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, AMDGPU::OpName::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::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));
154 bool R600InstrInfo::isTransOnly(unsigned Opcode) const {
155 return (get(Opcode).TSFlags & R600_InstFlag::TRANS_ONLY);
158 bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const {
159 return isTransOnly(MI->getOpcode());
162 bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
163 return ST.hasVertexCache() && IS_VTX(get(Opcode));
166 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const {
167 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
168 return MFI->ShaderType != ShaderType::COMPUTE && usesVertexCache(MI->getOpcode());
171 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
172 return (!ST.hasVertexCache() && IS_VTX(get(Opcode))) || IS_TEX(get(Opcode));
175 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
176 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
177 return (MFI->ShaderType == ShaderType::COMPUTE && usesVertexCache(MI->getOpcode())) ||
178 usesTextureCache(MI->getOpcode());
181 bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
184 case AMDGPU::GROUP_BARRIER:
191 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
192 R600InstrInfo::getSrcs(MachineInstr *MI) const {
193 SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
195 if (MI->getOpcode() == AMDGPU::DOT_4) {
196 static const unsigned OpTable[8][2] = {
197 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
198 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
199 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
200 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
201 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
202 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
203 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
204 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W},
207 for (unsigned j = 0; j < 8; j++) {
208 MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
210 unsigned Reg = MO.getReg();
211 if (Reg == AMDGPU::ALU_CONST) {
212 unsigned Sel = MI->getOperand(getOperandIdx(MI->getOpcode(),
213 OpTable[j][1])).getImm();
214 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
222 static const unsigned OpTable[3][2] = {
223 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
224 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
225 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
228 for (unsigned j = 0; j < 3; j++) {
229 int SrcIdx = getOperandIdx(MI->getOpcode(), OpTable[j][0]);
232 MachineOperand &MO = MI->getOperand(SrcIdx);
233 unsigned Reg = MI->getOperand(SrcIdx).getReg();
234 if (Reg == AMDGPU::ALU_CONST) {
235 unsigned Sel = MI->getOperand(
236 getOperandIdx(MI->getOpcode(), OpTable[j][1])).getImm();
237 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
240 if (Reg == AMDGPU::ALU_LITERAL_X) {
241 unsigned Imm = MI->getOperand(
242 getOperandIdx(MI->getOpcode(), AMDGPU::OpName::literal)).getImm();
243 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Imm));
246 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, 0));
251 std::vector<std::pair<int, unsigned> >
252 R600InstrInfo::ExtractSrcs(MachineInstr *MI,
253 const DenseMap<unsigned, unsigned> &PV,
254 unsigned &ConstCount) const {
256 const SmallVector<std::pair<MachineOperand *, int64_t>, 3> Srcs = getSrcs(MI);
257 const std::pair<int, unsigned> DummyPair(-1, 0);
258 std::vector<std::pair<int, unsigned> > Result;
260 for (unsigned n = Srcs.size(); i < n; ++i) {
261 unsigned Reg = Srcs[i].first->getReg();
262 unsigned Index = RI.getEncodingValue(Reg) & 0xff;
263 if (Reg == AMDGPU::OQAP) {
264 Result.push_back(std::pair<int, unsigned>(Index, 0));
266 if (PV.find(Reg) != PV.end()) {
267 // 255 is used to tells its a PS/PV reg
268 Result.push_back(std::pair<int, unsigned>(255, 0));
273 Result.push_back(DummyPair);
276 unsigned Chan = RI.getHWRegChan(Reg);
277 Result.push_back(std::pair<int, unsigned>(Index, Chan));
280 Result.push_back(DummyPair);
284 static std::vector<std::pair<int, unsigned> >
285 Swizzle(std::vector<std::pair<int, unsigned> > Src,
286 R600InstrInfo::BankSwizzle Swz) {
288 case R600InstrInfo::ALU_VEC_012_SCL_210:
290 case R600InstrInfo::ALU_VEC_021_SCL_122:
291 std::swap(Src[1], Src[2]);
293 case R600InstrInfo::ALU_VEC_102_SCL_221:
294 std::swap(Src[0], Src[1]);
296 case R600InstrInfo::ALU_VEC_120_SCL_212:
297 std::swap(Src[0], Src[1]);
298 std::swap(Src[0], Src[2]);
300 case R600InstrInfo::ALU_VEC_201:
301 std::swap(Src[0], Src[2]);
302 std::swap(Src[0], Src[1]);
304 case R600InstrInfo::ALU_VEC_210:
305 std::swap(Src[0], Src[2]);
312 getTransSwizzle(R600InstrInfo::BankSwizzle Swz, unsigned Op) {
314 case R600InstrInfo::ALU_VEC_012_SCL_210: {
315 unsigned Cycles[3] = { 2, 1, 0};
318 case R600InstrInfo::ALU_VEC_021_SCL_122: {
319 unsigned Cycles[3] = { 1, 2, 2};
322 case R600InstrInfo::ALU_VEC_120_SCL_212: {
323 unsigned Cycles[3] = { 2, 1, 2};
326 case R600InstrInfo::ALU_VEC_102_SCL_221: {
327 unsigned Cycles[3] = { 2, 2, 1};
331 llvm_unreachable("Wrong Swizzle for Trans Slot");
336 /// returns how many MIs (whose inputs are represented by IGSrcs) can be packed
337 /// in the same Instruction Group while meeting read port limitations given a
338 /// Swz swizzle sequence.
