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 return (get(Opcode).TSFlags & R600_InstFlag::TRANS_ONLY);
156 bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const {
157 return isTransOnly(MI->getOpcode());
160 bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
161 return ST.hasVertexCache() && IS_VTX(get(Opcode));
164 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const {
165 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
166 return MFI->ShaderType != ShaderType::COMPUTE && usesVertexCache(MI->getOpcode());
169 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
170 return (!ST.hasVertexCache() && IS_VTX(get(Opcode))) || IS_TEX(get(Opcode));
173 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
174 const R600MachineFunctionInfo *MFI = MI->getParent()->getParent()->getInfo<R600MachineFunctionInfo>();
175 return (MFI->ShaderType == ShaderType::COMPUTE && usesVertexCache(MI->getOpcode())) ||
176 usesTextureCache(MI->getOpcode());
179 bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
182 case AMDGPU::GROUP_BARRIER:
189 int R600InstrInfo::getSrcIdx(unsigned Opcode, unsigned SrcNum) const {
190 static const unsigned OpTable[] = {
191 AMDGPU::OpName::src0,
192 AMDGPU::OpName::src1,
197 return getOperandIdx(Opcode, OpTable[SrcNum]);
200 #define SRC_SEL_ROWS 11
201 int R600InstrInfo::getSelIdx(unsigned Opcode, unsigned SrcIdx) const {
202 static const unsigned SrcSelTable[SRC_SEL_ROWS][2] = {
203 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
204 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
205 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
206 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
207 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
208 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
209 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
210 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
211 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
212 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
213 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W}
216 for (unsigned i = 0; i < SRC_SEL_ROWS; ++i) {
217 if (getOperandIdx(Opcode, SrcSelTable[i][0]) == (int)SrcIdx) {
218 return getOperandIdx(Opcode, SrcSelTable[i][1]);
225 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
226 R600InstrInfo::getSrcs(MachineInstr *MI) const {
227 SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
229 if (MI->getOpcode() == AMDGPU::DOT_4) {
230 static const unsigned OpTable[8][2] = {
231 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
232 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
233 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
234 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
235 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
236 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
237 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
238 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W},
241 for (unsigned j = 0; j < 8; j++) {
242 MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
244 unsigned Reg = MO.getReg();
245 if (Reg == AMDGPU::ALU_CONST) {
246 unsigned Sel = MI->getOperand(getOperandIdx(MI->getOpcode(),
247 OpTable[j][1])).getImm();
248 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
256 static const unsigned OpTable[3][2] = {
257 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
258 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
259 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
262 for (unsigned j = 0; j < 3; j++) {
263 int SrcIdx = getOperandIdx(MI->getOpcode(), OpTable[j][0]);
