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 //===----------------------------------------------------------------------===//
15 #include "SIInstrInfo.h"
16 #include "AMDGPUTargetMachine.h"
17 #include "GCNHazardRecognizer.h"
18 #include "SIDefines.h"
19 #include "SIMachineFunctionInfo.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/CodeGen/ScheduleDAG.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/CodeGen/RegisterScavenging.h"
26 #include "llvm/MC/MCInstrDesc.h"
27 #include "llvm/Support/Debug.h"
31 // Must be at least 4 to be able to branch over minimum unconditional branch
32 // code. This is only for making it possible to write reasonably small tests for
34 static cl::opt<unsigned>
35 BranchOffsetBits("amdgpu-s-branch-bits", cl::ReallyHidden, cl::init(16),
36 cl::desc("Restrict range of branch instructions (DEBUG)"));
38 SIInstrInfo::SIInstrInfo(const SISubtarget &ST)
39 : AMDGPUInstrInfo(ST), RI(), ST(ST) {}
41 //===----------------------------------------------------------------------===//
42 // TargetInstrInfo callbacks
43 //===----------------------------------------------------------------------===//
45 static unsigned getNumOperandsNoGlue(SDNode *Node) {
46 unsigned N = Node->getNumOperands();
47 while (N && Node->getOperand(N - 1).getValueType() == MVT::Glue)
52 static SDValue findChainOperand(SDNode *Load) {
53 SDValue LastOp = Load->getOperand(getNumOperandsNoGlue(Load) - 1);
54 assert(LastOp.getValueType() == MVT::Other && "Chain missing from load node");
58 /// \brief Returns true if both nodes have the same value for the given
59 /// operand \p Op, or if both nodes do not have this operand.
60 static bool nodesHaveSameOperandValue(SDNode *N0, SDNode* N1, unsigned OpName) {
61 unsigned Opc0 = N0->getMachineOpcode();
62 unsigned Opc1 = N1->getMachineOpcode();
64 int Op0Idx = AMDGPU::getNamedOperandIdx(Opc0, OpName);
65 int Op1Idx = AMDGPU::getNamedOperandIdx(Opc1, OpName);
67 if (Op0Idx == -1 && Op1Idx == -1)
71 if ((Op0Idx == -1 && Op1Idx != -1) ||
72 (Op1Idx == -1 && Op0Idx != -1))
75 // getNamedOperandIdx returns the index for the MachineInstr's operands,
76 // which includes the result as the first operand. We are indexing into the
77 // MachineSDNode's operands, so we need to skip the result operand to get
82 return N0->getOperand(Op0Idx) == N1->getOperand(Op1Idx);
85 bool SIInstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
86 AliasAnalysis *AA) const {
87 // TODO: The generic check fails for VALU instructions that should be
88 // rematerializable due to implicit reads of exec. We really want all of the
89 // generic logic for this except for this.
90 switch (MI.getOpcode()) {
91 case AMDGPU::V_MOV_B32_e32:
92 case AMDGPU::V_MOV_B32_e64:
93 case AMDGPU::V_MOV_B64_PSEUDO:
100 bool SIInstrInfo::areLoadsFromSameBasePtr(SDNode *Load0, SDNode *Load1,
102 int64_t &Offset1) const {
103 if (!Load0->isMachineOpcode() || !Load1->isMachineOpcode())
106 unsigned Opc0 = Load0->getMachineOpcode();
107 unsigned Opc1 = Load1->getMachineOpcode();
109 // Make sure both are actually loads.
110 if (!get(Opc0).mayLoad() || !get(Opc1).mayLoad())
113 if (isDS(Opc0) && isDS(Opc1)) {
115 // FIXME: Handle this case:
116 if (getNumOperandsNoGlue(Load0) != getNumOperandsNoGlue(Load1))
120 if (Load0->getOperand(1) != Load1->getOperand(1))
124 if (findChainOperand(Load0) != findChainOperand(Load1))
127 // Skip read2 / write2 variants for simplicity.
128 // TODO: We should report true if the used offsets are adjacent (excluded
130 if (AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::data1) != -1 ||
131 AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::data1) != -1)
134 Offset0 = cast<ConstantSDNode>(Load0->getOperand(2))->getZExtValue();
135 Offset1 = cast<ConstantSDNode>(Load1->getOperand(2))->getZExtValue();
139 if (isSMRD(Opc0) && isSMRD(Opc1)) {
140 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
143 if (Load0->getOperand(0) != Load1->getOperand(0))
146 const ConstantSDNode *Load0Offset =
147 dyn_cast<ConstantSDNode>(Load0->getOperand(1));
148 const ConstantSDNode *Load1Offset =
149 dyn_cast<ConstantSDNode>(Load1->getOperand(1));
151 if (!Load0Offset || !Load1Offset)
155 if (findChainOperand(Load0) != findChainOperand(Load1))
158 Offset0 = Load0Offset->getZExtValue();
159 Offset1 = Load1Offset->getZExtValue();
163 // MUBUF and MTBUF can access the same addresses.
164 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1))) {
166 // MUBUF and MTBUF have vaddr at different indices.
167 if (!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::soffset) ||
168 findChainOperand(Load0) != findChainOperand(Load1) ||
169 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::vaddr) ||
170 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::srsrc))
173 int OffIdx0 = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
174 int OffIdx1 = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
176 if (OffIdx0 == -1 || OffIdx1 == -1)
179 // getNamedOperandIdx returns the index for MachineInstrs. Since they
180 // inlcude the output in the operand list, but SDNodes don't, we need to
181 // subtract the index by one.
185 SDValue Off0 = Load0->getOperand(OffIdx0);
186 SDValue Off1 = Load1->getOperand(OffIdx1);
188 // The offset might be a FrameIndexSDNode.
189 if (!isa<ConstantSDNode>(Off0) || !isa<ConstantSDNode>(Off1))
192 Offset0 = cast<ConstantSDNode>(Off0)->getZExtValue();
193 Offset1 = cast<ConstantSDNode>(Off1)->getZExtValue();
200 static bool isStride64(unsigned Opc) {
202 case AMDGPU::DS_READ2ST64_B32:
203 case AMDGPU::DS_READ2ST64_B64:
204 case AMDGPU::DS_WRITE2ST64_B32:
205 case AMDGPU::DS_WRITE2ST64_B64:
212 bool SIInstrInfo::getMemOpBaseRegImmOfs(MachineInstr &LdSt, unsigned &BaseReg,
214 const TargetRegisterInfo *TRI) const {
215 unsigned Opc = LdSt.getOpcode();
218 const MachineOperand *OffsetImm =
219 getNamedOperand(LdSt, AMDGPU::OpName::offset);
221 // Normal, single offset LDS instruction.
222 const MachineOperand *AddrReg =
223 getNamedOperand(LdSt, AMDGPU::OpName::addr);
225 BaseReg = AddrReg->getReg();
226 Offset = OffsetImm->getImm();
230 // The 2 offset instructions use offset0 and offset1 instead. We can treat
231 // these as a load with a single offset if the 2 offsets are consecutive. We
232 // will use this for some partially aligned loads.
233 const MachineOperand *Offset0Imm =
234 getNamedOperand(LdSt, AMDGPU::OpName::offset0);
235 const MachineOperand *Offset1Imm =
236 getNamedOperand(LdSt, AMDGPU::OpName::offset1);
238 uint8_t Offset0 = Offset0Imm->getImm();
239 uint8_t Offset1 = Offset1Imm->getImm();
241 if (Offset1 > Offset0 && Offset1 - Offset0 == 1) {
242 // Each of these offsets is in element sized units, so we need to convert
243 // to bytes of the individual reads.
247 EltSize = getOpRegClass(LdSt, 0)->getSize() / 2;
249 assert(LdSt.mayStore());
250 int Data0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
251 EltSize = getOpRegClass(LdSt, Data0Idx)->getSize();
257 const MachineOperand *AddrReg =
258 getNamedOperand(LdSt, AMDGPU::OpName::addr);
259 BaseReg = AddrReg->getReg();
260 Offset = EltSize * Offset0;
267 if (isMUBUF(LdSt) || isMTBUF(LdSt)) {
268 const MachineOperand *SOffset = getNamedOperand(LdSt, AMDGPU::OpName::soffset);
269 if (SOffset && SOffset->isReg())
272 const MachineOperand *AddrReg =
273 getNamedOperand(LdSt, AMDGPU::OpName::vaddr);
277 const MachineOperand *OffsetImm =
278 getNamedOperand(LdSt, AMDGPU::OpName::offset);
279 BaseReg = AddrReg->getReg();
280 Offset = OffsetImm->getImm();
282 if (SOffset) // soffset can be an inline immediate.
283 Offset += SOffset->getImm();
289 const MachineOperand *OffsetImm =
290 getNamedOperand(LdSt, AMDGPU::OpName::offset);
294 const MachineOperand *SBaseReg =
295 getNamedOperand(LdSt, AMDGPU::OpName::sbase);
296 BaseReg = SBaseReg->getReg();
297 Offset = OffsetImm->getImm();
302 const MachineOperand *AddrReg = getNamedOperand(LdSt, AMDGPU::OpName::addr);
303 BaseReg = AddrReg->getReg();
311 bool SIInstrInfo::shouldClusterMemOps(MachineInstr &FirstLdSt,
312 MachineInstr &SecondLdSt,
313 unsigned NumLoads) const {
314 const MachineOperand *FirstDst = nullptr;
315 const MachineOperand *SecondDst = nullptr;
317 if (isDS(FirstLdSt) && isDS(SecondLdSt)) {
318 FirstDst = getNamedOperand(FirstLdSt, AMDGPU::OpName::vdst);
319 SecondDst = getNamedOperand(SecondLdSt, AMDGPU::OpName::vdst);
322 if (isSMRD(FirstLdSt) && isSMRD(SecondLdSt)) {
323 FirstDst = getNamedOperand(FirstLdSt, AMDGPU::OpName::sdst);
324 SecondDst = getNamedOperand(SecondLdSt, AMDGPU::OpName::sdst);
327 if ((isMUBUF(FirstLdSt) && isMUBUF(SecondLdSt)) ||
328 (isMTBUF(FirstLdSt) && isMTBUF(SecondLdSt))) {
329 FirstDst = getNamedOperand(FirstLdSt, AMDGPU::OpName::vdata);
330 SecondDst = getNamedOperand(SecondLdSt, AMDGPU::OpName::vdata);
333 if (!FirstDst || !SecondDst)
336 // Try to limit clustering based on the total number of bytes loaded
337 // rather than the number of instructions. This is done to help reduce
338 // register pressure. The method used is somewhat inexact, though,
339 // because it assumes that all loads in the cluster will load the
340 // same number of bytes as FirstLdSt.
342 // The unit of this value is bytes.
343 // FIXME: This needs finer tuning.
344 unsigned LoadClusterThreshold = 16;
346 const MachineRegisterInfo &MRI =
347 FirstLdSt.getParent()->getParent()->getRegInfo();
348 const TargetRegisterClass *DstRC = MRI.getRegClass(FirstDst->getReg());
350 return (NumLoads * DstRC->getSize()) <= LoadClusterThreshold;
353 void SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
354 MachineBasicBlock::iterator MI,
355 const DebugLoc &DL, unsigned DestReg,
356 unsigned SrcReg, bool KillSrc) const {
357 const TargetRegisterClass *RC = RI.getPhysRegClass(DestReg);
359 if (RC == &AMDGPU::VGPR_32RegClass) {
360 assert(AMDGPU::VGPR_32RegClass.contains(SrcReg) ||
361 AMDGPU::SReg_32RegClass.contains(SrcReg));
362 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
363 .addReg(SrcReg, getKillRegState(KillSrc));
367 if (RC == &AMDGPU::SReg_32RegClass) {
368 if (SrcReg == AMDGPU::SCC) {
369 BuildMI(MBB, MI, DL, get(AMDGPU::S_CSELECT_B32), DestReg)
375 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
376 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
377 .addReg(SrcReg, getKillRegState(KillSrc));
381 if (RC == &AMDGPU::SReg_64RegClass) {
382 if (DestReg == AMDGPU::VCC) {
383 if (AMDGPU::SReg_64RegClass.contains(SrcReg)) {
384 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), AMDGPU::VCC)
385 .addReg(SrcReg, getKillRegState(KillSrc));
387 // FIXME: Hack until VReg_1 removed.
388 assert(AMDGPU::VGPR_32RegClass.contains(SrcReg));
389 BuildMI(MBB, MI, DL, get(AMDGPU::V_CMP_NE_U32_e32))
391 .addReg(SrcReg, getKillRegState(KillSrc));
397 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
398 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
399 .addReg(SrcReg, getKillRegState(KillSrc));
403 if (DestReg == AMDGPU::SCC) {
404 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
405 BuildMI(MBB, MI, DL, get(AMDGPU::S_CMP_LG_U32))
406 .addReg(SrcReg, getKillRegState(KillSrc))
411 unsigned EltSize = 4;
412 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
413 if (RI.isSGPRClass(RC)) {
414 if (RC->getSize() > 4) {
415 Opcode = AMDGPU::S_MOV_B64;
418 Opcode = AMDGPU::S_MOV_B32;
423 ArrayRef<int16_t> SubIndices = RI.getRegSplitParts(RC, EltSize);
424 bool Forward = RI.getHWRegIndex(DestReg) <= RI.getHWRegIndex(SrcReg);
426 for (unsigned Idx = 0; Idx < SubIndices.size(); ++Idx) {
429 SubIdx = SubIndices[Idx];
431 SubIdx = SubIndices[SubIndices.size() - Idx - 1];
433 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
434 get(Opcode), RI.getSubReg(DestReg, SubIdx));
436 Builder.addReg(RI.getSubReg(SrcReg, SubIdx));
438 if (Idx == SubIndices.size() - 1)
439 Builder.addReg(SrcReg, getKillRegState(KillSrc) | RegState::Implicit);
442 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
444 Builder.addReg(SrcReg, RegState::Implicit);
448 int SIInstrInfo::commuteOpcode(unsigned Opcode) const {
451 // Try to map original to commuted opcode
452 NewOpc = AMDGPU::getCommuteRev(Opcode);
454 // Check if the commuted (REV) opcode exists on the target.
455 return pseudoToMCOpcode(NewOpc) != -1 ? NewOpc : -1;
457 // Try to map commuted to original opcode
458 NewOpc = AMDGPU::getCommuteOrig(Opcode);
460 // Check if the original (non-REV) opcode exists on the target.
