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"
28 #define GET_INSTRINFO_CTOR_DTOR
29 #include "AMDGPUGenDFAPacketizer.inc"
31 R600InstrInfo::R600InstrInfo(const AMDGPUSubtarget &st)
32 : AMDGPUInstrInfo(st),
36 const R600RegisterInfo &R600InstrInfo::getRegisterInfo() const {
40 bool R600InstrInfo::isTrig(const MachineInstr &MI) const {
41 return get(MI.getOpcode()).TSFlags & R600_InstFlag::TRIG;
44 bool R600InstrInfo::isVector(const MachineInstr &MI) const {
45 return get(MI.getOpcode()).TSFlags & R600_InstFlag::VECTOR;
49 R600InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
50 MachineBasicBlock::iterator MI, DebugLoc DL,
51 unsigned DestReg, unsigned SrcReg,
53 unsigned VectorComponents = 0;
54 if ((AMDGPU::R600_Reg128RegClass.contains(DestReg) ||
55 AMDGPU::R600_Reg128VerticalRegClass.contains(DestReg)) &&
56 (AMDGPU::R600_Reg128RegClass.contains(SrcReg) ||
57 AMDGPU::R600_Reg128VerticalRegClass.contains(SrcReg))) {
59 } else if((AMDGPU::R600_Reg64RegClass.contains(DestReg) ||
60 AMDGPU::R600_Reg64VerticalRegClass.contains(DestReg)) &&
61 (AMDGPU::R600_Reg64RegClass.contains(SrcReg) ||
62 AMDGPU::R600_Reg64VerticalRegClass.contains(SrcReg))) {
66 if (VectorComponents > 0) {
67 for (unsigned I = 0; I < VectorComponents; I++) {
68 unsigned SubRegIndex = RI.getSubRegFromChannel(I);
69 buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
70 RI.getSubReg(DestReg, SubRegIndex),
71 RI.getSubReg(SrcReg, SubRegIndex))
73 RegState::Define | RegState::Implicit);
76 MachineInstr *NewMI = buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
78 NewMI->getOperand(getOperandIdx(*NewMI, AMDGPU::OpName::src0))
83 /// \returns true if \p MBBI can be moved into a new basic.
84 bool R600InstrInfo::isLegalToSplitMBBAt(MachineBasicBlock &MBB,
85 MachineBasicBlock::iterator MBBI) const {
86 for (MachineInstr::const_mop_iterator I = MBBI->operands_begin(),
87 E = MBBI->operands_end(); I != E; ++I) {
88 if (I->isReg() && !TargetRegisterInfo::isVirtualRegister(I->getReg()) &&
89 I->isUse() && RI.isPhysRegLiveAcrossClauses(I->getReg()))
95 bool R600InstrInfo::isMov(unsigned Opcode) const {
99 default: return false;
101 case AMDGPU::MOV_IMM_F32:
102 case AMDGPU::MOV_IMM_I32:
107 // Some instructions act as place holders to emulate operations that the GPU
108 // hardware does automatically. This function can be used to check if
109 // an opcode falls into this category.
110 bool R600InstrInfo::isPlaceHolderOpcode(unsigned Opcode) const {
112 default: return false;
118 bool R600InstrInfo::isReductionOp(unsigned Opcode) const {
122 bool R600InstrInfo::isCubeOp(unsigned Opcode) const {
124 default: return false;
125 case AMDGPU::CUBE_r600_pseudo:
126 case AMDGPU::CUBE_r600_real:
127 case AMDGPU::CUBE_eg_pseudo:
128 case AMDGPU::CUBE_eg_real:
133 bool R600InstrInfo::isALUInstr(unsigned Opcode) const {
134 unsigned TargetFlags = get(Opcode).TSFlags;
136 return (TargetFlags & R600_InstFlag::ALU_INST);
139 bool R600InstrInfo::hasInstrModifiers(unsigned Opcode) const {
140 unsigned TargetFlags = get(Opcode).TSFlags;
142 return ((TargetFlags & R600_InstFlag::OP1) |
143 (TargetFlags & R600_InstFlag::OP2) |
144 (TargetFlags & R600_InstFlag::OP3));
147 bool R600InstrInfo::isLDSInstr(unsigned Opcode) const {
148 unsigned TargetFlags = get(Opcode).TSFlags;
150 return ((TargetFlags & R600_InstFlag::LDS_1A) |
151 (TargetFlags & R600_InstFlag::LDS_1A1D) |
152 (TargetFlags & R600_InstFlag::LDS_1A2D));
155 bool R600InstrInfo::isLDSNoRetInstr(unsigned Opcode) const {
156 return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) == -1;
159 bool R600InstrInfo::isLDSRetInstr(unsigned Opcode) const {
160 return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) != -1;
163 bool R600InstrInfo::canBeConsideredALU(const MachineInstr *MI) const {
164 if (isALUInstr(MI->getOpcode()))
166 if (isVector(*MI) || isCubeOp(MI->getOpcode()))
168 switch (MI->getOpcode()) {
170 case AMDGPU::INTERP_PAIR_XY:
171 case AMDGPU::INTERP_PAIR_ZW:
172 case AMDGPU::INTERP_VEC_LOAD:
181 bool R600InstrInfo::isTransOnly(unsigned Opcode) const {
182 if (ST.hasCaymanISA())
184 return (get(Opcode).getSchedClass() == AMDGPU::Sched::TransALU);
187 bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const {
188 return isTransOnly(MI->getOpcode());
191 bool R600InstrInfo::isVectorOnly(unsigned Opcode) const {
192 return (get(Opcode).getSchedClass() == AMDGPU::Sched::VecALU);
195 bool R600InstrInfo::isVectorOnly(const MachineInstr *MI) const {
196 return isVectorOnly(MI->getOpcode());
199 bool R600InstrInfo::isExport(unsigned Opcode) const {