266 MachineOperand &MO = MI->getOperand(SrcIdx);
267 unsigned Reg = MI->getOperand(SrcIdx).getReg();
268 if (Reg == AMDGPU::ALU_CONST) {
269 unsigned Sel = MI->getOperand(
270 getOperandIdx(MI->getOpcode(), OpTable[j][1])).getImm();
271 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
274 if (Reg == AMDGPU::ALU_LITERAL_X) {
275 unsigned Imm = MI->getOperand(
276 getOperandIdx(MI->getOpcode(), AMDGPU::OpName::literal)).getImm();
277 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Imm));
280 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, 0));
285 std::vector<std::pair<int, unsigned> >
286 R600InstrInfo::ExtractSrcs(MachineInstr *MI,
287 const DenseMap<unsigned, unsigned> &PV,
288 unsigned &ConstCount) const {
290 const SmallVector<std::pair<MachineOperand *, int64_t>, 3> Srcs = getSrcs(MI);
291 const std::pair<int, unsigned> DummyPair(-1, 0);
292 std::vector<std::pair<int, unsigned> > Result;
294 for (unsigned n = Srcs.size(); i < n; ++i) {
295 unsigned Reg = Srcs[i].first->getReg();
296 unsigned Index = RI.getEncodingValue(Reg) & 0xff;
297 if (Reg == AMDGPU::OQAP) {
298 Result.push_back(std::pair<int, unsigned>(Index, 0));
300 if (PV.find(Reg) != PV.end()) {
301 // 255 is used to tells its a PS/PV reg
302 Result.push_back(std::pair<int, unsigned>(255, 0));
307 Result.push_back(DummyPair);
310 unsigned Chan = RI.getHWRegChan(Reg);
311 Result.push_back(std::pair<int, unsigned>(Index, Chan));
314 Result.push_back(DummyPair);
318 static std::vector<std::pair<int, unsigned> >
319 Swizzle(std::vector<std::pair<int, unsigned> > Src,
320 R600InstrInfo::BankSwizzle Swz) {
322 case R600InstrInfo::ALU_VEC_012_SCL_210:
324 case R600InstrInfo::ALU_VEC_021_SCL_122:
325 std::swap(Src[1], Src[2]);
327 case R600InstrInfo::ALU_VEC_102_SCL_221:
328 std::swap(Src[0], Src[1]);
330 case R600InstrInfo::ALU_VEC_120_SCL_212:
331 std::swap(Src[0], Src[1]);
332 std::swap(Src[0], Src[2]);
334 case R600InstrInfo::ALU_VEC_201:
335 std::swap(Src[0], Src[2]);
336 std::swap(Src[0], Src[1]);
338 case R600InstrInfo::ALU_VEC_210:
339 std::swap(Src[0], Src[2]);
346 getTransSwizzle(R600InstrInfo::BankSwizzle Swz, unsigned Op) {
348 case R600InstrInfo::ALU_VEC_012_SCL_210: {
349 unsigned Cycles[3] = { 2, 1, 0};
352 case R600InstrInfo::ALU_VEC_021_SCL_122: {
353 unsigned Cycles[3] = { 1, 2, 2};
356 case R600InstrInfo::ALU_VEC_120_SCL_212: {
357 unsigned Cycles[3] = { 2, 1, 2};
360 case R600InstrInfo::ALU_VEC_102_SCL_221: {
361 unsigned Cycles[3] = { 2, 2, 1};
365 llvm_unreachable("Wrong Swizzle for Trans Slot");
370 /// returns how many MIs (whose inputs are represented by IGSrcs) can be packed
371 /// in the same Instruction Group while meeting read port limitations given a
372 /// Swz swizzle sequence.
373 unsigned R600InstrInfo::isLegalUpTo(
374 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
375 const std::vector<R600InstrInfo::BankSwizzle> &Swz,
376 const std::vector<std::pair<int, unsigned> > &TransSrcs,
377 R600InstrInfo::BankSwizzle TransSwz) const {
379 memset(Vector, -1, sizeof(Vector));
380 for (unsigned i = 0, e = IGSrcs.size(); i < e; i++) {
381 const std::vector<std::pair<int, unsigned> > &Srcs =
382 Swizzle(IGSrcs[i], Swz[i]);
383 for (unsigned j = 0; j < 3; j++) {
384 const std::pair<int, unsigned> &Src = Srcs[j];
385 if (Src.first < 0 || Src.first == 255)
387 if (Src.first == GET_REG_INDEX(RI.getEncodingValue(AMDGPU::OQAP))) {
388 if (Swz[i] != R600InstrInfo::ALU_VEC_012_SCL_210 &&
389 Swz[i] != R600InstrInfo::ALU_VEC_021_SCL_122) {
390 // The value from output queue A (denoted by register OQAP) can
391 // only be fetched during the first cycle.
