1 //===-- SIInstrInfo.cpp - SI Instruction Information ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief SI Implementation of TargetInstrInfo.
13 //===----------------------------------------------------------------------===//
16 #include "SIInstrInfo.h"
17 #include "AMDGPUTargetMachine.h"
18 #include "SIDefines.h"
19 #include "SIMachineFunctionInfo.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/MachineRegisterInfo.h"
22 #include "llvm/IR/Function.h"
23 #include "llvm/MC/MCInstrDesc.h"
27 SIInstrInfo::SIInstrInfo(const AMDGPUSubtarget &st)
28 : AMDGPUInstrInfo(st),
31 //===----------------------------------------------------------------------===//
32 // TargetInstrInfo callbacks
33 //===----------------------------------------------------------------------===//
36 SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
37 MachineBasicBlock::iterator MI, DebugLoc DL,
38 unsigned DestReg, unsigned SrcReg,
41 // If we are trying to copy to or from SCC, there is a bug somewhere else in
42 // the backend. While it may be theoretically possible to do this, it should
43 // never be necessary.
44 assert(DestReg != AMDGPU::SCC && SrcReg != AMDGPU::SCC);
46 static const int16_t Sub0_15[] = {
47 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
48 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
49 AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
50 AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15, 0
53 static const int16_t Sub0_7[] = {
54 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
55 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7, 0
58 static const int16_t Sub0_3[] = {
59 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
62 static const int16_t Sub0_2[] = {
63 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
66 static const int16_t Sub0_1[] = {
67 AMDGPU::sub0, AMDGPU::sub1, 0
71 const int16_t *SubIndices;
73 if (AMDGPU::M0 == DestReg) {
74 // Check if M0 isn't already set to this value
75 for (MachineBasicBlock::reverse_iterator E = MBB.rend(),
76 I = MachineBasicBlock::reverse_iterator(MI); I != E; ++I) {
78 if (!I->definesRegister(AMDGPU::M0))
81 unsigned Opc = I->getOpcode();
82 if (Opc != TargetOpcode::COPY && Opc != AMDGPU::S_MOV_B32)
85 if (!I->readsRegister(SrcReg))
88 // The copy isn't necessary
93 if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
94 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
95 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
96 .addReg(SrcReg, getKillRegState(KillSrc));
99 } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
100 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
101 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
102 .addReg(SrcReg, getKillRegState(KillSrc));
105 } else if (AMDGPU::SReg_128RegClass.contains(DestReg)) {
106 assert(AMDGPU::SReg_128RegClass.contains(SrcReg));
107 Opcode = AMDGPU::S_MOV_B32;
110 } else if (AMDGPU::SReg_256RegClass.contains(DestReg)) {
111 assert(AMDGPU::SReg_256RegClass.contains(SrcReg));
112 Opcode = AMDGPU::S_MOV_B32;
115 } else if (AMDGPU::SReg_512RegClass.contains(DestReg)) {
116 assert(AMDGPU::SReg_512RegClass.contains(SrcReg));
117 Opcode = AMDGPU::S_MOV_B32;
118 SubIndices = Sub0_15;
120 } else if (AMDGPU::VReg_32RegClass.contains(DestReg)) {
121 assert(AMDGPU::VReg_32RegClass.contains(SrcReg) ||
122 AMDGPU::SReg_32RegClass.contains(SrcReg));
123 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
124 .addReg(SrcReg, getKillRegState(KillSrc));
127 } else if (AMDGPU::VReg_64RegClass.contains(DestReg)) {
128 assert(AMDGPU::VReg_64RegClass.contains(SrcReg) ||
129 AMDGPU::SReg_64RegClass.contains(SrcReg));
130 Opcode = AMDGPU::V_MOV_B32_e32;
133 } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
134 assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
135 Opcode = AMDGPU::V_MOV_B32_e32;
138 } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
139 assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
140 AMDGPU::SReg_128RegClass.contains(SrcReg));
141 Opcode = AMDGPU::V_MOV_B32_e32;
144 } else if (AMDGPU::VReg_256RegClass.contains(DestReg)) {
145 assert(AMDGPU::VReg_256RegClass.contains(SrcReg) ||
146 AMDGPU::SReg_256RegClass.contains(SrcReg));
147 Opcode = AMDGPU::V_MOV_B32_e32;
150 } else if (AMDGPU::VReg_512RegClass.contains(DestReg)) {
151 assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
152 AMDGPU::SReg_512RegClass.contains(SrcReg));
153 Opcode = AMDGPU::V_MOV_B32_e32;
154 SubIndices = Sub0_15;
157 llvm_unreachable("Can't copy register!");
160 while (unsigned SubIdx = *SubIndices++) {
161 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
162 get(Opcode), RI.getSubReg(DestReg, SubIdx));
164 Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
167 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
171 unsigned SIInstrInfo::commuteOpcode(unsigned Opcode) const {
174 // Try to map original to commuted opcode
175 if ((NewOpc = AMDGPU::getCommuteRev(Opcode)) != -1)
178 // Try to map commuted to original opcode
179 if ((NewOpc = AMDGPU::getCommuteOrig(Opcode)) != -1)
185 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
186 MachineBasicBlock::iterator MI,
187 unsigned SrcReg, bool isKill,
189 const TargetRegisterClass *RC,
190 const TargetRegisterInfo *TRI) const {
191 MachineFunction *MF = MBB.getParent();
192 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
193 MachineRegisterInfo &MRI = MF->getRegInfo();
194 DebugLoc DL = MBB.findDebugLoc(MI);
195 unsigned KillFlag = isKill ? RegState::Kill : 0;
197 if (RI.hasVGPRs(RC)) {
198 LLVMContext &Ctx = MF->getFunction()->getContext();
199 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Can't spill VGPR!");
200 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), AMDGPU::VGPR0)
202 } else if (TRI->getCommonSubClass(RC, &AMDGPU::SGPR_32RegClass)) {
203 unsigned Lane = MFI->SpillTracker.reserveLanes(MRI, MF);
204 unsigned TgtReg = MFI->SpillTracker.LaneVGPR;
206 BuildMI(MBB, MI, DL, get(AMDGPU::V_WRITELANE_B32), TgtReg)
207 .addReg(SrcReg, KillFlag)
209 MFI->SpillTracker.addSpilledReg(FrameIndex, TgtReg, Lane);
210 } else if (RI.isSGPRClass(RC)) {
211 // We are only allowed to create one new instruction when spilling
212 // registers, so we need to use pseudo instruction for vector
215 // Reserve a spot in the spill tracker for each sub-register of
216 // the vector register.
