1 //===-- SIInstrInfo.cpp - SI 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 SI Implementation of TargetInstrInfo.
13 //===----------------------------------------------------------------------===//
16 #include "SIInstrInfo.h"
17 #include "AMDGPUTargetMachine.h"
18 #include "SIDefines.h"
19 #include "SIMachineFunctionInfo.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/CodeGen/RegisterScavenging.h"
25 #include "llvm/MC/MCInstrDesc.h"
26 #include "llvm/Support/Debug.h"
30 SIInstrInfo::SIInstrInfo(const AMDGPUSubtarget &st)
31 : AMDGPUInstrInfo(st), RI(st) {}
33 //===----------------------------------------------------------------------===//
34 // TargetInstrInfo callbacks
35 //===----------------------------------------------------------------------===//
37 static unsigned getNumOperandsNoGlue(SDNode *Node) {
38 unsigned N = Node->getNumOperands();
39 while (N && Node->getOperand(N - 1).getValueType() == MVT::Glue)
44 static SDValue findChainOperand(SDNode *Load) {
45 SDValue LastOp = Load->getOperand(getNumOperandsNoGlue(Load) - 1);
46 assert(LastOp.getValueType() == MVT::Other && "Chain missing from load node");
50 /// \brief Returns true if both nodes have the same value for the given
51 /// operand \p Op, or if both nodes do not have this operand.
52 static bool nodesHaveSameOperandValue(SDNode *N0, SDNode* N1, unsigned OpName) {
53 unsigned Opc0 = N0->getMachineOpcode();
54 unsigned Opc1 = N1->getMachineOpcode();
56 int Op0Idx = AMDGPU::getNamedOperandIdx(Opc0, OpName);
57 int Op1Idx = AMDGPU::getNamedOperandIdx(Opc1, OpName);
59 if (Op0Idx == -1 && Op1Idx == -1)
63 if ((Op0Idx == -1 && Op1Idx != -1) ||
64 (Op1Idx == -1 && Op0Idx != -1))
67 // getNamedOperandIdx returns the index for the MachineInstr's operands,
68 // which includes the result as the first operand. We are indexing into the
69 // MachineSDNode's operands, so we need to skip the result operand to get
74 return N0->getOperand(Op0Idx) == N1->getOperand(Op1Idx);
77 bool SIInstrInfo::areLoadsFromSameBasePtr(SDNode *Load0, SDNode *Load1,
79 int64_t &Offset1) const {
80 if (!Load0->isMachineOpcode() || !Load1->isMachineOpcode())
83 unsigned Opc0 = Load0->getMachineOpcode();
84 unsigned Opc1 = Load1->getMachineOpcode();
86 // Make sure both are actually loads.
87 if (!get(Opc0).mayLoad() || !get(Opc1).mayLoad())
90 if (isDS(Opc0) && isDS(Opc1)) {
92 // FIXME: Handle this case:
93 if (getNumOperandsNoGlue(Load0) != getNumOperandsNoGlue(Load1))
97 if (Load0->getOperand(1) != Load1->getOperand(1))
101 if (findChainOperand(Load0) != findChainOperand(Load1))
104 // Skip read2 / write2 variants for simplicity.
105 // TODO: We should report true if the used offsets are adjacent (excluded
107 if (AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::data1) != -1 ||
108 AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::data1) != -1)
111 Offset0 = cast<ConstantSDNode>(Load0->getOperand(2))->getZExtValue();
112 Offset1 = cast<ConstantSDNode>(Load1->getOperand(2))->getZExtValue();
116 if (isSMRD(Opc0) && isSMRD(Opc1)) {
117 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
120 if (Load0->getOperand(0) != Load1->getOperand(0))
124 if (findChainOperand(Load0) != findChainOperand(Load1))
127 Offset0 = cast<ConstantSDNode>(Load0->getOperand(1))->getZExtValue();
128 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue();
132 // MUBUF and MTBUF can access the same addresses.
133 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1))) {
135 // MUBUF and MTBUF have vaddr at different indices.
136 if (!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::soffset) ||
137 findChainOperand(Load0) != findChainOperand(Load1) ||
138 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::vaddr) ||
139 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::srsrc))
142 int OffIdx0 = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
143 int OffIdx1 = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
145 if (OffIdx0 == -1 || OffIdx1 == -1)
148 // getNamedOperandIdx returns the index for MachineInstrs. Since they
149 // inlcude the output in the operand list, but SDNodes don't, we need to
150 // subtract the index by one.
154 SDValue Off0 = Load0->getOperand(OffIdx0);
155 SDValue Off1 = Load1->getOperand(OffIdx1);
157 // The offset might be a FrameIndexSDNode.
158 if (!isa<ConstantSDNode>(Off0) || !isa<ConstantSDNode>(Off1))
161 Offset0 = cast<ConstantSDNode>(Off0)->getZExtValue();
162 Offset1 = cast<ConstantSDNode>(Off1)->getZExtValue();
169 static bool isStride64(unsigned Opc) {
171 case AMDGPU::DS_READ2ST64_B32:
172 case AMDGPU::DS_READ2ST64_B64:
173 case AMDGPU::DS_WRITE2ST64_B32:
174 case AMDGPU::DS_WRITE2ST64_B64:
181 bool SIInstrInfo::getLdStBaseRegImmOfs(MachineInstr *LdSt,
182 unsigned &BaseReg, unsigned &Offset,
183 const TargetRegisterInfo *TRI) const {
184 unsigned Opc = LdSt->getOpcode();
186 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
187 AMDGPU::OpName::offset);
189 // Normal, single offset LDS instruction.
190 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
191 AMDGPU::OpName::addr);
193 BaseReg = AddrReg->getReg();
194 Offset = OffsetImm->getImm();
198 // The 2 offset instructions use offset0 and offset1 instead. We can treat
199 // these as a load with a single offset if the 2 offsets are consecutive. We
200 // will use this for some partially aligned loads.
201 const MachineOperand *Offset0Imm = getNamedOperand(*LdSt,
202 AMDGPU::OpName::offset0);
203 const MachineOperand *Offset1Imm = getNamedOperand(*LdSt,
204 AMDGPU::OpName::offset1);
206 uint8_t Offset0 = Offset0Imm->getImm();
207 uint8_t Offset1 = Offset1Imm->getImm();
208 assert(Offset1 > Offset0);
210 if (Offset1 - Offset0 == 1) {
211 // Each of these offsets is in element sized units, so we need to convert
212 // to bytes of the individual reads.
216 EltSize = getOpRegClass(*LdSt, 0)->getSize() / 2;
218 assert(LdSt->mayStore());
219 int Data0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
220 EltSize = getOpRegClass(*LdSt, Data0Idx)->getSize();
226 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
227 AMDGPU::OpName::addr);
228 BaseReg = AddrReg->getReg();
229 Offset = EltSize * Offset0;
236 if (isMUBUF(Opc) || isMTBUF(Opc)) {
237 if (AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::soffset) != -1)
240 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
241 AMDGPU::OpName::vaddr);
245 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
246 AMDGPU::OpName::offset);
247 BaseReg = AddrReg->getReg();
248 Offset = OffsetImm->getImm();
253 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
254 AMDGPU::OpName::offset);
258 const MachineOperand *SBaseReg = getNamedOperand(*LdSt,
259 AMDGPU::OpName::sbase);
260 BaseReg = SBaseReg->getReg();
261 Offset = OffsetImm->getImm();
268 bool SIInstrInfo::shouldClusterLoads(MachineInstr *FirstLdSt,
269 MachineInstr *SecondLdSt,
270 unsigned NumLoads) const {
271 unsigned Opc0 = FirstLdSt->getOpcode();
272 unsigned Opc1 = SecondLdSt->getOpcode();
274 // TODO: This needs finer tuning
278 if (isDS(Opc0) && isDS(Opc1))
281 if (isSMRD(Opc0) && isSMRD(Opc1))
284 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1)))
291 SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
292 MachineBasicBlock::iterator MI, DebugLoc DL,
293 unsigned DestReg, unsigned SrcReg,
294 bool KillSrc) const {
296 // If we are trying to copy to or from SCC, there is a bug somewhere else in
297 // the backend. While it may be theoretically possible to do this, it should
298 // never be necessary.
