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) {
339 if (Src[0] == Src[1])
342 case R600InstrInfo::ALU_VEC_012_SCL_210:
344 case R600InstrInfo::ALU_VEC_021_SCL_122:
345 std::swap(Src[1], Src[2]);
347 case R600InstrInfo::ALU_VEC_102_SCL_221:
348 std::swap(Src[0], Src[1]);
350 case R600InstrInfo::ALU_VEC_120_SCL_212:
351 std::swap(Src[0], Src[1]);
352 std::swap(Src[0], Src[2]);
354 case R600InstrInfo::ALU_VEC_201:
355 std::swap(Src[0], Src[2]);
356 std::swap(Src[0], Src[1]);
358 case R600InstrInfo::ALU_VEC_210:
359 std::swap(Src[0], Src[2]);
366 getTransSwizzle(R600InstrInfo::BankSwizzle Swz, unsigned Op) {
368 case R600InstrInfo::ALU_VEC_012_SCL_210: {
369 unsigned Cycles[3] = { 2, 1, 0};
372 case R600InstrInfo::ALU_VEC_021_SCL_122: {
373 unsigned Cycles[3] = { 1, 2, 2};
376 case R600InstrInfo::ALU_VEC_120_SCL_212: {
377 unsigned Cycles[3] = { 2, 1, 2};
380 case R600InstrInfo::ALU_VEC_102_SCL_221: {
381 unsigned Cycles[3] = { 2, 2, 1};
385 llvm_unreachable("Wrong Swizzle for Trans Slot");
390 /// returns how many MIs (whose inputs are represented by IGSrcs) can be packed
391 /// in the same Instruction Group while meeting read port limitations given a
392 /// Swz swizzle sequence.
393 unsigned R600InstrInfo::isLegalUpTo(
394 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
395 const std::vector<R600InstrInfo::BankSwizzle> &Swz,
396 const std::vector<std::pair<int, unsigned> > &TransSrcs,
397 R600InstrInfo::BankSwizzle TransSwz) const {
399 memset(Vector, -1, sizeof(Vector));
400 for (unsigned i = 0, e = IGSrcs.size(); i < e; i++) {
401 const std::vector<std::pair<int, unsigned> > &Srcs =
402 Swizzle(IGSrcs[i], Swz[i]);
403 for (unsigned j = 0; j < 3; j++) {
404 const std::pair<int, unsigned> &Src = Srcs[j];
405 if (Src.first < 0 || Src.first == 255)
407 if (Src.first == GET_REG_INDEX(RI.getEncodingValue(AMDGPU::OQAP))) {
408 if (Swz[i] != R600InstrInfo::ALU_VEC_012_SCL_210 &&
409 Swz[i] != R600InstrInfo::ALU_VEC_021_SCL_122) {
410 // The value from output queue A (denoted by register OQAP) can
411 // only be fetched during the first cycle.
414 // OQAP does not count towards the normal read port restrictions
417 if (Vector[Src.second][j] < 0)
418 Vector[Src.second][j] = Src.first;
419 if (Vector[Src.second][j] != Src.first)
423 // Now check Trans Alu
424 for (unsigned i = 0, e = TransSrcs.size(); i < e; ++i) {
425 const std::pair<int, unsigned> &Src = TransSrcs[i];
426 unsigned Cycle = getTransSwizzle(TransSwz, i);
429 if (Src.first == 255)
431 if (Vector[Src.second][Cycle] < 0)
432 Vector[Src.second][Cycle] = Src.first;
433 if (Vector[Src.second][Cycle] != Src.first)
434 return IGSrcs.size() - 1;
436 return IGSrcs.size();
439 /// Given a swizzle sequence SwzCandidate and an index Idx, returns the next
440 /// (in lexicographic term) swizzle sequence assuming that all swizzles after
441 /// Idx can be skipped
443 NextPossibleSolution(
444 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
446 assert(Idx < SwzCandidate.size());
448 while (ResetIdx > -1 && SwzCandidate[ResetIdx] == R600InstrInfo::ALU_VEC_210)
450 for (unsigned i = ResetIdx + 1, e = SwzCandidate.size(); i < e; i++) {
451 SwzCandidate[i] = R600InstrInfo::ALU_VEC_012_SCL_210;
455 int NextSwizzle = SwzCandidate[ResetIdx] + 1;
456 SwzCandidate[ResetIdx] = (R600InstrInfo::BankSwizzle)NextSwizzle;
460 /// Enumerate all possible Swizzle sequence to find one that can meet all
461 /// read port requirements.