200 return (get(Opcode).TSFlags & R600_InstFlag::IS_EXPORT);
203 bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
204 return ST.hasVertexCache() && IS_VTX(get(Opcode));
207 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const {
208 const MachineFunction *MF = MI->getParent()->getParent();
209 const R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
210 return MFI->getShaderType() != ShaderType::COMPUTE &&
211 usesVertexCache(MI->getOpcode());
214 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
215 return (!ST.hasVertexCache() && IS_VTX(get(Opcode))) || IS_TEX(get(Opcode));
218 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
219 const MachineFunction *MF = MI->getParent()->getParent();
220 const R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
221 return (MFI->getShaderType() == ShaderType::COMPUTE &&
222 usesVertexCache(MI->getOpcode())) ||
223 usesTextureCache(MI->getOpcode());
226 bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
229 case AMDGPU::GROUP_BARRIER:
236 bool R600InstrInfo::usesAddressRegister(MachineInstr *MI) const {
237 return MI->findRegisterUseOperandIdx(AMDGPU::AR_X) != -1;
240 bool R600InstrInfo::definesAddressRegister(MachineInstr *MI) const {
241 return MI->findRegisterDefOperandIdx(AMDGPU::AR_X) != -1;
244 bool R600InstrInfo::readsLDSSrcReg(const MachineInstr *MI) const {
245 if (!isALUInstr(MI->getOpcode())) {
248 for (MachineInstr::const_mop_iterator I = MI->operands_begin(),
249 E = MI->operands_end(); I != E; ++I) {
250 if (!I->isReg() || !I->isUse() ||
251 TargetRegisterInfo::isVirtualRegister(I->getReg()))
254 if (AMDGPU::R600_LDS_SRC_REGRegClass.contains(I->getReg()))
260 int R600InstrInfo::getSrcIdx(unsigned Opcode, unsigned SrcNum) const {
261 static const unsigned OpTable[] = {
262 AMDGPU::OpName::src0,
263 AMDGPU::OpName::src1,
268 return getOperandIdx(Opcode, OpTable[SrcNum]);
271 int R600InstrInfo::getSelIdx(unsigned Opcode, unsigned SrcIdx) const {
272 static const unsigned SrcSelTable[][2] = {
273 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
274 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
275 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
276 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
277 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
278 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
279 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
280 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
281 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
282 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
283 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W}
286 for (const auto &Row : SrcSelTable) {
287 if (getOperandIdx(Opcode, Row[0]) == (int)SrcIdx) {
288 return getOperandIdx(Opcode, Row[1]);
294 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
295 R600InstrInfo::getSrcs(MachineInstr *MI) const {
296 SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
298 if (MI->getOpcode() == AMDGPU::DOT_4) {
299 static const unsigned OpTable[8][2] = {
300 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
301 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
302 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
303 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
304 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
305 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
306 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
307 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W},
310 for (unsigned j = 0; j < 8; j++) {
311 MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
313 unsigned Reg = MO.getReg();
314 if (Reg == AMDGPU::ALU_CONST) {
315 unsigned Sel = MI->getOperand(getOperandIdx(MI->getOpcode(),
316 OpTable[j][1])).getImm();
317 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
325 static const unsigned OpTable[3][2] = {
326 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
327 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
328 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
331 for (unsigned j = 0; j < 3; j++) {
332 int SrcIdx = getOperandIdx(MI->getOpcode(), OpTable[j][0]);
335 MachineOperand &MO = MI->getOperand(SrcIdx);
336 unsigned Reg = MI->getOperand(SrcIdx).getReg();
337 if (Reg == AMDGPU::ALU_CONST) {
338 unsigned Sel = MI->getOperand(
339 getOperandIdx(MI->getOpcode(), OpTable[j][1])).getImm();
340 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
343 if (Reg == AMDGPU::ALU_LITERAL_X) {
344 unsigned Imm = MI->getOperand(
345 getOperandIdx(MI->getOpcode(), AMDGPU::OpName::literal)).getImm();
