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/MC/MCInstrDesc.h"
26 SIInstrInfo::SIInstrInfo(AMDGPUTargetMachine &tm)
27 : AMDGPUInstrInfo(tm),
30 //===----------------------------------------------------------------------===//
31 // TargetInstrInfo callbacks
32 //===----------------------------------------------------------------------===//
35 SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
36 MachineBasicBlock::iterator MI, DebugLoc DL,
37 unsigned DestReg, unsigned SrcReg,
40 // If we are trying to copy to or from SCC, there is a bug somewhere else in
41 // the backend. While it may be theoretically possible to do this, it should
42 // never be necessary.
43 assert(DestReg != AMDGPU::SCC && SrcReg != AMDGPU::SCC);
45 static const int16_t Sub0_15[] = {
46 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
47 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
48 AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
49 AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15, 0
52 static const int16_t Sub0_7[] = {
53 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
54 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7, 0
57 static const int16_t Sub0_3[] = {
58 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
61 static const int16_t Sub0_2[] = {
62 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
65 static const int16_t Sub0_1[] = {
66 AMDGPU::sub0, AMDGPU::sub1, 0
70 const int16_t *SubIndices;
72 if (AMDGPU::M0 == DestReg) {
73 // Check if M0 isn't already set to this value
74 for (MachineBasicBlock::reverse_iterator E = MBB.rend(),
75 I = MachineBasicBlock::reverse_iterator(MI); I != E; ++I) {
77 if (!I->definesRegister(AMDGPU::M0))
80 unsigned Opc = I->getOpcode();
81 if (Opc != TargetOpcode::COPY && Opc != AMDGPU::S_MOV_B32)
84 if (!I->readsRegister(SrcReg))
87 // The copy isn't necessary
92 if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
93 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
94 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
95 .addReg(SrcReg, getKillRegState(KillSrc));
98 } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
99 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
100 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
101 .addReg(SrcReg, getKillRegState(KillSrc));
104 } else if (AMDGPU::SReg_128RegClass.contains(DestReg)) {
105 assert(AMDGPU::SReg_128RegClass.contains(SrcReg));
106 Opcode = AMDGPU::S_MOV_B32;
109 } else if (AMDGPU::SReg_256RegClass.contains(DestReg)) {
110 assert(AMDGPU::SReg_256RegClass.contains(SrcReg));
111 Opcode = AMDGPU::S_MOV_B32;
114 } else if (AMDGPU::SReg_512RegClass.contains(DestReg)) {
115 assert(AMDGPU::SReg_512RegClass.contains(SrcReg));
116 Opcode = AMDGPU::S_MOV_B32;
117 SubIndices = Sub0_15;
119 } else if (AMDGPU::VReg_32RegClass.contains(DestReg)) {
120 assert(AMDGPU::VReg_32RegClass.contains(SrcReg) ||
121 AMDGPU::SReg_32RegClass.contains(SrcReg));
122 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
123 .addReg(SrcReg, getKillRegState(KillSrc));
126 } else if (AMDGPU::VReg_64RegClass.contains(DestReg)) {
127 assert(AMDGPU::VReg_64RegClass.contains(SrcReg) ||
128 AMDGPU::SReg_64RegClass.contains(SrcReg));
129 Opcode = AMDGPU::V_MOV_B32_e32;
132 } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
133 assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
134 Opcode = AMDGPU::V_MOV_B32_e32;
137 } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
138 assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
139 AMDGPU::SReg_128RegClass.contains(SrcReg));
140 Opcode = AMDGPU::V_MOV_B32_e32;
143 } else if (AMDGPU::VReg_256RegClass.contains(DestReg)) {
144 assert(AMDGPU::VReg_256RegClass.contains(SrcReg) ||
145 AMDGPU::SReg_256RegClass.contains(SrcReg));
146 Opcode = AMDGPU::V_MOV_B32_e32;
149 } else if (AMDGPU::VReg_512RegClass.contains(DestReg)) {
150 assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
151 AMDGPU::SReg_512RegClass.contains(SrcReg));
152 Opcode = AMDGPU::V_MOV_B32_e32;
153 SubIndices = Sub0_15;
156 llvm_unreachable("Can't copy register!");
159 while (unsigned SubIdx = *SubIndices++) {
160 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
161 get(Opcode), RI.getSubReg(DestReg, SubIdx));
163 Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