462 bool R600InstrInfo::FindSwizzleForVectorSlot(
463 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
464 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
465 const std::vector<std::pair<int, unsigned> > &TransSrcs,
466 R600InstrInfo::BankSwizzle TransSwz) const {
467 unsigned ValidUpTo = 0;
469 ValidUpTo = isLegalUpTo(IGSrcs, SwzCandidate, TransSrcs, TransSwz);
470 if (ValidUpTo == IGSrcs.size())
472 } while (NextPossibleSolution(SwzCandidate, ValidUpTo));
476 /// Instructions in Trans slot can't read gpr at cycle 0 if they also read
477 /// a const, and can't read a gpr at cycle 1 if they read 2 const.
479 isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,
480 const std::vector<std::pair<int, unsigned> > &TransOps,
481 unsigned ConstCount) {
482 // TransALU can't read 3 constants
485 for (unsigned i = 0, e = TransOps.size(); i < e; ++i) {
486 const std::pair<int, unsigned> &Src = TransOps[i];
487 unsigned Cycle = getTransSwizzle(TransSwz, i);
490 if (ConstCount > 0 && Cycle == 0)
492 if (ConstCount > 1 && Cycle == 1)
499 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG,
500 const DenseMap<unsigned, unsigned> &PV,
501 std::vector<BankSwizzle> &ValidSwizzle,
504 //Todo : support shared src0 - src1 operand
506 std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs;
507 ValidSwizzle.clear();
509 BankSwizzle TransBS = ALU_VEC_012_SCL_210;
510 for (unsigned i = 0, e = IG.size(); i < e; ++i) {
511 IGSrcs.push_back(ExtractSrcs(IG[i], PV, ConstCount));
512 unsigned Op = getOperandIdx(IG[i]->getOpcode(),
513 AMDGPU::OpName::bank_swizzle);
514 ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle)
515 IG[i]->getOperand(Op).getImm());
517 std::vector<std::pair<int, unsigned> > TransOps;
519 return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS);
521 TransOps = IGSrcs.back();
523 ValidSwizzle.pop_back();
525 static const R600InstrInfo::BankSwizzle TransSwz[] = {
531 for (unsigned i = 0; i < 4; i++) {
532 TransBS = TransSwz[i];
533 if (!isConstCompatible(TransBS, TransOps, ConstCount))
535 bool Result = FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps,
538 ValidSwizzle.push_back(TransBS);
548 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts)
550 assert (Consts.size() <= 12 && "Too many operands in instructions group");
551 unsigned Pair1 = 0, Pair2 = 0;
552 for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
553 unsigned ReadConstHalf = Consts[i] & 2;
554 unsigned ReadConstIndex = Consts[i] & (~3);
555 unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf;
557 Pair1 = ReadHalfConst;
560 if (Pair1 == ReadHalfConst)
563 Pair2 = ReadHalfConst;
566 if (Pair2 != ReadHalfConst)
573 R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
575 std::vector<unsigned> Consts;
576 SmallSet<int64_t, 4> Literals;
577 for (unsigned i = 0, n = MIs.size(); i < n; i++) {
578 MachineInstr *MI = MIs[i];
579 if (!isALUInstr(MI->getOpcode()))
582 const SmallVectorImpl<std::pair<MachineOperand *, int64_t> > &Srcs =
585 for (unsigned j = 0, e = Srcs.size(); j < e; j++) {
586 std::pair<MachineOperand *, unsigned> Src = Srcs[j];
587 if (Src.first->getReg() == AMDGPU::ALU_LITERAL_X)
588 Literals.insert(Src.second);