346 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Imm));
349 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, 0));
354 std::vector<std::pair<int, unsigned> >
355 R600InstrInfo::ExtractSrcs(MachineInstr *MI,
356 const DenseMap<unsigned, unsigned> &PV,
357 unsigned &ConstCount) const {
359 const SmallVector<std::pair<MachineOperand *, int64_t>, 3> Srcs = getSrcs(MI);
360 const std::pair<int, unsigned> DummyPair(-1, 0);
361 std::vector<std::pair<int, unsigned> > Result;
363 for (unsigned n = Srcs.size(); i < n; ++i) {
364 unsigned Reg = Srcs[i].first->getReg();
365 unsigned Index = RI.getEncodingValue(Reg) & 0xff;
366 if (Reg == AMDGPU::OQAP) {
367 Result.push_back(std::pair<int, unsigned>(Index, 0));
369 if (PV.find(Reg) != PV.end()) {
370 // 255 is used to tells its a PS/PV reg
371 Result.push_back(std::pair<int, unsigned>(255, 0));
376 Result.push_back(DummyPair);
379 unsigned Chan = RI.getHWRegChan(Reg);
380 Result.push_back(std::pair<int, unsigned>(Index, Chan));
383 Result.push_back(DummyPair);
387 static std::vector<std::pair<int, unsigned> >
388 Swizzle(std::vector<std::pair<int, unsigned> > Src,
389 R600InstrInfo::BankSwizzle Swz) {
390 if (Src[0] == Src[1])
393 case R600InstrInfo::ALU_VEC_012_SCL_210:
395 case R600InstrInfo::ALU_VEC_021_SCL_122:
396 std::swap(Src[1], Src[2]);
398 case R600InstrInfo::ALU_VEC_102_SCL_221:
399 std::swap(Src[0], Src[1]);
401 case R600InstrInfo::ALU_VEC_120_SCL_212:
402 std::swap(Src[0], Src[1]);
403 std::swap(Src[0], Src[2]);
405 case R600InstrInfo::ALU_VEC_201:
406 std::swap(Src[0], Src[2]);
407 std::swap(Src[0], Src[1]);
409 case R600InstrInfo::ALU_VEC_210:
410 std::swap(Src[0], Src[2]);
417 getTransSwizzle(R600InstrInfo::BankSwizzle Swz, unsigned Op) {
419 case R600InstrInfo::ALU_VEC_012_SCL_210: {
420 unsigned Cycles[3] = { 2, 1, 0};
423 case R600InstrInfo::ALU_VEC_021_SCL_122: {
424 unsigned Cycles[3] = { 1, 2, 2};
427 case R600InstrInfo::ALU_VEC_120_SCL_212: {
428 unsigned Cycles[3] = { 2, 1, 2};
431 case R600InstrInfo::ALU_VEC_102_SCL_221: {
432 unsigned Cycles[3] = { 2, 2, 1};
436 llvm_unreachable("Wrong Swizzle for Trans Slot");
441 /// returns how many MIs (whose inputs are represented by IGSrcs) can be packed
442 /// in the same Instruction Group while meeting read port limitations given a
443 /// Swz swizzle sequence.
444 unsigned R600InstrInfo::isLegalUpTo(
445 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
446 const std::vector<R600InstrInfo::BankSwizzle> &Swz,
447 const std::vector<std::pair<int, unsigned> > &TransSrcs,
448 R600InstrInfo::BankSwizzle TransSwz) const {
450 memset(Vector, -1, sizeof(Vector));
451 for (unsigned i = 0, e = IGSrcs.size(); i < e; i++) {
452 const std::vector<std::pair<int, unsigned> > &Srcs =
453 Swizzle(IGSrcs[i], Swz[i]);
454 for (unsigned j = 0; j < 3; j++) {
455 const std::pair<int, unsigned> &Src = Srcs[j];
456 if (Src.first < 0 || Src.first == 255)
458 if (Src.first == GET_REG_INDEX(RI.getEncodingValue(AMDGPU::OQAP))) {
459 if (Swz[i] != R600InstrInfo::ALU_VEC_012_SCL_210 &&
460 Swz[i] != R600InstrInfo::ALU_VEC_021_SCL_122) {
461 // The value from output queue A (denoted by register OQAP) can
462 // only be fetched during the first cycle.
465 // OQAP does not count towards the normal read port restrictions
468 if (Vector[Src.second][j] < 0)
469 Vector[Src.second][j] = Src.first;
470 if (Vector[Src.second][j] != Src.first)
474 // Now check Trans Alu
475 for (unsigned i = 0, e = TransSrcs.size(); i < e; ++i) {
476 const std::pair<int, unsigned> &Src = TransSrcs[i];
477 unsigned Cycle = getTransSwizzle(TransSwz, i);
480 if (Src.first == 255)
482 if (Vector[Src.second][Cycle] < 0)
483 Vector[Src.second][Cycle] = Src.first;
484 if (Vector[Src.second][Cycle] != Src.first)
485 return IGSrcs.size() - 1;
487 return IGSrcs.size();
490 /// Given a swizzle sequence SwzCandidate and an index Idx, returns the next
491 /// (in lexicographic term) swizzle sequence assuming that all swizzles after
492 /// Idx can be skipped
494 NextPossibleSolution(
495 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
497 assert(Idx < SwzCandidate.size());
499 while (ResetIdx > -1 && SwzCandidate[ResetIdx] == R600InstrInfo::ALU_VEC_210)
501 for (unsigned i = ResetIdx + 1, e = SwzCandidate.size(); i < e; i++) {
502 SwzCandidate[i] = R600InstrInfo::ALU_VEC_012_SCL_210;
506 int NextSwizzle = SwzCandidate[ResetIdx] + 1;
507 SwzCandidate[ResetIdx] = (R600InstrInfo::BankSwizzle)NextSwizzle;
511 /// Enumerate all possible Swizzle sequence to find one that can meet all
512 /// read port requirements.
