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 //===----------------------------------------------------------------------===//
35 static unsigned getNumOperandsNoGlue(SDNode *Node) {
36 unsigned N = Node->getNumOperands();
37 while (N && Node->getOperand(N - 1).getValueType() == MVT::Glue)
42 static SDValue findChainOperand(SDNode *Load) {
43 SDValue LastOp = Load->getOperand(getNumOperandsNoGlue(Load) - 1);
44 assert(LastOp.getValueType() == MVT::Other && "Chain missing from load node");
48 /// \brief Returns true if both nodes have the same value for the given
49 /// operand \p Op, or if both nodes do not have this operand.
50 static bool nodesHaveSameOperandValue(SDNode *N0, SDNode* N1, unsigned OpName) {
51 unsigned Opc0 = N0->getMachineOpcode();
52 unsigned Opc1 = N1->getMachineOpcode();
54 int Op0Idx = AMDGPU::getNamedOperandIdx(Opc0, OpName);
55 int Op1Idx = AMDGPU::getNamedOperandIdx(Opc1, OpName);
57 if (Op0Idx == -1 && Op1Idx == -1)
61 if ((Op0Idx == -1 && Op1Idx != -1) ||
62 (Op1Idx == -1 && Op0Idx != -1))
65 // getNamedOperandIdx returns the index for the MachineInstr's operands,
66 // which includes the result as the first operand. We are indexing into the
67 // MachineSDNode's operands, so we need to skip the result operand to get
72 return N0->getOperand(Op0Idx) == N0->getOperand(Op1Idx);
75 bool SIInstrInfo::areLoadsFromSameBasePtr(SDNode *Load0, SDNode *Load1,
77 int64_t &Offset1) const {
78 if (!Load0->isMachineOpcode() || !Load1->isMachineOpcode())
81 unsigned Opc0 = Load0->getMachineOpcode();
82 unsigned Opc1 = Load1->getMachineOpcode();
84 // Make sure both are actually loads.
85 if (!get(Opc0).mayLoad() || !get(Opc1).mayLoad())
88 if (isDS(Opc0) && isDS(Opc1)) {
89 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
91 // TODO: Also shouldn't see read2st
92 assert(Opc0 != AMDGPU::DS_READ2_B32 &&
93 Opc0 != AMDGPU::DS_READ2_B64 &&
94 Opc1 != AMDGPU::DS_READ2_B32 &&
95 Opc1 != AMDGPU::DS_READ2_B64);
98 if (Load0->getOperand(1) != Load1->getOperand(1))
102 if (findChainOperand(Load0) != findChainOperand(Load1))
105 Offset0 = cast<ConstantSDNode>(Load0->getOperand(2))->getZExtValue();
106 Offset1 = cast<ConstantSDNode>(Load1->getOperand(2))->getZExtValue();
110 if (isSMRD(Opc0) && isSMRD(Opc1)) {
111 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
114 if (Load0->getOperand(0) != Load1->getOperand(0))
118 if (findChainOperand(Load0) != findChainOperand(Load1))
121 Offset0 = cast<ConstantSDNode>(Load0->getOperand(1))->getZExtValue();
122 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue();
126 // MUBUF and MTBUF can access the same addresses.
127 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1))) {
129 // MUBUF and MTBUF have vaddr at different indices.
130 if (!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::soffset) ||
131 findChainOperand(Load0) != findChainOperand(Load1) ||
132 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::vaddr) ||
133 !nodesHaveSameOperandValue(Load1, Load1, AMDGPU::OpName::srsrc))
136 int OffIdx0 = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
137 int OffIdx1 = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
139 if (OffIdx0 == -1 || OffIdx1 == -1)
142 // getNamedOperandIdx returns the index for MachineInstrs. Since they
143 // inlcude the output in the operand list, but SDNodes don't, we need to
144 // subtract the index by one.
148 SDValue Off0 = Load0->getOperand(OffIdx0);
149 SDValue Off1 = Load1->getOperand(OffIdx1);
151 // The offset might be a FrameIndexSDNode.
152 if (!isa<ConstantSDNode>(Off0) || !isa<ConstantSDNode>(Off1))
155 Offset0 = cast<ConstantSDNode>(Off0)->getZExtValue();
156 Offset1 = cast<ConstantSDNode>(Off1)->getZExtValue();
163 bool SIInstrInfo::getLdStBaseRegImmOfs(MachineInstr *LdSt,
164 unsigned &BaseReg, unsigned &Offset,
165 const TargetRegisterInfo *TRI) const {
166 unsigned Opc = LdSt->getOpcode();
168 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
169 AMDGPU::OpName::offset);
171 // Normal, single offset LDS instruction.
172 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
173 AMDGPU::OpName::addr);
175 BaseReg = AddrReg->getReg();
176 Offset = OffsetImm->getImm();
180 // The 2 offset instructions use offset0 and offset1 instead. We can treat
181 // these as a load with a single offset if the 2 offsets are consecutive. We
182 // will use this for some partially aligned loads.
183 const MachineOperand *Offset0Imm = getNamedOperand(*LdSt,
184 AMDGPU::OpName::offset0);
185 const MachineOperand *Offset1Imm = getNamedOperand(*LdSt,
186 AMDGPU::OpName::offset1);
188 uint8_t Offset0 = Offset0Imm->getImm();
189 uint8_t Offset1 = Offset1Imm->getImm();
190 assert(Offset1 > Offset0);
192 if (Offset1 - Offset0 == 1) {
193 // Each of these offsets is in element sized units, so we need to convert
194 // to bytes of the individual reads.
198 EltSize = getOpRegClass(*LdSt, 0)->getSize() / 2;
200 assert(LdSt->mayStore());
201 int Data0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
202 EltSize = getOpRegClass(*LdSt, Data0Idx)->getSize();
205 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
206 AMDGPU::OpName::addr);
207 BaseReg = AddrReg->getReg();
208 Offset = EltSize * Offset0;
215 if (isMUBUF(Opc) || isMTBUF(Opc)) {
216 if (AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::soffset) != -1)
219 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
220 AMDGPU::OpName::vaddr);
224 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
225 AMDGPU::OpName::offset);
226 BaseReg = AddrReg->getReg();
227 Offset = OffsetImm->getImm();
232 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
233 AMDGPU::OpName::offset);
237 const MachineOperand *SBaseReg = getNamedOperand(*LdSt,
238 AMDGPU::OpName::sbase);
239 BaseReg = SBaseReg->getReg();
240 Offset = OffsetImm->getImm();
248 SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
249 MachineBasicBlock::iterator MI, DebugLoc DL,
250 unsigned DestReg, unsigned SrcReg,
251 bool KillSrc) const {
253 // If we are trying to copy to or from SCC, there is a bug somewhere else in
254 // the backend. While it may be theoretically possible to do this, it should
255 // never be necessary.