217 unsigned NumSubRegs = RC->getSize() / 4;
218 unsigned FirstLane = MFI->SpillTracker.reserveLanes(MRI, MF, NumSubRegs);
219 MFI->SpillTracker.addSpilledReg(FrameIndex, MFI->SpillTracker.LaneVGPR,
223 switch (RC->getSize() * 8) {
224 case 64: Opcode = AMDGPU::SI_SPILL_S64_SAVE; break;
225 case 128: Opcode = AMDGPU::SI_SPILL_S128_SAVE; break;
226 case 256: Opcode = AMDGPU::SI_SPILL_S256_SAVE; break;
227 case 512: Opcode = AMDGPU::SI_SPILL_S512_SAVE; break;
228 default: llvm_unreachable("Cannot spill register class");
231 BuildMI(MBB, MI, DL, get(Opcode), MFI->SpillTracker.LaneVGPR)
235 llvm_unreachable("VGPR spilling not supported");
239 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
240 MachineBasicBlock::iterator MI,
241 unsigned DestReg, int FrameIndex,
242 const TargetRegisterClass *RC,
243 const TargetRegisterInfo *TRI) const {
244 MachineFunction *MF = MBB.getParent();
245 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
246 DebugLoc DL = MBB.findDebugLoc(MI);
248 if (RI.hasVGPRs(RC)) {
249 LLVMContext &Ctx = MF->getFunction()->getContext();
250 Ctx.emitError("SIInstrInfo::loadRegToStackSlot - Can't retrieve spilled VGPR!");
251 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
253 } else if (RI.isSGPRClass(RC)){
255 switch(RC->getSize() * 8) {
256 case 32: Opcode = AMDGPU::SI_SPILL_S32_RESTORE; break;
257 case 64: Opcode = AMDGPU::SI_SPILL_S64_RESTORE; break;
258 case 128: Opcode = AMDGPU::SI_SPILL_S128_RESTORE; break;
259 case 256: Opcode = AMDGPU::SI_SPILL_S256_RESTORE; break;
260 case 512: Opcode = AMDGPU::SI_SPILL_S512_RESTORE; break;
261 default: llvm_unreachable("Cannot spill register class");
264 SIMachineFunctionInfo::SpilledReg Spill =
265 MFI->SpillTracker.getSpilledReg(FrameIndex);
267 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
271 llvm_unreachable("VGPR spilling not supported");
275 static unsigned getNumSubRegsForSpillOp(unsigned Op) {
278 case AMDGPU::SI_SPILL_S512_SAVE:
279 case AMDGPU::SI_SPILL_S512_RESTORE:
281 case AMDGPU::SI_SPILL_S256_SAVE:
282 case AMDGPU::SI_SPILL_S256_RESTORE:
284 case AMDGPU::SI_SPILL_S128_SAVE:
285 case AMDGPU::SI_SPILL_S128_RESTORE:
287 case AMDGPU::SI_SPILL_S64_SAVE:
288 case AMDGPU::SI_SPILL_S64_RESTORE:
290 case AMDGPU::SI_SPILL_S32_RESTORE:
292 default: llvm_unreachable("Invalid spill opcode");
296 void SIInstrInfo::insertNOPs(MachineBasicBlock::iterator MI,
305 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(AMDGPU::S_NOP))
310 bool SIInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
311 SIMachineFunctionInfo *MFI =
312 MI->getParent()->getParent()->getInfo<SIMachineFunctionInfo>();
313 MachineBasicBlock &MBB = *MI->getParent();
314 DebugLoc DL = MBB.findDebugLoc(MI);
315 switch (MI->getOpcode()) {
316 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
318 // SGPR register spill
319 case AMDGPU::SI_SPILL_S512_SAVE:
320 case AMDGPU::SI_SPILL_S256_SAVE:
321 case AMDGPU::SI_SPILL_S128_SAVE:
322 case AMDGPU::SI_SPILL_S64_SAVE: {
323 unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
324 unsigned FrameIndex = MI->getOperand(2).getImm();
326 for (unsigned i = 0, e = NumSubRegs; i < e; ++i) {
327 SIMachineFunctionInfo::SpilledReg Spill;
328 unsigned SubReg = RI.getPhysRegSubReg(MI->getOperand(1).getReg(),
329 &AMDGPU::SGPR_32RegClass, i);
330 Spill = MFI->SpillTracker.getSpilledReg(FrameIndex);
332 BuildMI(MBB, MI, DL, get(AMDGPU::V_WRITELANE_B32),
333 MI->getOperand(0).getReg())
335 .addImm(Spill.Lane + i);
337 MI->eraseFromParent();
341 // SGPR register restore
342 case AMDGPU::SI_SPILL_S512_RESTORE:
343 case AMDGPU::SI_SPILL_S256_RESTORE:
344 case AMDGPU::SI_SPILL_S128_RESTORE:
345 case AMDGPU::SI_SPILL_S64_RESTORE:
346 case AMDGPU::SI_SPILL_S32_RESTORE: {
347 unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
349 for (unsigned i = 0, e = NumSubRegs; i < e; ++i) {
350 SIMachineFunctionInfo::SpilledReg Spill;
351 unsigned FrameIndex = MI->getOperand(2).getImm();
352 unsigned SubReg = RI.getPhysRegSubReg(MI->getOperand(0).getReg(),
353 &AMDGPU::SGPR_32RegClass, i);
354 Spill = MFI->SpillTracker.getSpilledReg(FrameIndex);
356 BuildMI(MBB, MI, DL, get(AMDGPU::V_READLANE_B32), SubReg)
357 .addReg(MI->getOperand(1).getReg())
358 .addImm(Spill.Lane + i);
361 MI->eraseFromParent();
368 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
371 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
372 if (MI->getNumOperands() < 3 || !MI->getOperand(1).isReg())
375 // Cannot commute VOP2 if src0 is SGPR.