513 bool R600InstrInfo::FindSwizzleForVectorSlot(
514 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
515 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
516 const std::vector<std::pair<int, unsigned> > &TransSrcs,
517 R600InstrInfo::BankSwizzle TransSwz) const {
518 unsigned ValidUpTo = 0;
520 ValidUpTo = isLegalUpTo(IGSrcs, SwzCandidate, TransSrcs, TransSwz);
521 if (ValidUpTo == IGSrcs.size())
523 } while (NextPossibleSolution(SwzCandidate, ValidUpTo));
527 /// Instructions in Trans slot can't read gpr at cycle 0 if they also read
528 /// a const, and can't read a gpr at cycle 1 if they read 2 const.
530 isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,
531 const std::vector<std::pair<int, unsigned> > &TransOps,
532 unsigned ConstCount) {
533 // TransALU can't read 3 constants
536 for (unsigned i = 0, e = TransOps.size(); i < e; ++i) {
537 const std::pair<int, unsigned> &Src = TransOps[i];
538 unsigned Cycle = getTransSwizzle(TransSwz, i);
541 if (ConstCount > 0 && Cycle == 0)
543 if (ConstCount > 1 && Cycle == 1)
550 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG,
551 const DenseMap<unsigned, unsigned> &PV,
552 std::vector<BankSwizzle> &ValidSwizzle,
555 //Todo : support shared src0 - src1 operand
557 std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs;
558 ValidSwizzle.clear();
560 BankSwizzle TransBS = ALU_VEC_012_SCL_210;
561 for (unsigned i = 0, e = IG.size(); i < e; ++i) {
562 IGSrcs.push_back(ExtractSrcs(IG[i], PV, ConstCount));
563 unsigned Op = getOperandIdx(IG[i]->getOpcode(),
564 AMDGPU::OpName::bank_swizzle);
565 ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle)
566 IG[i]->getOperand(Op).getImm());
568 std::vector<std::pair<int, unsigned> > TransOps;
570 return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS);
572 TransOps = std::move(IGSrcs.back());
574 ValidSwizzle.pop_back();
576 static const R600InstrInfo::BankSwizzle TransSwz[] = {
582 for (unsigned i = 0; i < 4; i++) {
583 TransBS = TransSwz[i];
584 if (!isConstCompatible(TransBS, TransOps, ConstCount))
586 bool Result = FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps,
589 ValidSwizzle.push_back(TransBS);
599 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts)
601 assert (Consts.size() <= 12 && "Too many operands in instructions group");
602 unsigned Pair1 = 0, Pair2 = 0;
603 for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
604 unsigned ReadConstHalf = Consts[i] & 2;
605 unsigned ReadConstIndex = Consts[i] & (~3);
606 unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf;
608 Pair1 = ReadHalfConst;
611 if (Pair1 == ReadHalfConst)
614 Pair2 = ReadHalfConst;
617 if (Pair2 != ReadHalfConst)
624 R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
626 std::vector<unsigned> Consts;
627 SmallSet<int64_t, 4> Literals;
628 for (unsigned i = 0, n = MIs.size(); i < n; i++) {
629 MachineInstr *MI = MIs[i];
630 if (!isALUInstr(MI->getOpcode()))
633 const SmallVectorImpl<std::pair<MachineOperand *, int64_t> > &Srcs =
636 for (unsigned j = 0, e = Srcs.size(); j < e; j++) {
637 std::pair<MachineOperand *, unsigned> Src = Srcs[j];
638 if (Src.first->getReg() == AMDGPU::ALU_LITERAL_X)
639 Literals.insert(Src.second);
640 if (Literals.size() > 4)
642 if (Src.first->getReg() == AMDGPU::ALU_CONST)
643 Consts.push_back(Src.second);
644 if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) ||
645 AMDGPU::R600_KC1RegClass.contains(Src.first->getReg())) {
646 unsigned Index = RI.getEncodingValue(Src.first->getReg()) & 0xff;
647 unsigned Chan = RI.getHWRegChan(Src.first->getReg());
648 Consts.push_back((Index << 2) | Chan);
652 return fitsConstReadLimitations(Consts);
656 R600InstrInfo::CreateTargetScheduleState(const TargetSubtargetInfo &STI) const {
657 const InstrItineraryData *II = STI.getInstrItineraryData();
658 return static_cast<const AMDGPUSubtarget &>(STI).createDFAPacketizer(II);
662 isPredicateSetter(unsigned Opcode) {
671 static MachineInstr *
672 findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
673 MachineBasicBlock::iterator I) {
674 while (I != MBB.begin()) {
676 MachineInstr *MI = I;
677 if (isPredicateSetter(MI->getOpcode()))
685 bool isJump(unsigned Opcode) {
686 return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
689 static bool isBranch(unsigned Opcode) {
690 return Opcode == AMDGPU::BRANCH || Opcode == AMDGPU::BRANCH_COND_i32 ||
691 Opcode == AMDGPU::BRANCH_COND_f32;
695 R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
696 MachineBasicBlock *&TBB,
697 MachineBasicBlock *&FBB,
698 SmallVectorImpl<MachineOperand> &Cond,
699 bool AllowModify) const {
700 // Most of the following comes from the ARM implementation of AnalyzeBranch
702 // If the block has no terminators, it just falls into the block after it.