461 return pseudoToMCOpcode(NewOpc) != -1 ? NewOpc : -1;
466 unsigned SIInstrInfo::getMovOpcode(const TargetRegisterClass *DstRC) const {
468 if (DstRC->getSize() == 4) {
469 return RI.isSGPRClass(DstRC) ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32;
470 } else if (DstRC->getSize() == 8 && RI.isSGPRClass(DstRC)) {
471 return AMDGPU::S_MOV_B64;
472 } else if (DstRC->getSize() == 8 && !RI.isSGPRClass(DstRC)) {
473 return AMDGPU::V_MOV_B64_PSEUDO;
478 static unsigned getSGPRSpillSaveOpcode(unsigned Size) {
481 return AMDGPU::SI_SPILL_S32_SAVE;
483 return AMDGPU::SI_SPILL_S64_SAVE;
485 return AMDGPU::SI_SPILL_S128_SAVE;
487 return AMDGPU::SI_SPILL_S256_SAVE;
489 return AMDGPU::SI_SPILL_S512_SAVE;
491 llvm_unreachable("unknown register size");
495 static unsigned getVGPRSpillSaveOpcode(unsigned Size) {
498 return AMDGPU::SI_SPILL_V32_SAVE;
500 return AMDGPU::SI_SPILL_V64_SAVE;
502 return AMDGPU::SI_SPILL_V96_SAVE;
504 return AMDGPU::SI_SPILL_V128_SAVE;
506 return AMDGPU::SI_SPILL_V256_SAVE;
508 return AMDGPU::SI_SPILL_V512_SAVE;
510 llvm_unreachable("unknown register size");
514 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
515 MachineBasicBlock::iterator MI,
516 unsigned SrcReg, bool isKill,
518 const TargetRegisterClass *RC,
519 const TargetRegisterInfo *TRI) const {
520 MachineFunction *MF = MBB.getParent();
521 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
522 MachineFrameInfo &FrameInfo = MF->getFrameInfo();
523 DebugLoc DL = MBB.findDebugLoc(MI);
525 unsigned Size = FrameInfo.getObjectSize(FrameIndex);
526 unsigned Align = FrameInfo.getObjectAlignment(FrameIndex);
527 MachinePointerInfo PtrInfo
528 = MachinePointerInfo::getFixedStack(*MF, FrameIndex);
529 MachineMemOperand *MMO
530 = MF->getMachineMemOperand(PtrInfo, MachineMemOperand::MOStore,
533 if (RI.isSGPRClass(RC)) {
534 MFI->setHasSpilledSGPRs();
536 // We are only allowed to create one new instruction when spilling
537 // registers, so we need to use pseudo instruction for spilling SGPRs.
538 const MCInstrDesc &OpDesc = get(getSGPRSpillSaveOpcode(RC->getSize()));
540 // The SGPR spill/restore instructions only work on number sgprs, so we need
541 // to make sure we are using the correct register class.
542 if (TargetRegisterInfo::isVirtualRegister(SrcReg) && RC->getSize() == 4) {
543 MachineRegisterInfo &MRI = MF->getRegInfo();
544 MRI.constrainRegClass(SrcReg, &AMDGPU::SReg_32_XM0RegClass);
547 MachineInstrBuilder Spill = BuildMI(MBB, MI, DL, OpDesc)
548 .addReg(SrcReg, getKillRegState(isKill)) // data
549 .addFrameIndex(FrameIndex) // addr
551 .addReg(MFI->getScratchRSrcReg(), RegState::Implicit)
552 .addReg(MFI->getScratchWaveOffsetReg(), RegState::Implicit);
553 // Add the scratch resource registers as implicit uses because we may end up
554 // needing them, and need to ensure that the reserved registers are
555 // correctly handled.
557 if (ST.hasScalarStores()) {
558 // m0 is used for offset to scalar stores if used to spill.
559 Spill.addReg(AMDGPU::M0, RegState::ImplicitDefine);
565 if (!ST.isVGPRSpillingEnabled(*MF->getFunction())) {
566 LLVMContext &Ctx = MF->getFunction()->getContext();
567 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Do not know how to"
569 BuildMI(MBB, MI, DL, get(AMDGPU::KILL))
575 assert(RI.hasVGPRs(RC) && "Only VGPR spilling expected");
577 unsigned Opcode = getVGPRSpillSaveOpcode(RC->getSize());
578 MFI->setHasSpilledVGPRs();
579 BuildMI(MBB, MI, DL, get(Opcode))
580 .addReg(SrcReg, getKillRegState(isKill)) // data
581 .addFrameIndex(FrameIndex) // addr
582 .addReg(MFI->getScratchRSrcReg()) // scratch_rsrc
583 .addReg(MFI->getScratchWaveOffsetReg()) // scratch_offset
588 static unsigned getSGPRSpillRestoreOpcode(unsigned Size) {
591 return AMDGPU::SI_SPILL_S32_RESTORE;
593 return AMDGPU::SI_SPILL_S64_RESTORE;
595 return AMDGPU::SI_SPILL_S128_RESTORE;
597 return AMDGPU::SI_SPILL_S256_RESTORE;
599 return AMDGPU::SI_SPILL_S512_RESTORE;
601 llvm_unreachable("unknown register size");
605 static unsigned getVGPRSpillRestoreOpcode(unsigned Size) {
608 return AMDGPU::SI_SPILL_V32_RESTORE;
610 return AMDGPU::SI_SPILL_V64_RESTORE;
612 return AMDGPU::SI_SPILL_V96_RESTORE;
614 return AMDGPU::SI_SPILL_V128_RESTORE;
616 return AMDGPU::SI_SPILL_V256_RESTORE;
618 return AMDGPU::SI_SPILL_V512_RESTORE;
620 llvm_unreachable("unknown register size");
624 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
625 MachineBasicBlock::iterator MI,
626 unsigned DestReg, int FrameIndex,
627 const TargetRegisterClass *RC,
628 const TargetRegisterInfo *TRI) const {
629 MachineFunction *MF = MBB.getParent();
630 const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
631 MachineFrameInfo &FrameInfo = MF->getFrameInfo();
632 DebugLoc DL = MBB.findDebugLoc(MI);
633 unsigned Align = FrameInfo.getObjectAlignment(FrameIndex);
634 unsigned Size = FrameInfo.getObjectSize(FrameIndex);
636 MachinePointerInfo PtrInfo
637 = MachinePointerInfo::getFixedStack(*MF, FrameIndex);
639 MachineMemOperand *MMO = MF->getMachineMemOperand(
640 PtrInfo, MachineMemOperand::MOLoad, Size, Align);
642 if (RI.isSGPRClass(RC)) {
643 // FIXME: Maybe this should not include a memoperand because it will be
644 // lowered to non-memory instructions.
645 const MCInstrDesc &OpDesc = get(getSGPRSpillRestoreOpcode(RC->getSize()));
646 if (TargetRegisterInfo::isVirtualRegister(DestReg) && RC->getSize() == 4) {
647 MachineRegisterInfo &MRI = MF->getRegInfo();
648 MRI.constrainRegClass(DestReg, &AMDGPU::SReg_32_XM0RegClass);
651 MachineInstrBuilder Spill = BuildMI(MBB, MI, DL, OpDesc, DestReg)
652 .addFrameIndex(FrameIndex) // addr
654 .addReg(MFI->getScratchRSrcReg(), RegState::Implicit)
655 .addReg(MFI->getScratchWaveOffsetReg(), RegState::Implicit);
657 if (ST.hasScalarStores()) {
658 // m0 is used for offset to scalar stores if used to spill.
659 Spill.addReg(AMDGPU::M0, RegState::ImplicitDefine);
665 if (!ST.isVGPRSpillingEnabled(*MF->getFunction())) {
666 LLVMContext &Ctx = MF->getFunction()->getContext();
667 Ctx.emitError("SIInstrInfo::loadRegFromStackSlot - Do not know how to"
668 " restore register");
669 BuildMI(MBB, MI, DL, get(AMDGPU::IMPLICIT_DEF), DestReg);
674 assert(RI.hasVGPRs(RC) && "Only VGPR spilling expected");
676 unsigned Opcode = getVGPRSpillRestoreOpcode(RC->getSize());
677 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
678 .addFrameIndex(FrameIndex) // vaddr
679 .addReg(MFI->getScratchRSrcReg()) // scratch_rsrc
680 .addReg(MFI->getScratchWaveOffsetReg()) // scratch_offset
685 /// \param @Offset Offset in bytes of the FrameIndex being spilled
686 unsigned SIInstrInfo::calculateLDSSpillAddress(
687 MachineBasicBlock &MBB, MachineInstr &MI, RegScavenger *RS, unsigned TmpReg,
688 unsigned FrameOffset, unsigned Size) const {
689 MachineFunction *MF = MBB.getParent();
690 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
691 const SISubtarget &ST = MF->getSubtarget<SISubtarget>();
692 const SIRegisterInfo *TRI = ST.getRegisterInfo();
693 DebugLoc DL = MBB.findDebugLoc(MI);
694 unsigned WorkGroupSize = MFI->getMaxFlatWorkGroupSize();
695 unsigned WavefrontSize = ST.getWavefrontSize();
697 unsigned TIDReg = MFI->getTIDReg();
698 if (!MFI->hasCalculatedTID()) {
699 MachineBasicBlock &Entry = MBB.getParent()->front();
700 MachineBasicBlock::iterator Insert = Entry.front();
701 DebugLoc DL = Insert->getDebugLoc();
703 TIDReg = RI.findUnusedRegister(MF->getRegInfo(), &AMDGPU::VGPR_32RegClass,
705 if (TIDReg == AMDGPU::NoRegister)
708 if (!AMDGPU::isShader(MF->getFunction()->getCallingConv()) &&
709 WorkGroupSize > WavefrontSize) {
712 = TRI->getPreloadedValue(*MF, SIRegisterInfo::WORKGROUP_ID_X);
714 = TRI->getPreloadedValue(*MF, SIRegisterInfo::WORKGROUP_ID_Y);
716 = TRI->getPreloadedValue(*MF, SIRegisterInfo::WORKGROUP_ID_Z);
717 unsigned InputPtrReg =
718 TRI->getPreloadedValue(*MF, SIRegisterInfo::KERNARG_SEGMENT_PTR);
719 for (unsigned Reg : {TIDIGXReg, TIDIGYReg, TIDIGZReg}) {
720 if (!Entry.isLiveIn(Reg))
721 Entry.addLiveIn(Reg);
724 RS->enterBasicBlock(Entry);
725 // FIXME: Can we scavenge an SReg_64 and access the subregs?
726 unsigned STmp0 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
727 unsigned STmp1 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
728 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp0)
730 .addImm(SI::KernelInputOffsets::NGROUPS_Z);
731 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp1)
733 .addImm(SI::KernelInputOffsets::NGROUPS_Y);
735 // NGROUPS.X * NGROUPS.Y
736 BuildMI(Entry, Insert, DL, get(AMDGPU::S_MUL_I32), STmp1)
739 // (NGROUPS.X * NGROUPS.Y) * TIDIG.X
740 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MUL_U32_U24_e32), TIDReg)
743 // NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)
744 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MAD_U32_U24), TIDReg)
748 // (NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)) + TIDIG.Z
749 BuildMI(Entry, Insert, DL, get(AMDGPU::V_ADD_I32_e32), TIDReg)
754 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_LO_U32_B32_e64),
759 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_HI_U32_B32_e64),
765 BuildMI(Entry, Insert, DL, get(AMDGPU::V_LSHLREV_B32_e32),
769 MFI->setTIDReg(TIDReg);
772 // Add FrameIndex to LDS offset
773 unsigned LDSOffset = MFI->getLDSSize() + (FrameOffset * WorkGroupSize);
774 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_I32_e32), TmpReg)
781 void SIInstrInfo::insertWaitStates(MachineBasicBlock &MBB,
782 MachineBasicBlock::iterator MI,
784 DebugLoc DL = MBB.findDebugLoc(MI);
792 BuildMI(MBB, MI, DL, get(AMDGPU::S_NOP))
797 void SIInstrInfo::insertNoop(MachineBasicBlock &MBB,
798 MachineBasicBlock::iterator MI) const {
799 insertWaitStates(MBB, MI, 1);
802 unsigned SIInstrInfo::getNumWaitStates(const MachineInstr &MI) const {
803 switch (MI.getOpcode()) {
804 default: return 1; // FIXME: Do wait states equal cycles?
807 return MI.getOperand(0).getImm() + 1;
811 bool SIInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
812 MachineBasicBlock &MBB = *MI.getParent();
813 DebugLoc DL = MBB.findDebugLoc(MI);
814 switch (MI.getOpcode()) {
815 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
816 case AMDGPU::S_MOV_B64_term: {
817 // This is only a terminator to get the correct spill code placement during
818 // register allocation.
819 MI.setDesc(get(AMDGPU::S_MOV_B64));
822 case AMDGPU::S_XOR_B64_term: {
823 // This is only a terminator to get the correct spill code placement during
824 // register allocation.
825 MI.setDesc(get(AMDGPU::S_XOR_B64));
828 case AMDGPU::S_ANDN2_B64_term: {
829 // This is only a terminator to get the correct spill code placement during
830 // register allocation.
831 MI.setDesc(get(AMDGPU::S_ANDN2_B64));
834 case AMDGPU::V_MOV_B64_PSEUDO: {
835 unsigned Dst = MI.getOperand(0).getReg();
836 unsigned DstLo = RI.getSubReg(Dst, AMDGPU::sub0);
837 unsigned DstHi = RI.getSubReg(Dst, AMDGPU::sub1);
839 const MachineOperand &SrcOp = MI.getOperand(1);
840 // FIXME: Will this work for 64-bit floating point immediates?
841 assert(!SrcOp.isFPImm());
843 APInt Imm(64, SrcOp.getImm());
844 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
845 .addImm(Imm.getLoBits(32).getZExtValue())
846 .addReg(Dst, RegState::Implicit | RegState::Define);
847 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
848 .addImm(Imm.getHiBits(32).getZExtValue())
849 .addReg(Dst, RegState::Implicit | RegState::Define);
851 assert(SrcOp.isReg());
852 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
853 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub0))
854 .addReg(Dst, RegState::Implicit | RegState::Define);
855 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
856 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub1))
857 .addReg(Dst, RegState::Implicit | RegState::Define);
859 MI.eraseFromParent();
862 case AMDGPU::V_MOVRELD_B32_V1:
863 case AMDGPU::V_MOVRELD_B32_V2:
864 case AMDGPU::V_MOVRELD_B32_V4:
865 case AMDGPU::V_MOVRELD_B32_V8:
866 case AMDGPU::V_MOVRELD_B32_V16: {
867 const MCInstrDesc &MovRelDesc = get(AMDGPU::V_MOVRELD_B32_e32);
868 unsigned VecReg = MI.getOperand(0).getReg();
869 bool IsUndef = MI.getOperand(1).isUndef();
870 unsigned SubReg = AMDGPU::sub0 + MI.getOperand(3).getImm();
871 assert(VecReg == MI.getOperand(1).getReg());
873 MachineInstr *MovRel =
874 BuildMI(MBB, MI, DL, MovRelDesc)
875 .addReg(RI.getSubReg(VecReg, SubReg), RegState::Undef)
876 .addOperand(MI.getOperand(2))
877 .addReg(VecReg, RegState::ImplicitDefine)
878 .addReg(VecReg, RegState::Implicit | (IsUndef ? RegState::Undef : 0));
880 const int ImpDefIdx =
881 MovRelDesc.getNumOperands() + MovRelDesc.getNumImplicitUses();
882 const int ImpUseIdx = ImpDefIdx + 1;
883 MovRel->tieOperands(ImpDefIdx, ImpUseIdx);
885 MI.eraseFromParent();
888 case AMDGPU::SI_PC_ADD_REL_OFFSET: {
889 MachineFunction &MF = *MBB.getParent();
890 unsigned Reg = MI.getOperand(0).getReg();
891 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
892 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
894 // Create a bundle so these instructions won't be re-ordered by the
895 // post-RA scheduler.