589 if (Literals.size() > 4)
591 if (Src.first->getReg() == AMDGPU::ALU_CONST)
592 Consts.push_back(Src.second);
593 if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) ||
594 AMDGPU::R600_KC1RegClass.contains(Src.first->getReg())) {
595 unsigned Index = RI.getEncodingValue(Src.first->getReg()) & 0xff;
596 unsigned Chan = RI.getHWRegChan(Src.first->getReg());
597 Consts.push_back((Index << 2) | Chan);
601 return fitsConstReadLimitations(Consts);
604 DFAPacketizer *R600InstrInfo::CreateTargetScheduleState(const TargetMachine *TM,
605 const ScheduleDAG *DAG) const {
606 const InstrItineraryData *II = TM->getInstrItineraryData();
607 return TM->getSubtarget<AMDGPUSubtarget>().createDFAPacketizer(II);
611 isPredicateSetter(unsigned Opcode) {
620 static MachineInstr *
621 findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
622 MachineBasicBlock::iterator I) {
623 while (I != MBB.begin()) {
625 MachineInstr *MI = I;
626 if (isPredicateSetter(MI->getOpcode()))
634 bool isJump(unsigned Opcode) {
635 return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
639 R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
640 MachineBasicBlock *&TBB,
641 MachineBasicBlock *&FBB,
642 SmallVectorImpl<MachineOperand> &Cond,
643 bool AllowModify) const {
644 // Most of the following comes from the ARM implementation of AnalyzeBranch
646 // If the block has no terminators, it just falls into the block after it.
647 MachineBasicBlock::iterator I = MBB.end();
648 if (I == MBB.begin())
651 while (I->isDebugValue()) {
652 if (I == MBB.begin())
656 if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) {
660 // Get the last instruction in the block.
661 MachineInstr *LastInst = I;
663 // If there is only one terminator instruction, process it.
664 unsigned LastOpc = LastInst->getOpcode();
665 if (I == MBB.begin() ||
666 !isJump(static_cast<MachineInstr *>(--I)->getOpcode())) {
667 if (LastOpc == AMDGPU::JUMP) {
668 TBB = LastInst->getOperand(0).getMBB();
670 } else if (LastOpc == AMDGPU::JUMP_COND) {
671 MachineInstr *predSet = I;
672 while (!isPredicateSetter(predSet->getOpcode())) {
675 TBB = LastInst->getOperand(0).getMBB();
676 Cond.push_back(predSet->getOperand(1));
677 Cond.push_back(predSet->getOperand(2));
678 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
681 return true; // Can't handle indirect branch.
684 // Get the instruction before it if it is a terminator.
685 MachineInstr *SecondLastInst = I;
686 unsigned SecondLastOpc = SecondLastInst->getOpcode();
688 // If the block ends with a B and a Bcc, handle it.
689 if (SecondLastOpc == AMDGPU::JUMP_COND && LastOpc == AMDGPU::JUMP) {
690 MachineInstr *predSet = --I;
691 while (!isPredicateSetter(predSet->getOpcode())) {
694 TBB = SecondLastInst->getOperand(0).getMBB();
695 FBB = LastInst->getOperand(0).getMBB();
696 Cond.push_back(predSet->getOperand(1));
697 Cond.push_back(predSet->getOperand(2));
698 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
702 // Otherwise, can't handle this.
706 int R600InstrInfo::getBranchInstr(const MachineOperand &op) const {
707 const MachineInstr *MI = op.getParent();
709 switch (MI->getDesc().OpInfo->RegClass) {
710 default: // FIXME: fallthrough??
