1 //===-- R600InstrInfo.cpp - R600 Instruction Information ------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief R600 Implementation of TargetInstrInfo.
13 //===----------------------------------------------------------------------===//
15 #include "R600InstrInfo.h"
17 #include "AMDGPUSubtarget.h"
18 #include "AMDGPUTargetMachine.h"
19 #include "R600Defines.h"
20 #include "R600MachineFunctionInfo.h"
21 #include "R600RegisterInfo.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #define GET_INSTRINFO_CTOR
27 #include "AMDGPUGenDFAPacketizer.inc"
31 R600InstrInfo::R600InstrInfo(AMDGPUTargetMachine &tm)
32 : AMDGPUInstrInfo(tm),
34 ST(tm.getSubtarget<AMDGPUSubtarget>())
37 const R600RegisterInfo &R600InstrInfo::getRegisterInfo() const {
41 bool R600InstrInfo::isTrig(const MachineInstr &MI) const {
42 return get(MI.getOpcode()).TSFlags & R600_InstFlag::TRIG;
45 bool R600InstrInfo::isVector(const MachineInstr &MI) const {
46 return get(MI.getOpcode()).TSFlags & R600_InstFlag::VECTOR;
50 R600InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
51 MachineBasicBlock::iterator MI, DebugLoc DL,
52 unsigned DestReg, unsigned SrcReg,
54 unsigned VectorComponents = 0;
55 if (AMDGPU::R600_Reg128RegClass.contains(DestReg) &&
56 AMDGPU::R600_Reg128RegClass.contains(SrcReg)) {
58 } else if(AMDGPU::R600_Reg64RegClass.contains(DestReg) &&
59 AMDGPU::R600_Reg64RegClass.contains(SrcReg)) {
63 if (VectorComponents > 0) {
64 for (unsigned I = 0; I < VectorComponents; I++) {
65 unsigned SubRegIndex = RI.getSubRegFromChannel(I);
66 buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
67 RI.getSubReg(DestReg, SubRegIndex),
68 RI.getSubReg(SrcReg, SubRegIndex))
70 RegState::Define | RegState::Implicit);
73 MachineInstr *NewMI = buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
75 NewMI->getOperand(getOperandIdx(*NewMI, AMDGPU::OpName::src0))
80 MachineInstr * R600InstrInfo::getMovImmInstr(MachineFunction *MF,
81 unsigned DstReg, int64_t Imm) const {
82 MachineInstr * MI = MF->CreateMachineInstr(get(AMDGPU::MOV), DebugLoc());
83 MachineInstrBuilder MIB(*MF, MI);
84 MIB.addReg(DstReg, RegState::Define);
85 MIB.addReg(AMDGPU::ALU_LITERAL_X);
87 MIB.addReg(0); // PREDICATE_BIT
92 unsigned R600InstrInfo::getIEQOpcode() const {
93 return AMDGPU::SETE_INT;
96 bool R600InstrInfo::isMov(unsigned Opcode) const {
100 default: return false;
102 case AMDGPU::MOV_IMM_F32:
103 case AMDGPU::MOV_IMM_I32:
108 // Some instructions act as place holders to emulate operations that the GPU
109 // hardware does automatically. This function can be used to check if
110 // an opcode falls into this category.
111 bool R600InstrInfo::isPlaceHolderOpcode(unsigned Opcode) const {
113 default: return false;
119 bool R600InstrInfo::isReductionOp(unsigned Opcode) const {
123 bool R600InstrInfo::isCubeOp(unsigned Opcode) const {
125 default: return false;
126 case AMDGPU::CUBE_r600_pseudo:
127 case AMDGPU::CUBE_r600_real:
128 case AMDGPU::CUBE_eg_pseudo:
129 case AMDGPU::CUBE_eg_real:
134 bool R600InstrInfo::isALUInstr(unsigned Opcode) const {
135 unsigned TargetFlags = get(Opcode).TSFlags;
137 return (TargetFlags & R600_InstFlag::ALU_INST);
140 bool R600InstrInfo::hasInstrModifiers(unsigned Opcode) const {
141 unsigned TargetFlags = get(Opcode).TSFlags;
143 return ((TargetFlags & R600_InstFlag::OP1) |
144 (TargetFlags & R600_InstFlag::OP2) |
145 (TargetFlags & R600_InstFlag::OP3));
148 bool R600InstrInfo::isLDSInstr(unsigned Opcode) const {
149 unsigned TargetFlags = get(Opcode).TSFlags;
151 return ((TargetFlags & R600_InstFlag::LDS_1A) |
152 (TargetFlags & R600_InstFlag::LDS_1A1D) |
153 (TargetFlags & R600_InstFlag::LDS_1A2D));
156 bool R600InstrInfo::isTransOnly(unsigned Opcode) const {
157 if (ST.hasCaymanISA())
159 return (get(Opcode).getSchedClass() == AMDGPU::Sched::TransALU);
162 bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const {
163 return isTransOnly(MI->getOpcode());
166 bool R600InstrInfo::isVectorOnly(unsigned Opcode) const {
167 return (get(Opcode).getSchedClass() == AMDGPU::Sched::VecALU);
170 bool R600InstrInfo::isVectorOnly(const MachineInstr *MI) const {
171 return isVectorOnly(MI->getOpcode());
174 bool R600InstrInfo::isExport(unsigned Opcode) const {
