1 //===-- R600InstrInfo.cpp - R600 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 R600 Implementation of TargetInstrInfo.
13 //===----------------------------------------------------------------------===//
15 #include "R600InstrInfo.h"
17 #include "AMDGPUSubtarget.h"
18 #include "AMDGPUTargetMachine.h"
19 #include "R600Defines.h"
20 #include "R600MachineFunctionInfo.h"
21 #include "R600RegisterInfo.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #define GET_INSTRINFO_CTOR
27 #include "AMDGPUGenDFAPacketizer.inc"
31 R600InstrInfo::R600InstrInfo(AMDGPUTargetMachine &tm)
32 : AMDGPUInstrInfo(tm),
34 ST(tm.getSubtarget<AMDGPUSubtarget>())
37 const R600RegisterInfo &R600InstrInfo::getRegisterInfo() const {
41 bool R600InstrInfo::isTrig(const MachineInstr &MI) const {
42 return get(MI.getOpcode()).TSFlags & R600_InstFlag::TRIG;
45 bool R600InstrInfo::isVector(const MachineInstr &MI) const {
46 return get(MI.getOpcode()).TSFlags & R600_InstFlag::VECTOR;
50 R600InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
51 MachineBasicBlock::iterator MI, DebugLoc DL,
52 unsigned DestReg, unsigned SrcReg,
54 unsigned VectorComponents = 0;
55 if (AMDGPU::R600_Reg128RegClass.contains(DestReg) &&
56 AMDGPU::R600_Reg128RegClass.contains(SrcReg)) {
58 } else if(AMDGPU::R600_Reg64RegClass.contains(DestReg) &&
59 AMDGPU::R600_Reg64RegClass.contains(SrcReg)) {
63 if (VectorComponents > 0) {
64 for (unsigned I = 0; I < VectorComponents; I++) {
65 unsigned SubRegIndex = RI.getSubRegFromChannel(I);
66 buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
67 RI.getSubReg(DestReg, SubRegIndex),
68 RI.getSubReg(SrcReg, SubRegIndex))
70 RegState::Define | RegState::Implicit);
73 MachineInstr *NewMI = buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
75 NewMI->getOperand(getOperandIdx(*NewMI, AMDGPU::OpName::src0))
80 MachineInstr * R600InstrInfo::getMovImmInstr(MachineFunction *MF,
81 unsigned DstReg, int64_t Imm) const {
82 MachineInstr * MI = MF->CreateMachineInstr(get(AMDGPU::MOV), DebugLoc());
83 MachineInstrBuilder MIB(*MF, MI);
84 MIB.addReg(DstReg, RegState::Define);
85 MIB.addReg(AMDGPU::ALU_LITERAL_X);
87 MIB.addReg(0); // PREDICATE_BIT
92 unsigned R600InstrInfo::getIEQOpcode() const {
93 return AMDGPU::SETE_INT;
96 bool R600InstrInfo::isMov(unsigned Opcode) const {
100 default: return false;
102 case AMDGPU::MOV_IMM_F32:
103 case AMDGPU::MOV_IMM_I32:
108 // Some instructions act as place holders to emulate operations that the GPU
109 // hardware does automatically. This function can be used to check if
110 // an opcode falls into this category.
111 bool R600InstrInfo::isPlaceHolderOpcode(unsigned Opcode) const {
113 default: return false;
119 bool R600InstrInfo::isReductionOp(unsigned Opcode) const {
123 bool R600InstrInfo::isCubeOp(unsigned Opcode) const {
125 default: return false;
126 case AMDGPU::CUBE_r600_pseudo:
127 case AMDGPU::CUBE_r600_real:
128 case AMDGPU::CUBE_eg_pseudo:
129 case AMDGPU::CUBE_eg_real:
134 bool R600InstrInfo::isALUInstr(unsigned Opcode) const {
135 unsigned TargetFlags = get(Opcode).TSFlags;
137 return (TargetFlags & R600_InstFlag::ALU_INST);
140 bool R600InstrInfo::hasInstrModifiers(unsigned Opcode) const {
141 unsigned TargetFlags = get(Opcode).TSFlags;
143 return ((TargetFlags & R600_InstFlag::OP1) |
144 (TargetFlags & R600_InstFlag::OP2) |
145 (TargetFlags & R600_InstFlag::OP3));
148 bool R600InstrInfo::isLDSInstr(unsigned Opcode) const {
149 unsigned TargetFlags = get(Opcode).TSFlags;
151 return ((TargetFlags & R600_InstFlag::LDS_1A) |
152 (TargetFlags & R600_InstFlag::LDS_1A1D) |
153 (TargetFlags & R600_InstFlag::LDS_1A2D));
156 bool R600InstrInfo::canBeConsideredALU(const MachineInstr *MI) const {
157 if (isALUInstr(MI->getOpcode()))
159 if (isVector(*MI) || isCubeOp(MI->getOpcode()))
161 switch (MI->getOpcode()) {
163 case AMDGPU::INTERP_PAIR_XY:
164 case AMDGPU::INTERP_PAIR_ZW:
165 case AMDGPU::INTERP_VEC_LOAD:
174 bool R600InstrInfo::isTransOnly(unsigned Opcode) const {
175 if (ST.hasCaymanISA())
177 return (get(Opcode).getSchedClass() == AMDGPU::Sched::TransALU);
180 bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const {
181 return isTransOnly(MI->getOpcode());
184 bool R600InstrInfo::isVectorOnly(unsigned Opcode) const {
185 return (get(Opcode).getSchedClass() == AMDGPU::Sched::VecALU);
188 bool R600InstrInfo::isVectorOnly(const MachineInstr *MI) const {
189 return isVectorOnly(MI->getOpcode());