711 case AMDGPU::GPRI32RegClassID: return AMDGPU::BRANCH_COND_i32;
712 case AMDGPU::GPRF32RegClassID: return AMDGPU::BRANCH_COND_f32;
717 MachineBasicBlock::iterator FindLastAluClause(MachineBasicBlock &MBB) {
718 for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend();
720 if (It->getOpcode() == AMDGPU::CF_ALU ||
721 It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE)
722 return llvm::prior(It.base());
728 R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
729 MachineBasicBlock *TBB,
730 MachineBasicBlock *FBB,
731 const SmallVectorImpl<MachineOperand> &Cond,
733 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
737 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB);
740 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
741 assert(PredSet && "No previous predicate !");
742 addFlag(PredSet, 0, MO_FLAG_PUSH);
743 PredSet->getOperand(2).setImm(Cond[1].getImm());
745 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
747 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
748 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
749 if (CfAlu == MBB.end())
751 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
752 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
756 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
757 assert(PredSet && "No previous predicate !");
758 addFlag(PredSet, 0, MO_FLAG_PUSH);
759 PredSet->getOperand(2).setImm(Cond[1].getImm());
760 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
762 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
763 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB);
764 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
765 if (CfAlu == MBB.end())
767 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
768 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
774 R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
776 // Note : we leave PRED* instructions there.
777 // They may be needed when predicating instructions.
779 MachineBasicBlock::iterator I = MBB.end();
781 if (I == MBB.begin()) {
785 switch (I->getOpcode()) {
788 case AMDGPU::JUMP_COND: {
789 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
790 clearFlag(predSet, 0, MO_FLAG_PUSH);
791 I->eraseFromParent();
792 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
793 if (CfAlu == MBB.end())
795 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
796 CfAlu->setDesc(get(AMDGPU::CF_ALU));
800 I->eraseFromParent();
805 if (I == MBB.begin()) {
809 switch (I->getOpcode()) {
810 // FIXME: only one case??
813 case AMDGPU::JUMP_COND: {
814 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
815 clearFlag(predSet, 0, MO_FLAG_PUSH);
816 I->eraseFromParent();
817 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
818 if (CfAlu == MBB.end())
820 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
821 CfAlu->setDesc(get(AMDGPU::CF_ALU));
825 I->eraseFromParent();
832 R600InstrInfo::isPredicated(const MachineInstr *MI) const {
833 int idx = MI->findFirstPredOperandIdx();
837 unsigned Reg = MI->getOperand(idx).getReg();
839 default: return false;
840 case AMDGPU::PRED_SEL_ONE:
841 case AMDGPU::PRED_SEL_ZERO:
842 case AMDGPU::PREDICATE_BIT:
848 R600InstrInfo::isPredicable(MachineInstr *MI) const {
849 // XXX: KILL* instructions can be predicated, but they must be the last
850 // instruction in a clause, so this means any instructions after them cannot
851 // be predicated. Until we have proper support for instruction clauses in the
852 // backend, we will mark KILL* instructions as unpredicable.
854 if (MI->getOpcode() == AMDGPU::KILLGT) {
856 } else if (MI->getOpcode() == AMDGPU::CF_ALU) {
857 // If the clause start in the middle of MBB then the MBB has more
858 // than a single clause, unable to predicate several clauses.