896 MIBundleBuilder Bundler(MBB, MI);
897 Bundler.append(BuildMI(MF, DL, get(AMDGPU::S_GETPC_B64), Reg));
899 // Add 32-bit offset from this instruction to the start of the
901 Bundler.append(BuildMI(MF, DL, get(AMDGPU::S_ADD_U32), RegLo)
903 .addOperand(MI.getOperand(1)));
905 MachineInstrBuilder MIB = BuildMI(MF, DL, get(AMDGPU::S_ADDC_U32), RegHi)
907 if (MI.getOperand(2).getTargetFlags() == SIInstrInfo::MO_NONE)
910 MIB.addOperand(MI.getOperand(2));
913 llvm::finalizeBundle(MBB, Bundler.begin());
915 MI.eraseFromParent();
922 bool SIInstrInfo::swapSourceModifiers(MachineInstr &MI,
923 MachineOperand &Src0,
925 MachineOperand &Src1,
926 unsigned Src1OpName) const {
927 MachineOperand *Src0Mods = getNamedOperand(MI, Src0OpName);
931 MachineOperand *Src1Mods = getNamedOperand(MI, Src1OpName);
933 "All commutable instructions have both src0 and src1 modifiers");
935 int Src0ModsVal = Src0Mods->getImm();
936 int Src1ModsVal = Src1Mods->getImm();
938 Src1Mods->setImm(Src0ModsVal);
939 Src0Mods->setImm(Src1ModsVal);
943 static MachineInstr *swapRegAndNonRegOperand(MachineInstr &MI,
944 MachineOperand &RegOp,
945 MachineOperand &NonRegOp) {
946 unsigned Reg = RegOp.getReg();
947 unsigned SubReg = RegOp.getSubReg();
948 bool IsKill = RegOp.isKill();
949 bool IsDead = RegOp.isDead();
950 bool IsUndef = RegOp.isUndef();
951 bool IsDebug = RegOp.isDebug();
953 if (NonRegOp.isImm())
954 RegOp.ChangeToImmediate(NonRegOp.getImm());
955 else if (NonRegOp.isFI())
956 RegOp.ChangeToFrameIndex(NonRegOp.getIndex());
960 NonRegOp.ChangeToRegister(Reg, false, false, IsKill, IsDead, IsUndef, IsDebug);
961 NonRegOp.setSubReg(SubReg);
966 MachineInstr *SIInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
968 unsigned Src1Idx) const {
969 assert(!NewMI && "this should never be used");
971 unsigned Opc = MI.getOpcode();
972 int CommutedOpcode = commuteOpcode(Opc);
973 if (CommutedOpcode == -1)
976 assert(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0) ==
977 static_cast<int>(Src0Idx) &&
978 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1) ==
979 static_cast<int>(Src1Idx) &&
980 "inconsistency with findCommutedOpIndices");
982 MachineOperand &Src0 = MI.getOperand(Src0Idx);
983 MachineOperand &Src1 = MI.getOperand(Src1Idx);
985 MachineInstr *CommutedMI = nullptr;
986 if (Src0.isReg() && Src1.isReg()) {
987 if (isOperandLegal(MI, Src1Idx, &Src0)) {
988 // Be sure to copy the source modifiers to the right place.
990 = TargetInstrInfo::commuteInstructionImpl(MI, NewMI, Src0Idx, Src1Idx);
993 } else if (Src0.isReg() && !Src1.isReg()) {
994 // src0 should always be able to support any operand type, so no need to
995 // check operand legality.
996 CommutedMI = swapRegAndNonRegOperand(MI, Src0, Src1);
997 } else if (!Src0.isReg() && Src1.isReg()) {
998 if (isOperandLegal(MI, Src1Idx, &Src0))
999 CommutedMI = swapRegAndNonRegOperand(MI, Src1, Src0);
1001 // FIXME: Found two non registers to commute. This does happen.
1007 swapSourceModifiers(MI, Src0, AMDGPU::OpName::src0_modifiers,
1008 Src1, AMDGPU::OpName::src1_modifiers);
1010 CommutedMI->setDesc(get(CommutedOpcode));
1016 // This needs to be implemented because the source modifiers may be inserted
1017 // between the true commutable operands, and the base
1018 // TargetInstrInfo::commuteInstruction uses it.
1019 bool SIInstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx0,
1020 unsigned &SrcOpIdx1) const {
1021 if (!MI.isCommutable())
1024 unsigned Opc = MI.getOpcode();
1025 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
1029 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
1033 return fixCommutedOpIndices(SrcOpIdx0, SrcOpIdx1, Src0Idx, Src1Idx);
1036 bool SIInstrInfo::isBranchOffsetInRange(unsigned BranchOp,
1037 int64_t BrOffset) const {
1038 // BranchRelaxation should never have to check s_setpc_b64 because its dest
1039 // block is unanalyzable.
1040 assert(BranchOp != AMDGPU::S_SETPC_B64);
1042 // Convert to dwords.
1045 // The branch instructions do PC += signext(SIMM16 * 4) + 4, so the offset is
1046 // from the next instruction.
1049 return isIntN(BranchOffsetBits, BrOffset);
1052 MachineBasicBlock *SIInstrInfo::getBranchDestBlock(
1053 const MachineInstr &MI) const {
1054 if (MI.getOpcode() == AMDGPU::S_SETPC_B64) {
1055 // This would be a difficult analysis to perform, but can always be legal so
1056 // there's no need to analyze it.
1060 return MI.getOperand(0).getMBB();
1063 unsigned SIInstrInfo::insertIndirectBranch(MachineBasicBlock &MBB,
1064 MachineBasicBlock &DestBB,
1067 RegScavenger *RS) const {
1068 assert(RS && "RegScavenger required for long branching");
1069 assert(MBB.empty() &&
1070 "new block should be inserted for expanding unconditional branch");
1071 assert(MBB.pred_size() == 1);
1073 MachineFunction *MF = MBB.getParent();
1074 MachineRegisterInfo &MRI = MF->getRegInfo();
1076 // FIXME: Virtual register workaround for RegScavenger not working with empty
1078 unsigned PCReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1082 // We need to compute the offset relative to the instruction immediately after
1083 // s_getpc_b64. Insert pc arithmetic code before last terminator.
1084 MachineInstr *GetPC = BuildMI(MBB, I, DL, get(AMDGPU::S_GETPC_B64), PCReg);
1086 // TODO: Handle > 32-bit block address.
1087 if (BrOffset >= 0) {
1088 BuildMI(MBB, I, DL, get(AMDGPU::S_ADD_U32))
1089 .addReg(PCReg, RegState::Define, AMDGPU::sub0)
1090 .addReg(PCReg, 0, AMDGPU::sub0)
1091 .addMBB(&DestBB, AMDGPU::TF_LONG_BRANCH_FORWARD);
1092 BuildMI(MBB, I, DL, get(AMDGPU::S_ADDC_U32))
1093 .addReg(PCReg, RegState::Define, AMDGPU::sub1)
1094 .addReg(PCReg, 0, AMDGPU::sub1)
1097 // Backwards branch.
1098 BuildMI(MBB, I, DL, get(AMDGPU::S_SUB_U32))
1099 .addReg(PCReg, RegState::Define, AMDGPU::sub0)
1100 .addReg(PCReg, 0, AMDGPU::sub0)
1101 .addMBB(&DestBB, AMDGPU::TF_LONG_BRANCH_BACKWARD);
1102 BuildMI(MBB, I, DL, get(AMDGPU::S_SUBB_U32))
1103 .addReg(PCReg, RegState::Define, AMDGPU::sub1)
1104 .addReg(PCReg, 0, AMDGPU::sub1)
1108 // Insert the indirect branch after the other terminator.
1109 BuildMI(&MBB, DL, get(AMDGPU::S_SETPC_B64))
1112 // FIXME: If spilling is necessary, this will fail because this scavenger has
1113 // no emergency stack slots. It is non-trivial to spill in this situation,
1114 // because the restore code needs to be specially placed after the
1115 // jump. BranchRelaxation then needs to be made aware of the newly inserted
1118 // If a spill is needed for the pc register pair, we need to insert a spill
1119 // restore block right before the destination block, and insert a short branch
1120 // into the old destination block's fallthrough predecessor.
1123 // s_cbranch_scc0 skip_long_branch:
1127 // s_getpc_b64 s[8:9]
1128 // s_add_u32 s8, s8, restore_bb
1129 // s_addc_u32 s9, s9, 0
1130 // s_setpc_b64 s[8:9]
1132 // skip_long_branch:
1137 // dest_bb_fallthrough_predecessor:
1143 // fallthrough dest_bb
1148 RS->enterBasicBlockEnd(MBB);
1149 unsigned Scav = RS->scavengeRegister(&AMDGPU::SReg_64RegClass,
1150 MachineBasicBlock::iterator(GetPC), 0);
1151 MRI.replaceRegWith(PCReg, Scav);
1152 MRI.clearVirtRegs();
1153 RS->setRegUsed(Scav);
1155 return 4 + 8 + 4 + 4;
1158 unsigned SIInstrInfo::getBranchOpcode(SIInstrInfo::BranchPredicate Cond) {
1160 case SIInstrInfo::SCC_TRUE:
1161 return AMDGPU::S_CBRANCH_SCC1;
1162 case SIInstrInfo::SCC_FALSE:
1163 return AMDGPU::S_CBRANCH_SCC0;
1164 case SIInstrInfo::VCCNZ:
1165 return AMDGPU::S_CBRANCH_VCCNZ;
1166 case SIInstrInfo::VCCZ:
1167 return AMDGPU::S_CBRANCH_VCCZ;
1168 case SIInstrInfo::EXECNZ:
1169 return AMDGPU::S_CBRANCH_EXECNZ;
1170 case SIInstrInfo::EXECZ:
1171 return AMDGPU::S_CBRANCH_EXECZ;
1173 llvm_unreachable("invalid branch predicate");
1177 SIInstrInfo::BranchPredicate SIInstrInfo::getBranchPredicate(unsigned Opcode) {
1179 case AMDGPU::S_CBRANCH_SCC0:
1181 case AMDGPU::S_CBRANCH_SCC1:
1183 case AMDGPU::S_CBRANCH_VCCNZ:
1185 case AMDGPU::S_CBRANCH_VCCZ:
1187 case AMDGPU::S_CBRANCH_EXECNZ:
1189 case AMDGPU::S_CBRANCH_EXECZ:
1196 bool SIInstrInfo::analyzeBranchImpl(MachineBasicBlock &MBB,
1197 MachineBasicBlock::iterator I,
1198 MachineBasicBlock *&TBB,
1199 MachineBasicBlock *&FBB,
1200 SmallVectorImpl<MachineOperand> &Cond,
1201 bool AllowModify) const {
1202 if (I->getOpcode() == AMDGPU::S_BRANCH) {
1203 // Unconditional Branch
1204 TBB = I->getOperand(0).getMBB();
1208 BranchPredicate Pred = getBranchPredicate(I->getOpcode());
1209 if (Pred == INVALID_BR)
1212 MachineBasicBlock *CondBB = I->getOperand(0).getMBB();
1213 Cond.push_back(MachineOperand::CreateImm(Pred));
1214 Cond.push_back(I->getOperand(1)); // Save the branch register.
1218 if (I == MBB.end()) {
1219 // Conditional branch followed by fall-through.
1224 if (I->getOpcode() == AMDGPU::S_BRANCH) {
1226 FBB = I->getOperand(0).getMBB();
1233 bool SIInstrInfo::analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
1234 MachineBasicBlock *&FBB,
1235 SmallVectorImpl<MachineOperand> &Cond,
1236 bool AllowModify) const {
1237 MachineBasicBlock::iterator I = MBB.getFirstTerminator();
1241 if (I->getOpcode() != AMDGPU::SI_MASK_BRANCH)
1242 return analyzeBranchImpl(MBB, I, TBB, FBB, Cond, AllowModify);
1246 // TODO: Should be able to treat as fallthrough?
1250 if (analyzeBranchImpl(MBB, I, TBB, FBB, Cond, AllowModify))
1253 MachineBasicBlock *MaskBrDest = I->getOperand(0).getMBB();
1255 // Specifically handle the case where the conditional branch is to the same
1256 // destination as the mask branch. e.g.
1258 // si_mask_branch BB8
1259 // s_cbranch_execz BB8
1262 // This is required to understand divergent loops which may need the branches
1264 if (TBB != MaskBrDest || Cond.empty())
1267 auto Pred = Cond[0].getImm();
1268 return (Pred != EXECZ && Pred != EXECNZ);
1271 unsigned SIInstrInfo::removeBranch(MachineBasicBlock &MBB,
1272 int *BytesRemoved) const {
1273 MachineBasicBlock::iterator I = MBB.getFirstTerminator();
1276 unsigned RemovedSize = 0;
1277 while (I != MBB.end()) {
1278 MachineBasicBlock::iterator Next = std::next(I);
1279 if (I->getOpcode() == AMDGPU::SI_MASK_BRANCH) {
1284 RemovedSize += getInstSizeInBytes(*I);
1285 I->eraseFromParent();
1291 *BytesRemoved = RemovedSize;
1296 unsigned SIInstrInfo::insertBranch(MachineBasicBlock &MBB,
1297 MachineBasicBlock *TBB,
1298 MachineBasicBlock *FBB,
1299 ArrayRef<MachineOperand> Cond,
1301 int *BytesAdded) const {
1303 if (!FBB && Cond.empty()) {
1304 BuildMI(&MBB, DL, get(AMDGPU::S_BRANCH))
1311 assert(TBB && Cond[0].isImm());
1314 = getBranchOpcode(static_cast<BranchPredicate>(Cond[0].getImm()));
1318 MachineInstr *CondBr =
1319 BuildMI(&MBB, DL, get(Opcode))
1322 // Copy the flags onto the implicit condition register operand.
1323 MachineOperand &CondReg = CondBr->getOperand(1);
1324 CondReg.setIsUndef(Cond[1].isUndef());
1325 CondReg.setIsKill(Cond[1].isKill());
1334 MachineInstr *CondBr =
1335 BuildMI(&MBB, DL, get(Opcode))
1337 BuildMI(&MBB, DL, get(AMDGPU::S_BRANCH))
1340 MachineOperand &CondReg = CondBr->getOperand(1);
1341 CondReg.setIsUndef(Cond[1].isUndef());
1342 CondReg.setIsKill(Cond[1].isKill());
1350 bool SIInstrInfo::reverseBranchCondition(
1351 SmallVectorImpl<MachineOperand> &Cond) const {
1352 assert(Cond.size() == 2);
1353 Cond[0].setImm(-Cond[0].getImm());
1357 static void removeModOperands(MachineInstr &MI) {
1358 unsigned Opc = MI.getOpcode();
1359 int Src0ModIdx = AMDGPU::getNamedOperandIdx(Opc,
1360 AMDGPU::OpName::src0_modifiers);
1361 int Src1ModIdx = AMDGPU::getNamedOperandIdx(Opc,
1362 AMDGPU::OpName::src1_modifiers);
1363 int Src2ModIdx = AMDGPU::getNamedOperandIdx(Opc,
1364 AMDGPU::OpName::src2_modifiers);
1366 MI.RemoveOperand(Src2ModIdx);
1367 MI.RemoveOperand(Src1ModIdx);
1368 MI.RemoveOperand(Src0ModIdx);
1371 bool SIInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
1372 unsigned Reg, MachineRegisterInfo *MRI) const {
1373 if (!MRI->hasOneNonDBGUse(Reg))
1376 unsigned Opc = UseMI.getOpcode();
1377 if (Opc == AMDGPU::COPY) {
1378 bool isVGPRCopy = RI.isVGPR(*MRI, UseMI.getOperand(0).getReg());
1379 switch (DefMI.getOpcode()) {
1382 case AMDGPU::S_MOV_B64:
1383 // TODO: We could fold 64-bit immediates, but this get compilicated
1384 // when there are sub-registers.
1387 case AMDGPU::V_MOV_B32_e32:
1388 case AMDGPU::S_MOV_B32:
1391 unsigned NewOpc = isVGPRCopy ? AMDGPU::V_MOV_B32_e32 : AMDGPU::S_MOV_B32;
1392 const MachineOperand *ImmOp = getNamedOperand(DefMI, AMDGPU::OpName::src0);
1394 // FIXME: We could handle FrameIndex values here.