175 return (get(Opcode).TSFlags & R600_InstFlag::IS_EXPORT);
178 bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
179 return ST.hasVertexCache() && IS_VTX(get(Opcode));
182 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const {
183 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
184 return MFI->ShaderType != ShaderType::COMPUTE && usesVertexCache(MI->getOpcode());
187 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
188 return (!ST.hasVertexCache() && IS_VTX(get(Opcode))) || IS_TEX(get(Opcode));
191 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
192 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
193 return (MFI->ShaderType == ShaderType::COMPUTE && usesVertexCache(MI->getOpcode())) ||
194 usesTextureCache(MI->getOpcode());
197 bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
200 case AMDGPU::GROUP_BARRIER:
207 int R600InstrInfo::getSrcIdx(unsigned Opcode, unsigned SrcNum) const {
208 static const unsigned OpTable[] = {
209 AMDGPU::OpName::src0,
210 AMDGPU::OpName::src1,
215 return getOperandIdx(Opcode, OpTable[SrcNum]);
218 #define SRC_SEL_ROWS 11
219 int R600InstrInfo::getSelIdx(unsigned Opcode, unsigned SrcIdx) const {
220 static const unsigned SrcSelTable[SRC_SEL_ROWS][2] = {
221 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
222 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
223 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
224 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
225 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
226 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
227 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
228 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
229 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
230 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
231 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W}
234 for (unsigned i = 0; i < SRC_SEL_ROWS; ++i) {
235 if (getOperandIdx(Opcode, SrcSelTable[i][0]) == (int)SrcIdx) {
236 return getOperandIdx(Opcode, SrcSelTable[i][1]);
243 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
244 R600InstrInfo::getSrcs(MachineInstr *MI) const {
245 SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
247 if (MI->getOpcode() == AMDGPU::DOT_4) {
248 static const unsigned OpTable[8][2] = {
249 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
250 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
251 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
252 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
253 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
254 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
255 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
256 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W},
259 for (unsigned j = 0; j < 8; j++) {
260 MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
262 unsigned Reg = MO.getReg();
263 if (Reg == AMDGPU::ALU_CONST) {
264 unsigned Sel = MI->getOperand(getOperandIdx(MI->getOpcode(),
265 OpTable[j][1])).getImm();
266 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
274 static const unsigned OpTable[3][2] = {
275 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
276 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
277 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
280 for (unsigned j = 0; j < 3; j++) {
281 int SrcIdx = getOperandIdx(MI->getOpcode(), OpTable[j][0]);
284 MachineOperand &MO = MI->getOperand(SrcIdx);
285 unsigned Reg = MI->getOperand(SrcIdx).getReg();
286 if (Reg == AMDGPU::ALU_CONST) {
287 unsigned Sel = MI->getOperand(
288 getOperandIdx(MI->getOpcode(), OpTable[j][1])).getImm();
289 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
292 if (Reg == AMDGPU::ALU_LITERAL_X) {
293 unsigned Imm = MI->getOperand(
294 getOperandIdx(MI->getOpcode(), AMDGPU::OpName::literal)).getImm();
295 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Imm));
298 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, 0));
303 std::vector<std::pair<int, unsigned> >
304 R600InstrInfo::ExtractSrcs(MachineInstr *MI,
305 const DenseMap<unsigned, unsigned> &PV,
306 unsigned &ConstCount) const {