703 MachineBasicBlock::iterator I = MBB.end();
704 if (I == MBB.begin())
707 while (I->isDebugValue()) {
708 if (I == MBB.begin())
712 // AMDGPU::BRANCH* instructions are only available after isel and are not
714 if (isBranch(I->getOpcode()))
716 if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) {
720 // Remove successive JUMP
721 while (I != MBB.begin() && std::prev(I)->getOpcode() == AMDGPU::JUMP) {
722 MachineBasicBlock::iterator PriorI = std::prev(I);
724 I->removeFromParent();
727 MachineInstr *LastInst = I;
729 // If there is only one terminator instruction, process it.
730 unsigned LastOpc = LastInst->getOpcode();
731 if (I == MBB.begin() ||
732 !isJump(static_cast<MachineInstr *>(--I)->getOpcode())) {
733 if (LastOpc == AMDGPU::JUMP) {
734 TBB = LastInst->getOperand(0).getMBB();
736 } else if (LastOpc == AMDGPU::JUMP_COND) {
737 MachineInstr *predSet = I;
738 while (!isPredicateSetter(predSet->getOpcode())) {
741 TBB = LastInst->getOperand(0).getMBB();
742 Cond.push_back(predSet->getOperand(1));
743 Cond.push_back(predSet->getOperand(2));
744 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
747 return true; // Can't handle indirect branch.
750 // Get the instruction before it if it is a terminator.
751 MachineInstr *SecondLastInst = I;
752 unsigned SecondLastOpc = SecondLastInst->getOpcode();
754 // If the block ends with a B and a Bcc, handle it.
755 if (SecondLastOpc == AMDGPU::JUMP_COND && LastOpc == AMDGPU::JUMP) {
756 MachineInstr *predSet = --I;
757 while (!isPredicateSetter(predSet->getOpcode())) {
760 TBB = SecondLastInst->getOperand(0).getMBB();
761 FBB = LastInst->getOperand(0).getMBB();
762 Cond.push_back(predSet->getOperand(1));
763 Cond.push_back(predSet->getOperand(2));
764 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
768 // Otherwise, can't handle this.
773 MachineBasicBlock::iterator FindLastAluClause(MachineBasicBlock &MBB) {
774 for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend();
776 if (It->getOpcode() == AMDGPU::CF_ALU ||
777 It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE)
778 return std::prev(It.base());
784 R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
785 MachineBasicBlock *TBB,
786 MachineBasicBlock *FBB,
787 const SmallVectorImpl<MachineOperand> &Cond,
789 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
793 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB);
796 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
797 assert(PredSet && "No previous predicate !");
798 addFlag(PredSet, 0, MO_FLAG_PUSH);
799 PredSet->getOperand(2).setImm(Cond[1].getImm());
801 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
803 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
804 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
805 if (CfAlu == MBB.end())
807 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
808 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
812 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
813 assert(PredSet && "No previous predicate !");
814 addFlag(PredSet, 0, MO_FLAG_PUSH);
815 PredSet->getOperand(2).setImm(Cond[1].getImm());
816 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
818 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
819 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB);
820 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
821 if (CfAlu == MBB.end())
823 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
824 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
830 R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
832 // Note : we leave PRED* instructions there.
833 // They may be needed when predicating instructions.
835 MachineBasicBlock::iterator I = MBB.end();
837 if (I == MBB.begin()) {
841 switch (I->getOpcode()) {
844 case AMDGPU::JUMP_COND: {
845 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
846 clearFlag(predSet, 0, MO_FLAG_PUSH);
847 I->eraseFromParent();
848 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
849 if (CfAlu == MBB.end())
851 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
852 CfAlu->setDesc(get(AMDGPU::CF_ALU));
856 I->eraseFromParent();
861 if (I == MBB.begin()) {
865 switch (I->getOpcode()) {
866 // FIXME: only one case??