1395 if (!ImmOp->isImm()) {
1398 UseMI.setDesc(get(NewOpc));
1399 UseMI.getOperand(1).ChangeToImmediate(ImmOp->getImm());
1400 UseMI.addImplicitDefUseOperands(*UseMI.getParent()->getParent());
1404 if (Opc == AMDGPU::V_MAD_F32 || Opc == AMDGPU::V_MAC_F32_e64 ||
1405 Opc == AMDGPU::V_MAD_F16 || Opc == AMDGPU::V_MAC_F16_e64) {
1406 bool IsF32 = Opc == AMDGPU::V_MAD_F32 || Opc == AMDGPU::V_MAC_F32_e64;
1408 // Don't fold if we are using source modifiers. The new VOP2 instructions
1410 if (hasModifiersSet(UseMI, AMDGPU::OpName::src0_modifiers) ||
1411 hasModifiersSet(UseMI, AMDGPU::OpName::src1_modifiers) ||
1412 hasModifiersSet(UseMI, AMDGPU::OpName::src2_modifiers)) {
1416 const MachineOperand &ImmOp = DefMI.getOperand(1);
1418 // If this is a free constant, there's no reason to do this.
1419 // TODO: We could fold this here instead of letting SIFoldOperands do it
1421 if (isInlineConstant(ImmOp, 4))
1424 MachineOperand *Src0 = getNamedOperand(UseMI, AMDGPU::OpName::src0);
1425 MachineOperand *Src1 = getNamedOperand(UseMI, AMDGPU::OpName::src1);
1426 MachineOperand *Src2 = getNamedOperand(UseMI, AMDGPU::OpName::src2);
1428 // Multiplied part is the constant: Use v_madmk_{f16, f32}.
1429 // We should only expect these to be on src0 due to canonicalizations.
1430 if (Src0->isReg() && Src0->getReg() == Reg) {
1431 if (!Src1->isReg() || RI.isSGPRClass(MRI->getRegClass(Src1->getReg())))
1434 if (!Src2->isReg() || RI.isSGPRClass(MRI->getRegClass(Src2->getReg())))
1437 // We need to swap operands 0 and 1 since madmk constant is at operand 1.
1439 const int64_t Imm = DefMI.getOperand(1).getImm();
1441 // FIXME: This would be a lot easier if we could return a new instruction
1442 // instead of having to modify in place.
1444 // Remove these first since they are at the end.
1445 UseMI.RemoveOperand(
1446 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::omod));
1447 UseMI.RemoveOperand(
1448 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::clamp));
1450 unsigned Src1Reg = Src1->getReg();
1451 unsigned Src1SubReg = Src1->getSubReg();
1452 Src0->setReg(Src1Reg);
1453 Src0->setSubReg(Src1SubReg);
1454 Src0->setIsKill(Src1->isKill());
1456 if (Opc == AMDGPU::V_MAC_F32_e64 ||
1457 Opc == AMDGPU::V_MAC_F16_e64)
1458 UseMI.untieRegOperand(
1459 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2));
1461 Src1->ChangeToImmediate(Imm);
1463 removeModOperands(UseMI);
1464 UseMI.setDesc(get(IsF32 ? AMDGPU::V_MADMK_F32 : AMDGPU::V_MADMK_F16));
1466 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
1468 DefMI.eraseFromParent();
1473 // Added part is the constant: Use v_madak_{f16, f32}.
1474 if (Src2->isReg() && Src2->getReg() == Reg) {
1475 // Not allowed to use constant bus for another operand.
1476 // We can however allow an inline immediate as src0.
1477 if (!Src0->isImm() &&
1478 (Src0->isReg() && RI.isSGPRClass(MRI->getRegClass(Src0->getReg()))))
1481 if (!Src1->isReg() || RI.isSGPRClass(MRI->getRegClass(Src1->getReg())))
1484 const int64_t Imm = DefMI.getOperand(1).getImm();
1486 // FIXME: This would be a lot easier if we could return a new instruction
1487 // instead of having to modify in place.
1489 // Remove these first since they are at the end.
1490 UseMI.RemoveOperand(
1491 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::omod));
1492 UseMI.RemoveOperand(
1493 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::clamp));
1495 if (Opc == AMDGPU::V_MAC_F32_e64 ||
1496 Opc == AMDGPU::V_MAC_F16_e64)
1497 UseMI.untieRegOperand(
1498 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2));
1500 // ChangingToImmediate adds Src2 back to the instruction.
1501 Src2->ChangeToImmediate(Imm);
1503 // These come before src2.
1504 removeModOperands(UseMI);
1505 UseMI.setDesc(get(IsF32 ? AMDGPU::V_MADAK_F32 : AMDGPU::V_MADAK_F16));
1507 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
1509 DefMI.eraseFromParent();
1518 static bool offsetsDoNotOverlap(int WidthA, int OffsetA,
1519 int WidthB, int OffsetB) {
1520 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB;
1521 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA;
1522 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
1523 return LowOffset + LowWidth <= HighOffset;
1526 bool SIInstrInfo::checkInstOffsetsDoNotOverlap(MachineInstr &MIa,
1527 MachineInstr &MIb) const {
1528 unsigned BaseReg0, BaseReg1;
1529 int64_t Offset0, Offset1;
1531 if (getMemOpBaseRegImmOfs(MIa, BaseReg0, Offset0, &RI) &&
1532 getMemOpBaseRegImmOfs(MIb, BaseReg1, Offset1, &RI)) {
1534 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand()) {
1535 // FIXME: Handle ds_read2 / ds_write2.
1538 unsigned Width0 = (*MIa.memoperands_begin())->getSize();
1539 unsigned Width1 = (*MIb.memoperands_begin())->getSize();
1540 if (BaseReg0 == BaseReg1 &&
1541 offsetsDoNotOverlap(Width0, Offset0, Width1, Offset1)) {
1549 bool SIInstrInfo::areMemAccessesTriviallyDisjoint(MachineInstr &MIa,
1551 AliasAnalysis *AA) const {
1552 assert((MIa.mayLoad() || MIa.mayStore()) &&
1553 "MIa must load from or modify a memory location");
1554 assert((MIb.mayLoad() || MIb.mayStore()) &&
1555 "MIb must load from or modify a memory location");
1557 if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects())
1560 // XXX - Can we relax this between address spaces?
1561 if (MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef())
1564 if (AA && MIa.hasOneMemOperand() && MIb.hasOneMemOperand()) {
1565 const MachineMemOperand *MMOa = *MIa.memoperands_begin();
1566 const MachineMemOperand *MMOb = *MIb.memoperands_begin();
1567 if (MMOa->getValue() && MMOb->getValue()) {
1568 MemoryLocation LocA(MMOa->getValue(), MMOa->getSize(), MMOa->getAAInfo());
1569 MemoryLocation LocB(MMOb->getValue(), MMOb->getSize(), MMOb->getAAInfo());
1570 if (!AA->alias(LocA, LocB))
1575 // TODO: Should we check the address space from the MachineMemOperand? That
1576 // would allow us to distinguish objects we know don't alias based on the
1577 // underlying address space, even if it was lowered to a different one,
1578 // e.g. private accesses lowered to use MUBUF instructions on a scratch
1582 return checkInstOffsetsDoNotOverlap(MIa, MIb);
1584 return !isFLAT(MIb);
1587 if (isMUBUF(MIa) || isMTBUF(MIa)) {
1588 if (isMUBUF(MIb) || isMTBUF(MIb))
1589 return checkInstOffsetsDoNotOverlap(MIa, MIb);
1591 return !isFLAT(MIb) && !isSMRD(MIb);
1596 return checkInstOffsetsDoNotOverlap(MIa, MIb);
1598 return !isFLAT(MIb) && !isMUBUF(MIa) && !isMTBUF(MIa);
1603 return checkInstOffsetsDoNotOverlap(MIa, MIb);
1611 MachineInstr *SIInstrInfo::convertToThreeAddress(MachineFunction::iterator &MBB,
1613 LiveVariables *LV) const {
1616 switch (MI.getOpcode()) {
1619 case AMDGPU::V_MAC_F16_e64:
1621 case AMDGPU::V_MAC_F32_e64:
1623 case AMDGPU::V_MAC_F16_e32:
1625 case AMDGPU::V_MAC_F32_e32: {
1626 const MachineOperand *Src0 = getNamedOperand(MI, AMDGPU::OpName::src0);
1627 if (Src0->isImm() && !isInlineConstant(*Src0, 4))
1633 const MachineOperand *Dst = getNamedOperand(MI, AMDGPU::OpName::vdst);
1634 const MachineOperand *Src0 = getNamedOperand(MI, AMDGPU::OpName::src0);
1635 const MachineOperand *Src1 = getNamedOperand(MI, AMDGPU::OpName::src1);
1636 const MachineOperand *Src2 = getNamedOperand(MI, AMDGPU::OpName::src2);
1638 return BuildMI(*MBB, MI, MI.getDebugLoc(),
1639 get(IsF16 ? AMDGPU::V_MAD_F16 : AMDGPU::V_MAD_F32))
1641 .addImm(0) // Src0 mods
1643 .addImm(0) // Src1 mods
1645 .addImm(0) // Src mods
1651 // It's not generally safe to move VALU instructions across these since it will
1652 // start using the register as a base index rather than directly.
1653 // XXX - Why isn't hasSideEffects sufficient for these?
1654 static bool changesVGPRIndexingMode(const MachineInstr &MI) {
1655 switch (MI.getOpcode()) {
1656 case AMDGPU::S_SET_GPR_IDX_ON:
1657 case AMDGPU::S_SET_GPR_IDX_MODE:
1658 case AMDGPU::S_SET_GPR_IDX_OFF:
1665 bool SIInstrInfo::isSchedulingBoundary(const MachineInstr &MI,
1666 const MachineBasicBlock *MBB,
1667 const MachineFunction &MF) const {
1668 // XXX - Do we want the SP check in the base implementation?
1670 // Target-independent instructions do not have an implicit-use of EXEC, even
1671 // when they operate on VGPRs. Treating EXEC modifications as scheduling
1672 // boundaries prevents incorrect movements of such instructions.
1673 return TargetInstrInfo::isSchedulingBoundary(MI, MBB, MF) ||
1674 MI.modifiesRegister(AMDGPU::EXEC, &RI) ||
1675 changesVGPRIndexingMode(MI);
1678 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
1679 int64_t SVal = Imm.getSExtValue();
1680 if (SVal >= -16 && SVal <= 64)
1683 if (Imm.getBitWidth() == 64) {
1684 uint64_t Val = Imm.getZExtValue();
1685 return (DoubleToBits(0.0) == Val) ||
1686 (DoubleToBits(1.0) == Val) ||
1687 (DoubleToBits(-1.0) == Val) ||
1688 (DoubleToBits(0.5) == Val) ||
1689 (DoubleToBits(-0.5) == Val) ||
1690 (DoubleToBits(2.0) == Val) ||
1691 (DoubleToBits(-2.0) == Val) ||
1692 (DoubleToBits(4.0) == Val) ||
1693 (DoubleToBits(-4.0) == Val) ||
1694 (ST.hasInv2PiInlineImm() && Val == 0x3fc45f306dc9c882);
1697 // The actual type of the operand does not seem to matter as long
1698 // as the bits match one of the inline immediate values. For example:
1700 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
1701 // so it is a legal inline immediate.
1703 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
1704 // floating-point, so it is a legal inline immediate.
1705 uint32_t Val = Imm.getZExtValue();
1707 return (FloatToBits(0.0f) == Val) ||
1708 (FloatToBits(1.0f) == Val) ||
1709 (FloatToBits(-1.0f) == Val) ||
1710 (FloatToBits(0.5f) == Val) ||
1711 (FloatToBits(-0.5f) == Val) ||
1712 (FloatToBits(2.0f) == Val) ||
1713 (FloatToBits(-2.0f) == Val) ||
1714 (FloatToBits(4.0f) == Val) ||
1715 (FloatToBits(-4.0f) == Val) ||
1716 (ST.hasInv2PiInlineImm() && Val == 0x3e22f983);
1719 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO,
1720 unsigned OpSize) const {
1722 // MachineOperand provides no way to tell the true operand size, since it
1723 // only records a 64-bit value. We need to know the size to determine if a
1724 // 32-bit floating point immediate bit pattern is legal for an integer
1725 // immediate. It would be for any 32-bit integer operand, but would not be
1726 // for a 64-bit one.
1728 unsigned BitSize = 8 * OpSize;
1729 return isInlineConstant(APInt(BitSize, MO.getImm(), true));
1735 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO,
1736 unsigned OpSize) const {
1737 return MO.isImm() && !isInlineConstant(MO, OpSize);
1740 bool SIInstrInfo::isLiteralConstantLike(const MachineOperand &MO,
1741 unsigned OpSize) const {
1742 switch (MO.getType()) {
1743 case MachineOperand::MO_Register:
1745 case MachineOperand::MO_Immediate:
1746 return !isInlineConstant(MO, OpSize);
1747 case MachineOperand::MO_FrameIndex:
1748 case MachineOperand::MO_MachineBasicBlock:
1749 case MachineOperand::MO_ExternalSymbol:
1750 case MachineOperand::MO_GlobalAddress:
1751 case MachineOperand::MO_MCSymbol:
1754 llvm_unreachable("unexpected operand type");
1758 static bool compareMachineOp(const MachineOperand &Op0,
1759 const MachineOperand &Op1) {
1760 if (Op0.getType() != Op1.getType())
1763 switch (Op0.getType()) {
1764 case MachineOperand::MO_Register:
1765 return Op0.getReg() == Op1.getReg();
1766 case MachineOperand::MO_Immediate:
1767 return Op0.getImm() == Op1.getImm();
1769 llvm_unreachable("Didn't expect to be comparing these operand types");
1773 bool SIInstrInfo::isImmOperandLegal(const MachineInstr &MI, unsigned OpNo,
1774 const MachineOperand &MO) const {
1775 const MCOperandInfo &OpInfo = get(MI.getOpcode()).OpInfo[OpNo];
1777 assert(MO.isImm() || MO.isTargetIndex() || MO.isFI());
1779 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
1782 if (OpInfo.RegClass < 0)
1785 unsigned OpSize = RI.getRegClass(OpInfo.RegClass)->getSize();
1786 if (isLiteralConstant(MO, OpSize))
1787 return RI.opCanUseLiteralConstant(OpInfo.OperandType);
1789 return RI.opCanUseInlineConstant(OpInfo.OperandType);
1792 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
1793 int Op32 = AMDGPU::getVOPe32(Opcode);
1797 return pseudoToMCOpcode(Op32) != -1;
1800 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
1801 // The src0_modifier operand is present on all instructions
1802 // that have modifiers.
1804 return AMDGPU::getNamedOperandIdx(Opcode,
1805 AMDGPU::OpName::src0_modifiers) != -1;
1808 bool SIInstrInfo::hasModifiersSet(const MachineInstr &MI,
1809 unsigned OpName) const {
1810 const MachineOperand *Mods = getNamedOperand(MI, OpName);
1811 return Mods && Mods->getImm();
1814 bool SIInstrInfo::usesConstantBus(const MachineRegisterInfo &MRI,
1815 const MachineOperand &MO,
1816 unsigned OpSize) const {
1817 // Literal constants use the constant bus.
1818 if (isLiteralConstant(MO, OpSize))
1821 if (!MO.isReg() || !MO.isUse())
1824 if (TargetRegisterInfo::isVirtualRegister(MO.getReg()))
1825 return RI.isSGPRClass(MRI.getRegClass(MO.getReg()));
1827 // FLAT_SCR is just an SGPR pair.