394 // OQAP does not count towards the normal read port restrictions
397 if (Vector[Src.second][j] < 0)
398 Vector[Src.second][j] = Src.first;
399 if (Vector[Src.second][j] != Src.first)
403 // Now check Trans Alu
404 for (unsigned i = 0, e = TransSrcs.size(); i < e; ++i) {
405 const std::pair<int, unsigned> &Src = TransSrcs[i];
406 unsigned Cycle = getTransSwizzle(TransSwz, i);
409 if (Src.first == 255)
411 if (Vector[Src.second][Cycle] < 0)
412 Vector[Src.second][Cycle] = Src.first;
413 if (Vector[Src.second][Cycle] != Src.first)
414 return IGSrcs.size() - 1;
416 return IGSrcs.size();
419 /// Given a swizzle sequence SwzCandidate and an index Idx, returns the next
420 /// (in lexicographic term) swizzle sequence assuming that all swizzles after
421 /// Idx can be skipped
423 NextPossibleSolution(
424 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
426 assert(Idx < SwzCandidate.size());
428 while (ResetIdx > -1 && SwzCandidate[ResetIdx] == R600InstrInfo::ALU_VEC_210)
430 for (unsigned i = ResetIdx + 1, e = SwzCandidate.size(); i < e; i++) {
431 SwzCandidate[i] = R600InstrInfo::ALU_VEC_012_SCL_210;
435 int NextSwizzle = SwzCandidate[ResetIdx] + 1;
436 SwzCandidate[ResetIdx] = (R600InstrInfo::BankSwizzle)NextSwizzle;
440 /// Enumerate all possible Swizzle sequence to find one that can meet all
441 /// read port requirements.
442 bool R600InstrInfo::FindSwizzleForVectorSlot(
443 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
444 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
445 const std::vector<std::pair<int, unsigned> > &TransSrcs,
446 R600InstrInfo::BankSwizzle TransSwz) const {
447 unsigned ValidUpTo = 0;
449 ValidUpTo = isLegalUpTo(IGSrcs, SwzCandidate, TransSrcs, TransSwz);
450 if (ValidUpTo == IGSrcs.size())
452 } while (NextPossibleSolution(SwzCandidate, ValidUpTo));
456 /// Instructions in Trans slot can't read gpr at cycle 0 if they also read
457 /// a const, and can't read a gpr at cycle 1 if they read 2 const.
459 isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,
460 const std::vector<std::pair<int, unsigned> > &TransOps,
461 unsigned ConstCount) {
462 for (unsigned i = 0, e = TransOps.size(); i < e; ++i) {
463 const std::pair<int, unsigned> &Src = TransOps[i];
464 unsigned Cycle = getTransSwizzle(TransSwz, i);
467 if (ConstCount > 0 && Cycle == 0)
469 if (ConstCount > 1 && Cycle == 1)
476 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG,
477 const DenseMap<unsigned, unsigned> &PV,
478 std::vector<BankSwizzle> &ValidSwizzle,
481 //Todo : support shared src0 - src1 operand
483 std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs;
484 ValidSwizzle.clear();
486 BankSwizzle TransBS = ALU_VEC_012_SCL_210;
487 for (unsigned i = 0, e = IG.size(); i < e; ++i) {
488 IGSrcs.push_back(ExtractSrcs(IG[i], PV, ConstCount));
489 unsigned Op = getOperandIdx(IG[i]->getOpcode(),
490 AMDGPU::OpName::bank_swizzle);
491 ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle)
492 IG[i]->getOperand(Op).getImm());
494 std::vector<std::pair<int, unsigned> > TransOps;
496 return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS);
498 TransOps = IGSrcs.back();
500 ValidSwizzle.pop_back();
502 static const R600InstrInfo::BankSwizzle TransSwz[] = {
508 for (unsigned i = 0; i < 4; i++) {
509 TransBS = TransSwz[i];
510 if (!isConstCompatible(TransBS, TransOps, ConstCount))
512 bool Result = FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps,
515 ValidSwizzle.push_back(TransBS);
525 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts)
527 assert (Consts.size() <= 12 && "Too many operands in instructions group");
528 unsigned Pair1 = 0, Pair2 = 0;
529 for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
530 unsigned ReadConstHalf = Consts[i] & 2;
