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/RegisterScavenging.h"
24 #include "llvm/CodeGen/ScheduleDAG.h"
25 #include "llvm/IR/DiagnosticInfo.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/MC/MCInstrDesc.h"
28 #include "llvm/Support/Debug.h"
32 // Must be at least 4 to be able to branch over minimum unconditional branch
33 // code. This is only for making it possible to write reasonably small tests for
35 static cl::opt<unsigned>
36 BranchOffsetBits("amdgpu-s-branch-bits", cl::ReallyHidden, cl::init(16),
37 cl::desc("Restrict range of branch instructions (DEBUG)"));
39 SIInstrInfo::SIInstrInfo(const SISubtarget &ST)
40 : AMDGPUInstrInfo(ST), RI(ST), ST(ST) {}
42 //===----------------------------------------------------------------------===//
43 // TargetInstrInfo callbacks
44 //===----------------------------------------------------------------------===//
46 static unsigned getNumOperandsNoGlue(SDNode *Node) {
47 unsigned N = Node->getNumOperands();
48 while (N && Node->getOperand(N - 1).getValueType() == MVT::Glue)
53 static SDValue findChainOperand(SDNode *Load) {
54 SDValue LastOp = Load->getOperand(getNumOperandsNoGlue(Load) - 1);
55 assert(LastOp.getValueType() == MVT::Other && "Chain missing from load node");
59 /// \brief Returns true if both nodes have the same value for the given
60 /// operand \p Op, or if both nodes do not have this operand.
61 static bool nodesHaveSameOperandValue(SDNode *N0, SDNode* N1, unsigned OpName) {
62 unsigned Opc0 = N0->getMachineOpcode();
63 unsigned Opc1 = N1->getMachineOpcode();
65 int Op0Idx = AMDGPU::getNamedOperandIdx(Opc0, OpName);
66 int Op1Idx = AMDGPU::getNamedOperandIdx(Opc1, OpName);
68 if (Op0Idx == -1 && Op1Idx == -1)
72 if ((Op0Idx == -1 && Op1Idx != -1) ||
73 (Op1Idx == -1 && Op0Idx != -1))
76 // getNamedOperandIdx returns the index for the MachineInstr's operands,
77 // which includes the result as the first operand. We are indexing into the
78 // MachineSDNode's operands, so we need to skip the result operand to get
83 return N0->getOperand(Op0Idx) == N1->getOperand(Op1Idx);
86 bool SIInstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
87 AliasAnalysis *AA) const {
88 // TODO: The generic check fails for VALU instructions that should be
89 // rematerializable due to implicit reads of exec. We really want all of the
90 // generic logic for this except for this.
91 switch (MI.getOpcode()) {
92 case AMDGPU::V_MOV_B32_e32:
93 case AMDGPU::V_MOV_B32_e64:
94 case AMDGPU::V_MOV_B64_PSEUDO:
101 bool SIInstrInfo::areLoadsFromSameBasePtr(SDNode *Load0, SDNode *Load1,
103 int64_t &Offset1) const {
104 if (!Load0->isMachineOpcode() || !Load1->isMachineOpcode())
107 unsigned Opc0 = Load0->getMachineOpcode();
108 unsigned Opc1 = Load1->getMachineOpcode();
110 // Make sure both are actually loads.
111 if (!get(Opc0).mayLoad() || !get(Opc1).mayLoad())
114 if (isDS(Opc0) && isDS(Opc1)) {
116 // FIXME: Handle this case:
117 if (getNumOperandsNoGlue(Load0) != getNumOperandsNoGlue(Load1))
121 if (Load0->getOperand(1) != Load1->getOperand(1))
125 if (findChainOperand(Load0) != findChainOperand(Load1))
128 // Skip read2 / write2 variants for simplicity.
129 // TODO: We should report true if the used offsets are adjacent (excluded
131 if (AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::data1) != -1 ||
132 AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::data1) != -1)
135 Offset0 = cast<ConstantSDNode>(Load0->getOperand(2))->getZExtValue();
136 Offset1 = cast<ConstantSDNode>(Load1->getOperand(2))->getZExtValue();
140 if (isSMRD(Opc0) && isSMRD(Opc1)) {
141 // Skip time and cache invalidation instructions.
142 if (AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::sbase) == -1 ||
143 AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::sbase) == -1)
146 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
149 if (Load0->getOperand(0) != Load1->getOperand(0))
152 const ConstantSDNode *Load0Offset =
153 dyn_cast<ConstantSDNode>(Load0->getOperand(1));
154 const ConstantSDNode *Load1Offset =
155 dyn_cast<ConstantSDNode>(Load1->getOperand(1));
157 if (!Load0Offset || !Load1Offset)
161 if (findChainOperand(Load0) != findChainOperand(Load1))
164 Offset0 = Load0Offset->getZExtValue();
165 Offset1 = Load1Offset->getZExtValue();
169 // MUBUF and MTBUF can access the same addresses.
170 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1))) {
172 // MUBUF and MTBUF have vaddr at different indices.
173 if (!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::soffset) ||
174 findChainOperand(Load0) != findChainOperand(Load1) ||
175 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::vaddr) ||
176 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::srsrc))
179 int OffIdx0 = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
180 int OffIdx1 = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
182 if (OffIdx0 == -1 || OffIdx1 == -1)
185 // getNamedOperandIdx returns the index for MachineInstrs. Since they
186 // inlcude the output in the operand list, but SDNodes don't, we need to
187 // subtract the index by one.
191 SDValue Off0 = Load0->getOperand(OffIdx0);
192 SDValue Off1 = Load1->getOperand(OffIdx1);
194 // The offset might be a FrameIndexSDNode.
195 if (!isa<ConstantSDNode>(Off0) || !isa<ConstantSDNode>(Off1))
198 Offset0 = cast<ConstantSDNode>(Off0)->getZExtValue();
199 Offset1 = cast<ConstantSDNode>(Off1)->getZExtValue();
206 static bool isStride64(unsigned Opc) {
208 case AMDGPU::DS_READ2ST64_B32:
209 case AMDGPU::DS_READ2ST64_B64:
210 case AMDGPU::DS_WRITE2ST64_B32:
211 case AMDGPU::DS_WRITE2ST64_B64:
218 bool SIInstrInfo::getMemOpBaseRegImmOfs(MachineInstr &LdSt, unsigned &BaseReg,
220 const TargetRegisterInfo *TRI) const {
221 unsigned Opc = LdSt.getOpcode();
224 const MachineOperand *OffsetImm =
225 getNamedOperand(LdSt, AMDGPU::OpName::offset);
227 // Normal, single offset LDS instruction.
228 const MachineOperand *AddrReg =
229 getNamedOperand(LdSt, AMDGPU::OpName::addr);
231 BaseReg = AddrReg->getReg();
232 Offset = OffsetImm->getImm();
236 // The 2 offset instructions use offset0 and offset1 instead. We can treat
237 // these as a load with a single offset if the 2 offsets are consecutive. We
238 // will use this for some partially aligned loads.
239 const MachineOperand *Offset0Imm =
240 getNamedOperand(LdSt, AMDGPU::OpName::offset0);
241 const MachineOperand *Offset1Imm =
242 getNamedOperand(LdSt, AMDGPU::OpName::offset1);
244 uint8_t Offset0 = Offset0Imm->getImm();
245 uint8_t Offset1 = Offset1Imm->getImm();
247 if (Offset1 > Offset0 && Offset1 - Offset0 == 1) {
248 // Each of these offsets is in element sized units, so we need to convert
249 // to bytes of the individual reads.
253 EltSize = TRI->getRegSizeInBits(*getOpRegClass(LdSt, 0)) / 16;
255 assert(LdSt.mayStore());
256 int Data0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
257 EltSize = TRI->getRegSizeInBits(*getOpRegClass(LdSt, Data0Idx)) / 8;
263 const MachineOperand *AddrReg =
264 getNamedOperand(LdSt, AMDGPU::OpName::addr);
265 BaseReg = AddrReg->getReg();
266 Offset = EltSize * Offset0;
273 if (isMUBUF(LdSt) || isMTBUF(LdSt)) {
274 const MachineOperand *SOffset = getNamedOperand(LdSt, AMDGPU::OpName::soffset);
275 if (SOffset && SOffset->isReg())
278 const MachineOperand *AddrReg =
279 getNamedOperand(LdSt, AMDGPU::OpName::vaddr);
283 const MachineOperand *OffsetImm =
284 getNamedOperand(LdSt, AMDGPU::OpName::offset);
285 BaseReg = AddrReg->getReg();
286 Offset = OffsetImm->getImm();
288 if (SOffset) // soffset can be an inline immediate.
289 Offset += SOffset->getImm();
295 const MachineOperand *OffsetImm =
296 getNamedOperand(LdSt, AMDGPU::OpName::offset);
300 const MachineOperand *SBaseReg =
301 getNamedOperand(LdSt, AMDGPU::OpName::sbase);
302 BaseReg = SBaseReg->getReg();
303 Offset = OffsetImm->getImm();
308 const MachineOperand *AddrReg = getNamedOperand(LdSt, AMDGPU::OpName::vaddr);
309 BaseReg = AddrReg->getReg();
317 bool SIInstrInfo::shouldClusterMemOps(MachineInstr &FirstLdSt,
318 MachineInstr &SecondLdSt,
319 unsigned NumLoads) const {
320 const MachineOperand *FirstDst = nullptr;
321 const MachineOperand *SecondDst = nullptr;
323 if ((isMUBUF(FirstLdSt) && isMUBUF(SecondLdSt)) ||
324 (isMTBUF(FirstLdSt) && isMTBUF(SecondLdSt)) ||
325 (isFLAT(FirstLdSt) && isFLAT(SecondLdSt))) {
326 FirstDst = getNamedOperand(FirstLdSt, AMDGPU::OpName::vdata);
327 SecondDst = getNamedOperand(SecondLdSt, AMDGPU::OpName::vdata);
328 } else if (isSMRD(FirstLdSt) && isSMRD(SecondLdSt)) {
329 FirstDst = getNamedOperand(FirstLdSt, AMDGPU::OpName::sdst);
330 SecondDst = getNamedOperand(SecondLdSt, AMDGPU::OpName::sdst);
331 } else if (isDS(FirstLdSt) && isDS(SecondLdSt)) {
332 FirstDst = getNamedOperand(FirstLdSt, AMDGPU::OpName::vdst);
333 SecondDst = getNamedOperand(SecondLdSt, AMDGPU::OpName::vdst);
336 if (!FirstDst || !SecondDst)
339 // Try to limit clustering based on the total number of bytes loaded
340 // rather than the number of instructions. This is done to help reduce
341 // register pressure. The method used is somewhat inexact, though,
342 // because it assumes that all loads in the cluster will load the
343 // same number of bytes as FirstLdSt.
345 // The unit of this value is bytes.
346 // FIXME: This needs finer tuning.
347 unsigned LoadClusterThreshold = 16;
349 const MachineRegisterInfo &MRI =
350 FirstLdSt.getParent()->getParent()->getRegInfo();
351 const TargetRegisterClass *DstRC = MRI.getRegClass(FirstDst->getReg());
353 return (NumLoads * (RI.getRegSizeInBits(*DstRC) / 8)) <= LoadClusterThreshold;
356 static void reportIllegalCopy(const SIInstrInfo *TII, MachineBasicBlock &MBB,
357 MachineBasicBlock::iterator MI,
358 const DebugLoc &DL, unsigned DestReg,
359 unsigned SrcReg, bool KillSrc) {
360 MachineFunction *MF = MBB.getParent();
361 DiagnosticInfoUnsupported IllegalCopy(*MF->getFunction(),
362 "illegal SGPR to VGPR copy",
364 LLVMContext &C = MF->getFunction()->getContext();
365 C.diagnose(IllegalCopy);
367 BuildMI(MBB, MI, DL, TII->get(AMDGPU::SI_ILLEGAL_COPY), DestReg)
368 .addReg(SrcReg, getKillRegState(KillSrc));
371 void SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
372 MachineBasicBlock::iterator MI,
373 const DebugLoc &DL, unsigned DestReg,
374 unsigned SrcReg, bool KillSrc) const {
375 const TargetRegisterClass *RC = RI.getPhysRegClass(DestReg);
377 if (RC == &AMDGPU::VGPR_32RegClass) {
378 assert(AMDGPU::VGPR_32RegClass.contains(SrcReg) ||
379 AMDGPU::SReg_32RegClass.contains(SrcReg));
380 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
381 .addReg(SrcReg, getKillRegState(KillSrc));
385 if (RC == &AMDGPU::SReg_32_XM0RegClass ||
386 RC == &AMDGPU::SReg_32RegClass) {
387 if (SrcReg == AMDGPU::SCC) {
388 BuildMI(MBB, MI, DL, get(AMDGPU::S_CSELECT_B32), DestReg)
394 if (!AMDGPU::SReg_32RegClass.contains(SrcReg)) {
395 reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc);
399 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
400 .addReg(SrcReg, getKillRegState(KillSrc));
404 if (RC == &AMDGPU::SReg_64RegClass) {
405 if (DestReg == AMDGPU::VCC) {
406 if (AMDGPU::SReg_64RegClass.contains(SrcReg)) {
407 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), AMDGPU::VCC)
408 .addReg(SrcReg, getKillRegState(KillSrc));
410 // FIXME: Hack until VReg_1 removed.
411 assert(AMDGPU::VGPR_32RegClass.contains(SrcReg));
412 BuildMI(MBB, MI, DL, get(AMDGPU::V_CMP_NE_U32_e32))
414 .addReg(SrcReg, getKillRegState(KillSrc));
420 if (!AMDGPU::SReg_64RegClass.contains(SrcReg)) {
421 reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc);
425 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
426 .addReg(SrcReg, getKillRegState(KillSrc));
430 if (DestReg == AMDGPU::SCC) {
431 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
432 BuildMI(MBB, MI, DL, get(AMDGPU::S_CMP_LG_U32))
433 .addReg(SrcReg, getKillRegState(KillSrc))
438 unsigned EltSize = 4;
439 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
440 if (RI.isSGPRClass(RC)) {
441 if (RI.getRegSizeInBits(*RC) > 32) {
442 Opcode = AMDGPU::S_MOV_B64;
445 Opcode = AMDGPU::S_MOV_B32;
449 if (!RI.isSGPRClass(RI.getPhysRegClass(SrcReg))) {
450 reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc);
456 ArrayRef<int16_t> SubIndices = RI.getRegSplitParts(RC, EltSize);
457 bool Forward = RI.getHWRegIndex(DestReg) <= RI.getHWRegIndex(SrcReg);
459 for (unsigned Idx = 0; Idx < SubIndices.size(); ++Idx) {
462 SubIdx = SubIndices[Idx];
464 SubIdx = SubIndices[SubIndices.size() - Idx - 1];
466 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
467 get(Opcode), RI.getSubReg(DestReg, SubIdx));
469 Builder.addReg(RI.getSubReg(SrcReg, SubIdx));
472 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
474 bool UseKill = KillSrc && Idx == SubIndices.size() - 1;
475 Builder.addReg(SrcReg, getKillRegState(UseKill) | RegState::Implicit);
479 int SIInstrInfo::commuteOpcode(unsigned Opcode) const {
482 // Try to map original to commuted opcode
483 NewOpc = AMDGPU::getCommuteRev(Opcode);
485 // Check if the commuted (REV) opcode exists on the target.
486 return pseudoToMCOpcode(NewOpc) != -1 ? NewOpc : -1;
488 // Try to map commuted to original opcode
489 NewOpc = AMDGPU::getCommuteOrig(Opcode);
491 // Check if the original (non-REV) opcode exists on the target.
