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 {
120 bool R600InstrInfo::isCubeOp(unsigned Opcode) const {
122 default: return false;
123 case AMDGPU::CUBE_r600_pseudo:
124 case AMDGPU::CUBE_r600_real:
125 case AMDGPU::CUBE_eg_pseudo:
126 case AMDGPU::CUBE_eg_real:
131 bool R600InstrInfo::isALUInstr(unsigned Opcode) const {
132 unsigned TargetFlags = get(Opcode).TSFlags;
134 return (TargetFlags & R600_InstFlag::ALU_INST);
137 bool R600InstrInfo::hasInstrModifiers(unsigned Opcode) const {
138 unsigned TargetFlags = get(Opcode).TSFlags;
140 return ((TargetFlags & R600_InstFlag::OP1) |
141 (TargetFlags & R600_InstFlag::OP2) |
142 (TargetFlags & R600_InstFlag::OP3));
145 bool R600InstrInfo::isLDSInstr(unsigned Opcode) const {
146 unsigned TargetFlags = get(Opcode).TSFlags;
148 return ((TargetFlags & R600_InstFlag::LDS_1A) |
149 (TargetFlags & R600_InstFlag::LDS_1A1D));
152 bool R600InstrInfo::isTransOnly(unsigned Opcode) const {
153 if (ST.hasCaymanISA())
155 return (get(Opcode).getSchedClass() == AMDGPU::TransALU);
158 bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const {
159 return isTransOnly(MI->getOpcode());
162 bool R600InstrInfo::isVectorOnly(unsigned Opcode) const {
163 return (get(Opcode).getSchedClass() == AMDGPU::VecALU);
166 bool R600InstrInfo::isVectorOnly(const MachineInstr *MI) const {
167 return isVectorOnly(MI->getOpcode());
170 bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
171 return ST.hasVertexCache() && IS_VTX(get(Opcode));
174 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const {
175 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
176 return MFI->ShaderType != ShaderType::COMPUTE && usesVertexCache(MI->getOpcode());
179 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
180 return (!ST.hasVertexCache() && IS_VTX(get(Opcode))) || IS_TEX(get(Opcode));
183 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
184 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
185 return (MFI->ShaderType == ShaderType::COMPUTE && usesVertexCache(MI->getOpcode())) ||
186 usesTextureCache(MI->getOpcode());
189 bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
192 case AMDGPU::GROUP_BARRIER:
199 int R600InstrInfo::getSrcIdx(unsigned Opcode, unsigned SrcNum) const {
200 static const unsigned OpTable[] = {
201 AMDGPU::OpName::src0,
202 AMDGPU::OpName::src1,
207 return getOperandIdx(Opcode, OpTable[SrcNum]);
210 #define SRC_SEL_ROWS 11
211 int R600InstrInfo::getSelIdx(unsigned Opcode, unsigned SrcIdx) const {
212 static const unsigned SrcSelTable[SRC_SEL_ROWS][2] = {
213 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
214 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
215 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
216 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
217 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
218 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
219 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
220 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
221 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
222 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
223 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W}
226 for (unsigned i = 0; i < SRC_SEL_ROWS; ++i) {
227 if (getOperandIdx(Opcode, SrcSelTable[i][0]) == (int)SrcIdx) {
228 return getOperandIdx(Opcode, SrcSelTable[i][1]);
235 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
236 R600InstrInfo::getSrcs(MachineInstr *MI) const {
237 SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
239 if (MI->getOpcode() == AMDGPU::DOT_4) {
240 static const unsigned OpTable[8][2] = {
241 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
242 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