299 assert(DestReg != AMDGPU::SCC && SrcReg != AMDGPU::SCC);
301 static const int16_t Sub0_15[] = {
302 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
303 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
304 AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
305 AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15, 0
308 static const int16_t Sub0_7[] = {
309 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
310 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7, 0
313 static const int16_t Sub0_3[] = {
314 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
317 static const int16_t Sub0_2[] = {
318 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
321 static const int16_t Sub0_1[] = {
322 AMDGPU::sub0, AMDGPU::sub1, 0
326 const int16_t *SubIndices;
328 if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
329 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
330 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
331 .addReg(SrcReg, getKillRegState(KillSrc));
334 } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
335 if (DestReg == AMDGPU::VCC) {
336 // FIXME: Hack until VReg_1 removed.
338 assert(AMDGPU::VGPR_32RegClass.contains(SrcReg));
339 BuildMI(MBB, MI, DL, get(AMDGPU::V_CMP_NE_I32_e32), AMDGPU::VCC)
341 .addReg(SrcReg, getKillRegState(KillSrc));
345 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
346 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
347 .addReg(SrcReg, getKillRegState(KillSrc));
350 } else if (AMDGPU::SReg_128RegClass.contains(DestReg)) {
351 assert(AMDGPU::SReg_128RegClass.contains(SrcReg));
352 Opcode = AMDGPU::S_MOV_B32;
355 } else if (AMDGPU::SReg_256RegClass.contains(DestReg)) {
356 assert(AMDGPU::SReg_256RegClass.contains(SrcReg));
357 Opcode = AMDGPU::S_MOV_B32;
360 } else if (AMDGPU::SReg_512RegClass.contains(DestReg)) {
361 assert(AMDGPU::SReg_512RegClass.contains(SrcReg));
362 Opcode = AMDGPU::S_MOV_B32;
363 SubIndices = Sub0_15;
365 } else if (AMDGPU::VGPR_32RegClass.contains(DestReg)) {
366 assert(AMDGPU::VGPR_32RegClass.contains(SrcReg) ||
367 AMDGPU::SReg_32RegClass.contains(SrcReg));
368 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
369 .addReg(SrcReg, getKillRegState(KillSrc));
372 } else if (AMDGPU::VReg_64RegClass.contains(DestReg)) {
373 assert(AMDGPU::VReg_64RegClass.contains(SrcReg) ||
374 AMDGPU::SReg_64RegClass.contains(SrcReg));
375 Opcode = AMDGPU::V_MOV_B32_e32;
378 } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
379 assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
380 Opcode = AMDGPU::V_MOV_B32_e32;
383 } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
384 assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
385 AMDGPU::SReg_128RegClass.contains(SrcReg));
386 Opcode = AMDGPU::V_MOV_B32_e32;
389 } else if (AMDGPU::VReg_256RegClass.contains(DestReg)) {
390 assert(AMDGPU::VReg_256RegClass.contains(SrcReg) ||
391 AMDGPU::SReg_256RegClass.contains(SrcReg));
392 Opcode = AMDGPU::V_MOV_B32_e32;
395 } else if (AMDGPU::VReg_512RegClass.contains(DestReg)) {
396 assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
397 AMDGPU::SReg_512RegClass.contains(SrcReg));
398 Opcode = AMDGPU::V_MOV_B32_e32;
399 SubIndices = Sub0_15;
402 llvm_unreachable("Can't copy register!");
405 while (unsigned SubIdx = *SubIndices++) {
406 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
407 get(Opcode), RI.getSubReg(DestReg, SubIdx));
409 Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
412 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
416 unsigned SIInstrInfo::commuteOpcode(unsigned Opcode) const {
419 // Try to map original to commuted opcode
420 NewOpc = AMDGPU::getCommuteRev(Opcode);
421 // Check if the commuted (REV) opcode exists on the target.
422 if (NewOpc != -1 && pseudoToMCOpcode(NewOpc) != -1)
425 // Try to map commuted to original opcode
426 NewOpc = AMDGPU::getCommuteOrig(Opcode);
427 // Check if the original (non-REV) opcode exists on the target.
428 if (NewOpc != -1 && pseudoToMCOpcode(NewOpc) != -1)
434 unsigned SIInstrInfo::getMovOpcode(const TargetRegisterClass *DstRC) const {
436 if (DstRC->getSize() == 4) {
437 return RI.isSGPRClass(DstRC) ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32;
438 } else if (DstRC->getSize() == 8 && RI.isSGPRClass(DstRC)) {
439 return AMDGPU::S_MOV_B64;
440 } else if (DstRC->getSize() == 8 && !RI.isSGPRClass(DstRC)) {
441 return AMDGPU::V_MOV_B64_PSEUDO;
446 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
447 MachineBasicBlock::iterator MI,
448 unsigned SrcReg, bool isKill,
450 const TargetRegisterClass *RC,
451 const TargetRegisterInfo *TRI) const {
452 MachineFunction *MF = MBB.getParent();
453 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
454 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
455 DebugLoc DL = MBB.findDebugLoc(MI);
458 if (RI.isSGPRClass(RC)) {
459 // We are only allowed to create one new instruction when spilling
460 // registers, so we need to use pseudo instruction for spilling
462 switch (RC->getSize() * 8) {
463 case 32: Opcode = AMDGPU::SI_SPILL_S32_SAVE; break;
464 case 64: Opcode = AMDGPU::SI_SPILL_S64_SAVE; break;
465 case 128: Opcode = AMDGPU::SI_SPILL_S128_SAVE; break;
466 case 256: Opcode = AMDGPU::SI_SPILL_S256_SAVE; break;
467 case 512: Opcode = AMDGPU::SI_SPILL_S512_SAVE; break;
469 } else if(RI.hasVGPRs(RC) && ST.isVGPRSpillingEnabled(MFI)) {
470 MFI->setHasSpilledVGPRs();
472 switch(RC->getSize() * 8) {
473 case 32: Opcode = AMDGPU::SI_SPILL_V32_SAVE; break;
474 case 64: Opcode = AMDGPU::SI_SPILL_V64_SAVE; break;
475 case 96: Opcode = AMDGPU::SI_SPILL_V96_SAVE; break;
476 case 128: Opcode = AMDGPU::SI_SPILL_V128_SAVE; break;
477 case 256: Opcode = AMDGPU::SI_SPILL_V256_SAVE; break;
478 case 512: Opcode = AMDGPU::SI_SPILL_V512_SAVE; break;
483 FrameInfo->setObjectAlignment(FrameIndex, 4);
484 BuildMI(MBB, MI, DL, get(Opcode))
486 .addFrameIndex(FrameIndex)
487 // Place-holder registers, these will be filled in by
488 // SIPrepareScratchRegs.