256 assert(DestReg != AMDGPU::SCC && SrcReg != AMDGPU::SCC);
258 static const int16_t Sub0_15[] = {
259 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
260 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
261 AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
262 AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15, 0
265 static const int16_t Sub0_7[] = {
266 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
267 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7, 0
270 static const int16_t Sub0_3[] = {
271 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
274 static const int16_t Sub0_2[] = {
275 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
278 static const int16_t Sub0_1[] = {
279 AMDGPU::sub0, AMDGPU::sub1, 0
283 const int16_t *SubIndices;
285 if (AMDGPU::M0 == DestReg) {
286 // Check if M0 isn't already set to this value
287 for (MachineBasicBlock::reverse_iterator E = MBB.rend(),
288 I = MachineBasicBlock::reverse_iterator(MI); I != E; ++I) {
290 if (!I->definesRegister(AMDGPU::M0))
293 unsigned Opc = I->getOpcode();
294 if (Opc != TargetOpcode::COPY && Opc != AMDGPU::S_MOV_B32)
297 if (!I->readsRegister(SrcReg))
300 // The copy isn't necessary
305 if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
306 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
307 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
308 .addReg(SrcReg, getKillRegState(KillSrc));
311 } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
312 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
313 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
314 .addReg(SrcReg, getKillRegState(KillSrc));
317 } else if (AMDGPU::SReg_128RegClass.contains(DestReg)) {
318 assert(AMDGPU::SReg_128RegClass.contains(SrcReg));
319 Opcode = AMDGPU::S_MOV_B32;
322 } else if (AMDGPU::SReg_256RegClass.contains(DestReg)) {
323 assert(AMDGPU::SReg_256RegClass.contains(SrcReg));
324 Opcode = AMDGPU::S_MOV_B32;
327 } else if (AMDGPU::SReg_512RegClass.contains(DestReg)) {
328 assert(AMDGPU::SReg_512RegClass.contains(SrcReg));
329 Opcode = AMDGPU::S_MOV_B32;
330 SubIndices = Sub0_15;
332 } else if (AMDGPU::VReg_32RegClass.contains(DestReg)) {
333 assert(AMDGPU::VReg_32RegClass.contains(SrcReg) ||
334 AMDGPU::SReg_32RegClass.contains(SrcReg));
335 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
336 .addReg(SrcReg, getKillRegState(KillSrc));
339 } else if (AMDGPU::VReg_64RegClass.contains(DestReg)) {
340 assert(AMDGPU::VReg_64RegClass.contains(SrcReg) ||
341 AMDGPU::SReg_64RegClass.contains(SrcReg));
342 Opcode = AMDGPU::V_MOV_B32_e32;
345 } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
346 assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
347 Opcode = AMDGPU::V_MOV_B32_e32;
350 } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
351 assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
352 AMDGPU::SReg_128RegClass.contains(SrcReg));
353 Opcode = AMDGPU::V_MOV_B32_e32;
356 } else if (AMDGPU::VReg_256RegClass.contains(DestReg)) {
357 assert(AMDGPU::VReg_256RegClass.contains(SrcReg) ||
358 AMDGPU::SReg_256RegClass.contains(SrcReg));
359 Opcode = AMDGPU::V_MOV_B32_e32;
362 } else if (AMDGPU::VReg_512RegClass.contains(DestReg)) {
363 assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
364 AMDGPU::SReg_512RegClass.contains(SrcReg));
365 Opcode = AMDGPU::V_MOV_B32_e32;
366 SubIndices = Sub0_15;
369 llvm_unreachable("Can't copy register!");
372 while (unsigned SubIdx = *SubIndices++) {
373 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
374 get(Opcode), RI.getSubReg(DestReg, SubIdx));
376 Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
379 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
383 unsigned SIInstrInfo::commuteOpcode(unsigned Opcode) const {
386 // Try to map original to commuted opcode
387 if ((NewOpc = AMDGPU::getCommuteRev(Opcode)) != -1)
390 // Try to map commuted to original opcode
391 if ((NewOpc = AMDGPU::getCommuteOrig(Opcode)) != -1)
397 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
398 MachineBasicBlock::iterator MI,
399 unsigned SrcReg, bool isKill,
401 const TargetRegisterClass *RC,
402 const TargetRegisterInfo *TRI) const {
403 MachineFunction *MF = MBB.getParent();
404 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
405 MachineRegisterInfo &MRI = MF->getRegInfo();
406 DebugLoc DL = MBB.findDebugLoc(MI);
407 unsigned KillFlag = isKill ? RegState::Kill : 0;
409 if (RI.hasVGPRs(RC)) {
410 LLVMContext &Ctx = MF->getFunction()->getContext();
411 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Can't spill VGPR!");
412 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), AMDGPU::VGPR0)
414 } else if (TRI->getCommonSubClass(RC, &AMDGPU::SGPR_32RegClass)) {
415 unsigned Lane = MFI->SpillTracker.reserveLanes(MRI, MF);
416 unsigned TgtReg = MFI->SpillTracker.LaneVGPR;
418 BuildMI(MBB, MI, DL, get(AMDGPU::V_WRITELANE_B32), TgtReg)
419 .addReg(SrcReg, KillFlag)
421 MFI->SpillTracker.addSpilledReg(FrameIndex, TgtReg, Lane);
422 } else if (RI.isSGPRClass(RC)) {
423 // We are only allowed to create one new instruction when spilling
424 // registers, so we need to use pseudo instruction for vector
427 // Reserve a spot in the spill tracker for each sub-register of
428 // the vector register.