531 unsigned ReadConstIndex = Consts[i] & (~3);
532 unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf;
534 Pair1 = ReadHalfConst;
537 if (Pair1 == ReadHalfConst)
540 Pair2 = ReadHalfConst;
543 if (Pair2 != ReadHalfConst)
550 R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
552 std::vector<unsigned> Consts;
553 SmallSet<int64_t, 4> Literals;
554 for (unsigned i = 0, n = MIs.size(); i < n; i++) {
555 MachineInstr *MI = MIs[i];
556 if (!isALUInstr(MI->getOpcode()))
559 const SmallVectorImpl<std::pair<MachineOperand *, int64_t> > &Srcs =
562 for (unsigned j = 0, e = Srcs.size(); j < e; j++) {
563 std::pair<MachineOperand *, unsigned> Src = Srcs[j];
564 if (Src.first->getReg() == AMDGPU::ALU_LITERAL_X)
565 Literals.insert(Src.second);
566 if (Literals.size() > 4)
568 if (Src.first->getReg() == AMDGPU::ALU_CONST)
569 Consts.push_back(Src.second);
570 if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) ||
571 AMDGPU::R600_KC1RegClass.contains(Src.first->getReg())) {
572 unsigned Index = RI.getEncodingValue(Src.first->getReg()) & 0xff;
573 unsigned Chan = RI.getHWRegChan(Src.first->getReg());
574 Consts.push_back((Index << 2) | Chan);
578 return fitsConstReadLimitations(Consts);
581 DFAPacketizer *R600InstrInfo::CreateTargetScheduleState(const TargetMachine *TM,
582 const ScheduleDAG *DAG) const {
583 const InstrItineraryData *II = TM->getInstrItineraryData();
584 return TM->getSubtarget<AMDGPUSubtarget>().createDFAPacketizer(II);
588 isPredicateSetter(unsigned Opcode) {
597 static MachineInstr *
598 findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
599 MachineBasicBlock::iterator I) {
600 while (I != MBB.begin()) {
602 MachineInstr *MI = I;
603 if (isPredicateSetter(MI->getOpcode()))
611 bool isJump(unsigned Opcode) {
612 return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
616 R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
617 MachineBasicBlock *&TBB,
618 MachineBasicBlock *&FBB,
619 SmallVectorImpl<MachineOperand> &Cond,
620 bool AllowModify) const {
621 // Most of the following comes from the ARM implementation of AnalyzeBranch
623 // If the block has no terminators, it just falls into the block after it.
624 MachineBasicBlock::iterator I = MBB.end();
625 if (I == MBB.begin())
628 while (I->isDebugValue()) {
629 if (I == MBB.begin())
633 if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) {
637 // Get the last instruction in the block.
638 MachineInstr *LastInst = I;
640 // If there is only one terminator instruction, process it.
641 unsigned LastOpc = LastInst->getOpcode();
642 if (I == MBB.begin() ||
643 !isJump(static_cast<MachineInstr *>(--I)->getOpcode())) {
644 if (LastOpc == AMDGPU::JUMP) {
645 TBB = LastInst->getOperand(0).getMBB();
647 } else if (LastOpc == AMDGPU::JUMP_COND) {
648 MachineInstr *predSet = I;
649 while (!isPredicateSetter(predSet->getOpcode())) {
652 TBB = LastInst->getOperand(0).getMBB();
653 Cond.push_back(predSet->getOperand(1));
654 Cond.push_back(predSet->getOperand(2));
655 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
658 return true; // Can't handle indirect branch.
661 // Get the instruction before it if it is a terminator.
662 MachineInstr *SecondLastInst = I;
663 unsigned SecondLastOpc = SecondLastInst->getOpcode();
665 // If the block ends with a B and a Bcc, handle it.
666 if (SecondLastOpc == AMDGPU::JUMP_COND && LastOpc == AMDGPU::JUMP) {
667 MachineInstr *predSet = --I;
668 while (!isPredicateSetter(predSet->getOpcode())) {
671 TBB = SecondLastInst->getOperand(0).getMBB();
672 FBB = LastInst->getOperand(0).getMBB();
673 Cond.push_back(predSet->getOperand(1));
674 Cond.push_back(predSet->getOperand(2));
675 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
679 // Otherwise, can't handle this.