492 return pseudoToMCOpcode(NewOpc) != -1 ? NewOpc : -1;
497 void SIInstrInfo::materializeImmediate(MachineBasicBlock &MBB,
498 MachineBasicBlock::iterator MI,
499 const DebugLoc &DL, unsigned DestReg,
500 int64_t Value) const {
501 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
502 const TargetRegisterClass *RegClass = MRI.getRegClass(DestReg);
503 if (RegClass == &AMDGPU::SReg_32RegClass ||
504 RegClass == &AMDGPU::SGPR_32RegClass ||
505 RegClass == &AMDGPU::SReg_32_XM0RegClass ||
506 RegClass == &AMDGPU::SReg_32_XM0_XEXECRegClass) {
507 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
512 if (RegClass == &AMDGPU::SReg_64RegClass ||
513 RegClass == &AMDGPU::SGPR_64RegClass ||
514 RegClass == &AMDGPU::SReg_64_XEXECRegClass) {
515 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
520 if (RegClass == &AMDGPU::VGPR_32RegClass) {
521 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
525 if (RegClass == &AMDGPU::VReg_64RegClass) {
526 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B64_PSEUDO), DestReg)
531 unsigned EltSize = 4;
532 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
533 if (RI.isSGPRClass(RegClass)) {
534 if (RI.getRegSizeInBits(*RegClass) > 32) {
535 Opcode = AMDGPU::S_MOV_B64;
538 Opcode = AMDGPU::S_MOV_B32;
543 ArrayRef<int16_t> SubIndices = RI.getRegSplitParts(RegClass, EltSize);
544 for (unsigned Idx = 0; Idx < SubIndices.size(); ++Idx) {
545 int64_t IdxValue = Idx == 0 ? Value : 0;
547 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
548 get(Opcode), RI.getSubReg(DestReg, Idx));
549 Builder.addImm(IdxValue);
553 const TargetRegisterClass *
554 SIInstrInfo::getPreferredSelectRegClass(unsigned Size) const {
555 return &AMDGPU::VGPR_32RegClass;
558 void SIInstrInfo::insertVectorSelect(MachineBasicBlock &MBB,
559 MachineBasicBlock::iterator I,
560 const DebugLoc &DL, unsigned DstReg,
561 ArrayRef<MachineOperand> Cond,
563 unsigned FalseReg) const {
564 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
565 assert(MRI.getRegClass(DstReg) == &AMDGPU::VGPR_32RegClass &&
568 if (Cond.size() == 1) {
569 BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
573 } else if (Cond.size() == 2) {
574 assert(Cond[0].isImm() && "Cond[0] is not an immediate");
575 switch (Cond[0].getImm()) {
576 case SIInstrInfo::SCC_TRUE: {
577 unsigned SReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
578 BuildMI(MBB, I, DL, get(AMDGPU::S_CSELECT_B64), SReg)
581 BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
587 case SIInstrInfo::SCC_FALSE: {
588 unsigned SReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
589 BuildMI(MBB, I, DL, get(AMDGPU::S_CSELECT_B64), SReg)
592 BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
598 case SIInstrInfo::VCCNZ: {
599 MachineOperand RegOp = Cond[1];
600 RegOp.setImplicit(false);
601 BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
607 case SIInstrInfo::VCCZ: {
608 MachineOperand RegOp = Cond[1];
609 RegOp.setImplicit(false);
610 BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
616 case SIInstrInfo::EXECNZ: {
617 unsigned SReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
618 unsigned SReg2 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
619 BuildMI(MBB, I, DL, get(AMDGPU::S_OR_SAVEEXEC_B64), SReg2)
621 BuildMI(MBB, I, DL, get(AMDGPU::S_CSELECT_B64), SReg)
624 BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
630 case SIInstrInfo::EXECZ: {
631 unsigned SReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
632 unsigned SReg2 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
633 BuildMI(MBB, I, DL, get(AMDGPU::S_OR_SAVEEXEC_B64), SReg2)
635 BuildMI(MBB, I, DL, get(AMDGPU::S_CSELECT_B64), SReg)
638 BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
642 llvm_unreachable("Unhandled branch predicate EXECZ");
646 llvm_unreachable("invalid branch predicate");
649 llvm_unreachable("Can only handle Cond size 1 or 2");
653 unsigned SIInstrInfo::insertEQ(MachineBasicBlock *MBB,
654 MachineBasicBlock::iterator I,
656 unsigned SrcReg, int Value) const {
657 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
658 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
659 BuildMI(*MBB, I, DL, get(AMDGPU::V_CMP_EQ_I32_e64), Reg)
666 unsigned SIInstrInfo::insertNE(MachineBasicBlock *MBB,
667 MachineBasicBlock::iterator I,
669 unsigned SrcReg, int Value) const {
670 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
671 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
672 BuildMI(*MBB, I, DL, get(AMDGPU::V_CMP_NE_I32_e64), Reg)
679 unsigned SIInstrInfo::getMovOpcode(const TargetRegisterClass *DstRC) const {
681 if (RI.getRegSizeInBits(*DstRC) == 32) {
682 return RI.isSGPRClass(DstRC) ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32;
683 } else if (RI.getRegSizeInBits(*DstRC) == 64 && RI.isSGPRClass(DstRC)) {
684 return AMDGPU::S_MOV_B64;
685 } else if (RI.getRegSizeInBits(*DstRC) == 64 && !RI.isSGPRClass(DstRC)) {
686 return AMDGPU::V_MOV_B64_PSEUDO;
691 static unsigned getSGPRSpillSaveOpcode(unsigned Size) {
694 return AMDGPU::SI_SPILL_S32_SAVE;
696 return AMDGPU::SI_SPILL_S64_SAVE;
698 return AMDGPU::SI_SPILL_S128_SAVE;
700 return AMDGPU::SI_SPILL_S256_SAVE;
702 return AMDGPU::SI_SPILL_S512_SAVE;
704 llvm_unreachable("unknown register size");
708 static unsigned getVGPRSpillSaveOpcode(unsigned Size) {
711 return AMDGPU::SI_SPILL_V32_SAVE;
713 return AMDGPU::SI_SPILL_V64_SAVE;
715 return AMDGPU::SI_SPILL_V96_SAVE;
717 return AMDGPU::SI_SPILL_V128_SAVE;
719 return AMDGPU::SI_SPILL_V256_SAVE;
721 return AMDGPU::SI_SPILL_V512_SAVE;
723 llvm_unreachable("unknown register size");
727 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
728 MachineBasicBlock::iterator MI,
729 unsigned SrcReg, bool isKill,
731 const TargetRegisterClass *RC,
732 const TargetRegisterInfo *TRI) const {
733 MachineFunction *MF = MBB.getParent();
734 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
735 MachineFrameInfo &FrameInfo = MF->getFrameInfo();
736 DebugLoc DL = MBB.findDebugLoc(MI);
738 unsigned Size = FrameInfo.getObjectSize(FrameIndex);
739 unsigned Align = FrameInfo.getObjectAlignment(FrameIndex);
740 MachinePointerInfo PtrInfo
741 = MachinePointerInfo::getFixedStack(*MF, FrameIndex);
742 MachineMemOperand *MMO
743 = MF->getMachineMemOperand(PtrInfo, MachineMemOperand::MOStore,
745 unsigned SpillSize = TRI->getSpillSize(*RC);
747 if (RI.isSGPRClass(RC)) {
748 MFI->setHasSpilledSGPRs();
750 // We are only allowed to create one new instruction when spilling
751 // registers, so we need to use pseudo instruction for spilling SGPRs.
752 const MCInstrDesc &OpDesc = get(getSGPRSpillSaveOpcode(SpillSize));
754 // The SGPR spill/restore instructions only work on number sgprs, so we need
755 // to make sure we are using the correct register class.
756 if (TargetRegisterInfo::isVirtualRegister(SrcReg) && SpillSize == 4) {
757 MachineRegisterInfo &MRI = MF->getRegInfo();
758 MRI.constrainRegClass(SrcReg, &AMDGPU::SReg_32_XM0RegClass);
761 MachineInstrBuilder Spill = BuildMI(MBB, MI, DL, OpDesc)
762 .addReg(SrcReg, getKillRegState(isKill)) // data
763 .addFrameIndex(FrameIndex) // addr
765 .addReg(MFI->getScratchRSrcReg(), RegState::Implicit)
766 .addReg(MFI->getFrameOffsetReg(), RegState::Implicit);
767 // Add the scratch resource registers as implicit uses because we may end up
768 // needing them, and need to ensure that the reserved registers are
769 // correctly handled.
771 if (ST.hasScalarStores()) {
772 // m0 is used for offset to scalar stores if used to spill.
773 Spill.addReg(AMDGPU::M0, RegState::ImplicitDefine | RegState::Dead);
779 if (!ST.isVGPRSpillingEnabled(*MF->getFunction())) {
780 LLVMContext &Ctx = MF->getFunction()->getContext();
781 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Do not know how to"
783 BuildMI(MBB, MI, DL, get(AMDGPU::KILL))
789 assert(RI.hasVGPRs(RC) && "Only VGPR spilling expected");
791 unsigned Opcode = getVGPRSpillSaveOpcode(SpillSize);
792 MFI->setHasSpilledVGPRs();
793 BuildMI(MBB, MI, DL, get(Opcode))
794 .addReg(SrcReg, getKillRegState(isKill)) // data
795 .addFrameIndex(FrameIndex) // addr
796 .addReg(MFI->getScratchRSrcReg()) // scratch_rsrc
797 .addReg(MFI->getFrameOffsetReg()) // scratch_offset
802 static unsigned getSGPRSpillRestoreOpcode(unsigned Size) {
805 return AMDGPU::SI_SPILL_S32_RESTORE;
807 return AMDGPU::SI_SPILL_S64_RESTORE;
809 return AMDGPU::SI_SPILL_S128_RESTORE;
811 return AMDGPU::SI_SPILL_S256_RESTORE;
813 return AMDGPU::SI_SPILL_S512_RESTORE;
815 llvm_unreachable("unknown register size");
819 static unsigned getVGPRSpillRestoreOpcode(unsigned Size) {
822 return AMDGPU::SI_SPILL_V32_RESTORE;
824 return AMDGPU::SI_SPILL_V64_RESTORE;
826 return AMDGPU::SI_SPILL_V96_RESTORE;
828 return AMDGPU::SI_SPILL_V128_RESTORE;
830 return AMDGPU::SI_SPILL_V256_RESTORE;
832 return AMDGPU::SI_SPILL_V512_RESTORE;
834 llvm_unreachable("unknown register size");
838 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
839 MachineBasicBlock::iterator MI,
840 unsigned DestReg, int FrameIndex,
841 const TargetRegisterClass *RC,
842 const TargetRegisterInfo *TRI) const {
843 MachineFunction *MF = MBB.getParent();
844 const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
845 MachineFrameInfo &FrameInfo = MF->getFrameInfo();
846 DebugLoc DL = MBB.findDebugLoc(MI);
847 unsigned Align = FrameInfo.getObjectAlignment(FrameIndex);
848 unsigned Size = FrameInfo.getObjectSize(FrameIndex);
849 unsigned SpillSize = TRI->getSpillSize(*RC);
851 MachinePointerInfo PtrInfo
852 = MachinePointerInfo::getFixedStack(*MF, FrameIndex);
854 MachineMemOperand *MMO = MF->getMachineMemOperand(
855 PtrInfo, MachineMemOperand::MOLoad, Size, Align);
857 if (RI.isSGPRClass(RC)) {
858 // FIXME: Maybe this should not include a memoperand because it will be
859 // lowered to non-memory instructions.
860 const MCInstrDesc &OpDesc = get(getSGPRSpillRestoreOpcode(SpillSize));
861 if (TargetRegisterInfo::isVirtualRegister(DestReg) && SpillSize == 4) {
862 MachineRegisterInfo &MRI = MF->getRegInfo();
863 MRI.constrainRegClass(DestReg, &AMDGPU::SReg_32_XM0RegClass);
866 MachineInstrBuilder Spill = BuildMI(MBB, MI, DL, OpDesc, DestReg)
867 .addFrameIndex(FrameIndex) // addr
869 .addReg(MFI->getScratchRSrcReg(), RegState::Implicit)
870 .addReg(MFI->getFrameOffsetReg(), RegState::Implicit);
872 if (ST.hasScalarStores()) {
873 // m0 is used for offset to scalar stores if used to spill.
874 Spill.addReg(AMDGPU::M0, RegState::ImplicitDefine | RegState::Dead);
880 if (!ST.isVGPRSpillingEnabled(*MF->getFunction())) {
881 LLVMContext &Ctx = MF->getFunction()->getContext();
882 Ctx.emitError("SIInstrInfo::loadRegFromStackSlot - Do not know how to"
883 " restore register");
884 BuildMI(MBB, MI, DL, get(AMDGPU::IMPLICIT_DEF), DestReg);
889 assert(RI.hasVGPRs(RC) && "Only VGPR spilling expected");
891 unsigned Opcode = getVGPRSpillRestoreOpcode(SpillSize);
892 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
893 .addFrameIndex(FrameIndex) // vaddr
894 .addReg(MFI->getScratchRSrcReg()) // scratch_rsrc
895 .addReg(MFI->getFrameOffsetReg()) // scratch_offset
900 /// \param @Offset Offset in bytes of the FrameIndex being spilled
901 unsigned SIInstrInfo::calculateLDSSpillAddress(
902 MachineBasicBlock &MBB, MachineInstr &MI, RegScavenger *RS, unsigned TmpReg,
903 unsigned FrameOffset, unsigned Size) const {
904 MachineFunction *MF = MBB.getParent();
905 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
906 const SISubtarget &ST = MF->getSubtarget<SISubtarget>();
907 const SIRegisterInfo *TRI = ST.getRegisterInfo();
908 DebugLoc DL = MBB.findDebugLoc(MI);
909 unsigned WorkGroupSize = MFI->getMaxFlatWorkGroupSize();
910 unsigned WavefrontSize = ST.getWavefrontSize();
912 unsigned TIDReg = MFI->getTIDReg();
913 if (!MFI->hasCalculatedTID()) {
914 MachineBasicBlock &Entry = MBB.getParent()->front();
915 MachineBasicBlock::iterator Insert = Entry.front();
916 DebugLoc DL = Insert->getDebugLoc();
918 TIDReg = RI.findUnusedRegister(MF->getRegInfo(), &AMDGPU::VGPR_32RegClass,
920 if (TIDReg == AMDGPU::NoRegister)
923 if (!AMDGPU::isShader(MF->getFunction()->getCallingConv()) &&
924 WorkGroupSize > WavefrontSize) {
927 = TRI->getPreloadedValue(*MF, SIRegisterInfo::WORKGROUP_ID_X);
929 = TRI->getPreloadedValue(*MF, SIRegisterInfo::WORKGROUP_ID_Y);
931 = TRI->getPreloadedValue(*MF, SIRegisterInfo::WORKGROUP_ID_Z);
932 unsigned InputPtrReg =
933 TRI->getPreloadedValue(*MF, SIRegisterInfo::KERNARG_SEGMENT_PTR);
934 for (unsigned Reg : {TIDIGXReg, TIDIGYReg, TIDIGZReg}) {
935 if (!Entry.isLiveIn(Reg))
936 Entry.addLiveIn(Reg);
939 RS->enterBasicBlock(Entry);
940 // FIXME: Can we scavenge an SReg_64 and access the subregs?
941 unsigned STmp0 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
942 unsigned STmp1 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
943 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp0)
945 .addImm(SI::KernelInputOffsets::NGROUPS_Z);
946 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp1)
948 .addImm(SI::KernelInputOffsets::NGROUPS_Y);
950 // NGROUPS.X * NGROUPS.Y
951 BuildMI(Entry, Insert, DL, get(AMDGPU::S_MUL_I32), STmp1)
954 // (NGROUPS.X * NGROUPS.Y) * TIDIG.X
955 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MUL_U32_U24_e32), TIDReg)
958 // NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)
959 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MAD_U32_U24), TIDReg)
963 // (NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)) + TIDIG.Z
964 BuildMI(Entry, Insert, DL, get(AMDGPU::V_ADD_I32_e32), TIDReg)
969 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_LO_U32_B32_e64),
974 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_HI_U32_B32_e64),
980 BuildMI(Entry, Insert, DL, get(AMDGPU::V_LSHLREV_B32_e32),
984 MFI->setTIDReg(TIDReg);
987 // Add FrameIndex to LDS offset
988 unsigned LDSOffset = MFI->getLDSSize() + (FrameOffset * WorkGroupSize);
989 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_I32_e32), TmpReg)
996 void SIInstrInfo::insertWaitStates(MachineBasicBlock &MBB,
997 MachineBasicBlock::iterator MI,
999 DebugLoc DL = MBB.findDebugLoc(MI);
1007 BuildMI(MBB, MI, DL, get(AMDGPU::S_NOP))
1012 void SIInstrInfo::insertNoop(MachineBasicBlock &MBB,
1013 MachineBasicBlock::iterator MI) const {
1014 insertWaitStates(MBB, MI, 1);
1017 void SIInstrInfo::insertReturn(MachineBasicBlock &MBB) const {
1018 auto MF = MBB.getParent();
1019 SIMachineFunctionInfo *Info = MF->getInfo<SIMachineFunctionInfo>();
1021 assert(Info->isEntryFunction());
1023 if (MBB.succ_empty()) {
1024 bool HasNoTerminator = MBB.getFirstTerminator() == MBB.end();
1025 if (HasNoTerminator)
1026 BuildMI(MBB, MBB.end(), DebugLoc(),
1027 get(Info->returnsVoid() ? AMDGPU::S_ENDPGM : AMDGPU::SI_RETURN_TO_EPILOG));
1031 unsigned SIInstrInfo::getNumWaitStates(const MachineInstr &MI) const {
1032 switch (MI.getOpcode()) {
1033 default: return 1; // FIXME: Do wait states equal cycles?
1036 return MI.getOperand(0).getImm() + 1;
1040 bool SIInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
1041 MachineBasicBlock &MBB = *MI.getParent();
1042 DebugLoc DL = MBB.findDebugLoc(MI);
1043 switch (MI.getOpcode()) {
1044 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
1045 case AMDGPU::S_MOV_B64_term: {
1046 // This is only a terminator to get the correct spill code placement during
1047 // register allocation.
1048 MI.setDesc(get(AMDGPU::S_MOV_B64));
1051 case AMDGPU::S_XOR_B64_term: {
1052 // This is only a terminator to get the correct spill code placement during
1053 // register allocation.
1054 MI.setDesc(get(AMDGPU::S_XOR_B64));
1057 case AMDGPU::S_ANDN2_B64_term: {
1058 // This is only a terminator to get the correct spill code placement during
1059 // register allocation.
1060 MI.setDesc(get(AMDGPU::S_ANDN2_B64));
1063 case AMDGPU::V_MOV_B64_PSEUDO: {
1064 unsigned Dst = MI.getOperand(0).getReg();
1065 unsigned DstLo = RI.getSubReg(Dst, AMDGPU::sub0);
1066 unsigned DstHi = RI.getSubReg(Dst, AMDGPU::sub1);
1068 const MachineOperand &SrcOp = MI.getOperand(1);
1069 // FIXME: Will this work for 64-bit floating point immediates?
1070 assert(!SrcOp.isFPImm());
1071 if (SrcOp.isImm()) {
1072 APInt Imm(64, SrcOp.getImm());
1073 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
1074 .addImm(Imm.getLoBits(32).getZExtValue())
1075 .addReg(Dst, RegState::Implicit | RegState::Define);
1076 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
1077 .addImm(Imm.getHiBits(32).getZExtValue())
1078 .addReg(Dst, RegState::Implicit | RegState::Define);
1080 assert(SrcOp.isReg());
1081 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
1082 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub0))
1083 .addReg(Dst, RegState::Implicit | RegState::Define);
1084 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
1085 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub1))
1086 .addReg(Dst, RegState::Implicit | RegState::Define);
1088 MI.eraseFromParent();
1091 case AMDGPU::V_MOVRELD_B32_V1:
1092 case AMDGPU::V_MOVRELD_B32_V2:
1093 case AMDGPU::V_MOVRELD_B32_V4:
1094 case AMDGPU::V_MOVRELD_B32_V8:
1095 case AMDGPU::V_MOVRELD_B32_V16: {
1096 const MCInstrDesc &MovRelDesc = get(AMDGPU::V_MOVRELD_B32_e32);
1097 unsigned VecReg = MI.getOperand(0).getReg();
1098 bool IsUndef = MI.getOperand(1).isUndef();
1099 unsigned SubReg = AMDGPU::sub0 + MI.getOperand(3).getImm();
1100 assert(VecReg == MI.getOperand(1).getReg());
1102 MachineInstr *MovRel =
1103 BuildMI(MBB, MI, DL, MovRelDesc)
1104 .addReg(RI.getSubReg(VecReg, SubReg), RegState::Undef)
1105 .add(MI.getOperand(2))
1106 .addReg(VecReg, RegState::ImplicitDefine)
1108 RegState::Implicit | (IsUndef ? RegState::Undef : 0));
1110 const int ImpDefIdx =
1111 MovRelDesc.getNumOperands() + MovRelDesc.getNumImplicitUses();
1112 const int ImpUseIdx = ImpDefIdx + 1;
1113 MovRel->tieOperands(ImpDefIdx, ImpUseIdx);
1115 MI.eraseFromParent();
1118 case AMDGPU::SI_PC_ADD_REL_OFFSET: {
1119 MachineFunction &MF = *MBB.getParent();
1120 unsigned Reg = MI.getOperand(0).getReg();
1121 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
1122 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
1124 // Create a bundle so these instructions won't be re-ordered by the
1125 // post-RA scheduler.