308 const SmallVector<std::pair<MachineOperand *, int64_t>, 3> Srcs = getSrcs(MI);
309 const std::pair<int, unsigned> DummyPair(-1, 0);
310 std::vector<std::pair<int, unsigned> > Result;
312 for (unsigned n = Srcs.size(); i < n; ++i) {
313 unsigned Reg = Srcs[i].first->getReg();
314 unsigned Index = RI.getEncodingValue(Reg) & 0xff;
315 if (Reg == AMDGPU::OQAP) {
316 Result.push_back(std::pair<int, unsigned>(Index, 0));
318 if (PV.find(Reg) != PV.end()) {
319 // 255 is used to tells its a PS/PV reg
320 Result.push_back(std::pair<int, unsigned>(255, 0));
325 Result.push_back(DummyPair);
328 unsigned Chan = RI.getHWRegChan(Reg);
329 Result.push_back(std::pair<int, unsigned>(Index, Chan));
332 Result.push_back(DummyPair);
336 static std::vector<std::pair<int, unsigned> >
337 Swizzle(std::vector<std::pair<int, unsigned> > Src,
338 R600InstrInfo::BankSwizzle Swz) {
340 case R600InstrInfo::ALU_VEC_012_SCL_210:
342 case R600InstrInfo::ALU_VEC_021_SCL_122:
343 std::swap(Src[1], Src[2]);
345 case R600InstrInfo::ALU_VEC_102_SCL_221:
346 std::swap(Src[0], Src[1]);
348 case R600InstrInfo::ALU_VEC_120_SCL_212:
349 std::swap(Src[0], Src[1]);
350 std::swap(Src[0], Src[2]);
352 case R600InstrInfo::ALU_VEC_201:
353 std::swap(Src[0], Src[2]);
354 std::swap(Src[0], Src[1]);
356 case R600InstrInfo::ALU_VEC_210:
357 std::swap(Src[0], Src[2]);
364 getTransSwizzle(R600InstrInfo::BankSwizzle Swz, unsigned Op) {
366 case R600InstrInfo::ALU_VEC_012_SCL_210: {
367 unsigned Cycles[3] = { 2, 1, 0};
370 case R600InstrInfo::ALU_VEC_021_SCL_122: {
371 unsigned Cycles[3] = { 1, 2, 2};
374 case R600InstrInfo::ALU_VEC_120_SCL_212: {
375 unsigned Cycles[3] = { 2, 1, 2};
378 case R600InstrInfo::ALU_VEC_102_SCL_221: {
379 unsigned Cycles[3] = { 2, 2, 1};
383 llvm_unreachable("Wrong Swizzle for Trans Slot");
388 /// returns how many MIs (whose inputs are represented by IGSrcs) can be packed
389 /// in the same Instruction Group while meeting read port limitations given a
390 /// Swz swizzle sequence.
391 unsigned R600InstrInfo::isLegalUpTo(
392 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
393 const std::vector<R600InstrInfo::BankSwizzle> &Swz,
394 const std::vector<std::pair<int, unsigned> > &TransSrcs,
395 R600InstrInfo::BankSwizzle TransSwz) const {
397 memset(Vector, -1, sizeof(Vector));
398 for (unsigned i = 0, e = IGSrcs.size(); i < e; i++) {
399 const std::vector<std::pair<int, unsigned> > &Srcs =
400 Swizzle(IGSrcs[i], Swz[i]);
401 for (unsigned j = 0; j < 3; j++) {
402 const std::pair<int, unsigned> &Src = Srcs[j];
403 if (Src.first < 0 || Src.first == 255)
405 if (Src.first == GET_REG_INDEX(RI.getEncodingValue(AMDGPU::OQAP))) {
406 if (Swz[i] != R600InstrInfo::ALU_VEC_012_SCL_210 &&
407 Swz[i] != R600InstrInfo::ALU_VEC_021_SCL_122) {
408 // The value from output queue A (denoted by register OQAP) can
409 // only be fetched during the first cycle.
412 // OQAP does not count towards the normal read port restrictions
415 if (Vector[Src.second][j] < 0)
416 Vector[Src.second][j] = Src.first;
417 if (Vector[Src.second][j] != Src.first)
421 // Now check Trans Alu
422 for (unsigned i = 0, e = TransSrcs.size(); i < e; ++i) {
423 const std::pair<int, unsigned> &Src = TransSrcs[i];
424 unsigned Cycle = getTransSwizzle(TransSwz, i);
427 if (Src.first == 255)
429 if (Vector[Src.second][Cycle] < 0)
430 Vector[Src.second][Cycle] = Src.first;
431 if (Vector[Src.second][Cycle] != Src.first)
432 return IGSrcs.size() - 1;
434 return IGSrcs.size();
437 /// Given a swizzle sequence SwzCandidate and an index Idx, returns the next
438 /// (in lexicographic term) swizzle sequence assuming that all swizzles after
439 /// Idx can be skipped
441 NextPossibleSolution(
442 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
444 assert(Idx < SwzCandidate.size());
446 while (ResetIdx > -1 && SwzCandidate[ResetIdx] == R600InstrInfo::ALU_VEC_210)
448 for (unsigned i = ResetIdx + 1, e = SwzCandidate.size(); i < e; i++) {
449 SwzCandidate[i] = R600InstrInfo::ALU_VEC_012_SCL_210;
453 int NextSwizzle = SwzCandidate[ResetIdx] + 1;
454 SwzCandidate[ResetIdx] = (R600InstrInfo::BankSwizzle)NextSwizzle;
458 /// Enumerate all possible Swizzle sequence to find one that can meet all
459 /// read port requirements.