243 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
244 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
245 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
246 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
247 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
248 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W},
251 for (unsigned j = 0; j < 8; j++) {
252 MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
254 unsigned Reg = MO.getReg();
255 if (Reg == AMDGPU::ALU_CONST) {
256 unsigned Sel = MI->getOperand(getOperandIdx(MI->getOpcode(),
257 OpTable[j][1])).getImm();
258 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
266 static const unsigned OpTable[3][2] = {
267 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
268 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
269 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
272 for (unsigned j = 0; j < 3; j++) {
273 int SrcIdx = getOperandIdx(MI->getOpcode(), OpTable[j][0]);
276 MachineOperand &MO = MI->getOperand(SrcIdx);
277 unsigned Reg = MI->getOperand(SrcIdx).getReg();
278 if (Reg == AMDGPU::ALU_CONST) {
279 unsigned Sel = MI->getOperand(
280 getOperandIdx(MI->getOpcode(), OpTable[j][1])).getImm();
281 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
284 if (Reg == AMDGPU::ALU_LITERAL_X) {
285 unsigned Imm = MI->getOperand(
286 getOperandIdx(MI->getOpcode(), AMDGPU::OpName::literal)).getImm();
287 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Imm));
290 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, 0));
295 std::vector<std::pair<int, unsigned> >
296 R600InstrInfo::ExtractSrcs(MachineInstr *MI,
297 const DenseMap<unsigned, unsigned> &PV,
298 unsigned &ConstCount) const {
300 const SmallVector<std::pair<MachineOperand *, int64_t>, 3> Srcs = getSrcs(MI);
301 const std::pair<int, unsigned> DummyPair(-1, 0);
302 std::vector<std::pair<int, unsigned> > Result;
304 for (unsigned n = Srcs.size(); i < n; ++i) {
305 unsigned Reg = Srcs[i].first->getReg();
306 unsigned Index = RI.getEncodingValue(Reg) & 0xff;
307 if (Reg == AMDGPU::OQAP) {
308 Result.push_back(std::pair<int, unsigned>(Index, 0));
310 if (PV.find(Reg) != PV.end()) {
311 // 255 is used to tells its a PS/PV reg
312 Result.push_back(std::pair<int, unsigned>(255, 0));
317 Result.push_back(DummyPair);
320 unsigned Chan = RI.getHWRegChan(Reg);
321 Result.push_back(std::pair<int, unsigned>(Index, Chan));
324 Result.push_back(DummyPair);
328 static std::vector<std::pair<int, unsigned> >
329 Swizzle(std::vector<std::pair<int, unsigned> > Src,
330 R600InstrInfo::BankSwizzle Swz) {
332 case R600InstrInfo::ALU_VEC_012_SCL_210:
334 case R600InstrInfo::ALU_VEC_021_SCL_122:
335 std::swap(Src[1], Src[2]);
337 case R600InstrInfo::ALU_VEC_102_SCL_221:
338 std::swap(Src[0], Src[1]);
340 case R600InstrInfo::ALU_VEC_120_SCL_212:
341 std::swap(Src[0], Src[1]);
342 std::swap(Src[0], Src[2]);
344 case R600InstrInfo::ALU_VEC_201:
345 std::swap(Src[0], Src[2]);
346 std::swap(Src[0], Src[1]);
348 case R600InstrInfo::ALU_VEC_210:
349 std::swap(Src[0], Src[2]);
356 getTransSwizzle(R600InstrInfo::BankSwizzle Swz, unsigned Op) {
358 case R600InstrInfo::ALU_VEC_012_SCL_210: {
359 unsigned Cycles[3] = { 2, 1, 0};
362 case R600InstrInfo::ALU_VEC_021_SCL_122: {
363 unsigned Cycles[3] = { 1, 2, 2};
366 case R600InstrInfo::ALU_VEC_120_SCL_212: {
367 unsigned Cycles[3] = { 2, 1, 2};
370 case R600InstrInfo::ALU_VEC_102_SCL_221: {
371 unsigned Cycles[3] = { 2, 2, 1};
375 llvm_unreachable("Wrong Swizzle for Trans Slot");
380 /// returns how many MIs (whose inputs are represented by IGSrcs) can be packed
381 /// in the same Instruction Group while meeting read port limitations given a
382 /// Swz swizzle sequence.