339 unsigned R600InstrInfo::isLegalUpTo(
340 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
341 const std::vector<R600InstrInfo::BankSwizzle> &Swz,
342 const std::vector<std::pair<int, unsigned> > &TransSrcs,
343 R600InstrInfo::BankSwizzle TransSwz) const {
345 memset(Vector, -1, sizeof(Vector));
346 for (unsigned i = 0, e = IGSrcs.size(); i < e; i++) {
347 const std::vector<std::pair<int, unsigned> > &Srcs =
348 Swizzle(IGSrcs[i], Swz[i]);
349 for (unsigned j = 0; j < 3; j++) {
350 const std::pair<int, unsigned> &Src = Srcs[j];
351 if (Src.first < 0 || Src.first == 255)
353 if (Src.first == GET_REG_INDEX(RI.getEncodingValue(AMDGPU::OQAP))) {
354 if (Swz[i] != R600InstrInfo::ALU_VEC_012_SCL_210 &&
355 Swz[i] != R600InstrInfo::ALU_VEC_021_SCL_122) {
356 // The value from output queue A (denoted by register OQAP) can
357 // only be fetched during the first cycle.
360 // OQAP does not count towards the normal read port restrictions
363 if (Vector[Src.second][j] < 0)
364 Vector[Src.second][j] = Src.first;
365 if (Vector[Src.second][j] != Src.first)
369 // Now check Trans Alu
370 for (unsigned i = 0, e = TransSrcs.size(); i < e; ++i) {
371 const std::pair<int, unsigned> &Src = TransSrcs[i];
372 unsigned Cycle = getTransSwizzle(TransSwz, i);
375 if (Src.first == 255)
377 if (Vector[Src.second][Cycle] < 0)
378 Vector[Src.second][Cycle] = Src.first;
379 if (Vector[Src.second][Cycle] != Src.first)
380 return IGSrcs.size() - 1;
382 return IGSrcs.size();
385 /// Given a swizzle sequence SwzCandidate and an index Idx, returns the next
386 /// (in lexicographic term) swizzle sequence assuming that all swizzles after
387 /// Idx can be skipped
389 NextPossibleSolution(
390 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
392 assert(Idx < SwzCandidate.size());
394 while (ResetIdx > -1 && SwzCandidate[ResetIdx] == R600InstrInfo::ALU_VEC_210)
396 for (unsigned i = ResetIdx + 1, e = SwzCandidate.size(); i < e; i++) {
397 SwzCandidate[i] = R600InstrInfo::ALU_VEC_012_SCL_210;
401 int NextSwizzle = SwzCandidate[ResetIdx] + 1;
402 SwzCandidate[ResetIdx] = (R600InstrInfo::BankSwizzle)NextSwizzle;
406 /// Enumerate all possible Swizzle sequence to find one that can meet all
407 /// read port requirements.
408 bool R600InstrInfo::FindSwizzleForVectorSlot(
409 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
410 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
411 const std::vector<std::pair<int, unsigned> > &TransSrcs,
412 R600InstrInfo::BankSwizzle TransSwz) const {
413 unsigned ValidUpTo = 0;
415 ValidUpTo = isLegalUpTo(IGSrcs, SwzCandidate, TransSrcs, TransSwz);
416 if (ValidUpTo == IGSrcs.size())
418 } while (NextPossibleSolution(SwzCandidate, ValidUpTo));
422 /// Instructions in Trans slot can't read gpr at cycle 0 if they also read
423 /// a const, and can't read a gpr at cycle 1 if they read 2 const.