429 unsigned NumSubRegs = RC->getSize() / 4;
430 unsigned FirstLane = MFI->SpillTracker.reserveLanes(MRI, MF, NumSubRegs);
431 MFI->SpillTracker.addSpilledReg(FrameIndex, MFI->SpillTracker.LaneVGPR,
435 switch (RC->getSize() * 8) {
436 case 64: Opcode = AMDGPU::SI_SPILL_S64_SAVE; break;
437 case 128: Opcode = AMDGPU::SI_SPILL_S128_SAVE; break;
438 case 256: Opcode = AMDGPU::SI_SPILL_S256_SAVE; break;
439 case 512: Opcode = AMDGPU::SI_SPILL_S512_SAVE; break;
440 default: llvm_unreachable("Cannot spill register class");
443 BuildMI(MBB, MI, DL, get(Opcode), MFI->SpillTracker.LaneVGPR)
447 llvm_unreachable("VGPR spilling not supported");
451 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
452 MachineBasicBlock::iterator MI,
453 unsigned DestReg, int FrameIndex,
454 const TargetRegisterClass *RC,
455 const TargetRegisterInfo *TRI) const {
456 MachineFunction *MF = MBB.getParent();
457 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
458 DebugLoc DL = MBB.findDebugLoc(MI);
460 if (RI.hasVGPRs(RC)) {
461 LLVMContext &Ctx = MF->getFunction()->getContext();
462 Ctx.emitError("SIInstrInfo::loadRegToStackSlot - Can't retrieve spilled VGPR!");
463 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
465 } else if (RI.isSGPRClass(RC)){
467 switch(RC->getSize() * 8) {
468 case 32: Opcode = AMDGPU::SI_SPILL_S32_RESTORE; break;
469 case 64: Opcode = AMDGPU::SI_SPILL_S64_RESTORE; break;
470 case 128: Opcode = AMDGPU::SI_SPILL_S128_RESTORE; break;
471 case 256: Opcode = AMDGPU::SI_SPILL_S256_RESTORE; break;
472 case 512: Opcode = AMDGPU::SI_SPILL_S512_RESTORE; break;
473 default: llvm_unreachable("Cannot spill register class");
476 SIMachineFunctionInfo::SpilledReg Spill =
477 MFI->SpillTracker.getSpilledReg(FrameIndex);
479 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
483 llvm_unreachable("VGPR spilling not supported");
487 static unsigned getNumSubRegsForSpillOp(unsigned Op) {
490 case AMDGPU::SI_SPILL_S512_SAVE:
491 case AMDGPU::SI_SPILL_S512_RESTORE:
493 case AMDGPU::SI_SPILL_S256_SAVE:
494 case AMDGPU::SI_SPILL_S256_RESTORE:
496 case AMDGPU::SI_SPILL_S128_SAVE:
497 case AMDGPU::SI_SPILL_S128_RESTORE:
499 case AMDGPU::SI_SPILL_S64_SAVE:
500 case AMDGPU::SI_SPILL_S64_RESTORE:
502 case AMDGPU::SI_SPILL_S32_RESTORE:
504 default: llvm_unreachable("Invalid spill opcode");
508 void SIInstrInfo::insertNOPs(MachineBasicBlock::iterator MI,
517 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(AMDGPU::S_NOP))
522 bool SIInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
523 SIMachineFunctionInfo *MFI =
524 MI->getParent()->getParent()->getInfo<SIMachineFunctionInfo>();
525 MachineBasicBlock &MBB = *MI->getParent();
526 DebugLoc DL = MBB.findDebugLoc(MI);
527 switch (MI->getOpcode()) {
528 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
530 // SGPR register spill
531 case AMDGPU::SI_SPILL_S512_SAVE:
532 case AMDGPU::SI_SPILL_S256_SAVE:
533 case AMDGPU::SI_SPILL_S128_SAVE:
534 case AMDGPU::SI_SPILL_S64_SAVE: {
535 unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
536 unsigned FrameIndex = MI->getOperand(2).getImm();
538 for (unsigned i = 0, e = NumSubRegs; i < e; ++i) {
539 SIMachineFunctionInfo::SpilledReg Spill;
540 unsigned SubReg = RI.getPhysRegSubReg(MI->getOperand(1).getReg(),
541 &AMDGPU::SGPR_32RegClass, i);
542 Spill = MFI->SpillTracker.getSpilledReg(FrameIndex);
544 BuildMI(MBB, MI, DL, get(AMDGPU::V_WRITELANE_B32),
545 MI->getOperand(0).getReg())
547 .addImm(Spill.Lane + i);
549 MI->eraseFromParent();
553 // SGPR register restore
554 case AMDGPU::SI_SPILL_S512_RESTORE:
555 case AMDGPU::SI_SPILL_S256_RESTORE:
556 case AMDGPU::SI_SPILL_S128_RESTORE:
557 case AMDGPU::SI_SPILL_S64_RESTORE:
558 case AMDGPU::SI_SPILL_S32_RESTORE: {
559 unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
561 for (unsigned i = 0, e = NumSubRegs; i < e; ++i) {
562 SIMachineFunctionInfo::SpilledReg Spill;
563 unsigned FrameIndex = MI->getOperand(2).getImm();
564 unsigned SubReg = RI.getPhysRegSubReg(MI->getOperand(0).getReg(),
565 &AMDGPU::SGPR_32RegClass, i);
566 Spill = MFI->SpillTracker.getSpilledReg(FrameIndex);
568 BuildMI(MBB, MI, DL, get(AMDGPU::V_READLANE_B32), SubReg)
569 .addReg(MI->getOperand(1).getReg())
570 .addImm(Spill.Lane + i);
573 MI->eraseFromParent();
576 case AMDGPU::SI_CONSTDATA_PTR: {
577 unsigned Reg = MI->getOperand(0).getReg();
578 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
579 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
581 BuildMI(MBB, MI, DL, get(AMDGPU::S_GETPC_B64), Reg);
583 // Add 32-bit offset from this instruction to the start of the constant data.
584 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADD_I32), RegLo)
586 .addTargetIndex(AMDGPU::TI_CONSTDATA_START)
587 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit);
588 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADDC_U32), RegHi)
591 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit)
592 .addReg(AMDGPU::SCC, RegState::Implicit);
593 MI->eraseFromParent();
600 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
603 if (MI->getNumOperands() < 3 || !MI->getOperand(1).isReg())
606 // Make sure it s legal to commute operands for VOP2.
607 if (isVOP2(MI->getOpcode()) &&
608 (!isOperandLegal(MI, 1, &MI->getOperand(2)) ||
609 !isOperandLegal(MI, 2, &MI->getOperand(1))))
612 if (!MI->getOperand(2).isReg()) {
613 // XXX: Commute instructions with FPImm operands
614 if (NewMI || MI->getOperand(2).isFPImm() ||
615 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
619 // XXX: Commute VOP3 instructions with abs and neg set .
620 const MachineOperand *Abs = getNamedOperand(*MI, AMDGPU::OpName::abs);
621 const MachineOperand *Neg = getNamedOperand(*MI, AMDGPU::OpName::neg);
622 const MachineOperand *Src0Mods = getNamedOperand(*MI,
623 AMDGPU::OpName::src0_modifiers);
624 const MachineOperand *Src1Mods = getNamedOperand(*MI,
625 AMDGPU::OpName::src1_modifiers);
626 const MachineOperand *Src2Mods = getNamedOperand(*MI,
627 AMDGPU::OpName::src2_modifiers);
629 if ((Abs && Abs->getImm()) || (Neg && Neg->getImm()) ||
630 (Src0Mods && Src0Mods->getImm()) || (Src1Mods && Src1Mods->getImm()) ||
631 (Src2Mods && Src2Mods->getImm()))
634 unsigned Reg = MI->getOperand(1).getReg();
635 unsigned SubReg = MI->getOperand(1).getSubReg();
636 MI->getOperand(1).ChangeToImmediate(MI->getOperand(2).getImm());
637 MI->getOperand(2).ChangeToRegister(Reg, false);
638 MI->getOperand(2).setSubReg(SubReg);
640 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
644 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
649 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
650 MachineBasicBlock::iterator I,
652 unsigned SrcReg) const {
653 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
654 DstReg) .addReg(SrcReg);
657 bool SIInstrInfo::isMov(unsigned Opcode) const {
659 default: return false;
660 case AMDGPU::S_MOV_B32:
661 case AMDGPU::S_MOV_B64:
662 case AMDGPU::V_MOV_B32_e32:
663 case AMDGPU::V_MOV_B32_e64:
669 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
670 return RC != &AMDGPU::EXECRegRegClass;
674 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
675 AliasAnalysis *AA) const {
676 switch(MI->getOpcode()) {
677 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
678 case AMDGPU::S_MOV_B32:
679 case AMDGPU::S_MOV_B64:
680 case AMDGPU::V_MOV_B32_e32:
681 return MI->getOperand(1).isImm();
687 // Helper function generated by tablegen. We are wrapping this with
688 // an SIInstrInfo function that returns bool rather than int.