1126 MIBundleBuilder Bundler(MBB, MI);
1127 Bundler.append(BuildMI(MF, DL, get(AMDGPU::S_GETPC_B64), Reg));
1129 // Add 32-bit offset from this instruction to the start of the
1131 Bundler.append(BuildMI(MF, DL, get(AMDGPU::S_ADD_U32), RegLo)
1133 .add(MI.getOperand(1)));
1135 MachineInstrBuilder MIB = BuildMI(MF, DL, get(AMDGPU::S_ADDC_U32), RegHi)
1137 if (MI.getOperand(2).getTargetFlags() == SIInstrInfo::MO_NONE)
1140 MIB.add(MI.getOperand(2));
1142 Bundler.append(MIB);
1143 llvm::finalizeBundle(MBB, Bundler.begin());
1145 MI.eraseFromParent();
1152 bool SIInstrInfo::swapSourceModifiers(MachineInstr &MI,
1153 MachineOperand &Src0,
1154 unsigned Src0OpName,
1155 MachineOperand &Src1,
1156 unsigned Src1OpName) const {
1157 MachineOperand *Src0Mods = getNamedOperand(MI, Src0OpName);
1161 MachineOperand *Src1Mods = getNamedOperand(MI, Src1OpName);
1163 "All commutable instructions have both src0 and src1 modifiers");
1165 int Src0ModsVal = Src0Mods->getImm();
1166 int Src1ModsVal = Src1Mods->getImm();
1168 Src1Mods->setImm(Src0ModsVal);
1169 Src0Mods->setImm(Src1ModsVal);
1173 static MachineInstr *swapRegAndNonRegOperand(MachineInstr &MI,
1174 MachineOperand &RegOp,
1175 MachineOperand &NonRegOp) {
1176 unsigned Reg = RegOp.getReg();
1177 unsigned SubReg = RegOp.getSubReg();
1178 bool IsKill = RegOp.isKill();
1179 bool IsDead = RegOp.isDead();
1180 bool IsUndef = RegOp.isUndef();
1181 bool IsDebug = RegOp.isDebug();
1183 if (NonRegOp.isImm())
1184 RegOp.ChangeToImmediate(NonRegOp.getImm());
1185 else if (NonRegOp.isFI())
1186 RegOp.ChangeToFrameIndex(NonRegOp.getIndex());
1190 NonRegOp.ChangeToRegister(Reg, false, false, IsKill, IsDead, IsUndef, IsDebug);
1191 NonRegOp.setSubReg(SubReg);
1196 MachineInstr *SIInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
1198 unsigned Src1Idx) const {
1199 assert(!NewMI && "this should never be used");
1201 unsigned Opc = MI.getOpcode();
1202 int CommutedOpcode = commuteOpcode(Opc);
1203 if (CommutedOpcode == -1)
1206 assert(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0) ==
1207 static_cast<int>(Src0Idx) &&
1208 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1) ==
1209 static_cast<int>(Src1Idx) &&
1210 "inconsistency with findCommutedOpIndices");
1212 MachineOperand &Src0 = MI.getOperand(Src0Idx);
1213 MachineOperand &Src1 = MI.getOperand(Src1Idx);
1215 MachineInstr *CommutedMI = nullptr;
1216 if (Src0.isReg() && Src1.isReg()) {
1217 if (isOperandLegal(MI, Src1Idx, &Src0)) {
1218 // Be sure to copy the source modifiers to the right place.
1220 = TargetInstrInfo::commuteInstructionImpl(MI, NewMI, Src0Idx, Src1Idx);
1223 } else if (Src0.isReg() && !Src1.isReg()) {
1224 // src0 should always be able to support any operand type, so no need to
1225 // check operand legality.
1226 CommutedMI = swapRegAndNonRegOperand(MI, Src0, Src1);
1227 } else if (!Src0.isReg() && Src1.isReg()) {
1228 if (isOperandLegal(MI, Src1Idx, &Src0))
1229 CommutedMI = swapRegAndNonRegOperand(MI, Src1, Src0);
1231 // FIXME: Found two non registers to commute. This does happen.
1237 swapSourceModifiers(MI, Src0, AMDGPU::OpName::src0_modifiers,
1238 Src1, AMDGPU::OpName::src1_modifiers);
1240 CommutedMI->setDesc(get(CommutedOpcode));
1246 // This needs to be implemented because the source modifiers may be inserted
1247 // between the true commutable operands, and the base
1248 // TargetInstrInfo::commuteInstruction uses it.
1249 bool SIInstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx0,
1250 unsigned &SrcOpIdx1) const {
1251 if (!MI.isCommutable())
1254 unsigned Opc = MI.getOpcode();
1255 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
1259 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
1263 return fixCommutedOpIndices(SrcOpIdx0, SrcOpIdx1, Src0Idx, Src1Idx);
1266 bool SIInstrInfo::isBranchOffsetInRange(unsigned BranchOp,
1267 int64_t BrOffset) const {
1268 // BranchRelaxation should never have to check s_setpc_b64 because its dest
1269 // block is unanalyzable.
1270 assert(BranchOp != AMDGPU::S_SETPC_B64);
1272 // Convert to dwords.
1275 // The branch instructions do PC += signext(SIMM16 * 4) + 4, so the offset is
1276 // from the next instruction.
1279 return isIntN(BranchOffsetBits, BrOffset);
1282 MachineBasicBlock *SIInstrInfo::getBranchDestBlock(
1283 const MachineInstr &MI) const {
1284 if (MI.getOpcode() == AMDGPU::S_SETPC_B64) {
1285 // This would be a difficult analysis to perform, but can always be legal so
1286 // there's no need to analyze it.
1290 return MI.getOperand(0).getMBB();
1293 unsigned SIInstrInfo::insertIndirectBranch(MachineBasicBlock &MBB,
1294 MachineBasicBlock &DestBB,
1297 RegScavenger *RS) const {
1298 assert(RS && "RegScavenger required for long branching");
1299 assert(MBB.empty() &&
1300 "new block should be inserted for expanding unconditional branch");
1301 assert(MBB.pred_size() == 1);
1303 MachineFunction *MF = MBB.getParent();
1304 MachineRegisterInfo &MRI = MF->getRegInfo();
1306 // FIXME: Virtual register workaround for RegScavenger not working with empty
1308 unsigned PCReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1312 // We need to compute the offset relative to the instruction immediately after
1313 // s_getpc_b64. Insert pc arithmetic code before last terminator.
1314 MachineInstr *GetPC = BuildMI(MBB, I, DL, get(AMDGPU::S_GETPC_B64), PCReg);
1316 // TODO: Handle > 32-bit block address.
1317 if (BrOffset >= 0) {
1318 BuildMI(MBB, I, DL, get(AMDGPU::S_ADD_U32))
1319 .addReg(PCReg, RegState::Define, AMDGPU::sub0)
1320 .addReg(PCReg, 0, AMDGPU::sub0)
1321 .addMBB(&DestBB, AMDGPU::TF_LONG_BRANCH_FORWARD);
1322 BuildMI(MBB, I, DL, get(AMDGPU::S_ADDC_U32))
1323 .addReg(PCReg, RegState::Define, AMDGPU::sub1)
1324 .addReg(PCReg, 0, AMDGPU::sub1)
1327 // Backwards branch.
1328 BuildMI(MBB, I, DL, get(AMDGPU::S_SUB_U32))
1329 .addReg(PCReg, RegState::Define, AMDGPU::sub0)
1330 .addReg(PCReg, 0, AMDGPU::sub0)
1331 .addMBB(&DestBB, AMDGPU::TF_LONG_BRANCH_BACKWARD);
1332 BuildMI(MBB, I, DL, get(AMDGPU::S_SUBB_U32))
1333 .addReg(PCReg, RegState::Define, AMDGPU::sub1)
1334 .addReg(PCReg, 0, AMDGPU::sub1)
1338 // Insert the indirect branch after the other terminator.
1339 BuildMI(&MBB, DL, get(AMDGPU::S_SETPC_B64))
1342 // FIXME: If spilling is necessary, this will fail because this scavenger has
1343 // no emergency stack slots. It is non-trivial to spill in this situation,
1344 // because the restore code needs to be specially placed after the
1345 // jump. BranchRelaxation then needs to be made aware of the newly inserted
1348 // If a spill is needed for the pc register pair, we need to insert a spill
1349 // restore block right before the destination block, and insert a short branch
1350 // into the old destination block's fallthrough predecessor.
1353 // s_cbranch_scc0 skip_long_branch:
1357 // s_getpc_b64 s[8:9]
1358 // s_add_u32 s8, s8, restore_bb
1359 // s_addc_u32 s9, s9, 0
1360 // s_setpc_b64 s[8:9]
1362 // skip_long_branch:
1367 // dest_bb_fallthrough_predecessor:
1373 // fallthrough dest_bb
1378 RS->enterBasicBlockEnd(MBB);
1379 unsigned Scav = RS->scavengeRegister(&AMDGPU::SReg_64RegClass,
1380 MachineBasicBlock::iterator(GetPC), 0);
1381 MRI.replaceRegWith(PCReg, Scav);
1382 MRI.clearVirtRegs();
1383 RS->setRegUsed(Scav);
1385 return 4 + 8 + 4 + 4;
1388 unsigned SIInstrInfo::getBranchOpcode(SIInstrInfo::BranchPredicate Cond) {
1390 case SIInstrInfo::SCC_TRUE:
1391 return AMDGPU::S_CBRANCH_SCC1;
1392 case SIInstrInfo::SCC_FALSE:
1393 return AMDGPU::S_CBRANCH_SCC0;
1394 case SIInstrInfo::VCCNZ:
1395 return AMDGPU::S_CBRANCH_VCCNZ;
1396 case SIInstrInfo::VCCZ:
1397 return AMDGPU::S_CBRANCH_VCCZ;
1398 case SIInstrInfo::EXECNZ:
1399 return AMDGPU::S_CBRANCH_EXECNZ;
1400 case SIInstrInfo::EXECZ:
1401 return AMDGPU::S_CBRANCH_EXECZ;
1403 llvm_unreachable("invalid branch predicate");
1407 SIInstrInfo::BranchPredicate SIInstrInfo::getBranchPredicate(unsigned Opcode) {
1409 case AMDGPU::S_CBRANCH_SCC0:
1411 case AMDGPU::S_CBRANCH_SCC1:
1413 case AMDGPU::S_CBRANCH_VCCNZ:
1415 case AMDGPU::S_CBRANCH_VCCZ:
1417 case AMDGPU::S_CBRANCH_EXECNZ:
1419 case AMDGPU::S_CBRANCH_EXECZ:
1426 bool SIInstrInfo::analyzeBranchImpl(MachineBasicBlock &MBB,
1427 MachineBasicBlock::iterator I,
1428 MachineBasicBlock *&TBB,
1429 MachineBasicBlock *&FBB,
1430 SmallVectorImpl<MachineOperand> &Cond,
1431 bool AllowModify) const {
1432 if (I->getOpcode() == AMDGPU::S_BRANCH) {
1433 // Unconditional Branch
1434 TBB = I->getOperand(0).getMBB();
1438 MachineBasicBlock *CondBB = nullptr;
1440 if (I->getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO) {
1441 CondBB = I->getOperand(1).getMBB();
1442 Cond.push_back(I->getOperand(0));
1444 BranchPredicate Pred = getBranchPredicate(I->getOpcode());
1445 if (Pred == INVALID_BR)
1448 CondBB = I->getOperand(0).getMBB();
1449 Cond.push_back(MachineOperand::CreateImm(Pred));
1450 Cond.push_back(I->getOperand(1)); // Save the branch register.
1454 if (I == MBB.end()) {
1455 // Conditional branch followed by fall-through.
1460 if (I->getOpcode() == AMDGPU::S_BRANCH) {
1462 FBB = I->getOperand(0).getMBB();
1469 bool SIInstrInfo::analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
1470 MachineBasicBlock *&FBB,
1471 SmallVectorImpl<MachineOperand> &Cond,
1472 bool AllowModify) const {
1473 MachineBasicBlock::iterator I = MBB.getFirstTerminator();
1477 if (I->getOpcode() != AMDGPU::SI_MASK_BRANCH)
1478 return analyzeBranchImpl(MBB, I, TBB, FBB, Cond, AllowModify);
1482 // TODO: Should be able to treat as fallthrough?
1486 if (analyzeBranchImpl(MBB, I, TBB, FBB, Cond, AllowModify))
1489 MachineBasicBlock *MaskBrDest = I->getOperand(0).getMBB();
1491 // Specifically handle the case where the conditional branch is to the same
1492 // destination as the mask branch. e.g.
1494 // si_mask_branch BB8
1495 // s_cbranch_execz BB8
1498 // This is required to understand divergent loops which may need the branches
1500 if (TBB != MaskBrDest || Cond.empty())
1503 auto Pred = Cond[0].getImm();
1504 return (Pred != EXECZ && Pred != EXECNZ);
1507 unsigned SIInstrInfo::removeBranch(MachineBasicBlock &MBB,
1508 int *BytesRemoved) const {
1509 MachineBasicBlock::iterator I = MBB.getFirstTerminator();
1512 unsigned RemovedSize = 0;
1513 while (I != MBB.end()) {
1514 MachineBasicBlock::iterator Next = std::next(I);
1515 if (I->getOpcode() == AMDGPU::SI_MASK_BRANCH) {
1520 RemovedSize += getInstSizeInBytes(*I);
1521 I->eraseFromParent();
1527 *BytesRemoved = RemovedSize;
1532 // Copy the flags onto the implicit condition register operand.
1533 static void preserveCondRegFlags(MachineOperand &CondReg,
1534 const MachineOperand &OrigCond) {
1535 CondReg.setIsUndef(OrigCond.isUndef());
1536 CondReg.setIsKill(OrigCond.isKill());
1539 unsigned SIInstrInfo::insertBranch(MachineBasicBlock &MBB,
1540 MachineBasicBlock *TBB,
1541 MachineBasicBlock *FBB,
1542 ArrayRef<MachineOperand> Cond,
1544 int *BytesAdded) const {
1546 if (!FBB && Cond.empty()) {
1547 BuildMI(&MBB, DL, get(AMDGPU::S_BRANCH))
1554 if(Cond.size() == 1 && Cond[0].isReg()) {
1555 BuildMI(&MBB, DL, get(AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO))
1561 assert(TBB && Cond[0].isImm());
1564 = getBranchOpcode(static_cast<BranchPredicate>(Cond[0].getImm()));
1568 MachineInstr *CondBr =
1569 BuildMI(&MBB, DL, get(Opcode))
1572 // Copy the flags onto the implicit condition register operand.
1573 preserveCondRegFlags(CondBr->getOperand(1), Cond[1]);
1582 MachineInstr *CondBr =
1583 BuildMI(&MBB, DL, get(Opcode))
1585 BuildMI(&MBB, DL, get(AMDGPU::S_BRANCH))
1588 MachineOperand &CondReg = CondBr->getOperand(1);
1589 CondReg.setIsUndef(Cond[1].isUndef());
1590 CondReg.setIsKill(Cond[1].isKill());
1598 bool SIInstrInfo::reverseBranchCondition(
1599 SmallVectorImpl<MachineOperand> &Cond) const {
1600 if (Cond.size() != 2) {
1604 if (Cond[0].isImm()) {
1605 Cond[0].setImm(-Cond[0].getImm());
1612 bool SIInstrInfo::canInsertSelect(const MachineBasicBlock &MBB,
1613 ArrayRef<MachineOperand> Cond,
1614 unsigned TrueReg, unsigned FalseReg,
1616 int &TrueCycles, int &FalseCycles) const {
1617 switch (Cond[0].getImm()) {
1620 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1621 const TargetRegisterClass *RC = MRI.getRegClass(TrueReg);
1622 assert(MRI.getRegClass(FalseReg) == RC);
1624 int NumInsts = AMDGPU::getRegBitWidth(RC->getID()) / 32;
1625 CondCycles = TrueCycles = FalseCycles = NumInsts; // ???
1627 // Limit to equal cost for branch vs. N v_cndmask_b32s.
1628 return !RI.isSGPRClass(RC) && NumInsts <= 6;
1632 // FIXME: We could insert for VGPRs if we could replace the original compare
1633 // with a vector one.
1634 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1635 const TargetRegisterClass *RC = MRI.getRegClass(TrueReg);
1636 assert(MRI.getRegClass(FalseReg) == RC);
1638 int NumInsts = AMDGPU::getRegBitWidth(RC->getID()) / 32;
1640 // Multiples of 8 can do s_cselect_b64
1641 if (NumInsts % 2 == 0)
1644 CondCycles = TrueCycles = FalseCycles = NumInsts; // ???
1645 return RI.isSGPRClass(RC);
1652 void SIInstrInfo::insertSelect(MachineBasicBlock &MBB,
1653 MachineBasicBlock::iterator I, const DebugLoc &DL,
1654 unsigned DstReg, ArrayRef<MachineOperand> Cond,
1655 unsigned TrueReg, unsigned FalseReg) const {
1656 BranchPredicate Pred = static_cast<BranchPredicate>(Cond[0].getImm());
1657 if (Pred == VCCZ || Pred == SCC_FALSE) {
1658 Pred = static_cast<BranchPredicate>(-Pred);
1659 std::swap(TrueReg, FalseReg);
1662 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1663 const TargetRegisterClass *DstRC = MRI.getRegClass(DstReg);
1664 unsigned DstSize = RI.getRegSizeInBits(*DstRC);
1666 if (DstSize == 32) {
1667 unsigned SelOp = Pred == SCC_TRUE ?
1668 AMDGPU::S_CSELECT_B32 : AMDGPU::V_CNDMASK_B32_e32;
1670 // Instruction's operands are backwards from what is expected.
1671 MachineInstr *Select =
1672 BuildMI(MBB, I, DL, get(SelOp), DstReg)
1676 preserveCondRegFlags(Select->getOperand(3), Cond[1]);
1680 if (DstSize == 64 && Pred == SCC_TRUE) {
1681 MachineInstr *Select =
1682 BuildMI(MBB, I, DL, get(AMDGPU::S_CSELECT_B64), DstReg)
1686 preserveCondRegFlags(Select->getOperand(3), Cond[1]);
1690 static const int16_t Sub0_15[] = {
1691 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
1692 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
1693 AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
1694 AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15,
1697 static const int16_t Sub0_15_64[] = {
1698 AMDGPU::sub0_sub1, AMDGPU::sub2_sub3,
1699 AMDGPU::sub4_sub5, AMDGPU::sub6_sub7,
1700 AMDGPU::sub8_sub9, AMDGPU::sub10_sub11,
1701 AMDGPU::sub12_sub13, AMDGPU::sub14_sub15,
1704 unsigned SelOp = AMDGPU::V_CNDMASK_B32_e32;
1705 const TargetRegisterClass *EltRC = &AMDGPU::VGPR_32RegClass;
1706 const int16_t *SubIndices = Sub0_15;
1707 int NElts = DstSize / 32;
1709 // 64-bit select is only avaialble for SALU.