489 .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
490 .addReg(AMDGPU::SGPR0, RegState::Undef);
492 LLVMContext &Ctx = MF->getFunction()->getContext();
493 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Do not know how to"
495 BuildMI(MBB, MI, DL, get(AMDGPU::KILL))
500 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
501 MachineBasicBlock::iterator MI,
502 unsigned DestReg, int FrameIndex,
503 const TargetRegisterClass *RC,
504 const TargetRegisterInfo *TRI) const {
505 MachineFunction *MF = MBB.getParent();
506 const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
507 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
508 DebugLoc DL = MBB.findDebugLoc(MI);
511 if (RI.isSGPRClass(RC)){
512 switch(RC->getSize() * 8) {
513 case 32: Opcode = AMDGPU::SI_SPILL_S32_RESTORE; break;
514 case 64: Opcode = AMDGPU::SI_SPILL_S64_RESTORE; break;
515 case 128: Opcode = AMDGPU::SI_SPILL_S128_RESTORE; break;
516 case 256: Opcode = AMDGPU::SI_SPILL_S256_RESTORE; break;
517 case 512: Opcode = AMDGPU::SI_SPILL_S512_RESTORE; break;
519 } else if(RI.hasVGPRs(RC) && ST.isVGPRSpillingEnabled(MFI)) {
520 switch(RC->getSize() * 8) {
521 case 32: Opcode = AMDGPU::SI_SPILL_V32_RESTORE; break;
522 case 64: Opcode = AMDGPU::SI_SPILL_V64_RESTORE; break;
523 case 96: Opcode = AMDGPU::SI_SPILL_V96_RESTORE; break;
524 case 128: Opcode = AMDGPU::SI_SPILL_V128_RESTORE; break;
525 case 256: Opcode = AMDGPU::SI_SPILL_V256_RESTORE; break;
526 case 512: Opcode = AMDGPU::SI_SPILL_V512_RESTORE; break;
531 FrameInfo->setObjectAlignment(FrameIndex, 4);
532 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
533 .addFrameIndex(FrameIndex)
534 // Place-holder registers, these will be filled in by
535 // SIPrepareScratchRegs.
536 .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
537 .addReg(AMDGPU::SGPR0, RegState::Undef);
540 LLVMContext &Ctx = MF->getFunction()->getContext();
541 Ctx.emitError("SIInstrInfo::loadRegFromStackSlot - Do not know how to"
542 " restore register");
543 BuildMI(MBB, MI, DL, get(AMDGPU::IMPLICIT_DEF), DestReg);
547 /// \param @Offset Offset in bytes of the FrameIndex being spilled
548 unsigned SIInstrInfo::calculateLDSSpillAddress(MachineBasicBlock &MBB,
549 MachineBasicBlock::iterator MI,
550 RegScavenger *RS, unsigned TmpReg,
551 unsigned FrameOffset,
552 unsigned Size) const {
553 MachineFunction *MF = MBB.getParent();
554 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
555 const AMDGPUSubtarget &ST = MF->getSubtarget<AMDGPUSubtarget>();
556 const SIRegisterInfo *TRI =
557 static_cast<const SIRegisterInfo*>(ST.getRegisterInfo());
558 DebugLoc DL = MBB.findDebugLoc(MI);
559 unsigned WorkGroupSize = MFI->getMaximumWorkGroupSize(*MF);
560 unsigned WavefrontSize = ST.getWavefrontSize();
562 unsigned TIDReg = MFI->getTIDReg();
563 if (!MFI->hasCalculatedTID()) {
564 MachineBasicBlock &Entry = MBB.getParent()->front();
565 MachineBasicBlock::iterator Insert = Entry.front();
566 DebugLoc DL = Insert->getDebugLoc();
568 TIDReg = RI.findUnusedRegister(MF->getRegInfo(), &AMDGPU::VGPR_32RegClass);
569 if (TIDReg == AMDGPU::NoRegister)
573 if (MFI->getShaderType() == ShaderType::COMPUTE &&
574 WorkGroupSize > WavefrontSize) {
576 unsigned TIDIGXReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_X);
577 unsigned TIDIGYReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Y);
578 unsigned TIDIGZReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Z);
579 unsigned InputPtrReg =
580 TRI->getPreloadedValue(*MF, SIRegisterInfo::INPUT_PTR);
581 static const unsigned TIDIGRegs[3] = {
582 TIDIGXReg, TIDIGYReg, TIDIGZReg
584 for (unsigned Reg : TIDIGRegs) {
585 if (!Entry.isLiveIn(Reg))
586 Entry.addLiveIn(Reg);
589 RS->enterBasicBlock(&Entry);
590 unsigned STmp0 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
591 unsigned STmp1 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
592 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp0)
594 .addImm(SI::KernelInputOffsets::NGROUPS_Z);
595 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp1)
597 .addImm(SI::KernelInputOffsets::NGROUPS_Y);
599 // NGROUPS.X * NGROUPS.Y
600 BuildMI(Entry, Insert, DL, get(AMDGPU::S_MUL_I32), STmp1)
603 // (NGROUPS.X * NGROUPS.Y) * TIDIG.X
604 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MUL_U32_U24_e32), TIDReg)
607 // NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)
608 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MAD_U32_U24), TIDReg)
612 // (NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)) + TIDIG.Z
613 BuildMI(Entry, Insert, DL, get(AMDGPU::V_ADD_I32_e32), TIDReg)
618 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_LO_U32_B32_e64),
623 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_HI_U32_B32_e64),
629 BuildMI(Entry, Insert, DL, get(AMDGPU::V_LSHLREV_B32_e32),
633 MFI->setTIDReg(TIDReg);
636 // Add FrameIndex to LDS offset
637 unsigned LDSOffset = MFI->LDSSize + (FrameOffset * WorkGroupSize);
638 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_I32_e32), TmpReg)
645 void SIInstrInfo::insertNOPs(MachineBasicBlock::iterator MI,
654 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(AMDGPU::S_NOP))
659 bool SIInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
660 MachineBasicBlock &MBB = *MI->getParent();
661 DebugLoc DL = MBB.findDebugLoc(MI);
662 switch (MI->getOpcode()) {
663 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
665 case AMDGPU::SI_CONSTDATA_PTR: {
666 unsigned Reg = MI->getOperand(0).getReg();
667 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
668 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
670 BuildMI(MBB, MI, DL, get(AMDGPU::S_GETPC_B64), Reg);
672 // Add 32-bit offset from this instruction to the start of the constant data.
673 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADD_U32), RegLo)
675 .addTargetIndex(AMDGPU::TI_CONSTDATA_START)
676 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit);
677 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADDC_U32), RegHi)
680 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit)
681 .addReg(AMDGPU::SCC, RegState::Implicit);
682 MI->eraseFromParent();
685 case AMDGPU::SGPR_USE:
686 // This is just a placeholder for register allocation.
687 MI->eraseFromParent();
690 case AMDGPU::V_MOV_B64_PSEUDO: {
691 unsigned Dst = MI->getOperand(0).getReg();
692 unsigned DstLo = RI.getSubReg(Dst, AMDGPU::sub0);
693 unsigned DstHi = RI.getSubReg(Dst, AMDGPU::sub1);
695 const MachineOperand &SrcOp = MI->getOperand(1);
696 // FIXME: Will this work for 64-bit floating point immediates?
