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 if (AMDGPU::R600_Reg128RegClass.contains(DestReg)
55 && AMDGPU::R600_Reg128RegClass.contains(SrcReg)) {
56 for (unsigned I = 0; I < 4; I++) {
57 unsigned SubRegIndex = RI.getSubRegFromChannel(I);
58 buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
59 RI.getSubReg(DestReg, SubRegIndex),
60 RI.getSubReg(SrcReg, SubRegIndex))
62 RegState::Define | RegState::Implicit);
66 // We can't copy vec4 registers
67 assert(!AMDGPU::R600_Reg128RegClass.contains(DestReg)
68 && !AMDGPU::R600_Reg128RegClass.contains(SrcReg));
70 MachineInstr *NewMI = buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
72 NewMI->getOperand(getOperandIdx(*NewMI, R600Operands::SRC0))
77 MachineInstr * R600InstrInfo::getMovImmInstr(MachineFunction *MF,
78 unsigned DstReg, int64_t Imm) const {
79 MachineInstr * MI = MF->CreateMachineInstr(get(AMDGPU::MOV), DebugLoc());
80 MachineInstrBuilder MIB(*MF, MI);
81 MIB.addReg(DstReg, RegState::Define);
82 MIB.addReg(AMDGPU::ALU_LITERAL_X);
84 MIB.addReg(0); // PREDICATE_BIT
89 unsigned R600InstrInfo::getIEQOpcode() const {
90 return AMDGPU::SETE_INT;
93 bool R600InstrInfo::isMov(unsigned Opcode) const {
97 default: return false;
99 case AMDGPU::MOV_IMM_F32:
100 case AMDGPU::MOV_IMM_I32:
105 // Some instructions act as place holders to emulate operations that the GPU
106 // hardware does automatically. This function can be used to check if
107 // an opcode falls into this category.
108 bool R600InstrInfo::isPlaceHolderOpcode(unsigned Opcode) const {
110 default: return false;
116 bool R600InstrInfo::isReductionOp(unsigned Opcode) const {
118 default: return false;
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::OP1) |
137 (TargetFlags & R600_InstFlag::OP2) |
138 (TargetFlags & R600_InstFlag::OP3));
141 bool R600InstrInfo::isTransOnly(unsigned Opcode) const {
142 return (get(Opcode).TSFlags & R600_InstFlag::TRANS_ONLY);
145 bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const {
146 return isTransOnly(MI->getOpcode());
149 bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
150 return ST.hasVertexCache() && IS_VTX(get(Opcode));
153 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const {
154 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
155 return MFI->ShaderType != ShaderType::COMPUTE && usesVertexCache(MI->getOpcode());
158 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
159 return (!ST.hasVertexCache() && IS_VTX(get(Opcode))) || IS_TEX(get(Opcode));
162 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
163 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
164 return (MFI->ShaderType == ShaderType::COMPUTE && usesVertexCache(MI->getOpcode())) ||
165 usesTextureCache(MI->getOpcode());
168 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
169 R600InstrInfo::getSrcs(MachineInstr *MI) const {
170 SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
172 if (MI->getOpcode() == AMDGPU::DOT_4) {
173 static const R600Operands::VecOps OpTable[8][2] = {
174 {R600Operands::SRC0_X, R600Operands::SRC0_SEL_X},
175 {R600Operands::SRC0_Y, R600Operands::SRC0_SEL_Y},
176 {R600Operands::SRC0_Z, R600Operands::SRC0_SEL_Z},
177 {R600Operands::SRC0_W, R600Operands::SRC0_SEL_W},
178 {R600Operands::SRC1_X, R600Operands::SRC1_SEL_X},
179 {R600Operands::SRC1_Y, R600Operands::SRC1_SEL_Y},
180 {R600Operands::SRC1_Z, R600Operands::SRC1_SEL_Z},
181 {R600Operands::SRC1_W, R600Operands::SRC1_SEL_W},
184 for (unsigned j = 0; j < 8; j++) {
185 MachineOperand &MO = MI->getOperand(OpTable[j][0] + 1);
186 unsigned Reg = MO.getReg();
187 if (Reg == AMDGPU::ALU_CONST) {