1828 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
1831 // EXEC register uses the constant bus.
1832 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
1835 // SGPRs use the constant bus
1836 return (MO.getReg() == AMDGPU::VCC || MO.getReg() == AMDGPU::M0 ||
1837 (!MO.isImplicit() &&
1838 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
1839 AMDGPU::SGPR_64RegClass.contains(MO.getReg()))));
1842 static unsigned findImplicitSGPRRead(const MachineInstr &MI) {
1843 for (const MachineOperand &MO : MI.implicit_operands()) {
1844 // We only care about reads.
1848 switch (MO.getReg()) {
1851 case AMDGPU::FLAT_SCR:
1859 return AMDGPU::NoRegister;
1862 static bool shouldReadExec(const MachineInstr &MI) {
1863 if (SIInstrInfo::isVALU(MI)) {
1864 switch (MI.getOpcode()) {
1865 case AMDGPU::V_READLANE_B32:
1866 case AMDGPU::V_READLANE_B32_si:
1867 case AMDGPU::V_READLANE_B32_vi:
1868 case AMDGPU::V_WRITELANE_B32:
1869 case AMDGPU::V_WRITELANE_B32_si:
1870 case AMDGPU::V_WRITELANE_B32_vi:
1877 if (SIInstrInfo::isGenericOpcode(MI.getOpcode()) ||
1878 SIInstrInfo::isSALU(MI) ||
1879 SIInstrInfo::isSMRD(MI))
1885 static bool isSubRegOf(const SIRegisterInfo &TRI,
1886 const MachineOperand &SuperVec,
1887 const MachineOperand &SubReg) {
1888 if (TargetRegisterInfo::isPhysicalRegister(SubReg.getReg()))
1889 return TRI.isSubRegister(SuperVec.getReg(), SubReg.getReg());
1891 return SubReg.getSubReg() != AMDGPU::NoSubRegister &&
1892 SubReg.getReg() == SuperVec.getReg();
1895 bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
1896 StringRef &ErrInfo) const {
1897 uint16_t Opcode = MI.getOpcode();
1898 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1899 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
1900 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
1901 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
1903 // Make sure the number of operands is correct.
1904 const MCInstrDesc &Desc = get(Opcode);
1905 if (!Desc.isVariadic() &&
1906 Desc.getNumOperands() != MI.getNumExplicitOperands()) {
1907 ErrInfo = "Instruction has wrong number of operands.";
1911 if (MI.isInlineAsm()) {
1912 // Verify register classes for inlineasm constraints.
1913 for (unsigned I = InlineAsm::MIOp_FirstOperand, E = MI.getNumOperands();
1915 const TargetRegisterClass *RC = MI.getRegClassConstraint(I, this, &RI);
1919 const MachineOperand &Op = MI.getOperand(I);
1923 unsigned Reg = Op.getReg();
1924 if (!TargetRegisterInfo::isVirtualRegister(Reg) && !RC->contains(Reg)) {
1925 ErrInfo = "inlineasm operand has incorrect register class.";
1933 // Make sure the register classes are correct.
1934 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
1935 if (MI.getOperand(i).isFPImm()) {
1936 ErrInfo = "FPImm Machine Operands are not supported. ISel should bitcast "
1937 "all fp values to integers.";
1941 int RegClass = Desc.OpInfo[i].RegClass;
1943 switch (Desc.OpInfo[i].OperandType) {
1944 case MCOI::OPERAND_REGISTER:
1945 if (MI.getOperand(i).isImm()) {
1946 ErrInfo = "Illegal immediate value for operand.";
1950 case AMDGPU::OPERAND_REG_IMM32_INT:
1951 case AMDGPU::OPERAND_REG_IMM32_FP:
1953 case AMDGPU::OPERAND_REG_INLINE_C_INT:
1954 case AMDGPU::OPERAND_REG_INLINE_C_FP:
1955 if (isLiteralConstant(MI.getOperand(i),
1956 RI.getRegClass(RegClass)->getSize())) {
1957 ErrInfo = "Illegal immediate value for operand.";
1961 case MCOI::OPERAND_IMMEDIATE:
1962 case AMDGPU::OPERAND_KIMM32:
1963 // Check if this operand is an immediate.
1964 // FrameIndex operands will be replaced by immediates, so they are
1966 if (!MI.getOperand(i).isImm() && !MI.getOperand(i).isFI()) {
1967 ErrInfo = "Expected immediate, but got non-immediate";
1975 if (!MI.getOperand(i).isReg())
1978 if (RegClass != -1) {
1979 unsigned Reg = MI.getOperand(i).getReg();
1980 if (Reg == AMDGPU::NoRegister ||
1981 TargetRegisterInfo::isVirtualRegister(Reg))
1984 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
1985 if (!RC->contains(Reg)) {
1986 ErrInfo = "Operand has incorrect register class.";
1993 if (isVOP1(MI) || isVOP2(MI) || isVOP3(MI) || isVOPC(MI)) {
1994 // Only look at the true operands. Only a real operand can use the constant
1995 // bus, and we don't want to check pseudo-operands like the source modifier
1997 const int OpIndices[] = { Src0Idx, Src1Idx, Src2Idx };
1999 unsigned ConstantBusCount = 0;
2001 if (AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::imm) != -1)
2004 unsigned SGPRUsed = findImplicitSGPRRead(MI);
2005 if (SGPRUsed != AMDGPU::NoRegister)
2008 for (int OpIdx : OpIndices) {
2011 const MachineOperand &MO = MI.getOperand(OpIdx);
2012 if (usesConstantBus(MRI, MO, getOpSize(Opcode, OpIdx))) {
2014 if (MO.getReg() != SGPRUsed)
2016 SGPRUsed = MO.getReg();
2022 if (ConstantBusCount > 1) {
2023 ErrInfo = "VOP* instruction uses the constant bus more than once";
2028 // Verify misc. restrictions on specific instructions.
2029 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
2030 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
2031 const MachineOperand &Src0 = MI.getOperand(Src0Idx);
2032 const MachineOperand &Src1 = MI.getOperand(Src1Idx);
2033 const MachineOperand &Src2 = MI.getOperand(Src2Idx);
2034 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
2035 if (!compareMachineOp(Src0, Src1) &&
2036 !compareMachineOp(Src0, Src2)) {
2037 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
2044 int64_t Imm = getNamedOperand(MI, AMDGPU::OpName::simm16)->getImm();
2045 if (sopkIsZext(MI)) {
2046 if (!isUInt<16>(Imm)) {
2047 ErrInfo = "invalid immediate for SOPK instruction";
2051 if (!isInt<16>(Imm)) {
2052 ErrInfo = "invalid immediate for SOPK instruction";
2058 if (Desc.getOpcode() == AMDGPU::V_MOVRELS_B32_e32 ||
2059 Desc.getOpcode() == AMDGPU::V_MOVRELS_B32_e64 ||
2060 Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e32 ||
2061 Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e64) {
2062 const bool IsDst = Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e32 ||
2063 Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e64;
2065 const unsigned StaticNumOps = Desc.getNumOperands() +
2066 Desc.getNumImplicitUses();
2067 const unsigned NumImplicitOps = IsDst ? 2 : 1;
2069 // Allow additional implicit operands. This allows a fixup done by the post
2070 // RA scheduler where the main implicit operand is killed and implicit-defs
2071 // are added for sub-registers that remain live after this instruction.
2072 if (MI.getNumOperands() < StaticNumOps + NumImplicitOps) {
2073 ErrInfo = "missing implicit register operands";
2077 const MachineOperand *Dst = getNamedOperand(MI, AMDGPU::OpName::vdst);
2079 if (!Dst->isUse()) {
2080 ErrInfo = "v_movreld_b32 vdst should be a use operand";
2085 if (!MI.isRegTiedToUseOperand(StaticNumOps, &UseOpIdx) ||
2086 UseOpIdx != StaticNumOps + 1) {
2087 ErrInfo = "movrel implicit operands should be tied";
2092 const MachineOperand &Src0 = MI.getOperand(Src0Idx);
2093 const MachineOperand &ImpUse
2094 = MI.getOperand(StaticNumOps + NumImplicitOps - 1);
2095 if (!ImpUse.isReg() || !ImpUse.isUse() ||
2096 !isSubRegOf(RI, ImpUse, IsDst ? *Dst : Src0)) {
2097 ErrInfo = "src0 should be subreg of implicit vector use";
2102 // Make sure we aren't losing exec uses in the td files. This mostly requires
2103 // being careful when using let Uses to try to add other use registers.
2104 if (shouldReadExec(MI)) {
2105 if (!MI.hasRegisterImplicitUseOperand(AMDGPU::EXEC)) {
2106 ErrInfo = "VALU instruction does not implicitly read exec mask";
2112 if (MI.mayStore()) {
2113 // The register offset form of scalar stores may only use m0 as the
2114 // soffset register.
2115 const MachineOperand *Soff = getNamedOperand(MI, AMDGPU::OpName::soff);
2116 if (Soff && Soff->getReg() != AMDGPU::M0) {
2117 ErrInfo = "scalar stores must use m0 as offset register";
2126 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
2127 switch (MI.getOpcode()) {
2128 default: return AMDGPU::INSTRUCTION_LIST_END;
2129 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
2130 case AMDGPU::COPY: return AMDGPU::COPY;
2131 case AMDGPU::PHI: return AMDGPU::PHI;
2132 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
2133 case AMDGPU::S_MOV_B32:
2134 return MI.getOperand(1).isReg() ?
2135 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
2136 case AMDGPU::S_ADD_I32:
2137 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
2138 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
2139 case AMDGPU::S_SUB_I32:
2140 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
2141 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
2142 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
2143 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e64;
2144 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e64;
2145 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e64;
2146 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e64;
2147 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e64;
2148 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e64;
2149 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e64;
2150 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
2151 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
2152 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
2153 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
2154 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
2155 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
2156 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
2157 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
2158 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
2159 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
2160 case AMDGPU::S_BFM_B32: return AMDGPU::V_BFM_B32_e64;
2161 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
2162 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
2163 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
2164 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
2165 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
2166 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
2167 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
2168 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
2169 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
2170 case AMDGPU::S_CMP_EQ_U32: return AMDGPU::V_CMP_EQ_U32_e32;
2171 case AMDGPU::S_CMP_LG_U32: return AMDGPU::V_CMP_NE_U32_e32;
2172 case AMDGPU::S_CMP_GT_U32: return AMDGPU::V_CMP_GT_U32_e32;
2173 case AMDGPU::S_CMP_GE_U32: return AMDGPU::V_CMP_GE_U32_e32;
2174 case AMDGPU::S_CMP_LT_U32: return AMDGPU::V_CMP_LT_U32_e32;
2175 case AMDGPU::S_CMP_LE_U32: return AMDGPU::V_CMP_LE_U32_e32;
2176 case AMDGPU::S_CMP_EQ_U64: return AMDGPU::V_CMP_EQ_U64_e32;
2177 case AMDGPU::S_CMP_LG_U64: return AMDGPU::V_CMP_NE_U64_e32;
2178 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e64;
2179 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
2180 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
2181 case AMDGPU::S_FLBIT_I32: return AMDGPU::V_FFBH_I32_e64;
2182 case AMDGPU::S_CBRANCH_SCC0: return AMDGPU::S_CBRANCH_VCCZ;
2183 case AMDGPU::S_CBRANCH_SCC1: return AMDGPU::S_CBRANCH_VCCNZ;
2187 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
2188 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
2191 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
2192 unsigned OpNo) const {
2193 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
2194 const MCInstrDesc &Desc = get(MI.getOpcode());
2195 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
2196 Desc.OpInfo[OpNo].RegClass == -1) {
2197 unsigned Reg = MI.getOperand(OpNo).getReg();
2199 if (TargetRegisterInfo::isVirtualRegister(Reg))
2200 return MRI.getRegClass(Reg);
2201 return RI.getPhysRegClass(Reg);
2204 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
2205 return RI.getRegClass(RCID);
2208 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
2209 switch (MI.getOpcode()) {
2211 case AMDGPU::REG_SEQUENCE:
2213 case AMDGPU::INSERT_SUBREG:
2214 return RI.hasVGPRs(getOpRegClass(MI, 0));
2216 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
2220 void SIInstrInfo::legalizeOpWithMove(MachineInstr &MI, unsigned OpIdx) const {
2221 MachineBasicBlock::iterator I = MI;
2222 MachineBasicBlock *MBB = MI.getParent();
2223 MachineOperand &MO = MI.getOperand(OpIdx);
2224 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
2225 unsigned RCID = get(MI.getOpcode()).OpInfo[OpIdx].RegClass;
2226 const TargetRegisterClass *RC = RI.getRegClass(RCID);
2227 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
2229 Opcode = AMDGPU::COPY;
2230 else if (RI.isSGPRClass(RC))
2231 Opcode = AMDGPU::S_MOV_B32;
2233 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
2234 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC))
2235 VRC = &AMDGPU::VReg_64RegClass;
2237 VRC = &AMDGPU::VGPR_32RegClass;
2239 unsigned Reg = MRI.createVirtualRegister(VRC);
2240 DebugLoc DL = MBB->findDebugLoc(I);
2241 BuildMI(*MI.getParent(), I, DL, get(Opcode), Reg).addOperand(MO);
2242 MO.ChangeToRegister(Reg, false);
2245 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
2246 MachineRegisterInfo &MRI,
2247 MachineOperand &SuperReg,
2248 const TargetRegisterClass *SuperRC,
2250 const TargetRegisterClass *SubRC)
2252 MachineBasicBlock *MBB = MI->getParent();
2253 DebugLoc DL = MI->getDebugLoc();
2254 unsigned SubReg = MRI.createVirtualRegister(SubRC);
2256 if (SuperReg.getSubReg() == AMDGPU::NoSubRegister) {
2257 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
2258 .addReg(SuperReg.getReg(), 0, SubIdx);
2262 // Just in case the super register is itself a sub-register, copy it to a new
2263 // value so we don't need to worry about merging its subreg index with the
2264 // SubIdx passed to this function. The register coalescer should be able to
2265 // eliminate this extra copy.
2266 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
2268 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), NewSuperReg)
2269 .addReg(SuperReg.getReg(), 0, SuperReg.getSubReg());
2271 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
2272 .addReg(NewSuperReg, 0, SubIdx);
2277 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
2278 MachineBasicBlock::iterator MII,
2279 MachineRegisterInfo &MRI,
2281 const TargetRegisterClass *SuperRC,
2283 const TargetRegisterClass *SubRC) const {
2285 if (SubIdx == AMDGPU::sub0)
2286 return MachineOperand::CreateImm(static_cast<int32_t>(Op.getImm()));
2287 if (SubIdx == AMDGPU::sub1)
2288 return MachineOperand::CreateImm(static_cast<int32_t>(Op.getImm() >> 32));
2290 llvm_unreachable("Unhandled register index for immediate");
2293 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
2295 return MachineOperand::CreateReg(SubReg, false);
2298 // Change the order of operands from (0, 1, 2) to (0, 2, 1)
2299 void SIInstrInfo::swapOperands(MachineInstr &Inst) const {
2300 assert(Inst.getNumExplicitOperands() == 3);
2301 MachineOperand Op1 = Inst.getOperand(1);
2302 Inst.RemoveOperand(1);
2303 Inst.addOperand(Op1);
2306 bool SIInstrInfo::isLegalRegOperand(const MachineRegisterInfo &MRI,
2307 const MCOperandInfo &OpInfo,
2308 const MachineOperand &MO) const {
2312 unsigned Reg = MO.getReg();
2313 const TargetRegisterClass *RC =
2314 TargetRegisterInfo::isVirtualRegister(Reg) ?
2315 MRI.getRegClass(Reg) :
2316 RI.getPhysRegClass(Reg);
2318 const SIRegisterInfo *TRI =
2319 static_cast<const SIRegisterInfo*>(MRI.getTargetRegisterInfo());
2320 RC = TRI->getSubRegClass(RC, MO.getSubReg());
2322 // In order to be legal, the common sub-class must be equal to the
2323 // class of the current operand. For example:
2325 // v_mov_b32 s0 ; Operand defined as vsrc_b32
2326 // ; RI.getCommonSubClass(s0,vsrc_b32) = sgpr ; LEGAL
2328 // s_sendmsg 0, s0 ; Operand defined as m0reg
2329 // ; RI.getCommonSubClass(s0,m0reg) = m0reg ; NOT LEGAL
2331 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass)) == RC;
2334 bool SIInstrInfo::isLegalVSrcOperand(const MachineRegisterInfo &MRI,
2335 const MCOperandInfo &OpInfo,
2336 const MachineOperand &MO) const {
2338 return isLegalRegOperand(MRI, OpInfo, MO);
2340 // Handle non-register types that are treated like immediates.