697 assert(!SrcOp.isFPImm());
699 APInt Imm(64, SrcOp.getImm());
700 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
701 .addImm(Imm.getLoBits(32).getZExtValue())
702 .addReg(Dst, RegState::Implicit);
703 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
704 .addImm(Imm.getHiBits(32).getZExtValue())
705 .addReg(Dst, RegState::Implicit);
707 assert(SrcOp.isReg());
708 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
709 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub0))
710 .addReg(Dst, RegState::Implicit);
711 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
712 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub1))
713 .addReg(Dst, RegState::Implicit);
715 MI->eraseFromParent();
722 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
725 if (MI->getNumOperands() < 3)
728 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
729 AMDGPU::OpName::src0);
730 assert(Src0Idx != -1 && "Should always have src0 operand");
732 MachineOperand &Src0 = MI->getOperand(Src0Idx);
736 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
737 AMDGPU::OpName::src1);
741 MachineOperand &Src1 = MI->getOperand(Src1Idx);
743 // Make sure it's legal to commute operands for VOP2.
744 if (isVOP2(MI->getOpcode()) &&
745 (!isOperandLegal(MI, Src0Idx, &Src1) ||
746 !isOperandLegal(MI, Src1Idx, &Src0))) {
751 // Allow commuting instructions with Imm operands.
752 if (NewMI || !Src1.isImm() ||
753 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
757 // Be sure to copy the source modifiers to the right place.
758 if (MachineOperand *Src0Mods
759 = getNamedOperand(*MI, AMDGPU::OpName::src0_modifiers)) {
760 MachineOperand *Src1Mods
761 = getNamedOperand(*MI, AMDGPU::OpName::src1_modifiers);
763 int Src0ModsVal = Src0Mods->getImm();
764 if (!Src1Mods && Src0ModsVal != 0)
767 // XXX - This assert might be a lie. It might be useful to have a neg
768 // modifier with 0.0.
769 int Src1ModsVal = Src1Mods->getImm();
770 assert((Src1ModsVal == 0) && "Not expecting modifiers with immediates");
772 Src1Mods->setImm(Src0ModsVal);
773 Src0Mods->setImm(Src1ModsVal);
776 unsigned Reg = Src0.getReg();
777 unsigned SubReg = Src0.getSubReg();
779 Src0.ChangeToImmediate(Src1.getImm());
781 llvm_unreachable("Should only have immediates");
783 Src1.ChangeToRegister(Reg, false);
784 Src1.setSubReg(SubReg);
786 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
790 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
795 // This needs to be implemented because the source modifiers may be inserted
796 // between the true commutable operands, and the base
797 // TargetInstrInfo::commuteInstruction uses it.
798 bool SIInstrInfo::findCommutedOpIndices(MachineInstr *MI,
800 unsigned &SrcOpIdx2) const {
801 const MCInstrDesc &MCID = MI->getDesc();
802 if (!MCID.isCommutable())
805 unsigned Opc = MI->getOpcode();
806 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
810 // FIXME: Workaround TargetInstrInfo::commuteInstruction asserting on
812 if (!MI->getOperand(Src0Idx).isReg())
815 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
819 if (!MI->getOperand(Src1Idx).isReg())
822 // If any source modifiers are set, the generic instruction commuting won't
823 // understand how to copy the source modifiers.
824 if (hasModifiersSet(*MI, AMDGPU::OpName::src0_modifiers) ||
825 hasModifiersSet(*MI, AMDGPU::OpName::src1_modifiers))
833 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
834 MachineBasicBlock::iterator I,
836 unsigned SrcReg) const {
837 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
838 DstReg) .addReg(SrcReg);
841 bool SIInstrInfo::isMov(unsigned Opcode) const {
843 default: return false;
844 case AMDGPU::S_MOV_B32:
845 case AMDGPU::S_MOV_B64:
846 case AMDGPU::V_MOV_B32_e32:
847 case AMDGPU::V_MOV_B32_e64:
853 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
854 return RC != &AMDGPU::EXECRegRegClass;
858 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
859 AliasAnalysis *AA) const {
860 switch(MI->getOpcode()) {
861 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
862 case AMDGPU::S_MOV_B32:
863 case AMDGPU::S_MOV_B64:
864 case AMDGPU::V_MOV_B32_e32:
865 return MI->getOperand(1).isImm();
869 static bool offsetsDoNotOverlap(int WidthA, int OffsetA,
870 int WidthB, int OffsetB) {
871 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB;
872 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA;
873 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
874 return LowOffset + LowWidth <= HighOffset;
877 bool SIInstrInfo::checkInstOffsetsDoNotOverlap(MachineInstr *MIa,
878 MachineInstr *MIb) const {
879 unsigned BaseReg0, Offset0;
880 unsigned BaseReg1, Offset1;
882 if (getLdStBaseRegImmOfs(MIa, BaseReg0, Offset0, &RI) &&
883 getLdStBaseRegImmOfs(MIb, BaseReg1, Offset1, &RI)) {
884 assert(MIa->hasOneMemOperand() && MIb->hasOneMemOperand() &&
885 "read2 / write2 not expected here yet");
886 unsigned Width0 = (*MIa->memoperands_begin())->getSize();
887 unsigned Width1 = (*MIb->memoperands_begin())->getSize();
888 if (BaseReg0 == BaseReg1 &&
889 offsetsDoNotOverlap(Width0, Offset0, Width1, Offset1)) {
897 bool SIInstrInfo::areMemAccessesTriviallyDisjoint(MachineInstr *MIa,
899 AliasAnalysis *AA) const {
900 unsigned Opc0 = MIa->getOpcode();
901 unsigned Opc1 = MIb->getOpcode();
903 assert(MIa && (MIa->mayLoad() || MIa->mayStore()) &&
904 "MIa must load from or modify a memory location");
905 assert(MIb && (MIb->mayLoad() || MIb->mayStore()) &&
906 "MIb must load from or modify a memory location");
908 if (MIa->hasUnmodeledSideEffects() || MIb->hasUnmodeledSideEffects())
911 // XXX - Can we relax this between address spaces?
912 if (MIa->hasOrderedMemoryRef() || MIb->hasOrderedMemoryRef())
915 // TODO: Should we check the address space from the MachineMemOperand? That
916 // would allow us to distinguish objects we know don't alias based on the
917 // underlying addres space, even if it was lowered to a different one,
918 // e.g. private accesses lowered to use MUBUF instructions on a scratch
922 return checkInstOffsetsDoNotOverlap(MIa, MIb);
924 return !isFLAT(Opc1);
927 if (isMUBUF(Opc0) || isMTBUF(Opc0)) {
928 if (isMUBUF(Opc1) || isMTBUF(Opc1))
929 return checkInstOffsetsDoNotOverlap(MIa, MIb);
931 return !isFLAT(Opc1) && !isSMRD(Opc1);
936 return checkInstOffsetsDoNotOverlap(MIa, MIb);
938 return !isFLAT(Opc1) && !isMUBUF(Opc0) && !isMTBUF(Opc0);
943 return checkInstOffsetsDoNotOverlap(MIa, MIb);
951 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
952 int64_t SVal = Imm.getSExtValue();
953 if (SVal >= -16 && SVal <= 64)
956 if (Imm.getBitWidth() == 64) {
957 uint64_t Val = Imm.getZExtValue();
958 return (DoubleToBits(0.0) == Val) ||
959 (DoubleToBits(1.0) == Val) ||
960 (DoubleToBits(-1.0) == Val) ||
961 (DoubleToBits(0.5) == Val) ||
962 (DoubleToBits(-0.5) == Val) ||
963 (DoubleToBits(2.0) == Val) ||
964 (DoubleToBits(-2.0) == Val) ||
965 (DoubleToBits(4.0) == Val) ||
966 (DoubleToBits(-4.0) == Val);
969 // The actual type of the operand does not seem to matter as long
970 // as the bits match one of the inline immediate values. For example:
972 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
973 // so it is a legal inline immediate.