683 int R600InstrInfo::getBranchInstr(const MachineOperand &op) const {
684 const MachineInstr *MI = op.getParent();
686 switch (MI->getDesc().OpInfo->RegClass) {
687 default: // FIXME: fallthrough??
688 case AMDGPU::GPRI32RegClassID: return AMDGPU::BRANCH_COND_i32;
689 case AMDGPU::GPRF32RegClassID: return AMDGPU::BRANCH_COND_f32;
694 MachineBasicBlock::iterator FindLastAluClause(MachineBasicBlock &MBB) {
695 for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend();
697 if (It->getOpcode() == AMDGPU::CF_ALU ||
698 It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE)
699 return llvm::prior(It.base());
705 R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
706 MachineBasicBlock *TBB,
707 MachineBasicBlock *FBB,
708 const SmallVectorImpl<MachineOperand> &Cond,
710 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
714 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB);
717 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
718 assert(PredSet && "No previous predicate !");
719 addFlag(PredSet, 0, MO_FLAG_PUSH);
720 PredSet->getOperand(2).setImm(Cond[1].getImm());
722 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
724 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
725 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
726 if (CfAlu == MBB.end())
728 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
729 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
733 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
734 assert(PredSet && "No previous predicate !");
735 addFlag(PredSet, 0, MO_FLAG_PUSH);
736 PredSet->getOperand(2).setImm(Cond[1].getImm());
737 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
739 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
740 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB);
741 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
742 if (CfAlu == MBB.end())
744 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
745 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
751 R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
753 // Note : we leave PRED* instructions there.
754 // They may be needed when predicating instructions.
756 MachineBasicBlock::iterator I = MBB.end();
758 if (I == MBB.begin()) {
762 switch (I->getOpcode()) {
765 case AMDGPU::JUMP_COND: {
766 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
767 clearFlag(predSet, 0, MO_FLAG_PUSH);
768 I->eraseFromParent();
769 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
770 if (CfAlu == MBB.end())
772 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
773 CfAlu->setDesc(get(AMDGPU::CF_ALU));
777 I->eraseFromParent();
782 if (I == MBB.begin()) {
786 switch (I->getOpcode()) {
787 // FIXME: only one case??
790 case AMDGPU::JUMP_COND: {
791 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
792 clearFlag(predSet, 0, MO_FLAG_PUSH);
793 I->eraseFromParent();
794 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
795 if (CfAlu == MBB.end())
797 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
798 CfAlu->setDesc(get(AMDGPU::CF_ALU));
802 I->eraseFromParent();
809 R600InstrInfo::isPredicated(const MachineInstr *MI) const {
810 int idx = MI->findFirstPredOperandIdx();
814 unsigned Reg = MI->getOperand(idx).getReg();
816 default: return false;
817 case AMDGPU::PRED_SEL_ONE:
818 case AMDGPU::PRED_SEL_ZERO:
819 case AMDGPU::PREDICATE_BIT:
825 R600InstrInfo::isPredicable(MachineInstr *MI) const {
826 // XXX: KILL* instructions can be predicated, but they must be the last
827 // instruction in a clause, so this means any instructions after them cannot
828 // be predicated. Until we have proper support for instruction clauses in the
829 // backend, we will mark KILL* instructions as unpredicable.
831 if (MI->getOpcode() == AMDGPU::KILLGT) {
833 } else if (MI->getOpcode() == AMDGPU::CF_ALU) {
834 // If the clause start in the middle of MBB then the MBB has more
835 // than a single clause, unable to predicate several clauses.