689 int isDS(uint16_t Opcode);
693 bool SIInstrInfo::isDS(uint16_t Opcode) const {
694 return ::AMDGPU::isDS(Opcode) != -1;
697 bool SIInstrInfo::isMIMG(uint16_t Opcode) const {
698 return get(Opcode).TSFlags & SIInstrFlags::MIMG;
701 bool SIInstrInfo::isSMRD(uint16_t Opcode) const {
702 return get(Opcode).TSFlags & SIInstrFlags::SMRD;
705 bool SIInstrInfo::isMUBUF(uint16_t Opcode) const {
706 return get(Opcode).TSFlags & SIInstrFlags::MUBUF;
709 bool SIInstrInfo::isMTBUF(uint16_t Opcode) const {
710 return get(Opcode).TSFlags & SIInstrFlags::MTBUF;
713 bool SIInstrInfo::isVOP1(uint16_t Opcode) const {
714 return get(Opcode).TSFlags & SIInstrFlags::VOP1;
717 bool SIInstrInfo::isVOP2(uint16_t Opcode) const {
718 return get(Opcode).TSFlags & SIInstrFlags::VOP2;
721 bool SIInstrInfo::isVOP3(uint16_t Opcode) const {
722 return get(Opcode).TSFlags & SIInstrFlags::VOP3;
725 bool SIInstrInfo::isVOPC(uint16_t Opcode) const {
726 return get(Opcode).TSFlags & SIInstrFlags::VOPC;
729 bool SIInstrInfo::isSALUInstr(const MachineInstr &MI) const {
730 return get(MI.getOpcode()).TSFlags & SIInstrFlags::SALU;
733 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
734 int32_t Val = Imm.getSExtValue();
735 if (Val >= -16 && Val <= 64)
738 // The actual type of the operand does not seem to matter as long
739 // as the bits match one of the inline immediate values. For example:
741 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
742 // so it is a legal inline immediate.
744 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
745 // floating-point, so it is a legal inline immediate.
747 return (APInt::floatToBits(0.0f) == Imm) ||
748 (APInt::floatToBits(1.0f) == Imm) ||
749 (APInt::floatToBits(-1.0f) == Imm) ||
750 (APInt::floatToBits(0.5f) == Imm) ||
751 (APInt::floatToBits(-0.5f) == Imm) ||
752 (APInt::floatToBits(2.0f) == Imm) ||
753 (APInt::floatToBits(-2.0f) == Imm) ||
754 (APInt::floatToBits(4.0f) == Imm) ||
755 (APInt::floatToBits(-4.0f) == Imm);
758 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO) const {
760 return isInlineConstant(APInt(32, MO.getImm(), true));
763 APFloat FpImm = MO.getFPImm()->getValueAPF();
764 return isInlineConstant(FpImm.bitcastToAPInt());
770 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO) const {
771 return (MO.isImm() || MO.isFPImm()) && !isInlineConstant(MO);
774 static bool compareMachineOp(const MachineOperand &Op0,
775 const MachineOperand &Op1) {
776 if (Op0.getType() != Op1.getType())
779 switch (Op0.getType()) {
780 case MachineOperand::MO_Register:
781 return Op0.getReg() == Op1.getReg();
782 case MachineOperand::MO_Immediate:
783 return Op0.getImm() == Op1.getImm();
784 case MachineOperand::MO_FPImmediate:
785 return Op0.getFPImm() == Op1.getFPImm();
787 llvm_unreachable("Didn't expect to be comparing these operand types");
791 bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
792 const MachineOperand &MO) const {
793 const MCOperandInfo &OpInfo = get(MI->getOpcode()).OpInfo[OpNo];
795 assert(MO.isImm() || MO.isFPImm());
797 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
800 if (OpInfo.RegClass < 0)
803 return RI.regClassCanUseImmediate(OpInfo.RegClass);
806 bool SIInstrInfo::canFoldOffset(unsigned OffsetSize, unsigned AS) {
808 case AMDGPUAS::GLOBAL_ADDRESS: {
809 // MUBUF instructions a 12-bit offset in bytes.
810 return isUInt<12>(OffsetSize);
812 case AMDGPUAS::CONSTANT_ADDRESS: {
813 // SMRD instructions have an 8-bit offset in dwords.
814 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
816 case AMDGPUAS::LOCAL_ADDRESS:
817 case AMDGPUAS::REGION_ADDRESS: {
818 // The single offset versions have a 16-bit offset in bytes.
819 return isUInt<16>(OffsetSize);
821 case AMDGPUAS::PRIVATE_ADDRESS:
822 // Indirect register addressing does not use any offsets.
828 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
829 return AMDGPU::getVOPe32(Opcode) != -1;
832 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
833 // The src0_modifier operand is present on all instructions
834 // that have modifiers.
836 return AMDGPU::getNamedOperandIdx(Opcode,
837 AMDGPU::OpName::src0_modifiers) != -1;
840 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
841 StringRef &ErrInfo) const {
842 uint16_t Opcode = MI->getOpcode();
843 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
844 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
845 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
847 // Make sure the number of operands is correct.
848 const MCInstrDesc &Desc = get(Opcode);
849 if (!Desc.isVariadic() &&
850 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
851 ErrInfo = "Instruction has wrong number of operands.";
855 // Make sure the register classes are correct
856 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
857 switch (Desc.OpInfo[i].OperandType) {
858 case MCOI::OPERAND_REGISTER: {
859 int RegClass = Desc.OpInfo[i].RegClass;
860 if (!RI.regClassCanUseImmediate(RegClass) &&
861 (MI->getOperand(i).isImm() || MI->getOperand(i).isFPImm())) {
862 // Handle some special cases:
863 // Src0 can of VOP1, VOP2, VOPC can be an immediate no matter what
864 // the register class.
865 if (i != Src0Idx || (!isVOP1(Opcode) && !isVOP2(Opcode) &&
867 ErrInfo = "Expected register, but got immediate";
873 case MCOI::OPERAND_IMMEDIATE:
874 // Check if this operand is an immediate.
