKImmOperand(const MachineOperand &Src) const {
161 return isInt<16>(SignExtend64(Src.getImm(), 32)) &&
162 !TII->isInlineConstant(*Src.getParent(), Src.getOperandNo());
163 }
164
isKUImmOperand(const MachineOperand & Src) const165 bool SIShrinkInstructions::isKUImmOperand(const MachineOperand &Src) const {
166 return isUInt<16>(Src.getImm()) &&
167 !TII->isInlineConstant(*Src.getParent(), Src.getOperandNo());
168 }
169
isKImmOrKUImmOperand(const MachineOperand & Src,bool & IsUnsigned) const170 bool SIShrinkInstructions::isKImmOrKUImmOperand(const MachineOperand &Src,
171 bool &IsUnsigned) const {
172 if (isInt<16>(SignExtend64(Src.getImm(), 32))) {
173 IsUnsigned = false;
174 return !TII->isInlineConstant(Src);
175 }
176
177 if (isUInt<16>(Src.getImm())) {
178 IsUnsigned = true;
179 return !TII->isInlineConstant(Src);
180 }
181
182 return false;
183 }
184
185 /// \returns the opcode of an instruction a move immediate of the constant \p
186 /// Src can be replaced with if the constant is replaced with \p ModifiedImm.
187 /// i.e.
188 ///
189 /// If the bitreverse of a constant is an inline immediate, reverse the
190 /// immediate and return the bitreverse opcode.
191 ///
192 /// If the bitwise negation of a constant is an inline immediate, reverse the
193 /// immediate and return the bitwise not opcode.
canModifyToInlineImmOp32(const SIInstrInfo * TII,const MachineOperand & Src,int32_t & ModifiedImm,bool Scalar)194 static unsigned canModifyToInlineImmOp32(const SIInstrInfo *TII,
195 const MachineOperand &Src,
196 int32_t &ModifiedImm, bool Scalar) {
197 if (TII->isInlineConstant(Src))
198 return 0;
199 int32_t SrcImm = static_cast<int32_t>(Src.getImm());
200
201 if (!Scalar) {
202 // We could handle the scalar case with here, but we would need to check
203 // that SCC is not live as S_NOT_B32 clobbers it. It's probably not worth
204 // it, as the reasonable values are already covered by s_movk_i32.
205 ModifiedImm = ~SrcImm;
206 if (TII->isInlineConstant(APInt(32, ModifiedImm)))
207 return AMDGPU::V_NOT_B32_e32;
208 }
209
210 ModifiedImm = reverseBits<int32_t>(SrcImm);
211 if (TII->isInlineConstant(APInt(32, ModifiedImm)))
212 return Scalar ? AMDGPU::S_BREV_B32 : AMDGPU::V_BFREV_B32_e32;
213
214 return 0;
215 }
216
217 /// Copy implicit register operands from specified instruction to this
218 /// instruction that are not part of the instruction definition.
copyExtraImplicitOps(MachineInstr & NewMI,MachineInstr & MI) const219 void SIShrinkInstructions::copyExtraImplicitOps(MachineInstr &NewMI,
220 MachineInstr &MI) const {
221 MachineFunction &MF = *MI.getMF();
222 for (unsigned i = MI.getDesc().getNumOperands() +
223 MI.getDesc().implicit_uses().size() +
224 MI.getDesc().implicit_defs().size(),
225 e = MI.getNumOperands();
226 i != e; ++i) {
227 const MachineOperand &MO = MI.getOperand(i);
228 if ((MO.isReg() && MO.isImplicit()) || MO.isRegMask())
229 NewMI.addOperand(MF, MO);
230 }
231 }
232
shrinkScalarCompare(MachineInstr & MI) const233 void SIShrinkInstructions::shrinkScalarCompare(MachineInstr &MI) const {
234 if (!ST->hasSCmpK())
235 return;
236
237 // cmpk instructions do scc = dst <cc op> imm16, so commute the instruction to
238 // get constants on the RHS.
239 if (!MI.getOperand(0).isReg())
240 TII->commuteInstruction(MI, false, 0, 1);
241
242 // cmpk requires src0 to be a register
243 const MachineOperand &Src0 = MI.getOperand(0);
244 if (!Src0.isReg())
245 return;
246
247 MachineOperand &Src1 = MI.getOperand(1);
248 if (!Src1.isImm())
249 return;
250
251 int SOPKOpc = AMDGPU::getSOPKOp(MI.getOpcode());
252 if (SOPKOpc == -1)
253 return;
254
255 // eq/ne is special because the imm16 can be treated as signed or unsigned,
256 // and initially selected to the unsigned versions.