836 if (MI->getParent()->begin() != MachineBasicBlock::iterator(MI))
838 // TODO: We don't support KC merging atm
839 if (MI->getOperand(3).getImm() != 0 || MI->getOperand(4).getImm() != 0)
842 } else if (isVector(*MI)) {
845 return AMDGPUInstrInfo::isPredicable(MI);
851 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
853 unsigned ExtraPredCycles,
854 const BranchProbability &Probability) const{
859 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
861 unsigned ExtraTCycles,
862 MachineBasicBlock &FMBB,
864 unsigned ExtraFCycles,
865 const BranchProbability &Probability) const {
870 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
872 const BranchProbability &Probability)
878 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
879 MachineBasicBlock &FMBB) const {
885 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
886 MachineOperand &MO = Cond[1];
887 switch (MO.getImm()) {
888 case OPCODE_IS_ZERO_INT:
889 MO.setImm(OPCODE_IS_NOT_ZERO_INT);
891 case OPCODE_IS_NOT_ZERO_INT:
892 MO.setImm(OPCODE_IS_ZERO_INT);
895 MO.setImm(OPCODE_IS_NOT_ZERO);
897 case OPCODE_IS_NOT_ZERO:
898 MO.setImm(OPCODE_IS_ZERO);
904 MachineOperand &MO2 = Cond[2];
905 switch (MO2.getReg()) {
906 case AMDGPU::PRED_SEL_ZERO:
907 MO2.setReg(AMDGPU::PRED_SEL_ONE);
909 case AMDGPU::PRED_SEL_ONE:
910 MO2.setReg(AMDGPU::PRED_SEL_ZERO);
919 R600InstrInfo::DefinesPredicate(MachineInstr *MI,
920 std::vector<MachineOperand> &Pred) const {
921 return isPredicateSetter(MI->getOpcode());
926 R600InstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
927 const SmallVectorImpl<MachineOperand> &Pred2) const {
933 R600InstrInfo::PredicateInstruction(MachineInstr *MI,
934 const SmallVectorImpl<MachineOperand> &Pred) const {
935 int PIdx = MI->findFirstPredOperandIdx();
937 if (MI->getOpcode() == AMDGPU::CF_ALU) {
938 MI->getOperand(8).setImm(0);
943 MachineOperand &PMO = MI->getOperand(PIdx);
944 PMO.setReg(Pred[2].getReg());
945 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
946 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
953 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
954 const MachineInstr *MI,
955 unsigned *PredCost) const {
961 int R600InstrInfo::getIndirectIndexBegin(const MachineFunction &MF) const {
962 const MachineRegisterInfo &MRI = MF.getRegInfo();
963 const MachineFrameInfo *MFI = MF.getFrameInfo();
966 if (MFI->getNumObjects() == 0) {
970 if (MRI.livein_empty()) {
974 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
975 LE = MRI.livein_end();
977 Offset = std::max(Offset,
978 GET_REG_INDEX(RI.getEncodingValue(LI->first)));
984 int R600InstrInfo::getIndirectIndexEnd(const MachineFunction &MF) const {
986 const MachineFrameInfo *MFI = MF.getFrameInfo();
988 // Variable sized objects are not supported
989 assert(!MFI->hasVarSizedObjects());
991 if (MFI->getNumObjects() == 0) {
995 Offset = TM.getFrameLowering()->getFrameIndexOffset(MF, -1);
997 return getIndirectIndexBegin(MF) + Offset;
1000 std::vector<unsigned> R600InstrInfo::getIndirectReservedRegs(
1001 const MachineFunction &MF) const {
1002 const AMDGPUFrameLowering *TFL =
1003 static_cast<const AMDGPUFrameLowering*>(TM.getFrameLowering());
1004 std::vector<unsigned> Regs;
1006 unsigned StackWidth = TFL->getStackWidth(MF);
1007 int End = getIndirectIndexEnd(MF);
1013 for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
1014 unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
1015 Regs.push_back(SuperReg);
1016 for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
1017 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
1018 Regs.push_back(Reg);
1024 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
1025 unsigned Channel) const {
1026 // XXX: Remove when we support a stack width > 2
1027 assert(Channel == 0);
1031 const TargetRegisterClass * R600InstrInfo::getIndirectAddrStoreRegClass(
1032 unsigned SourceReg) const {
1033 return &AMDGPU::R600_TReg32RegClass;
1036 const TargetRegisterClass *R600InstrInfo::getIndirectAddrLoadRegClass() const {
1037 return &AMDGPU::TRegMemRegClass;