376 if (isVOP2(MI->getOpcode()) && MI->getOperand(1).isReg() &&
377 RI.isSGPRClass(MRI.getRegClass(MI->getOperand(1).getReg())))
380 if (!MI->getOperand(2).isReg()) {
381 // XXX: Commute instructions with FPImm operands
382 if (NewMI || MI->getOperand(2).isFPImm() ||
383 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
387 // XXX: Commute VOP3 instructions with abs and neg set.
388 if (isVOP3(MI->getOpcode()) &&
389 (MI->getOperand(AMDGPU::getNamedOperandIdx(MI->getOpcode(),
390 AMDGPU::OpName::abs)).getImm() ||
391 MI->getOperand(AMDGPU::getNamedOperandIdx(MI->getOpcode(),
392 AMDGPU::OpName::neg)).getImm()))
395 unsigned Reg = MI->getOperand(1).getReg();
396 unsigned SubReg = MI->getOperand(1).getSubReg();
397 MI->getOperand(1).ChangeToImmediate(MI->getOperand(2).getImm());
398 MI->getOperand(2).ChangeToRegister(Reg, false);
399 MI->getOperand(2).setSubReg(SubReg);
401 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
405 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
410 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
411 MachineBasicBlock::iterator I,
413 unsigned SrcReg) const {
414 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
415 DstReg) .addReg(SrcReg);
418 bool SIInstrInfo::isMov(unsigned Opcode) const {
420 default: return false;
421 case AMDGPU::S_MOV_B32:
422 case AMDGPU::S_MOV_B64:
423 case AMDGPU::V_MOV_B32_e32:
424 case AMDGPU::V_MOV_B32_e64:
430 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
431 return RC != &AMDGPU::EXECRegRegClass;
435 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
436 AliasAnalysis *AA) const {
437 switch(MI->getOpcode()) {
438 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
439 case AMDGPU::S_MOV_B32:
440 case AMDGPU::S_MOV_B64:
441 case AMDGPU::V_MOV_B32_e32:
442 return MI->getOperand(1).isImm();
448 // Helper function generated by tablegen. We are wrapping this with
449 // an SIInstrInfo function that reutrns bool rather than int.
450 int isDS(uint16_t Opcode);
454 bool SIInstrInfo::isDS(uint16_t Opcode) const {
455 return ::AMDGPU::isDS(Opcode) != -1;
458 int SIInstrInfo::isMIMG(uint16_t Opcode) const {
459 return get(Opcode).TSFlags & SIInstrFlags::MIMG;
462 int SIInstrInfo::isSMRD(uint16_t Opcode) const {
463 return get(Opcode).TSFlags & SIInstrFlags::SMRD;
466 bool SIInstrInfo::isVOP1(uint16_t Opcode) const {
467 return get(Opcode).TSFlags & SIInstrFlags::VOP1;
470 bool SIInstrInfo::isVOP2(uint16_t Opcode) const {
471 return get(Opcode).TSFlags & SIInstrFlags::VOP2;
474 bool SIInstrInfo::isVOP3(uint16_t Opcode) const {
475 return get(Opcode).TSFlags & SIInstrFlags::VOP3;
478 bool SIInstrInfo::isVOPC(uint16_t Opcode) const {
479 return get(Opcode).TSFlags & SIInstrFlags::VOPC;
482 bool SIInstrInfo::isSALUInstr(const MachineInstr &MI) const {
483 return get(MI.getOpcode()).TSFlags & SIInstrFlags::SALU;
486 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
487 int32_t Val = Imm.getSExtValue();
488 if (Val >= -16 && Val <= 64)
491 // The actual type of the operand does not seem to matter as long
492 // as the bits match one of the inline immediate values. For example:
494 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
495 // so it is a legal inline immediate.
497 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
498 // floating-point, so it is a legal inline immediate.
500 return (APInt::floatToBits(0.0f) == Imm) ||
501 (APInt::floatToBits(1.0f) == Imm) ||
502 (APInt::floatToBits(-1.0f) == Imm) ||
503 (APInt::floatToBits(0.5f) == Imm) ||
504 (APInt::floatToBits(-0.5f) == Imm) ||
505 (APInt::floatToBits(2.0f) == Imm) ||
506 (APInt::floatToBits(-2.0f) == Imm) ||
507 (APInt::floatToBits(4.0f) == Imm) ||
508 (APInt::floatToBits(-4.0f) == Imm);
511 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO) const {
513 return isInlineConstant(APInt(32, MO.getImm(), true));
516 APFloat FpImm = MO.getFPImm()->getValueAPF();
517 return isInlineConstant(FpImm.bitcastToAPInt());
523 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO) const {
524 return (MO.isImm() || MO.isFPImm()) && !isInlineConstant(MO);
527 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
528 StringRef &ErrInfo) const {
529 uint16_t Opcode = MI->getOpcode();
530 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
531 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
532 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
534 // Make sure the number of operands is correct.