257 if (SOPKOpc == AMDGPU::S_CMPK_EQ_U32 || SOPKOpc == AMDGPU::S_CMPK_LG_U32) {
258 bool HasUImm;
259 if (isKImmOrKUImmOperand(Src1, HasUImm)) {
260 if (!HasUImm) {
261 SOPKOpc = (SOPKOpc == AMDGPU::S_CMPK_EQ_U32) ?
262 AMDGPU::S_CMPK_EQ_I32 : AMDGPU::S_CMPK_LG_I32;
263 Src1.setImm(SignExtend32(Src1.getImm(), 32));
264 }
265
266 MI.setDesc(TII->get(SOPKOpc));
267 }
268
269 return;
270 }
271
272 const MCInstrDesc &NewDesc = TII->get(SOPKOpc);
273
274 if ((SIInstrInfo::sopkIsZext(SOPKOpc) && isKUImmOperand(Src1)) ||
275 (!SIInstrInfo::sopkIsZext(SOPKOpc) && isKImmOperand(Src1))) {
276 if (!SIInstrInfo::sopkIsZext(SOPKOpc))
277 Src1.setImm(SignExtend64(Src1.getImm(), 32));
278 MI.setDesc(NewDesc);
279 }
280 }
281
282 // Shrink NSA encoded instructions with contiguous VGPRs to non-NSA encoding.
shrinkMIMG(MachineInstr & MI) const283 void SIShrinkInstructions::shrinkMIMG(MachineInstr &MI) const {
284 const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(MI.getOpcode());
285 if (!Info)
286 return;
287
288 uint8_t NewEncoding;
289 switch (Info->MIMGEncoding) {
290 case AMDGPU::MIMGEncGfx10NSA:
291 NewEncoding = AMDGPU::MIMGEncGfx10Default;
292 break;
293 case AMDGPU::MIMGEncGfx11NSA:
294 NewEncoding = AMDGPU::MIMGEncGfx11Default;
295 break;
296 default:
297 return;
298 }
299
300 int VAddr0Idx =
301 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vaddr0);
302 unsigned NewAddrDwords = Info->VAddrDwords;
303 const TargetRegisterClass *RC;
304
305 if (Info->VAddrDwords == 2) {
306 RC = &AMDGPU::VReg_64RegClass;
307 } else if (Info->VAddrDwords == 3) {
308 RC = &AMDGPU::VReg_96RegClass;
309 } else if (Info->VAddrDwords == 4) {
310 RC = &AMDGPU::VReg_128RegClass;
311 } else if (Info->VAddrDwords == 5) {
312 RC = &AMDGPU::VReg_160RegClass;
313 } else if (Info->VAddrDwords == 6) {
314 RC = &AMDGPU::VReg_192RegClass;
315 } else if (Info->VAddrDwords == 7) {
316 RC = &AMDGPU::VReg_224RegClass;
317 } else if (Info->VAddrDwords == 8) {
318 RC = &AMDGPU::VReg_256RegClass;
319 } else if (Info->VAddrDwords == 9) {
320 RC = &AMDGPU::VReg_288RegClass;
321 } else if (Info->VAddrDwords == 10) {
322 RC = &AMDGPU::VReg_320RegClass;
323 } else if (Info->VAddrDwords == 11) {
324 RC = &AMDGPU::VReg_352RegClass;
325 } else if (Info->VAddrDwords == 12) {
326 RC = &AMDGPU::VReg_384RegClass;
327 } else {
328 RC = &AMDGPU::VReg_512RegClass;
329 NewAddrDwords = 16;
330 }
331
332 unsigned VgprBase = 0;
333 unsigned NextVgpr = 0;
334 bool IsUndef = true;
335 bool IsKill = NewAddrDwords == Info->VAddrDwords;
336 const unsigned NSAMaxSize = ST->getNSAMaxSize();
337 const bool IsPartialNSA = NewAddrDwords > NSAMaxSize;
338 const unsigned EndVAddr = IsPartialNSA ? NSAMaxSize : Info->VAddrOperands;
339 for (unsigned Idx = 0; Idx < EndVAddr; ++Idx) {
340 const MachineOperand &Op = MI.getOperand(VAddr0Idx + Idx);
341 unsigned Vgpr = TRI->getHWRegIndex(Op.getReg());
342 unsigned Dwords = TRI->getRegSizeInBits(Op.getReg(), *MRI) / 32;
343 assert(Dwords > 0 && "Un-implemented for less than 32 bit regs");
344
345 if (Idx == 0) {
346 VgprBase = Vgpr;
347 NextVgpr = Vgpr + Dwords;
348 } else if (Vgpr == NextVgpr) {
349 NextVgpr = Vgpr + Dwords;
350 } else {
351 return;
352 }
353
354 if (!Op.isUndef())
355 IsUndef = false;
356 if (!Op.isKill())
357 IsKill = false;
358 }
359
360 if (VgprBase + NewAddrDwords > 256)
361 return;
362
363 // Further check for implicit tied operands - this may be present if TFE is
364 // enabled
365 int TFEIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::tfe);
366 int LWEIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::lwe);
367 unsigned TFEVal = (TFEIdx == -1) ? 0 : MI.getOperand(TFEIdx).getImm();
368 unsigned LWEVal = (LWEIdx == -1) ? 0 : MI.getOperand(LWEIdx).getImm();
369 int ToUntie = -1;
370 if (TFEVal || LWEVal) {
371 // TFE/LWE is enabled so we need to deal with an implicit tied operand
372 for (unsigned i = LWEIdx + 1, e = MI.getNumOperands(); i != e; ++i) {
373 if (MI.getOperand(i).isReg() && MI.getOperand(i).isTied() &&
374 MI.getOperand(i).isImplicit()) {
375 // This is the tied operand
376 assert(
377 ToUntie == -1 &&
378 "found more than one tied implicit operand when expecting only 1");
379 ToUntie = i;
380 MI.untieRegOperand(ToUntie);
381 }
382 }
383 }
384
385 unsigned NewOpcode = AMDGPU::getMIMGOpcode(Info->BaseOpcode, NewEncoding,
386 Info->VDataDwords, NewAddrDwords);
387 MI.setDesc(TII->get(NewOpcode));
388 MI.getOperand(VAddr0Idx).setReg(RC->getRegister(VgprBase));
389 MI.getOperand(VAddr0Idx).setIsUndef(IsUndef);
390 MI.getOperand(VAddr0Idx).setIsKill(IsKill);
391
392 for (unsigned i = 1; i < EndVAddr; ++i)
393 MI.removeOperand(VAddr0Idx + 1);
394
395 if (ToUntie >= 0) {
396 MI.tieOperands(
397 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vdata),
398 ToUntie - (EndVAddr - 1));
399 }
400 }
401
402 // Shrink MAD to MADAK/MADMK and FMA to FMAAK/FMAMK.