875 // FrameIndex operands will be replaced by immediates, so they are
877 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFPImm() &&
878 !MI->getOperand(i).isFI()) {
879 ErrInfo = "Expected immediate, but got non-immediate";
887 if (!MI->getOperand(i).isReg())
890 int RegClass = Desc.OpInfo[i].RegClass;
891 if (RegClass != -1) {
892 unsigned Reg = MI->getOperand(i).getReg();
893 if (TargetRegisterInfo::isVirtualRegister(Reg))
896 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
897 if (!RC->contains(Reg)) {
898 ErrInfo = "Operand has incorrect register class.";
906 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
907 unsigned ConstantBusCount = 0;
908 unsigned SGPRUsed = AMDGPU::NoRegister;
909 for (int i = 0, e = MI->getNumOperands(); i != e; ++i) {
910 const MachineOperand &MO = MI->getOperand(i);
911 if (MO.isReg() && MO.isUse() &&
912 !TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
914 // EXEC register uses the constant bus.
915 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
918 // SGPRs use the constant bus
919 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
921 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
922 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
923 if (SGPRUsed != MO.getReg()) {
925 SGPRUsed = MO.getReg();
929 // Literal constants use the constant bus.
930 if (isLiteralConstant(MO))
933 if (ConstantBusCount > 1) {
934 ErrInfo = "VOP* instruction uses the constant bus more than once";
939 // Verify SRC1 for VOP2 and VOPC
940 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
941 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
942 if (Src1.isImm() || Src1.isFPImm()) {
943 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
949 if (isVOP3(Opcode)) {
950 if (Src0Idx != -1 && isLiteralConstant(MI->getOperand(Src0Idx))) {
951 ErrInfo = "VOP3 src0 cannot be a literal constant.";
954 if (Src1Idx != -1 && isLiteralConstant(MI->getOperand(Src1Idx))) {
955 ErrInfo = "VOP3 src1 cannot be a literal constant.";
958 if (Src2Idx != -1 && isLiteralConstant(MI->getOperand(Src2Idx))) {
959 ErrInfo = "VOP3 src2 cannot be a literal constant.";
964 // Verify misc. restrictions on specific instructions.
965 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
966 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
969 const MachineOperand &Src0 = MI->getOperand(2);
970 const MachineOperand &Src1 = MI->getOperand(3);
971 const MachineOperand &Src2 = MI->getOperand(4);
972 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
973 if (!compareMachineOp(Src0, Src1) &&
974 !compareMachineOp(Src0, Src2)) {
975 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
984 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
985 switch (MI.getOpcode()) {
986 default: return AMDGPU::INSTRUCTION_LIST_END;
987 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
988 case AMDGPU::COPY: return AMDGPU::COPY;
989 case AMDGPU::PHI: return AMDGPU::PHI;
990 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
991 case AMDGPU::S_MOV_B32:
992 return MI.getOperand(1).isReg() ?
993 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
994 case AMDGPU::S_ADD_I32: return AMDGPU::V_ADD_I32_e32;
995 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
996 case AMDGPU::S_SUB_I32: return AMDGPU::V_SUB_I32_e32;
997 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
998 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
999 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
1000 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
1001 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
1002 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
1003 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
1004 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
1005 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
1006 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
1007 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
1008 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
1009 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
1010 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
1011 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
1012 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
1013 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
1014 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
1015 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
1016 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
1017 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
1018 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
1019 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
1020 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
1021 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
1022 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
1023 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
1024 case AMDGPU::S_LOAD_DWORD_IMM:
1025 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1026 case AMDGPU::S_LOAD_DWORDX2_IMM:
1027 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1028 case AMDGPU::S_LOAD_DWORDX4_IMM:
1029 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1030 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e32;
1031 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
1032 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
1036 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
1037 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
1040 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
1041 unsigned OpNo) const {
1042 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1043 const MCInstrDesc &Desc = get(MI.getOpcode());
1044 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
1045 Desc.OpInfo[OpNo].RegClass == -1)
1046 return MRI.getRegClass(MI.getOperand(OpNo).getReg());
1048 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
1049 return RI.getRegClass(RCID);
1052 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
1053 switch (MI.getOpcode()) {
1055 case AMDGPU::REG_SEQUENCE:
1057 case AMDGPU::INSERT_SUBREG:
1058 return RI.hasVGPRs(getOpRegClass(MI, 0));
1060 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
1064 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
1065 MachineBasicBlock::iterator I = MI;
1066 MachineOperand &MO = MI->getOperand(OpIdx);
1067 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1068 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
1069 const TargetRegisterClass *RC = RI.getRegClass(RCID);
1070 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1072 Opcode = AMDGPU::COPY;
1073 } else if (RI.isSGPRClass(RC)) {
1074 Opcode = AMDGPU::S_MOV_B32;
1077 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
1078 unsigned Reg = MRI.createVirtualRegister(VRC);
1079 BuildMI(*MI->getParent(), I, MI->getParent()->findDebugLoc(I), get(Opcode),
1080 Reg).addOperand(MO);
1081 MO.ChangeToRegister(Reg, false);
1084 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
1085 MachineRegisterInfo &MRI,
1086 MachineOperand &SuperReg,
1087 const TargetRegisterClass *SuperRC,
1089 const TargetRegisterClass *SubRC)
1091 assert(SuperReg.isReg());
1093 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
1094 unsigned SubReg = MRI.createVirtualRegister(SubRC);
1096 // Just in case the super register is itself a sub-register, copy it to a new
1097 // value so we don't need to worry about merging its subreg index with the
1098 // SubIdx passed to this function. The register coalescer should be able to
1099 // eliminate this extra copy.
1100 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1102 .addOperand(SuperReg);
1104 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1106 .addReg(NewSuperReg, 0, SubIdx);
1110 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
1111 MachineBasicBlock::iterator MII,
1112 MachineRegisterInfo &MRI,
1114 const TargetRegisterClass *SuperRC,
1116 const TargetRegisterClass *SubRC) const {
1118 // XXX - Is there a better way to do this?
1119 if (SubIdx == AMDGPU::sub0)
1120 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
1121 if (SubIdx == AMDGPU::sub1)
1122 return MachineOperand::CreateImm(Op.getImm() >> 32);
1124 llvm_unreachable("Unhandled register index for immediate");
1127 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
1129 return MachineOperand::CreateReg(SubReg, false);
1132 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
1133 MachineBasicBlock::iterator MI,
1134 MachineRegisterInfo &MRI,
1135 const TargetRegisterClass *RC,
1136 const MachineOperand &Op) const {
1137 MachineBasicBlock *MBB = MI->getParent();
1138 DebugLoc DL = MI->getDebugLoc();
1139 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1140 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1141 unsigned Dst = MRI.createVirtualRegister(RC);
1143 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1145 .addImm(Op.getImm() & 0xFFFFFFFF);
1146 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1148 .addImm(Op.getImm() >> 32);
1150 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
1152 .addImm(AMDGPU::sub0)
1154 .addImm(AMDGPU::sub1);
1156 Worklist.push_back(Lo);
1157 Worklist.push_back(Hi);
1162 bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
1163 const MachineOperand *MO) const {
1164 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1165 const MCInstrDesc &InstDesc = get(MI->getOpcode());
1166 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
1167 const TargetRegisterClass *DefinedRC =
1168 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
1170 MO = &MI->getOperand(OpIdx);
1174 const TargetRegisterClass *RC = MRI.getRegClass(MO->getReg());
1175 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass));
1179 // Handle non-register types that are treated like immediates.