192 bool R600InstrInfo::isExport(unsigned Opcode) const {
193 return (get(Opcode).TSFlags & R600_InstFlag::IS_EXPORT);
196 bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
197 return ST.hasVertexCache() && IS_VTX(get(Opcode));
200 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const {
201 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
202 return MFI->ShaderType != ShaderType::COMPUTE && usesVertexCache(MI->getOpcode());
205 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
206 return (!ST.hasVertexCache() && IS_VTX(get(Opcode))) || IS_TEX(get(Opcode));
209 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
210 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
211 return (MFI->ShaderType == ShaderType::COMPUTE && usesVertexCache(MI->getOpcode())) ||
212 usesTextureCache(MI->getOpcode());
215 bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
218 case AMDGPU::GROUP_BARRIER:
225 bool R600InstrInfo::readsLDSSrcReg(const MachineInstr *MI) const {
226 if (!isALUInstr(MI->getOpcode())) {
229 for (MachineInstr::const_mop_iterator I = MI->operands_begin(),
230 E = MI->operands_end(); I != E; ++I) {
231 if (!I->isReg() || !I->isUse() ||
232 TargetRegisterInfo::isVirtualRegister(I->getReg()))
235 if (AMDGPU::R600_LDS_SRC_REGRegClass.contains(I->getReg()))
241 int R600InstrInfo::getSrcIdx(unsigned Opcode, unsigned SrcNum) const {
242 static const unsigned OpTable[] = {
243 AMDGPU::OpName::src0,
244 AMDGPU::OpName::src1,
249 return getOperandIdx(Opcode, OpTable[SrcNum]);
252 #define SRC_SEL_ROWS 11
253 int R600InstrInfo::getSelIdx(unsigned Opcode, unsigned SrcIdx) const {
254 static const unsigned SrcSelTable[SRC_SEL_ROWS][2] = {
255 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
256 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
257 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
258 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
259 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
260 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
261 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
262 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
263 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
264 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
265 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W}
268 for (unsigned i = 0; i < SRC_SEL_ROWS; ++i) {
269 if (getOperandIdx(Opcode, SrcSelTable[i][0]) == (int)SrcIdx) {
270 return getOperandIdx(Opcode, SrcSelTable[i][1]);
277 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
278 R600InstrInfo::getSrcs(MachineInstr *MI) const {
279 SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
281 if (MI->getOpcode() == AMDGPU::DOT_4) {
282 static const unsigned OpTable[8][2] = {
283 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
284 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
285 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
286 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
287 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
288 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
289 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
290 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W},
293 for (unsigned j = 0; j < 8; j++) {
294 MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
296 unsigned Reg = MO.getReg();
297 if (Reg == AMDGPU::ALU_CONST) {
298 unsigned Sel = MI->getOperand(getOperandIdx(MI->getOpcode(),
299 OpTable[j][1])).getImm();
300 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
308 static const unsigned OpTable[3][2] = {
309 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
310 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
311 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
314 for (unsigned j = 0; j < 3; j++) {
315 int SrcIdx = getOperandIdx(MI->getOpcode(), OpTable[j][0]);
318 MachineOperand &MO = MI->getOperand(SrcIdx);
319 unsigned Reg = MI->getOperand(SrcIdx).getReg();
320 if (Reg == AMDGPU::ALU_CONST) {
321 unsigned Sel = MI->getOperand(
322 getOperandIdx(MI->getOpcode(), OpTable[j][1])).getImm();
323 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
326 if (Reg == AMDGPU::ALU_LITERAL_X) {
327 unsigned Imm = MI->getOperand(
328 getOperandIdx(MI->getOpcode(), AMDGPU::OpName::literal)).getImm();
329 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Imm));
332 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, 0));