383 unsigned R600InstrInfo::isLegalUpTo(
384 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
385 const std::vector<R600InstrInfo::BankSwizzle> &Swz,
386 const std::vector<std::pair<int, unsigned> > &TransSrcs,
387 R600InstrInfo::BankSwizzle TransSwz) const {
389 memset(Vector, -1, sizeof(Vector));
390 for (unsigned i = 0, e = IGSrcs.size(); i < e; i++) {
391 const std::vector<std::pair<int, unsigned> > &Srcs =
392 Swizzle(IGSrcs[i], Swz[i]);
393 for (unsigned j = 0; j < 3; j++) {
394 const std::pair<int, unsigned> &Src = Srcs[j];
395 if (Src.first < 0 || Src.first == 255)
397 if (Src.first == GET_REG_INDEX(RI.getEncodingValue(AMDGPU::OQAP))) {
398 if (Swz[i] != R600InstrInfo::ALU_VEC_012_SCL_210 &&
399 Swz[i] != R600InstrInfo::ALU_VEC_021_SCL_122) {
400 // The value from output queue A (denoted by register OQAP) can
401 // only be fetched during the first cycle.
404 // OQAP does not count towards the normal read port restrictions
407 if (Vector[Src.second][j] < 0)
408 Vector[Src.second][j] = Src.first;
409 if (Vector[Src.second][j] != Src.first)
413 // Now check Trans Alu
414 for (unsigned i = 0, e = TransSrcs.size(); i < e; ++i) {
415 const std::pair<int, unsigned> &Src = TransSrcs[i];
416 unsigned Cycle = getTransSwizzle(TransSwz, i);
419 if (Src.first == 255)
421 if (Vector[Src.second][Cycle] < 0)
422 Vector[Src.second][Cycle] = Src.first;
423 if (Vector[Src.second][Cycle] != Src.first)
424 return IGSrcs.size() - 1;
426 return IGSrcs.size();
429 /// Given a swizzle sequence SwzCandidate and an index Idx, returns the next
430 /// (in lexicographic term) swizzle sequence assuming that all swizzles after
431 /// Idx can be skipped
433 NextPossibleSolution(
434 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
436 assert(Idx < SwzCandidate.size());
438 while (ResetIdx > -1 && SwzCandidate[ResetIdx] == R600InstrInfo::ALU_VEC_210)
440 for (unsigned i = ResetIdx + 1, e = SwzCandidate.size(); i < e; i++) {
441 SwzCandidate[i] = R600InstrInfo::ALU_VEC_012_SCL_210;
445 int NextSwizzle = SwzCandidate[ResetIdx] + 1;
446 SwzCandidate[ResetIdx] = (R600InstrInfo::BankSwizzle)NextSwizzle;
450 /// Enumerate all possible Swizzle sequence to find one that can meet all
451 /// read port requirements.
452 bool R600InstrInfo::FindSwizzleForVectorSlot(
453 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
454 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
455 const std::vector<std::pair<int, unsigned> > &TransSrcs,
456 R600InstrInfo::BankSwizzle TransSwz) const {
457 unsigned ValidUpTo = 0;
459 ValidUpTo = isLegalUpTo(IGSrcs, SwzCandidate, TransSrcs, TransSwz);
460 if (ValidUpTo == IGSrcs.size())
462 } while (NextPossibleSolution(SwzCandidate, ValidUpTo));
466 /// Instructions in Trans slot can't read gpr at cycle 0 if they also read
467 /// a const, and can't read a gpr at cycle 1 if they read 2 const.