460 bool R600InstrInfo::FindSwizzleForVectorSlot(
461 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
462 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
463 const std::vector<std::pair<int, unsigned> > &TransSrcs,
464 R600InstrInfo::BankSwizzle TransSwz) const {
465 unsigned ValidUpTo = 0;
467 ValidUpTo = isLegalUpTo(IGSrcs, SwzCandidate, TransSrcs, TransSwz);
468 if (ValidUpTo == IGSrcs.size())
470 } while (NextPossibleSolution(SwzCandidate, ValidUpTo));
474 /// Instructions in Trans slot can't read gpr at cycle 0 if they also read
475 /// a const, and can't read a gpr at cycle 1 if they read 2 const.
477 isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,
478 const std::vector<std::pair<int, unsigned> > &TransOps,
479 unsigned ConstCount) {
480 for (unsigned i = 0, e = TransOps.size(); i < e; ++i) {
481 const std::pair<int, unsigned> &Src = TransOps[i];
482 unsigned Cycle = getTransSwizzle(TransSwz, i);
485 if (ConstCount > 0 && Cycle == 0)
487 if (ConstCount > 1 && Cycle == 1)
494 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG,
495 const DenseMap<unsigned, unsigned> &PV,
496 std::vector<BankSwizzle> &ValidSwizzle,
499 //Todo : support shared src0 - src1 operand
501 std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs;
502 ValidSwizzle.clear();
504 BankSwizzle TransBS = ALU_VEC_012_SCL_210;
505 for (unsigned i = 0, e = IG.size(); i < e; ++i) {
506 IGSrcs.push_back(ExtractSrcs(IG[i], PV, ConstCount));
507 unsigned Op = getOperandIdx(IG[i]->getOpcode(),
508 AMDGPU::OpName::bank_swizzle);
509 ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle)
510 IG[i]->getOperand(Op).getImm());
512 std::vector<std::pair<int, unsigned> > TransOps;
514 return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS);
516 TransOps = IGSrcs.back();
518 ValidSwizzle.pop_back();
520 static const R600InstrInfo::BankSwizzle TransSwz[] = {
526 for (unsigned i = 0; i < 4; i++) {
527 TransBS = TransSwz[i];
528 if (!isConstCompatible(TransBS, TransOps, ConstCount))
530 bool Result = FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps,
533 ValidSwizzle.push_back(TransBS);
543 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts)
545 assert (Consts.size() <= 12 && "Too many operands in instructions group");
546 unsigned Pair1 = 0, Pair2 = 0;
547 for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
548 unsigned ReadConstHalf = Consts[i] & 2;
549 unsigned ReadConstIndex = Consts[i] & (~3);
550 unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf;
552 Pair1 = ReadHalfConst;
555 if (Pair1 == ReadHalfConst)
558 Pair2 = ReadHalfConst;
561 if (Pair2 != ReadHalfConst)
568 R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
570 std::vector<unsigned> Consts;
571 SmallSet<int64_t, 4> Literals;
572 for (unsigned i = 0, n = MIs.size(); i < n; i++) {
573 MachineInstr *MI = MIs[i];
574 if (!isALUInstr(MI->getOpcode()))
577 const SmallVectorImpl<std::pair<MachineOperand *, int64_t> > &Srcs =
580 for (unsigned j = 0, e = Srcs.size(); j < e; j++) {
581 std::pair<MachineOperand *, unsigned> Src = Srcs[j];
582 if (Src.first->getReg() == AMDGPU::ALU_LITERAL_X)
583 Literals.insert(Src.second);
584 if (Literals.size() > 4)
586 if (Src.first->getReg() == AMDGPU::ALU_CONST)
587 Consts.push_back(Src.second);
588 if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) ||
589 AMDGPU::R600_KC1RegClass.contains(Src.first->getReg())) {
590 unsigned Index = RI.getEncodingValue(Src.first->getReg()) & 0xff;
591 unsigned Chan = RI.getHWRegChan(Src.first->getReg());
592 Consts.push_back((Index << 2) | Chan);
596 return fitsConstReadLimitations(Consts);
599 DFAPacketizer *R600InstrInfo::CreateTargetScheduleState(const TargetMachine *TM,
600 const ScheduleDAG *DAG) const {
601 const InstrItineraryData *II = TM->getInstrItineraryData();
602 return TM->getSubtarget<AMDGPUSubtarget>().createDFAPacketizer(II);
606 isPredicateSetter(unsigned Opcode) {
615 static MachineInstr *
616 findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
617 MachineBasicBlock::iterator I) {
618 while (I != MBB.begin()) {
620 MachineInstr *MI = I;
621 if (isPredicateSetter(MI->getOpcode()))
629 bool isJump(unsigned Opcode) {
630 return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
634 R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
635 MachineBasicBlock *&TBB,
636 MachineBasicBlock *&FBB,
637 SmallVectorImpl<MachineOperand> &Cond,
638 bool AllowModify) const {
639 // Most of the following comes from the ARM implementation of AnalyzeBranch
641 // If the block has no terminators, it just falls into the block after it.