2341 assert(MO.isImm() || MO.isTargetIndex() || MO.isFI());
2345 bool SIInstrInfo::isOperandLegal(const MachineInstr &MI, unsigned OpIdx,
2346 const MachineOperand *MO) const {
2347 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
2348 const MCInstrDesc &InstDesc = MI.getDesc();
2349 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
2350 const TargetRegisterClass *DefinedRC =
2351 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
2353 MO = &MI.getOperand(OpIdx);
2355 if (isVALU(MI) && usesConstantBus(MRI, *MO, DefinedRC->getSize())) {
2357 RegSubRegPair SGPRUsed;
2359 SGPRUsed = RegSubRegPair(MO->getReg(), MO->getSubReg());
2361 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
2364 const MachineOperand &Op = MI.getOperand(i);
2366 if ((Op.getReg() != SGPRUsed.Reg || Op.getSubReg() != SGPRUsed.SubReg) &&
2367 usesConstantBus(MRI, Op, getOpSize(MI, i))) {
2370 } else if (InstDesc.OpInfo[i].OperandType == AMDGPU::OPERAND_KIMM32) {
2378 return isLegalRegOperand(MRI, OpInfo, *MO);
2381 // Handle non-register types that are treated like immediates.
2382 assert(MO->isImm() || MO->isTargetIndex() || MO->isFI());
2385 // This operand expects an immediate.
2389 return isImmOperandLegal(MI, OpIdx, *MO);
2392 void SIInstrInfo::legalizeOperandsVOP2(MachineRegisterInfo &MRI,
2393 MachineInstr &MI) const {
2394 unsigned Opc = MI.getOpcode();
2395 const MCInstrDesc &InstrDesc = get(Opc);
2397 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
2398 MachineOperand &Src1 = MI.getOperand(Src1Idx);
2400 // If there is an implicit SGPR use such as VCC use for v_addc_u32/v_subb_u32
2401 // we need to only have one constant bus use.
2403 // Note we do not need to worry about literal constants here. They are
2404 // disabled for the operand type for instructions because they will always
2405 // violate the one constant bus use rule.
2406 bool HasImplicitSGPR = findImplicitSGPRRead(MI) != AMDGPU::NoRegister;
2407 if (HasImplicitSGPR) {
2408 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
2409 MachineOperand &Src0 = MI.getOperand(Src0Idx);
2411 if (Src0.isReg() && RI.isSGPRReg(MRI, Src0.getReg()))
2412 legalizeOpWithMove(MI, Src0Idx);
2415 // VOP2 src0 instructions support all operand types, so we don't need to check
2416 // their legality. If src1 is already legal, we don't need to do anything.
2417 if (isLegalRegOperand(MRI, InstrDesc.OpInfo[Src1Idx], Src1))
2420 // We do not use commuteInstruction here because it is too aggressive and will
2421 // commute if it is possible. We only want to commute here if it improves
2422 // legality. This can be called a fairly large number of times so don't waste
2423 // compile time pointlessly swapping and checking legality again.
2424 if (HasImplicitSGPR || !MI.isCommutable()) {
2425 legalizeOpWithMove(MI, Src1Idx);
2429 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
2430 MachineOperand &Src0 = MI.getOperand(Src0Idx);
2432 // If src0 can be used as src1, commuting will make the operands legal.
2433 // Otherwise we have to give up and insert a move.
2435 // TODO: Other immediate-like operand kinds could be commuted if there was a
2436 // MachineOperand::ChangeTo* for them.
2437 if ((!Src1.isImm() && !Src1.isReg()) ||
2438 !isLegalRegOperand(MRI, InstrDesc.OpInfo[Src1Idx], Src0)) {
2439 legalizeOpWithMove(MI, Src1Idx);
2443 int CommutedOpc = commuteOpcode(MI);
2444 if (CommutedOpc == -1) {
2445 legalizeOpWithMove(MI, Src1Idx);
2449 MI.setDesc(get(CommutedOpc));
2451 unsigned Src0Reg = Src0.getReg();
2452 unsigned Src0SubReg = Src0.getSubReg();
2453 bool Src0Kill = Src0.isKill();
2456 Src0.ChangeToImmediate(Src1.getImm());
2457 else if (Src1.isReg()) {
2458 Src0.ChangeToRegister(Src1.getReg(), false, false, Src1.isKill());
2459 Src0.setSubReg(Src1.getSubReg());
2461 llvm_unreachable("Should only have register or immediate operands");
2463 Src1.ChangeToRegister(Src0Reg, false, false, Src0Kill);
2464 Src1.setSubReg(Src0SubReg);
2467 // Legalize VOP3 operands. Because all operand types are supported for any
2468 // operand, and since literal constants are not allowed and should never be
2469 // seen, we only need to worry about inserting copies if we use multiple SGPR
2471 void SIInstrInfo::legalizeOperandsVOP3(MachineRegisterInfo &MRI,
2472 MachineInstr &MI) const {
2473 unsigned Opc = MI.getOpcode();
2476 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0),
2477 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1),
2478 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2)
2481 // Find the one SGPR operand we are allowed to use.
2482 unsigned SGPRReg = findUsedSGPR(MI, VOP3Idx);
2484 for (unsigned i = 0; i < 3; ++i) {
2485 int Idx = VOP3Idx[i];
2488 MachineOperand &MO = MI.getOperand(Idx);
2490 // We should never see a VOP3 instruction with an illegal immediate operand.
2494 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
2495 continue; // VGPRs are legal
2497 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
2498 SGPRReg = MO.getReg();
2499 // We can use one SGPR in each VOP3 instruction.
2503 // If we make it this far, then the operand is not legal and we must
2505 legalizeOpWithMove(MI, Idx);
2509 unsigned SIInstrInfo::readlaneVGPRToSGPR(unsigned SrcReg, MachineInstr &UseMI,
2510 MachineRegisterInfo &MRI) const {
2511 const TargetRegisterClass *VRC = MRI.getRegClass(SrcReg);
2512 const TargetRegisterClass *SRC = RI.getEquivalentSGPRClass(VRC);
2513 unsigned DstReg = MRI.createVirtualRegister(SRC);
2514 unsigned SubRegs = VRC->getSize() / 4;
2516 SmallVector<unsigned, 8> SRegs;
2517 for (unsigned i = 0; i < SubRegs; ++i) {
2518 unsigned SGPR = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
2519 BuildMI(*UseMI.getParent(), UseMI, UseMI.getDebugLoc(),
2520 get(AMDGPU::V_READFIRSTLANE_B32), SGPR)
2521 .addReg(SrcReg, 0, RI.getSubRegFromChannel(i));
2522 SRegs.push_back(SGPR);
2525 MachineInstrBuilder MIB =
2526 BuildMI(*UseMI.getParent(), UseMI, UseMI.getDebugLoc(),
2527 get(AMDGPU::REG_SEQUENCE), DstReg);
2528 for (unsigned i = 0; i < SubRegs; ++i) {
2529 MIB.addReg(SRegs[i]);
2530 MIB.addImm(RI.getSubRegFromChannel(i));
2535 void SIInstrInfo::legalizeOperandsSMRD(MachineRegisterInfo &MRI,
2536 MachineInstr &MI) const {
2538 // If the pointer is store in VGPRs, then we need to move them to
2539 // SGPRs using v_readfirstlane. This is safe because we only select
2540 // loads with uniform pointers to SMRD instruction so we know the
2541 // pointer value is uniform.
2542 MachineOperand *SBase = getNamedOperand(MI, AMDGPU::OpName::sbase);
2543 if (SBase && !RI.isSGPRClass(MRI.getRegClass(SBase->getReg()))) {
2544 unsigned SGPR = readlaneVGPRToSGPR(SBase->getReg(), MI, MRI);
2545 SBase->setReg(SGPR);
2549 void SIInstrInfo::legalizeGenericOperand(MachineBasicBlock &InsertMBB,
2550 MachineBasicBlock::iterator I,
2551 const TargetRegisterClass *DstRC,
2553 MachineRegisterInfo &MRI,
2554 const DebugLoc &DL) const {
2556 unsigned OpReg = Op.getReg();
2557 unsigned OpSubReg = Op.getSubReg();
2559 const TargetRegisterClass *OpRC = RI.getSubClassWithSubReg(
2560 RI.getRegClassForReg(MRI, OpReg), OpSubReg);
2562 // Check if operand is already the correct register class.
2566 unsigned DstReg = MRI.createVirtualRegister(DstRC);
2567 MachineInstr *Copy = BuildMI(InsertMBB, I, DL, get(AMDGPU::COPY), DstReg)
2573 MachineInstr *Def = MRI.getVRegDef(OpReg);
2577 // Try to eliminate the copy if it is copying an immediate value.
2578 if (Def->isMoveImmediate())
2579 FoldImmediate(*Copy, *Def, OpReg, &MRI);
2582 void SIInstrInfo::legalizeOperands(MachineInstr &MI) const {
2583 MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
2586 if (isVOP2(MI) || isVOPC(MI)) {
2587 legalizeOperandsVOP2(MRI, MI);
2593 legalizeOperandsVOP3(MRI, MI);
2599 legalizeOperandsSMRD(MRI, MI);
2603 // Legalize REG_SEQUENCE and PHI
2604 // The register class of the operands much be the same type as the register
2605 // class of the output.
2606 if (MI.getOpcode() == AMDGPU::PHI) {
2607 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
2608 for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2) {
2609 if (!MI.getOperand(i).isReg() ||
2610 !TargetRegisterInfo::isVirtualRegister(MI.getOperand(i).getReg()))
2612 const TargetRegisterClass *OpRC =
2613 MRI.getRegClass(MI.getOperand(i).getReg());
2614 if (RI.hasVGPRs(OpRC)) {
2621 // If any of the operands are VGPR registers, then they all most be
2622 // otherwise we will create illegal VGPR->SGPR copies when legalizing
2624 if (VRC || !RI.isSGPRClass(getOpRegClass(MI, 0))) {
2627 VRC = RI.getEquivalentVGPRClass(SRC);
2634 // Update all the operands so they have the same type.
2635 for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) {
2636 MachineOperand &Op = MI.getOperand(I);
2637 if (!Op.isReg() || !TargetRegisterInfo::isVirtualRegister(Op.getReg()))
2640 // MI is a PHI instruction.
2641 MachineBasicBlock *InsertBB = MI.getOperand(I + 1).getMBB();
2642 MachineBasicBlock::iterator Insert = InsertBB->getFirstTerminator();
2644 // Avoid creating no-op copies with the same src and dst reg class. These
2645 // confuse some of the machine passes.
2646 legalizeGenericOperand(*InsertBB, Insert, RC, Op, MRI, MI.getDebugLoc());
2650 // REG_SEQUENCE doesn't really require operand legalization, but if one has a
2651 // VGPR dest type and SGPR sources, insert copies so all operands are
2652 // VGPRs. This seems to help operand folding / the register coalescer.
2653 if (MI.getOpcode() == AMDGPU::REG_SEQUENCE) {
2654 MachineBasicBlock *MBB = MI.getParent();
2655 const TargetRegisterClass *DstRC = getOpRegClass(MI, 0);
2656 if (RI.hasVGPRs(DstRC)) {
2657 // Update all the operands so they are VGPR register classes. These may
2658 // not be the same register class because REG_SEQUENCE supports mixing
2659 // subregister index types e.g. sub0_sub1 + sub2 + sub3
2660 for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) {
2661 MachineOperand &Op = MI.getOperand(I);
2662 if (!Op.isReg() || !TargetRegisterInfo::isVirtualRegister(Op.getReg()))
2665 const TargetRegisterClass *OpRC = MRI.getRegClass(Op.getReg());
2666 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(OpRC);
2670 legalizeGenericOperand(*MBB, MI, VRC, Op, MRI, MI.getDebugLoc());
2678 // Legalize INSERT_SUBREG
2679 // src0 must have the same register class as dst
2680 if (MI.getOpcode() == AMDGPU::INSERT_SUBREG) {
2681 unsigned Dst = MI.getOperand(0).getReg();
2682 unsigned Src0 = MI.getOperand(1).getReg();
2683 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
2684 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
2685 if (DstRC != Src0RC) {
2686 MachineBasicBlock *MBB = MI.getParent();
2687 MachineOperand &Op = MI.getOperand(1);
2688 legalizeGenericOperand(*MBB, MI, DstRC, Op, MRI, MI.getDebugLoc());
2695 MachineOperand *SRsrc = getNamedOperand(MI, AMDGPU::OpName::srsrc);
2696 if (SRsrc && !RI.isSGPRClass(MRI.getRegClass(SRsrc->getReg()))) {
2697 unsigned SGPR = readlaneVGPRToSGPR(SRsrc->getReg(), MI, MRI);
2698 SRsrc->setReg(SGPR);
2701 MachineOperand *SSamp = getNamedOperand(MI, AMDGPU::OpName::ssamp);
2702 if (SSamp && !RI.isSGPRClass(MRI.getRegClass(SSamp->getReg()))) {
2703 unsigned SGPR = readlaneVGPRToSGPR(SSamp->getReg(), MI, MRI);
2704 SSamp->setReg(SGPR);
2709 // Legalize MUBUF* instructions
2710 // FIXME: If we start using the non-addr64 instructions for compute, we
2711 // may need to legalize them here.
2713 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::srsrc);
2714 if (SRsrcIdx != -1) {
2715 // We have an MUBUF instruction
2716 MachineOperand *SRsrc = &MI.getOperand(SRsrcIdx);
2717 unsigned SRsrcRC = get(MI.getOpcode()).OpInfo[SRsrcIdx].RegClass;
2718 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
2719 RI.getRegClass(SRsrcRC))) {
2720 // The operands are legal.
2721 // FIXME: We may need to legalize operands besided srsrc.
2725 MachineBasicBlock &MBB = *MI.getParent();
2727 // Extract the ptr from the resource descriptor.