469 isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,
470 const std::vector<std::pair<int, unsigned> > &TransOps,
471 unsigned ConstCount) {
472 for (unsigned i = 0, e = TransOps.size(); i < e; ++i) {
473 const std::pair<int, unsigned> &Src = TransOps[i];
474 unsigned Cycle = getTransSwizzle(TransSwz, i);
477 if (ConstCount > 0 && Cycle == 0)
479 if (ConstCount > 1 && Cycle == 1)
486 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG,
487 const DenseMap<unsigned, unsigned> &PV,
488 std::vector<BankSwizzle> &ValidSwizzle,
491 //Todo : support shared src0 - src1 operand
493 std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs;
494 ValidSwizzle.clear();
496 BankSwizzle TransBS = ALU_VEC_012_SCL_210;
497 for (unsigned i = 0, e = IG.size(); i < e; ++i) {
498 IGSrcs.push_back(ExtractSrcs(IG[i], PV, ConstCount));
499 unsigned Op = getOperandIdx(IG[i]->getOpcode(),
500 AMDGPU::OpName::bank_swizzle);
501 ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle)
502 IG[i]->getOperand(Op).getImm());
504 std::vector<std::pair<int, unsigned> > TransOps;
506 return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS);
508 TransOps = IGSrcs.back();
510 ValidSwizzle.pop_back();
512 static const R600InstrInfo::BankSwizzle TransSwz[] = {
518 for (unsigned i = 0; i < 4; i++) {
519 TransBS = TransSwz[i];
520 if (!isConstCompatible(TransBS, TransOps, ConstCount))
522 bool Result = FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps,
525 ValidSwizzle.push_back(TransBS);
535 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts)
537 assert (Consts.size() <= 12 && "Too many operands in instructions group");
538 unsigned Pair1 = 0, Pair2 = 0;
539 for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
540 unsigned ReadConstHalf = Consts[i] & 2;
541 unsigned ReadConstIndex = Consts[i] & (~3);
542 unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf;
544 Pair1 = ReadHalfConst;
547 if (Pair1 == ReadHalfConst)
550 Pair2 = ReadHalfConst;
553 if (Pair2 != ReadHalfConst)
560 R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
562 std::vector<unsigned> Consts;
563 for (unsigned i = 0, n = MIs.size(); i < n; i++) {
564 MachineInstr *MI = MIs[i];
565 if (!isALUInstr(MI->getOpcode()))
568 const SmallVectorImpl<std::pair<MachineOperand *, int64_t> > &Srcs =
571 for (unsigned j = 0, e = Srcs.size(); j < e; j++) {
572 std::pair<MachineOperand *, unsigned> Src = Srcs[j];
573 if (Src.first->getReg() == AMDGPU::ALU_CONST)
574 Consts.push_back(Src.second);
575 if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) ||
576 AMDGPU::R600_KC1RegClass.contains(Src.first->getReg())) {
577 unsigned Index = RI.getEncodingValue(Src.first->getReg()) & 0xff;
578 unsigned Chan = RI.getHWRegChan(Src.first->getReg());
579 Consts.push_back((Index << 2) | Chan);
583 return fitsConstReadLimitations(Consts);
586 DFAPacketizer *R600InstrInfo::CreateTargetScheduleState(const TargetMachine *TM,
587 const ScheduleDAG *DAG) const {
588 const InstrItineraryData *II = TM->getInstrItineraryData();
589 return TM->getSubtarget<AMDGPUSubtarget>().createDFAPacketizer(II);
593 isPredicateSetter(unsigned Opcode) {
602 static MachineInstr *
603 findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
604 MachineBasicBlock::iterator I) {
605 while (I != MBB.begin()) {
607 MachineInstr *MI = I;
608 if (isPredicateSetter(MI->getOpcode()))
616 bool isJump(unsigned Opcode) {
617 return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
621 R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
622 MachineBasicBlock *&TBB,
623 MachineBasicBlock *&FBB,
624 SmallVectorImpl<MachineOperand> &Cond,
625 bool AllowModify) const {
626 // Most of the following comes from the ARM implementation of AnalyzeBranch
628 // If the block has no terminators, it just falls into the block after it.
629 MachineBasicBlock::iterator I = MBB.end();
630 if (I == MBB.begin())
633 while (I->isDebugValue()) {
634 if (I == MBB.begin())
638 if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) {
642 // Get the last instruction in the block.
643 MachineInstr *LastInst = I;
645 // If there is only one terminator instruction, process it.
646 unsigned LastOpc = LastInst->getOpcode();
647 if (I == MBB.begin() ||
648 !isJump(static_cast<MachineInstr *>(--I)->getOpcode())) {
649 if (LastOpc == AMDGPU::JUMP) {
650 TBB = LastInst->getOperand(0).getMBB();
652 } else if (LastOpc == AMDGPU::JUMP_COND) {
653 MachineInstr *predSet = I;
654 while (!isPredicateSetter(predSet->getOpcode())) {
657 TBB = LastInst->getOperand(0).getMBB();
658 Cond.push_back(predSet->getOperand(1));
659 Cond.push_back(predSet->getOperand(2));
660 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
663 return true; // Can't handle indirect branch.
666 // Get the instruction before it if it is a terminator.