1710 if (Pred == SCC_TRUE) {
1711 SelOp = AMDGPU::S_CSELECT_B64;
1712 EltRC = &AMDGPU::SGPR_64RegClass;
1713 SubIndices = Sub0_15_64;
1715 assert(NElts % 2 == 0);
1719 MachineInstrBuilder MIB = BuildMI(
1720 MBB, I, DL, get(AMDGPU::REG_SEQUENCE), DstReg);
1722 I = MIB->getIterator();
1724 SmallVector<unsigned, 8> Regs;
1725 for (int Idx = 0; Idx != NElts; ++Idx) {
1726 unsigned DstElt = MRI.createVirtualRegister(EltRC);
1727 Regs.push_back(DstElt);
1729 unsigned SubIdx = SubIndices[Idx];
1731 MachineInstr *Select =
1732 BuildMI(MBB, I, DL, get(SelOp), DstElt)
1733 .addReg(FalseReg, 0, SubIdx)
1734 .addReg(TrueReg, 0, SubIdx);
1735 preserveCondRegFlags(Select->getOperand(3), Cond[1]);
1742 bool SIInstrInfo::isFoldableCopy(const MachineInstr &MI) const {
1743 switch (MI.getOpcode()) {
1744 case AMDGPU::V_MOV_B32_e32:
1745 case AMDGPU::V_MOV_B32_e64:
1746 case AMDGPU::V_MOV_B64_PSEUDO: {
1747 // If there are additional implicit register operands, this may be used for
1748 // register indexing so the source register operand isn't simply copied.
1749 unsigned NumOps = MI.getDesc().getNumOperands() +
1750 MI.getDesc().getNumImplicitUses();
1752 return MI.getNumOperands() == NumOps;
1754 case AMDGPU::S_MOV_B32:
1755 case AMDGPU::S_MOV_B64:
1763 static void removeModOperands(MachineInstr &MI) {
1764 unsigned Opc = MI.getOpcode();
1765 int Src0ModIdx = AMDGPU::getNamedOperandIdx(Opc,
1766 AMDGPU::OpName::src0_modifiers);
1767 int Src1ModIdx = AMDGPU::getNamedOperandIdx(Opc,
1768 AMDGPU::OpName::src1_modifiers);
1769 int Src2ModIdx = AMDGPU::getNamedOperandIdx(Opc,
1770 AMDGPU::OpName::src2_modifiers);
1772 MI.RemoveOperand(Src2ModIdx);
1773 MI.RemoveOperand(Src1ModIdx);
1774 MI.RemoveOperand(Src0ModIdx);
1777 bool SIInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
1778 unsigned Reg, MachineRegisterInfo *MRI) const {
1779 if (!MRI->hasOneNonDBGUse(Reg))
1782 unsigned Opc = UseMI.getOpcode();
1783 if (Opc == AMDGPU::COPY) {
1784 bool isVGPRCopy = RI.isVGPR(*MRI, UseMI.getOperand(0).getReg());
1785 switch (DefMI.getOpcode()) {
1788 case AMDGPU::S_MOV_B64:
1789 // TODO: We could fold 64-bit immediates, but this get compilicated
1790 // when there are sub-registers.
1793 case AMDGPU::V_MOV_B32_e32:
1794 case AMDGPU::S_MOV_B32:
1797 unsigned NewOpc = isVGPRCopy ? AMDGPU::V_MOV_B32_e32 : AMDGPU::S_MOV_B32;
1798 const MachineOperand *ImmOp = getNamedOperand(DefMI, AMDGPU::OpName::src0);
1800 // FIXME: We could handle FrameIndex values here.
1801 if (!ImmOp->isImm()) {
1804 UseMI.setDesc(get(NewOpc));
1805 UseMI.getOperand(1).ChangeToImmediate(ImmOp->getImm());
1806 UseMI.addImplicitDefUseOperands(*UseMI.getParent()->getParent());
1810 if (Opc == AMDGPU::V_MAD_F32 || Opc == AMDGPU::V_MAC_F32_e64 ||
1811 Opc == AMDGPU::V_MAD_F16 || Opc == AMDGPU::V_MAC_F16_e64) {
1812 // Don't fold if we are using source or output modifiers. The new VOP2
1813 // instructions don't have them.
1814 if (hasAnyModifiersSet(UseMI))
1817 const MachineOperand &ImmOp = DefMI.getOperand(1);
1819 // If this is a free constant, there's no reason to do this.
1820 // TODO: We could fold this here instead of letting SIFoldOperands do it
1822 MachineOperand *Src0 = getNamedOperand(UseMI, AMDGPU::OpName::src0);
1824 // Any src operand can be used for the legality check.
1825 if (isInlineConstant(UseMI, *Src0, ImmOp))
1828 bool IsF32 = Opc == AMDGPU::V_MAD_F32 || Opc == AMDGPU::V_MAC_F32_e64;
1829 MachineOperand *Src1 = getNamedOperand(UseMI, AMDGPU::OpName::src1);
1830 MachineOperand *Src2 = getNamedOperand(UseMI, AMDGPU::OpName::src2);
1832 // Multiplied part is the constant: Use v_madmk_{f16, f32}.
1833 // We should only expect these to be on src0 due to canonicalizations.
1834 if (Src0->isReg() && Src0->getReg() == Reg) {
1835 if (!Src1->isReg() || RI.isSGPRClass(MRI->getRegClass(Src1->getReg())))
1838 if (!Src2->isReg() || RI.isSGPRClass(MRI->getRegClass(Src2->getReg())))
1841 // We need to swap operands 0 and 1 since madmk constant is at operand 1.
1843 const int64_t Imm = DefMI.getOperand(1).getImm();
1845 // FIXME: This would be a lot easier if we could return a new instruction
1846 // instead of having to modify in place.
1848 // Remove these first since they are at the end.
1849 UseMI.RemoveOperand(
1850 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::omod));
1851 UseMI.RemoveOperand(
1852 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::clamp));
1854 unsigned Src1Reg = Src1->getReg();
1855 unsigned Src1SubReg = Src1->getSubReg();
1856 Src0->setReg(Src1Reg);
1857 Src0->setSubReg(Src1SubReg);
1858 Src0->setIsKill(Src1->isKill());
1860 if (Opc == AMDGPU::V_MAC_F32_e64 ||
1861 Opc == AMDGPU::V_MAC_F16_e64)
1862 UseMI.untieRegOperand(
1863 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2));
1865 Src1->ChangeToImmediate(Imm);
1867 removeModOperands(UseMI);
1868 UseMI.setDesc(get(IsF32 ? AMDGPU::V_MADMK_F32 : AMDGPU::V_MADMK_F16));
1870 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
1872 DefMI.eraseFromParent();
1877 // Added part is the constant: Use v_madak_{f16, f32}.
1878 if (Src2->isReg() && Src2->getReg() == Reg) {
1879 // Not allowed to use constant bus for another operand.
1880 // We can however allow an inline immediate as src0.
1881 if (!Src0->isImm() &&
1882 (Src0->isReg() && RI.isSGPRClass(MRI->getRegClass(Src0->getReg()))))
1885 if (!Src1->isReg() || RI.isSGPRClass(MRI->getRegClass(Src1->getReg())))
1888 const int64_t Imm = DefMI.getOperand(1).getImm();
1890 // FIXME: This would be a lot easier if we could return a new instruction
1891 // instead of having to modify in place.
1893 // Remove these first since they are at the end.
1894 UseMI.RemoveOperand(
1895 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::omod));
1896 UseMI.RemoveOperand(
1897 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::clamp));
1899 if (Opc == AMDGPU::V_MAC_F32_e64 ||
1900 Opc == AMDGPU::V_MAC_F16_e64)
1901 UseMI.untieRegOperand(
1902 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2));
1904 // ChangingToImmediate adds Src2 back to the instruction.
1905 Src2->ChangeToImmediate(Imm);
1907 // These come before src2.
1908 removeModOperands(UseMI);
1909 UseMI.setDesc(get(IsF32 ? AMDGPU::V_MADAK_F32 : AMDGPU::V_MADAK_F16));
1911 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
1913 DefMI.eraseFromParent();
1922 static bool offsetsDoNotOverlap(int WidthA, int OffsetA,
1923 int WidthB, int OffsetB) {
1924 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB;
1925 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA;
1926 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
1927 return LowOffset + LowWidth <= HighOffset;
1930 bool SIInstrInfo::checkInstOffsetsDoNotOverlap(MachineInstr &MIa,
1931 MachineInstr &MIb) const {
1932 unsigned BaseReg0, BaseReg1;
1933 int64_t Offset0, Offset1;
1935 if (getMemOpBaseRegImmOfs(MIa, BaseReg0, Offset0, &RI) &&
1936 getMemOpBaseRegImmOfs(MIb, BaseReg1, Offset1, &RI)) {
1938 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand()) {
1939 // FIXME: Handle ds_read2 / ds_write2.
1942 unsigned Width0 = (*MIa.memoperands_begin())->getSize();
1943 unsigned Width1 = (*MIb.memoperands_begin())->getSize();
1944 if (BaseReg0 == BaseReg1 &&
1945 offsetsDoNotOverlap(Width0, Offset0, Width1, Offset1)) {
1953 bool SIInstrInfo::areMemAccessesTriviallyDisjoint(MachineInstr &MIa,
1955 AliasAnalysis *AA) const {
1956 assert((MIa.mayLoad() || MIa.mayStore()) &&
1957 "MIa must load from or modify a memory location");
1958 assert((MIb.mayLoad() || MIb.mayStore()) &&
1959 "MIb must load from or modify a memory location");
1961 if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects())
1964 // XXX - Can we relax this between address spaces?
1965 if (MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef())
1968 if (AA && MIa.hasOneMemOperand() && MIb.hasOneMemOperand()) {
1969 const MachineMemOperand *MMOa = *MIa.memoperands_begin();
1970 const MachineMemOperand *MMOb = *MIb.memoperands_begin();
1971 if (MMOa->getValue() && MMOb->getValue()) {
1972 MemoryLocation LocA(MMOa->getValue(), MMOa->getSize(), MMOa->getAAInfo());
1973 MemoryLocation LocB(MMOb->getValue(), MMOb->getSize(), MMOb->getAAInfo());
1974 if (!AA->alias(LocA, LocB))
1979 // TODO: Should we check the address space from the MachineMemOperand? That
1980 // would allow us to distinguish objects we know don't alias based on the
1981 // underlying address space, even if it was lowered to a different one,
1982 // e.g. private accesses lowered to use MUBUF instructions on a scratch
1986 return checkInstOffsetsDoNotOverlap(MIa, MIb);
1988 return !isFLAT(MIb);
1991 if (isMUBUF(MIa) || isMTBUF(MIa)) {
1992 if (isMUBUF(MIb) || isMTBUF(MIb))
1993 return checkInstOffsetsDoNotOverlap(MIa, MIb);
1995 return !isFLAT(MIb) && !isSMRD(MIb);
2000 return checkInstOffsetsDoNotOverlap(MIa, MIb);
2002 return !isFLAT(MIb) && !isMUBUF(MIa) && !isMTBUF(MIa);
2007 return checkInstOffsetsDoNotOverlap(MIa, MIb);
2015 MachineInstr *SIInstrInfo::convertToThreeAddress(MachineFunction::iterator &MBB,
2017 LiveVariables *LV) const {
2020 switch (MI.getOpcode()) {
2023 case AMDGPU::V_MAC_F16_e64:
2026 case AMDGPU::V_MAC_F32_e64:
2028 case AMDGPU::V_MAC_F16_e32:
2031 case AMDGPU::V_MAC_F32_e32: {
2032 int Src0Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
2033 AMDGPU::OpName::src0);
2034 const MachineOperand *Src0 = &MI.getOperand(Src0Idx);
2035 if (Src0->isImm() && !isInlineConstant(MI, Src0Idx, *Src0))
2041 const MachineOperand *Dst = getNamedOperand(MI, AMDGPU::OpName::vdst);
2042 const MachineOperand *Src0 = getNamedOperand(MI, AMDGPU::OpName::src0);
2043 const MachineOperand *Src0Mods =
2044 getNamedOperand(MI, AMDGPU::OpName::src0_modifiers);
2045 const MachineOperand *Src1 = getNamedOperand(MI, AMDGPU::OpName::src1);
2046 const MachineOperand *Src1Mods =
2047 getNamedOperand(MI, AMDGPU::OpName::src1_modifiers);
2048 const MachineOperand *Src2 = getNamedOperand(MI, AMDGPU::OpName::src2);
2049 const MachineOperand *Clamp = getNamedOperand(MI, AMDGPU::OpName::clamp);
2050 const MachineOperand *Omod = getNamedOperand(MI, AMDGPU::OpName::omod);
2052 return BuildMI(*MBB, MI, MI.getDebugLoc(),
2053 get(IsF16 ? AMDGPU::V_MAD_F16 : AMDGPU::V_MAD_F32))
2055 .addImm(Src0Mods ? Src0Mods->getImm() : 0)
2057 .addImm(Src1Mods ? Src1Mods->getImm() : 0)
2059 .addImm(0) // Src mods
2061 .addImm(Clamp ? Clamp->getImm() : 0)
2062 .addImm(Omod ? Omod->getImm() : 0);
2065 // It's not generally safe to move VALU instructions across these since it will
2066 // start using the register as a base index rather than directly.
2067 // XXX - Why isn't hasSideEffects sufficient for these?
2068 static bool changesVGPRIndexingMode(const MachineInstr &MI) {
2069 switch (MI.getOpcode()) {
2070 case AMDGPU::S_SET_GPR_IDX_ON:
2071 case AMDGPU::S_SET_GPR_IDX_MODE:
2072 case AMDGPU::S_SET_GPR_IDX_OFF:
2079 bool SIInstrInfo::isSchedulingBoundary(const MachineInstr &MI,
2080 const MachineBasicBlock *MBB,
2081 const MachineFunction &MF) const {
2082 // XXX - Do we want the SP check in the base implementation?
2084 // Target-independent instructions do not have an implicit-use of EXEC, even
2085 // when they operate on VGPRs. Treating EXEC modifications as scheduling
2086 // boundaries prevents incorrect movements of such instructions.
2087 return TargetInstrInfo::isSchedulingBoundary(MI, MBB, MF) ||
2088 MI.modifiesRegister(AMDGPU::EXEC, &RI) ||
2089 MI.getOpcode() == AMDGPU::S_SETREG_IMM32_B32 ||
2090 MI.getOpcode() == AMDGPU::S_SETREG_B32 ||
2091 changesVGPRIndexingMode(MI);
2094 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
2095 switch (Imm.getBitWidth()) {
2097 return AMDGPU::isInlinableLiteral32(Imm.getSExtValue(),
2098 ST.hasInv2PiInlineImm());
2100 return AMDGPU::isInlinableLiteral64(Imm.getSExtValue(),
2101 ST.hasInv2PiInlineImm());
2103 return ST.has16BitInsts() &&
2104 AMDGPU::isInlinableLiteral16(Imm.getSExtValue(),
2105 ST.hasInv2PiInlineImm());
2107 llvm_unreachable("invalid bitwidth");
2111 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO,
2112 uint8_t OperandType) const {
2114 OperandType < AMDGPU::OPERAND_SRC_FIRST ||
2115 OperandType > AMDGPU::OPERAND_SRC_LAST)
2118 // MachineOperand provides no way to tell the true operand size, since it only
2119 // records a 64-bit value. We need to know the size to determine if a 32-bit
2120 // floating point immediate bit pattern is legal for an integer immediate. It
2121 // would be for any 32-bit integer operand, but would not be for a 64-bit one.
2123 int64_t Imm = MO.getImm();
2124 switch (OperandType) {
2125 case AMDGPU::OPERAND_REG_IMM_INT32:
2126 case AMDGPU::OPERAND_REG_IMM_FP32:
2127 case AMDGPU::OPERAND_REG_INLINE_C_INT32:
2128 case AMDGPU::OPERAND_REG_INLINE_C_FP32: {
2129 int32_t Trunc = static_cast<int32_t>(Imm);
2130 return Trunc == Imm &&
2131 AMDGPU::isInlinableLiteral32(Trunc, ST.hasInv2PiInlineImm());
2133 case AMDGPU::OPERAND_REG_IMM_INT64:
2134 case AMDGPU::OPERAND_REG_IMM_FP64:
2135 case AMDGPU::OPERAND_REG_INLINE_C_INT64:
2136 case AMDGPU::OPERAND_REG_INLINE_C_FP64: {
2137 return AMDGPU::isInlinableLiteral64(MO.getImm(),
2138 ST.hasInv2PiInlineImm());
2140 case AMDGPU::OPERAND_REG_IMM_INT16:
2141 case AMDGPU::OPERAND_REG_IMM_FP16:
2142 case AMDGPU::OPERAND_REG_INLINE_C_INT16:
2143 case AMDGPU::OPERAND_REG_INLINE_C_FP16: {
2144 if (isInt<16>(Imm) || isUInt<16>(Imm)) {
2145 // A few special case instructions have 16-bit operands on subtargets
2146 // where 16-bit instructions are not legal.
2147 // TODO: Do the 32-bit immediates work? We shouldn't really need to handle
2148 // constants in these cases
2149 int16_t Trunc = static_cast<int16_t>(Imm);
2150 return ST.has16BitInsts() &&
2151 AMDGPU::isInlinableLiteral16(Trunc, ST.hasInv2PiInlineImm());
2156 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
2157 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: {
2158 uint32_t Trunc = static_cast<uint32_t>(Imm);
2159 return AMDGPU::isInlinableLiteralV216(Trunc, ST.hasInv2PiInlineImm());
2162 llvm_unreachable("invalid bitwidth");
2166 bool SIInstrInfo::isLiteralConstantLike(const MachineOperand &MO,
2167 const MCOperandInfo &OpInfo) const {
2168 switch (MO.getType()) {
2169 case MachineOperand::MO_Register:
2171 case MachineOperand::MO_Immediate:
2172 return !isInlineConstant(MO, OpInfo);
2173 case MachineOperand::MO_FrameIndex:
2174 case MachineOperand::MO_MachineBasicBlock:
2175 case MachineOperand::MO_ExternalSymbol:
2176 case MachineOperand::MO_GlobalAddress:
2177 case MachineOperand::MO_MCSymbol:
2180 llvm_unreachable("unexpected operand type");
2184 static bool compareMachineOp(const MachineOperand &Op0,
2185 const MachineOperand &Op1) {
2186 if (Op0.getType() != Op1.getType())
2189 switch (Op0.getType()) {
2190 case MachineOperand::MO_Register:
2191 return Op0.getReg() == Op1.getReg();
2192 case MachineOperand::MO_Immediate:
2193 return Op0.getImm() == Op1.getImm();
2195 llvm_unreachable("Didn't expect to be comparing these operand types");
2199 bool SIInstrInfo::isImmOperandLegal(const MachineInstr &MI, unsigned OpNo,
2200 const MachineOperand &MO) const {
2201 const MCOperandInfo &OpInfo = get(MI.getOpcode()).OpInfo[OpNo];
2203 assert(MO.isImm() || MO.isTargetIndex() || MO.isFI());
2205 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
2208 if (OpInfo.RegClass < 0)
2211 if (MO.isImm() && isInlineConstant(MO, OpInfo))
2212 return RI.opCanUseInlineConstant(OpInfo.OperandType);
2214 return RI.opCanUseLiteralConstant(OpInfo.OperandType);
2217 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
2218 int Op32 = AMDGPU::getVOPe32(Opcode);
2222 return pseudoToMCOpcode(Op32) != -1;
2225 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
2226 // The src0_modifier operand is present on all instructions
2227 // that have modifiers.