642 MachineBasicBlock::iterator I = MBB.end();
643 if (I == MBB.begin())
646 while (I->isDebugValue()) {
647 if (I == MBB.begin())
651 if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) {
655 // Get the last instruction in the block.
656 MachineInstr *LastInst = I;
658 // If there is only one terminator instruction, process it.
659 unsigned LastOpc = LastInst->getOpcode();
660 if (I == MBB.begin() ||
661 !isJump(static_cast<MachineInstr *>(--I)->getOpcode())) {
662 if (LastOpc == AMDGPU::JUMP) {
663 TBB = LastInst->getOperand(0).getMBB();
665 } else if (LastOpc == AMDGPU::JUMP_COND) {
666 MachineInstr *predSet = I;
667 while (!isPredicateSetter(predSet->getOpcode())) {
670 TBB = LastInst->getOperand(0).getMBB();
671 Cond.push_back(predSet->getOperand(1));
672 Cond.push_back(predSet->getOperand(2));
673 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
676 return true; // Can't handle indirect branch.
679 // Get the instruction before it if it is a terminator.
680 MachineInstr *SecondLastInst = I;
681 unsigned SecondLastOpc = SecondLastInst->getOpcode();
683 // If the block ends with a B and a Bcc, handle it.
684 if (SecondLastOpc == AMDGPU::JUMP_COND && LastOpc == AMDGPU::JUMP) {
685 MachineInstr *predSet = --I;
686 while (!isPredicateSetter(predSet->getOpcode())) {
689 TBB = SecondLastInst->getOperand(0).getMBB();
690 FBB = LastInst->getOperand(0).getMBB();
691 Cond.push_back(predSet->getOperand(1));
692 Cond.push_back(predSet->getOperand(2));
693 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
697 // Otherwise, can't handle this.
701 int R600InstrInfo::getBranchInstr(const MachineOperand &op) const {
702 const MachineInstr *MI = op.getParent();
704 switch (MI->getDesc().OpInfo->RegClass) {
705 default: // FIXME: fallthrough??
706 case AMDGPU::GPRI32RegClassID: return AMDGPU::BRANCH_COND_i32;
707 case AMDGPU::GPRF32RegClassID: return AMDGPU::BRANCH_COND_f32;
712 MachineBasicBlock::iterator FindLastAluClause(MachineBasicBlock &MBB) {
713 for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend();
715 if (It->getOpcode() == AMDGPU::CF_ALU ||
716 It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE)
717 return llvm::prior(It.base());
723 R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
724 MachineBasicBlock *TBB,
725 MachineBasicBlock *FBB,
726 const SmallVectorImpl<MachineOperand> &Cond,
728 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
732 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB);
735 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
736 assert(PredSet && "No previous predicate !");
737 addFlag(PredSet, 0, MO_FLAG_PUSH);
738 PredSet->getOperand(2).setImm(Cond[1].getImm());
740 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
742 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
743 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
744 if (CfAlu == MBB.end())
746 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
747 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
751 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
752 assert(PredSet && "No previous predicate !");
753 addFlag(PredSet, 0, MO_FLAG_PUSH);
754 PredSet->getOperand(2).setImm(Cond[1].getImm());
755 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
757 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
758 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB);
759 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
760 if (CfAlu == MBB.end())
762 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
763 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
769 R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
771 // Note : we leave PRED* instructions there.
772 // They may be needed when predicating instructions.
774 MachineBasicBlock::iterator I = MBB.end();
776 if (I == MBB.begin()) {
780 switch (I->getOpcode()) {
783 case AMDGPU::JUMP_COND: {
784 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
785 clearFlag(predSet, 0, MO_FLAG_PUSH);
786 I->eraseFromParent();
787 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
788 if (CfAlu == MBB.end())
790 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
791 CfAlu->setDesc(get(AMDGPU::CF_ALU));
795 I->eraseFromParent();
800 if (I == MBB.begin()) {
804 switch (I->getOpcode()) {
805 // FIXME: only one case??