337 std::vector<std::pair<int, unsigned> >
338 R600InstrInfo::ExtractSrcs(MachineInstr *MI,
339 const DenseMap<unsigned, unsigned> &PV,
340 unsigned &ConstCount) const {
342 const SmallVector<std::pair<MachineOperand *, int64_t>, 3> Srcs = getSrcs(MI);
343 const std::pair<int, unsigned> DummyPair(-1, 0);
344 std::vector<std::pair<int, unsigned> > Result;
346 for (unsigned n = Srcs.size(); i < n; ++i) {
347 unsigned Reg = Srcs[i].first->getReg();
348 unsigned Index = RI.getEncodingValue(Reg) & 0xff;
349 if (Reg == AMDGPU::OQAP) {
350 Result.push_back(std::pair<int, unsigned>(Index, 0));
352 if (PV.find(Reg) != PV.end()) {
353 // 255 is used to tells its a PS/PV reg
354 Result.push_back(std::pair<int, unsigned>(255, 0));
359 Result.push_back(DummyPair);
362 unsigned Chan = RI.getHWRegChan(Reg);
363 Result.push_back(std::pair<int, unsigned>(Index, Chan));
366 Result.push_back(DummyPair);
370 static std::vector<std::pair<int, unsigned> >
371 Swizzle(std::vector<std::pair<int, unsigned> > Src,
372 R600InstrInfo::BankSwizzle Swz) {
373 if (Src[0] == Src[1])
376 case R600InstrInfo::ALU_VEC_012_SCL_210:
378 case R600InstrInfo::ALU_VEC_021_SCL_122:
379 std::swap(Src[1], Src[2]);
381 case R600InstrInfo::ALU_VEC_102_SCL_221:
382 std::swap(Src[0], Src[1]);
384 case R600InstrInfo::ALU_VEC_120_SCL_212:
385 std::swap(Src[0], Src[1]);
386 std::swap(Src[0], Src[2]);
388 case R600InstrInfo::ALU_VEC_201:
389 std::swap(Src[0], Src[2]);
390 std::swap(Src[0], Src[1]);
392 case R600InstrInfo::ALU_VEC_210:
393 std::swap(Src[0], Src[2]);
400 getTransSwizzle(R600InstrInfo::BankSwizzle Swz, unsigned Op) {
402 case R600InstrInfo::ALU_VEC_012_SCL_210: {
403 unsigned Cycles[3] = { 2, 1, 0};
406 case R600InstrInfo::ALU_VEC_021_SCL_122: {
407 unsigned Cycles[3] = { 1, 2, 2};
410 case R600InstrInfo::ALU_VEC_120_SCL_212: {
411 unsigned Cycles[3] = { 2, 1, 2};
414 case R600InstrInfo::ALU_VEC_102_SCL_221: {
415 unsigned Cycles[3] = { 2, 2, 1};
419 llvm_unreachable("Wrong Swizzle for Trans Slot");
424 /// returns how many MIs (whose inputs are represented by IGSrcs) can be packed
425 /// in the same Instruction Group while meeting read port limitations given a
426 /// Swz swizzle sequence.
427 unsigned R600InstrInfo::isLegalUpTo(
428 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
429 const std::vector<R600InstrInfo::BankSwizzle> &Swz,
430 const std::vector<std::pair<int, unsigned> > &TransSrcs,
431 R600InstrInfo::BankSwizzle TransSwz) const {
433 memset(Vector, -1, sizeof(Vector));
434 for (unsigned i = 0, e = IGSrcs.size(); i < e; i++) {
435 const std::vector<std::pair<int, unsigned> > &Srcs =
436 Swizzle(IGSrcs[i], Swz[i]);
437 for (unsigned j = 0; j < 3; j++) {
438 const std::pair<int, unsigned> &Src = Srcs[j];
439 if (Src.first < 0 || Src.first == 255)
441 if (Src.first == GET_REG_INDEX(RI.getEncodingValue(AMDGPU::OQAP))) {
442 if (Swz[i] != R600InstrInfo::ALU_VEC_012_SCL_210 &&
443 Swz[i] != R600InstrInfo::ALU_VEC_021_SCL_122) {
444 // The value from output queue A (denoted by register OQAP) can
445 // only be fetched during the first cycle.
448 // OQAP does not count towards the normal read port restrictions
451 if (Vector[Src.second][j] < 0)
452 Vector[Src.second][j] = Src.first;
453 if (Vector[Src.second][j] != Src.first)
457 // Now check Trans Alu
458 for (unsigned i = 0, e = TransSrcs.size(); i < e; ++i) {
459 const std::pair<int, unsigned> &Src = TransSrcs[i];
460 unsigned Cycle = getTransSwizzle(TransSwz, i);
463 if (Src.first == 255)
465 if (Vector[Src.second][Cycle] < 0)
466 Vector[Src.second][Cycle] = Src.first;
467 if (Vector[Src.second][Cycle] != Src.first)
468 return IGSrcs.size() - 1;
470 return IGSrcs.size();
473 /// Given a swizzle sequence SwzCandidate and an index Idx, returns the next
474 /// (in lexicographic term) swizzle sequence assuming that all swizzles after
475 /// Idx can be skipped
477 NextPossibleSolution(
478 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
480 assert(Idx < SwzCandidate.size());
482 while (ResetIdx > -1 && SwzCandidate[ResetIdx] == R600InstrInfo::ALU_VEC_210)
484 for (unsigned i = ResetIdx + 1, e = SwzCandidate.size(); i < e; i++) {
485 SwzCandidate[i] = R600InstrInfo::ALU_VEC_012_SCL_210;
489 int NextSwizzle = SwzCandidate[ResetIdx] + 1;
490 SwzCandidate[ResetIdx] = (R600InstrInfo::BankSwizzle)NextSwizzle;
494 /// Enumerate all possible Swizzle sequence to find one that can meet all
495 /// read port requirements.