425 isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,
426 const std::vector<std::pair<int, unsigned> > &TransOps,
427 unsigned ConstCount) {
428 for (unsigned i = 0, e = TransOps.size(); i < e; ++i) {
429 const std::pair<int, unsigned> &Src = TransOps[i];
430 unsigned Cycle = getTransSwizzle(TransSwz, i);
433 if (ConstCount > 0 && Cycle == 0)
435 if (ConstCount > 1 && Cycle == 1)
442 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG,
443 const DenseMap<unsigned, unsigned> &PV,
444 std::vector<BankSwizzle> &ValidSwizzle,
447 //Todo : support shared src0 - src1 operand
449 std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs;
450 ValidSwizzle.clear();
452 BankSwizzle TransBS = ALU_VEC_012_SCL_210;
453 for (unsigned i = 0, e = IG.size(); i < e; ++i) {
454 IGSrcs.push_back(ExtractSrcs(IG[i], PV, ConstCount));
455 unsigned Op = getOperandIdx(IG[i]->getOpcode(),
456 AMDGPU::OpName::bank_swizzle);
457 ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle)
458 IG[i]->getOperand(Op).getImm());
460 std::vector<std::pair<int, unsigned> > TransOps;
462 return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS);
464 TransOps = IGSrcs.back();
466 ValidSwizzle.pop_back();
468 static const R600InstrInfo::BankSwizzle TransSwz[] = {
474 for (unsigned i = 0; i < 4; i++) {
475 TransBS = TransSwz[i];
476 if (!isConstCompatible(TransBS, TransOps, ConstCount))
478 bool Result = FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps,
481 ValidSwizzle.push_back(TransBS);
491 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts)
493 assert (Consts.size() <= 12 && "Too many operands in instructions group");
494 unsigned Pair1 = 0, Pair2 = 0;
495 for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
496 unsigned ReadConstHalf = Consts[i] & 2;
497 unsigned ReadConstIndex = Consts[i] & (~3);
498 unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf;
500 Pair1 = ReadHalfConst;
503 if (Pair1 == ReadHalfConst)
506 Pair2 = ReadHalfConst;
509 if (Pair2 != ReadHalfConst)
516 R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
518 std::vector<unsigned> Consts;
519 for (unsigned i = 0, n = MIs.size(); i < n; i++) {
520 MachineInstr *MI = MIs[i];
521 if (!isALUInstr(MI->getOpcode()))
524 const SmallVector<std::pair<MachineOperand *, int64_t>, 3> &Srcs =
527 for (unsigned j = 0, e = Srcs.size(); j < e; j++) {
528 std::pair<MachineOperand *, unsigned> Src = Srcs[j];
529 if (Src.first->getReg() == AMDGPU::ALU_CONST)
530 Consts.push_back(Src.second);
531 if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) ||
532 AMDGPU::R600_KC1RegClass.contains(Src.first->getReg())) {
533 unsigned Index = RI.getEncodingValue(Src.first->getReg()) & 0xff;
534 unsigned Chan = RI.getHWRegChan(Src.first->getReg());
535 Consts.push_back((Index << 2) | Chan);
539 return fitsConstReadLimitations(Consts);
542 DFAPacketizer *R600InstrInfo::CreateTargetScheduleState(const TargetMachine *TM,
543 const ScheduleDAG *DAG) const {
544 const InstrItineraryData *II = TM->getInstrItineraryData();
545 return TM->getSubtarget<AMDGPUSubtarget>().createDFAPacketizer(II);
549 isPredicateSetter(unsigned Opcode) {
558 static MachineInstr *
559 findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
560 MachineBasicBlock::iterator I) {
561 while (I != MBB.begin()) {
563 MachineInstr *MI = I;
564 if (isPredicateSetter(MI->getOpcode()))
572 bool isJump(unsigned Opcode) {
573 return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
577 R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
578 MachineBasicBlock *&TBB,
579 MachineBasicBlock *&FBB,
580 SmallVectorImpl<MachineOperand> &Cond,
581 bool AllowModify) const {
582 // Most of the following comes from the ARM implementation of AnalyzeBranch
584 // If the block has no terminators, it just falls into the block after it.