1040 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1041 MachineBasicBlock::iterator I,
1042 unsigned ValueReg, unsigned Address,
1043 unsigned OffsetReg) const {
1044 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1045 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1046 AMDGPU::AR_X, OffsetReg);
1047 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1049 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1051 .addReg(AMDGPU::AR_X,
1052 RegState::Implicit | RegState::Kill);
1053 setImmOperand(Mov, AMDGPU::OpName::dst_rel, 1);
1057 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1058 MachineBasicBlock::iterator I,
1059 unsigned ValueReg, unsigned Address,
1060 unsigned OffsetReg) const {
1061 unsigned AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address);
1062 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1065 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1066 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1069 .addReg(AMDGPU::AR_X,
1070 RegState::Implicit | RegState::Kill);
1071 setImmOperand(Mov, AMDGPU::OpName::src0_rel, 1);
1076 const TargetRegisterClass *R600InstrInfo::getSuperIndirectRegClass() const {
1077 return &AMDGPU::IndirectRegRegClass;
1080 unsigned R600InstrInfo::getMaxAlusPerClause() const {
1084 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
1085 MachineBasicBlock::iterator I,
1089 unsigned Src1Reg) const {
1090 MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
1094 MIB.addImm(0) // $update_exec_mask
1095 .addImm(0); // $update_predicate
1097 MIB.addImm(1) // $write
1099 .addImm(0) // $dst_rel
1100 .addImm(0) // $dst_clamp
1101 .addReg(Src0Reg) // $src0
1102 .addImm(0) // $src0_neg
1103 .addImm(0) // $src0_rel
1104 .addImm(0) // $src0_abs
1105 .addImm(-1); // $src0_sel
1108 MIB.addReg(Src1Reg) // $src1
1109 .addImm(0) // $src1_neg
1110 .addImm(0) // $src1_rel
1111 .addImm(0) // $src1_abs
1112 .addImm(-1); // $src1_sel
1115 //XXX: The r600g finalizer expects this to be 1, once we've moved the
1116 //scheduling to the backend, we can change the default to 0.
1117 MIB.addImm(1) // $last
1118 .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
1119 .addImm(0) // $literal
1120 .addImm(0); // $bank_swizzle
1125 #define OPERAND_CASE(Label) \
1127 static const unsigned Ops[] = \
1137 static unsigned getSlotedOps(unsigned Op, unsigned Slot) {
1139 OPERAND_CASE(AMDGPU::OpName::update_exec_mask)
1140 OPERAND_CASE(AMDGPU::OpName::update_pred)
1141 OPERAND_CASE(AMDGPU::OpName::write)
1142 OPERAND_CASE(AMDGPU::OpName::omod)
1143 OPERAND_CASE(AMDGPU::OpName::dst_rel)
1144 OPERAND_CASE(AMDGPU::OpName::clamp)
1145 OPERAND_CASE(AMDGPU::OpName::src0)
1146 OPERAND_CASE(AMDGPU::OpName::src0_neg)
1147 OPERAND_CASE(AMDGPU::OpName::src0_rel)
1148 OPERAND_CASE(AMDGPU::OpName::src0_abs)
1149 OPERAND_CASE(AMDGPU::OpName::src0_sel)
1150 OPERAND_CASE(AMDGPU::OpName::src1)
1151 OPERAND_CASE(AMDGPU::OpName::src1_neg)
1152 OPERAND_CASE(AMDGPU::OpName::src1_rel)
1153 OPERAND_CASE(AMDGPU::OpName::src1_abs)
1154 OPERAND_CASE(AMDGPU::OpName::src1_sel)
1155 OPERAND_CASE(AMDGPU::OpName::pred_sel)
1157 llvm_unreachable("Wrong Operand");
1163 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
1164 MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg)
1166 assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
1168 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
1169 if (ST.getGeneration() <= AMDGPUSubtarget::R700)
1170 Opcode = AMDGPU::DOT4_r600;
1172 Opcode = AMDGPU::DOT4_eg;
1173 MachineBasicBlock::iterator I = MI;
1174 MachineOperand &Src0 = MI->getOperand(
1175 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src0, Slot)));
1176 MachineOperand &Src1 = MI->getOperand(
1177 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src1, Slot)));
1178 MachineInstr *MIB = buildDefaultInstruction(
1179 MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg());
1180 static const unsigned Operands[14] = {
1181 AMDGPU::OpName::update_exec_mask,
1182 AMDGPU::OpName::update_pred,
1183 AMDGPU::OpName::write,
1184 AMDGPU::OpName::omod,
1185 AMDGPU::OpName::dst_rel,