188 unsigned Sel = MI->getOperand(OpTable[j][1] + 1).getImm();
189 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
197 static const R600Operands::Ops OpTable[3][2] = {
198 {R600Operands::SRC0, R600Operands::SRC0_SEL},
199 {R600Operands::SRC1, R600Operands::SRC1_SEL},
200 {R600Operands::SRC2, R600Operands::SRC2_SEL},
203 for (unsigned j = 0; j < 3; j++) {
204 int SrcIdx = getOperandIdx(MI->getOpcode(), OpTable[j][0]);
207 MachineOperand &MO = MI->getOperand(SrcIdx);
208 unsigned Reg = MI->getOperand(SrcIdx).getReg();
209 if (Reg == AMDGPU::ALU_CONST) {
210 unsigned Sel = MI->getOperand(
211 getOperandIdx(MI->getOpcode(), OpTable[j][1])).getImm();
212 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
215 if (Reg == AMDGPU::ALU_LITERAL_X) {
216 unsigned Imm = MI->getOperand(
217 getOperandIdx(MI->getOpcode(), R600Operands::IMM)).getImm();
218 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Imm));
221 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, 0));
226 std::vector<std::pair<int, unsigned> >
227 R600InstrInfo::ExtractSrcs(MachineInstr *MI,
228 const DenseMap<unsigned, unsigned> &PV)
230 const SmallVector<std::pair<MachineOperand *, int64_t>, 3> Srcs = getSrcs(MI);
231 const std::pair<int, unsigned> DummyPair(-1, 0);
232 std::vector<std::pair<int, unsigned> > Result;
234 for (unsigned n = Srcs.size(); i < n; ++i) {
235 unsigned Reg = Srcs[i].first->getReg();
236 unsigned Index = RI.getEncodingValue(Reg) & 0xff;
237 unsigned Chan = RI.getHWRegChan(Reg);
239 Result.push_back(DummyPair);
242 if (PV.find(Reg) != PV.end()) {
243 Result.push_back(DummyPair);
246 Result.push_back(std::pair<int, unsigned>(Index, Chan));
249 Result.push_back(DummyPair);
253 static std::vector<std::pair<int, unsigned> >
254 Swizzle(std::vector<std::pair<int, unsigned> > Src,
255 R600InstrInfo::BankSwizzle Swz) {
257 case R600InstrInfo::ALU_VEC_012:
259 case R600InstrInfo::ALU_VEC_021:
260 std::swap(Src[1], Src[2]);
262 case R600InstrInfo::ALU_VEC_102:
263 std::swap(Src[0], Src[1]);
265 case R600InstrInfo::ALU_VEC_120:
266 std::swap(Src[0], Src[1]);
267 std::swap(Src[0], Src[2]);
269 case R600InstrInfo::ALU_VEC_201:
270 std::swap(Src[0], Src[2]);
271 std::swap(Src[0], Src[1]);
273 case R600InstrInfo::ALU_VEC_210:
274 std::swap(Src[0], Src[2]);
281 isLegal(const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
282 const std::vector<R600InstrInfo::BankSwizzle> &Swz,
283 unsigned CheckedSize) {
285 memset(Vector, -1, sizeof(Vector));
286 for (unsigned i = 0; i < CheckedSize; i++) {
287 const std::vector<std::pair<int, unsigned> > &Srcs =
288 Swizzle(IGSrcs[i], Swz[i]);
289 for (unsigned j = 0; j < 3; j++) {
290 const std::pair<int, unsigned> &Src = Srcs[j];
293 if (Vector[Src.second][j] < 0)
294 Vector[Src.second][j] = Src.first;
295 if (Vector[Src.second][j] != Src.first)
302 static bool recursiveFitsFPLimitation(
303 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
304 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
305 unsigned Depth = 0) {
306 if (!isLegal(IGSrcs, SwzCandidate, Depth))
308 if (IGSrcs.size() == Depth)
310 unsigned i = SwzCandidate[Depth];
312 SwzCandidate[Depth] = (R600InstrInfo::BankSwizzle) i;
313 if (recursiveFitsFPLimitation(IGSrcs, SwzCandidate, Depth + 1))
316 SwzCandidate[Depth] = R600InstrInfo::ALU_VEC_012;
321 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG,
322 const DenseMap<unsigned, unsigned> &PV,
323 std::vector<BankSwizzle> &ValidSwizzle)
325 //Todo : support shared src0 - src1 operand
327 std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs;
328 ValidSwizzle.clear();
329 for (unsigned i = 0, e = IG.size(); i < e; ++i) {