585 MachineBasicBlock::iterator I = MBB.end();
586 if (I == MBB.begin())
589 while (I->isDebugValue()) {
590 if (I == MBB.begin())
594 if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) {
598 // Get the last instruction in the block.
599 MachineInstr *LastInst = I;
601 // If there is only one terminator instruction, process it.
602 unsigned LastOpc = LastInst->getOpcode();
603 if (I == MBB.begin() ||
604 !isJump(static_cast<MachineInstr *>(--I)->getOpcode())) {
605 if (LastOpc == AMDGPU::JUMP) {
606 TBB = LastInst->getOperand(0).getMBB();
608 } else if (LastOpc == AMDGPU::JUMP_COND) {
609 MachineInstr *predSet = I;
610 while (!isPredicateSetter(predSet->getOpcode())) {
613 TBB = LastInst->getOperand(0).getMBB();
614 Cond.push_back(predSet->getOperand(1));
615 Cond.push_back(predSet->getOperand(2));
616 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
619 return true; // Can't handle indirect branch.
622 // Get the instruction before it if it is a terminator.
623 MachineInstr *SecondLastInst = I;
624 unsigned SecondLastOpc = SecondLastInst->getOpcode();
626 // If the block ends with a B and a Bcc, handle it.
627 if (SecondLastOpc == AMDGPU::JUMP_COND && LastOpc == AMDGPU::JUMP) {
628 MachineInstr *predSet = --I;
629 while (!isPredicateSetter(predSet->getOpcode())) {
632 TBB = SecondLastInst->getOperand(0).getMBB();
633 FBB = LastInst->getOperand(0).getMBB();
634 Cond.push_back(predSet->getOperand(1));
635 Cond.push_back(predSet->getOperand(2));
636 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
640 // Otherwise, can't handle this.
644 int R600InstrInfo::getBranchInstr(const MachineOperand &op) const {
645 const MachineInstr *MI = op.getParent();
647 switch (MI->getDesc().OpInfo->RegClass) {
648 default: // FIXME: fallthrough??
649 case AMDGPU::GPRI32RegClassID: return AMDGPU::BRANCH_COND_i32;
650 case AMDGPU::GPRF32RegClassID: return AMDGPU::BRANCH_COND_f32;
655 MachineBasicBlock::iterator FindLastAluClause(MachineBasicBlock &MBB) {
656 for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend();
658 if (It->getOpcode() == AMDGPU::CF_ALU ||
659 It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE)
660 return llvm::prior(It.base());
666 R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
667 MachineBasicBlock *TBB,
668 MachineBasicBlock *FBB,
669 const SmallVectorImpl<MachineOperand> &Cond,
671 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
675 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB);
678 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
679 assert(PredSet && "No previous predicate !");
680 addFlag(PredSet, 0, MO_FLAG_PUSH);
681 PredSet->getOperand(2).setImm(Cond[1].getImm());
683 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
685 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
686 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
687 if (CfAlu == MBB.end())
689 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
690 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
694 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
695 assert(PredSet && "No previous predicate !");
696 addFlag(PredSet, 0, MO_FLAG_PUSH);
697 PredSet->getOperand(2).setImm(Cond[1].getImm());
698 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
700 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
701 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB);
702 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
703 if (CfAlu == MBB.end())
705 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
706 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
712 R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
714 // Note : we leave PRED* instructions there.
715 // They may be needed when predicating instructions.
717 MachineBasicBlock::iterator I = MBB.end();
719 if (I == MBB.begin()) {
723 switch (I->getOpcode()) {
726 case AMDGPU::JUMP_COND: {
727 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
728 clearFlag(predSet, 0, MO_FLAG_PUSH);
729 I->eraseFromParent();
730 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
731 if (CfAlu == MBB.end())
733 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
734 CfAlu->setDesc(get(AMDGPU::CF_ALU));
738 I->eraseFromParent();
743 if (I == MBB.begin()) {
747 switch (I->getOpcode()) {
748 // FIXME: only one case??
751 case AMDGPU::JUMP_COND: {
752 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
753 clearFlag(predSet, 0, MO_FLAG_PUSH);
754 I->eraseFromParent();
755 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
756 if (CfAlu == MBB.end())
758 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
759 CfAlu->setDesc(get(AMDGPU::CF_ALU));
763 I->eraseFromParent();
770 R600InstrInfo::isPredicated(const MachineInstr *MI) const {
771 int idx = MI->findFirstPredOperandIdx();
775 unsigned Reg = MI->getOperand(idx).getReg();
777 default: return false;
778 case AMDGPU::PRED_SEL_ONE:
779 case AMDGPU::PRED_SEL_ZERO:
780 case AMDGPU::PREDICATE_BIT:
786 R600InstrInfo::isPredicable(MachineInstr *MI) const {
787 // XXX: KILL* instructions can be predicated, but they must be the last
788 // instruction in a clause, so this means any instructions after them cannot
789 // be predicated. Until we have proper support for instruction clauses in the
790 // backend, we will mark KILL* instructions as unpredicable.