496 bool R600InstrInfo::FindSwizzleForVectorSlot(
497 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
498 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
499 const std::vector<std::pair<int, unsigned> > &TransSrcs,
500 R600InstrInfo::BankSwizzle TransSwz) const {
501 unsigned ValidUpTo = 0;
503 ValidUpTo = isLegalUpTo(IGSrcs, SwzCandidate, TransSrcs, TransSwz);
504 if (ValidUpTo == IGSrcs.size())
506 } while (NextPossibleSolution(SwzCandidate, ValidUpTo));
510 /// Instructions in Trans slot can't read gpr at cycle 0 if they also read
511 /// a const, and can't read a gpr at cycle 1 if they read 2 const.
513 isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,
514 const std::vector<std::pair<int, unsigned> > &TransOps,
515 unsigned ConstCount) {
516 // TransALU can't read 3 constants
519 for (unsigned i = 0, e = TransOps.size(); i < e; ++i) {
520 const std::pair<int, unsigned> &Src = TransOps[i];
521 unsigned Cycle = getTransSwizzle(TransSwz, i);
524 if (ConstCount > 0 && Cycle == 0)
526 if (ConstCount > 1 && Cycle == 1)
533 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG,
534 const DenseMap<unsigned, unsigned> &PV,
535 std::vector<BankSwizzle> &ValidSwizzle,
538 //Todo : support shared src0 - src1 operand
540 std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs;
541 ValidSwizzle.clear();
543 BankSwizzle TransBS = ALU_VEC_012_SCL_210;
544 for (unsigned i = 0, e = IG.size(); i < e; ++i) {
545 IGSrcs.push_back(ExtractSrcs(IG[i], PV, ConstCount));
546 unsigned Op = getOperandIdx(IG[i]->getOpcode(),
547 AMDGPU::OpName::bank_swizzle);
548 ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle)
549 IG[i]->getOperand(Op).getImm());
551 std::vector<std::pair<int, unsigned> > TransOps;
553 return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS);
555 TransOps = IGSrcs.back();
557 ValidSwizzle.pop_back();
559 static const R600InstrInfo::BankSwizzle TransSwz[] = {
565 for (unsigned i = 0; i < 4; i++) {
566 TransBS = TransSwz[i];
567 if (!isConstCompatible(TransBS, TransOps, ConstCount))
569 bool Result = FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps,
572 ValidSwizzle.push_back(TransBS);
582 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts)
584 assert (Consts.size() <= 12 && "Too many operands in instructions group");
585 unsigned Pair1 = 0, Pair2 = 0;
586 for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
587 unsigned ReadConstHalf = Consts[i] & 2;
588 unsigned ReadConstIndex = Consts[i] & (~3);
589 unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf;
591 Pair1 = ReadHalfConst;
594 if (Pair1 == ReadHalfConst)
597 Pair2 = ReadHalfConst;
600 if (Pair2 != ReadHalfConst)
607 R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
609 std::vector<unsigned> Consts;
610 SmallSet<int64_t, 4> Literals;
611 for (unsigned i = 0, n = MIs.size(); i < n; i++) {
612 MachineInstr *MI = MIs[i];
613 if (!isALUInstr(MI->getOpcode()))
616 const SmallVectorImpl<std::pair<MachineOperand *, int64_t> > &Srcs =
619 for (unsigned j = 0, e = Srcs.size(); j < e; j++) {
620 std::pair<MachineOperand *, unsigned> Src = Srcs[j];
621 if (Src.first->getReg() == AMDGPU::ALU_LITERAL_X)
622 Literals.insert(Src.second);
623 if (Literals.size() > 4)
625 if (Src.first->getReg() == AMDGPU::ALU_CONST)
626 Consts.push_back(Src.second);
627 if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) ||
628 AMDGPU::R600_KC1RegClass.contains(Src.first->getReg())) {
629 unsigned Index = RI.getEncodingValue(Src.first->getReg()) & 0xff;
630 unsigned Chan = RI.getHWRegChan(Src.first->getReg());
631 Consts.push_back((Index << 2) | Chan);
635 return fitsConstReadLimitations(Consts);
638 DFAPacketizer *R600InstrInfo::CreateTargetScheduleState(const TargetMachine *TM,
639 const ScheduleDAG *DAG) const {
640 const InstrItineraryData *II = TM->getInstrItineraryData();
641 return TM->getSubtarget<AMDGPUSubtarget>().createDFAPacketizer(II);
645 isPredicateSetter(unsigned Opcode) {
654 static MachineInstr *
655 findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
656 MachineBasicBlock::iterator I) {
657 while (I != MBB.begin()) {
659 MachineInstr *MI = I;
660 if (isPredicateSetter(MI->getOpcode()))
668 bool isJump(unsigned Opcode) {
669 return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
672 static bool isBranch(unsigned Opcode) {
673 return Opcode == AMDGPU::BRANCH || Opcode == AMDGPU::BRANCH_COND_i32 ||
674 Opcode == AMDGPU::BRANCH_COND_f32;
678 R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
679 MachineBasicBlock *&TBB,
680 MachineBasicBlock *&FBB,
681 SmallVectorImpl<MachineOperand> &Cond,
682 bool AllowModify) const {
683 // Most of the following comes from the ARM implementation of AnalyzeBranch
685 // If the block has no terminators, it just falls into the block after it.