667 MachineInstr *SecondLastInst = I;
668 unsigned SecondLastOpc = SecondLastInst->getOpcode();
670 // If the block ends with a B and a Bcc, handle it.
671 if (SecondLastOpc == AMDGPU::JUMP_COND && LastOpc == AMDGPU::JUMP) {
672 MachineInstr *predSet = --I;
673 while (!isPredicateSetter(predSet->getOpcode())) {
676 TBB = SecondLastInst->getOperand(0).getMBB();
677 FBB = LastInst->getOperand(0).getMBB();
678 Cond.push_back(predSet->getOperand(1));
679 Cond.push_back(predSet->getOperand(2));
680 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
684 // Otherwise, can't handle this.
688 int R600InstrInfo::getBranchInstr(const MachineOperand &op) const {
689 const MachineInstr *MI = op.getParent();
691 switch (MI->getDesc().OpInfo->RegClass) {
692 default: // FIXME: fallthrough??
693 case AMDGPU::GPRI32RegClassID: return AMDGPU::BRANCH_COND_i32;
694 case AMDGPU::GPRF32RegClassID: return AMDGPU::BRANCH_COND_f32;
699 MachineBasicBlock::iterator FindLastAluClause(MachineBasicBlock &MBB) {
700 for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend();
702 if (It->getOpcode() == AMDGPU::CF_ALU ||
703 It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE)
704 return llvm::prior(It.base());
710 R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
711 MachineBasicBlock *TBB,
712 MachineBasicBlock *FBB,
713 const SmallVectorImpl<MachineOperand> &Cond,
715 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
719 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB);
722 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
723 assert(PredSet && "No previous predicate !");
724 addFlag(PredSet, 0, MO_FLAG_PUSH);
725 PredSet->getOperand(2).setImm(Cond[1].getImm());
727 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
729 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
730 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
731 if (CfAlu == MBB.end())
733 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
734 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
738 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
739 assert(PredSet && "No previous predicate !");
740 addFlag(PredSet, 0, MO_FLAG_PUSH);
741 PredSet->getOperand(2).setImm(Cond[1].getImm());
742 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
744 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
745 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB);
746 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
747 if (CfAlu == MBB.end())
749 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
750 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
756 R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
758 // Note : we leave PRED* instructions there.
759 // They may be needed when predicating instructions.
761 MachineBasicBlock::iterator I = MBB.end();
763 if (I == MBB.begin()) {
767 switch (I->getOpcode()) {
770 case AMDGPU::JUMP_COND: {
771 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
772 clearFlag(predSet, 0, MO_FLAG_PUSH);
773 I->eraseFromParent();
774 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
775 if (CfAlu == MBB.end())
777 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
778 CfAlu->setDesc(get(AMDGPU::CF_ALU));
782 I->eraseFromParent();
787 if (I == MBB.begin()) {
791 switch (I->getOpcode()) {
792 // FIXME: only one case??
795 case AMDGPU::JUMP_COND: {
796 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
797 clearFlag(predSet, 0, MO_FLAG_PUSH);
798 I->eraseFromParent();
799 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
800 if (CfAlu == MBB.end())
802 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
803 CfAlu->setDesc(get(AMDGPU::CF_ALU));
807 I->eraseFromParent();
814 R600InstrInfo::isPredicated(const MachineInstr *MI) const {
815 int idx = MI->findFirstPredOperandIdx();
819 unsigned Reg = MI->getOperand(idx).getReg();
821 default: return false;
822 case AMDGPU::PRED_SEL_ONE:
823 case AMDGPU::PRED_SEL_ZERO:
824 case AMDGPU::PREDICATE_BIT:
830 R600InstrInfo::isPredicable(MachineInstr *MI) const {
831 // XXX: KILL* instructions can be predicated, but they must be the last
832 // instruction in a clause, so this means any instructions after them cannot
833 // be predicated. Until we have proper support for instruction clauses in the
834 // backend, we will mark KILL* instructions as unpredicable.
836 if (MI->getOpcode() == AMDGPU::KILLGT) {
838 } else if (MI->getOpcode() == AMDGPU::CF_ALU) {
839 // If the clause start in the middle of MBB then the MBB has more
840 // than a single clause, unable to predicate several clauses.