166 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
170 unsigned SIInstrInfo::commuteOpcode(unsigned Opcode) const {
173 // Try to map original to commuted opcode
174 if ((NewOpc = AMDGPU::getCommuteRev(Opcode)) != -1)
177 // Try to map commuted to original opcode
178 if ((NewOpc = AMDGPU::getCommuteOrig(Opcode)) != -1)
184 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
185 MachineBasicBlock::iterator MI,
186 unsigned SrcReg, bool isKill,
188 const TargetRegisterClass *RC,
189 const TargetRegisterInfo *TRI) const {
190 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
191 SIMachineFunctionInfo *MFI = MBB.getParent()->getInfo<SIMachineFunctionInfo>();
192 DebugLoc DL = MBB.findDebugLoc(MI);
193 unsigned KillFlag = isKill ? RegState::Kill : 0;
195 if (TRI->getCommonSubClass(RC, &AMDGPU::SGPR_32RegClass)) {
196 unsigned Lane = MFI->SpillTracker.getNextLane(MRI);
197 BuildMI(MBB, MI, DL, get(AMDGPU::V_WRITELANE_B32),
198 MFI->SpillTracker.LaneVGPR)
199 .addReg(SrcReg, KillFlag)
201 MFI->SpillTracker.addSpilledReg(FrameIndex, MFI->SpillTracker.LaneVGPR,
204 for (unsigned i = 0, e = RC->getSize() / 4; i != e; ++i) {
205 unsigned SubReg = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
206 BuildMI(MBB, MI, MBB.findDebugLoc(MI), get(AMDGPU::COPY), SubReg)
207 .addReg(SrcReg, 0, RI.getSubRegFromChannel(i));
208 storeRegToStackSlot(MBB, MI, SubReg, isKill, FrameIndex + i,
209 &AMDGPU::SReg_32RegClass, TRI);
214 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
215 MachineBasicBlock::iterator MI,
216 unsigned DestReg, int FrameIndex,
217 const TargetRegisterClass *RC,
218 const TargetRegisterInfo *TRI) const {
219 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
220 SIMachineFunctionInfo *MFI = MBB.getParent()->getInfo<SIMachineFunctionInfo>();
221 DebugLoc DL = MBB.findDebugLoc(MI);
222 if (TRI->getCommonSubClass(RC, &AMDGPU::SReg_32RegClass)) {
223 SIMachineFunctionInfo::SpilledReg Spill =
224 MFI->SpillTracker.getSpilledReg(FrameIndex);
226 BuildMI(MBB, MI, DL, get(AMDGPU::V_READLANE_B32), DestReg)
230 for (unsigned i = 0, e = RC->getSize() / 4; i != e; ++i) {
231 unsigned Flags = RegState::Define;
233 Flags |= RegState::Undef;
235 unsigned SubReg = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
236 loadRegFromStackSlot(MBB, MI, SubReg, FrameIndex + i,
237 &AMDGPU::SReg_32RegClass, TRI);
238 BuildMI(MBB, MI, DL, get(AMDGPU::COPY))
239 .addReg(DestReg, Flags, RI.getSubRegFromChannel(i))
245 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
248 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
249 if (MI->getNumOperands() < 3 || !MI->getOperand(1).isReg())
252 // Cannot commute VOP2 if src0 is SGPR.
253 if (isVOP2(MI->getOpcode()) && MI->getOperand(1).isReg() &&
254 RI.isSGPRClass(MRI.getRegClass(MI->getOperand(1).getReg())))
257 if (!MI->getOperand(2).isReg()) {
258 // XXX: Commute instructions with FPImm operands
259 if (NewMI || MI->getOperand(2).isFPImm() ||
260 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
264 // XXX: Commute VOP3 instructions with abs and neg set.
265 if (isVOP3(MI->getOpcode()) &&
266 (MI->getOperand(AMDGPU::getNamedOperandIdx(MI->getOpcode(),
267 AMDGPU::OpName::abs)).getImm() ||
268 MI->getOperand(AMDGPU::getNamedOperandIdx(MI->getOpcode(),
269 AMDGPU::OpName::neg)).getImm()))
272 unsigned Reg = MI->getOperand(1).getReg();
273 unsigned SubReg = MI->getOperand(1).getSubReg();
274 MI->getOperand(1).ChangeToImmediate(MI->getOperand(2).getImm());
275 MI->getOperand(2).ChangeToRegister(Reg, false);
276 MI->getOperand(2).setSubReg(SubReg);
278 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
282 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
287 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
288 MachineBasicBlock::iterator I,
290 unsigned SrcReg) const {
291 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
292 DstReg) .addReg(SrcReg);
295 bool SIInstrInfo::isMov(unsigned Opcode) const {
297 default: return false;
298 case AMDGPU::S_MOV_B32:
299 case AMDGPU::S_MOV_B64:
300 case AMDGPU::V_MOV_B32_e32:
301 case AMDGPU::V_MOV_B32_e64:
307 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
308 return RC != &AMDGPU::EXECRegRegClass;
313 // Helper function generated by tablegen. We are wrapping this with
314 // an SIInstrInfo function that reutrns bool rather than int.