535 const MCInstrDesc &Desc = get(Opcode);
536 if (!Desc.isVariadic() &&
537 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
538 ErrInfo = "Instruction has wrong number of operands.";
542 // Make sure the register classes are correct
543 for (unsigned i = 0, e = Desc.getNumOperands(); i != e; ++i) {
544 switch (Desc.OpInfo[i].OperandType) {
545 case MCOI::OPERAND_REGISTER:
547 case MCOI::OPERAND_IMMEDIATE:
548 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFPImm()) {
549 ErrInfo = "Expected immediate, but got non-immediate";
557 if (!MI->getOperand(i).isReg())
560 int RegClass = Desc.OpInfo[i].RegClass;
561 if (RegClass != -1) {
562 unsigned Reg = MI->getOperand(i).getReg();
563 if (TargetRegisterInfo::isVirtualRegister(Reg))
566 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
567 if (!RC->contains(Reg)) {
568 ErrInfo = "Operand has incorrect register class.";
576 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
577 unsigned ConstantBusCount = 0;
578 unsigned SGPRUsed = AMDGPU::NoRegister;
579 for (int i = 0, e = MI->getNumOperands(); i != e; ++i) {
580 const MachineOperand &MO = MI->getOperand(i);
581 if (MO.isReg() && MO.isUse() &&
582 !TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
584 // EXEC register uses the constant bus.
585 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
588 // SGPRs use the constant bus
589 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
591 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
592 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
593 if (SGPRUsed != MO.getReg()) {
595 SGPRUsed = MO.getReg();
599 // Literal constants use the constant bus.
600 if (isLiteralConstant(MO))
603 if (ConstantBusCount > 1) {
604 ErrInfo = "VOP* instruction uses the constant bus more than once";
609 // Verify SRC1 for VOP2 and VOPC
610 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
611 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
612 if (Src1.isImm() || Src1.isFPImm()) {
613 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
619 if (isVOP3(Opcode)) {
620 if (Src0Idx != -1 && isLiteralConstant(MI->getOperand(Src0Idx))) {
621 ErrInfo = "VOP3 src0 cannot be a literal constant.";
624 if (Src1Idx != -1 && isLiteralConstant(MI->getOperand(Src1Idx))) {
625 ErrInfo = "VOP3 src1 cannot be a literal constant.";
628 if (Src2Idx != -1 && isLiteralConstant(MI->getOperand(Src2Idx))) {
629 ErrInfo = "VOP3 src2 cannot be a literal constant.";
636 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
637 switch (MI.getOpcode()) {
638 default: return AMDGPU::INSTRUCTION_LIST_END;
639 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
640 case AMDGPU::COPY: return AMDGPU::COPY;
641 case AMDGPU::PHI: return AMDGPU::PHI;
642 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
643 case AMDGPU::S_MOV_B32:
644 return MI.getOperand(1).isReg() ?
645 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
646 case AMDGPU::S_ADD_I32: return AMDGPU::V_ADD_I32_e32;
647 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
648 case AMDGPU::S_SUB_I32: return AMDGPU::V_SUB_I32_e32;
649 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
650 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
651 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
652 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
653 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
654 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
655 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
656 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
657 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
658 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
659 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
660 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
661 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
662 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
663 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
664 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
665 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
666 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
667 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
668 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
669 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
670 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
671 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
672 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
673 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
674 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
675 case AMDGPU::S_LOAD_DWORD_IMM:
676 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
677 case AMDGPU::S_LOAD_DWORDX2_IMM:
678 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
679 case AMDGPU::S_LOAD_DWORDX4_IMM:
680 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
681 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e32;
682 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
686 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
687 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
690 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
691 unsigned OpNo) const {
692 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
693 const MCInstrDesc &Desc = get(MI.getOpcode());
694 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
695 Desc.OpInfo[OpNo].RegClass == -1)
696 return MRI.getRegClass(MI.getOperand(OpNo).getReg());
698 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
699 return RI.getRegClass(RCID);
702 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
703 switch (MI.getOpcode()) {
705 case AMDGPU::REG_SEQUENCE:
707 case AMDGPU::INSERT_SUBREG:
708 return RI.hasVGPRs(getOpRegClass(MI, 0));
710 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
714 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
715 MachineBasicBlock::iterator I = MI;
716 MachineOperand &MO = MI->getOperand(OpIdx);
717 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
718 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
719 const TargetRegisterClass *RC = RI.getRegClass(RCID);
720 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
722 Opcode = AMDGPU::COPY;
723 } else if (RI.isSGPRClass(RC)) {
724 Opcode = AMDGPU::S_MOV_B32;
727 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
728 unsigned Reg = MRI.createVirtualRegister(VRC);
729 BuildMI(*MI->getParent(), I, MI->getParent()->findDebugLoc(I), get(Opcode),
731 MO.ChangeToRegister(Reg, false);
734 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
735 MachineRegisterInfo &MRI,
736 MachineOperand &SuperReg,
737 const TargetRegisterClass *SuperRC,
739 const TargetRegisterClass *SubRC)
741 assert(SuperReg.isReg());
743 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
744 unsigned SubReg = MRI.createVirtualRegister(SubRC);
746 // Just in case the super register is itself a sub-register, copy it to a new
747 // value so we don't need to worry about merging its subreg index with the
748 // SubIdx passed to this function. The register coalescer should be able to
749 // eliminate this extra copy.