330 IGSrcs.push_back(ExtractSrcs(IG[i], PV));
331 unsigned Op = getOperandIdx(IG[i]->getOpcode(),
332 R600Operands::BANK_SWIZZLE);
333 ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle)
334 IG[i]->getOperand(Op).getImm());
336 bool Result = recursiveFitsFPLimitation(IGSrcs, ValidSwizzle);
344 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts)
346 assert (Consts.size() <= 12 && "Too many operands in instructions group");
347 unsigned Pair1 = 0, Pair2 = 0;
348 for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
349 unsigned ReadConstHalf = Consts[i] & 2;
350 unsigned ReadConstIndex = Consts[i] & (~3);
351 unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf;
353 Pair1 = ReadHalfConst;
356 if (Pair1 == ReadHalfConst)
359 Pair2 = ReadHalfConst;
362 if (Pair2 != ReadHalfConst)
369 R600InstrInfo::canBundle(const std::vector<MachineInstr *> &MIs) const {
370 std::vector<unsigned> Consts;
371 for (unsigned i = 0, n = MIs.size(); i < n; i++) {
372 MachineInstr *MI = MIs[i];
373 if (!isALUInstr(MI->getOpcode()))
376 const SmallVector<std::pair<MachineOperand *, int64_t>, 3> &Srcs =
379 for (unsigned j = 0, e = Srcs.size(); j < e; j++) {
380 std::pair<MachineOperand *, unsigned> Src = Srcs[j];
381 if (Src.first->getReg() == AMDGPU::ALU_CONST)
382 Consts.push_back(Src.second);
383 if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) ||
384 AMDGPU::R600_KC1RegClass.contains(Src.first->getReg())) {
385 unsigned Index = RI.getEncodingValue(Src.first->getReg()) & 0xff;
386 unsigned Chan = RI.getHWRegChan(Src.first->getReg());
387 Consts.push_back((Index << 2) | Chan);
391 return fitsConstReadLimitations(Consts);
394 DFAPacketizer *R600InstrInfo::CreateTargetScheduleState(const TargetMachine *TM,
395 const ScheduleDAG *DAG) const {
396 const InstrItineraryData *II = TM->getInstrItineraryData();
397 return TM->getSubtarget<AMDGPUSubtarget>().createDFAPacketizer(II);
401 isPredicateSetter(unsigned Opcode) {
410 static MachineInstr *
411 findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
412 MachineBasicBlock::iterator I) {
413 while (I != MBB.begin()) {
415 MachineInstr *MI = I;
416 if (isPredicateSetter(MI->getOpcode()))
424 bool isJump(unsigned Opcode) {
425 return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
429 R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
430 MachineBasicBlock *&TBB,
431 MachineBasicBlock *&FBB,
432 SmallVectorImpl<MachineOperand> &Cond,
433 bool AllowModify) const {
434 // Most of the following comes from the ARM implementation of AnalyzeBranch
436 // If the block has no terminators, it just falls into the block after it.
437 MachineBasicBlock::iterator I = MBB.end();
438 if (I == MBB.begin())
441 while (I->isDebugValue()) {
442 if (I == MBB.begin())
446 if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) {
450 // Get the last instruction in the block.
451 MachineInstr *LastInst = I;
453 // If there is only one terminator instruction, process it.
454 unsigned LastOpc = LastInst->getOpcode();
455 if (I == MBB.begin() ||
456 !isJump(static_cast<MachineInstr *>(--I)->getOpcode())) {
457 if (LastOpc == AMDGPU::JUMP) {
458 TBB = LastInst->getOperand(0).getMBB();
460 } else if (LastOpc == AMDGPU::JUMP_COND) {
461 MachineInstr *predSet = I;
462 while (!isPredicateSetter(predSet->getOpcode())) {
465 TBB = LastInst->getOperand(0).getMBB();
466 Cond.push_back(predSet->getOperand(1));
467 Cond.push_back(predSet->getOperand(2));
468 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
471 return true; // Can't handle indirect branch.
474 // Get the instruction before it if it is a terminator.