2229 return AMDGPU::getNamedOperandIdx(Opcode,
2230 AMDGPU::OpName::src0_modifiers) != -1;
2233 bool SIInstrInfo::hasModifiersSet(const MachineInstr &MI,
2234 unsigned OpName) const {
2235 const MachineOperand *Mods = getNamedOperand(MI, OpName);
2236 return Mods && Mods->getImm();
2239 bool SIInstrInfo::hasAnyModifiersSet(const MachineInstr &MI) const {
2240 return hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers) ||
2241 hasModifiersSet(MI, AMDGPU::OpName::src1_modifiers) ||
2242 hasModifiersSet(MI, AMDGPU::OpName::src2_modifiers) ||
2243 hasModifiersSet(MI, AMDGPU::OpName::clamp) ||
2244 hasModifiersSet(MI, AMDGPU::OpName::omod);
2247 bool SIInstrInfo::usesConstantBus(const MachineRegisterInfo &MRI,
2248 const MachineOperand &MO,
2249 const MCOperandInfo &OpInfo) const {
2250 // Literal constants use the constant bus.
2251 //if (isLiteralConstantLike(MO, OpInfo))
2254 return !isInlineConstant(MO, OpInfo);
2257 return true; // Misc other operands like FrameIndex
2262 if (TargetRegisterInfo::isVirtualRegister(MO.getReg()))
2263 return RI.isSGPRClass(MRI.getRegClass(MO.getReg()));
2265 // FLAT_SCR is just an SGPR pair.
2266 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
2269 // EXEC register uses the constant bus.
2270 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
2273 // SGPRs use the constant bus
2274 return (MO.getReg() == AMDGPU::VCC || MO.getReg() == AMDGPU::M0 ||
2275 (!MO.isImplicit() &&
2276 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
2277 AMDGPU::SGPR_64RegClass.contains(MO.getReg()))));
2280 static unsigned findImplicitSGPRRead(const MachineInstr &MI) {
2281 for (const MachineOperand &MO : MI.implicit_operands()) {
2282 // We only care about reads.
2286 switch (MO.getReg()) {
2289 case AMDGPU::FLAT_SCR:
2297 return AMDGPU::NoRegister;
2300 static bool shouldReadExec(const MachineInstr &MI) {
2301 if (SIInstrInfo::isVALU(MI)) {
2302 switch (MI.getOpcode()) {
2303 case AMDGPU::V_READLANE_B32:
2304 case AMDGPU::V_READLANE_B32_si:
2305 case AMDGPU::V_READLANE_B32_vi:
2306 case AMDGPU::V_WRITELANE_B32:
2307 case AMDGPU::V_WRITELANE_B32_si:
2308 case AMDGPU::V_WRITELANE_B32_vi:
2315 if (SIInstrInfo::isGenericOpcode(MI.getOpcode()) ||
2316 SIInstrInfo::isSALU(MI) ||
2317 SIInstrInfo::isSMRD(MI))
2323 static bool isSubRegOf(const SIRegisterInfo &TRI,
2324 const MachineOperand &SuperVec,
2325 const MachineOperand &SubReg) {
2326 if (TargetRegisterInfo::isPhysicalRegister(SubReg.getReg()))
2327 return TRI.isSubRegister(SuperVec.getReg(), SubReg.getReg());
2329 return SubReg.getSubReg() != AMDGPU::NoSubRegister &&
2330 SubReg.getReg() == SuperVec.getReg();
2333 bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
2334 StringRef &ErrInfo) const {
2335 uint16_t Opcode = MI.getOpcode();
2336 if (SIInstrInfo::isGenericOpcode(MI.getOpcode()))
2339 const MachineFunction *MF = MI.getParent()->getParent();
2340 const MachineRegisterInfo &MRI = MF->getRegInfo();
2342 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
2343 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
2344 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
2346 // Make sure the number of operands is correct.
2347 const MCInstrDesc &Desc = get(Opcode);
2348 if (!Desc.isVariadic() &&
2349 Desc.getNumOperands() != MI.getNumExplicitOperands()) {
2350 ErrInfo = "Instruction has wrong number of operands.";
2354 if (MI.isInlineAsm()) {
2355 // Verify register classes for inlineasm constraints.
2356 for (unsigned I = InlineAsm::MIOp_FirstOperand, E = MI.getNumOperands();
2358 const TargetRegisterClass *RC = MI.getRegClassConstraint(I, this, &RI);
2362 const MachineOperand &Op = MI.getOperand(I);
2366 unsigned Reg = Op.getReg();
2367 if (!TargetRegisterInfo::isVirtualRegister(Reg) && !RC->contains(Reg)) {
2368 ErrInfo = "inlineasm operand has incorrect register class.";
2376 // Make sure the register classes are correct.
2377 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
2378 if (MI.getOperand(i).isFPImm()) {
2379 ErrInfo = "FPImm Machine Operands are not supported. ISel should bitcast "
2380 "all fp values to integers.";
2384 int RegClass = Desc.OpInfo[i].RegClass;
2386 switch (Desc.OpInfo[i].OperandType) {
2387 case MCOI::OPERAND_REGISTER:
2388 if (MI.getOperand(i).isImm()) {
2389 ErrInfo = "Illegal immediate value for operand.";
2393 case AMDGPU::OPERAND_REG_IMM_INT32:
2394 case AMDGPU::OPERAND_REG_IMM_FP32:
2396 case AMDGPU::OPERAND_REG_INLINE_C_INT32:
2397 case AMDGPU::OPERAND_REG_INLINE_C_FP32:
2398 case AMDGPU::OPERAND_REG_INLINE_C_INT64:
2399 case AMDGPU::OPERAND_REG_INLINE_C_FP64:
2400 case AMDGPU::OPERAND_REG_INLINE_C_INT16:
2401 case AMDGPU::OPERAND_REG_INLINE_C_FP16: {
2402 const MachineOperand &MO = MI.getOperand(i);
2403 if (!MO.isReg() && (!MO.isImm() || !isInlineConstant(MI, i))) {
2404 ErrInfo = "Illegal immediate value for operand.";
2409 case MCOI::OPERAND_IMMEDIATE:
2410 case AMDGPU::OPERAND_KIMM32:
2411 // Check if this operand is an immediate.
2412 // FrameIndex operands will be replaced by immediates, so they are
2414 if (!MI.getOperand(i).isImm() && !MI.getOperand(i).isFI()) {
2415 ErrInfo = "Expected immediate, but got non-immediate";
2423 if (!MI.getOperand(i).isReg())
2426 if (RegClass != -1) {
2427 unsigned Reg = MI.getOperand(i).getReg();
2428 if (Reg == AMDGPU::NoRegister ||
2429 TargetRegisterInfo::isVirtualRegister(Reg))
2432 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
2433 if (!RC->contains(Reg)) {
2434 ErrInfo = "Operand has incorrect register class.";
2443 if (!ST.hasSDWA()) {
2444 ErrInfo = "SDWA is not supported on this target";
2448 int DstIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vdst);
2450 const int OpIndicies[] = { DstIdx, Src0Idx, Src1Idx, Src2Idx };
2452 for (int OpIdx: OpIndicies) {
2455 const MachineOperand &MO = MI.getOperand(OpIdx);
2457 if (!ST.hasSDWAScalar()) {
2459 if (!MO.isReg() || !RI.hasVGPRs(RI.getRegClassForReg(MRI, MO.getReg()))) {
2460 ErrInfo = "Only VGPRs allowed as operands in SDWA instructions on VI";
2464 // No immediates on GFX9
2466 ErrInfo = "Only reg allowed as operands in SDWA instructions on GFX9";
2472 if (!ST.hasSDWAOmod()) {
2473 // No omod allowed on VI
2474 const MachineOperand *OMod = getNamedOperand(MI, AMDGPU::OpName::omod);
2475 if (OMod != nullptr &&
2476 (!OMod->isImm() || OMod->getImm() != 0)) {
2477 ErrInfo = "OMod not allowed in SDWA instructions on VI";
2482 uint16_t BasicOpcode = AMDGPU::getBasicFromSDWAOp(Opcode);
2483 if (isVOPC(BasicOpcode)) {
2484 if (!ST.hasSDWASdst() && DstIdx != -1) {
2485 // Only vcc allowed as dst on VI for VOPC
2486 const MachineOperand &Dst = MI.getOperand(DstIdx);
2487 if (!Dst.isReg() || Dst.getReg() != AMDGPU::VCC) {
2488 ErrInfo = "Only VCC allowed as dst in SDWA instructions on VI";
2491 } else if (!ST.hasSDWAOutModsVOPC()) {
2492 // No clamp allowed on GFX9 for VOPC
2493 const MachineOperand *Clamp = getNamedOperand(MI, AMDGPU::OpName::clamp);
2494 if (Clamp && (!Clamp->isImm() || Clamp->getImm() != 0)) {
2495 ErrInfo = "Clamp not allowed in VOPC SDWA instructions on VI";
2499 // No omod allowed on GFX9 for VOPC
2500 const MachineOperand *OMod = getNamedOperand(MI, AMDGPU::OpName::omod);
2501 if (OMod && (!OMod->isImm() || OMod->getImm() != 0)) {
2502 ErrInfo = "OMod not allowed in VOPC SDWA instructions on VI";
2510 if (isVOP1(MI) || isVOP2(MI) || isVOP3(MI) || isVOPC(MI) || isSDWA(MI)) {
2511 // Only look at the true operands. Only a real operand can use the constant
2512 // bus, and we don't want to check pseudo-operands like the source modifier
2514 const int OpIndices[] = { Src0Idx, Src1Idx, Src2Idx };
2516 unsigned ConstantBusCount = 0;
2518 if (AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::imm) != -1)
2521 unsigned SGPRUsed = findImplicitSGPRRead(MI);
2522 if (SGPRUsed != AMDGPU::NoRegister)
2525 for (int OpIdx : OpIndices) {
2528 const MachineOperand &MO = MI.getOperand(OpIdx);
2529 if (usesConstantBus(MRI, MO, MI.getDesc().OpInfo[OpIdx])) {
2531 if (MO.getReg() != SGPRUsed)
2533 SGPRUsed = MO.getReg();
2539 if (ConstantBusCount > 1) {
2540 ErrInfo = "VOP* instruction uses the constant bus more than once";
2545 // Verify misc. restrictions on specific instructions.
2546 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
2547 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
2548 const MachineOperand &Src0 = MI.getOperand(Src0Idx);
2549 const MachineOperand &Src1 = MI.getOperand(Src1Idx);
2550 const MachineOperand &Src2 = MI.getOperand(Src2Idx);
2551 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
2552 if (!compareMachineOp(Src0, Src1) &&
2553 !compareMachineOp(Src0, Src2)) {
2554 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
2561 int64_t Imm = getNamedOperand(MI, AMDGPU::OpName::simm16)->getImm();
2562 if (sopkIsZext(MI)) {
2563 if (!isUInt<16>(Imm)) {
2564 ErrInfo = "invalid immediate for SOPK instruction";
2568 if (!isInt<16>(Imm)) {
2569 ErrInfo = "invalid immediate for SOPK instruction";
2575 if (Desc.getOpcode() == AMDGPU::V_MOVRELS_B32_e32 ||
2576 Desc.getOpcode() == AMDGPU::V_MOVRELS_B32_e64 ||
2577 Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e32 ||
2578 Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e64) {
2579 const bool IsDst = Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e32 ||
2580 Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e64;
2582 const unsigned StaticNumOps = Desc.getNumOperands() +
2583 Desc.getNumImplicitUses();
2584 const unsigned NumImplicitOps = IsDst ? 2 : 1;
2586 // Allow additional implicit operands. This allows a fixup done by the post
2587 // RA scheduler where the main implicit operand is killed and implicit-defs
2588 // are added for sub-registers that remain live after this instruction.
2589 if (MI.getNumOperands() < StaticNumOps + NumImplicitOps) {
2590 ErrInfo = "missing implicit register operands";
2594 const MachineOperand *Dst = getNamedOperand(MI, AMDGPU::OpName::vdst);
2596 if (!Dst->isUse()) {
2597 ErrInfo = "v_movreld_b32 vdst should be a use operand";
2602 if (!MI.isRegTiedToUseOperand(StaticNumOps, &UseOpIdx) ||
2603 UseOpIdx != StaticNumOps + 1) {
2604 ErrInfo = "movrel implicit operands should be tied";
2609 const MachineOperand &Src0 = MI.getOperand(Src0Idx);
2610 const MachineOperand &ImpUse
2611 = MI.getOperand(StaticNumOps + NumImplicitOps - 1);
2612 if (!ImpUse.isReg() || !ImpUse.isUse() ||
2613 !isSubRegOf(RI, ImpUse, IsDst ? *Dst : Src0)) {
2614 ErrInfo = "src0 should be subreg of implicit vector use";
2619 // Make sure we aren't losing exec uses in the td files. This mostly requires
2620 // being careful when using let Uses to try to add other use registers.
2621 if (shouldReadExec(MI)) {
2622 if (!MI.hasRegisterImplicitUseOperand(AMDGPU::EXEC)) {
2623 ErrInfo = "VALU instruction does not implicitly read exec mask";
2629 if (MI.mayStore()) {
2630 // The register offset form of scalar stores may only use m0 as the
2631 // soffset register.
2632 const MachineOperand *Soff = getNamedOperand(MI, AMDGPU::OpName::soff);
2633 if (Soff && Soff->getReg() != AMDGPU::M0) {
2634 ErrInfo = "scalar stores must use m0 as offset register";
2640 if (isFLAT(MI) && !MF->getSubtarget<SISubtarget>().hasFlatInstOffsets()) {
2641 const MachineOperand *Offset = getNamedOperand(MI, AMDGPU::OpName::offset);
2642 if (Offset->getImm() != 0) {
2643 ErrInfo = "subtarget does not support offsets in flat instructions";
2651 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
2652 switch (MI.getOpcode()) {
2653 default: return AMDGPU::INSTRUCTION_LIST_END;
2654 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
2655 case AMDGPU::COPY: return AMDGPU::COPY;
2656 case AMDGPU::PHI: return AMDGPU::PHI;
2657 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
2658 case AMDGPU::S_MOV_B32:
2659 return MI.getOperand(1).isReg() ?
2660 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
2661 case AMDGPU::S_ADD_I32:
2662 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
2663 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
2664 case AMDGPU::S_SUB_I32:
2665 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
2666 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
2667 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
2668 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e64;
2669 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e64;
2670 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e64;
2671 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e64;
2672 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e64;
2673 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e64;
2674 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e64;
2675 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
2676 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
2677 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
2678 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
2679 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
2680 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
2681 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
2682 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
2683 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
2684 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
2685 case AMDGPU::S_BFM_B32: return AMDGPU::V_BFM_B32_e64;
2686 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
2687 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
2688 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
2689 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
2690 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
2691 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
2692 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
2693 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
2694 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
2695 case AMDGPU::S_CMP_EQ_U32: return AMDGPU::V_CMP_EQ_U32_e32;
2696 case AMDGPU::S_CMP_LG_U32: return AMDGPU::V_CMP_NE_U32_e32;
2697 case AMDGPU::S_CMP_GT_U32: return AMDGPU::V_CMP_GT_U32_e32;
2698 case AMDGPU::S_CMP_GE_U32: return AMDGPU::V_CMP_GE_U32_e32;
2699 case AMDGPU::S_CMP_LT_U32: return AMDGPU::V_CMP_LT_U32_e32;
2700 case AMDGPU::S_CMP_LE_U32: return AMDGPU::V_CMP_LE_U32_e32;
2701 case AMDGPU::S_CMP_EQ_U64: return AMDGPU::V_CMP_EQ_U64_e32;
2702 case AMDGPU::S_CMP_LG_U64: return AMDGPU::V_CMP_NE_U64_e32;
2703 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e64;
2704 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
2705 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
2706 case AMDGPU::S_FLBIT_I32: return AMDGPU::V_FFBH_I32_e64;
2707 case AMDGPU::S_CBRANCH_SCC0: return AMDGPU::S_CBRANCH_VCCZ;
2708 case AMDGPU::S_CBRANCH_SCC1: return AMDGPU::S_CBRANCH_VCCNZ;
2712 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
2713 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
2716 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
2717 unsigned OpNo) const {
2718 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
2719 const MCInstrDesc &Desc = get(MI.getOpcode());
2720 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
2721 Desc.OpInfo[OpNo].RegClass == -1) {
2722 unsigned Reg = MI.getOperand(OpNo).getReg();
2724 if (TargetRegisterInfo::isVirtualRegister(Reg))
2725 return MRI.getRegClass(Reg);
2726 return RI.getPhysRegClass(Reg);
2729 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
2730 return RI.getRegClass(RCID);
2733 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
2734 switch (MI.getOpcode()) {
2736 case AMDGPU::REG_SEQUENCE:
2738 case AMDGPU::INSERT_SUBREG:
2739 return RI.hasVGPRs(getOpRegClass(MI, 0));
2741 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
2745 void SIInstrInfo::legalizeOpWithMove(MachineInstr &MI, unsigned OpIdx) const {
2746 MachineBasicBlock::iterator I = MI;
2747 MachineBasicBlock *MBB = MI.getParent();
2748 MachineOperand &MO = MI.getOperand(OpIdx);
2749 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
2750 unsigned RCID = get(MI.getOpcode()).OpInfo[OpIdx].RegClass;
2751 const TargetRegisterClass *RC = RI.getRegClass(RCID);
2752 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
2754 Opcode = AMDGPU::COPY;
2755 else if (RI.isSGPRClass(RC))
2756 Opcode = AMDGPU::S_MOV_B32;
2758 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
2759 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC))
2760 VRC = &AMDGPU::VReg_64RegClass;
2762 VRC = &AMDGPU::VGPR_32RegClass;
2764 unsigned Reg = MRI.createVirtualRegister(VRC);
2765 DebugLoc DL = MBB->findDebugLoc(I);
2766 BuildMI(*MI.getParent(), I, DL, get(Opcode), Reg).add(MO);
2767 MO.ChangeToRegister(Reg, false);
2770 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
2771 MachineRegisterInfo &MRI,
2772 MachineOperand &SuperReg,
2773 const TargetRegisterClass *SuperRC,
2775 const TargetRegisterClass *SubRC)
2777 MachineBasicBlock *MBB = MI->getParent();
2778 DebugLoc DL = MI->getDebugLoc();
2779 unsigned SubReg = MRI.createVirtualRegister(SubRC);
2781 if (SuperReg.getSubReg() == AMDGPU::NoSubRegister) {
2782 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
2783 .addReg(SuperReg.getReg(), 0, SubIdx);
2787 // Just in case the super register is itself a sub-register, copy it to a new
2788 // value so we don't need to worry about merging its subreg index with the
2789 // SubIdx passed to this function. The register coalescer should be able to
2790 // eliminate this extra copy.