869 case AMDGPU::JUMP_COND: {
870 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
871 clearFlag(predSet, 0, MO_FLAG_PUSH);
872 I->eraseFromParent();
873 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
874 if (CfAlu == MBB.end())
876 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
877 CfAlu->setDesc(get(AMDGPU::CF_ALU));
881 I->eraseFromParent();
888 R600InstrInfo::isPredicated(const MachineInstr *MI) const {
889 int idx = MI->findFirstPredOperandIdx();
893 unsigned Reg = MI->getOperand(idx).getReg();
895 default: return false;
896 case AMDGPU::PRED_SEL_ONE:
897 case AMDGPU::PRED_SEL_ZERO:
898 case AMDGPU::PREDICATE_BIT:
904 R600InstrInfo::isPredicable(MachineInstr *MI) const {
905 // XXX: KILL* instructions can be predicated, but they must be the last
906 // instruction in a clause, so this means any instructions after them cannot
907 // be predicated. Until we have proper support for instruction clauses in the
908 // backend, we will mark KILL* instructions as unpredicable.
910 if (MI->getOpcode() == AMDGPU::KILLGT) {
912 } else if (MI->getOpcode() == AMDGPU::CF_ALU) {
913 // If the clause start in the middle of MBB then the MBB has more
914 // than a single clause, unable to predicate several clauses.
915 if (MI->getParent()->begin() != MachineBasicBlock::iterator(MI))
917 // TODO: We don't support KC merging atm
918 if (MI->getOperand(3).getImm() != 0 || MI->getOperand(4).getImm() != 0)
921 } else if (isVector(*MI)) {
924 return AMDGPUInstrInfo::isPredicable(MI);
930 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
932 unsigned ExtraPredCycles,
933 const BranchProbability &Probability) const{
938 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
940 unsigned ExtraTCycles,
941 MachineBasicBlock &FMBB,
943 unsigned ExtraFCycles,
944 const BranchProbability &Probability) const {
949 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
951 const BranchProbability &Probability)
957 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
958 MachineBasicBlock &FMBB) const {
964 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
965 MachineOperand &MO = Cond[1];
966 switch (MO.getImm()) {
967 case OPCODE_IS_ZERO_INT:
968 MO.setImm(OPCODE_IS_NOT_ZERO_INT);
970 case OPCODE_IS_NOT_ZERO_INT:
971 MO.setImm(OPCODE_IS_ZERO_INT);
974 MO.setImm(OPCODE_IS_NOT_ZERO);
976 case OPCODE_IS_NOT_ZERO:
977 MO.setImm(OPCODE_IS_ZERO);
983 MachineOperand &MO2 = Cond[2];
984 switch (MO2.getReg()) {
985 case AMDGPU::PRED_SEL_ZERO:
986 MO2.setReg(AMDGPU::PRED_SEL_ONE);
988 case AMDGPU::PRED_SEL_ONE:
989 MO2.setReg(AMDGPU::PRED_SEL_ZERO);
998 R600InstrInfo::DefinesPredicate(MachineInstr *MI,
999 std::vector<MachineOperand> &Pred) const {
1000 return isPredicateSetter(MI->getOpcode());
1005 R600InstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
1006 const SmallVectorImpl<MachineOperand> &Pred2) const {
1012 R600InstrInfo::PredicateInstruction(MachineInstr *MI,
1013 const SmallVectorImpl<MachineOperand> &Pred) const {
1014 int PIdx = MI->findFirstPredOperandIdx();
1016 if (MI->getOpcode() == AMDGPU::CF_ALU) {
1017 MI->getOperand(8).setImm(0);
1021 if (MI->getOpcode() == AMDGPU::DOT_4) {
1022 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_X))
1023 .setReg(Pred[2].getReg());
1024 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Y))
1025 .setReg(Pred[2].getReg());
1026 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Z))
1027 .setReg(Pred[2].getReg());
1028 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_W))
1029 .setReg(Pred[2].getReg());
1030 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1031 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
1036 MachineOperand &PMO = MI->getOperand(PIdx);
1037 PMO.setReg(Pred[2].getReg());
1038 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1039 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
1046 unsigned int R600InstrInfo::getPredicationCost(const MachineInstr *) const {
1050 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1051 const MachineInstr *MI,
1052 unsigned *PredCost) const {
1058 bool R600InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
1060 switch(MI->getOpcode()) {
1061 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
1062 case AMDGPU::R600_EXTRACT_ELT_V2:
1063 case AMDGPU::R600_EXTRACT_ELT_V4:
1064 buildIndirectRead(MI->getParent(), MI, MI->getOperand(0).getReg(),
1065 RI.getHWRegIndex(MI->getOperand(1).getReg()), // Address
1066 MI->getOperand(2).getReg(),
1067 RI.getHWRegChan(MI->getOperand(1).getReg()));
1069 case AMDGPU::R600_INSERT_ELT_V2:
1070 case AMDGPU::R600_INSERT_ELT_V4:
1071 buildIndirectWrite(MI->getParent(), MI, MI->getOperand(2).getReg(), // Value
1072 RI.getHWRegIndex(MI->getOperand(1).getReg()), // Address
1073 MI->getOperand(3).getReg(), // Offset
1074 RI.getHWRegChan(MI->getOperand(1).getReg())); // Channel
1077 MI->eraseFromParent();
1081 void R600InstrInfo::reserveIndirectRegisters(BitVector &Reserved,
1082 const MachineFunction &MF) const {
1083 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