475 MachineInstr *SecondLastInst = I;
476 unsigned SecondLastOpc = SecondLastInst->getOpcode();
478 // If the block ends with a B and a Bcc, handle it.
479 if (SecondLastOpc == AMDGPU::JUMP_COND && LastOpc == AMDGPU::JUMP) {
480 MachineInstr *predSet = --I;
481 while (!isPredicateSetter(predSet->getOpcode())) {
484 TBB = SecondLastInst->getOperand(0).getMBB();
485 FBB = LastInst->getOperand(0).getMBB();
486 Cond.push_back(predSet->getOperand(1));
487 Cond.push_back(predSet->getOperand(2));
488 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
492 // Otherwise, can't handle this.
496 int R600InstrInfo::getBranchInstr(const MachineOperand &op) const {
497 const MachineInstr *MI = op.getParent();
499 switch (MI->getDesc().OpInfo->RegClass) {
500 default: // FIXME: fallthrough??
501 case AMDGPU::GPRI32RegClassID: return AMDGPU::BRANCH_COND_i32;
502 case AMDGPU::GPRF32RegClassID: return AMDGPU::BRANCH_COND_f32;
507 R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
508 MachineBasicBlock *TBB,
509 MachineBasicBlock *FBB,
510 const SmallVectorImpl<MachineOperand> &Cond,
512 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
516 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB);
519 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
520 assert(PredSet && "No previous predicate !");
521 addFlag(PredSet, 0, MO_FLAG_PUSH);
522 PredSet->getOperand(2).setImm(Cond[1].getImm());
524 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
526 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
530 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
531 assert(PredSet && "No previous predicate !");
532 addFlag(PredSet, 0, MO_FLAG_PUSH);
533 PredSet->getOperand(2).setImm(Cond[1].getImm());
534 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
536 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
537 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB);
543 R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
545 // Note : we leave PRED* instructions there.
546 // They may be needed when predicating instructions.
548 MachineBasicBlock::iterator I = MBB.end();
550 if (I == MBB.begin()) {
554 switch (I->getOpcode()) {
557 case AMDGPU::JUMP_COND: {
558 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
559 clearFlag(predSet, 0, MO_FLAG_PUSH);
560 I->eraseFromParent();
564 I->eraseFromParent();
569 if (I == MBB.begin()) {
573 switch (I->getOpcode()) {
574 // FIXME: only one case??
577 case AMDGPU::JUMP_COND: {
578 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
579 clearFlag(predSet, 0, MO_FLAG_PUSH);
580 I->eraseFromParent();
584 I->eraseFromParent();
591 R600InstrInfo::isPredicated(const MachineInstr *MI) const {
592 int idx = MI->findFirstPredOperandIdx();
596 unsigned Reg = MI->getOperand(idx).getReg();
598 default: return false;
599 case AMDGPU::PRED_SEL_ONE:
600 case AMDGPU::PRED_SEL_ZERO:
601 case AMDGPU::PREDICATE_BIT:
607 R600InstrInfo::isPredicable(MachineInstr *MI) const {
608 // XXX: KILL* instructions can be predicated, but they must be the last
609 // instruction in a clause, so this means any instructions after them cannot
610 // be predicated. Until we have proper support for instruction clauses in the
611 // backend, we will mark KILL* instructions as unpredicable.