2728 unsigned SRsrcPtr = buildExtractSubReg(MI, MRI, *SRsrc,
2729 &AMDGPU::VReg_128RegClass, AMDGPU::sub0_sub1, &AMDGPU::VReg_64RegClass);
2731 // Create an empty resource descriptor
2732 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
2733 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
2734 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
2735 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
2736 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
2739 BuildMI(MBB, MI, MI.getDebugLoc(), get(AMDGPU::S_MOV_B64), Zero64)
2742 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
2743 BuildMI(MBB, MI, MI.getDebugLoc(), get(AMDGPU::S_MOV_B32), SRsrcFormatLo)
2744 .addImm(RsrcDataFormat & 0xFFFFFFFF);
2746 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
2747 BuildMI(MBB, MI, MI.getDebugLoc(), get(AMDGPU::S_MOV_B32), SRsrcFormatHi)
2748 .addImm(RsrcDataFormat >> 32);
2750 // NewSRsrc = {Zero64, SRsrcFormat}
2751 BuildMI(MBB, MI, MI.getDebugLoc(), get(AMDGPU::REG_SEQUENCE), NewSRsrc)
2753 .addImm(AMDGPU::sub0_sub1)
2754 .addReg(SRsrcFormatLo)
2755 .addImm(AMDGPU::sub2)
2756 .addReg(SRsrcFormatHi)
2757 .addImm(AMDGPU::sub3);
2759 MachineOperand *VAddr = getNamedOperand(MI, AMDGPU::OpName::vaddr);
2760 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
2762 // This is already an ADDR64 instruction so we need to add the pointer
2763 // extracted from the resource descriptor to the current value of VAddr.
2764 unsigned NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2765 unsigned NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2767 // NewVaddrLo = SRsrcPtr:sub0 + VAddr:sub0
2768 DebugLoc DL = MI.getDebugLoc();
2769 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_I32_e32), NewVAddrLo)
2770 .addReg(SRsrcPtr, 0, AMDGPU::sub0)
2771 .addReg(VAddr->getReg(), 0, AMDGPU::sub0);
2773 // NewVaddrHi = SRsrcPtr:sub1 + VAddr:sub1
2774 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADDC_U32_e32), NewVAddrHi)
2775 .addReg(SRsrcPtr, 0, AMDGPU::sub1)
2776 .addReg(VAddr->getReg(), 0, AMDGPU::sub1);
2778 // NewVaddr = {NewVaddrHi, NewVaddrLo}
2779 BuildMI(MBB, MI, MI.getDebugLoc(), get(AMDGPU::REG_SEQUENCE), NewVAddr)
2781 .addImm(AMDGPU::sub0)
2783 .addImm(AMDGPU::sub1);
2785 // This instructions is the _OFFSET variant, so we need to convert it to
2787 assert(MBB.getParent()->getSubtarget<SISubtarget>().getGeneration()
2788 < SISubtarget::VOLCANIC_ISLANDS &&
2789 "FIXME: Need to emit flat atomics here");
2791 MachineOperand *VData = getNamedOperand(MI, AMDGPU::OpName::vdata);
2792 MachineOperand *Offset = getNamedOperand(MI, AMDGPU::OpName::offset);
2793 MachineOperand *SOffset = getNamedOperand(MI, AMDGPU::OpName::soffset);
2794 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI.getOpcode());
2796 // Atomics rith return have have an additional tied operand and are
2797 // missing some of the special bits.
2798 MachineOperand *VDataIn = getNamedOperand(MI, AMDGPU::OpName::vdata_in);
2799 MachineInstr *Addr64;
2802 // Regular buffer load / store.
2803 MachineInstrBuilder MIB =
2804 BuildMI(MBB, MI, MI.getDebugLoc(), get(Addr64Opcode))
2806 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
2807 // This will be replaced later
2808 // with the new value of vaddr.
2810 .addOperand(*SOffset)
2811 .addOperand(*Offset);
2813 // Atomics do not have this operand.
2814 if (const MachineOperand *GLC =
2815 getNamedOperand(MI, AMDGPU::OpName::glc)) {
2816 MIB.addImm(GLC->getImm());
2819 MIB.addImm(getNamedImmOperand(MI, AMDGPU::OpName::slc));
2821 if (const MachineOperand *TFE =
2822 getNamedOperand(MI, AMDGPU::OpName::tfe)) {
2823 MIB.addImm(TFE->getImm());
2826 MIB.setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
2829 // Atomics with return.
2830 Addr64 = BuildMI(MBB, MI, MI.getDebugLoc(), get(Addr64Opcode))
2832 .addOperand(*VDataIn)
2833 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
2834 // This will be replaced later
2835 // with the new value of vaddr.
2837 .addOperand(*SOffset)
2838 .addOperand(*Offset)
2839 .addImm(getNamedImmOperand(MI, AMDGPU::OpName::slc))
2840 .setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
2843 MI.removeFromParent();
2845 // NewVaddr = {NewVaddrHi, NewVaddrLo}
2846 BuildMI(MBB, Addr64, Addr64->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
2848 .addReg(SRsrcPtr, 0, AMDGPU::sub0)
2849 .addImm(AMDGPU::sub0)
2850 .addReg(SRsrcPtr, 0, AMDGPU::sub1)
2851 .addImm(AMDGPU::sub1);
2853 VAddr = getNamedOperand(*Addr64, AMDGPU::OpName::vaddr);
2854 SRsrc = getNamedOperand(*Addr64, AMDGPU::OpName::srsrc);
2857 // Update the instruction to use NewVaddr
2858 VAddr->setReg(NewVAddr);
2859 // Update the instruction to use NewSRsrc
2860 SRsrc->setReg(NewSRsrc);
2864 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
2865 SmallVector<MachineInstr *, 128> Worklist;
2866 Worklist.push_back(&TopInst);
2868 while (!Worklist.empty()) {
2869 MachineInstr &Inst = *Worklist.pop_back_val();
2870 MachineBasicBlock *MBB = Inst.getParent();
2871 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
2873 unsigned Opcode = Inst.getOpcode();
2874 unsigned NewOpcode = getVALUOp(Inst);
2876 // Handle some special cases
2880 case AMDGPU::S_AND_B64:
2881 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::V_AND_B32_e64);
2882 Inst.eraseFromParent();
2885 case AMDGPU::S_OR_B64:
2886 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::V_OR_B32_e64);
2887 Inst.eraseFromParent();
2890 case AMDGPU::S_XOR_B64:
2891 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::V_XOR_B32_e64);
2892 Inst.eraseFromParent();
2895 case AMDGPU::S_NOT_B64:
2896 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::V_NOT_B32_e32);
2897 Inst.eraseFromParent();
2900 case AMDGPU::S_BCNT1_I32_B64:
2901 splitScalar64BitBCNT(Worklist, Inst);
2902 Inst.eraseFromParent();
2905 case AMDGPU::S_BFE_I64: {
2906 splitScalar64BitBFE(Worklist, Inst);
2907 Inst.eraseFromParent();
2911 case AMDGPU::S_LSHL_B32:
2912 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
2913 NewOpcode = AMDGPU::V_LSHLREV_B32_e64;
2917 case AMDGPU::S_ASHR_I32:
2918 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
2919 NewOpcode = AMDGPU::V_ASHRREV_I32_e64;
2923 case AMDGPU::S_LSHR_B32:
2924 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
2925 NewOpcode = AMDGPU::V_LSHRREV_B32_e64;
2929 case AMDGPU::S_LSHL_B64:
2930 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
2931 NewOpcode = AMDGPU::V_LSHLREV_B64;
2935 case AMDGPU::S_ASHR_I64:
2936 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
2937 NewOpcode = AMDGPU::V_ASHRREV_I64;
2941 case AMDGPU::S_LSHR_B64:
2942 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
2943 NewOpcode = AMDGPU::V_LSHRREV_B64;
2948 case AMDGPU::S_ABS_I32:
2949 lowerScalarAbs(Worklist, Inst);
2950 Inst.eraseFromParent();
2953 case AMDGPU::S_CBRANCH_SCC0:
2954 case AMDGPU::S_CBRANCH_SCC1:
2955 // Clear unused bits of vcc
2956 BuildMI(*MBB, Inst, Inst.getDebugLoc(), get(AMDGPU::S_AND_B64),
2958 .addReg(AMDGPU::EXEC)
2959 .addReg(AMDGPU::VCC);
2962 case AMDGPU::S_BFE_U64:
2963 case AMDGPU::S_BFM_B64:
2964 llvm_unreachable("Moving this op to VALU not implemented");
2967 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
2968 // We cannot move this instruction to the VALU, so we should try to
2969 // legalize its operands instead.
2970 legalizeOperands(Inst);
2974 // Use the new VALU Opcode.
2975 const MCInstrDesc &NewDesc = get(NewOpcode);
2976 Inst.setDesc(NewDesc);
2978 // Remove any references to SCC. Vector instructions can't read from it, and
2979 // We're just about to add the implicit use / defs of VCC, and we don't want
2981 for (unsigned i = Inst.getNumOperands() - 1; i > 0; --i) {
2982 MachineOperand &Op = Inst.getOperand(i);
2983 if (Op.isReg() && Op.getReg() == AMDGPU::SCC) {
2984 Inst.RemoveOperand(i);
2985 addSCCDefUsersToVALUWorklist(Inst, Worklist);
2989 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
2990 // We are converting these to a BFE, so we need to add the missing
2991 // operands for the size and offset.
2992 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
2993 Inst.addOperand(MachineOperand::CreateImm(0));
2994 Inst.addOperand(MachineOperand::CreateImm(Size));
2996 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
2997 // The VALU version adds the second operand to the result, so insert an
2999 Inst.addOperand(MachineOperand::CreateImm(0));
3002 Inst.addImplicitDefUseOperands(*Inst.getParent()->getParent());
3004 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
3005 const MachineOperand &OffsetWidthOp = Inst.getOperand(2);
3006 // If we need to move this to VGPRs, we need to unpack the second operand
3007 // back into the 2 separate ones for bit offset and width.
3008 assert(OffsetWidthOp.isImm() &&
3009 "Scalar BFE is only implemented for constant width and offset");
3010 uint32_t Imm = OffsetWidthOp.getImm();
3012 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
3013 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
3014 Inst.RemoveOperand(2); // Remove old immediate.
3015 Inst.addOperand(MachineOperand::CreateImm(Offset));
3016 Inst.addOperand(MachineOperand::CreateImm(BitWidth));
3019 bool HasDst = Inst.getOperand(0).isReg() && Inst.getOperand(0).isDef();
3020 unsigned NewDstReg = AMDGPU::NoRegister;
3022 // Update the destination register class.
3023 const TargetRegisterClass *NewDstRC = getDestEquivalentVGPRClass(Inst);
3027 unsigned DstReg = Inst.getOperand(0).getReg();
3028 if (Inst.isCopy() &&
3029 TargetRegisterInfo::isVirtualRegister(Inst.getOperand(1).getReg()) &&
3030 NewDstRC == RI.getRegClassForReg(MRI, Inst.getOperand(1).getReg())) {
3031 // Instead of creating a copy where src and dst are the same register
3032 // class, we just replace all uses of dst with src. These kinds of
3033 // copies interfere with the heuristics MachineSink uses to decide
3034 // whether or not to split a critical edge. Since the pass assumes
3035 // that copies will end up as machine instructions and not be
3037 addUsersToMoveToVALUWorklist(DstReg, MRI, Worklist);
3038 MRI.replaceRegWith(DstReg, Inst.getOperand(1).getReg());
3039 MRI.clearKillFlags(Inst.getOperand(1).getReg());
3040 Inst.getOperand(0).setReg(DstReg);
3044 NewDstReg = MRI.createVirtualRegister(NewDstRC);
3045 MRI.replaceRegWith(DstReg, NewDstReg);
3048 // Legalize the operands
3049 legalizeOperands(Inst);
3052 addUsersToMoveToVALUWorklist(NewDstReg, MRI, Worklist);
3056 void SIInstrInfo::lowerScalarAbs(SmallVectorImpl<MachineInstr *> &Worklist,
3057 MachineInstr &Inst) const {
3058 MachineBasicBlock &MBB = *Inst.getParent();
3059 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3060 MachineBasicBlock::iterator MII = Inst;
3061 DebugLoc DL = Inst.getDebugLoc();
3063 MachineOperand &Dest = Inst.getOperand(0);
3064 MachineOperand &Src = Inst.getOperand(1);
3065 unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3066 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3068 BuildMI(MBB, MII, DL, get(AMDGPU::V_SUB_I32_e32), TmpReg)
3070 .addReg(Src.getReg());
3072 BuildMI(MBB, MII, DL, get(AMDGPU::V_MAX_I32_e64), ResultReg)
3073 .addReg(Src.getReg())
3076 MRI.replaceRegWith(Dest.getReg(), ResultReg);
3077 addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
3080 void SIInstrInfo::splitScalar64BitUnaryOp(
3081 SmallVectorImpl<MachineInstr *> &Worklist, MachineInstr &Inst,
3082 unsigned Opcode) const {
3083 MachineBasicBlock &MBB = *Inst.getParent();
3084 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3086 MachineOperand &Dest = Inst.getOperand(0);
3087 MachineOperand &Src0 = Inst.getOperand(1);
3088 DebugLoc DL = Inst.getDebugLoc();
3090 MachineBasicBlock::iterator MII = Inst;
3092 const MCInstrDesc &InstDesc = get(Opcode);
3093 const TargetRegisterClass *Src0RC = Src0.isReg() ?
3094 MRI.getRegClass(Src0.getReg()) :
3095 &AMDGPU::SGPR_32RegClass;
3097 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
3099 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
3100 AMDGPU::sub0, Src0SubRC);
3102 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
3103 const TargetRegisterClass *NewDestRC = RI.getEquivalentVGPRClass(DestRC);
3104 const TargetRegisterClass *NewDestSubRC = RI.getSubRegClass(NewDestRC, AMDGPU::sub0);
3106 unsigned DestSub0 = MRI.createVirtualRegister(NewDestSubRC);
3107 BuildMI(MBB, MII, DL, InstDesc, DestSub0)
3108 .addOperand(SrcReg0Sub0);
3110 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
3111 AMDGPU::sub1, Src0SubRC);
3113 unsigned DestSub1 = MRI.createVirtualRegister(NewDestSubRC);
3114 BuildMI(MBB, MII, DL, InstDesc, DestSub1)
3115 .addOperand(SrcReg0Sub1);
3117 unsigned FullDestReg = MRI.createVirtualRegister(NewDestRC);
3118 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
3120 .addImm(AMDGPU::sub0)
3122 .addImm(AMDGPU::sub1);
3124 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
3126 // We don't need to legalizeOperands here because for a single operand, src0
3127 // will support any kind of input.
3129 // Move all users of this moved value.
3130 addUsersToMoveToVALUWorklist(FullDestReg, MRI, Worklist);
3133 void SIInstrInfo::splitScalar64BitBinaryOp(
3134 SmallVectorImpl<MachineInstr *> &Worklist, MachineInstr &Inst,
3135 unsigned Opcode) const {
3136 MachineBasicBlock &MBB = *Inst.getParent();
3137 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3139 MachineOperand &Dest = Inst.getOperand(0);
3140 MachineOperand &Src0 = Inst.getOperand(1);
3141 MachineOperand &Src1 = Inst.getOperand(2);
3142 DebugLoc DL = Inst.getDebugLoc();
3144 MachineBasicBlock::iterator MII = Inst;
3146 const MCInstrDesc &InstDesc = get(Opcode);
3147 const TargetRegisterClass *Src0RC = Src0.isReg() ?
3148 MRI.getRegClass(Src0.getReg()) :
3149 &AMDGPU::SGPR_32RegClass;
3151 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
3152 const TargetRegisterClass *Src1RC = Src1.isReg() ?