750 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
752 .addOperand(SuperReg);
754 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
756 .addReg(NewSuperReg, 0, SubIdx);
760 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
761 MachineBasicBlock::iterator MII,
762 MachineRegisterInfo &MRI,
764 const TargetRegisterClass *SuperRC,
766 const TargetRegisterClass *SubRC) const {
768 // XXX - Is there a better way to do this?
769 if (SubIdx == AMDGPU::sub0)
770 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
771 if (SubIdx == AMDGPU::sub1)
772 return MachineOperand::CreateImm(Op.getImm() >> 32);
774 llvm_unreachable("Unhandled register index for immediate");
777 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
779 return MachineOperand::CreateReg(SubReg, false);
782 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
783 MachineBasicBlock::iterator MI,
784 MachineRegisterInfo &MRI,
785 const TargetRegisterClass *RC,
786 const MachineOperand &Op) const {
787 MachineBasicBlock *MBB = MI->getParent();
788 DebugLoc DL = MI->getDebugLoc();
789 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
790 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
791 unsigned Dst = MRI.createVirtualRegister(RC);
793 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
795 .addImm(Op.getImm() & 0xFFFFFFFF);
796 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
798 .addImm(Op.getImm() >> 32);
800 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
802 .addImm(AMDGPU::sub0)
804 .addImm(AMDGPU::sub1);
806 Worklist.push_back(Lo);
807 Worklist.push_back(Hi);
812 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
813 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
814 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
815 AMDGPU::OpName::src0);
816 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
817 AMDGPU::OpName::src1);
818 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
819 AMDGPU::OpName::src2);
822 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
823 MachineOperand &Src0 = MI->getOperand(Src0Idx);
824 MachineOperand &Src1 = MI->getOperand(Src1Idx);
826 // If the instruction implicitly reads VCC, we can't have any SGPR operands,
828 bool ReadsVCC = MI->readsRegister(AMDGPU::VCC, &RI);
829 if (ReadsVCC && Src0.isReg() &&
830 RI.isSGPRClass(MRI.getRegClass(Src0.getReg()))) {
831 legalizeOpWithMove(MI, Src0Idx);
835 if (ReadsVCC && Src1.isReg() &&
836 RI.isSGPRClass(MRI.getRegClass(Src1.getReg()))) {
837 legalizeOpWithMove(MI, Src1Idx);
841 // Legalize VOP2 instructions where src1 is not a VGPR. An SGPR input must
842 // be the first operand, and there can only be one.
843 if (Src1.isImm() || Src1.isFPImm() ||
844 (Src1.isReg() && RI.isSGPRClass(MRI.getRegClass(Src1.getReg())))) {
845 if (MI->isCommutable()) {
846 if (commuteInstruction(MI))
849 legalizeOpWithMove(MI, Src1Idx);
853 // XXX - Do any VOP3 instructions read VCC?
855 if (isVOP3(MI->getOpcode())) {
856 int VOP3Idx[3] = {Src0Idx, Src1Idx, Src2Idx};
857 unsigned SGPRReg = AMDGPU::NoRegister;
858 for (unsigned i = 0; i < 3; ++i) {
859 int Idx = VOP3Idx[i];
862 MachineOperand &MO = MI->getOperand(Idx);
865 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
866 continue; // VGPRs are legal
868 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
870 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
871 SGPRReg = MO.getReg();
872 // We can use one SGPR in each VOP3 instruction.
875 } else if (!isLiteralConstant(MO)) {
876 // If it is not a register and not a literal constant, then it must be
877 // an inline constant which is always legal.
880 // If we make it this far, then the operand is not legal and we must
882 legalizeOpWithMove(MI, Idx);
886 // Legalize REG_SEQUENCE and PHI
887 // The register class of the operands much be the same type as the register
888 // class of the output.
889 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
890 MI->getOpcode() == AMDGPU::PHI) {
891 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
892 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
893 if (!MI->getOperand(i).isReg() ||
894 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
896 const TargetRegisterClass *OpRC =
897 MRI.getRegClass(MI->getOperand(i).getReg());
898 if (RI.hasVGPRs(OpRC)) {
905 // If any of the operands are VGPR registers, then they all most be
906 // otherwise we will create illegal VGPR->SGPR copies when legalizing
908 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
911 VRC = RI.getEquivalentVGPRClass(SRC);
918 // Update all the operands so they have the same type.
919 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
920 if (!MI->getOperand(i).isReg() ||
921 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
923 unsigned DstReg = MRI.createVirtualRegister(RC);
924 MachineBasicBlock *InsertBB;
925 MachineBasicBlock::iterator Insert;
926 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
927 InsertBB = MI->getParent();
930 // MI is a PHI instruction.
931 InsertBB = MI->getOperand(i + 1).getMBB();
932 Insert = InsertBB->getFirstTerminator();
934 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
935 get(AMDGPU::COPY), DstReg)
936 .addOperand(MI->getOperand(i));
937 MI->getOperand(i).setReg(DstReg);
941 // Legalize INSERT_SUBREG
942 // src0 must have the same register class as dst
943 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
944 unsigned Dst = MI->getOperand(0).getReg();
945 unsigned Src0 = MI->getOperand(1).getReg();
946 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
947 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
948 if (DstRC != Src0RC) {
949 MachineBasicBlock &MBB = *MI->getParent();
950 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
951 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
953 MI->getOperand(1).setReg(NewSrc0);
958 // Legalize MUBUF* instructions
959 // FIXME: If we start using the non-addr64 instructions for compute, we
960 // may need to legalize them here.