808 case AMDGPU::JUMP_COND: {
809 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
810 clearFlag(predSet, 0, MO_FLAG_PUSH);
811 I->eraseFromParent();
812 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
813 if (CfAlu == MBB.end())
815 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
816 CfAlu->setDesc(get(AMDGPU::CF_ALU));
820 I->eraseFromParent();
827 R600InstrInfo::isPredicated(const MachineInstr *MI) const {
828 int idx = MI->findFirstPredOperandIdx();
832 unsigned Reg = MI->getOperand(idx).getReg();
834 default: return false;
835 case AMDGPU::PRED_SEL_ONE:
836 case AMDGPU::PRED_SEL_ZERO:
837 case AMDGPU::PREDICATE_BIT:
843 R600InstrInfo::isPredicable(MachineInstr *MI) const {
844 // XXX: KILL* instructions can be predicated, but they must be the last
845 // instruction in a clause, so this means any instructions after them cannot
846 // be predicated. Until we have proper support for instruction clauses in the
847 // backend, we will mark KILL* instructions as unpredicable.
849 if (MI->getOpcode() == AMDGPU::KILLGT) {
851 } else if (MI->getOpcode() == AMDGPU::CF_ALU) {
852 // If the clause start in the middle of MBB then the MBB has more
853 // than a single clause, unable to predicate several clauses.
854 if (MI->getParent()->begin() != MachineBasicBlock::iterator(MI))
856 // TODO: We don't support KC merging atm
857 if (MI->getOperand(3).getImm() != 0 || MI->getOperand(4).getImm() != 0)
860 } else if (isVector(*MI)) {
863 return AMDGPUInstrInfo::isPredicable(MI);
869 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
871 unsigned ExtraPredCycles,
872 const BranchProbability &Probability) const{
877 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
879 unsigned ExtraTCycles,
880 MachineBasicBlock &FMBB,
882 unsigned ExtraFCycles,
883 const BranchProbability &Probability) const {
888 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
890 const BranchProbability &Probability)
896 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
897 MachineBasicBlock &FMBB) const {
903 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
904 MachineOperand &MO = Cond[1];
905 switch (MO.getImm()) {
906 case OPCODE_IS_ZERO_INT:
907 MO.setImm(OPCODE_IS_NOT_ZERO_INT);
909 case OPCODE_IS_NOT_ZERO_INT:
910 MO.setImm(OPCODE_IS_ZERO_INT);
913 MO.setImm(OPCODE_IS_NOT_ZERO);
915 case OPCODE_IS_NOT_ZERO:
916 MO.setImm(OPCODE_IS_ZERO);
922 MachineOperand &MO2 = Cond[2];
923 switch (MO2.getReg()) {
924 case AMDGPU::PRED_SEL_ZERO:
925 MO2.setReg(AMDGPU::PRED_SEL_ONE);
927 case AMDGPU::PRED_SEL_ONE:
928 MO2.setReg(AMDGPU::PRED_SEL_ZERO);
937 R600InstrInfo::DefinesPredicate(MachineInstr *MI,
938 std::vector<MachineOperand> &Pred) const {
939 return isPredicateSetter(MI->getOpcode());
944 R600InstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
945 const SmallVectorImpl<MachineOperand> &Pred2) const {
951 R600InstrInfo::PredicateInstruction(MachineInstr *MI,
952 const SmallVectorImpl<MachineOperand> &Pred) const {
953 int PIdx = MI->findFirstPredOperandIdx();
955 if (MI->getOpcode() == AMDGPU::CF_ALU) {
956 MI->getOperand(8).setImm(0);
961 MachineOperand &PMO = MI->getOperand(PIdx);
962 PMO.setReg(Pred[2].getReg());
963 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
964 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
971 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
972 const MachineInstr *MI,
973 unsigned *PredCost) const {
979 int R600InstrInfo::getIndirectIndexBegin(const MachineFunction &MF) const {
980 const MachineRegisterInfo &MRI = MF.getRegInfo();
981 const MachineFrameInfo *MFI = MF.getFrameInfo();
984 if (MFI->getNumObjects() == 0) {
988 if (MRI.livein_empty()) {