686 MachineBasicBlock::iterator I = MBB.end();
687 if (I == MBB.begin())
690 while (I->isDebugValue()) {
691 if (I == MBB.begin())
695 // AMDGPU::BRANCH* instructions are only available after isel and are not
697 if (isBranch(I->getOpcode()))
699 if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) {
703 // Get the last instruction in the block.
704 MachineInstr *LastInst = I;
706 // If there is only one terminator instruction, process it.
707 unsigned LastOpc = LastInst->getOpcode();
708 if (I == MBB.begin() ||
709 !isJump(static_cast<MachineInstr *>(--I)->getOpcode())) {
710 if (LastOpc == AMDGPU::JUMP) {
711 TBB = LastInst->getOperand(0).getMBB();
713 } else if (LastOpc == AMDGPU::JUMP_COND) {
714 MachineInstr *predSet = I;
715 while (!isPredicateSetter(predSet->getOpcode())) {
718 TBB = LastInst->getOperand(0).getMBB();
719 Cond.push_back(predSet->getOperand(1));
720 Cond.push_back(predSet->getOperand(2));
721 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
724 return true; // Can't handle indirect branch.
727 // Get the instruction before it if it is a terminator.
728 MachineInstr *SecondLastInst = I;
729 unsigned SecondLastOpc = SecondLastInst->getOpcode();
731 // If the block ends with a B and a Bcc, handle it.
732 if (SecondLastOpc == AMDGPU::JUMP_COND && LastOpc == AMDGPU::JUMP) {
733 MachineInstr *predSet = --I;
734 while (!isPredicateSetter(predSet->getOpcode())) {
737 TBB = SecondLastInst->getOperand(0).getMBB();
738 FBB = LastInst->getOperand(0).getMBB();
739 Cond.push_back(predSet->getOperand(1));
740 Cond.push_back(predSet->getOperand(2));
741 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
745 // Otherwise, can't handle this.
749 int R600InstrInfo::getBranchInstr(const MachineOperand &op) const {
750 const MachineInstr *MI = op.getParent();
752 switch (MI->getDesc().OpInfo->RegClass) {
753 default: // FIXME: fallthrough??
754 case AMDGPU::GPRI32RegClassID: return AMDGPU::BRANCH_COND_i32;
755 case AMDGPU::GPRF32RegClassID: return AMDGPU::BRANCH_COND_f32;
760 MachineBasicBlock::iterator FindLastAluClause(MachineBasicBlock &MBB) {
761 for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend();
763 if (It->getOpcode() == AMDGPU::CF_ALU ||
764 It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE)
765 return llvm::prior(It.base());
771 R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
772 MachineBasicBlock *TBB,
773 MachineBasicBlock *FBB,
774 const SmallVectorImpl<MachineOperand> &Cond,
776 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
780 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB);
783 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
784 assert(PredSet && "No previous predicate !");
785 addFlag(PredSet, 0, MO_FLAG_PUSH);
786 PredSet->getOperand(2).setImm(Cond[1].getImm());
788 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
790 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
791 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
792 if (CfAlu == MBB.end())
794 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
795 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
799 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
800 assert(PredSet && "No previous predicate !");
801 addFlag(PredSet, 0, MO_FLAG_PUSH);
802 PredSet->getOperand(2).setImm(Cond[1].getImm());
803 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
805 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
806 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB);
807 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
808 if (CfAlu == MBB.end())
810 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
811 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
817 R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
819 // Note : we leave PRED* instructions there.
820 // They may be needed when predicating instructions.