962 int SRsrcIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
963 AMDGPU::OpName::srsrc);
964 int VAddrIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
965 AMDGPU::OpName::vaddr);
966 if (SRsrcIdx != -1 && VAddrIdx != -1) {
967 const TargetRegisterClass *VAddrRC =
968 RI.getRegClass(get(MI->getOpcode()).OpInfo[VAddrIdx].RegClass);
970 if(VAddrRC->getSize() == 8 &&
971 MRI.getRegClass(MI->getOperand(SRsrcIdx).getReg()) != VAddrRC) {
972 // We have a MUBUF instruction that uses a 64-bit vaddr register and
973 // srsrc has the incorrect register class. In order to fix this, we
974 // need to extract the pointer from the resource descriptor (srsrc),
975 // add it to the value of vadd, then store the result in the vaddr
976 // operand. Then, we need to set the pointer field of the resource
977 // descriptor to zero.
979 MachineBasicBlock &MBB = *MI->getParent();
980 MachineOperand &SRsrcOp = MI->getOperand(SRsrcIdx);
981 MachineOperand &VAddrOp = MI->getOperand(VAddrIdx);
982 unsigned SRsrcPtrLo, SRsrcPtrHi, VAddrLo, VAddrHi;
983 unsigned NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
984 unsigned NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
985 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
986 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
987 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
988 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
989 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
991 // SRsrcPtrLo = srsrc:sub0
992 SRsrcPtrLo = buildExtractSubReg(MI, MRI, SRsrcOp,
993 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VReg_32RegClass);
995 // SRsrcPtrHi = srsrc:sub1
996 SRsrcPtrHi = buildExtractSubReg(MI, MRI, SRsrcOp,
997 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VReg_32RegClass);
999 // VAddrLo = vaddr:sub0
1000 VAddrLo = buildExtractSubReg(MI, MRI, VAddrOp,
1001 &AMDGPU::VReg_64RegClass, AMDGPU::sub0, &AMDGPU::VReg_32RegClass);
1003 // VAddrHi = vaddr:sub1
1004 VAddrHi = buildExtractSubReg(MI, MRI, VAddrOp,
1005 &AMDGPU::VReg_64RegClass, AMDGPU::sub1, &AMDGPU::VReg_32RegClass);
1007 // NewVaddrLo = SRsrcPtrLo + VAddrLo
1008 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1012 .addReg(AMDGPU::VCC, RegState::Define | RegState::Implicit);
1014 // NewVaddrHi = SRsrcPtrHi + VAddrHi
1015 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1019 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1020 .addReg(AMDGPU::VCC, RegState::Implicit);
1022 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1023 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1026 .addImm(AMDGPU::sub0)
1028 .addImm(AMDGPU::sub1);
1031 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1035 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1036 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1038 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1040 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1041 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1043 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1045 // NewSRsrc = {Zero64, SRsrcFormat}
1046 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1049 .addImm(AMDGPU::sub0_sub1)
1050 .addReg(SRsrcFormatLo)
1051 .addImm(AMDGPU::sub2)
1052 .addReg(SRsrcFormatHi)
1053 .addImm(AMDGPU::sub3);
1055 // Update the instruction to use NewVaddr
1056 MI->getOperand(VAddrIdx).setReg(NewVAddr);
1057 // Update the instruction to use NewSRsrc
1058 MI->getOperand(SRsrcIdx).setReg(NewSRsrc);
1063 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
1064 MachineBasicBlock *MBB = MI->getParent();
1065 switch (MI->getOpcode()) {
1066 case AMDGPU::S_LOAD_DWORD_IMM:
1067 case AMDGPU::S_LOAD_DWORD_SGPR:
1068 case AMDGPU::S_LOAD_DWORDX2_IMM:
1069 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1070 case AMDGPU::S_LOAD_DWORDX4_IMM:
1071 case AMDGPU::S_LOAD_DWORDX4_SGPR:
1072 unsigned NewOpcode = getVALUOp(*MI);
1076 if (MI->getOperand(2).isReg()) {
1077 RegOffset = MI->getOperand(2).getReg();
1080 assert(MI->getOperand(2).isImm());
1081 // SMRD instructions take a dword offsets and MUBUF instructions
1082 // take a byte offset.
1083 ImmOffset = MI->getOperand(2).getImm() << 2;
1084 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1085 if (isUInt<12>(ImmOffset)) {
1086 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1090 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1097 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1098 unsigned DWord0 = RegOffset;
1099 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1100 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1101 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1103 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
1105 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
1106 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1107 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
1108 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1109 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
1111 .addImm(AMDGPU::sub0)
1113 .addImm(AMDGPU::sub1)
1115 .addImm(AMDGPU::sub2)
1117 .addImm(AMDGPU::sub3);
1118 MI->setDesc(get(NewOpcode));
1119 if (MI->getOperand(2).isReg()) {
1120 MI->getOperand(2).setReg(MI->getOperand(1).getReg());
1122 MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
1124 MI->getOperand(1).setReg(SRsrc);
1125 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
1129 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
1130 SmallVector<MachineInstr *, 128> Worklist;
1131 Worklist.push_back(&TopInst);
1133 while (!Worklist.empty()) {
1134 MachineInstr *Inst = Worklist.pop_back_val();
1135 MachineBasicBlock *MBB = Inst->getParent();
1136 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1138 unsigned Opcode = Inst->getOpcode();
1139 unsigned NewOpcode = getVALUOp(*Inst);
1141 // Handle some special cases
1144 if (isSMRD(Inst->getOpcode())) {
1145 moveSMRDToVALU(Inst, MRI);
1148 case AMDGPU::S_MOV_B64: {
1149 DebugLoc DL = Inst->getDebugLoc();
1151 // If the source operand is a register we can replace this with a
1153 if (Inst->getOperand(1).isReg()) {
1154 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
1155 .addOperand(Inst->getOperand(0))
1156 .addOperand(Inst->getOperand(1));
1157 Worklist.push_back(Copy);
1159 // Otherwise, we need to split this into two movs, because there is
1160 // no 64-bit VALU move instruction.