992 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
993 LE = MRI.livein_end();
995 Offset = std::max(Offset,
996 GET_REG_INDEX(RI.getEncodingValue(LI->first)));
1002 int R600InstrInfo::getIndirectIndexEnd(const MachineFunction &MF) const {
1004 const MachineFrameInfo *MFI = MF.getFrameInfo();
1006 // Variable sized objects are not supported
1007 assert(!MFI->hasVarSizedObjects());
1009 if (MFI->getNumObjects() == 0) {
1013 Offset = TM.getFrameLowering()->getFrameIndexOffset(MF, -1);
1015 return getIndirectIndexBegin(MF) + Offset;
1018 std::vector<unsigned> R600InstrInfo::getIndirectReservedRegs(
1019 const MachineFunction &MF) const {
1020 const AMDGPUFrameLowering *TFL =
1021 static_cast<const AMDGPUFrameLowering*>(TM.getFrameLowering());
1022 std::vector<unsigned> Regs;
1024 unsigned StackWidth = TFL->getStackWidth(MF);
1025 int End = getIndirectIndexEnd(MF);
1031 for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
1032 unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
1033 Regs.push_back(SuperReg);
1034 for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
1035 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
1036 Regs.push_back(Reg);
1042 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
1043 unsigned Channel) const {
1044 // XXX: Remove when we support a stack width > 2
1045 assert(Channel == 0);
1049 const TargetRegisterClass * R600InstrInfo::getIndirectAddrStoreRegClass(
1050 unsigned SourceReg) const {
1051 return &AMDGPU::R600_TReg32RegClass;
1054 const TargetRegisterClass *R600InstrInfo::getIndirectAddrLoadRegClass() const {
1055 return &AMDGPU::TRegMemRegClass;
1058 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1059 MachineBasicBlock::iterator I,
1060 unsigned ValueReg, unsigned Address,
1061 unsigned OffsetReg) const {
1062 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1063 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1064 AMDGPU::AR_X, OffsetReg);
1065 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1067 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1069 .addReg(AMDGPU::AR_X,
1070 RegState::Implicit | RegState::Kill);
1071 setImmOperand(Mov, AMDGPU::OpName::dst_rel, 1);
1075 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1076 MachineBasicBlock::iterator I,
1077 unsigned ValueReg, unsigned Address,
1078 unsigned OffsetReg) const {
1079 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1080 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1083 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1084 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1087 .addReg(AMDGPU::AR_X,
1088 RegState::Implicit | RegState::Kill);
1089 setImmOperand(Mov, AMDGPU::OpName::src0_rel, 1);
1094 const TargetRegisterClass *R600InstrInfo::getSuperIndirectRegClass() const {
1095 return &AMDGPU::IndirectRegRegClass;
1098 unsigned R600InstrInfo::getMaxAlusPerClause() const {
1102 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
1103 MachineBasicBlock::iterator I,
1107 unsigned Src1Reg) const {
1108 MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
1112 MIB.addImm(0) // $update_exec_mask
1113 .addImm(0); // $update_predicate
1115 MIB.addImm(1) // $write
1117 .addImm(0) // $dst_rel
1118 .addImm(0) // $dst_clamp
1119 .addReg(Src0Reg) // $src0
1120 .addImm(0) // $src0_neg
1121 .addImm(0) // $src0_rel
1122 .addImm(0) // $src0_abs
1123 .addImm(-1); // $src0_sel
1126 MIB.addReg(Src1Reg) // $src1
1127 .addImm(0) // $src1_neg
1128 .addImm(0) // $src1_rel
1129 .addImm(0) // $src1_abs
1130 .addImm(-1); // $src1_sel
1133 //XXX: The r600g finalizer expects this to be 1, once we've moved the
1134 //scheduling to the backend, we can change the default to 0.