315 int isDS(uint16_t Opcode);
319 bool SIInstrInfo::isDS(uint16_t Opcode) const {
320 return ::AMDGPU::isDS(Opcode) != -1;
323 int SIInstrInfo::isMIMG(uint16_t Opcode) const {
324 return get(Opcode).TSFlags & SIInstrFlags::MIMG;
327 int SIInstrInfo::isSMRD(uint16_t Opcode) const {
328 return get(Opcode).TSFlags & SIInstrFlags::SMRD;
331 bool SIInstrInfo::isVOP1(uint16_t Opcode) const {
332 return get(Opcode).TSFlags & SIInstrFlags::VOP1;
335 bool SIInstrInfo::isVOP2(uint16_t Opcode) const {
336 return get(Opcode).TSFlags & SIInstrFlags::VOP2;
339 bool SIInstrInfo::isVOP3(uint16_t Opcode) const {
340 return get(Opcode).TSFlags & SIInstrFlags::VOP3;
343 bool SIInstrInfo::isVOPC(uint16_t Opcode) const {
344 return get(Opcode).TSFlags & SIInstrFlags::VOPC;
347 bool SIInstrInfo::isSALUInstr(const MachineInstr &MI) const {
348 return get(MI.getOpcode()).TSFlags & SIInstrFlags::SALU;
351 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO) const {
353 return MO.getImm() >= -16 && MO.getImm() <= 64;
356 return MO.getFPImm()->isExactlyValue(0.0) ||
357 MO.getFPImm()->isExactlyValue(0.5) ||
358 MO.getFPImm()->isExactlyValue(-0.5) ||
359 MO.getFPImm()->isExactlyValue(1.0) ||
360 MO.getFPImm()->isExactlyValue(-1.0) ||
361 MO.getFPImm()->isExactlyValue(2.0) ||
362 MO.getFPImm()->isExactlyValue(-2.0) ||
363 MO.getFPImm()->isExactlyValue(4.0) ||
364 MO.getFPImm()->isExactlyValue(-4.0);
369 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO) const {
370 return (MO.isImm() || MO.isFPImm()) && !isInlineConstant(MO);
373 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
374 StringRef &ErrInfo) const {
375 uint16_t Opcode = MI->getOpcode();
376 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
377 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
378 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
380 // Make sure the number of operands is correct.
381 const MCInstrDesc &Desc = get(Opcode);
382 if (!Desc.isVariadic() &&
383 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
384 ErrInfo = "Instruction has wrong number of operands.";
388 // Make sure the register classes are correct
389 for (unsigned i = 0, e = Desc.getNumOperands(); i != e; ++i) {
390 switch (Desc.OpInfo[i].OperandType) {
391 case MCOI::OPERAND_REGISTER:
393 case MCOI::OPERAND_IMMEDIATE:
394 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFPImm()) {
395 ErrInfo = "Expected immediate, but got non-immediate";
403 if (!MI->getOperand(i).isReg())
406 int RegClass = Desc.OpInfo[i].RegClass;
407 if (RegClass != -1) {
408 unsigned Reg = MI->getOperand(i).getReg();
409 if (TargetRegisterInfo::isVirtualRegister(Reg))
412 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
413 if (!RC->contains(Reg)) {
414 ErrInfo = "Operand has incorrect register class.";
422 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
423 unsigned ConstantBusCount = 0;
424 unsigned SGPRUsed = AMDGPU::NoRegister;
425 for (int i = 0, e = MI->getNumOperands(); i != e; ++i) {
426 const MachineOperand &MO = MI->getOperand(i);
427 if (MO.isReg() && MO.isUse() &&
428 !TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
430 // EXEC register uses the constant bus.
431 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
434 // SGPRs use the constant bus
435 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
437 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
438 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
439 if (SGPRUsed != MO.getReg()) {
441 SGPRUsed = MO.getReg();
445 // Literal constants use the constant bus.