shrinkMadFma(MachineInstr & MI) const403 void SIShrinkInstructions::shrinkMadFma(MachineInstr &MI) const {
404 // Pre-GFX10 VOP3 instructions like MAD/FMA cannot take a literal operand so
405 // there is no reason to try to shrink them.
406 if (!ST->hasVOP3Literal())
407 return;
408
409 // There is no advantage to doing this pre-RA.
410 if (!MF->getProperties().hasProperty(
411 MachineFunctionProperties::Property::NoVRegs))
412 return;
413
414 if (TII->hasAnyModifiersSet(MI))
415 return;
416
417 const unsigned Opcode = MI.getOpcode();
418 MachineOperand &Src0 = *TII->getNamedOperand(MI, AMDGPU::OpName::src0);
419 MachineOperand &Src1 = *TII->getNamedOperand(MI, AMDGPU::OpName::src1);
420 MachineOperand &Src2 = *TII->getNamedOperand(MI, AMDGPU::OpName::src2);
421 unsigned NewOpcode = AMDGPU::INSTRUCTION_LIST_END;
422
423 bool Swap;
424
425 // Detect "Dst = VSrc * VGPR + Imm" and convert to AK form.
426 if (Src2.isImm() && !TII->isInlineConstant(Src2)) {
427 if (Src1.isReg() && TRI->isVGPR(*MRI, Src1.getReg()))
428 Swap = false;
429 else if (Src0.isReg() && TRI->isVGPR(*MRI, Src0.getReg()))
430 Swap = true;
431 else
432 return;
433
434 switch (Opcode) {
435 default:
436 llvm_unreachable("Unexpected mad/fma opcode!");
437 case AMDGPU::V_MAD_F32_e64:
438 NewOpcode = AMDGPU::V_MADAK_F32;
439 break;
440 case AMDGPU::V_FMA_F32_e64:
441 NewOpcode = AMDGPU::V_FMAAK_F32;
442 break;
443 case AMDGPU::V_MAD_F16_e64:
444 NewOpcode = AMDGPU::V_MADAK_F16;
445 break;
446 case AMDGPU::V_FMA_F16_e64:
447 case AMDGPU::V_FMA_F16_gfx9_e64:
448 NewOpcode = ST->hasTrue16BitInsts() ? AMDGPU::V_FMAAK_F16_t16
449 : AMDGPU::V_FMAAK_F16;
450 break;
451 }
452 }
453
454 // Detect "Dst = VSrc * Imm + VGPR" and convert to MK form.
455 if (Src2.isReg() && TRI->isVGPR(*MRI, Src2.getReg())) {
456 if (Src1.isImm() && !TII->isInlineConstant(Src1))
457 Swap = false;
458 else if (Src0.isImm() && !TII->isInlineConstant(Src0))
459 Swap = true;
460 else
461 return;
462
463 switch (Opcode) {
464 default:
465 llvm_unreachable("Unexpected mad/fma opcode!");
466 case AMDGPU::V_MAD_F32_e64:
467 NewOpcode = AMDGPU::V_MADMK_F32;
468 break;
469 case AMDGPU::V_FMA_F32_e64:
470 NewOpcode = AMDGPU::V_FMAMK_F32;
471 break;
472 case AMDGPU::V_MAD_F16_e64:
473 NewOpcode = AMDGPU::V_MADMK_F16;
474 break;
475 case AMDGPU::V_FMA_F16_e64:
476 case AMDGPU::V_FMA_F16_gfx9_e64:
477 NewOpcode = ST->hasTrue16BitInsts() ? AMDGPU::V_FMAMK_F16_t16
478 : AMDGPU::V_FMAMK_F16;
479 break;
480 }
481 }
482
483 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END)
484 return;
485
486 if (AMDGPU::isTrue16Inst(NewOpcode) && !shouldShrinkTrue16(MI))
487 return;
488
489 if (Swap) {
490 // Swap Src0 and Src1 by building a new instruction.