613 if (MI->getOpcode() == AMDGPU::KILLGT) {
615 } else if (isVector(*MI)) {
618 return AMDGPUInstrInfo::isPredicable(MI);
624 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
626 unsigned ExtraPredCycles,
627 const BranchProbability &Probability) const{
632 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
634 unsigned ExtraTCycles,
635 MachineBasicBlock &FMBB,
637 unsigned ExtraFCycles,
638 const BranchProbability &Probability) const {
643 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
645 const BranchProbability &Probability)
651 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
652 MachineBasicBlock &FMBB) const {
658 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
659 MachineOperand &MO = Cond[1];
660 switch (MO.getImm()) {
661 case OPCODE_IS_ZERO_INT:
662 MO.setImm(OPCODE_IS_NOT_ZERO_INT);
664 case OPCODE_IS_NOT_ZERO_INT:
665 MO.setImm(OPCODE_IS_ZERO_INT);
668 MO.setImm(OPCODE_IS_NOT_ZERO);
670 case OPCODE_IS_NOT_ZERO:
671 MO.setImm(OPCODE_IS_ZERO);
677 MachineOperand &MO2 = Cond[2];
678 switch (MO2.getReg()) {
679 case AMDGPU::PRED_SEL_ZERO:
680 MO2.setReg(AMDGPU::PRED_SEL_ONE);
682 case AMDGPU::PRED_SEL_ONE:
683 MO2.setReg(AMDGPU::PRED_SEL_ZERO);
692 R600InstrInfo::DefinesPredicate(MachineInstr *MI,
693 std::vector<MachineOperand> &Pred) const {
694 return isPredicateSetter(MI->getOpcode());
699 R600InstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
700 const SmallVectorImpl<MachineOperand> &Pred2) const {
706 R600InstrInfo::PredicateInstruction(MachineInstr *MI,
707 const SmallVectorImpl<MachineOperand> &Pred) const {
708 int PIdx = MI->findFirstPredOperandIdx();
711 MachineOperand &PMO = MI->getOperand(PIdx);
712 PMO.setReg(Pred[2].getReg());
713 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
714 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
721 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
722 const MachineInstr *MI,
723 unsigned *PredCost) const {
729 int R600InstrInfo::getIndirectIndexBegin(const MachineFunction &MF) const {
730 const MachineRegisterInfo &MRI = MF.getRegInfo();
731 const MachineFrameInfo *MFI = MF.getFrameInfo();
734 if (MFI->getNumObjects() == 0) {
738 if (MRI.livein_empty()) {
742 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
743 LE = MRI.livein_end();
745 Offset = std::max(Offset,
746 GET_REG_INDEX(RI.getEncodingValue(LI->first)));
752 int R600InstrInfo::getIndirectIndexEnd(const MachineFunction &MF) const {
754 const MachineFrameInfo *MFI = MF.getFrameInfo();
756 // Variable sized objects are not supported
757 assert(!MFI->hasVarSizedObjects());
759 if (MFI->getNumObjects() == 0) {
763 Offset = TM.getFrameLowering()->getFrameIndexOffset(MF, -1);
765 return getIndirectIndexBegin(MF) + Offset;
768 std::vector<unsigned> R600InstrInfo::getIndirectReservedRegs(
769 const MachineFunction &MF) const {
770 const AMDGPUFrameLowering *TFL =
771 static_cast<const AMDGPUFrameLowering*>(TM.getFrameLowering());
772 std::vector<unsigned> Regs;
774 unsigned StackWidth = TFL->getStackWidth(MF);
775 int End = getIndirectIndexEnd(MF);
781 for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
782 unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
783 Regs.push_back(SuperReg);
784 for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
785 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
792 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
793 unsigned Channel) const {
794 // XXX: Remove when we support a stack width > 2
795 assert(Channel == 0);
799 const TargetRegisterClass * R600InstrInfo::getIndirectAddrStoreRegClass(
800 unsigned SourceReg) const {
801 return &AMDGPU::R600_TReg32RegClass;
804 const TargetRegisterClass *R600InstrInfo::getIndirectAddrLoadRegClass() const {
805 return &AMDGPU::TRegMemRegClass;
808 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
809 MachineBasicBlock::iterator I,
810 unsigned ValueReg, unsigned Address,
811 unsigned OffsetReg) const {
812 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
813 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
814 AMDGPU::AR_X, OffsetReg);
815 setImmOperand(MOVA, R600Operands::WRITE, 0);
817 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
819 .addReg(AMDGPU::AR_X,
820 RegState::Implicit | RegState::Kill);
821 setImmOperand(Mov, R600Operands::DST_REL, 1);
825 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
826 MachineBasicBlock::iterator I,
827 unsigned ValueReg, unsigned Address,
828 unsigned OffsetReg) const {
829 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
830 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
833 setImmOperand(MOVA, R600Operands::WRITE, 0);
834 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
837 .addReg(AMDGPU::AR_X,
838 RegState::Implicit | RegState::Kill);
839 setImmOperand(Mov, R600Operands::SRC0_REL, 1);
844 const TargetRegisterClass *R600InstrInfo::getSuperIndirectRegClass() const {
845 return &AMDGPU::IndirectRegRegClass;
848 unsigned R600InstrInfo::getMaxAlusPerClause() const {
852 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
853 MachineBasicBlock::iterator I,
857 unsigned Src1Reg) const {
858 MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
862 MIB.addImm(0) // $update_exec_mask
863 .addImm(0); // $update_predicate
865 MIB.addImm(1) // $write
867 .addImm(0) // $dst_rel
868 .addImm(0) // $dst_clamp
869 .addReg(Src0Reg) // $src0
870 .addImm(0) // $src0_neg
871 .addImm(0) // $src0_rel
872 .addImm(0) // $src0_abs
873 .addImm(-1); // $src0_sel
876 MIB.addReg(Src1Reg) // $src1
877 .addImm(0) // $src1_neg
878 .addImm(0) // $src1_rel
879 .addImm(0) // $src1_abs
880 .addImm(-1); // $src1_sel
883 //XXX: The r600g finalizer expects this to be 1, once we've moved the
884 //scheduling to the backend, we can change the default to 0.