975 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
976 // floating-point, so it is a legal inline immediate.
977 uint32_t Val = Imm.getZExtValue();
979 return (FloatToBits(0.0f) == Val) ||
980 (FloatToBits(1.0f) == Val) ||
981 (FloatToBits(-1.0f) == Val) ||
982 (FloatToBits(0.5f) == Val) ||
983 (FloatToBits(-0.5f) == Val) ||
984 (FloatToBits(2.0f) == Val) ||
985 (FloatToBits(-2.0f) == Val) ||
986 (FloatToBits(4.0f) == Val) ||
987 (FloatToBits(-4.0f) == Val);
990 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO,
991 unsigned OpSize) const {
993 // MachineOperand provides no way to tell the true operand size, since it
994 // only records a 64-bit value. We need to know the size to determine if a
995 // 32-bit floating point immediate bit pattern is legal for an integer
996 // immediate. It would be for any 32-bit integer operand, but would not be
999 unsigned BitSize = 8 * OpSize;
1000 return isInlineConstant(APInt(BitSize, MO.getImm(), true));
1006 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO,
1007 unsigned OpSize) const {
1008 return MO.isImm() && !isInlineConstant(MO, OpSize);
1011 static bool compareMachineOp(const MachineOperand &Op0,
1012 const MachineOperand &Op1) {
1013 if (Op0.getType() != Op1.getType())
1016 switch (Op0.getType()) {
1017 case MachineOperand::MO_Register:
1018 return Op0.getReg() == Op1.getReg();
1019 case MachineOperand::MO_Immediate:
1020 return Op0.getImm() == Op1.getImm();
1022 llvm_unreachable("Didn't expect to be comparing these operand types");
1026 bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
1027 const MachineOperand &MO) const {
1028 const MCOperandInfo &OpInfo = get(MI->getOpcode()).OpInfo[OpNo];
1030 assert(MO.isImm() || MO.isTargetIndex() || MO.isFI());
1032 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
1035 if (OpInfo.RegClass < 0)
1038 unsigned OpSize = RI.getRegClass(OpInfo.RegClass)->getSize();
1039 if (isLiteralConstant(MO, OpSize))
1040 return RI.opCanUseLiteralConstant(OpInfo.OperandType);
1042 return RI.opCanUseInlineConstant(OpInfo.OperandType);
1045 bool SIInstrInfo::canFoldOffset(unsigned OffsetSize, unsigned AS) const {
1047 case AMDGPUAS::GLOBAL_ADDRESS: {
1048 // MUBUF instructions a 12-bit offset in bytes.
1049 return isUInt<12>(OffsetSize);
1051 case AMDGPUAS::CONSTANT_ADDRESS: {
1052 // SMRD instructions have an 8-bit offset in dwords on SI and
1053 // a 20-bit offset in bytes on VI.
1054 if (RI.ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1055 return isUInt<20>(OffsetSize);
1057 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
1059 case AMDGPUAS::LOCAL_ADDRESS:
1060 case AMDGPUAS::REGION_ADDRESS: {
1061 // The single offset versions have a 16-bit offset in bytes.
1062 return isUInt<16>(OffsetSize);
1064 case AMDGPUAS::PRIVATE_ADDRESS:
1065 // Indirect register addressing does not use any offsets.
1071 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
1072 int Op32 = AMDGPU::getVOPe32(Opcode);
1076 return pseudoToMCOpcode(Op32) != -1;
1079 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
1080 // The src0_modifier operand is present on all instructions
1081 // that have modifiers.
1083 return AMDGPU::getNamedOperandIdx(Opcode,
1084 AMDGPU::OpName::src0_modifiers) != -1;
1087 bool SIInstrInfo::hasModifiersSet(const MachineInstr &MI,
1088 unsigned OpName) const {
1089 const MachineOperand *Mods = getNamedOperand(MI, OpName);
1090 return Mods && Mods->getImm();
1093 bool SIInstrInfo::usesConstantBus(const MachineRegisterInfo &MRI,
1094 const MachineOperand &MO,
1095 unsigned OpSize) const {
1096 // Literal constants use the constant bus.
1097 if (isLiteralConstant(MO, OpSize))
1100 if (!MO.isReg() || !MO.isUse())
1103 if (TargetRegisterInfo::isVirtualRegister(MO.getReg()))
1104 return RI.isSGPRClass(MRI.getRegClass(MO.getReg()));
1106 // FLAT_SCR is just an SGPR pair.
1107 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
1110 // EXEC register uses the constant bus.
1111 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
1114 // SGPRs use the constant bus
1115 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
1116 (!MO.isImplicit() &&
1117 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
1118 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
1125 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
1126 StringRef &ErrInfo) const {
1127 uint16_t Opcode = MI->getOpcode();
1128 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1129 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
1130 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
1131 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
1133 // Make sure the number of operands is correct.
1134 const MCInstrDesc &Desc = get(Opcode);
1135 if (!Desc.isVariadic() &&
1136 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
1137 ErrInfo = "Instruction has wrong number of operands.";
1141 // Make sure the register classes are correct
1142 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
1143 if (MI->getOperand(i).isFPImm()) {
1144 ErrInfo = "FPImm Machine Operands are not supported. ISel should bitcast "
1145 "all fp values to integers.";
1149 switch (Desc.OpInfo[i].OperandType) {
1150 case MCOI::OPERAND_REGISTER:
1151 if (MI->getOperand(i).isImm()) {
1152 ErrInfo = "Illegal immediate value for operand.";
1156 case AMDGPU::OPERAND_REG_IMM32:
1158 case AMDGPU::OPERAND_REG_INLINE_C:
1159 if (MI->getOperand(i).isImm()) {
1160 int RegClass = Desc.OpInfo[i].RegClass;
1161 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
1162 if (!isInlineConstant(MI->getOperand(i), RC->getSize())) {
1163 ErrInfo = "Illegal immediate value for operand.";
1168 case MCOI::OPERAND_IMMEDIATE:
1169 // Check if this operand is an immediate.