822 MachineBasicBlock::iterator I = MBB.end();
824 if (I == MBB.begin()) {
828 switch (I->getOpcode()) {
831 case AMDGPU::JUMP_COND: {
832 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
833 clearFlag(predSet, 0, MO_FLAG_PUSH);
834 I->eraseFromParent();
835 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
836 if (CfAlu == MBB.end())
838 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
839 CfAlu->setDesc(get(AMDGPU::CF_ALU));
843 I->eraseFromParent();
848 if (I == MBB.begin()) {
852 switch (I->getOpcode()) {
853 // FIXME: only one case??
856 case AMDGPU::JUMP_COND: {
857 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
858 clearFlag(predSet, 0, MO_FLAG_PUSH);
859 I->eraseFromParent();
860 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
861 if (CfAlu == MBB.end())
863 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
864 CfAlu->setDesc(get(AMDGPU::CF_ALU));
868 I->eraseFromParent();
875 R600InstrInfo::isPredicated(const MachineInstr *MI) const {
876 int idx = MI->findFirstPredOperandIdx();
880 unsigned Reg = MI->getOperand(idx).getReg();
882 default: return false;
883 case AMDGPU::PRED_SEL_ONE:
884 case AMDGPU::PRED_SEL_ZERO:
885 case AMDGPU::PREDICATE_BIT:
891 R600InstrInfo::isPredicable(MachineInstr *MI) const {
892 // XXX: KILL* instructions can be predicated, but they must be the last
893 // instruction in a clause, so this means any instructions after them cannot
894 // be predicated. Until we have proper support for instruction clauses in the
895 // backend, we will mark KILL* instructions as unpredicable.
897 if (MI->getOpcode() == AMDGPU::KILLGT) {
899 } else if (MI->getOpcode() == AMDGPU::CF_ALU) {
900 // If the clause start in the middle of MBB then the MBB has more
901 // than a single clause, unable to predicate several clauses.
902 if (MI->getParent()->begin() != MachineBasicBlock::iterator(MI))
904 // TODO: We don't support KC merging atm
905 if (MI->getOperand(3).getImm() != 0 || MI->getOperand(4).getImm() != 0)
908 } else if (isVector(*MI)) {
911 return AMDGPUInstrInfo::isPredicable(MI);
917 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
919 unsigned ExtraPredCycles,
920 const BranchProbability &Probability) const{
925 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
927 unsigned ExtraTCycles,
928 MachineBasicBlock &FMBB,
930 unsigned ExtraFCycles,
931 const BranchProbability &Probability) const {
936 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
938 const BranchProbability &Probability)
944 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
945 MachineBasicBlock &FMBB) const {
951 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
952 MachineOperand &MO = Cond[1];
953 switch (MO.getImm()) {
954 case OPCODE_IS_ZERO_INT:
955 MO.setImm(OPCODE_IS_NOT_ZERO_INT);
957 case OPCODE_IS_NOT_ZERO_INT:
958 MO.setImm(OPCODE_IS_ZERO_INT);
961 MO.setImm(OPCODE_IS_NOT_ZERO);
963 case OPCODE_IS_NOT_ZERO:
964 MO.setImm(OPCODE_IS_ZERO);
970 MachineOperand &MO2 = Cond[2];
971 switch (MO2.getReg()) {
972 case AMDGPU::PRED_SEL_ZERO:
973 MO2.setReg(AMDGPU::PRED_SEL_ONE);
975 case AMDGPU::PRED_SEL_ONE:
976 MO2.setReg(AMDGPU::PRED_SEL_ZERO);
985 R600InstrInfo::DefinesPredicate(MachineInstr *MI,
986 std::vector<MachineOperand> &Pred) const {
987 return isPredicateSetter(MI->getOpcode());
992 R600InstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
993 const SmallVectorImpl<MachineOperand> &Pred2) const {
999 R600InstrInfo::PredicateInstruction(MachineInstr *MI,
1000 const SmallVectorImpl<MachineOperand> &Pred) const {
1001 int PIdx = MI->findFirstPredOperandIdx();
1003 if (MI->getOpcode() == AMDGPU::CF_ALU) {
1004 MI->getOperand(8).setImm(0);
1009 MachineOperand &PMO = MI->getOperand(PIdx);
1010 PMO.setReg(Pred[2].getReg());
1011 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1012 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
1019 unsigned int R600InstrInfo::getPredicationCost(const MachineInstr *) const {
1023 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1024 const MachineInstr *MI,