885 MIB.addImm(1) // $last
886 .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
887 .addImm(0) // $literal
888 .addImm(0); // $bank_swizzle
893 #define OPERAND_CASE(Label) \
895 static const R600Operands::VecOps Ops[] = \
905 static R600Operands::VecOps
906 getSlotedOps(R600Operands::Ops Op, unsigned Slot) {
908 OPERAND_CASE(R600Operands::UPDATE_EXEC_MASK)
909 OPERAND_CASE(R600Operands::UPDATE_PREDICATE)
910 OPERAND_CASE(R600Operands::WRITE)
911 OPERAND_CASE(R600Operands::OMOD)
912 OPERAND_CASE(R600Operands::DST_REL)
913 OPERAND_CASE(R600Operands::CLAMP)
914 OPERAND_CASE(R600Operands::SRC0)
915 OPERAND_CASE(R600Operands::SRC0_NEG)
916 OPERAND_CASE(R600Operands::SRC0_REL)
917 OPERAND_CASE(R600Operands::SRC0_ABS)
918 OPERAND_CASE(R600Operands::SRC0_SEL)
919 OPERAND_CASE(R600Operands::SRC1)
920 OPERAND_CASE(R600Operands::SRC1_NEG)
921 OPERAND_CASE(R600Operands::SRC1_REL)
922 OPERAND_CASE(R600Operands::SRC1_ABS)
923 OPERAND_CASE(R600Operands::SRC1_SEL)
924 OPERAND_CASE(R600Operands::PRED_SEL)
926 llvm_unreachable("Wrong Operand");
933 getVecOperandIdx(R600Operands::VecOps Op) {
938 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
939 MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg)
941 assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
943 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
944 if (ST.getGeneration() <= AMDGPUSubtarget::R700)
945 Opcode = AMDGPU::DOT4_r600;
947 Opcode = AMDGPU::DOT4_eg;
948 MachineBasicBlock::iterator I = MI;
949 MachineOperand &Src0 = MI->getOperand(
950 getVecOperandIdx(getSlotedOps(R600Operands::SRC0, Slot)));
951 MachineOperand &Src1 = MI->getOperand(
952 getVecOperandIdx(getSlotedOps(R600Operands::SRC1, Slot)));
953 MachineInstr *MIB = buildDefaultInstruction(
954 MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg());
955 static const R600Operands::Ops Operands[14] = {
956 R600Operands::UPDATE_EXEC_MASK,
957 R600Operands::UPDATE_PREDICATE,
960 R600Operands::DST_REL,
962 R600Operands::SRC0_NEG,
963 R600Operands::SRC0_REL,
964 R600Operands::SRC0_ABS,
965 R600Operands::SRC0_SEL,
966 R600Operands::SRC1_NEG,
967 R600Operands::SRC1_REL,
968 R600Operands::SRC1_ABS,
969 R600Operands::SRC1_SEL,
972 for (unsigned i = 0; i < 14; i++) {
973 MachineOperand &MO = MI->getOperand(
974 getVecOperandIdx(getSlotedOps(Operands[i], Slot)));
976 setImmOperand(MIB, Operands[i], MO.getImm());
978 MIB->getOperand(20).setImm(0);
982 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
983 MachineBasicBlock::iterator I,
985 uint64_t Imm) const {
986 MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
987 AMDGPU::ALU_LITERAL_X);
988 setImmOperand(MovImm, R600Operands::IMM, Imm);
992 int R600InstrInfo::getOperandIdx(const MachineInstr &MI,
993 R600Operands::Ops Op) const {
994 return getOperandIdx(MI.getOpcode(), Op);
997 int R600InstrInfo::getOperandIdx(const MachineInstr &MI,
998 R600Operands::VecOps Op) const {
999 return getOperandIdx(MI.getOpcode(), Op);
1002 int R600InstrInfo::getOperandIdx(unsigned Opcode,
1003 R600Operands::Ops Op) const {
1004 unsigned TargetFlags = get(Opcode).TSFlags;
1005 unsigned OpTableIdx;
1007 if (!HAS_NATIVE_OPERANDS(TargetFlags)) {
1009 case R600Operands::DST: return 0;
1010 case R600Operands::SRC0: return 1;
1011 case R600Operands::SRC1: return 2;
1012 case R600Operands::SRC2: return 3;
1014 assert(!"Unknown operand type for instruction");