859 if (MI->getParent()->begin() != MachineBasicBlock::iterator(MI))
861 // TODO: We don't support KC merging atm
862 if (MI->getOperand(3).getImm() != 0 || MI->getOperand(4).getImm() != 0)
865 } else if (isVector(*MI)) {
868 return AMDGPUInstrInfo::isPredicable(MI);
874 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
876 unsigned ExtraPredCycles,
877 const BranchProbability &Probability) const{
882 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
884 unsigned ExtraTCycles,
885 MachineBasicBlock &FMBB,
887 unsigned ExtraFCycles,
888 const BranchProbability &Probability) const {
893 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
895 const BranchProbability &Probability)
901 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
902 MachineBasicBlock &FMBB) const {
908 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
909 MachineOperand &MO = Cond[1];
910 switch (MO.getImm()) {
911 case OPCODE_IS_ZERO_INT:
912 MO.setImm(OPCODE_IS_NOT_ZERO_INT);
914 case OPCODE_IS_NOT_ZERO_INT:
915 MO.setImm(OPCODE_IS_ZERO_INT);
918 MO.setImm(OPCODE_IS_NOT_ZERO);
920 case OPCODE_IS_NOT_ZERO:
921 MO.setImm(OPCODE_IS_ZERO);
927 MachineOperand &MO2 = Cond[2];
928 switch (MO2.getReg()) {
929 case AMDGPU::PRED_SEL_ZERO:
930 MO2.setReg(AMDGPU::PRED_SEL_ONE);
932 case AMDGPU::PRED_SEL_ONE:
933 MO2.setReg(AMDGPU::PRED_SEL_ZERO);
942 R600InstrInfo::DefinesPredicate(MachineInstr *MI,
943 std::vector<MachineOperand> &Pred) const {
944 return isPredicateSetter(MI->getOpcode());
949 R600InstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
950 const SmallVectorImpl<MachineOperand> &Pred2) const {
956 R600InstrInfo::PredicateInstruction(MachineInstr *MI,
957 const SmallVectorImpl<MachineOperand> &Pred) const {
958 int PIdx = MI->findFirstPredOperandIdx();
960 if (MI->getOpcode() == AMDGPU::CF_ALU) {
961 MI->getOperand(8).setImm(0);
966 MachineOperand &PMO = MI->getOperand(PIdx);
967 PMO.setReg(Pred[2].getReg());
968 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
969 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
976 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
977 const MachineInstr *MI,
978 unsigned *PredCost) const {
984 int R600InstrInfo::getIndirectIndexBegin(const MachineFunction &MF) const {
985 const MachineRegisterInfo &MRI = MF.getRegInfo();
986 const MachineFrameInfo *MFI = MF.getFrameInfo();
989 if (MFI->getNumObjects() == 0) {
993 if (MRI.livein_empty()) {
997 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
998 LE = MRI.livein_end();
1000 Offset = std::max(Offset,
1001 GET_REG_INDEX(RI.getEncodingValue(LI->first)));
1007 int R600InstrInfo::getIndirectIndexEnd(const MachineFunction &MF) const {
1009 const MachineFrameInfo *MFI = MF.getFrameInfo();
1011 // Variable sized objects are not supported
1012 assert(!MFI->hasVarSizedObjects());
1014 if (MFI->getNumObjects() == 0) {
1018 Offset = TM.getFrameLowering()->getFrameIndexOffset(MF, -1);
1020 return getIndirectIndexBegin(MF) + Offset;
1023 std::vector<unsigned> R600InstrInfo::getIndirectReservedRegs(
1024 const MachineFunction &MF) const {
1025 const AMDGPUFrameLowering *TFL =
1026 static_cast<const AMDGPUFrameLowering*>(TM.getFrameLowering());
1027 std::vector<unsigned> Regs;
1029 unsigned StackWidth = TFL->getStackWidth(MF);
1030 int End = getIndirectIndexEnd(MF);
1036 for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
1037 unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
1038 Regs.push_back(SuperReg);
1039 for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
1040 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
1041 Regs.push_back(Reg);
1047 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
1048 unsigned Channel) const {
1049 // XXX: Remove when we support a stack width > 2
1050 assert(Channel == 0);
1054 const TargetRegisterClass * R600InstrInfo::getIndirectAddrStoreRegClass(
1055 unsigned SourceReg) const {
1056 return &AMDGPU::R600_TReg32RegClass;
1059 const TargetRegisterClass *R600InstrInfo::getIndirectAddrLoadRegClass() const {
1060 return &AMDGPU::TRegMemRegClass;