491 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), TII->get(NewOpcode),
492 MI.getOperand(0).getReg())
493 .add(Src1)
494 .add(Src0)
495 .add(Src2)
496 .setMIFlags(MI.getFlags());
497 MI.eraseFromParent();
498 } else {
499 TII->removeModOperands(MI);
500 MI.setDesc(TII->get(NewOpcode));
501 }
502 }
503
504 /// Attempt to shrink AND/OR/XOR operations requiring non-inlineable literals.
505 /// For AND or OR, try using S_BITSET{0,1} to clear or set bits.
506 /// If the inverse of the immediate is legal, use ANDN2, ORN2 or
507 /// XNOR (as a ^ b == ~(a ^ ~b)).
508 /// \returns true if the caller should continue the machine function iterator
shrinkScalarLogicOp(MachineInstr & MI) const509 bool SIShrinkInstructions::shrinkScalarLogicOp(MachineInstr &MI) const {
510 unsigned Opc = MI.getOpcode();
511 const MachineOperand *Dest = &MI.getOperand(0);
512 MachineOperand *Src0 = &MI.getOperand(1);
513 MachineOperand *Src1 = &MI.getOperand(2);
514 MachineOperand *SrcReg = Src0;
515 MachineOperand *SrcImm = Src1;
516
517 if (!SrcImm->isImm() ||
518 AMDGPU::isInlinableLiteral32(SrcImm->getImm(), ST->hasInv2PiInlineImm()))
519 return false;
520
521 uint32_t Imm = static_cast<uint32_t>(SrcImm->getImm());
522 uint32_t NewImm = 0;
523
524 if (Opc == AMDGPU::S_AND_B32) {
525 if (isPowerOf2_32(~Imm)) {
526 NewImm = llvm::countr_one(Imm);
527 Opc = AMDGPU::S_BITSET0_B32;
528 } else if (AMDGPU::isInlinableLiteral32(~Imm, ST->hasInv2PiInlineImm())) {
529 NewImm = ~Imm;
530 Opc = AMDGPU::S_ANDN2_B32;
531 }
532 } else if (Opc == AMDGPU::S_OR_B32) {
533 if (isPowerOf2_32(Imm)) {
534 NewImm = llvm::countr_zero(Imm);
535 Opc = AMDGPU::S_BITSET1_B32;
536 } else if (AMDGPU::isInlinableLiteral32(~Imm, ST->hasInv2PiInlineImm())) {
537 NewImm = ~Imm;
538 Opc = AMDGPU::S_ORN2_B32;
539 }
540 } else if (Opc == AMDGPU::S_XOR_B32) {
541 if (AMDGPU::isInlinableLiteral32(~Imm, ST->hasInv2PiInlineImm())) {
542 NewImm = ~Imm;
543 Opc = AMDGPU::S_XNOR_B32;
544 }
545 } else {
546 llvm_unreachable("unexpected opcode");
547 }
548
549 if (NewImm != 0) {
550 if (Dest->getReg().isVirtual() && SrcReg->isReg()) {
551 MRI->setRegAllocationHint(Dest->getReg(), 0, SrcReg->getReg());
552 MRI->setRegAllocationHint(SrcReg->getReg(), 0, Dest->getReg());
553 return true;
554 }
555
556 if (SrcReg->isReg() && SrcReg->getReg() == Dest->getReg()) {
557 const bool IsUndef = SrcReg->isUndef();
558 const bool IsKill = SrcReg->isKill();
559 MI.setDesc(TII->get(Opc));
560 if (Opc == AMDGPU::S_BITSET0_B32 ||
561 Opc == AMDGPU::S_BITSET1_B32) {
562 Src0->ChangeToImmediate(NewImm);
563 // Remove the immediate and add the tied input.
564 MI.getOperand(2).ChangeToRegister(Dest->getReg(), /*IsDef*/ false,
565 /*isImp*/ false, IsKill,
566 /*isDead*/ false, IsUndef);
567 MI.tieOperands(0, 2);
568 } else {
569 SrcImm->setImm(NewImm);
570 }
571 }
572 }
573
574 return false;
575 }
576
577 // This is the same as MachineInstr::readsRegister/modifiesRegister except
578 // it takes subregs into account.