3153 MRI.getRegClass(Src1.getReg()) :
3154 &AMDGPU::SGPR_32RegClass;
3156 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
3158 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
3159 AMDGPU::sub0, Src0SubRC);
3160 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
3161 AMDGPU::sub0, Src1SubRC);
3163 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
3164 const TargetRegisterClass *NewDestRC = RI.getEquivalentVGPRClass(DestRC);
3165 const TargetRegisterClass *NewDestSubRC = RI.getSubRegClass(NewDestRC, AMDGPU::sub0);
3167 unsigned DestSub0 = MRI.createVirtualRegister(NewDestSubRC);
3168 MachineInstr &LoHalf = *BuildMI(MBB, MII, DL, InstDesc, DestSub0)
3169 .addOperand(SrcReg0Sub0)
3170 .addOperand(SrcReg1Sub0);
3172 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
3173 AMDGPU::sub1, Src0SubRC);
3174 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
3175 AMDGPU::sub1, Src1SubRC);
3177 unsigned DestSub1 = MRI.createVirtualRegister(NewDestSubRC);
3178 MachineInstr &HiHalf = *BuildMI(MBB, MII, DL, InstDesc, DestSub1)
3179 .addOperand(SrcReg0Sub1)
3180 .addOperand(SrcReg1Sub1);
3182 unsigned FullDestReg = MRI.createVirtualRegister(NewDestRC);
3183 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
3185 .addImm(AMDGPU::sub0)
3187 .addImm(AMDGPU::sub1);
3189 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
3191 // Try to legalize the operands in case we need to swap the order to keep it
3193 legalizeOperands(LoHalf);
3194 legalizeOperands(HiHalf);
3196 // Move all users of this moved vlaue.
3197 addUsersToMoveToVALUWorklist(FullDestReg, MRI, Worklist);
3200 void SIInstrInfo::splitScalar64BitBCNT(
3201 SmallVectorImpl<MachineInstr *> &Worklist, MachineInstr &Inst) const {
3202 MachineBasicBlock &MBB = *Inst.getParent();
3203 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3205 MachineBasicBlock::iterator MII = Inst;
3206 DebugLoc DL = Inst.getDebugLoc();
3208 MachineOperand &Dest = Inst.getOperand(0);
3209 MachineOperand &Src = Inst.getOperand(1);
3211 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e64);
3212 const TargetRegisterClass *SrcRC = Src.isReg() ?
3213 MRI.getRegClass(Src.getReg()) :
3214 &AMDGPU::SGPR_32RegClass;
3216 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3217 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3219 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
3221 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
3222 AMDGPU::sub0, SrcSubRC);
3223 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
3224 AMDGPU::sub1, SrcSubRC);
3226 BuildMI(MBB, MII, DL, InstDesc, MidReg)
3227 .addOperand(SrcRegSub0)
3230 BuildMI(MBB, MII, DL, InstDesc, ResultReg)
3231 .addOperand(SrcRegSub1)
3234 MRI.replaceRegWith(Dest.getReg(), ResultReg);
3236 // We don't need to legalize operands here. src0 for etiher instruction can be
3237 // an SGPR, and the second input is unused or determined here.
3238 addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
3241 void SIInstrInfo::splitScalar64BitBFE(SmallVectorImpl<MachineInstr *> &Worklist,
3242 MachineInstr &Inst) const {
3243 MachineBasicBlock &MBB = *Inst.getParent();
3244 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3245 MachineBasicBlock::iterator MII = Inst;
3246 DebugLoc DL = Inst.getDebugLoc();
3248 MachineOperand &Dest = Inst.getOperand(0);
3249 uint32_t Imm = Inst.getOperand(2).getImm();
3250 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
3251 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
3255 // Only sext_inreg cases handled.
3256 assert(Inst.getOpcode() == AMDGPU::S_BFE_I64 && BitWidth <= 32 &&
3257 Offset == 0 && "Not implemented");
3259 if (BitWidth < 32) {
3260 unsigned MidRegLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3261 unsigned MidRegHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3262 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
3264 BuildMI(MBB, MII, DL, get(AMDGPU::V_BFE_I32), MidRegLo)
3265 .addReg(Inst.getOperand(1).getReg(), 0, AMDGPU::sub0)
3269 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e32), MidRegHi)
3273 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
3275 .addImm(AMDGPU::sub0)
3277 .addImm(AMDGPU::sub1);
3279 MRI.replaceRegWith(Dest.getReg(), ResultReg);
3280 addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
3284 MachineOperand &Src = Inst.getOperand(1);
3285 unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3286 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
3288 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e64), TmpReg)
3290 .addReg(Src.getReg(), 0, AMDGPU::sub0);
3292 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
3293 .addReg(Src.getReg(), 0, AMDGPU::sub0)
3294 .addImm(AMDGPU::sub0)
3296 .addImm(AMDGPU::sub1);
3298 MRI.replaceRegWith(Dest.getReg(), ResultReg);
3299 addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
3302 void SIInstrInfo::addUsersToMoveToVALUWorklist(
3304 MachineRegisterInfo &MRI,
3305 SmallVectorImpl<MachineInstr *> &Worklist) const {
3306 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(DstReg),
3307 E = MRI.use_end(); I != E; ++I) {
3308 MachineInstr &UseMI = *I->getParent();
3309 if (!canReadVGPR(UseMI, I.getOperandNo())) {
3310 Worklist.push_back(&UseMI);
3315 void SIInstrInfo::addSCCDefUsersToVALUWorklist(
3316 MachineInstr &SCCDefInst, SmallVectorImpl<MachineInstr *> &Worklist) const {
3317 // This assumes that all the users of SCC are in the same block
3319 for (MachineInstr &MI :
3320 llvm::make_range(MachineBasicBlock::iterator(SCCDefInst),
3321 SCCDefInst.getParent()->end())) {
3322 // Exit if we find another SCC def.
3323 if (MI.findRegisterDefOperandIdx(AMDGPU::SCC) != -1)
3326 if (MI.findRegisterUseOperandIdx(AMDGPU::SCC) != -1)
3327 Worklist.push_back(&MI);
3331 const TargetRegisterClass *SIInstrInfo::getDestEquivalentVGPRClass(
3332 const MachineInstr &Inst) const {
3333 const TargetRegisterClass *NewDstRC = getOpRegClass(Inst, 0);
3335 switch (Inst.getOpcode()) {
3336 // For target instructions, getOpRegClass just returns the virtual register
3337 // class associated with the operand, so we need to find an equivalent VGPR
3338 // register class in order to move the instruction to the VALU.
3341 case AMDGPU::REG_SEQUENCE:
3342 case AMDGPU::INSERT_SUBREG:
3343 if (RI.hasVGPRs(NewDstRC))
3346 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
3355 // Find the one SGPR operand we are allowed to use.
3356 unsigned SIInstrInfo::findUsedSGPR(const MachineInstr &MI,
3357 int OpIndices[3]) const {
3358 const MCInstrDesc &Desc = MI.getDesc();
3360 // Find the one SGPR operand we are allowed to use.
3362 // First we need to consider the instruction's operand requirements before
3363 // legalizing. Some operands are required to be SGPRs, such as implicit uses
3364 // of VCC, but we are still bound by the constant bus requirement to only use
3367 // If the operand's class is an SGPR, we can never move it.
3369 unsigned SGPRReg = findImplicitSGPRRead(MI);
3370 if (SGPRReg != AMDGPU::NoRegister)
3373 unsigned UsedSGPRs[3] = { AMDGPU::NoRegister };
3374 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
3376 for (unsigned i = 0; i < 3; ++i) {
3377 int Idx = OpIndices[i];
3381 const MachineOperand &MO = MI.getOperand(Idx);
3385 // Is this operand statically required to be an SGPR based on the operand
3387 const TargetRegisterClass *OpRC = RI.getRegClass(Desc.OpInfo[Idx].RegClass);
3388 bool IsRequiredSGPR = RI.isSGPRClass(OpRC);
3392 // If this could be a VGPR or an SGPR, Check the dynamic register class.
3393 unsigned Reg = MO.getReg();
3394 const TargetRegisterClass *RegRC = MRI.getRegClass(Reg);
3395 if (RI.isSGPRClass(RegRC))
3399 // We don't have a required SGPR operand, so we have a bit more freedom in
3400 // selecting operands to move.
3402 // Try to select the most used SGPR. If an SGPR is equal to one of the
3403 // others, we choose that.
3406 // V_FMA_F32 v0, s0, s0, s0 -> No moves
3407 // V_FMA_F32 v0, s0, s1, s0 -> Move s1
3409 // TODO: If some of the operands are 64-bit SGPRs and some 32, we should
3412 if (UsedSGPRs[0] != AMDGPU::NoRegister) {
3413 if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
3414 SGPRReg = UsedSGPRs[0];
3417 if (SGPRReg == AMDGPU::NoRegister && UsedSGPRs[1] != AMDGPU::NoRegister) {
3418 if (UsedSGPRs[1] == UsedSGPRs[2])
3419 SGPRReg = UsedSGPRs[1];
3425 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
3426 unsigned OperandName) const {
3427 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
3431 return &MI.getOperand(Idx);
3434 uint64_t SIInstrInfo::getDefaultRsrcDataFormat() const {
3435 uint64_t RsrcDataFormat = AMDGPU::RSRC_DATA_FORMAT;
3436 if (ST.isAmdHsaOS()) {
3437 RsrcDataFormat |= (1ULL << 56);
3439 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS)
3441 RsrcDataFormat |= (2ULL << 59);
3444 return RsrcDataFormat;
3447 uint64_t SIInstrInfo::getScratchRsrcWords23() const {
3448 uint64_t Rsrc23 = getDefaultRsrcDataFormat() |
3449 AMDGPU::RSRC_TID_ENABLE |
3450 0xffffffff; // Size;
3452 uint64_t EltSizeValue = Log2_32(ST.getMaxPrivateElementSize()) - 1;
3454 Rsrc23 |= (EltSizeValue << AMDGPU::RSRC_ELEMENT_SIZE_SHIFT) |
3456 (UINT64_C(3) << AMDGPU::RSRC_INDEX_STRIDE_SHIFT);
3458 // If TID_ENABLE is set, DATA_FORMAT specifies stride bits [14:17].
3459 // Clear them unless we want a huge stride.
3460 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS)
3461 Rsrc23 &= ~AMDGPU::RSRC_DATA_FORMAT;
3466 bool SIInstrInfo::isLowLatencyInstruction(const MachineInstr &MI) const {
3467 unsigned Opc = MI.getOpcode();
3472 bool SIInstrInfo::isHighLatencyInstruction(const MachineInstr &MI) const {
3473 unsigned Opc = MI.getOpcode();
3475 return isMUBUF(Opc) || isMTBUF(Opc) || isMIMG(Opc);
3478 unsigned SIInstrInfo::isStackAccess(const MachineInstr &MI,
3479 int &FrameIndex) const {
3480 const MachineOperand *Addr = getNamedOperand(MI, AMDGPU::OpName::vaddr);
3481 if (!Addr || !Addr->isFI())
3482 return AMDGPU::NoRegister;
3484 assert(!MI.memoperands_empty() &&
3485 (*MI.memoperands_begin())->getAddrSpace() == AMDGPUAS::PRIVATE_ADDRESS);
3487 FrameIndex = Addr->getIndex();
3488 return getNamedOperand(MI, AMDGPU::OpName::vdata)->getReg();
3491 unsigned SIInstrInfo::isSGPRStackAccess(const MachineInstr &MI,
3492 int &FrameIndex) const {
3493 const MachineOperand *Addr = getNamedOperand(MI, AMDGPU::OpName::addr);
3494 assert(Addr && Addr->isFI());
3495 FrameIndex = Addr->getIndex();
3496 return getNamedOperand(MI, AMDGPU::OpName::data)->getReg();
3499 unsigned SIInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
3500 int &FrameIndex) const {
3503 return AMDGPU::NoRegister;
3505 if (isMUBUF(MI) || isVGPRSpill(MI))
3506 return isStackAccess(MI, FrameIndex);
3508 if (isSGPRSpill(MI))
3509 return isSGPRStackAccess(MI, FrameIndex);
3511 return AMDGPU::NoRegister;
3514 unsigned SIInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
3515 int &FrameIndex) const {
3517 return AMDGPU::NoRegister;
3519 if (isMUBUF(MI) || isVGPRSpill(MI))
3520 return isStackAccess(MI, FrameIndex);
3522 if (isSGPRSpill(MI))
3523 return isSGPRStackAccess(MI, FrameIndex);
3525 return AMDGPU::NoRegister;
3528 unsigned SIInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
3529 unsigned Opc = MI.getOpcode();
3530 const MCInstrDesc &Desc = getMCOpcodeFromPseudo(Opc);
3531 unsigned DescSize = Desc.getSize();
3533 // If we have a definitive size, we can use it. Otherwise we need to inspect
3534 // the operands to know the size.
3536 // FIXME: Instructions that have a base 32-bit encoding report their size as
3537 // 4, even though they are really 8 bytes if they have a literal operand.
3538 if (DescSize != 0 && DescSize != 4)
3541 if (Opc == AMDGPU::WAVE_BARRIER)
3544 // 4-byte instructions may have a 32-bit literal encoded after them. Check
3545 // operands that coud ever be literals.
3546 if (isVALU(MI) || isSALU(MI)) {
3547 if (isFixedSize(MI)) {
3548 assert(DescSize == 4);
3552 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
3554 return 4; // No operands.
3556 if (isLiteralConstantLike(MI.getOperand(Src0Idx), getOpSize(MI, Src0Idx)))
3559 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
3563 if (isLiteralConstantLike(MI.getOperand(Src1Idx), getOpSize(MI, Src1Idx)))
3573 case AMDGPU::SI_MASK_BRANCH:
3574 case TargetOpcode::IMPLICIT_DEF:
3575 case TargetOpcode::KILL:
3576 case TargetOpcode::DBG_VALUE:
3577 case TargetOpcode::BUNDLE:
3578 case TargetOpcode::EH_LABEL:
3580 case TargetOpcode::INLINEASM: {
3581 const MachineFunction *MF = MI.getParent()->getParent();
3582 const char *AsmStr = MI.getOperand(0).getSymbolName();
3583 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
3586 llvm_unreachable("unable to find instruction size");
3590 bool SIInstrInfo::mayAccessFlatAddressSpace(const MachineInstr &MI) const {
3594 if (MI.memoperands_empty())
3597 for (const MachineMemOperand *MMO : MI.memoperands()) {
3598 if (MMO->getAddrSpace() == AMDGPUAS::FLAT_ADDRESS)
3604 ArrayRef<std::pair<int, const char *>>
3605 SIInstrInfo::getSerializableTargetIndices() const {
3606 static const std::pair<int, const char *> TargetIndices[] = {
3607 {AMDGPU::TI_CONSTDATA_START, "amdgpu-constdata-start"},
3608 {AMDGPU::TI_SCRATCH_RSRC_DWORD0, "amdgpu-scratch-rsrc-dword0"},
3609 {AMDGPU::TI_SCRATCH_RSRC_DWORD1, "amdgpu-scratch-rsrc-dword1"},
3610 {AMDGPU::TI_SCRATCH_RSRC_DWORD2, "amdgpu-scratch-rsrc-dword2"},
3611 {AMDGPU::TI_SCRATCH_RSRC_DWORD3, "amdgpu-scratch-rsrc-dword3"}};
3612 return makeArrayRef(TargetIndices);
3615 /// This is used by the post-RA scheduler (SchedulePostRAList.cpp). The
3616 /// post-RA version of misched uses CreateTargetMIHazardRecognizer.
3617 ScheduleHazardRecognizer *
3618 SIInstrInfo::CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
3619 const ScheduleDAG *DAG) const {
3620 return new GCNHazardRecognizer(DAG->MF);
3623 /// This is the hazard recognizer used at -O0 by the PostRAHazardRecognizer
3625 ScheduleHazardRecognizer *
3626 SIInstrInfo::CreateTargetPostRAHazardRecognizer(const MachineFunction &MF) const {
3627 return new GCNHazardRecognizer(MF);