1180 assert(MO->isImm() || MO->isFPImm() || MO->isTargetIndex() || MO->isFI());
1183 // This opperand expects an immediate
1186 return RI.regClassCanUseImmediate(DefinedRC);
1189 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
1190 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1192 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1193 AMDGPU::OpName::src0);
1194 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1195 AMDGPU::OpName::src1);
1196 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1197 AMDGPU::OpName::src2);
1200 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
1202 if (!isOperandLegal(MI, Src0Idx))
1203 legalizeOpWithMove(MI, Src0Idx);
1206 if (isOperandLegal(MI, Src1Idx))
1209 // Usually src0 of VOP2 instructions allow more types of inputs
1210 // than src1, so try to commute the instruction to decrease our
1211 // chances of having to insert a MOV instruction to legalize src1.
1212 if (MI->isCommutable()) {
1213 if (commuteInstruction(MI))
1214 // If we are successful in commuting, then we know MI is legal, so
1219 legalizeOpWithMove(MI, Src1Idx);
1223 // XXX - Do any VOP3 instructions read VCC?
1225 if (isVOP3(MI->getOpcode())) {
1226 int VOP3Idx[3] = {Src0Idx, Src1Idx, Src2Idx};
1227 unsigned SGPRReg = AMDGPU::NoRegister;
1228 for (unsigned i = 0; i < 3; ++i) {
1229 int Idx = VOP3Idx[i];
1232 MachineOperand &MO = MI->getOperand(Idx);
1235 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
1236 continue; // VGPRs are legal
1238 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
1240 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
1241 SGPRReg = MO.getReg();
1242 // We can use one SGPR in each VOP3 instruction.
1245 } else if (!isLiteralConstant(MO)) {
1246 // If it is not a register and not a literal constant, then it must be
1247 // an inline constant which is always legal.
1250 // If we make it this far, then the operand is not legal and we must
1252 legalizeOpWithMove(MI, Idx);
1256 // Legalize REG_SEQUENCE and PHI
1257 // The register class of the operands much be the same type as the register
1258 // class of the output.
1259 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
1260 MI->getOpcode() == AMDGPU::PHI) {
1261 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
1262 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1263 if (!MI->getOperand(i).isReg() ||
1264 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1266 const TargetRegisterClass *OpRC =
1267 MRI.getRegClass(MI->getOperand(i).getReg());
1268 if (RI.hasVGPRs(OpRC)) {
1275 // If any of the operands are VGPR registers, then they all most be
1276 // otherwise we will create illegal VGPR->SGPR copies when legalizing
1278 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
1281 VRC = RI.getEquivalentVGPRClass(SRC);
1288 // Update all the operands so they have the same type.
1289 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1290 if (!MI->getOperand(i).isReg() ||
1291 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1293 unsigned DstReg = MRI.createVirtualRegister(RC);
1294 MachineBasicBlock *InsertBB;
1295 MachineBasicBlock::iterator Insert;
1296 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
1297 InsertBB = MI->getParent();
1300 // MI is a PHI instruction.
1301 InsertBB = MI->getOperand(i + 1).getMBB();
1302 Insert = InsertBB->getFirstTerminator();
1304 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
1305 get(AMDGPU::COPY), DstReg)
1306 .addOperand(MI->getOperand(i));
1307 MI->getOperand(i).setReg(DstReg);
1311 // Legalize INSERT_SUBREG
1312 // src0 must have the same register class as dst
1313 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
1314 unsigned Dst = MI->getOperand(0).getReg();
1315 unsigned Src0 = MI->getOperand(1).getReg();
1316 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
1317 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
1318 if (DstRC != Src0RC) {
1319 MachineBasicBlock &MBB = *MI->getParent();
1320 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
1321 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
1323 MI->getOperand(1).setReg(NewSrc0);
1328 // Legalize MUBUF* instructions
1329 // FIXME: If we start using the non-addr64 instructions for compute, we
1330 // may need to legalize them here.
1332 AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::srsrc);
1333 if (SRsrcIdx != -1) {
1334 // We have an MUBUF instruction
1335 MachineOperand *SRsrc = &MI->getOperand(SRsrcIdx);
1336 unsigned SRsrcRC = get(MI->getOpcode()).OpInfo[SRsrcIdx].RegClass;
1337 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
1338 RI.getRegClass(SRsrcRC))) {
1339 // The operands are legal.
1340 // FIXME: We may need to legalize operands besided srsrc.
1344 MachineBasicBlock &MBB = *MI->getParent();
1345 // Extract the the ptr from the resource descriptor.
1347 // SRsrcPtrLo = srsrc:sub0
1348 unsigned SRsrcPtrLo = buildExtractSubReg(MI, MRI, *SRsrc,
1349 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VReg_32RegClass);
1351 // SRsrcPtrHi = srsrc:sub1
1352 unsigned SRsrcPtrHi = buildExtractSubReg(MI, MRI, *SRsrc,
1353 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VReg_32RegClass);
1355 // Create an empty resource descriptor
1356 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1357 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1358 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1359 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1362 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1366 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1367 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1369 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1371 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1372 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1374 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1376 // NewSRsrc = {Zero64, SRsrcFormat}
1377 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1380 .addImm(AMDGPU::sub0_sub1)
1381 .addReg(SRsrcFormatLo)
1382 .addImm(AMDGPU::sub2)
1383 .addReg(SRsrcFormatHi)
1384 .addImm(AMDGPU::sub3);
1386 MachineOperand *VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1387 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
1388 unsigned NewVAddrLo;
1389 unsigned NewVAddrHi;
1391 // This is already an ADDR64 instruction so we need to add the pointer
1392 // extracted from the resource descriptor to the current value of VAddr.
1393 NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1394 NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1396 // NewVaddrLo = SRsrcPtrLo + VAddr:sub0
1397 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1400 .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
1401 .addReg(AMDGPU::VCC, RegState::ImplicitDefine);
1403 // NewVaddrHi = SRsrcPtrHi + VAddr:sub1
1404 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1407 .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
1408 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1409 .addReg(AMDGPU::VCC, RegState::Implicit);
1412 // This instructions is the _OFFSET variant, so we need to convert it to
1414 MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
1415 MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
1416 MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
1417 assert(SOffset->isImm() && SOffset->getImm() == 0 && "Legalizing MUBUF "
1418 "with non-zero soffset is not implemented");
1421 // Create the new instruction.