1170 // FrameIndex operands will be replaced by immediates, so they are
1172 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFI()) {
1173 ErrInfo = "Expected immediate, but got non-immediate";
1181 if (!MI->getOperand(i).isReg())
1184 int RegClass = Desc.OpInfo[i].RegClass;
1185 if (RegClass != -1) {
1186 unsigned Reg = MI->getOperand(i).getReg();
1187 if (TargetRegisterInfo::isVirtualRegister(Reg))
1190 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
1191 if (!RC->contains(Reg)) {
1192 ErrInfo = "Operand has incorrect register class.";
1200 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
1201 // Only look at the true operands. Only a real operand can use the constant
1202 // bus, and we don't want to check pseudo-operands like the source modifier
1204 const int OpIndices[] = { Src0Idx, Src1Idx, Src2Idx };
1206 unsigned ConstantBusCount = 0;
1207 unsigned SGPRUsed = AMDGPU::NoRegister;
1208 for (int OpIdx : OpIndices) {
1211 const MachineOperand &MO = MI->getOperand(OpIdx);
1212 if (usesConstantBus(MRI, MO, getOpSize(Opcode, OpIdx))) {
1214 if (MO.getReg() != SGPRUsed)
1216 SGPRUsed = MO.getReg();
1222 if (ConstantBusCount > 1) {
1223 ErrInfo = "VOP* instruction uses the constant bus more than once";
1228 // Verify SRC1 for VOP2 and VOPC
1229 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
1230 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1232 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
1238 if (isVOP3(Opcode)) {
1239 if (Src0Idx != -1 &&
1240 isLiteralConstant(MI->getOperand(Src0Idx), getOpSize(Opcode, Src0Idx))) {
1241 ErrInfo = "VOP3 src0 cannot be a literal constant.";
1244 if (Src1Idx != -1 &&
1245 isLiteralConstant(MI->getOperand(Src1Idx), getOpSize(Opcode, Src1Idx))) {
1246 ErrInfo = "VOP3 src1 cannot be a literal constant.";
1249 if (Src2Idx != -1 &&
1250 isLiteralConstant(MI->getOperand(Src2Idx), getOpSize(Opcode, Src2Idx))) {
1251 ErrInfo = "VOP3 src2 cannot be a literal constant.";
1256 // Verify misc. restrictions on specific instructions.
1257 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
1258 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
1259 const MachineOperand &Src0 = MI->getOperand(Src0Idx);
1260 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1261 const MachineOperand &Src2 = MI->getOperand(Src2Idx);
1262 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
1263 if (!compareMachineOp(Src0, Src1) &&
1264 !compareMachineOp(Src0, Src2)) {
1265 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
1274 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
1275 switch (MI.getOpcode()) {
1276 default: return AMDGPU::INSTRUCTION_LIST_END;
1277 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
1278 case AMDGPU::COPY: return AMDGPU::COPY;
1279 case AMDGPU::PHI: return AMDGPU::PHI;
1280 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
1281 case AMDGPU::S_MOV_B32:
1282 return MI.getOperand(1).isReg() ?
1283 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
1284 case AMDGPU::S_ADD_I32:
1285 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
1286 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
1287 case AMDGPU::S_SUB_I32:
1288 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
1289 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
1290 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
1291 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
1292 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
1293 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
1294 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
1295 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
1296 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
1297 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
1298 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
1299 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
1300 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
1301 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
1302 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
1303 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
1304 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
1305 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
1306 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
1307 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
1308 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
1309 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
1310 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
1311 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
1312 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
1313 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
1314 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
1315 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
1316 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
1317 case AMDGPU::S_LOAD_DWORD_IMM:
1318 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1319 case AMDGPU::S_LOAD_DWORDX2_IMM:
1320 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1321 case AMDGPU::S_LOAD_DWORDX4_IMM:
1322 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1323 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e64;
1324 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
1325 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
1329 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
1330 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
1333 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
1334 unsigned OpNo) const {
1335 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1336 const MCInstrDesc &Desc = get(MI.getOpcode());
1337 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
1338 Desc.OpInfo[OpNo].RegClass == -1) {
1339 unsigned Reg = MI.getOperand(OpNo).getReg();
1341 if (TargetRegisterInfo::isVirtualRegister(Reg))
1342 return MRI.getRegClass(Reg);
1343 return RI.getPhysRegClass(Reg);
1346 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
1347 return RI.getRegClass(RCID);
1350 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
1351 switch (MI.getOpcode()) {
1353 case AMDGPU::REG_SEQUENCE:
1355 case AMDGPU::INSERT_SUBREG:
1356 return RI.hasVGPRs(getOpRegClass(MI, 0));
1358 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
1362 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
1363 MachineBasicBlock::iterator I = MI;
1364 MachineBasicBlock *MBB = MI->getParent();
1365 MachineOperand &MO = MI->getOperand(OpIdx);
1366 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1367 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
1368 const TargetRegisterClass *RC = RI.getRegClass(RCID);
1369 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1371 Opcode = AMDGPU::COPY;
1372 else if (RI.isSGPRClass(RC))
1373 Opcode = AMDGPU::S_MOV_B32;
1376 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
1377 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC))
1378 VRC = &AMDGPU::VReg_64RegClass;
1380 VRC = &AMDGPU::VGPR_32RegClass;
1382 unsigned Reg = MRI.createVirtualRegister(VRC);
1383 DebugLoc DL = MBB->findDebugLoc(I);
1384 BuildMI(*MI->getParent(), I, DL, get(Opcode), Reg)
1386 MO.ChangeToRegister(Reg, false);
1389 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
1390 MachineRegisterInfo &MRI,
1391 MachineOperand &SuperReg,
1392 const TargetRegisterClass *SuperRC,
1394 const TargetRegisterClass *SubRC)
1396 assert(SuperReg.isReg());
1398 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
1399 unsigned SubReg = MRI.createVirtualRegister(SubRC);
1401 // Just in case the super register is itself a sub-register, copy it to a new
1402 // value so we don't need to worry about merging its subreg index with the
1403 // SubIdx passed to this function. The register coalescer should be able to
1404 // eliminate this extra copy.
1405 MachineBasicBlock *MBB = MI->getParent();
1406 DebugLoc DL = MI->getDebugLoc();
1408 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), NewSuperReg)
1409 .addReg(SuperReg.getReg(), 0, SuperReg.getSubReg());
1411 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
1412 .addReg(NewSuperReg, 0, SubIdx);
1417 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
1418 MachineBasicBlock::iterator MII,
1419 MachineRegisterInfo &MRI,
1421 const TargetRegisterClass *SuperRC,
1423 const TargetRegisterClass *SubRC) const {
1425 // XXX - Is there a better way to do this?
1426 if (SubIdx == AMDGPU::sub0)
1427 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
1428 if (SubIdx == AMDGPU::sub1)
1429 return MachineOperand::CreateImm(Op.getImm() >> 32);
1431 llvm_unreachable("Unhandled register index for immediate");
1434 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
1436 return MachineOperand::CreateReg(SubReg, false);
1439 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
1440 MachineBasicBlock::iterator MI,
1441 MachineRegisterInfo &MRI,
1442 const TargetRegisterClass *RC,
1443 const MachineOperand &Op) const {
1444 MachineBasicBlock *MBB = MI->getParent();
1445 DebugLoc DL = MI->getDebugLoc();
1446 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1447 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1448 unsigned Dst = MRI.createVirtualRegister(RC);
1450 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1452 .addImm(Op.getImm() & 0xFFFFFFFF);
1453 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1455 .addImm(Op.getImm() >> 32);
1457 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
1459 .addImm(AMDGPU::sub0)
1461 .addImm(AMDGPU::sub1);
1463 Worklist.push_back(Lo);
1464 Worklist.push_back(Hi);
1469 // Change the order of operands from (0, 1, 2) to (0, 2, 1)
1470 void SIInstrInfo::swapOperands(MachineBasicBlock::iterator Inst) const {
1471 assert(Inst->getNumExplicitOperands() == 3);
1472 MachineOperand Op1 = Inst->getOperand(1);
1473 Inst->RemoveOperand(1);
1474 Inst->addOperand(Op1);
1477 bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
1478 const MachineOperand *MO) const {
1479 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1480 const MCInstrDesc &InstDesc = get(MI->getOpcode());
1481 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
1482 const TargetRegisterClass *DefinedRC =
1483 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
1485 MO = &MI->getOperand(OpIdx);
1487 if (isVALU(InstDesc.Opcode) &&
1488 usesConstantBus(MRI, *MO, DefinedRC->getSize())) {
1490 MO->isReg() ? MO->getReg() : (unsigned)AMDGPU::NoRegister;
1491 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1494 const MachineOperand &Op = MI->getOperand(i);
1495 if (Op.isReg() && Op.getReg() != SGPRUsed &&
1496 usesConstantBus(MRI, Op, getOpSize(*MI, i))) {
1504 const TargetRegisterClass *RC = MRI.getRegClass(MO->getReg());
1506 // In order to be legal, the common sub-class must be equal to the
1507 // class of the current operand. For example:
1509 // v_mov_b32 s0 ; Operand defined as vsrc_32
1510 // ; RI.getCommonSubClass(s0,vsrc_32) = sgpr ; LEGAL
1512 // s_sendmsg 0, s0 ; Operand defined as m0reg
1513 // ; RI.getCommonSubClass(s0,m0reg) = m0reg ; NOT LEGAL
1515 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass)) == RC;
1519 // Handle non-register types that are treated like immediates.