841 if (MI->getParent()->begin() != MachineBasicBlock::iterator(MI))
843 // TODO: We don't support KC merging atm
844 if (MI->getOperand(3).getImm() != 0 || MI->getOperand(4).getImm() != 0)
847 } else if (isVector(*MI)) {
850 return AMDGPUInstrInfo::isPredicable(MI);
856 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
858 unsigned ExtraPredCycles,
859 const BranchProbability &Probability) const{
864 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
866 unsigned ExtraTCycles,
867 MachineBasicBlock &FMBB,
869 unsigned ExtraFCycles,
870 const BranchProbability &Probability) const {
875 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
877 const BranchProbability &Probability)
883 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
884 MachineBasicBlock &FMBB) const {
890 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
891 MachineOperand &MO = Cond[1];
892 switch (MO.getImm()) {
893 case OPCODE_IS_ZERO_INT:
894 MO.setImm(OPCODE_IS_NOT_ZERO_INT);
896 case OPCODE_IS_NOT_ZERO_INT:
897 MO.setImm(OPCODE_IS_ZERO_INT);
900 MO.setImm(OPCODE_IS_NOT_ZERO);
902 case OPCODE_IS_NOT_ZERO:
903 MO.setImm(OPCODE_IS_ZERO);
909 MachineOperand &MO2 = Cond[2];
910 switch (MO2.getReg()) {
911 case AMDGPU::PRED_SEL_ZERO:
912 MO2.setReg(AMDGPU::PRED_SEL_ONE);
914 case AMDGPU::PRED_SEL_ONE:
915 MO2.setReg(AMDGPU::PRED_SEL_ZERO);
924 R600InstrInfo::DefinesPredicate(MachineInstr *MI,
925 std::vector<MachineOperand> &Pred) const {
926 return isPredicateSetter(MI->getOpcode());
931 R600InstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
932 const SmallVectorImpl<MachineOperand> &Pred2) const {
938 R600InstrInfo::PredicateInstruction(MachineInstr *MI,
939 const SmallVectorImpl<MachineOperand> &Pred) const {
940 int PIdx = MI->findFirstPredOperandIdx();
942 if (MI->getOpcode() == AMDGPU::CF_ALU) {
943 MI->getOperand(8).setImm(0);
948 MachineOperand &PMO = MI->getOperand(PIdx);
949 PMO.setReg(Pred[2].getReg());
950 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
951 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
958 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
959 const MachineInstr *MI,
960 unsigned *PredCost) const {
966 int R600InstrInfo::getIndirectIndexBegin(const MachineFunction &MF) const {
967 const MachineRegisterInfo &MRI = MF.getRegInfo();
968 const MachineFrameInfo *MFI = MF.getFrameInfo();
971 if (MFI->getNumObjects() == 0) {
975 if (MRI.livein_empty()) {
979 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
980 LE = MRI.livein_end();
982 Offset = std::max(Offset,
983 GET_REG_INDEX(RI.getEncodingValue(LI->first)));
989 int R600InstrInfo::getIndirectIndexEnd(const MachineFunction &MF) const {
991 const MachineFrameInfo *MFI = MF.getFrameInfo();
993 // Variable sized objects are not supported
994 assert(!MFI->hasVarSizedObjects());
996 if (MFI->getNumObjects() == 0) {
1000 Offset = TM.getFrameLowering()->getFrameIndexOffset(MF, -1);
1002 return getIndirectIndexBegin(MF) + Offset;
1005 std::vector<unsigned> R600InstrInfo::getIndirectReservedRegs(
1006 const MachineFunction &MF) const {
1007 const AMDGPUFrameLowering *TFL =
1008 static_cast<const AMDGPUFrameLowering*>(TM.getFrameLowering());
1009 std::vector<unsigned> Regs;
1011 unsigned StackWidth = TFL->getStackWidth(MF);
1012 int End = getIndirectIndexEnd(MF);
1018 for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
1019 unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
1020 Regs.push_back(SuperReg);
1021 for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
1022 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
1023 Regs.push_back(Reg);
1029 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
1030 unsigned Channel) const {
1031 // XXX: Remove when we support a stack width > 2
1032 assert(Channel == 0);
1036 const TargetRegisterClass * R600InstrInfo::getIndirectAddrStoreRegClass(
1037 unsigned SourceReg) const {
1038 return &AMDGPU::R600_TReg32RegClass;
1041 const TargetRegisterClass *R600InstrInfo::getIndirectAddrLoadRegClass() const {
1042 return &AMDGPU::TRegMemRegClass;