1019 if (TargetFlags & R600_InstFlag::OP1) {
1021 } else if (TargetFlags & R600_InstFlag::OP2) {
1024 assert((TargetFlags & R600_InstFlag::OP3) && "OP1, OP2, or OP3 not defined "
1025 "for this instruction");
1029 return R600Operands::ALUOpTable[OpTableIdx][Op];
1032 int R600InstrInfo::getOperandIdx(unsigned Opcode,
1033 R600Operands::VecOps Op) const {
1037 void R600InstrInfo::setImmOperand(MachineInstr *MI, R600Operands::Ops Op,
1038 int64_t Imm) const {
1039 int Idx = getOperandIdx(*MI, Op);
1040 assert(Idx != -1 && "Operand not supported for this instruction.");
1041 assert(MI->getOperand(Idx).isImm());
1042 MI->getOperand(Idx).setImm(Imm);
1045 //===----------------------------------------------------------------------===//
1046 // Instruction flag getters/setters
1047 //===----------------------------------------------------------------------===//
1049 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
1050 return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
1053 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
1054 unsigned Flag) const {
1055 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1058 // If we pass something other than the default value of Flag to this
1059 // function, it means we are want to set a flag on an instruction
1060 // that uses native encoding.
1061 assert(HAS_NATIVE_OPERANDS(TargetFlags));
1062 bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
1065 FlagIndex = getOperandIdx(*MI, R600Operands::CLAMP);
1068 FlagIndex = getOperandIdx(*MI, R600Operands::WRITE);
1070 case MO_FLAG_NOT_LAST:
1072 FlagIndex = getOperandIdx(*MI, R600Operands::LAST);
1076 case 0: FlagIndex = getOperandIdx(*MI, R600Operands::SRC0_NEG); break;
1077 case 1: FlagIndex = getOperandIdx(*MI, R600Operands::SRC1_NEG); break;
1078 case 2: FlagIndex = getOperandIdx(*MI, R600Operands::SRC2_NEG); break;
1083 assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
1087 case 0: FlagIndex = getOperandIdx(*MI, R600Operands::SRC0_ABS); break;
1088 case 1: FlagIndex = getOperandIdx(*MI, R600Operands::SRC1_ABS); break;
1096 assert(FlagIndex != -1 && "Flag not supported for this instruction");
1098 FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
1099 assert(FlagIndex != 0 &&
1100 "Instruction flags not supported for this instruction");
1103 MachineOperand &FlagOp = MI->getOperand(FlagIndex);
1104 assert(FlagOp.isImm());
1108 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
1109 unsigned Flag) const {
1110 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1114 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1115 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1116 if (Flag == MO_FLAG_NOT_LAST) {
1117 clearFlag(MI, Operand, MO_FLAG_LAST);
1118 } else if (Flag == MO_FLAG_MASK) {
1119 clearFlag(MI, Operand, Flag);
1124 MachineOperand &FlagOp = getFlagOp(MI, Operand);
1125 FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
1129 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
1130 unsigned Flag) const {
1131 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1132 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1133 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1136 MachineOperand &FlagOp = getFlagOp(MI);
1137 unsigned InstFlags = FlagOp.getImm();
1138 InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
1139 FlagOp.setImm(InstFlags);