instAccessReg(iterator_range<MachineInstr::const_mop_iterator> && R,Register Reg,unsigned SubReg) const579 bool SIShrinkInstructions::instAccessReg(
580 iterator_range<MachineInstr::const_mop_iterator> &&R, Register Reg,
581 unsigned SubReg) const {
582 for (const MachineOperand &MO : R) {
583 if (!MO.isReg())
584 continue;
585
586 if (Reg.isPhysical() && MO.getReg().isPhysical()) {
587 if (TRI->regsOverlap(Reg, MO.getReg()))
588 return true;
589 } else if (MO.getReg() == Reg && Reg.isVirtual()) {
590 LaneBitmask Overlap = TRI->getSubRegIndexLaneMask(SubReg) &
591 TRI->getSubRegIndexLaneMask(MO.getSubReg());
592 if (Overlap.any())
593 return true;
594 }
595 }
596 return false;
597 }
598
instReadsReg(const MachineInstr * MI,unsigned Reg,unsigned SubReg) const599 bool SIShrinkInstructions::instReadsReg(const MachineInstr *MI, unsigned Reg,
600 unsigned SubReg) const {
601 return instAccessReg(MI->uses(), Reg, SubReg);
602 }
603
instModifiesReg(const MachineInstr * MI,unsigned Reg,unsigned SubReg) const604 bool SIShrinkInstructions::instModifiesReg(const MachineInstr *MI, unsigned Reg,
605 unsigned SubReg) const {
606 return instAccessReg(MI->defs(), Reg, SubReg);
607 }
608
609 TargetInstrInfo::RegSubRegPair
getSubRegForIndex(Register Reg,unsigned Sub,unsigned I) const610 SIShrinkInstructions::getSubRegForIndex(Register Reg, unsigned Sub,
611 unsigned I) const {
612 if (TRI->getRegSizeInBits(Reg, *MRI) != 32) {
613 if (Reg.isPhysical()) {
614 Reg = TRI->getSubReg(Reg, TRI->getSubRegFromChannel(I));
615 } else {
616 Sub = TRI->getSubRegFromChannel(I + TRI->getChannelFromSubReg(Sub));
617 }
618 }
619 return TargetInstrInfo::RegSubRegPair(Reg, Sub);
620 }
621
dropInstructionKeepingImpDefs(MachineInstr & MI) const622 void SIShrinkInstructions::dropInstructionKeepingImpDefs(
623 MachineInstr &MI) const {
624 for (unsigned i = MI.getDesc().getNumOperands() +
625 MI.getDesc().implicit_uses().size() +
626 MI.getDesc().implicit_defs().size(),
627 e = MI.getNumOperands();
628 i != e; ++i) {
629 const MachineOperand &Op = MI.getOperand(i);
630 if (!Op.isDef())
631 continue;
632 BuildMI(*MI.getParent(), MI.getIterator(), MI.getDebugLoc(),
633 TII->get(AMDGPU::IMPLICIT_DEF), Op.getReg());
634 }
635
636 MI.eraseFromParent();
637 }
638
639 // Match:
640 // mov t, x
641 // mov x, y
642 // mov y, t
643 //
644 // =>
645 //
646 // mov t, x (t is potentially dead and move eliminated)
647 // v_swap_b32 x, y
648 //
649 // Returns next valid instruction pointer if was able to create v_swap_b32.
650 //
651 // This shall not be done too early not to prevent possible folding which may
652 // remove matched moves, and this should preferably be done before RA to
653 // release saved registers and also possibly after RA which can insert copies
654 // too.
655 //
656 // This is really just a generic peephole that is not a canonical shrinking,
657 // although requirements match the pass placement and it reduces code size too.
matchSwap(MachineInstr & MovT) const658 MachineInstr *SIShrinkInstructions::matchSwap(MachineInstr &MovT) const {
659 assert(MovT.getOpcode() == AMDGPU::V_MOV_B32_e32 ||
660 MovT.getOpcode() == AMDGPU::COPY);
661
662 Register T = MovT.getOperand(0).getReg();
663 unsigned Tsub = MovT.getOperand(0).getSubReg();
664 MachineOperand &Xop = MovT.getOperand(1);
665
666 if (!Xop.isReg())
667 return nullptr;
668 Register X = Xop.getReg();
669 unsigned Xsub = Xop.getSubReg();
670
671 unsigned Size = TII->getOpSize(MovT, 0) / 4;
672
673 if (!TRI->isVGPR(*MRI, X))
674 return nullptr;
675
676 const unsigned SearchLimit = 16;
677 unsigned Count = 0;
678 bool KilledT = false;
679 for (auto Iter = std::next(MovT.getIterator()),
680 E = MovT.getParent()->instr_end();
681 Iter != E && Count < SearchLimit && !KilledT; ++Iter, ++Count) {
682
683 MachineInstr *MovY = &*Iter;
684 KilledT = MovY->killsRegister(T, TRI);
685
686 if ((MovY->getOpcode() != AMDGPU::V_MOV_B32_e32 &&
687 MovY->getOpcode() != AMDGPU::COPY) ||
688 !MovY->getOperand(1).isReg() ||
689 MovY->getOperand(1).getReg() != T ||