1045 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1046 MachineBasicBlock::iterator I,
1047 unsigned ValueReg, unsigned Address,
1048 unsigned OffsetReg) const {
1049 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1050 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1051 AMDGPU::AR_X, OffsetReg);
1052 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1054 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1056 .addReg(AMDGPU::AR_X,
1057 RegState::Implicit | RegState::Kill);
1058 setImmOperand(Mov, AMDGPU::OpName::dst_rel, 1);
1062 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1063 MachineBasicBlock::iterator I,
1064 unsigned ValueReg, unsigned Address,
1065 unsigned OffsetReg) const {
1066 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1067 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1070 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1071 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1074 .addReg(AMDGPU::AR_X,
1075 RegState::Implicit | RegState::Kill);
1076 setImmOperand(Mov, AMDGPU::OpName::src0_rel, 1);
1081 const TargetRegisterClass *R600InstrInfo::getSuperIndirectRegClass() const {
1082 return &AMDGPU::IndirectRegRegClass;
1085 unsigned R600InstrInfo::getMaxAlusPerClause() const {
1089 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
1090 MachineBasicBlock::iterator I,
1094 unsigned Src1Reg) const {
1095 MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
1099 MIB.addImm(0) // $update_exec_mask
1100 .addImm(0); // $update_predicate
1102 MIB.addImm(1) // $write
1104 .addImm(0) // $dst_rel
1105 .addImm(0) // $dst_clamp
1106 .addReg(Src0Reg) // $src0
1107 .addImm(0) // $src0_neg
1108 .addImm(0) // $src0_rel
1109 .addImm(0) // $src0_abs
1110 .addImm(-1); // $src0_sel
1113 MIB.addReg(Src1Reg) // $src1
1114 .addImm(0) // $src1_neg
1115 .addImm(0) // $src1_rel
1116 .addImm(0) // $src1_abs
1117 .addImm(-1); // $src1_sel
1120 //XXX: The r600g finalizer expects this to be 1, once we've moved the
1121 //scheduling to the backend, we can change the default to 0.
1122 MIB.addImm(1) // $last
1123 .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
1124 .addImm(0) // $literal
1125 .addImm(0); // $bank_swizzle
1130 #define OPERAND_CASE(Label) \
1132 static const unsigned Ops[] = \
1142 static unsigned getSlotedOps(unsigned Op, unsigned Slot) {
1144 OPERAND_CASE(AMDGPU::OpName::update_exec_mask)
1145 OPERAND_CASE(AMDGPU::OpName::update_pred)
1146 OPERAND_CASE(AMDGPU::OpName::write)
1147 OPERAND_CASE(AMDGPU::OpName::omod)
1148 OPERAND_CASE(AMDGPU::OpName::dst_rel)
1149 OPERAND_CASE(AMDGPU::OpName::clamp)
1150 OPERAND_CASE(AMDGPU::OpName::src0)
1151 OPERAND_CASE(AMDGPU::OpName::src0_neg)
1152 OPERAND_CASE(AMDGPU::OpName::src0_rel)
1153 OPERAND_CASE(AMDGPU::OpName::src0_abs)
1154 OPERAND_CASE(AMDGPU::OpName::src0_sel)
1155 OPERAND_CASE(AMDGPU::OpName::src1)
1156 OPERAND_CASE(AMDGPU::OpName::src1_neg)
1157 OPERAND_CASE(AMDGPU::OpName::src1_rel)
1158 OPERAND_CASE(AMDGPU::OpName::src1_abs)
1159 OPERAND_CASE(AMDGPU::OpName::src1_sel)
1160 OPERAND_CASE(AMDGPU::OpName::pred_sel)
1162 llvm_unreachable("Wrong Operand");
1168 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
1169 MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg)
1171 assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
1173 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
1174 if (ST.getGeneration() <= AMDGPUSubtarget::R700)
1175 Opcode = AMDGPU::DOT4_r600;
1177 Opcode = AMDGPU::DOT4_eg;
1178 MachineBasicBlock::iterator I = MI;
1179 MachineOperand &Src0 = MI->getOperand(
1180 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src0, Slot)));
1181 MachineOperand &Src1 = MI->getOperand(
1182 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src1, Slot)));
1183 MachineInstr *MIB = buildDefaultInstruction(
1184 MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg());
1185 static const unsigned Operands[14] = {
1186 AMDGPU::OpName::update_exec_mask,
1187 AMDGPU::OpName::update_pred,
1188 AMDGPU::OpName::write,
1189 AMDGPU::OpName::omod,
1190 AMDGPU::OpName::dst_rel,