1520 assert(MO->isImm() || MO->isTargetIndex() || MO->isFI());
1523 // This operand expects an immediate.
1527 return isImmOperandLegal(MI, OpIdx, *MO);
1530 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
1531 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1533 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1534 AMDGPU::OpName::src0);
1535 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1536 AMDGPU::OpName::src1);
1537 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1538 AMDGPU::OpName::src2);
1541 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
1543 if (!isOperandLegal(MI, Src0Idx))
1544 legalizeOpWithMove(MI, Src0Idx);
1547 if (isOperandLegal(MI, Src1Idx))
1550 // Usually src0 of VOP2 instructions allow more types of inputs
1551 // than src1, so try to commute the instruction to decrease our
1552 // chances of having to insert a MOV instruction to legalize src1.
1553 if (MI->isCommutable()) {
1554 if (commuteInstruction(MI))
1555 // If we are successful in commuting, then we know MI is legal, so
1560 legalizeOpWithMove(MI, Src1Idx);
1564 // XXX - Do any VOP3 instructions read VCC?
1566 if (isVOP3(MI->getOpcode())) {
1567 int VOP3Idx[3] = { Src0Idx, Src1Idx, Src2Idx };
1569 // Find the one SGPR operand we are allowed to use.
1570 unsigned SGPRReg = findUsedSGPR(MI, VOP3Idx);
1572 for (unsigned i = 0; i < 3; ++i) {
1573 int Idx = VOP3Idx[i];
1576 MachineOperand &MO = MI->getOperand(Idx);
1579 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
1580 continue; // VGPRs are legal
1582 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
1584 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
1585 SGPRReg = MO.getReg();
1586 // We can use one SGPR in each VOP3 instruction.
1589 } else if (!isLiteralConstant(MO, getOpSize(MI->getOpcode(), Idx))) {
1590 // If it is not a register and not a literal constant, then it must be
1591 // an inline constant which is always legal.
1594 // If we make it this far, then the operand is not legal and we must
1596 legalizeOpWithMove(MI, Idx);
1600 // Legalize REG_SEQUENCE and PHI
1601 // The register class of the operands much be the same type as the register
1602 // class of the output.
1603 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
1604 MI->getOpcode() == AMDGPU::PHI) {
1605 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
1606 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1607 if (!MI->getOperand(i).isReg() ||
1608 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1610 const TargetRegisterClass *OpRC =
1611 MRI.getRegClass(MI->getOperand(i).getReg());
1612 if (RI.hasVGPRs(OpRC)) {
1619 // If any of the operands are VGPR registers, then they all most be
1620 // otherwise we will create illegal VGPR->SGPR copies when legalizing
1622 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
1625 VRC = RI.getEquivalentVGPRClass(SRC);
1632 // Update all the operands so they have the same type.
1633 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1634 if (!MI->getOperand(i).isReg() ||
1635 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1637 unsigned DstReg = MRI.createVirtualRegister(RC);
1638 MachineBasicBlock *InsertBB;
1639 MachineBasicBlock::iterator Insert;
1640 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
1641 InsertBB = MI->getParent();
1644 // MI is a PHI instruction.
1645 InsertBB = MI->getOperand(i + 1).getMBB();
1646 Insert = InsertBB->getFirstTerminator();
1648 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
1649 get(AMDGPU::COPY), DstReg)
1650 .addOperand(MI->getOperand(i));
1651 MI->getOperand(i).setReg(DstReg);
1655 // Legalize INSERT_SUBREG
1656 // src0 must have the same register class as dst
1657 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
1658 unsigned Dst = MI->getOperand(0).getReg();
1659 unsigned Src0 = MI->getOperand(1).getReg();
1660 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
1661 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
1662 if (DstRC != Src0RC) {
1663 MachineBasicBlock &MBB = *MI->getParent();
1664 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
1665 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
1667 MI->getOperand(1).setReg(NewSrc0);
1672 // Legalize MUBUF* instructions
1673 // FIXME: If we start using the non-addr64 instructions for compute, we
1674 // may need to legalize them here.
1676 AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::srsrc);
1677 if (SRsrcIdx != -1) {
1678 // We have an MUBUF instruction
1679 MachineOperand *SRsrc = &MI->getOperand(SRsrcIdx);
1680 unsigned SRsrcRC = get(MI->getOpcode()).OpInfo[SRsrcIdx].RegClass;
1681 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
1682 RI.getRegClass(SRsrcRC))) {
1683 // The operands are legal.
1684 // FIXME: We may need to legalize operands besided srsrc.
1688 MachineBasicBlock &MBB = *MI->getParent();
1689 // Extract the the ptr from the resource descriptor.
1691 // SRsrcPtrLo = srsrc:sub0
1692 unsigned SRsrcPtrLo = buildExtractSubReg(MI, MRI, *SRsrc,
1693 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VGPR_32RegClass);
1695 // SRsrcPtrHi = srsrc:sub1
1696 unsigned SRsrcPtrHi = buildExtractSubReg(MI, MRI, *SRsrc,
1697 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VGPR_32RegClass);
1699 // Create an empty resource descriptor
1700 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1701 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1702 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1703 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1704 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
1707 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1711 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1712 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1714 .addImm(RsrcDataFormat & 0xFFFFFFFF);
1716 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1717 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1719 .addImm(RsrcDataFormat >> 32);
1721 // NewSRsrc = {Zero64, SRsrcFormat}
1722 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1725 .addImm(AMDGPU::sub0_sub1)
1726 .addReg(SRsrcFormatLo)
1727 .addImm(AMDGPU::sub2)
1728 .addReg(SRsrcFormatHi)
1729 .addImm(AMDGPU::sub3);
1731 MachineOperand *VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1732 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
1733 unsigned NewVAddrLo;
1734 unsigned NewVAddrHi;
1736 // This is already an ADDR64 instruction so we need to add the pointer
1737 // extracted from the resource descriptor to the current value of VAddr.