1084 MF.getSubtarget().getFrameLowering());
1086 unsigned StackWidth = TFL->getStackWidth(MF);
1087 int End = getIndirectIndexEnd(MF);
1092 for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
1093 unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
1094 Reserved.set(SuperReg);
1095 for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
1096 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
1102 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
1103 unsigned Channel) const {
1104 // XXX: Remove when we support a stack width > 2
1105 assert(Channel == 0);
1109 const TargetRegisterClass *R600InstrInfo::getIndirectAddrRegClass() const {
1110 return &AMDGPU::R600_TReg32_XRegClass;
1113 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1114 MachineBasicBlock::iterator I,
1115 unsigned ValueReg, unsigned Address,
1116 unsigned OffsetReg) const {
1117 return buildIndirectWrite(MBB, I, ValueReg, Address, OffsetReg, 0);
1120 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1121 MachineBasicBlock::iterator I,
1122 unsigned ValueReg, unsigned Address,
1124 unsigned AddrChan) const {
1127 default: llvm_unreachable("Invalid Channel");
1128 case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break;
1129 case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break;
1130 case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break;
1131 case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break;
1133 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1134 AMDGPU::AR_X, OffsetReg);
1135 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1137 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1139 .addReg(AMDGPU::AR_X,
1140 RegState::Implicit | RegState::Kill);
1141 setImmOperand(Mov, AMDGPU::OpName::dst_rel, 1);
1145 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1146 MachineBasicBlock::iterator I,
1147 unsigned ValueReg, unsigned Address,
1148 unsigned OffsetReg) const {
1149 return buildIndirectRead(MBB, I, ValueReg, Address, OffsetReg, 0);
1152 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1153 MachineBasicBlock::iterator I,
1154 unsigned ValueReg, unsigned Address,
1156 unsigned AddrChan) const {
1159 default: llvm_unreachable("Invalid Channel");
1160 case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break;
1161 case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break;
1162 case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break;
1163 case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break;
1165 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1168 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1169 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1172 .addReg(AMDGPU::AR_X,
1173 RegState::Implicit | RegState::Kill);
1174 setImmOperand(Mov, AMDGPU::OpName::src0_rel, 1);
1179 unsigned R600InstrInfo::getMaxAlusPerClause() const {
1183 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
1184 MachineBasicBlock::iterator I,
1188 unsigned Src1Reg) const {
1189 MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
1193 MIB.addImm(0) // $update_exec_mask
1194 .addImm(0); // $update_predicate
1196 MIB.addImm(1) // $write
1198 .addImm(0) // $dst_rel
1199 .addImm(0) // $dst_clamp
1200 .addReg(Src0Reg) // $src0
1201 .addImm(0) // $src0_neg
1202 .addImm(0) // $src0_rel
1203 .addImm(0) // $src0_abs
1204 .addImm(-1); // $src0_sel
1207 MIB.addReg(Src1Reg) // $src1
1208 .addImm(0) // $src1_neg
1209 .addImm(0) // $src1_rel
1210 .addImm(0) // $src1_abs
1211 .addImm(-1); // $src1_sel
1214 //XXX: The r600g finalizer expects this to be 1, once we've moved the
1215 //scheduling to the backend, we can change the default to 0.
1216 MIB.addImm(1) // $last
1217 .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
1218 .addImm(0) // $literal
1219 .addImm(0); // $bank_swizzle
1224 #define OPERAND_CASE(Label) \
1226 static const unsigned Ops[] = \
1236 static unsigned getSlotedOps(unsigned Op, unsigned Slot) {
1238 OPERAND_CASE(AMDGPU::OpName::update_exec_mask)
1239 OPERAND_CASE(AMDGPU::OpName::update_pred)
1240 OPERAND_CASE(AMDGPU::OpName::write)
1241 OPERAND_CASE(AMDGPU::OpName::omod)
1242 OPERAND_CASE(AMDGPU::OpName::dst_rel)
1243 OPERAND_CASE(AMDGPU::OpName::clamp)
1244 OPERAND_CASE(AMDGPU::OpName::src0)
1245 OPERAND_CASE(AMDGPU::OpName::src0_neg)
1246 OPERAND_CASE(AMDGPU::OpName::src0_rel)
1247 OPERAND_CASE(AMDGPU::OpName::src0_abs)
1248 OPERAND_CASE(AMDGPU::OpName::src0_sel)
1249 OPERAND_CASE(AMDGPU::OpName::src1)
1250 OPERAND_CASE(AMDGPU::OpName::src1_neg)
1251 OPERAND_CASE(AMDGPU::OpName::src1_rel)
1252 OPERAND_CASE(AMDGPU::OpName::src1_abs)
1253 OPERAND_CASE(AMDGPU::OpName::src1_sel)
1254 OPERAND_CASE(AMDGPU::OpName::pred_sel)
1256 llvm_unreachable("Wrong Operand");
1262 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
1263 MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg)
1265 assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
1267 if (ST.getGeneration() <= AMDGPUSubtarget::R700)
1268 Opcode = AMDGPU::DOT4_r600;
1270 Opcode = AMDGPU::DOT4_eg;
1271 MachineBasicBlock::iterator I = MI;