1161 unsigned Reg = Inst->getOperand(0).getReg();
1162 unsigned Dst = split64BitImm(Worklist,
1165 MRI.getRegClass(Reg),
1166 Inst->getOperand(1));
1167 MRI.replaceRegWith(Reg, Dst);
1169 Inst->eraseFromParent();
1172 case AMDGPU::S_AND_B64:
1173 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
1174 Inst->eraseFromParent();
1177 case AMDGPU::S_OR_B64:
1178 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
1179 Inst->eraseFromParent();
1182 case AMDGPU::S_XOR_B64:
1183 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
1184 Inst->eraseFromParent();
1187 case AMDGPU::S_NOT_B64:
1188 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
1189 Inst->eraseFromParent();
1192 case AMDGPU::S_BCNT1_I32_B64:
1193 splitScalar64BitBCNT(Worklist, Inst);
1194 Inst->eraseFromParent();
1197 case AMDGPU::S_BFE_U64:
1198 case AMDGPU::S_BFE_I64:
1199 case AMDGPU::S_BFM_B64:
1200 llvm_unreachable("Moving this op to VALU not implemented");
1203 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
1204 // We cannot move this instruction to the VALU, so we should try to
1205 // legalize its operands instead.
1206 legalizeOperands(Inst);
1210 // Use the new VALU Opcode.
1211 const MCInstrDesc &NewDesc = get(NewOpcode);
1212 Inst->setDesc(NewDesc);
1214 // Remove any references to SCC. Vector instructions can't read from it, and
1215 // We're just about to add the implicit use / defs of VCC, and we don't want
1217 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
1218 MachineOperand &Op = Inst->getOperand(i);
1219 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
1220 Inst->RemoveOperand(i);
1223 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
1224 // We are converting these to a BFE, so we need to add the missing
1225 // operands for the size and offset.
1226 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
1227 Inst->addOperand(Inst->getOperand(1));
1228 Inst->getOperand(1).ChangeToImmediate(0);
1229 Inst->addOperand(MachineOperand::CreateImm(0));
1230 Inst->addOperand(MachineOperand::CreateImm(0));
1231 Inst->addOperand(MachineOperand::CreateImm(0));
1232 Inst->addOperand(MachineOperand::CreateImm(Size));
1234 // XXX - Other pointless operands. There are 4, but it seems you only need
1235 // 3 to not hit an assertion later in MCInstLower.
1236 Inst->addOperand(MachineOperand::CreateImm(0));
1237 Inst->addOperand(MachineOperand::CreateImm(0));
1238 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
1239 // The VALU version adds the second operand to the result, so insert an
1241 Inst->addOperand(MachineOperand::CreateImm(0));
1244 addDescImplicitUseDef(NewDesc, Inst);
1246 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
1247 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
1248 // If we need to move this to VGPRs, we need to unpack the second operand
1249 // back into the 2 separate ones for bit offset and width.
1250 assert(OffsetWidthOp.isImm() &&
1251 "Scalar BFE is only implemented for constant width and offset");
1252 uint32_t Imm = OffsetWidthOp.getImm();
1254 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
1255 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
1257 Inst->RemoveOperand(2); // Remove old immediate.
1258 Inst->addOperand(Inst->getOperand(1));
1259 Inst->getOperand(1).ChangeToImmediate(0);
1260 Inst->addOperand(MachineOperand::CreateImm(0));
1261 Inst->addOperand(MachineOperand::CreateImm(Offset));
1262 Inst->addOperand(MachineOperand::CreateImm(0));
1263 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
1264 Inst->addOperand(MachineOperand::CreateImm(0));
1265 Inst->addOperand(MachineOperand::CreateImm(0));
1268 // Update the destination register class.
1270 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
1273 // For target instructions, getOpRegClass just returns the virtual
1274 // register class associated with the operand, so we need to find an
1275 // equivalent VGPR register class in order to move the instruction to the
1279 case AMDGPU::REG_SEQUENCE:
1280 case AMDGPU::INSERT_SUBREG:
1281 if (RI.hasVGPRs(NewDstRC))
1283 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
1291 unsigned DstReg = Inst->getOperand(0).getReg();
1292 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1293 MRI.replaceRegWith(DstReg, NewDstReg);
1295 // Legalize the operands
1296 legalizeOperands(Inst);
1298 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
1299 E = MRI.use_end(); I != E; ++I) {
1300 MachineInstr &UseMI = *I->getParent();
1301 if (!canReadVGPR(UseMI, I.getOperandNo())) {
1302 Worklist.push_back(&UseMI);
1308 //===----------------------------------------------------------------------===//
1309 // Indirect addressing callbacks
1310 //===----------------------------------------------------------------------===//
1312 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
1313 unsigned Channel) const {
1314 assert(Channel == 0);
1318 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
1319 return &AMDGPU::VReg_32RegClass;
1322 void SIInstrInfo::splitScalar64BitUnaryOp(
1323 SmallVectorImpl<MachineInstr *> &Worklist,
1325 unsigned Opcode) const {
1326 MachineBasicBlock &MBB = *Inst->getParent();
1327 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1329 MachineOperand &Dest = Inst->getOperand(0);
1330 MachineOperand &Src0 = Inst->getOperand(1);
1331 DebugLoc DL = Inst->getDebugLoc();
1333 MachineBasicBlock::iterator MII = Inst;
1335 const MCInstrDesc &InstDesc = get(Opcode);
1336 const TargetRegisterClass *Src0RC = Src0.isReg() ?