446 if (isLiteralConstant(MO))
449 if (ConstantBusCount > 1) {
450 ErrInfo = "VOP* instruction uses the constant bus more than once";
455 // Verify SRC1 for VOP2 and VOPC
456 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
457 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
458 if (Src1.isImm() || Src1.isFPImm()) {
459 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
465 if (isVOP3(Opcode)) {
466 if (Src0Idx != -1 && isLiteralConstant(MI->getOperand(Src0Idx))) {
467 ErrInfo = "VOP3 src0 cannot be a literal constant.";
470 if (Src1Idx != -1 && isLiteralConstant(MI->getOperand(Src1Idx))) {
471 ErrInfo = "VOP3 src1 cannot be a literal constant.";
474 if (Src2Idx != -1 && isLiteralConstant(MI->getOperand(Src2Idx))) {
475 ErrInfo = "VOP3 src2 cannot be a literal constant.";
482 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
483 switch (MI.getOpcode()) {
484 default: return AMDGPU::INSTRUCTION_LIST_END;
485 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
486 case AMDGPU::COPY: return AMDGPU::COPY;
487 case AMDGPU::PHI: return AMDGPU::PHI;
488 case AMDGPU::S_ADD_I32: return AMDGPU::V_ADD_I32_e32;
489 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
490 case AMDGPU::S_SUB_I32: return AMDGPU::V_SUB_I32_e32;
491 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
492 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
493 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
494 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
495 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
496 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
497 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
501 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
502 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
505 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
506 unsigned OpNo) const {
507 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
508 const MCInstrDesc &Desc = get(MI.getOpcode());
509 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
510 Desc.OpInfo[OpNo].RegClass == -1)
511 return MRI.getRegClass(MI.getOperand(OpNo).getReg());
513 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
514 return RI.getRegClass(RCID);
517 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
518 switch (MI.getOpcode()) {
520 case AMDGPU::REG_SEQUENCE:
521 return RI.hasVGPRs(getOpRegClass(MI, 0));
523 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
527 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
528 MachineBasicBlock::iterator I = MI;
529 MachineOperand &MO = MI->getOperand(OpIdx);
530 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
531 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
532 const TargetRegisterClass *RC = RI.getRegClass(RCID);
533 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
535 Opcode = AMDGPU::COPY;
536 } else if (RI.isSGPRClass(RC)) {
537 Opcode = AMDGPU::S_MOV_B32;
540 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
541 unsigned Reg = MRI.createVirtualRegister(VRC);
542 BuildMI(*MI->getParent(), I, MI->getParent()->findDebugLoc(I), get(Opcode),
544 MO.ChangeToRegister(Reg, false);
547 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
548 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
549 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
550 AMDGPU::OpName::src0);
551 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
552 AMDGPU::OpName::src1);
553 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
554 AMDGPU::OpName::src2);
557 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
558 MachineOperand &Src0 = MI->getOperand(Src0Idx);
559 MachineOperand &Src1 = MI->getOperand(Src1Idx);
561 // If the instruction implicitly reads VCC, we can't have any SGPR operands,
563 bool ReadsVCC = MI->readsRegister(AMDGPU::VCC, &RI);
564 if (ReadsVCC && Src0.isReg() &&
565 RI.isSGPRClass(MRI.getRegClass(Src0.getReg()))) {
566 legalizeOpWithMove(MI, Src0Idx);
570 if (ReadsVCC && Src1.isReg() &&
571 RI.isSGPRClass(MRI.getRegClass(Src1.getReg()))) {
572 legalizeOpWithMove(MI, Src1Idx);
576 // Legalize VOP2 instructions where src1 is not a VGPR. An SGPR input must
577 // be the first operand, and there can only be one.
578 if (Src1.isImm() || Src1.isFPImm() ||
579 (Src1.isReg() && RI.isSGPRClass(MRI.getRegClass(Src1.getReg())))) {
580 if (MI->isCommutable()) {
581 if (commuteInstruction(MI))
584 legalizeOpWithMove(MI, Src1Idx);
588 // XXX - Do any VOP3 instructions read VCC?
590 if (isVOP3(MI->getOpcode())) {
591 int VOP3Idx[3] = {Src0Idx, Src1Idx, Src2Idx};
592 unsigned SGPRReg = AMDGPU::NoRegister;
593 for (unsigned i = 0; i < 3; ++i) {
594 int Idx = VOP3Idx[i];
597 MachineOperand &MO = MI->getOperand(Idx);
600 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
601 continue; // VGPRs are legal
603 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
605 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
606 SGPRReg = MO.getReg();
607 // We can use one SGPR in each VOP3 instruction.
610 } else if (!isLiteralConstant(MO)) {
611 // If it is not a register and not a literal constant, then it must be
612 // an inline constant which is always legal.
615 // If we make it this far, then the operand is not legal and we must
617 legalizeOpWithMove(MI, Idx);
621 // Legalize REG_SEQUENCE
622 // The register class of the operands much be the same type as the register
623 // class of the output.