2791 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
2793 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), NewSuperReg)
2794 .addReg(SuperReg.getReg(), 0, SuperReg.getSubReg());
2796 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
2797 .addReg(NewSuperReg, 0, SubIdx);
2802 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
2803 MachineBasicBlock::iterator MII,
2804 MachineRegisterInfo &MRI,
2806 const TargetRegisterClass *SuperRC,
2808 const TargetRegisterClass *SubRC) const {
2810 if (SubIdx == AMDGPU::sub0)
2811 return MachineOperand::CreateImm(static_cast<int32_t>(Op.getImm()));
2812 if (SubIdx == AMDGPU::sub1)
2813 return MachineOperand::CreateImm(static_cast<int32_t>(Op.getImm() >> 32));
2815 llvm_unreachable("Unhandled register index for immediate");
2818 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
2820 return MachineOperand::CreateReg(SubReg, false);
2823 // Change the order of operands from (0, 1, 2) to (0, 2, 1)
2824 void SIInstrInfo::swapOperands(MachineInstr &Inst) const {
2825 assert(Inst.getNumExplicitOperands() == 3);
2826 MachineOperand Op1 = Inst.getOperand(1);
2827 Inst.RemoveOperand(1);
2828 Inst.addOperand(Op1);
2831 bool SIInstrInfo::isLegalRegOperand(const MachineRegisterInfo &MRI,
2832 const MCOperandInfo &OpInfo,
2833 const MachineOperand &MO) const {
2837 unsigned Reg = MO.getReg();
2838 const TargetRegisterClass *RC =
2839 TargetRegisterInfo::isVirtualRegister(Reg) ?
2840 MRI.getRegClass(Reg) :
2841 RI.getPhysRegClass(Reg);
2843 const SIRegisterInfo *TRI =
2844 static_cast<const SIRegisterInfo*>(MRI.getTargetRegisterInfo());
2845 RC = TRI->getSubRegClass(RC, MO.getSubReg());
2847 // In order to be legal, the common sub-class must be equal to the
2848 // class of the current operand. For example:
2850 // v_mov_b32 s0 ; Operand defined as vsrc_b32
2851 // ; RI.getCommonSubClass(s0,vsrc_b32) = sgpr ; LEGAL
2853 // s_sendmsg 0, s0 ; Operand defined as m0reg
2854 // ; RI.getCommonSubClass(s0,m0reg) = m0reg ; NOT LEGAL
2856 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass)) == RC;
2859 bool SIInstrInfo::isLegalVSrcOperand(const MachineRegisterInfo &MRI,
2860 const MCOperandInfo &OpInfo,
2861 const MachineOperand &MO) const {
2863 return isLegalRegOperand(MRI, OpInfo, MO);
2865 // Handle non-register types that are treated like immediates.
2866 assert(MO.isImm() || MO.isTargetIndex() || MO.isFI());
2870 bool SIInstrInfo::isOperandLegal(const MachineInstr &MI, unsigned OpIdx,
2871 const MachineOperand *MO) const {
2872 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
2873 const MCInstrDesc &InstDesc = MI.getDesc();
2874 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
2875 const TargetRegisterClass *DefinedRC =
2876 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
2878 MO = &MI.getOperand(OpIdx);
2880 if (isVALU(MI) && usesConstantBus(MRI, *MO, OpInfo)) {
2882 RegSubRegPair SGPRUsed;
2884 SGPRUsed = RegSubRegPair(MO->getReg(), MO->getSubReg());
2886 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
2889 const MachineOperand &Op = MI.getOperand(i);
2891 if ((Op.getReg() != SGPRUsed.Reg || Op.getSubReg() != SGPRUsed.SubReg) &&
2892 usesConstantBus(MRI, Op, InstDesc.OpInfo[i])) {
2895 } else if (InstDesc.OpInfo[i].OperandType == AMDGPU::OPERAND_KIMM32) {
2903 return isLegalRegOperand(MRI, OpInfo, *MO);
2906 // Handle non-register types that are treated like immediates.
2907 assert(MO->isImm() || MO->isTargetIndex() || MO->isFI());
2910 // This operand expects an immediate.
2914 return isImmOperandLegal(MI, OpIdx, *MO);
2917 void SIInstrInfo::legalizeOperandsVOP2(MachineRegisterInfo &MRI,
2918 MachineInstr &MI) const {
2919 unsigned Opc = MI.getOpcode();
2920 const MCInstrDesc &InstrDesc = get(Opc);
2922 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
2923 MachineOperand &Src1 = MI.getOperand(Src1Idx);
2925 // If there is an implicit SGPR use such as VCC use for v_addc_u32/v_subb_u32
2926 // we need to only have one constant bus use.
2928 // Note we do not need to worry about literal constants here. They are
2929 // disabled for the operand type for instructions because they will always
2930 // violate the one constant bus use rule.
2931 bool HasImplicitSGPR = findImplicitSGPRRead(MI) != AMDGPU::NoRegister;
2932 if (HasImplicitSGPR) {
2933 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
2934 MachineOperand &Src0 = MI.getOperand(Src0Idx);
2936 if (Src0.isReg() && RI.isSGPRReg(MRI, Src0.getReg()))
2937 legalizeOpWithMove(MI, Src0Idx);
2940 // VOP2 src0 instructions support all operand types, so we don't need to check
2941 // their legality. If src1 is already legal, we don't need to do anything.
2942 if (isLegalRegOperand(MRI, InstrDesc.OpInfo[Src1Idx], Src1))
2945 // Special case: V_READLANE_B32 accepts only immediate or SGPR operands for
2946 // lane select. Fix up using V_READFIRSTLANE, since we assume that the lane
2947 // select is uniform.
2948 if (Opc == AMDGPU::V_READLANE_B32 && Src1.isReg() &&
2949 RI.isVGPR(MRI, Src1.getReg())) {
2950 unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
2951 const DebugLoc &DL = MI.getDebugLoc();
2952 BuildMI(*MI.getParent(), MI, DL, get(AMDGPU::V_READFIRSTLANE_B32), Reg)
2954 Src1.ChangeToRegister(Reg, false);
2958 // We do not use commuteInstruction here because it is too aggressive and will
2959 // commute if it is possible. We only want to commute here if it improves
2960 // legality. This can be called a fairly large number of times so don't waste
2961 // compile time pointlessly swapping and checking legality again.
2962 if (HasImplicitSGPR || !MI.isCommutable()) {
2963 legalizeOpWithMove(MI, Src1Idx);
2967 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
2968 MachineOperand &Src0 = MI.getOperand(Src0Idx);
2970 // If src0 can be used as src1, commuting will make the operands legal.
2971 // Otherwise we have to give up and insert a move.
2973 // TODO: Other immediate-like operand kinds could be commuted if there was a
2974 // MachineOperand::ChangeTo* for them.
2975 if ((!Src1.isImm() && !Src1.isReg()) ||
2976 !isLegalRegOperand(MRI, InstrDesc.OpInfo[Src1Idx], Src0)) {
2977 legalizeOpWithMove(MI, Src1Idx);
2981 int CommutedOpc = commuteOpcode(MI);
2982 if (CommutedOpc == -1) {
2983 legalizeOpWithMove(MI, Src1Idx);
2987 MI.setDesc(get(CommutedOpc));
2989 unsigned Src0Reg = Src0.getReg();
2990 unsigned Src0SubReg = Src0.getSubReg();
2991 bool Src0Kill = Src0.isKill();
2994 Src0.ChangeToImmediate(Src1.getImm());
2995 else if (Src1.isReg()) {
2996 Src0.ChangeToRegister(Src1.getReg(), false, false, Src1.isKill());
2997 Src0.setSubReg(Src1.getSubReg());
2999 llvm_unreachable("Should only have register or immediate operands");
3001 Src1.ChangeToRegister(Src0Reg, false, false, Src0Kill);
3002 Src1.setSubReg(Src0SubReg);
3005 // Legalize VOP3 operands. Because all operand types are supported for any
3006 // operand, and since literal constants are not allowed and should never be
3007 // seen, we only need to worry about inserting copies if we use multiple SGPR
3009 void SIInstrInfo::legalizeOperandsVOP3(MachineRegisterInfo &MRI,
3010 MachineInstr &MI) const {
3011 unsigned Opc = MI.getOpcode();
3014 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0),
3015 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1),
3016 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2)
3019 // Find the one SGPR operand we are allowed to use.
3020 unsigned SGPRReg = findUsedSGPR(MI, VOP3Idx);
3022 for (unsigned i = 0; i < 3; ++i) {
3023 int Idx = VOP3Idx[i];
3026 MachineOperand &MO = MI.getOperand(Idx);
3028 // We should never see a VOP3 instruction with an illegal immediate operand.
3032 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
3033 continue; // VGPRs are legal
3035 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
3036 SGPRReg = MO.getReg();
3037 // We can use one SGPR in each VOP3 instruction.
3041 // If we make it this far, then the operand is not legal and we must
3043 legalizeOpWithMove(MI, Idx);
3047 unsigned SIInstrInfo::readlaneVGPRToSGPR(unsigned SrcReg, MachineInstr &UseMI,
3048 MachineRegisterInfo &MRI) const {
3049 const TargetRegisterClass *VRC = MRI.getRegClass(SrcReg);
3050 const TargetRegisterClass *SRC = RI.getEquivalentSGPRClass(VRC);
3051 unsigned DstReg = MRI.createVirtualRegister(SRC);
3052 unsigned SubRegs = RI.getRegSizeInBits(*VRC) / 32;
3054 SmallVector<unsigned, 8> SRegs;
3055 for (unsigned i = 0; i < SubRegs; ++i) {
3056 unsigned SGPR = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
3057 BuildMI(*UseMI.getParent(), UseMI, UseMI.getDebugLoc(),
3058 get(AMDGPU::V_READFIRSTLANE_B32), SGPR)
3059 .addReg(SrcReg, 0, RI.getSubRegFromChannel(i));
3060 SRegs.push_back(SGPR);
3063 MachineInstrBuilder MIB =
3064 BuildMI(*UseMI.getParent(), UseMI, UseMI.getDebugLoc(),
3065 get(AMDGPU::REG_SEQUENCE), DstReg);
3066 for (unsigned i = 0; i < SubRegs; ++i) {
3067 MIB.addReg(SRegs[i]);
3068 MIB.addImm(RI.getSubRegFromChannel(i));
3073 void SIInstrInfo::legalizeOperandsSMRD(MachineRegisterInfo &MRI,
3074 MachineInstr &MI) const {
3076 // If the pointer is store in VGPRs, then we need to move them to
3077 // SGPRs using v_readfirstlane. This is safe because we only select
3078 // loads with uniform pointers to SMRD instruction so we know the
3079 // pointer value is uniform.
3080 MachineOperand *SBase = getNamedOperand(MI, AMDGPU::OpName::sbase);
3081 if (SBase && !RI.isSGPRClass(MRI.getRegClass(SBase->getReg()))) {
3082 unsigned SGPR = readlaneVGPRToSGPR(SBase->getReg(), MI, MRI);
3083 SBase->setReg(SGPR);
3087 void SIInstrInfo::legalizeGenericOperand(MachineBasicBlock &InsertMBB,
3088 MachineBasicBlock::iterator I,
3089 const TargetRegisterClass *DstRC,
3091 MachineRegisterInfo &MRI,
3092 const DebugLoc &DL) const {
3094 unsigned OpReg = Op.getReg();
3095 unsigned OpSubReg = Op.getSubReg();
3097 const TargetRegisterClass *OpRC = RI.getSubClassWithSubReg(
3098 RI.getRegClassForReg(MRI, OpReg), OpSubReg);
3100 // Check if operand is already the correct register class.
3104 unsigned DstReg = MRI.createVirtualRegister(DstRC);
3105 MachineInstr *Copy =
3106 BuildMI(InsertMBB, I, DL, get(AMDGPU::COPY), DstReg).add(Op);
3111 MachineInstr *Def = MRI.getVRegDef(OpReg);
3115 // Try to eliminate the copy if it is copying an immediate value.
3116 if (Def->isMoveImmediate())
3117 FoldImmediate(*Copy, *Def, OpReg, &MRI);
3120 void SIInstrInfo::legalizeOperands(MachineInstr &MI) const {
3121 MachineFunction &MF = *MI.getParent()->getParent();
3122 MachineRegisterInfo &MRI = MF.getRegInfo();
3125 if (isVOP2(MI) || isVOPC(MI)) {
3126 legalizeOperandsVOP2(MRI, MI);
3132 legalizeOperandsVOP3(MRI, MI);
3138 legalizeOperandsSMRD(MRI, MI);
3142 // Legalize REG_SEQUENCE and PHI
3143 // The register class of the operands much be the same type as the register
3144 // class of the output.
3145 if (MI.getOpcode() == AMDGPU::PHI) {
3146 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
3147 for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2) {
3148 if (!MI.getOperand(i).isReg() ||
3149 !TargetRegisterInfo::isVirtualRegister(MI.getOperand(i).getReg()))
3151 const TargetRegisterClass *OpRC =
3152 MRI.getRegClass(MI.getOperand(i).getReg());
3153 if (RI.hasVGPRs(OpRC)) {
3160 // If any of the operands are VGPR registers, then they all most be
3161 // otherwise we will create illegal VGPR->SGPR copies when legalizing
3163 if (VRC || !RI.isSGPRClass(getOpRegClass(MI, 0))) {
3166 VRC = RI.getEquivalentVGPRClass(SRC);
3173 // Update all the operands so they have the same type.
3174 for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) {
3175 MachineOperand &Op = MI.getOperand(I);
3176 if (!Op.isReg() || !TargetRegisterInfo::isVirtualRegister(Op.getReg()))
3179 // MI is a PHI instruction.
3180 MachineBasicBlock *InsertBB = MI.getOperand(I + 1).getMBB();
3181 MachineBasicBlock::iterator Insert = InsertBB->getFirstTerminator();
3183 // Avoid creating no-op copies with the same src and dst reg class. These
3184 // confuse some of the machine passes.
3185 legalizeGenericOperand(*InsertBB, Insert, RC, Op, MRI, MI.getDebugLoc());
3189 // REG_SEQUENCE doesn't really require operand legalization, but if one has a
3190 // VGPR dest type and SGPR sources, insert copies so all operands are
3191 // VGPRs. This seems to help operand folding / the register coalescer.
3192 if (MI.getOpcode() == AMDGPU::REG_SEQUENCE) {
3193 MachineBasicBlock *MBB = MI.getParent();
3194 const TargetRegisterClass *DstRC = getOpRegClass(MI, 0);
3195 if (RI.hasVGPRs(DstRC)) {
3196 // Update all the operands so they are VGPR register classes. These may
3197 // not be the same register class because REG_SEQUENCE supports mixing
3198 // subregister index types e.g. sub0_sub1 + sub2 + sub3
3199 for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) {
3200 MachineOperand &Op = MI.getOperand(I);
3201 if (!Op.isReg() || !TargetRegisterInfo::isVirtualRegister(Op.getReg()))
3204 const TargetRegisterClass *OpRC = MRI.getRegClass(Op.getReg());
3205 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(OpRC);
3209 legalizeGenericOperand(*MBB, MI, VRC, Op, MRI, MI.getDebugLoc());
3217 // Legalize INSERT_SUBREG
3218 // src0 must have the same register class as dst
3219 if (MI.getOpcode() == AMDGPU::INSERT_SUBREG) {
3220 unsigned Dst = MI.getOperand(0).getReg();
3221 unsigned Src0 = MI.getOperand(1).getReg();
3222 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
3223 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
3224 if (DstRC != Src0RC) {
3225 MachineBasicBlock *MBB = MI.getParent();
3226 MachineOperand &Op = MI.getOperand(1);
3227 legalizeGenericOperand(*MBB, MI, DstRC, Op, MRI, MI.getDebugLoc());
3232 // Legalize MIMG and MUBUF/MTBUF for shaders.
3234 // Shaders only generate MUBUF/MTBUF instructions via intrinsics or via
3235 // scratch memory access. In both cases, the legalization never involves
3236 // conversion to the addr64 form.
3238 (AMDGPU::isShader(MF.getFunction()->getCallingConv()) &&
3239 (isMUBUF(MI) || isMTBUF(MI)))) {
3240 MachineOperand *SRsrc = getNamedOperand(MI, AMDGPU::OpName::srsrc);
3241 if (SRsrc && !RI.isSGPRClass(MRI.getRegClass(SRsrc->getReg()))) {
3242 unsigned SGPR = readlaneVGPRToSGPR(SRsrc->getReg(), MI, MRI);
3243 SRsrc->setReg(SGPR);
3246 MachineOperand *SSamp = getNamedOperand(MI, AMDGPU::OpName::ssamp);
3247 if (SSamp && !RI.isSGPRClass(MRI.getRegClass(SSamp->getReg()))) {
3248 unsigned SGPR = readlaneVGPRToSGPR(SSamp->getReg(), MI, MRI);
3249 SSamp->setReg(SGPR);
3254 // Legalize MUBUF* instructions by converting to addr64 form.
3255 // FIXME: If we start using the non-addr64 instructions for compute, we
3256 // may need to legalize them as above. This especially applies to the
3257 // buffer_load_format_* variants and variants with idxen (or bothen).
3259 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::srsrc);
3260 if (SRsrcIdx != -1) {
3261 // We have an MUBUF instruction
3262 MachineOperand *SRsrc = &MI.getOperand(SRsrcIdx);
3263 unsigned SRsrcRC = get(MI.getOpcode()).OpInfo[SRsrcIdx].RegClass;
3264 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
3265 RI.getRegClass(SRsrcRC))) {
3266 // The operands are legal.
3267 // FIXME: We may need to legalize operands besided srsrc.
3271 MachineBasicBlock &MBB = *MI.getParent();
3273 // Extract the ptr from the resource descriptor.
3274 unsigned SRsrcPtr = buildExtractSubReg(MI, MRI, *SRsrc,
3275 &AMDGPU::VReg_128RegClass, AMDGPU::sub0_sub1, &AMDGPU::VReg_64RegClass);
3277 // Create an empty resource descriptor
3278 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
3279 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
3280 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
3281 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
3282 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
3285 BuildMI(MBB, MI, MI.getDebugLoc(), get(AMDGPU::S_MOV_B64), Zero64)
3288 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
3289 BuildMI(MBB, MI, MI.getDebugLoc(), get(AMDGPU::S_MOV_B32), SRsrcFormatLo)
3290 .addImm(RsrcDataFormat & 0xFFFFFFFF);
3292 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
3293 BuildMI(MBB, MI, MI.getDebugLoc(), get(AMDGPU::S_MOV_B32), SRsrcFormatHi)
3294 .addImm(RsrcDataFormat >> 32);
3296 // NewSRsrc = {Zero64, SRsrcFormat}
3297 BuildMI(MBB, MI, MI.getDebugLoc(), get(AMDGPU::REG_SEQUENCE), NewSRsrc)
3299 .addImm(AMDGPU::sub0_sub1)
3300 .addReg(SRsrcFormatLo)
3301 .addImm(AMDGPU::sub2)
3302 .addReg(SRsrcFormatHi)
3303 .addImm(AMDGPU::sub3);
3305 MachineOperand *VAddr = getNamedOperand(MI, AMDGPU::OpName::vaddr);
3306 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
3308 // This is already an ADDR64 instruction so we need to add the pointer
3309 // extracted from the resource descriptor to the current value of VAddr.