1422 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
1423 MachineInstr *Addr64 =
1424 BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
1427 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
1428 // This will be replaced later
1429 // with the new value of vaddr.
1430 .addOperand(*Offset);
1432 MI->removeFromParent();
1435 NewVAddrLo = SRsrcPtrLo;
1436 NewVAddrHi = SRsrcPtrHi;
1437 VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1438 SRsrc = getNamedOperand(*MI, AMDGPU::OpName::srsrc);
1441 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1442 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1445 .addImm(AMDGPU::sub0)
1447 .addImm(AMDGPU::sub1);
1450 // Update the instruction to use NewVaddr
1451 VAddr->setReg(NewVAddr);
1452 // Update the instruction to use NewSRsrc
1453 SRsrc->setReg(NewSRsrc);
1457 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
1458 MachineBasicBlock *MBB = MI->getParent();
1459 switch (MI->getOpcode()) {
1460 case AMDGPU::S_LOAD_DWORD_IMM:
1461 case AMDGPU::S_LOAD_DWORD_SGPR:
1462 case AMDGPU::S_LOAD_DWORDX2_IMM:
1463 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1464 case AMDGPU::S_LOAD_DWORDX4_IMM:
1465 case AMDGPU::S_LOAD_DWORDX4_SGPR:
1466 unsigned NewOpcode = getVALUOp(*MI);
1470 if (MI->getOperand(2).isReg()) {
1471 RegOffset = MI->getOperand(2).getReg();
1474 assert(MI->getOperand(2).isImm());
1475 // SMRD instructions take a dword offsets and MUBUF instructions
1476 // take a byte offset.
1477 ImmOffset = MI->getOperand(2).getImm() << 2;
1478 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1479 if (isUInt<12>(ImmOffset)) {
1480 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1484 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1491 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1492 unsigned DWord0 = RegOffset;
1493 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1494 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1495 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1497 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
1499 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
1500 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1501 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
1502 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1503 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
1505 .addImm(AMDGPU::sub0)
1507 .addImm(AMDGPU::sub1)
1509 .addImm(AMDGPU::sub2)
1511 .addImm(AMDGPU::sub3);
1512 MI->setDesc(get(NewOpcode));
1513 if (MI->getOperand(2).isReg()) {
1514 MI->getOperand(2).setReg(MI->getOperand(1).getReg());
1516 MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
1518 MI->getOperand(1).setReg(SRsrc);
1519 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
1523 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
1524 SmallVector<MachineInstr *, 128> Worklist;
1525 Worklist.push_back(&TopInst);
1527 while (!Worklist.empty()) {
1528 MachineInstr *Inst = Worklist.pop_back_val();
1529 MachineBasicBlock *MBB = Inst->getParent();
1530 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1532 unsigned Opcode = Inst->getOpcode();
1533 unsigned NewOpcode = getVALUOp(*Inst);
1535 // Handle some special cases
1538 if (isSMRD(Inst->getOpcode())) {
1539 moveSMRDToVALU(Inst, MRI);
1542 case AMDGPU::S_MOV_B64: {
1543 DebugLoc DL = Inst->getDebugLoc();
1545 // If the source operand is a register we can replace this with a
1547 if (Inst->getOperand(1).isReg()) {
1548 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
1549 .addOperand(Inst->getOperand(0))
1550 .addOperand(Inst->getOperand(1));
1551 Worklist.push_back(Copy);
1553 // Otherwise, we need to split this into two movs, because there is
1554 // no 64-bit VALU move instruction.
1555 unsigned Reg = Inst->getOperand(0).getReg();
1556 unsigned Dst = split64BitImm(Worklist,
1559 MRI.getRegClass(Reg),
1560 Inst->getOperand(1));
1561 MRI.replaceRegWith(Reg, Dst);
1563 Inst->eraseFromParent();
1566 case AMDGPU::S_AND_B64:
1567 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
1568 Inst->eraseFromParent();
1571 case AMDGPU::S_OR_B64:
1572 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
1573 Inst->eraseFromParent();
1576 case AMDGPU::S_XOR_B64:
1577 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
1578 Inst->eraseFromParent();
1581 case AMDGPU::S_NOT_B64:
1582 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
1583 Inst->eraseFromParent();
1586 case AMDGPU::S_BCNT1_I32_B64:
1587 splitScalar64BitBCNT(Worklist, Inst);
1588 Inst->eraseFromParent();
1591 case AMDGPU::S_BFE_U64:
1592 case AMDGPU::S_BFE_I64:
1593 case AMDGPU::S_BFM_B64:
1594 llvm_unreachable("Moving this op to VALU not implemented");
1597 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
1598 // We cannot move this instruction to the VALU, so we should try to
1599 // legalize its operands instead.
1600 legalizeOperands(Inst);
1604 // Use the new VALU Opcode.
1605 const MCInstrDesc &NewDesc = get(NewOpcode);
1606 Inst->setDesc(NewDesc);
1608 // Remove any references to SCC. Vector instructions can't read from it, and
1609 // We're just about to add the implicit use / defs of VCC, and we don't want
1611 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
1612 MachineOperand &Op = Inst->getOperand(i);
1613 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
1614 Inst->RemoveOperand(i);
1617 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
1618 // We are converting these to a BFE, so we need to add the missing
1619 // operands for the size and offset.
1620 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
1621 Inst->addOperand(MachineOperand::CreateImm(0));
1622 Inst->addOperand(MachineOperand::CreateImm(Size));
1624 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
1625 // The VALU version adds the second operand to the result, so insert an
1627 Inst->addOperand(MachineOperand::CreateImm(0));
1630 addDescImplicitUseDef(NewDesc, Inst);
1632 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
1633 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
1634 // If we need to move this to VGPRs, we need to unpack the second operand
1635 // back into the 2 separate ones for bit offset and width.
1636 assert(OffsetWidthOp.isImm() &&
1637 "Scalar BFE is only implemented for constant width and offset");
1638 uint32_t Imm = OffsetWidthOp.getImm();
1640 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
1641 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
1642 Inst->RemoveOperand(2); // Remove old immediate.
1643 Inst->addOperand(MachineOperand::CreateImm(Offset));
1644 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
1647 // Update the destination register class.