1186 AMDGPU::OpName::clamp,
1187 AMDGPU::OpName::src0_neg,
1188 AMDGPU::OpName::src0_rel,
1189 AMDGPU::OpName::src0_abs,
1190 AMDGPU::OpName::src0_sel,
1191 AMDGPU::OpName::src1_neg,
1192 AMDGPU::OpName::src1_rel,
1193 AMDGPU::OpName::src1_abs,
1194 AMDGPU::OpName::src1_sel,
1197 for (unsigned i = 0; i < 14; i++) {
1198 MachineOperand &MO = MI->getOperand(
1199 getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));
1200 assert (MO.isImm());
1201 setImmOperand(MIB, Operands[i], MO.getImm());
1203 MIB->getOperand(20).setImm(0);
1207 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
1208 MachineBasicBlock::iterator I,
1210 uint64_t Imm) const {
1211 MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
1212 AMDGPU::ALU_LITERAL_X);
1213 setImmOperand(MovImm, AMDGPU::OpName::literal, Imm);
1217 int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const {
1218 return getOperandIdx(MI.getOpcode(), Op);
1221 int R600InstrInfo::getOperandIdx(unsigned Opcode, unsigned Op) const {
1222 return AMDGPU::getNamedOperandIdx(Opcode, Op);
1225 void R600InstrInfo::setImmOperand(MachineInstr *MI, unsigned Op,
1226 int64_t Imm) const {
1227 int Idx = getOperandIdx(*MI, Op);
1228 assert(Idx != -1 && "Operand not supported for this instruction.");
1229 assert(MI->getOperand(Idx).isImm());
1230 MI->getOperand(Idx).setImm(Imm);
1233 //===----------------------------------------------------------------------===//
1234 // Instruction flag getters/setters
1235 //===----------------------------------------------------------------------===//
1237 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
1238 return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
1241 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
1242 unsigned Flag) const {
1243 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1246 // If we pass something other than the default value of Flag to this
1247 // function, it means we are want to set a flag on an instruction
1248 // that uses native encoding.
1249 assert(HAS_NATIVE_OPERANDS(TargetFlags));
1250 bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
1253 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::clamp);
1256 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::write);
1258 case MO_FLAG_NOT_LAST:
1260 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::last);
1264 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_neg); break;
1265 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_neg); break;
1266 case 2: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src2_neg); break;
1271 assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
1275 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_abs); break;
1276 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_abs); break;
1284 assert(FlagIndex != -1 && "Flag not supported for this instruction");
1286 FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
1287 assert(FlagIndex != 0 &&
1288 "Instruction flags not supported for this instruction");
1291 MachineOperand &FlagOp = MI->getOperand(FlagIndex);
1292 assert(FlagOp.isImm());
1296 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
1297 unsigned Flag) const {
1298 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1302 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1303 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1304 if (Flag == MO_FLAG_NOT_LAST) {
1305 clearFlag(MI, Operand, MO_FLAG_LAST);
1306 } else if (Flag == MO_FLAG_MASK) {
1307 clearFlag(MI, Operand, Flag);
1312 MachineOperand &FlagOp = getFlagOp(MI, Operand);
1313 FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
1317 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
1318 unsigned Flag) const {
1319 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1320 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1321 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1324 MachineOperand &FlagOp = getFlagOp(MI);
1325 unsigned InstFlags = FlagOp.getImm();
1326 InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
1327 FlagOp.setImm(InstFlags);