690 MovY->getOperand(1).getSubReg() != Tsub)
691 continue;
692
693 Register Y = MovY->getOperand(0).getReg();
694 unsigned Ysub = MovY->getOperand(0).getSubReg();
695
696 if (!TRI->isVGPR(*MRI, Y))
697 continue;
698
699 MachineInstr *MovX = nullptr;
700 for (auto IY = MovY->getIterator(), I = std::next(MovT.getIterator());
701 I != IY; ++I) {
702 if (instReadsReg(&*I, X, Xsub) || instModifiesReg(&*I, Y, Ysub) ||
703 instModifiesReg(&*I, T, Tsub) ||
704 (MovX && instModifiesReg(&*I, X, Xsub))) {
705 MovX = nullptr;
706 break;
707 }
708 if (!instReadsReg(&*I, Y, Ysub)) {
709 if (!MovX && instModifiesReg(&*I, X, Xsub)) {
710 MovX = nullptr;
711 break;
712 }
713 continue;
714 }
715 if (MovX ||
716 (I->getOpcode() != AMDGPU::V_MOV_B32_e32 &&
717 I->getOpcode() != AMDGPU::COPY) ||
718 I->getOperand(0).getReg() != X ||
719 I->getOperand(0).getSubReg() != Xsub) {
720 MovX = nullptr;
721 break;
722 }
723
724 if (Size > 1 && (I->getNumImplicitOperands() > (I->isCopy() ? 0U : 1U)))
725 continue;
726
727 MovX = &*I;
728 }
729
730 if (!MovX)
731 continue;
732
733 LLVM_DEBUG(dbgs() << "Matched v_swap_b32:\n" << MovT << *MovX << *MovY);
734
735 for (unsigned I = 0; I < Size; ++I) {
736 TargetInstrInfo::RegSubRegPair X1, Y1;
737 X1 = getSubRegForIndex(X, Xsub, I);
738 Y1 = getSubRegForIndex(Y, Ysub, I);
739 MachineBasicBlock &MBB = *MovT.getParent();
740 auto MIB = BuildMI(MBB, MovX->getIterator(), MovT.getDebugLoc(),
741 TII->get(AMDGPU::V_SWAP_B32))
742 .addDef(X1.Reg, 0, X1.SubReg)
743 .addDef(Y1.Reg, 0, Y1.SubReg)
744 .addReg(Y1.Reg, 0, Y1.SubReg)
745 .addReg(X1.Reg, 0, X1.SubReg).getInstr();
746 if (MovX->hasRegisterImplicitUseOperand(AMDGPU::EXEC)) {
747 // Drop implicit EXEC.
748 MIB->removeOperand(MIB->getNumExplicitOperands());
749 MIB->copyImplicitOps(*MBB.getParent(), *MovX);
750 }
751 }
752 MovX->eraseFromParent();
753 dropInstructionKeepingImpDefs(*MovY);
754 MachineInstr *Next = &*std::next(MovT.getIterator());
755
756 if (T.isVirtual() && MRI->use_nodbg_empty(T)) {
757 dropInstructionKeepingImpDefs(MovT);
758 } else {
759 Xop.setIsKill(false);
760 for (int I = MovT.getNumImplicitOperands() - 1; I >= 0; --I ) {
761 unsigned OpNo = MovT.getNumExplicitOperands() + I;
762 const MachineOperand &Op = MovT.getOperand(OpNo);
763 if (Op.isKill() && TRI->regsOverlap(X, Op.getReg()))
764 MovT.removeOperand(OpNo);
765 }
766 }
767
768 return Next;
769 }
770
771 return nullptr;
772 }
773
774 // If an instruction has dead sdst replace it with NULL register on gfx1030+
tryReplaceDeadSDST(MachineInstr & MI) const775 bool SIShrinkInstructions::tryReplaceDeadSDST(MachineInstr &MI) const {
776 if (!ST->hasGFX10_3Insts())
777 return false;
778
779 MachineOperand *Op = TII->getNamedOperand(MI, AMDGPU::OpName::sdst);
780 if (!Op)
781 return false;
782 Register SDstReg = Op->getReg();
783 if (SDstReg.isPhysical() || !MRI->use_nodbg_empty(SDstReg))
784 return false;
785
786 Op->setReg(ST->isWave32() ? AMDGPU::SGPR_NULL : AMDGPU::SGPR_NULL64);
787 return true;
788 }
789
runOnMachineFunction(MachineFunction & MF)790 bool SIShrinkInstructions::runOnMachineFunction(MachineFunction &MF) {
791 if (skipFunction(MF.getFunction()))
792 return false;
793
794 this->MF = &MF;
795 MRI = &MF.getRegInfo();
796 ST = &MF.getSubtarget<GCNSubtarget>();
797 TII = ST->getInstrInfo();
798 TRI = &TII->getRegisterInfo();
799
800 unsigned VCCReg = ST->isWave32() ? AMDGPU::VCC_LO : AMDGPU::VCC;
801
802 std::vector<unsigned> I1Defs;
803
804 for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
805 BI != BE; ++BI) {
806
807 MachineBasicBlock &MBB = *BI;
808 MachineBasicBlock::iterator I, Next;
809 for (I = MBB.begin(); I != MBB.end(); I = Next) {
810 Next = std::next(I);
811 MachineInstr &MI = *I;
812
813 if (MI.getOpcode() == AMDGPU::V_MOV_B32_e32) {
814 // If this has a literal constant source that is the same as the
815 // reversed bits of an inline immediate, replace with a bitreverse of
816 // that constant. This saves 4 bytes in the common case of materializing
817 // sign bits.