792 if (MI->getOpcode() == AMDGPU::KILLGT) {
794 } else if (MI->getOpcode() == AMDGPU::CF_ALU) {
795 // If the clause start in the middle of MBB then the MBB has more
796 // than a single clause, unable to predicate several clauses.
797 if (MI->getParent()->begin() != MachineBasicBlock::iterator(MI))
799 // TODO: We don't support KC merging atm
800 if (MI->getOperand(3).getImm() != 0 || MI->getOperand(4).getImm() != 0)
803 } else if (isVector(*MI)) {
806 return AMDGPUInstrInfo::isPredicable(MI);
812 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
814 unsigned ExtraPredCycles,
815 const BranchProbability &Probability) const{
820 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
822 unsigned ExtraTCycles,
823 MachineBasicBlock &FMBB,
825 unsigned ExtraFCycles,
826 const BranchProbability &Probability) const {
831 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
833 const BranchProbability &Probability)
839 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
840 MachineBasicBlock &FMBB) const {
846 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
847 MachineOperand &MO = Cond[1];
848 switch (MO.getImm()) {
849 case OPCODE_IS_ZERO_INT:
850 MO.setImm(OPCODE_IS_NOT_ZERO_INT);
852 case OPCODE_IS_NOT_ZERO_INT:
853 MO.setImm(OPCODE_IS_ZERO_INT);
856 MO.setImm(OPCODE_IS_NOT_ZERO);
858 case OPCODE_IS_NOT_ZERO:
859 MO.setImm(OPCODE_IS_ZERO);
865 MachineOperand &MO2 = Cond[2];
866 switch (MO2.getReg()) {
867 case AMDGPU::PRED_SEL_ZERO:
868 MO2.setReg(AMDGPU::PRED_SEL_ONE);
870 case AMDGPU::PRED_SEL_ONE:
871 MO2.setReg(AMDGPU::PRED_SEL_ZERO);
880 R600InstrInfo::DefinesPredicate(MachineInstr *MI,
881 std::vector<MachineOperand> &Pred) const {
882 return isPredicateSetter(MI->getOpcode());
887 R600InstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
888 const SmallVectorImpl<MachineOperand> &Pred2) const {
894 R600InstrInfo::PredicateInstruction(MachineInstr *MI,
895 const SmallVectorImpl<MachineOperand> &Pred) const {
896 int PIdx = MI->findFirstPredOperandIdx();
898 if (MI->getOpcode() == AMDGPU::CF_ALU) {
899 MI->getOperand(8).setImm(0);
904 MachineOperand &PMO = MI->getOperand(PIdx);
905 PMO.setReg(Pred[2].getReg());
906 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
907 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
914 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
915 const MachineInstr *MI,
916 unsigned *PredCost) const {
922 int R600InstrInfo::getIndirectIndexBegin(const MachineFunction &MF) const {
923 const MachineRegisterInfo &MRI = MF.getRegInfo();
924 const MachineFrameInfo *MFI = MF.getFrameInfo();
927 if (MFI->getNumObjects() == 0) {
931 if (MRI.livein_empty()) {
935 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
936 LE = MRI.livein_end();
938 Offset = std::max(Offset,
939 GET_REG_INDEX(RI.getEncodingValue(LI->first)));
945 int R600InstrInfo::getIndirectIndexEnd(const MachineFunction &MF) const {
947 const MachineFrameInfo *MFI = MF.getFrameInfo();
949 // Variable sized objects are not supported
950 assert(!MFI->hasVarSizedObjects());
952 if (MFI->getNumObjects() == 0) {
956 Offset = TM.getFrameLowering()->getFrameIndexOffset(MF, -1);
958 return getIndirectIndexBegin(MF) + Offset;
961 std::vector<unsigned> R600InstrInfo::getIndirectReservedRegs(
962 const MachineFunction &MF) const {
963 const AMDGPUFrameLowering *TFL =
964 static_cast<const AMDGPUFrameLowering*>(TM.getFrameLowering());
965 std::vector<unsigned> Regs;
967 unsigned StackWidth = TFL->getStackWidth(MF);
968 int End = getIndirectIndexEnd(MF);
974 for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
975 unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
976 Regs.push_back(SuperReg);
977 for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
978 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
985 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
986 unsigned Channel) const {
987 // XXX: Remove when we support a stack width > 2
988 assert(Channel == 0);
992 const TargetRegisterClass * R600InstrInfo::getIndirectAddrStoreRegClass(
993 unsigned SourceReg) const {
994 return &AMDGPU::R600_TReg32RegClass;