1191 AMDGPU::OpName::clamp,
1192 AMDGPU::OpName::src0_neg,
1193 AMDGPU::OpName::src0_rel,
1194 AMDGPU::OpName::src0_abs,
1195 AMDGPU::OpName::src0_sel,
1196 AMDGPU::OpName::src1_neg,
1197 AMDGPU::OpName::src1_rel,
1198 AMDGPU::OpName::src1_abs,
1199 AMDGPU::OpName::src1_sel,
1202 for (unsigned i = 0; i < 14; i++) {
1203 MachineOperand &MO = MI->getOperand(
1204 getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));
1205 assert (MO.isImm());
1206 setImmOperand(MIB, Operands[i], MO.getImm());
1208 MIB->getOperand(20).setImm(0);
1212 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
1213 MachineBasicBlock::iterator I,
1215 uint64_t Imm) const {
1216 MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
1217 AMDGPU::ALU_LITERAL_X);
1218 setImmOperand(MovImm, AMDGPU::OpName::literal, Imm);
1222 int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const {
1223 return getOperandIdx(MI.getOpcode(), Op);
1226 int R600InstrInfo::getOperandIdx(unsigned Opcode, unsigned Op) const {
1227 return AMDGPU::getNamedOperandIdx(Opcode, Op);
1230 void R600InstrInfo::setImmOperand(MachineInstr *MI, unsigned Op,
1231 int64_t Imm) const {
1232 int Idx = getOperandIdx(*MI, Op);
1233 assert(Idx != -1 && "Operand not supported for this instruction.");
1234 assert(MI->getOperand(Idx).isImm());
1235 MI->getOperand(Idx).setImm(Imm);
1238 //===----------------------------------------------------------------------===//
1239 // Instruction flag getters/setters
1240 //===----------------------------------------------------------------------===//
1242 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
1243 return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
1246 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
1247 unsigned Flag) const {
1248 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1251 // If we pass something other than the default value of Flag to this
1252 // function, it means we are want to set a flag on an instruction
1253 // that uses native encoding.
1254 assert(HAS_NATIVE_OPERANDS(TargetFlags));
1255 bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
1258 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::clamp);
1261 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::write);
1263 case MO_FLAG_NOT_LAST:
1265 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::last);
1269 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_neg); break;
1270 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_neg); break;
1271 case 2: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src2_neg); break;
1276 assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
1280 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_abs); break;
1281 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_abs); break;
1289 assert(FlagIndex != -1 && "Flag not supported for this instruction");
1291 FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
1292 assert(FlagIndex != 0 &&
1293 "Instruction flags not supported for this instruction");
1296 MachineOperand &FlagOp = MI->getOperand(FlagIndex);
1297 assert(FlagOp.isImm());
1301 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
1302 unsigned Flag) const {
1303 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1307 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1308 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1309 if (Flag == MO_FLAG_NOT_LAST) {
1310 clearFlag(MI, Operand, MO_FLAG_LAST);
1311 } else if (Flag == MO_FLAG_MASK) {
1312 clearFlag(MI, Operand, Flag);
1317 MachineOperand &FlagOp = getFlagOp(MI, Operand);
1318 FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
1322 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
1323 unsigned Flag) const {
1324 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1325 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1326 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1329 MachineOperand &FlagOp = getFlagOp(MI);
1330 unsigned InstFlags = FlagOp.getImm();
1331 InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
1332 FlagOp.setImm(InstFlags);