1135 MIB.addImm(1) // $last
1136 .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
1137 .addImm(0) // $literal
1138 .addImm(0); // $bank_swizzle
1143 #define OPERAND_CASE(Label) \
1145 static const unsigned Ops[] = \
1155 static unsigned getSlotedOps(unsigned Op, unsigned Slot) {
1157 OPERAND_CASE(AMDGPU::OpName::update_exec_mask)
1158 OPERAND_CASE(AMDGPU::OpName::update_pred)
1159 OPERAND_CASE(AMDGPU::OpName::write)
1160 OPERAND_CASE(AMDGPU::OpName::omod)
1161 OPERAND_CASE(AMDGPU::OpName::dst_rel)
1162 OPERAND_CASE(AMDGPU::OpName::clamp)
1163 OPERAND_CASE(AMDGPU::OpName::src0)
1164 OPERAND_CASE(AMDGPU::OpName::src0_neg)
1165 OPERAND_CASE(AMDGPU::OpName::src0_rel)
1166 OPERAND_CASE(AMDGPU::OpName::src0_abs)
1167 OPERAND_CASE(AMDGPU::OpName::src0_sel)
1168 OPERAND_CASE(AMDGPU::OpName::src1)
1169 OPERAND_CASE(AMDGPU::OpName::src1_neg)
1170 OPERAND_CASE(AMDGPU::OpName::src1_rel)
1171 OPERAND_CASE(AMDGPU::OpName::src1_abs)
1172 OPERAND_CASE(AMDGPU::OpName::src1_sel)
1173 OPERAND_CASE(AMDGPU::OpName::pred_sel)
1175 llvm_unreachable("Wrong Operand");
1181 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
1182 MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg)
1184 assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
1186 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
1187 if (ST.getGeneration() <= AMDGPUSubtarget::R700)
1188 Opcode = AMDGPU::DOT4_r600;
1190 Opcode = AMDGPU::DOT4_eg;
1191 MachineBasicBlock::iterator I = MI;
1192 MachineOperand &Src0 = MI->getOperand(
1193 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src0, Slot)));
1194 MachineOperand &Src1 = MI->getOperand(
1195 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src1, Slot)));
1196 MachineInstr *MIB = buildDefaultInstruction(
1197 MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg());
1198 static const unsigned Operands[14] = {
1199 AMDGPU::OpName::update_exec_mask,
1200 AMDGPU::OpName::update_pred,
1201 AMDGPU::OpName::write,
1202 AMDGPU::OpName::omod,
1203 AMDGPU::OpName::dst_rel,
1204 AMDGPU::OpName::clamp,
1205 AMDGPU::OpName::src0_neg,
1206 AMDGPU::OpName::src0_rel,
1207 AMDGPU::OpName::src0_abs,
1208 AMDGPU::OpName::src0_sel,
1209 AMDGPU::OpName::src1_neg,
1210 AMDGPU::OpName::src1_rel,
1211 AMDGPU::OpName::src1_abs,
1212 AMDGPU::OpName::src1_sel,
1215 for (unsigned i = 0; i < 14; i++) {
1216 MachineOperand &MO = MI->getOperand(
1217 getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));
1218 assert (MO.isImm());
1219 setImmOperand(MIB, Operands[i], MO.getImm());
1221 MIB->getOperand(20).setImm(0);
1225 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
1226 MachineBasicBlock::iterator I,
1228 uint64_t Imm) const {
1229 MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
1230 AMDGPU::ALU_LITERAL_X);
1231 setImmOperand(MovImm, AMDGPU::OpName::literal, Imm);
1235 int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const {
1236 return getOperandIdx(MI.getOpcode(), Op);
1239 int R600InstrInfo::getOperandIdx(unsigned Opcode, unsigned Op) const {
1240 return AMDGPU::getNamedOperandIdx(Opcode, Op);
1243 void R600InstrInfo::setImmOperand(MachineInstr *MI, unsigned Op,
1244 int64_t Imm) const {
1245 int Idx = getOperandIdx(*MI, Op);
1246 assert(Idx != -1 && "Operand not supported for this instruction.");
1247 assert(MI->getOperand(Idx).isImm());
1248 MI->getOperand(Idx).setImm(Imm);
1251 //===----------------------------------------------------------------------===//
1252 // Instruction flag getters/setters
1253 //===----------------------------------------------------------------------===//
1255 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
1256 return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
1259 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
1260 unsigned Flag) const {
1261 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1264 // If we pass something other than the default value of Flag to this
1265 // function, it means we are want to set a flag on an instruction
1266 // that uses native encoding.
1267 assert(HAS_NATIVE_OPERANDS(TargetFlags));
1268 bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
1271 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::clamp);
1274 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::write);
1276 case MO_FLAG_NOT_LAST:
1278 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::last);
1282 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_neg); break;
1283 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_neg); break;
1284 case 2: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src2_neg); break;
1289 assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
1293 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_abs); break;
1294 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_abs); break;
1302 assert(FlagIndex != -1 && "Flag not supported for this instruction");
1304 FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
1305 assert(FlagIndex != 0 &&
1306 "Instruction flags not supported for this instruction");
1309 MachineOperand &FlagOp = MI->getOperand(FlagIndex);
1310 assert(FlagOp.isImm());
1314 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
1315 unsigned Flag) const {
1316 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1320 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1321 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1322 if (Flag == MO_FLAG_NOT_LAST) {
1323 clearFlag(MI, Operand, MO_FLAG_LAST);
1324 } else if (Flag == MO_FLAG_MASK) {
1325 clearFlag(MI, Operand, Flag);
1330 MachineOperand &FlagOp = getFlagOp(MI, Operand);
1331 FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
1335 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
1336 unsigned Flag) const {
1337 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1338 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1339 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1342 MachineOperand &FlagOp = getFlagOp(MI);
1343 unsigned InstFlags = FlagOp.getImm();
1344 InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
1345 FlagOp.setImm(InstFlags);