1063 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1064 MachineBasicBlock::iterator I,
1065 unsigned ValueReg, unsigned Address,
1066 unsigned OffsetReg) const {
1067 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1068 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1069 AMDGPU::AR_X, OffsetReg);
1070 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1072 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1074 .addReg(AMDGPU::AR_X,
1075 RegState::Implicit | RegState::Kill);
1076 setImmOperand(Mov, AMDGPU::OpName::dst_rel, 1);
1080 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1081 MachineBasicBlock::iterator I,
1082 unsigned ValueReg, unsigned Address,
1083 unsigned OffsetReg) const {
1084 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1085 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1088 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1089 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1092 .addReg(AMDGPU::AR_X,
1093 RegState::Implicit | RegState::Kill);
1094 setImmOperand(Mov, AMDGPU::OpName::src0_rel, 1);
1099 const TargetRegisterClass *R600InstrInfo::getSuperIndirectRegClass() const {
1100 return &AMDGPU::IndirectRegRegClass;
1103 unsigned R600InstrInfo::getMaxAlusPerClause() const {
1107 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
1108 MachineBasicBlock::iterator I,
1112 unsigned Src1Reg) const {
1113 MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
1117 MIB.addImm(0) // $update_exec_mask
1118 .addImm(0); // $update_predicate
1120 MIB.addImm(1) // $write
1122 .addImm(0) // $dst_rel
1123 .addImm(0) // $dst_clamp
1124 .addReg(Src0Reg) // $src0
1125 .addImm(0) // $src0_neg
1126 .addImm(0) // $src0_rel
1127 .addImm(0) // $src0_abs
1128 .addImm(-1); // $src0_sel
1131 MIB.addReg(Src1Reg) // $src1
1132 .addImm(0) // $src1_neg
1133 .addImm(0) // $src1_rel
1134 .addImm(0) // $src1_abs
1135 .addImm(-1); // $src1_sel
1138 //XXX: The r600g finalizer expects this to be 1, once we've moved the
1139 //scheduling to the backend, we can change the default to 0.
1140 MIB.addImm(1) // $last
1141 .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
1142 .addImm(0) // $literal
1143 .addImm(0); // $bank_swizzle
1148 #define OPERAND_CASE(Label) \
1150 static const unsigned Ops[] = \
1160 static unsigned getSlotedOps(unsigned Op, unsigned Slot) {
1162 OPERAND_CASE(AMDGPU::OpName::update_exec_mask)
1163 OPERAND_CASE(AMDGPU::OpName::update_pred)
1164 OPERAND_CASE(AMDGPU::OpName::write)
1165 OPERAND_CASE(AMDGPU::OpName::omod)
1166 OPERAND_CASE(AMDGPU::OpName::dst_rel)
1167 OPERAND_CASE(AMDGPU::OpName::clamp)
1168 OPERAND_CASE(AMDGPU::OpName::src0)
1169 OPERAND_CASE(AMDGPU::OpName::src0_neg)
1170 OPERAND_CASE(AMDGPU::OpName::src0_rel)
1171 OPERAND_CASE(AMDGPU::OpName::src0_abs)
1172 OPERAND_CASE(AMDGPU::OpName::src0_sel)
1173 OPERAND_CASE(AMDGPU::OpName::src1)
1174 OPERAND_CASE(AMDGPU::OpName::src1_neg)
1175 OPERAND_CASE(AMDGPU::OpName::src1_rel)
1176 OPERAND_CASE(AMDGPU::OpName::src1_abs)
1177 OPERAND_CASE(AMDGPU::OpName::src1_sel)
1178 OPERAND_CASE(AMDGPU::OpName::pred_sel)
1180 llvm_unreachable("Wrong Operand");
1186 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
1187 MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg)
1189 assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
1191 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
1192 if (ST.getGeneration() <= AMDGPUSubtarget::R700)
1193 Opcode = AMDGPU::DOT4_r600;
1195 Opcode = AMDGPU::DOT4_eg;
1196 MachineBasicBlock::iterator I = MI;
1197 MachineOperand &Src0 = MI->getOperand(
1198 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src0, Slot)));
1199 MachineOperand &Src1 = MI->getOperand(
1200 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src1, Slot)));
1201 MachineInstr *MIB = buildDefaultInstruction(
1202 MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg());
1203 static const unsigned Operands[14] = {
1204 AMDGPU::OpName::update_exec_mask,
1205 AMDGPU::OpName::update_pred,
1206 AMDGPU::OpName::write,
1207 AMDGPU::OpName::omod,
1208 AMDGPU::OpName::dst_rel,