1337 MRI.getRegClass(Src0.getReg()) :
1338 &AMDGPU::SGPR_32RegClass;
1340 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
1342 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1343 AMDGPU::sub0, Src0SubRC);
1345 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
1346 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
1348 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
1349 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
1350 .addOperand(SrcReg0Sub0);
1352 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1353 AMDGPU::sub1, Src0SubRC);
1355 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
1356 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
1357 .addOperand(SrcReg0Sub1);
1359 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
1360 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
1362 .addImm(AMDGPU::sub0)
1364 .addImm(AMDGPU::sub1);
1366 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
1368 // Try to legalize the operands in case we need to swap the order to keep it
1370 Worklist.push_back(LoHalf);
1371 Worklist.push_back(HiHalf);
1374 void SIInstrInfo::splitScalar64BitBinaryOp(
1375 SmallVectorImpl<MachineInstr *> &Worklist,
1377 unsigned Opcode) const {
1378 MachineBasicBlock &MBB = *Inst->getParent();
1379 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1381 MachineOperand &Dest = Inst->getOperand(0);
1382 MachineOperand &Src0 = Inst->getOperand(1);
1383 MachineOperand &Src1 = Inst->getOperand(2);
1384 DebugLoc DL = Inst->getDebugLoc();
1386 MachineBasicBlock::iterator MII = Inst;
1388 const MCInstrDesc &InstDesc = get(Opcode);
1389 const TargetRegisterClass *Src0RC = Src0.isReg() ?
1390 MRI.getRegClass(Src0.getReg()) :
1391 &AMDGPU::SGPR_32RegClass;
1393 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
1394 const TargetRegisterClass *Src1RC = Src1.isReg() ?
1395 MRI.getRegClass(Src1.getReg()) :
1396 &AMDGPU::SGPR_32RegClass;
1398 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
1400 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1401 AMDGPU::sub0, Src0SubRC);
1402 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
1403 AMDGPU::sub0, Src1SubRC);
1405 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
1406 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
1408 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
1409 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
1410 .addOperand(SrcReg0Sub0)
1411 .addOperand(SrcReg1Sub0);
1413 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1414 AMDGPU::sub1, Src0SubRC);
1415 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
1416 AMDGPU::sub1, Src1SubRC);
1418 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
1419 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
1420 .addOperand(SrcReg0Sub1)
1421 .addOperand(SrcReg1Sub1);
1423 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
1424 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
1426 .addImm(AMDGPU::sub0)
1428 .addImm(AMDGPU::sub1);
1430 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
1432 // Try to legalize the operands in case we need to swap the order to keep it
1434 Worklist.push_back(LoHalf);
1435 Worklist.push_back(HiHalf);
1438 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
1439 MachineInstr *Inst) const {
1440 MachineBasicBlock &MBB = *Inst->getParent();
1441 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1443 MachineBasicBlock::iterator MII = Inst;
1444 DebugLoc DL = Inst->getDebugLoc();
1446 MachineOperand &Dest = Inst->getOperand(0);
1447 MachineOperand &Src = Inst->getOperand(1);
1449 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e32);
1450 const TargetRegisterClass *SrcRC = Src.isReg() ?
1451 MRI.getRegClass(Src.getReg()) :
1452 &AMDGPU::SGPR_32RegClass;
1454 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1455 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1457 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
1459 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
1460 AMDGPU::sub0, SrcSubRC);
1461 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
1462 AMDGPU::sub1, SrcSubRC);
1464 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
1465 .addOperand(SrcRegSub0)
1468 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
1469 .addOperand(SrcRegSub1)
1472 MRI.replaceRegWith(Dest.getReg(), ResultReg);
1474 Worklist.push_back(First);
1475 Worklist.push_back(Second);
1478 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
1479 MachineInstr *Inst) const {
1480 // Add the implict and explicit register definitions.
1481 if (NewDesc.ImplicitUses) {
1482 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
1483 unsigned Reg = NewDesc.ImplicitUses[i];
1484 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
1488 if (NewDesc.ImplicitDefs) {
1489 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
1490 unsigned Reg = NewDesc.ImplicitDefs[i];
1491 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
1496 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
1497 MachineBasicBlock *MBB,
1498 MachineBasicBlock::iterator I,
1500 unsigned Address, unsigned OffsetReg) const {
1501 const DebugLoc &DL = MBB->findDebugLoc(I);
1502 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
1503 getIndirectIndexBegin(*MBB->getParent()));
1505 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
1506 .addReg(IndirectBaseReg, RegState::Define)
1507 .addOperand(I->getOperand(0))
1508 .addReg(IndirectBaseReg)
1514 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
1515 MachineBasicBlock *MBB,
1516 MachineBasicBlock::iterator I,
1518 unsigned Address, unsigned OffsetReg) const {
1519 const DebugLoc &DL = MBB->findDebugLoc(I);
1520 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
1521 getIndirectIndexBegin(*MBB->getParent()));
1523 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
1524 .addOperand(I->getOperand(0))
1525 .addOperand(I->getOperand(1))
1526 .addReg(IndirectBaseReg)
1532 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
1533 const MachineFunction &MF) const {
1534 int End = getIndirectIndexEnd(MF);
1535 int Begin = getIndirectIndexBegin(MF);
1541 for (int Index = Begin; Index <= End; ++Index)
1542 Reserved.set(AMDGPU::VReg_32RegClass.getRegister(Index));
1544 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
1545 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
1547 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
1548 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
1550 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
1551 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
1553 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
1554 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
1556 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
1557 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));