3310 unsigned NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3311 unsigned NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3313 // NewVaddrLo = SRsrcPtr:sub0 + VAddr:sub0
3314 DebugLoc DL = MI.getDebugLoc();
3315 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_I32_e32), NewVAddrLo)
3316 .addReg(SRsrcPtr, 0, AMDGPU::sub0)
3317 .addReg(VAddr->getReg(), 0, AMDGPU::sub0);
3319 // NewVaddrHi = SRsrcPtr:sub1 + VAddr:sub1
3320 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADDC_U32_e32), NewVAddrHi)
3321 .addReg(SRsrcPtr, 0, AMDGPU::sub1)
3322 .addReg(VAddr->getReg(), 0, AMDGPU::sub1);
3324 // NewVaddr = {NewVaddrHi, NewVaddrLo}
3325 BuildMI(MBB, MI, MI.getDebugLoc(), get(AMDGPU::REG_SEQUENCE), NewVAddr)
3327 .addImm(AMDGPU::sub0)
3329 .addImm(AMDGPU::sub1);
3331 // This instructions is the _OFFSET variant, so we need to convert it to
3333 assert(MBB.getParent()->getSubtarget<SISubtarget>().getGeneration()
3334 < SISubtarget::VOLCANIC_ISLANDS &&
3335 "FIXME: Need to emit flat atomics here");
3337 MachineOperand *VData = getNamedOperand(MI, AMDGPU::OpName::vdata);
3338 MachineOperand *Offset = getNamedOperand(MI, AMDGPU::OpName::offset);
3339 MachineOperand *SOffset = getNamedOperand(MI, AMDGPU::OpName::soffset);
3340 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI.getOpcode());
3342 // Atomics rith return have have an additional tied operand and are
3343 // missing some of the special bits.
3344 MachineOperand *VDataIn = getNamedOperand(MI, AMDGPU::OpName::vdata_in);
3345 MachineInstr *Addr64;
3348 // Regular buffer load / store.
3349 MachineInstrBuilder MIB =
3350 BuildMI(MBB, MI, MI.getDebugLoc(), get(Addr64Opcode))
3352 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
3353 // This will be replaced later
3354 // with the new value of vaddr.
3359 // Atomics do not have this operand.
3360 if (const MachineOperand *GLC =
3361 getNamedOperand(MI, AMDGPU::OpName::glc)) {
3362 MIB.addImm(GLC->getImm());
3365 MIB.addImm(getNamedImmOperand(MI, AMDGPU::OpName::slc));
3367 if (const MachineOperand *TFE =
3368 getNamedOperand(MI, AMDGPU::OpName::tfe)) {
3369 MIB.addImm(TFE->getImm());
3372 MIB.setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
3375 // Atomics with return.
3376 Addr64 = BuildMI(MBB, MI, MI.getDebugLoc(), get(Addr64Opcode))
3379 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
3380 // This will be replaced later
3381 // with the new value of vaddr.
3385 .addImm(getNamedImmOperand(MI, AMDGPU::OpName::slc))
3386 .setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
3389 MI.removeFromParent();
3391 // NewVaddr = {NewVaddrHi, NewVaddrLo}
3392 BuildMI(MBB, Addr64, Addr64->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
3394 .addReg(SRsrcPtr, 0, AMDGPU::sub0)
3395 .addImm(AMDGPU::sub0)
3396 .addReg(SRsrcPtr, 0, AMDGPU::sub1)
3397 .addImm(AMDGPU::sub1);
3399 VAddr = getNamedOperand(*Addr64, AMDGPU::OpName::vaddr);
3400 SRsrc = getNamedOperand(*Addr64, AMDGPU::OpName::srsrc);
3403 // Update the instruction to use NewVaddr
3404 VAddr->setReg(NewVAddr);
3405 // Update the instruction to use NewSRsrc
3406 SRsrc->setReg(NewSRsrc);
3410 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
3411 SmallVector<MachineInstr *, 128> Worklist;
3412 Worklist.push_back(&TopInst);
3414 while (!Worklist.empty()) {
3415 MachineInstr &Inst = *Worklist.pop_back_val();
3416 MachineBasicBlock *MBB = Inst.getParent();
3417 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
3419 unsigned Opcode = Inst.getOpcode();
3420 unsigned NewOpcode = getVALUOp(Inst);
3422 // Handle some special cases
3426 case AMDGPU::S_AND_B64:
3427 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::V_AND_B32_e64);
3428 Inst.eraseFromParent();
3431 case AMDGPU::S_OR_B64:
3432 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::V_OR_B32_e64);
3433 Inst.eraseFromParent();
3436 case AMDGPU::S_XOR_B64:
3437 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::V_XOR_B32_e64);
3438 Inst.eraseFromParent();
3441 case AMDGPU::S_NOT_B64:
3442 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::V_NOT_B32_e32);
3443 Inst.eraseFromParent();
3446 case AMDGPU::S_BCNT1_I32_B64:
3447 splitScalar64BitBCNT(Worklist, Inst);
3448 Inst.eraseFromParent();
3451 case AMDGPU::S_BFE_I64: {
3452 splitScalar64BitBFE(Worklist, Inst);
3453 Inst.eraseFromParent();
3457 case AMDGPU::S_LSHL_B32:
3458 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
3459 NewOpcode = AMDGPU::V_LSHLREV_B32_e64;
3463 case AMDGPU::S_ASHR_I32:
3464 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
3465 NewOpcode = AMDGPU::V_ASHRREV_I32_e64;
3469 case AMDGPU::S_LSHR_B32:
3470 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
3471 NewOpcode = AMDGPU::V_LSHRREV_B32_e64;
3475 case AMDGPU::S_LSHL_B64:
3476 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
3477 NewOpcode = AMDGPU::V_LSHLREV_B64;
3481 case AMDGPU::S_ASHR_I64:
3482 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
3483 NewOpcode = AMDGPU::V_ASHRREV_I64;
3487 case AMDGPU::S_LSHR_B64:
3488 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
3489 NewOpcode = AMDGPU::V_LSHRREV_B64;
3494 case AMDGPU::S_ABS_I32:
3495 lowerScalarAbs(Worklist, Inst);
3496 Inst.eraseFromParent();
3499 case AMDGPU::S_CBRANCH_SCC0:
3500 case AMDGPU::S_CBRANCH_SCC1:
3501 // Clear unused bits of vcc
3502 BuildMI(*MBB, Inst, Inst.getDebugLoc(), get(AMDGPU::S_AND_B64),
3504 .addReg(AMDGPU::EXEC)
3505 .addReg(AMDGPU::VCC);
3508 case AMDGPU::S_BFE_U64:
3509 case AMDGPU::S_BFM_B64:
3510 llvm_unreachable("Moving this op to VALU not implemented");
3512 case AMDGPU::S_PACK_LL_B32_B16:
3513 case AMDGPU::S_PACK_LH_B32_B16:
3514 case AMDGPU::S_PACK_HH_B32_B16: {
3515 movePackToVALU(Worklist, MRI, Inst);
3516 Inst.eraseFromParent();
3521 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
3522 // We cannot move this instruction to the VALU, so we should try to
3523 // legalize its operands instead.
3524 legalizeOperands(Inst);
3528 // Use the new VALU Opcode.
3529 const MCInstrDesc &NewDesc = get(NewOpcode);
3530 Inst.setDesc(NewDesc);
3532 // Remove any references to SCC. Vector instructions can't read from it, and
3533 // We're just about to add the implicit use / defs of VCC, and we don't want
3535 for (unsigned i = Inst.getNumOperands() - 1; i > 0; --i) {
3536 MachineOperand &Op = Inst.getOperand(i);
3537 if (Op.isReg() && Op.getReg() == AMDGPU::SCC) {
3538 Inst.RemoveOperand(i);
3539 addSCCDefUsersToVALUWorklist(Inst, Worklist);
3543 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
3544 // We are converting these to a BFE, so we need to add the missing
3545 // operands for the size and offset.
3546 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
3547 Inst.addOperand(MachineOperand::CreateImm(0));
3548 Inst.addOperand(MachineOperand::CreateImm(Size));
3550 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
3551 // The VALU version adds the second operand to the result, so insert an
3553 Inst.addOperand(MachineOperand::CreateImm(0));
3556 Inst.addImplicitDefUseOperands(*Inst.getParent()->getParent());
3558 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
3559 const MachineOperand &OffsetWidthOp = Inst.getOperand(2);
3560 // If we need to move this to VGPRs, we need to unpack the second operand
3561 // back into the 2 separate ones for bit offset and width.
3562 assert(OffsetWidthOp.isImm() &&
3563 "Scalar BFE is only implemented for constant width and offset");
3564 uint32_t Imm = OffsetWidthOp.getImm();
3566 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
3567 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
3568 Inst.RemoveOperand(2); // Remove old immediate.
3569 Inst.addOperand(MachineOperand::CreateImm(Offset));
3570 Inst.addOperand(MachineOperand::CreateImm(BitWidth));
3573 bool HasDst = Inst.getOperand(0).isReg() && Inst.getOperand(0).isDef();
3574 unsigned NewDstReg = AMDGPU::NoRegister;
3576 unsigned DstReg = Inst.getOperand(0).getReg();
3577 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
3580 // Update the destination register class.
3581 const TargetRegisterClass *NewDstRC = getDestEquivalentVGPRClass(Inst);
3585 if (Inst.isCopy() &&
3586 TargetRegisterInfo::isVirtualRegister(Inst.getOperand(1).getReg()) &&
3587 NewDstRC == RI.getRegClassForReg(MRI, Inst.getOperand(1).getReg())) {
3588 // Instead of creating a copy where src and dst are the same register
3589 // class, we just replace all uses of dst with src. These kinds of
3590 // copies interfere with the heuristics MachineSink uses to decide
3591 // whether or not to split a critical edge. Since the pass assumes
3592 // that copies will end up as machine instructions and not be
3594 addUsersToMoveToVALUWorklist(DstReg, MRI, Worklist);
3595 MRI.replaceRegWith(DstReg, Inst.getOperand(1).getReg());
3596 MRI.clearKillFlags(Inst.getOperand(1).getReg());
3597 Inst.getOperand(0).setReg(DstReg);
3601 NewDstReg = MRI.createVirtualRegister(NewDstRC);
3602 MRI.replaceRegWith(DstReg, NewDstReg);
3605 // Legalize the operands
3606 legalizeOperands(Inst);
3609 addUsersToMoveToVALUWorklist(NewDstReg, MRI, Worklist);
3613 void SIInstrInfo::lowerScalarAbs(SmallVectorImpl<MachineInstr *> &Worklist,
3614 MachineInstr &Inst) const {
3615 MachineBasicBlock &MBB = *Inst.getParent();
3616 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3617 MachineBasicBlock::iterator MII = Inst;
3618 DebugLoc DL = Inst.getDebugLoc();
3620 MachineOperand &Dest = Inst.getOperand(0);
3621 MachineOperand &Src = Inst.getOperand(1);
3622 unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3623 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3625 BuildMI(MBB, MII, DL, get(AMDGPU::V_SUB_I32_e32), TmpReg)
3627 .addReg(Src.getReg());
3629 BuildMI(MBB, MII, DL, get(AMDGPU::V_MAX_I32_e64), ResultReg)
3630 .addReg(Src.getReg())
3633 MRI.replaceRegWith(Dest.getReg(), ResultReg);
3634 addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
3637 void SIInstrInfo::splitScalar64BitUnaryOp(
3638 SmallVectorImpl<MachineInstr *> &Worklist, MachineInstr &Inst,
3639 unsigned Opcode) const {
3640 MachineBasicBlock &MBB = *Inst.getParent();
3641 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3643 MachineOperand &Dest = Inst.getOperand(0);
3644 MachineOperand &Src0 = Inst.getOperand(1);
3645 DebugLoc DL = Inst.getDebugLoc();
3647 MachineBasicBlock::iterator MII = Inst;
3649 const MCInstrDesc &InstDesc = get(Opcode);
3650 const TargetRegisterClass *Src0RC = Src0.isReg() ?
3651 MRI.getRegClass(Src0.getReg()) :
3652 &AMDGPU::SGPR_32RegClass;
3654 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
3656 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
3657 AMDGPU::sub0, Src0SubRC);
3659 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
3660 const TargetRegisterClass *NewDestRC = RI.getEquivalentVGPRClass(DestRC);
3661 const TargetRegisterClass *NewDestSubRC = RI.getSubRegClass(NewDestRC, AMDGPU::sub0);
3663 unsigned DestSub0 = MRI.createVirtualRegister(NewDestSubRC);
3664 BuildMI(MBB, MII, DL, InstDesc, DestSub0).add(SrcReg0Sub0);
3666 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
3667 AMDGPU::sub1, Src0SubRC);
3669 unsigned DestSub1 = MRI.createVirtualRegister(NewDestSubRC);
3670 BuildMI(MBB, MII, DL, InstDesc, DestSub1).add(SrcReg0Sub1);
3672 unsigned FullDestReg = MRI.createVirtualRegister(NewDestRC);
3673 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
3675 .addImm(AMDGPU::sub0)
3677 .addImm(AMDGPU::sub1);
3679 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
3681 // We don't need to legalizeOperands here because for a single operand, src0
3682 // will support any kind of input.
3684 // Move all users of this moved value.
3685 addUsersToMoveToVALUWorklist(FullDestReg, MRI, Worklist);
3688 void SIInstrInfo::splitScalar64BitBinaryOp(
3689 SmallVectorImpl<MachineInstr *> &Worklist, MachineInstr &Inst,
3690 unsigned Opcode) const {
3691 MachineBasicBlock &MBB = *Inst.getParent();
3692 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3694 MachineOperand &Dest = Inst.getOperand(0);
3695 MachineOperand &Src0 = Inst.getOperand(1);
3696 MachineOperand &Src1 = Inst.getOperand(2);
3697 DebugLoc DL = Inst.getDebugLoc();
3699 MachineBasicBlock::iterator MII = Inst;
3701 const MCInstrDesc &InstDesc = get(Opcode);
3702 const TargetRegisterClass *Src0RC = Src0.isReg() ?
3703 MRI.getRegClass(Src0.getReg()) :
3704 &AMDGPU::SGPR_32RegClass;
3706 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
3707 const TargetRegisterClass *Src1RC = Src1.isReg() ?
3708 MRI.getRegClass(Src1.getReg()) :
3709 &AMDGPU::SGPR_32RegClass;
3711 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
3713 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
3714 AMDGPU::sub0, Src0SubRC);
3715 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
3716 AMDGPU::sub0, Src1SubRC);
3718 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
3719 const TargetRegisterClass *NewDestRC = RI.getEquivalentVGPRClass(DestRC);
3720 const TargetRegisterClass *NewDestSubRC = RI.getSubRegClass(NewDestRC, AMDGPU::sub0);
3722 unsigned DestSub0 = MRI.createVirtualRegister(NewDestSubRC);
3723 MachineInstr &LoHalf = *BuildMI(MBB, MII, DL, InstDesc, DestSub0)
3727 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
3728 AMDGPU::sub1, Src0SubRC);
3729 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
3730 AMDGPU::sub1, Src1SubRC);
3732 unsigned DestSub1 = MRI.createVirtualRegister(NewDestSubRC);
3733 MachineInstr &HiHalf = *BuildMI(MBB, MII, DL, InstDesc, DestSub1)
3737 unsigned FullDestReg = MRI.createVirtualRegister(NewDestRC);
3738 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
3740 .addImm(AMDGPU::sub0)
3742 .addImm(AMDGPU::sub1);
3744 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
3746 // Try to legalize the operands in case we need to swap the order to keep it
3748 legalizeOperands(LoHalf);
3749 legalizeOperands(HiHalf);
3751 // Move all users of this moved vlaue.
3752 addUsersToMoveToVALUWorklist(FullDestReg, MRI, Worklist);
3755 void SIInstrInfo::splitScalar64BitBCNT(
3756 SmallVectorImpl<MachineInstr *> &Worklist, MachineInstr &Inst) const {
3757 MachineBasicBlock &MBB = *Inst.getParent();
3758 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3760 MachineBasicBlock::iterator MII = Inst;
3761 DebugLoc DL = Inst.getDebugLoc();
3763 MachineOperand &Dest = Inst.getOperand(0);
3764 MachineOperand &Src = Inst.getOperand(1);
3766 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e64);
3767 const TargetRegisterClass *SrcRC = Src.isReg() ?
3768 MRI.getRegClass(Src.getReg()) :
3769 &AMDGPU::SGPR_32RegClass;
3771 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3772 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3774 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
3776 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
3777 AMDGPU::sub0, SrcSubRC);
3778 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
3779 AMDGPU::sub1, SrcSubRC);
3781 BuildMI(MBB, MII, DL, InstDesc, MidReg).add(SrcRegSub0).addImm(0);
3783 BuildMI(MBB, MII, DL, InstDesc, ResultReg).add(SrcRegSub1).addReg(MidReg);
3785 MRI.replaceRegWith(Dest.getReg(), ResultReg);
3787 // We don't need to legalize operands here. src0 for etiher instruction can be
3788 // an SGPR, and the second input is unused or determined here.
3789 addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
3792 void SIInstrInfo::splitScalar64BitBFE(SmallVectorImpl<MachineInstr *> &Worklist,
3793 MachineInstr &Inst) const {
3794 MachineBasicBlock &MBB = *Inst.getParent();
3795 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3796 MachineBasicBlock::iterator MII = Inst;
3797 DebugLoc DL = Inst.getDebugLoc();
3799 MachineOperand &Dest = Inst.getOperand(0);
3800 uint32_t Imm = Inst.getOperand(2).getImm();
3801 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
3802 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
3806 // Only sext_inreg cases handled.