1209 AMDGPU::OpName::clamp,
1210 AMDGPU::OpName::src0_neg,
1211 AMDGPU::OpName::src0_rel,
1212 AMDGPU::OpName::src0_abs,
1213 AMDGPU::OpName::src0_sel,
1214 AMDGPU::OpName::src1_neg,
1215 AMDGPU::OpName::src1_rel,
1216 AMDGPU::OpName::src1_abs,
1217 AMDGPU::OpName::src1_sel,
1220 for (unsigned i = 0; i < 14; i++) {
1221 MachineOperand &MO = MI->getOperand(
1222 getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));
1223 assert (MO.isImm());
1224 setImmOperand(MIB, Operands[i], MO.getImm());
1226 MIB->getOperand(20).setImm(0);
1230 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
1231 MachineBasicBlock::iterator I,
1233 uint64_t Imm) const {
1234 MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
1235 AMDGPU::ALU_LITERAL_X);
1236 setImmOperand(MovImm, AMDGPU::OpName::literal, Imm);
1240 int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const {
1241 return getOperandIdx(MI.getOpcode(), Op);
1244 int R600InstrInfo::getOperandIdx(unsigned Opcode, unsigned Op) const {
1245 return AMDGPU::getNamedOperandIdx(Opcode, Op);
1248 void R600InstrInfo::setImmOperand(MachineInstr *MI, unsigned Op,
1249 int64_t Imm) const {
1250 int Idx = getOperandIdx(*MI, Op);
1251 assert(Idx != -1 && "Operand not supported for this instruction.");
1252 assert(MI->getOperand(Idx).isImm());
1253 MI->getOperand(Idx).setImm(Imm);
1256 //===----------------------------------------------------------------------===//
1257 // Instruction flag getters/setters
1258 //===----------------------------------------------------------------------===//
1260 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
1261 return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
1264 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
1265 unsigned Flag) const {
1266 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1269 // If we pass something other than the default value of Flag to this
1270 // function, it means we are want to set a flag on an instruction
1271 // that uses native encoding.
1272 assert(HAS_NATIVE_OPERANDS(TargetFlags));
1273 bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
1276 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::clamp);
1279 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::write);
1281 case MO_FLAG_NOT_LAST:
1283 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::last);
1287 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_neg); break;
1288 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_neg); break;
1289 case 2: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src2_neg); break;
1294 assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
1298 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_abs); break;
1299 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_abs); break;
1307 assert(FlagIndex != -1 && "Flag not supported for this instruction");
1309 FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
1310 assert(FlagIndex != 0 &&
1311 "Instruction flags not supported for this instruction");
1314 MachineOperand &FlagOp = MI->getOperand(FlagIndex);
1315 assert(FlagOp.isImm());
1319 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
1320 unsigned Flag) const {
1321 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1325 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1326 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1327 if (Flag == MO_FLAG_NOT_LAST) {
1328 clearFlag(MI, Operand, MO_FLAG_LAST);
1329 } else if (Flag == MO_FLAG_MASK) {
1330 clearFlag(MI, Operand, Flag);
1335 MachineOperand &FlagOp = getFlagOp(MI, Operand);
1336 FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
1340 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
1341 unsigned Flag) const {
1342 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1343 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1344 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1347 MachineOperand &FlagOp = getFlagOp(MI);
1348 unsigned InstFlags = FlagOp.getImm();
1349 InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
1350 FlagOp.setImm(InstFlags);