997 const TargetRegisterClass *R600InstrInfo::getIndirectAddrLoadRegClass() const {
998 return &AMDGPU::TRegMemRegClass;
1001 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1002 MachineBasicBlock::iterator I,
1003 unsigned ValueReg, unsigned Address,
1004 unsigned OffsetReg) const {
1005 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1006 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1007 AMDGPU::AR_X, OffsetReg);
1008 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1010 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1012 .addReg(AMDGPU::AR_X,
1013 RegState::Implicit | RegState::Kill);
1014 setImmOperand(Mov, AMDGPU::OpName::dst_rel, 1);
1018 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1019 MachineBasicBlock::iterator I,
1020 unsigned ValueReg, unsigned Address,
1021 unsigned OffsetReg) const {
1022 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1023 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1026 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1027 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1030 .addReg(AMDGPU::AR_X,
1031 RegState::Implicit | RegState::Kill);
1032 setImmOperand(Mov, AMDGPU::OpName::src0_rel, 1);
1037 const TargetRegisterClass *R600InstrInfo::getSuperIndirectRegClass() const {
1038 return &AMDGPU::IndirectRegRegClass;
1041 unsigned R600InstrInfo::getMaxAlusPerClause() const {
1045 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
1046 MachineBasicBlock::iterator I,
1050 unsigned Src1Reg) const {
1051 MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
1055 MIB.addImm(0) // $update_exec_mask
1056 .addImm(0); // $update_predicate
1058 MIB.addImm(1) // $write
1060 .addImm(0) // $dst_rel
1061 .addImm(0) // $dst_clamp
1062 .addReg(Src0Reg) // $src0
1063 .addImm(0) // $src0_neg
1064 .addImm(0) // $src0_rel
1065 .addImm(0) // $src0_abs
1066 .addImm(-1); // $src0_sel
1069 MIB.addReg(Src1Reg) // $src1
1070 .addImm(0) // $src1_neg
1071 .addImm(0) // $src1_rel
1072 .addImm(0) // $src1_abs
1073 .addImm(-1); // $src1_sel
1076 //XXX: The r600g finalizer expects this to be 1, once we've moved the
1077 //scheduling to the backend, we can change the default to 0.
1078 MIB.addImm(1) // $last
1079 .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
1080 .addImm(0) // $literal
1081 .addImm(0); // $bank_swizzle
1086 #define OPERAND_CASE(Label) \
1088 static const unsigned Ops[] = \
1098 static unsigned getSlotedOps(unsigned Op, unsigned Slot) {
1100 OPERAND_CASE(AMDGPU::OpName::update_exec_mask)
1101 OPERAND_CASE(AMDGPU::OpName::update_pred)
1102 OPERAND_CASE(AMDGPU::OpName::write)
1103 OPERAND_CASE(AMDGPU::OpName::omod)
1104 OPERAND_CASE(AMDGPU::OpName::dst_rel)
1105 OPERAND_CASE(AMDGPU::OpName::clamp)
1106 OPERAND_CASE(AMDGPU::OpName::src0)
1107 OPERAND_CASE(AMDGPU::OpName::src0_neg)
1108 OPERAND_CASE(AMDGPU::OpName::src0_rel)
1109 OPERAND_CASE(AMDGPU::OpName::src0_abs)
1110 OPERAND_CASE(AMDGPU::OpName::src0_sel)
1111 OPERAND_CASE(AMDGPU::OpName::src1)
1112 OPERAND_CASE(AMDGPU::OpName::src1_neg)
1113 OPERAND_CASE(AMDGPU::OpName::src1_rel)
1114 OPERAND_CASE(AMDGPU::OpName::src1_abs)
1115 OPERAND_CASE(AMDGPU::OpName::src1_sel)
1116 OPERAND_CASE(AMDGPU::OpName::pred_sel)
1118 llvm_unreachable("Wrong Operand");
1124 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
1125 MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg)
1127 assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
1129 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
1130 if (ST.getGeneration() <= AMDGPUSubtarget::R700)
1131 Opcode = AMDGPU::DOT4_r600;
1133 Opcode = AMDGPU::DOT4_eg;
1134 MachineBasicBlock::iterator I = MI;
1135 MachineOperand &Src0 = MI->getOperand(
1136 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src0, Slot)));
1137 MachineOperand &Src1 = MI->getOperand(
1138 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src1, Slot)));
1139 MachineInstr *MIB = buildDefaultInstruction(
1140 MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg());
1141 static const unsigned Operands[14] = {
1142 AMDGPU::OpName::update_exec_mask,
1143 AMDGPU::OpName::update_pred,
1144 AMDGPU::OpName::write,