818
819 // Test if we are after regalloc. We only want to do this after any
820 // optimizations happen because this will confuse them.
821 // XXX - not exactly a check for post-regalloc run.
822 MachineOperand &Src = MI.getOperand(1);
823 if (Src.isImm() && MI.getOperand(0).getReg().isPhysical()) {
824 int32_t ModImm;
825 unsigned ModOpcode =
826 canModifyToInlineImmOp32(TII, Src, ModImm, /*Scalar=*/false);
827 if (ModOpcode != 0) {
828 MI.setDesc(TII->get(ModOpcode));
829 Src.setImm(static_cast<int64_t>(ModImm));
830 continue;
831 }
832 }
833 }
834
835 if (ST->hasSwap() && (MI.getOpcode() == AMDGPU::V_MOV_B32_e32 ||
836 MI.getOpcode() == AMDGPU::COPY)) {
837 if (auto *NextMI = matchSwap(MI)) {
838 Next = NextMI->getIterator();
839 continue;
840 }
841 }
842
843 // Try to use S_ADDK_I32 and S_MULK_I32.
844 if (MI.getOpcode() == AMDGPU::S_ADD_I32 ||
845 MI.getOpcode() == AMDGPU::S_MUL_I32) {
846 const MachineOperand *Dest = &MI.getOperand(0);
847 MachineOperand *Src0 = &MI.getOperand(1);
848 MachineOperand *Src1 = &MI.getOperand(2);
849
850 if (!Src0->isReg() && Src1->isReg()) {
851 if (TII->commuteInstruction(MI, false, 1, 2))
852 std::swap(Src0, Src1);
853 }
854
855 // FIXME: This could work better if hints worked with subregisters. If
856 // we have a vector add of a constant, we usually don't get the correct
857 // allocation due to the subregister usage.
858 if (Dest->getReg().isVirtual() && Src0->isReg()) {
859 MRI->setRegAllocationHint(Dest->getReg(), 0, Src0->getReg());
860 MRI->setRegAllocationHint(Src0->getReg(), 0, Dest->getReg());
861 continue;
862 }
863
864 if (Src0->isReg() && Src0->getReg() == Dest->getReg()) {
865 if (Src1->isImm() && isKImmOperand(*Src1)) {
866 unsigned Opc = (MI.getOpcode() == AMDGPU::S_ADD_I32) ?
867 AMDGPU::S_ADDK_I32 : AMDGPU::S_MULK_I32;
868
869 Src1->setImm(SignExtend64(Src1->getImm(), 32));
870 MI.setDesc(TII->get(Opc));
871 MI.tieOperands(0, 1);
872 }
873 }
874 }
875
876 // Try to use s_cmpk_*
877 if (MI.isCompare() && TII->isSOPC(MI)) {
878 shrinkScalarCompare(MI);
879 continue;
880 }
881
882 // Try to use S_MOVK_I32, which will save 4 bytes for small immediates.
883 if (MI.getOpcode() == AMDGPU::S_MOV_B32) {
884 const MachineOperand &Dst = MI.getOperand(0);
885 MachineOperand &Src = MI.getOperand(1);
886
887 if (Src.isImm() && Dst.getReg().isPhysical()) {
888 unsigned ModOpc;
889 int32_t ModImm;
890 if (isKImmOperand(Src)) {
891 MI.setDesc(TII->get(AMDGPU::S_MOVK_I32));
892 Src.setImm(SignExtend64(Src.getImm(), 32));
893 } else if ((ModOpc = canModifyToInlineImmOp32(TII, Src, ModImm,
894 /*Scalar=*/true))) {
895 MI.setDesc(TII->get(ModOpc));
896 Src.setImm(static_cast<int64_t>(ModImm));
897 }
898 }
899
900 continue;
901 }
902
903 // Shrink scalar logic operations.
904 if (MI.getOpcode() == AMDGPU::S_AND_B32 ||
905 MI.getOpcode() == AMDGPU::S_OR_B32 ||
906 MI.getOpcode() == AMDGPU::S_XOR_B32) {
907 if (shrinkScalarLogicOp(MI))
908 continue;
909 }
910
911 if (TII->isMIMG(MI.getOpcode()) &&
912 ST->getGeneration() >= AMDGPUSubtarget::GFX10 &&
913 MF.getProperties().hasProperty(
914 MachineFunctionProperties::Property::NoVRegs)) {
915 shrinkMIMG(MI);
916 continue;
917 }
918
919 if (!TII->isVOP3(MI))
920 continue;
921
922 if (MI.getOpcode() == AMDGPU::V_MAD_F32_e64 ||
923 MI.getOpcode() == AMDGPU::V_FMA_F32_e64 ||
924 MI.getOpcode() == AMDGPU::V_MAD_F16_e64 ||
925 MI.getOpcode() == AMDGPU::V_FMA_F16_e64 ||
926 MI.getOpcode() == AMDGPU::V_FMA_F16_gfx9_e64) {
927 shrinkMadFma(MI);
928 continue;
929 }
930
931 if (!TII->hasVALU32BitEncoding(MI.getOpcode())) {
932 // If there is no chance we will shrink it and use VCC as sdst to get
933 // a 32 bit form try to replace dead sdst with NULL.