624 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
625 const TargetRegisterClass *RC = NULL, *SRC = NULL, *VRC = NULL;
626 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
627 if (!MI->getOperand(i).isReg() ||
628 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
630 const TargetRegisterClass *OpRC =
631 MRI.getRegClass(MI->getOperand(i).getReg());
632 if (RI.hasVGPRs(OpRC)) {
639 // If any of the operands are VGPR registers, then they all most be
640 // otherwise we will create illegal VGPR->SGPR copies when legalizing
642 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
645 VRC = RI.getEquivalentVGPRClass(SRC);
652 // Update all the operands so they have the same type.
653 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
654 if (!MI->getOperand(i).isReg() ||
655 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
657 unsigned DstReg = MRI.createVirtualRegister(RC);
658 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
659 get(AMDGPU::COPY), DstReg)
660 .addOperand(MI->getOperand(i));
661 MI->getOperand(i).setReg(DstReg);
666 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
667 SmallVector<MachineInstr *, 128> Worklist;
668 Worklist.push_back(&TopInst);
670 while (!Worklist.empty()) {
671 MachineInstr *Inst = Worklist.pop_back_val();
672 unsigned NewOpcode = getVALUOp(*Inst);
673 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END)
676 MachineRegisterInfo &MRI = Inst->getParent()->getParent()->getRegInfo();
678 // Use the new VALU Opcode.
679 const MCInstrDesc &NewDesc = get(NewOpcode);
680 Inst->setDesc(NewDesc);
682 // Remove any references to SCC. Vector instructions can't read from it, and
683 // We're just about to add the implicit use / defs of VCC, and we don't want
685 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
686 MachineOperand &Op = Inst->getOperand(i);
687 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
688 Inst->RemoveOperand(i);
691 // Add the implict and explicit register definitions.
692 if (NewDesc.ImplicitUses) {
693 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
694 unsigned Reg = NewDesc.ImplicitUses[i];
695 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
699 if (NewDesc.ImplicitDefs) {
700 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
701 unsigned Reg = NewDesc.ImplicitDefs[i];
702 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
706 legalizeOperands(Inst);
708 // Update the destination register class.
709 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
711 switch (Inst->getOpcode()) {
712 // For target instructions, getOpRegClass just returns the virtual
713 // register class associated with the operand, so we need to find an
714 // equivalent VGPR register class in order to move the instruction to the
718 case AMDGPU::REG_SEQUENCE:
719 if (RI.hasVGPRs(NewDstRC))
721 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
729 unsigned DstReg = Inst->getOperand(0).getReg();
730 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
731 MRI.replaceRegWith(DstReg, NewDstReg);
733 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
734 E = MRI.use_end(); I != E; ++I) {
735 MachineInstr &UseMI = *I->getParent();
736 if (!canReadVGPR(UseMI, I.getOperandNo())) {
737 Worklist.push_back(&UseMI);
743 //===----------------------------------------------------------------------===//
744 // Indirect addressing callbacks
745 //===----------------------------------------------------------------------===//
747 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
748 unsigned Channel) const {
749 assert(Channel == 0);
753 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
754 return &AMDGPU::VReg_32RegClass;
757 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
758 MachineBasicBlock *MBB,
759 MachineBasicBlock::iterator I,
761 unsigned Address, unsigned OffsetReg) const {
762 const DebugLoc &DL = MBB->findDebugLoc(I);
763 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
764 getIndirectIndexBegin(*MBB->getParent()));
766 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
767 .addReg(IndirectBaseReg, RegState::Define)
768 .addOperand(I->getOperand(0))
769 .addReg(IndirectBaseReg)
775 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
776 MachineBasicBlock *MBB,
777 MachineBasicBlock::iterator I,
779 unsigned Address, unsigned OffsetReg) const {
780 const DebugLoc &DL = MBB->findDebugLoc(I);
781 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
782 getIndirectIndexBegin(*MBB->getParent()));
784 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
785 .addOperand(I->getOperand(0))
786 .addOperand(I->getOperand(1))
787 .addReg(IndirectBaseReg)
793 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
794 const MachineFunction &MF) const {
795 int End = getIndirectIndexEnd(MF);
796 int Begin = getIndirectIndexBegin(MF);
802 for (int Index = Begin; Index <= End; ++Index)
803 Reserved.set(AMDGPU::VReg_32RegClass.getRegister(Index));
805 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
806 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
808 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
809 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
811 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
812 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
814 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
815 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
817 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
818 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));