1025 unsigned *PredCost) const {
1031 int R600InstrInfo::getIndirectIndexBegin(const MachineFunction &MF) const {
1032 const MachineRegisterInfo &MRI = MF.getRegInfo();
1033 const MachineFrameInfo *MFI = MF.getFrameInfo();
1036 if (MFI->getNumObjects() == 0) {
1040 if (MRI.livein_empty()) {
1044 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
1045 LE = MRI.livein_end();
1047 Offset = std::max(Offset,
1048 GET_REG_INDEX(RI.getEncodingValue(LI->first)));
1054 int R600InstrInfo::getIndirectIndexEnd(const MachineFunction &MF) const {
1056 const MachineFrameInfo *MFI = MF.getFrameInfo();
1058 // Variable sized objects are not supported
1059 assert(!MFI->hasVarSizedObjects());
1061 if (MFI->getNumObjects() == 0) {
1065 Offset = TM.getFrameLowering()->getFrameIndexOffset(MF, -1);
1067 return getIndirectIndexBegin(MF) + Offset;
1070 std::vector<unsigned> R600InstrInfo::getIndirectReservedRegs(
1071 const MachineFunction &MF) const {
1072 const AMDGPUFrameLowering *TFL =
1073 static_cast<const AMDGPUFrameLowering*>(TM.getFrameLowering());
1074 std::vector<unsigned> Regs;
1076 unsigned StackWidth = TFL->getStackWidth(MF);
1077 int End = getIndirectIndexEnd(MF);
1083 for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
1084 unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
1085 Regs.push_back(SuperReg);
1086 for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
1087 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
1088 Regs.push_back(Reg);
1094 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
1095 unsigned Channel) const {
1096 // XXX: Remove when we support a stack width > 2
1097 assert(Channel == 0);
1101 const TargetRegisterClass * R600InstrInfo::getIndirectAddrStoreRegClass(
1102 unsigned SourceReg) const {
1103 return &AMDGPU::R600_TReg32RegClass;
1106 const TargetRegisterClass *R600InstrInfo::getIndirectAddrLoadRegClass() const {
1107 return &AMDGPU::TRegMemRegClass;
1110 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1111 MachineBasicBlock::iterator I,
1112 unsigned ValueReg, unsigned Address,
1113 unsigned OffsetReg) const {
1114 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1115 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1116 AMDGPU::AR_X, OffsetReg);
1117 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1119 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1121 .addReg(AMDGPU::AR_X,
1122 RegState::Implicit | RegState::Kill);
1123 setImmOperand(Mov, AMDGPU::OpName::dst_rel, 1);
1127 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1128 MachineBasicBlock::iterator I,
1129 unsigned ValueReg, unsigned Address,
1130 unsigned OffsetReg) const {
1131 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1132 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1135 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1136 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1139 .addReg(AMDGPU::AR_X,
1140 RegState::Implicit | RegState::Kill);
1141 setImmOperand(Mov, AMDGPU::OpName::src0_rel, 1);
1146 const TargetRegisterClass *R600InstrInfo::getSuperIndirectRegClass() const {
1147 return &AMDGPU::IndirectRegRegClass;
1150 unsigned R600InstrInfo::getMaxAlusPerClause() const {
1154 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
1155 MachineBasicBlock::iterator I,
1159 unsigned Src1Reg) const {
1160 MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
1164 MIB.addImm(0) // $update_exec_mask
1165 .addImm(0); // $update_predicate
1167 MIB.addImm(1) // $write
1169 .addImm(0) // $dst_rel
1170 .addImm(0) // $dst_clamp
1171 .addReg(Src0Reg) // $src0
1172 .addImm(0) // $src0_neg
1173 .addImm(0) // $src0_rel
1174 .addImm(0) // $src0_abs
1175 .addImm(-1); // $src0_sel
1178 MIB.addReg(Src1Reg) // $src1
1179 .addImm(0) // $src1_neg
1180 .addImm(0) // $src1_rel
1181 .addImm(0) // $src1_abs
1182 .addImm(-1); // $src1_sel
1185 //XXX: The r600g finalizer expects this to be 1, once we've moved the
1186 //scheduling to the backend, we can change the default to 0.