934 tryReplaceDeadSDST(MI);
935 continue;
936 }
937
938 if (!TII->canShrink(MI, *MRI)) {
939 // Try commuting the instruction and see if that enables us to shrink
940 // it.
941 if (!MI.isCommutable() || !TII->commuteInstruction(MI) ||
942 !TII->canShrink(MI, *MRI)) {
943 tryReplaceDeadSDST(MI);
944 continue;
945 }
946 }
947
948 int Op32 = AMDGPU::getVOPe32(MI.getOpcode());
949
950 if (TII->isVOPC(Op32)) {
951 MachineOperand &Op0 = MI.getOperand(0);
952 if (Op0.isReg()) {
953 // Exclude VOPCX instructions as these don't explicitly write a
954 // dst.
955 Register DstReg = Op0.getReg();
956 if (DstReg.isVirtual()) {
957 // VOPC instructions can only write to the VCC register. We can't
958 // force them to use VCC here, because this is only one register and
959 // cannot deal with sequences which would require multiple copies of
960 // VCC, e.g. S_AND_B64 (vcc = V_CMP_...), (vcc = V_CMP_...)
961 //
962 // So, instead of forcing the instruction to write to VCC, we
963 // provide a hint to the register allocator to use VCC and then we
964 // will run this pass again after RA and shrink it if it outputs to
965 // VCC.
966 MRI->setRegAllocationHint(DstReg, 0, VCCReg);
967 continue;
968 }
969 if (DstReg != VCCReg)
970 continue;
971 }
972 }
973
974 if (Op32 == AMDGPU::V_CNDMASK_B32_e32) {
975 // We shrink V_CNDMASK_B32_e64 using regalloc hints like we do for VOPC
976 // instructions.
977 const MachineOperand *Src2 =
978 TII->getNamedOperand(MI, AMDGPU::OpName::src2);
979 if (!Src2->isReg())
980 continue;
981 Register SReg = Src2->getReg();
982 if (SReg.isVirtual()) {
983 MRI->setRegAllocationHint(SReg, 0, VCCReg);
984 continue;
985 }
986 if (SReg != VCCReg)
987 continue;
988 }
989
990 // Check for the bool flag output for instructions like V_ADD_I32_e64.
991 const MachineOperand *SDst = TII->getNamedOperand(MI,
992 AMDGPU::OpName::sdst);
993
994 if (SDst) {
995 bool Next = false;
996
997 if (SDst->getReg() != VCCReg) {
998 if (SDst->getReg().isVirtual())
999 MRI->setRegAllocationHint(SDst->getReg(), 0, VCCReg);
1000 Next = true;
1001 }
1002
1003 // All of the instructions with carry outs also have an SGPR input in
1004 // src2.
1005 const MachineOperand *Src2 = TII->getNamedOperand(MI,
1006 AMDGPU::OpName::src2);
1007 if (Src2 && Src2->getReg() != VCCReg) {
1008 if (Src2->getReg().isVirtual())
1009 MRI->setRegAllocationHint(Src2->getReg(), 0, VCCReg);
1010 Next = true;
1011 }
1012
1013 if (Next)
1014 continue;
1015 }
1016
1017 // Pre-GFX10, shrinking VOP3 instructions pre-RA gave us the chance to
1018 // fold an immediate into the shrunk instruction as a literal operand. In
1019 // GFX10 VOP3 instructions can take a literal operand anyway, so there is
1020 // no advantage to doing this.
1021 if (ST->hasVOP3Literal() &&
1022 !MF.getProperties().hasProperty(
1023 MachineFunctionProperties::Property::NoVRegs))
1024 continue;
1025
1026 if (ST->hasTrue16BitInsts() && AMDGPU::isTrue16Inst(MI.getOpcode()) &&
1027 !shouldShrinkTrue16(MI))
1028 continue;
1029
1030 // We can shrink this instruction
1031 LLVM_DEBUG(dbgs() << "Shrinking " << MI);
1032
1033 MachineInstr *Inst32 = TII->buildShrunkInst(MI, Op32);
1034 ++NumInstructionsShrunk;
1035
1036 // Copy extra operands not present in the instruction definition.
1037 copyExtraImplicitOps(*Inst32, MI);
1038
1039 // Copy deadness from the old explicit vcc def to the new implicit def.
1040 if (SDst && SDst->isDead())
1041 Inst32->findRegisterDefOperand(VCCReg, /*TRI=*/nullptr)->setIsDead();
1042
1043 MI.eraseFromParent();
1044 foldImmediates(*Inst32);
1045
1046 LLVM_DEBUG(dbgs() << "e32 MI = " << *Inst32 << '\n');
1047 }
1048 }
1049 return false;
1050 }
1051