1649 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
1652 // For target instructions, getOpRegClass just returns the virtual
1653 // register class associated with the operand, so we need to find an
1654 // equivalent VGPR register class in order to move the instruction to the
1658 case AMDGPU::REG_SEQUENCE:
1659 case AMDGPU::INSERT_SUBREG:
1660 if (RI.hasVGPRs(NewDstRC))
1662 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
1670 unsigned DstReg = Inst->getOperand(0).getReg();
1671 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1672 MRI.replaceRegWith(DstReg, NewDstReg);
1674 // Legalize the operands
1675 legalizeOperands(Inst);
1677 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
1678 E = MRI.use_end(); I != E; ++I) {
1679 MachineInstr &UseMI = *I->getParent();
1680 if (!canReadVGPR(UseMI, I.getOperandNo())) {
1681 Worklist.push_back(&UseMI);
1687 //===----------------------------------------------------------------------===//
1688 // Indirect addressing callbacks
1689 //===----------------------------------------------------------------------===//
1691 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
1692 unsigned Channel) const {
1693 assert(Channel == 0);
1697 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
1698 return &AMDGPU::VReg_32RegClass;
1701 void SIInstrInfo::splitScalar64BitUnaryOp(
1702 SmallVectorImpl<MachineInstr *> &Worklist,
1704 unsigned Opcode) const {
1705 MachineBasicBlock &MBB = *Inst->getParent();
1706 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1708 MachineOperand &Dest = Inst->getOperand(0);
1709 MachineOperand &Src0 = Inst->getOperand(1);
1710 DebugLoc DL = Inst->getDebugLoc();
1712 MachineBasicBlock::iterator MII = Inst;
1714 const MCInstrDesc &InstDesc = get(Opcode);
1715 const TargetRegisterClass *Src0RC = Src0.isReg() ?
1716 MRI.getRegClass(Src0.getReg()) :
1717 &AMDGPU::SGPR_32RegClass;
1719 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
1721 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1722 AMDGPU::sub0, Src0SubRC);
1724 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
1725 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
1727 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
1728 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
1729 .addOperand(SrcReg0Sub0);
1731 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1732 AMDGPU::sub1, Src0SubRC);
1734 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
1735 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
1736 .addOperand(SrcReg0Sub1);
1738 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
1739 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
1741 .addImm(AMDGPU::sub0)
1743 .addImm(AMDGPU::sub1);
1745 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
1747 // Try to legalize the operands in case we need to swap the order to keep it
1749 Worklist.push_back(LoHalf);
1750 Worklist.push_back(HiHalf);
1753 void SIInstrInfo::splitScalar64BitBinaryOp(
1754 SmallVectorImpl<MachineInstr *> &Worklist,
1756 unsigned Opcode) const {
1757 MachineBasicBlock &MBB = *Inst->getParent();
1758 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1760 MachineOperand &Dest = Inst->getOperand(0);
1761 MachineOperand &Src0 = Inst->getOperand(1);
1762 MachineOperand &Src1 = Inst->getOperand(2);
1763 DebugLoc DL = Inst->getDebugLoc();
1765 MachineBasicBlock::iterator MII = Inst;
1767 const MCInstrDesc &InstDesc = get(Opcode);
1768 const TargetRegisterClass *Src0RC = Src0.isReg() ?
1769 MRI.getRegClass(Src0.getReg()) :
1770 &AMDGPU::SGPR_32RegClass;
1772 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
1773 const TargetRegisterClass *Src1RC = Src1.isReg() ?
1774 MRI.getRegClass(Src1.getReg()) :
1775 &AMDGPU::SGPR_32RegClass;
1777 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
1779 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1780 AMDGPU::sub0, Src0SubRC);
1781 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
1782 AMDGPU::sub0, Src1SubRC);
1784 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
1785 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
1787 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
1788 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
1789 .addOperand(SrcReg0Sub0)
1790 .addOperand(SrcReg1Sub0);
1792 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1793 AMDGPU::sub1, Src0SubRC);
1794 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
1795 AMDGPU::sub1, Src1SubRC);
1797 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
1798 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
1799 .addOperand(SrcReg0Sub1)
1800 .addOperand(SrcReg1Sub1);
1802 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
1803 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
1805 .addImm(AMDGPU::sub0)
1807 .addImm(AMDGPU::sub1);
1809 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
1811 // Try to legalize the operands in case we need to swap the order to keep it
1813 Worklist.push_back(LoHalf);
1814 Worklist.push_back(HiHalf);
1817 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
1818 MachineInstr *Inst) const {
1819 MachineBasicBlock &MBB = *Inst->getParent();
1820 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1822 MachineBasicBlock::iterator MII = Inst;
1823 DebugLoc DL = Inst->getDebugLoc();
1825 MachineOperand &Dest = Inst->getOperand(0);
1826 MachineOperand &Src = Inst->getOperand(1);
1828 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e32);
1829 const TargetRegisterClass *SrcRC = Src.isReg() ?
1830 MRI.getRegClass(Src.getReg()) :
1831 &AMDGPU::SGPR_32RegClass;
1833 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1834 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1836 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
1838 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
1839 AMDGPU::sub0, SrcSubRC);
1840 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
1841 AMDGPU::sub1, SrcSubRC);
1843 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
1844 .addOperand(SrcRegSub0)
1847 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
1848 .addOperand(SrcRegSub1)
1851 MRI.replaceRegWith(Dest.getReg(), ResultReg);
1853 Worklist.push_back(First);
1854 Worklist.push_back(Second);
1857 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
1858 MachineInstr *Inst) const {
1859 // Add the implict and explicit register definitions.
1860 if (NewDesc.ImplicitUses) {
1861 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
1862 unsigned Reg = NewDesc.ImplicitUses[i];
1863 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
1867 if (NewDesc.ImplicitDefs) {
1868 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
1869 unsigned Reg = NewDesc.ImplicitDefs[i];
1870 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
1875 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
1876 MachineBasicBlock *MBB,
1877 MachineBasicBlock::iterator I,
1879 unsigned Address, unsigned OffsetReg) const {
1880 const DebugLoc &DL = MBB->findDebugLoc(I);
1881 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
1882 getIndirectIndexBegin(*MBB->getParent()));
1884 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
1885 .addReg(IndirectBaseReg, RegState::Define)
1886 .addOperand(I->getOperand(0))
1887 .addReg(IndirectBaseReg)
1893 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
1894 MachineBasicBlock *MBB,
1895 MachineBasicBlock::iterator I,
1897 unsigned Address, unsigned OffsetReg) const {
1898 const DebugLoc &DL = MBB->findDebugLoc(I);
1899 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
1900 getIndirectIndexBegin(*MBB->getParent()));
1902 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
1903 .addOperand(I->getOperand(0))
1904 .addOperand(I->getOperand(1))
1905 .addReg(IndirectBaseReg)
1911 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
1912 const MachineFunction &MF) const {
1913 int End = getIndirectIndexEnd(MF);
1914 int Begin = getIndirectIndexBegin(MF);
1920 for (int Index = Begin; Index <= End; ++Index)
1921 Reserved.set(AMDGPU::VReg_32RegClass.getRegister(Index));
1923 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
1924 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
1926 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
1927 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
1929 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
1930 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
1932 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
1933 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
1935 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
1936 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));
1939 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
1940 unsigned OperandName) const {
1941 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
1945 return &MI.getOperand(Idx);