3807 assert(Inst.getOpcode() == AMDGPU::S_BFE_I64 && BitWidth <= 32 &&
3808 Offset == 0 && "Not implemented");
3810 if (BitWidth < 32) {
3811 unsigned MidRegLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3812 unsigned MidRegHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3813 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
3815 BuildMI(MBB, MII, DL, get(AMDGPU::V_BFE_I32), MidRegLo)
3816 .addReg(Inst.getOperand(1).getReg(), 0, AMDGPU::sub0)
3820 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e32), MidRegHi)
3824 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
3826 .addImm(AMDGPU::sub0)
3828 .addImm(AMDGPU::sub1);
3830 MRI.replaceRegWith(Dest.getReg(), ResultReg);
3831 addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
3835 MachineOperand &Src = Inst.getOperand(1);
3836 unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3837 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
3839 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e64), TmpReg)
3841 .addReg(Src.getReg(), 0, AMDGPU::sub0);
3843 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
3844 .addReg(Src.getReg(), 0, AMDGPU::sub0)
3845 .addImm(AMDGPU::sub0)
3847 .addImm(AMDGPU::sub1);
3849 MRI.replaceRegWith(Dest.getReg(), ResultReg);
3850 addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
3853 void SIInstrInfo::addUsersToMoveToVALUWorklist(
3855 MachineRegisterInfo &MRI,
3856 SmallVectorImpl<MachineInstr *> &Worklist) const {
3857 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(DstReg),
3858 E = MRI.use_end(); I != E;) {
3859 MachineInstr &UseMI = *I->getParent();
3860 if (!canReadVGPR(UseMI, I.getOperandNo())) {
3861 Worklist.push_back(&UseMI);
3865 } while (I != E && I->getParent() == &UseMI);
3872 void SIInstrInfo::movePackToVALU(SmallVectorImpl<MachineInstr *> &Worklist,
3873 MachineRegisterInfo &MRI,
3874 MachineInstr &Inst) const {
3875 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3876 MachineBasicBlock *MBB = Inst.getParent();
3877 MachineOperand &Src0 = Inst.getOperand(1);
3878 MachineOperand &Src1 = Inst.getOperand(2);
3879 const DebugLoc &DL = Inst.getDebugLoc();
3881 switch (Inst.getOpcode()) {
3882 case AMDGPU::S_PACK_LL_B32_B16: {
3883 unsigned ImmReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3884 unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3886 // FIXME: Can do a lot better if we know the high bits of src0 or src1 are
3888 BuildMI(*MBB, Inst, DL, get(AMDGPU::V_MOV_B32_e32), ImmReg)
3891 BuildMI(*MBB, Inst, DL, get(AMDGPU::V_AND_B32_e64), TmpReg)
3892 .addReg(ImmReg, RegState::Kill)
3895 BuildMI(*MBB, Inst, DL, get(AMDGPU::V_LSHL_OR_B32), ResultReg)
3898 .addReg(TmpReg, RegState::Kill);
3901 case AMDGPU::S_PACK_LH_B32_B16: {
3902 unsigned ImmReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3903 BuildMI(*MBB, Inst, DL, get(AMDGPU::V_MOV_B32_e32), ImmReg)
3905 BuildMI(*MBB, Inst, DL, get(AMDGPU::V_BFI_B32), ResultReg)
3906 .addReg(ImmReg, RegState::Kill)
3911 case AMDGPU::S_PACK_HH_B32_B16: {
3912 unsigned ImmReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3913 unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
3914 BuildMI(*MBB, Inst, DL, get(AMDGPU::V_LSHRREV_B32_e64), TmpReg)
3917 BuildMI(*MBB, Inst, DL, get(AMDGPU::V_MOV_B32_e32), ImmReg)
3918 .addImm(0xffff0000);
3919 BuildMI(*MBB, Inst, DL, get(AMDGPU::V_AND_OR_B32), ResultReg)
3921 .addReg(ImmReg, RegState::Kill)
3922 .addReg(TmpReg, RegState::Kill);
3926 llvm_unreachable("unhandled s_pack_* instruction");
3929 MachineOperand &Dest = Inst.getOperand(0);
3930 MRI.replaceRegWith(Dest.getReg(), ResultReg);
3931 addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
3934 void SIInstrInfo::addSCCDefUsersToVALUWorklist(
3935 MachineInstr &SCCDefInst, SmallVectorImpl<MachineInstr *> &Worklist) const {
3936 // This assumes that all the users of SCC are in the same block
3938 for (MachineInstr &MI :
3939 llvm::make_range(MachineBasicBlock::iterator(SCCDefInst),
3940 SCCDefInst.getParent()->end())) {
3941 // Exit if we find another SCC def.
3942 if (MI.findRegisterDefOperandIdx(AMDGPU::SCC) != -1)
3945 if (MI.findRegisterUseOperandIdx(AMDGPU::SCC) != -1)
3946 Worklist.push_back(&MI);
3950 const TargetRegisterClass *SIInstrInfo::getDestEquivalentVGPRClass(
3951 const MachineInstr &Inst) const {
3952 const TargetRegisterClass *NewDstRC = getOpRegClass(Inst, 0);
3954 switch (Inst.getOpcode()) {
3955 // For target instructions, getOpRegClass just returns the virtual register
3956 // class associated with the operand, so we need to find an equivalent VGPR
3957 // register class in order to move the instruction to the VALU.
3960 case AMDGPU::REG_SEQUENCE:
3961 case AMDGPU::INSERT_SUBREG:
3962 if (RI.hasVGPRs(NewDstRC))
3965 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
3974 // Find the one SGPR operand we are allowed to use.
3975 unsigned SIInstrInfo::findUsedSGPR(const MachineInstr &MI,
3976 int OpIndices[3]) const {
3977 const MCInstrDesc &Desc = MI.getDesc();
3979 // Find the one SGPR operand we are allowed to use.
3981 // First we need to consider the instruction's operand requirements before
3982 // legalizing. Some operands are required to be SGPRs, such as implicit uses
3983 // of VCC, but we are still bound by the constant bus requirement to only use
3986 // If the operand's class is an SGPR, we can never move it.
3988 unsigned SGPRReg = findImplicitSGPRRead(MI);
3989 if (SGPRReg != AMDGPU::NoRegister)
3992 unsigned UsedSGPRs[3] = { AMDGPU::NoRegister };
3993 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
3995 for (unsigned i = 0; i < 3; ++i) {
3996 int Idx = OpIndices[i];
4000 const MachineOperand &MO = MI.getOperand(Idx);
4004 // Is this operand statically required to be an SGPR based on the operand
4006 const TargetRegisterClass *OpRC = RI.getRegClass(Desc.OpInfo[Idx].RegClass);
4007 bool IsRequiredSGPR = RI.isSGPRClass(OpRC);
4011 // If this could be a VGPR or an SGPR, Check the dynamic register class.
4012 unsigned Reg = MO.getReg();
4013 const TargetRegisterClass *RegRC = MRI.getRegClass(Reg);
4014 if (RI.isSGPRClass(RegRC))
4018 // We don't have a required SGPR operand, so we have a bit more freedom in
4019 // selecting operands to move.
4021 // Try to select the most used SGPR. If an SGPR is equal to one of the
4022 // others, we choose that.
4025 // V_FMA_F32 v0, s0, s0, s0 -> No moves
4026 // V_FMA_F32 v0, s0, s1, s0 -> Move s1
4028 // TODO: If some of the operands are 64-bit SGPRs and some 32, we should
4031 if (UsedSGPRs[0] != AMDGPU::NoRegister) {
4032 if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
4033 SGPRReg = UsedSGPRs[0];
4036 if (SGPRReg == AMDGPU::NoRegister && UsedSGPRs[1] != AMDGPU::NoRegister) {
4037 if (UsedSGPRs[1] == UsedSGPRs[2])
4038 SGPRReg = UsedSGPRs[1];
4044 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
4045 unsigned OperandName) const {
4046 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
4050 return &MI.getOperand(Idx);
4053 uint64_t SIInstrInfo::getDefaultRsrcDataFormat() const {
4054 uint64_t RsrcDataFormat = AMDGPU::RSRC_DATA_FORMAT;
4055 if (ST.isAmdHsaOS()) {
4056 // Set ATC = 1. GFX9 doesn't have this bit.
4057 if (ST.getGeneration() <= SISubtarget::VOLCANIC_ISLANDS)
4058 RsrcDataFormat |= (1ULL << 56);
4060 // Set MTYPE = 2 (MTYPE_UC = uncached). GFX9 doesn't have this.
4061 // BTW, it disables TC L2 and therefore decreases performance.
4062 if (ST.getGeneration() == SISubtarget::VOLCANIC_ISLANDS)
4063 RsrcDataFormat |= (2ULL << 59);
4066 return RsrcDataFormat;
4069 uint64_t SIInstrInfo::getScratchRsrcWords23() const {
4070 uint64_t Rsrc23 = getDefaultRsrcDataFormat() |
4071 AMDGPU::RSRC_TID_ENABLE |
4072 0xffffffff; // Size;
4074 // GFX9 doesn't have ELEMENT_SIZE.
4075 if (ST.getGeneration() <= SISubtarget::VOLCANIC_ISLANDS) {
4076 uint64_t EltSizeValue = Log2_32(ST.getMaxPrivateElementSize()) - 1;
4077 Rsrc23 |= EltSizeValue << AMDGPU::RSRC_ELEMENT_SIZE_SHIFT;
4080 // IndexStride = 64.
4081 Rsrc23 |= UINT64_C(3) << AMDGPU::RSRC_INDEX_STRIDE_SHIFT;
4083 // If TID_ENABLE is set, DATA_FORMAT specifies stride bits [14:17].
4084 // Clear them unless we want a huge stride.
4085 if (ST.getGeneration() >= SISubtarget::VOLCANIC_ISLANDS)
4086 Rsrc23 &= ~AMDGPU::RSRC_DATA_FORMAT;
4091 bool SIInstrInfo::isLowLatencyInstruction(const MachineInstr &MI) const {
4092 unsigned Opc = MI.getOpcode();
4097 bool SIInstrInfo::isHighLatencyInstruction(const MachineInstr &MI) const {
4098 unsigned Opc = MI.getOpcode();
4100 return isMUBUF(Opc) || isMTBUF(Opc) || isMIMG(Opc);
4103 unsigned SIInstrInfo::isStackAccess(const MachineInstr &MI,
4104 int &FrameIndex) const {
4105 const MachineOperand *Addr = getNamedOperand(MI, AMDGPU::OpName::vaddr);
4106 if (!Addr || !Addr->isFI())
4107 return AMDGPU::NoRegister;
4109 assert(!MI.memoperands_empty() &&
4110 (*MI.memoperands_begin())->getAddrSpace() == AMDGPUASI.PRIVATE_ADDRESS);
4112 FrameIndex = Addr->getIndex();
4113 return getNamedOperand(MI, AMDGPU::OpName::vdata)->getReg();
4116 unsigned SIInstrInfo::isSGPRStackAccess(const MachineInstr &MI,
4117 int &FrameIndex) const {
4118 const MachineOperand *Addr = getNamedOperand(MI, AMDGPU::OpName::addr);
4119 assert(Addr && Addr->isFI());
4120 FrameIndex = Addr->getIndex();
4121 return getNamedOperand(MI, AMDGPU::OpName::data)->getReg();
4124 unsigned SIInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
4125 int &FrameIndex) const {
4128 return AMDGPU::NoRegister;
4130 if (isMUBUF(MI) || isVGPRSpill(MI))
4131 return isStackAccess(MI, FrameIndex);
4133 if (isSGPRSpill(MI))
4134 return isSGPRStackAccess(MI, FrameIndex);
4136 return AMDGPU::NoRegister;
4139 unsigned SIInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
4140 int &FrameIndex) const {
4142 return AMDGPU::NoRegister;
4144 if (isMUBUF(MI) || isVGPRSpill(MI))
4145 return isStackAccess(MI, FrameIndex);
4147 if (isSGPRSpill(MI))
4148 return isSGPRStackAccess(MI, FrameIndex);
4150 return AMDGPU::NoRegister;
4153 unsigned SIInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
4154 unsigned Opc = MI.getOpcode();
4155 const MCInstrDesc &Desc = getMCOpcodeFromPseudo(Opc);
4156 unsigned DescSize = Desc.getSize();
4158 // If we have a definitive size, we can use it. Otherwise we need to inspect
4159 // the operands to know the size.
4161 // FIXME: Instructions that have a base 32-bit encoding report their size as
4162 // 4, even though they are really 8 bytes if they have a literal operand.
4163 if (DescSize != 0 && DescSize != 4)
4166 // 4-byte instructions may have a 32-bit literal encoded after them. Check
4167 // operands that coud ever be literals.
4168 if (isVALU(MI) || isSALU(MI)) {
4169 if (isFixedSize(MI))
4172 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
4174 return 4; // No operands.
4176 if (isLiteralConstantLike(MI.getOperand(Src0Idx), Desc.OpInfo[Src0Idx]))
4179 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
4183 if (isLiteralConstantLike(MI.getOperand(Src1Idx), Desc.OpInfo[Src1Idx]))
4193 case TargetOpcode::IMPLICIT_DEF:
4194 case TargetOpcode::KILL:
4195 case TargetOpcode::DBG_VALUE:
4196 case TargetOpcode::BUNDLE:
4197 case TargetOpcode::EH_LABEL:
4199 case TargetOpcode::INLINEASM: {
4200 const MachineFunction *MF = MI.getParent()->getParent();
4201 const char *AsmStr = MI.getOperand(0).getSymbolName();
4202 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
4205 llvm_unreachable("unable to find instruction size");
4209 bool SIInstrInfo::mayAccessFlatAddressSpace(const MachineInstr &MI) const {
4213 if (MI.memoperands_empty())
4216 for (const MachineMemOperand *MMO : MI.memoperands()) {
4217 if (MMO->getAddrSpace() == AMDGPUASI.FLAT_ADDRESS)
4223 bool SIInstrInfo::isNonUniformBranchInstr(MachineInstr &Branch) const {
4224 return Branch.getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO;
4227 void SIInstrInfo::convertNonUniformIfRegion(MachineBasicBlock *IfEntry,
4228 MachineBasicBlock *IfEnd) const {
4229 MachineBasicBlock::iterator TI = IfEntry->getFirstTerminator();
4230 assert(TI != IfEntry->end());
4232 MachineInstr *Branch = &(*TI);
4233 MachineFunction *MF = IfEntry->getParent();
4234 MachineRegisterInfo &MRI = IfEntry->getParent()->getRegInfo();
4236 if (Branch->getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO) {
4237 unsigned DstReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
4238 MachineInstr *SIIF =
4239 BuildMI(*MF, Branch->getDebugLoc(), get(AMDGPU::SI_IF), DstReg)
4240 .add(Branch->getOperand(0))
4241 .add(Branch->getOperand(1));
4242 MachineInstr *SIEND =
4243 BuildMI(*MF, Branch->getDebugLoc(), get(AMDGPU::SI_END_CF))
4247 IfEntry->insert(IfEntry->end(), SIIF);
4248 IfEnd->insert(IfEnd->getFirstNonPHI(), SIEND);
4252 void SIInstrInfo::convertNonUniformLoopRegion(
4253 MachineBasicBlock *LoopEntry, MachineBasicBlock *LoopEnd) const {
4254 MachineBasicBlock::iterator TI = LoopEnd->getFirstTerminator();
4255 // We expect 2 terminators, one conditional and one unconditional.
4256 assert(TI != LoopEnd->end());
4258 MachineInstr *Branch = &(*TI);
4259 MachineFunction *MF = LoopEnd->getParent();
4260 MachineRegisterInfo &MRI = LoopEnd->getParent()->getRegInfo();
4262 if (Branch->getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO) {
4264 unsigned DstReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
4265 unsigned BackEdgeReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
4266 MachineInstrBuilder HeaderPHIBuilder =
4267 BuildMI(*(MF), Branch->getDebugLoc(), get(TargetOpcode::PHI), DstReg);
4268 for (MachineBasicBlock::pred_iterator PI = LoopEntry->pred_begin(),
4269 E = LoopEntry->pred_end();
4271 if (*PI == LoopEnd) {
4272 HeaderPHIBuilder.addReg(BackEdgeReg);
4274 MachineBasicBlock *PMBB = *PI;
4275 unsigned ZeroReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
4276 materializeImmediate(*PMBB, PMBB->getFirstTerminator(), DebugLoc(),
4278 HeaderPHIBuilder.addReg(ZeroReg);
4280 HeaderPHIBuilder.addMBB(*PI);
4282 MachineInstr *HeaderPhi = HeaderPHIBuilder;
4283 MachineInstr *SIIFBREAK = BuildMI(*(MF), Branch->getDebugLoc(),
4284 get(AMDGPU::SI_IF_BREAK), BackEdgeReg)
4286 .add(Branch->getOperand(0));
4287 MachineInstr *SILOOP =
4288 BuildMI(*(MF), Branch->getDebugLoc(), get(AMDGPU::SI_LOOP))
4289 .addReg(BackEdgeReg)
4292 LoopEntry->insert(LoopEntry->begin(), HeaderPhi);
4294 LoopEnd->insert(LoopEnd->end(), SIIFBREAK);
4295 LoopEnd->insert(LoopEnd->end(), SILOOP);
4299 ArrayRef<std::pair<int, const char *>>
4300 SIInstrInfo::getSerializableTargetIndices() const {
4301 static const std::pair<int, const char *> TargetIndices[] = {
4302 {AMDGPU::TI_CONSTDATA_START, "amdgpu-constdata-start"},
4303 {AMDGPU::TI_SCRATCH_RSRC_DWORD0, "amdgpu-scratch-rsrc-dword0"},
4304 {AMDGPU::TI_SCRATCH_RSRC_DWORD1, "amdgpu-scratch-rsrc-dword1"},
4305 {AMDGPU::TI_SCRATCH_RSRC_DWORD2, "amdgpu-scratch-rsrc-dword2"},
4306 {AMDGPU::TI_SCRATCH_RSRC_DWORD3, "amdgpu-scratch-rsrc-dword3"}};
4307 return makeArrayRef(TargetIndices);
4310 /// This is used by the post-RA scheduler (SchedulePostRAList.cpp). The
4311 /// post-RA version of misched uses CreateTargetMIHazardRecognizer.
4312 ScheduleHazardRecognizer *
4313 SIInstrInfo::CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
4314 const ScheduleDAG *DAG) const {
4315 return new GCNHazardRecognizer(DAG->MF);
4318 /// This is the hazard recognizer used at -O0 by the PostRAHazardRecognizer
4320 ScheduleHazardRecognizer *
4321 SIInstrInfo::CreateTargetPostRAHazardRecognizer(const MachineFunction &MF) const {
4322 return new GCNHazardRecognizer(MF);
4325 std::pair<unsigned, unsigned>
4326 SIInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
4327 return std::make_pair(TF & MO_MASK, TF & ~MO_MASK);
4330 ArrayRef<std::pair<unsigned, const char *>>
4331 SIInstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
4332 static const std::pair<unsigned, const char *> TargetFlags[] = {
4333 { MO_GOTPCREL, "amdgpu-gotprel" },
4334 { MO_GOTPCREL32_LO, "amdgpu-gotprel32-lo" },
4335 { MO_GOTPCREL32_HI, "amdgpu-gotprel32-hi" },
4336 { MO_REL32_LO, "amdgpu-rel32-lo" },
4337 { MO_REL32_HI, "amdgpu-rel32-hi" }
4340 return makeArrayRef(TargetFlags);
4343 bool SIInstrInfo::isBasicBlockPrologue(const MachineInstr &MI) const {
4344 return !MI.isTerminator() && MI.getOpcode() != AMDGPU::COPY &&
4345 MI.modifiesRegister(AMDGPU::EXEC, &RI);
4349 SIInstrInfo::getAddNoCarry(MachineBasicBlock &MBB,
4350 MachineBasicBlock::iterator I,
4352 unsigned DestReg) const {
4353 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
4355 unsigned UnusedCarry = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
4357 return BuildMI(MBB, I, DL, get(AMDGPU::V_ADD_I32_e64), DestReg)
4358 .addReg(UnusedCarry, RegState::Define | RegState::Dead);