1272 MachineOperand &Src0 = MI->getOperand(
1273 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src0, Slot)));
1274 MachineOperand &Src1 = MI->getOperand(
1275 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src1, Slot)));
1276 MachineInstr *MIB = buildDefaultInstruction(
1277 MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg());
1278 static const unsigned Operands[14] = {
1279 AMDGPU::OpName::update_exec_mask,
1280 AMDGPU::OpName::update_pred,
1281 AMDGPU::OpName::write,
1282 AMDGPU::OpName::omod,
1283 AMDGPU::OpName::dst_rel,
1284 AMDGPU::OpName::clamp,
1285 AMDGPU::OpName::src0_neg,
1286 AMDGPU::OpName::src0_rel,
1287 AMDGPU::OpName::src0_abs,
1288 AMDGPU::OpName::src0_sel,
1289 AMDGPU::OpName::src1_neg,
1290 AMDGPU::OpName::src1_rel,
1291 AMDGPU::OpName::src1_abs,
1292 AMDGPU::OpName::src1_sel,
1295 MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
1296 getSlotedOps(AMDGPU::OpName::pred_sel, Slot)));
1297 MIB->getOperand(getOperandIdx(Opcode, AMDGPU::OpName::pred_sel))
1298 .setReg(MO.getReg());
1300 for (unsigned i = 0; i < 14; i++) {
1301 MachineOperand &MO = MI->getOperand(
1302 getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));
1303 assert (MO.isImm());
1304 setImmOperand(MIB, Operands[i], MO.getImm());
1306 MIB->getOperand(20).setImm(0);
1310 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
1311 MachineBasicBlock::iterator I,
1313 uint64_t Imm) const {
1314 MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
1315 AMDGPU::ALU_LITERAL_X);
1316 setImmOperand(MovImm, AMDGPU::OpName::literal, Imm);
1320 MachineInstr *R600InstrInfo::buildMovInstr(MachineBasicBlock *MBB,
1321 MachineBasicBlock::iterator I,
1322 unsigned DstReg, unsigned SrcReg) const {
1323 return buildDefaultInstruction(*MBB, I, AMDGPU::MOV, DstReg, SrcReg);
1326 int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const {
1327 return getOperandIdx(MI.getOpcode(), Op);
1330 int R600InstrInfo::getOperandIdx(unsigned Opcode, unsigned Op) const {
1331 return AMDGPU::getNamedOperandIdx(Opcode, Op);
1334 void R600InstrInfo::setImmOperand(MachineInstr *MI, unsigned Op,
1335 int64_t Imm) const {
1336 int Idx = getOperandIdx(*MI, Op);
1337 assert(Idx != -1 && "Operand not supported for this instruction.");
1338 assert(MI->getOperand(Idx).isImm());
1339 MI->getOperand(Idx).setImm(Imm);
1342 //===----------------------------------------------------------------------===//
1343 // Instruction flag getters/setters
1344 //===----------------------------------------------------------------------===//
1346 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
1347 return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
1350 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
1351 unsigned Flag) const {
1352 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1355 // If we pass something other than the default value of Flag to this
1356 // function, it means we are want to set a flag on an instruction
1357 // that uses native encoding.
1358 assert(HAS_NATIVE_OPERANDS(TargetFlags));
1359 bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
1362 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::clamp);
1365 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::write);
1367 case MO_FLAG_NOT_LAST:
1369 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::last);
1373 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_neg); break;
1374 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_neg); break;
1375 case 2: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src2_neg); break;
1380 assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
1384 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_abs); break;
1385 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_abs); break;
1393 assert(FlagIndex != -1 && "Flag not supported for this instruction");
1395 FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
1396 assert(FlagIndex != 0 &&
1397 "Instruction flags not supported for this instruction");
1400 MachineOperand &FlagOp = MI->getOperand(FlagIndex);
1401 assert(FlagOp.isImm());
1405 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
1406 unsigned Flag) const {
1407 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1411 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1412 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1413 if (Flag == MO_FLAG_NOT_LAST) {
1414 clearFlag(MI, Operand, MO_FLAG_LAST);
1415 } else if (Flag == MO_FLAG_MASK) {
1416 clearFlag(MI, Operand, Flag);
1421 MachineOperand &FlagOp = getFlagOp(MI, Operand);
1422 FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
1426 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
1427 unsigned Flag) const {
1428 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1429 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1430 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1433 MachineOperand &FlagOp = getFlagOp(MI);
1434 unsigned InstFlags = FlagOp.getImm();
1435 InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
1436 FlagOp.setImm(InstFlags);