1187 MIB.addImm(1) // $last
1188 .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
1189 .addImm(0) // $literal
1190 .addImm(0); // $bank_swizzle
1195 #define OPERAND_CASE(Label) \
1197 static const unsigned Ops[] = \
1207 static unsigned getSlotedOps(unsigned Op, unsigned Slot) {
1209 OPERAND_CASE(AMDGPU::OpName::update_exec_mask)
1210 OPERAND_CASE(AMDGPU::OpName::update_pred)
1211 OPERAND_CASE(AMDGPU::OpName::write)
1212 OPERAND_CASE(AMDGPU::OpName::omod)
1213 OPERAND_CASE(AMDGPU::OpName::dst_rel)
1214 OPERAND_CASE(AMDGPU::OpName::clamp)
1215 OPERAND_CASE(AMDGPU::OpName::src0)
1216 OPERAND_CASE(AMDGPU::OpName::src0_neg)
1217 OPERAND_CASE(AMDGPU::OpName::src0_rel)
1218 OPERAND_CASE(AMDGPU::OpName::src0_abs)
1219 OPERAND_CASE(AMDGPU::OpName::src0_sel)
1220 OPERAND_CASE(AMDGPU::OpName::src1)
1221 OPERAND_CASE(AMDGPU::OpName::src1_neg)
1222 OPERAND_CASE(AMDGPU::OpName::src1_rel)
1223 OPERAND_CASE(AMDGPU::OpName::src1_abs)
1224 OPERAND_CASE(AMDGPU::OpName::src1_sel)
1225 OPERAND_CASE(AMDGPU::OpName::pred_sel)
1227 llvm_unreachable("Wrong Operand");
1233 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
1234 MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg)
1236 assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
1238 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
1239 if (ST.getGeneration() <= AMDGPUSubtarget::R700)
1240 Opcode = AMDGPU::DOT4_r600;
1242 Opcode = AMDGPU::DOT4_eg;
1243 MachineBasicBlock::iterator I = MI;
1244 MachineOperand &Src0 = MI->getOperand(
1245 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src0, Slot)));
1246 MachineOperand &Src1 = MI->getOperand(
1247 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src1, Slot)));
1248 MachineInstr *MIB = buildDefaultInstruction(
1249 MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg());
1250 static const unsigned Operands[14] = {
1251 AMDGPU::OpName::update_exec_mask,
1252 AMDGPU::OpName::update_pred,
1253 AMDGPU::OpName::write,
1254 AMDGPU::OpName::omod,
1255 AMDGPU::OpName::dst_rel,
1256 AMDGPU::OpName::clamp,
1257 AMDGPU::OpName::src0_neg,
1258 AMDGPU::OpName::src0_rel,
1259 AMDGPU::OpName::src0_abs,
1260 AMDGPU::OpName::src0_sel,
1261 AMDGPU::OpName::src1_neg,
1262 AMDGPU::OpName::src1_rel,
1263 AMDGPU::OpName::src1_abs,
1264 AMDGPU::OpName::src1_sel,
1267 for (unsigned i = 0; i < 14; i++) {
1268 MachineOperand &MO = MI->getOperand(
1269 getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));
1270 assert (MO.isImm());
1271 setImmOperand(MIB, Operands[i], MO.getImm());
1273 MIB->getOperand(20).setImm(0);
1277 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
1278 MachineBasicBlock::iterator I,
1280 uint64_t Imm) const {
1281 MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
1282 AMDGPU::ALU_LITERAL_X);
1283 setImmOperand(MovImm, AMDGPU::OpName::literal, Imm);
1287 int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const {
1288 return getOperandIdx(MI.getOpcode(), Op);
1291 int R600InstrInfo::getOperandIdx(unsigned Opcode, unsigned Op) const {
1292 return AMDGPU::getNamedOperandIdx(Opcode, Op);
1295 void R600InstrInfo::setImmOperand(MachineInstr *MI, unsigned Op,
1296 int64_t Imm) const {
1297 int Idx = getOperandIdx(*MI, Op);
1298 assert(Idx != -1 && "Operand not supported for this instruction.");
1299 assert(MI->getOperand(Idx).isImm());
1300 MI->getOperand(Idx).setImm(Imm);
1303 //===----------------------------------------------------------------------===//
1304 // Instruction flag getters/setters
1305 //===----------------------------------------------------------------------===//
1307 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
1308 return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
1311 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
1312 unsigned Flag) const {
1313 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1316 // If we pass something other than the default value of Flag to this
1317 // function, it means we are want to set a flag on an instruction
1318 // that uses native encoding.
1319 assert(HAS_NATIVE_OPERANDS(TargetFlags));
1320 bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
1323 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::clamp);
1326 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::write);
1328 case MO_FLAG_NOT_LAST:
1330 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::last);
1334 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_neg); break;
1335 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_neg); break;
1336 case 2: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src2_neg); break;
1341 assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
1345 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_abs); break;
1346 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_abs); break;
1354 assert(FlagIndex != -1 && "Flag not supported for this instruction");
1356 FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
1357 assert(FlagIndex != 0 &&
1358 "Instruction flags not supported for this instruction");
1361 MachineOperand &FlagOp = MI->getOperand(FlagIndex);
1362 assert(FlagOp.isImm());
1366 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
1367 unsigned Flag) const {
1368 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1372 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1373 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1374 if (Flag == MO_FLAG_NOT_LAST) {
1375 clearFlag(MI, Operand, MO_FLAG_LAST);
1376 } else if (Flag == MO_FLAG_MASK) {
1377 clearFlag(MI, Operand, Flag);
1382 MachineOperand &FlagOp = getFlagOp(MI, Operand);
1383 FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
1387 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
1388 unsigned Flag) const {
1389 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1390 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1391 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1394 MachineOperand &FlagOp = getFlagOp(MI);
1395 unsigned InstFlags = FlagOp.getImm();
1396 InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
1397 FlagOp.setImm(InstFlags);