1145 AMDGPU::OpName::omod,
1146 AMDGPU::OpName::dst_rel,
1147 AMDGPU::OpName::clamp,
1148 AMDGPU::OpName::src0_neg,
1149 AMDGPU::OpName::src0_rel,
1150 AMDGPU::OpName::src0_abs,
1151 AMDGPU::OpName::src0_sel,
1152 AMDGPU::OpName::src1_neg,
1153 AMDGPU::OpName::src1_rel,
1154 AMDGPU::OpName::src1_abs,
1155 AMDGPU::OpName::src1_sel,
1158 for (unsigned i = 0; i < 14; i++) {
1159 MachineOperand &MO = MI->getOperand(
1160 getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));
1161 assert (MO.isImm());
1162 setImmOperand(MIB, Operands[i], MO.getImm());
1164 MIB->getOperand(20).setImm(0);
1168 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
1169 MachineBasicBlock::iterator I,
1171 uint64_t Imm) const {
1172 MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
1173 AMDGPU::ALU_LITERAL_X);
1174 setImmOperand(MovImm, AMDGPU::OpName::literal, Imm);
1178 int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const {
1179 return getOperandIdx(MI.getOpcode(), Op);
1182 int R600InstrInfo::getOperandIdx(unsigned Opcode, unsigned Op) const {
1183 return AMDGPU::getNamedOperandIdx(Opcode, Op);
1186 void R600InstrInfo::setImmOperand(MachineInstr *MI, unsigned Op,
1187 int64_t Imm) const {
1188 int Idx = getOperandIdx(*MI, Op);
1189 assert(Idx != -1 && "Operand not supported for this instruction.");
1190 assert(MI->getOperand(Idx).isImm());
1191 MI->getOperand(Idx).setImm(Imm);
1194 //===----------------------------------------------------------------------===//
1195 // Instruction flag getters/setters
1196 //===----------------------------------------------------------------------===//
1198 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
1199 return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
1202 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
1203 unsigned Flag) const {
1204 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1207 // If we pass something other than the default value of Flag to this
1208 // function, it means we are want to set a flag on an instruction
1209 // that uses native encoding.
1210 assert(HAS_NATIVE_OPERANDS(TargetFlags));
1211 bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
1214 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::clamp);
1217 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::write);
1219 case MO_FLAG_NOT_LAST:
1221 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::last);
1225 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_neg); break;
1226 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_neg); break;
1227 case 2: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src2_neg); break;
1232 assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
1236 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_abs); break;
1237 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_abs); break;
1245 assert(FlagIndex != -1 && "Flag not supported for this instruction");
1247 FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
1248 assert(FlagIndex != 0 &&
1249 "Instruction flags not supported for this instruction");
1252 MachineOperand &FlagOp = MI->getOperand(FlagIndex);
1253 assert(FlagOp.isImm());
1257 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
1258 unsigned Flag) const {
1259 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1263 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1264 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1265 if (Flag == MO_FLAG_NOT_LAST) {
1266 clearFlag(MI, Operand, MO_FLAG_LAST);
1267 } else if (Flag == MO_FLAG_MASK) {
1268 clearFlag(MI, Operand, Flag);
1273 MachineOperand &FlagOp = getFlagOp(MI, Operand);
1274 FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
1278 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
1279 unsigned Flag) const {
1280 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1281 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1282 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1285 MachineOperand &FlagOp = getFlagOp(MI);
1286 unsigned InstFlags = FlagOp.getImm();
1287 InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
1288 FlagOp.setImm(InstFlags);