1738 NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1739 NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1741 // NewVaddrLo = SRsrcPtrLo + VAddr:sub0
1742 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1745 .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
1746 .addReg(AMDGPU::VCC, RegState::ImplicitDefine);
1748 // NewVaddrHi = SRsrcPtrHi + VAddr:sub1
1749 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1752 .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
1753 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1754 .addReg(AMDGPU::VCC, RegState::Implicit);
1757 // This instructions is the _OFFSET variant, so we need to convert it to
1759 MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
1760 MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
1761 MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
1763 // Create the new instruction.
1764 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
1765 MachineInstr *Addr64 =
1766 BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
1769 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
1770 // This will be replaced later
1771 // with the new value of vaddr.
1772 .addOperand(*SOffset)
1773 .addOperand(*Offset);
1775 MI->removeFromParent();
1778 NewVAddrLo = SRsrcPtrLo;
1779 NewVAddrHi = SRsrcPtrHi;
1780 VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1781 SRsrc = getNamedOperand(*MI, AMDGPU::OpName::srsrc);
1784 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1785 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1788 .addImm(AMDGPU::sub0)
1790 .addImm(AMDGPU::sub1);
1793 // Update the instruction to use NewVaddr
1794 VAddr->setReg(NewVAddr);
1795 // Update the instruction to use NewSRsrc
1796 SRsrc->setReg(NewSRsrc);
1800 void SIInstrInfo::splitSMRD(MachineInstr *MI,
1801 const TargetRegisterClass *HalfRC,
1802 unsigned HalfImmOp, unsigned HalfSGPROp,
1803 MachineInstr *&Lo, MachineInstr *&Hi) const {
1805 DebugLoc DL = MI->getDebugLoc();
1806 MachineBasicBlock *MBB = MI->getParent();
1807 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1808 unsigned RegLo = MRI.createVirtualRegister(HalfRC);
1809 unsigned RegHi = MRI.createVirtualRegister(HalfRC);
1810 unsigned HalfSize = HalfRC->getSize();
1811 const MachineOperand *OffOp =
1812 getNamedOperand(*MI, AMDGPU::OpName::offset);
1813 const MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
1815 // The SMRD has an 8-bit offset in dwords on SI and a 20-bit offset in bytes
1818 bool isVI = RI.ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS;
1819 unsigned OffScale = isVI ? 1 : 4;
1820 // Handle the _IMM variant
1821 unsigned LoOffset = OffOp->getImm() * OffScale;
1822 unsigned HiOffset = LoOffset + HalfSize;
1823 Lo = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegLo)
1825 .addImm(LoOffset / OffScale);
1827 if (!isUInt<20>(HiOffset) || (!isVI && !isUInt<8>(HiOffset / OffScale))) {
1828 unsigned OffsetSGPR =
1829 MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1830 BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32), OffsetSGPR)
1831 .addImm(HiOffset); // The offset in register is in bytes.
1832 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
1834 .addReg(OffsetSGPR);
1836 Hi = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegHi)
1838 .addImm(HiOffset / OffScale);
1841 // Handle the _SGPR variant
1842 MachineOperand *SOff = getNamedOperand(*MI, AMDGPU::OpName::soff);
1843 Lo = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegLo)
1846 unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1847 BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
1850 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp))
1852 .addReg(OffsetSGPR);
1855 unsigned SubLo, SubHi;
1858 SubLo = AMDGPU::sub0;
1859 SubHi = AMDGPU::sub1;
1862 SubLo = AMDGPU::sub0_sub1;
1863 SubHi = AMDGPU::sub2_sub3;
1866 SubLo = AMDGPU::sub0_sub1_sub2_sub3;
1867 SubHi = AMDGPU::sub4_sub5_sub6_sub7;
1870 SubLo = AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
1871 SubHi = AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15;
1874 llvm_unreachable("Unhandled HalfSize");
1877 BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE))
1878 .addOperand(MI->getOperand(0))
1885 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
1886 MachineBasicBlock *MBB = MI->getParent();
1887 switch (MI->getOpcode()) {
1888 case AMDGPU::S_LOAD_DWORD_IMM:
1889 case AMDGPU::S_LOAD_DWORD_SGPR:
1890 case AMDGPU::S_LOAD_DWORDX2_IMM:
1891 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1892 case AMDGPU::S_LOAD_DWORDX4_IMM:
1893 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1894 unsigned NewOpcode = getVALUOp(*MI);
1898 if (MI->getOperand(2).isReg()) {
1899 RegOffset = MI->getOperand(2).getReg();
1902 assert(MI->getOperand(2).isImm());
1903 // SMRD instructions take a dword offsets on SI and byte offset on VI
1904 // and MUBUF instructions always take a byte offset.
1905 ImmOffset = MI->getOperand(2).getImm();
1906 if (RI.ST.getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS)
1908 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1910 if (isUInt<12>(ImmOffset)) {
1911 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1915 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1922 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1923 unsigned DWord0 = RegOffset;
1924 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1925 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1926 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1927 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
1929 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
1931 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
1932 .addImm(RsrcDataFormat & 0xFFFFFFFF);
1933 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
1934 .addImm(RsrcDataFormat >> 32);
1935 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
1937 .addImm(AMDGPU::sub0)
1939 .addImm(AMDGPU::sub1)
1941 .addImm(AMDGPU::sub2)
1943 .addImm(AMDGPU::sub3);
1944 MI->setDesc(get(NewOpcode));
1945 if (MI->getOperand(2).isReg()) {
1946 MI->getOperand(2).setReg(MI->getOperand(1).getReg());
1948 MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
1950 MI->getOperand(1).setReg(SRsrc);
1951 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(0));
1952 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
1954 const TargetRegisterClass *NewDstRC =
1955 RI.getRegClass(get(NewOpcode).OpInfo[0].RegClass);
1957 unsigned DstReg = MI->getOperand(0).getReg();
1958 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1959 MRI.replaceRegWith(DstReg, NewDstReg);
1962 case AMDGPU::S_LOAD_DWORDX8_IMM:
1963 case AMDGPU::S_LOAD_DWORDX8_SGPR: {
1964 MachineInstr *Lo, *Hi;
1965 splitSMRD(MI, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM,
1966 AMDGPU::S_LOAD_DWORDX4_SGPR, Lo, Hi);
1967 MI->eraseFromParent();
1968 moveSMRDToVALU(Lo, MRI);
1969 moveSMRDToVALU(Hi, MRI);
1973 case AMDGPU::S_LOAD_DWORDX16_IMM:
1974 case AMDGPU::S_LOAD_DWORDX16_SGPR: {
1975 MachineInstr *Lo, *Hi;
1976 splitSMRD(MI, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM,
1977 AMDGPU::S_LOAD_DWORDX8_SGPR, Lo, Hi);
1978 MI->eraseFromParent();
1979 moveSMRDToVALU(Lo, MRI);
1980 moveSMRDToVALU(Hi, MRI);
1986 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
1987 SmallVector<MachineInstr *, 128> Worklist;
1988 Worklist.push_back(&TopInst);
1990 while (!Worklist.empty()) {
1991 MachineInstr *Inst = Worklist.pop_back_val();
1992 MachineBasicBlock *MBB = Inst->getParent();
1993 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1995 unsigned Opcode = Inst->getOpcode();
1996 unsigned NewOpcode = getVALUOp(*Inst);
1998 // Handle some special cases
2001 if (isSMRD(Inst->getOpcode())) {
2002 moveSMRDToVALU(Inst, MRI);
2005 case AMDGPU::S_MOV_B64: {
2006 DebugLoc DL = Inst->getDebugLoc();
2008 // If the source operand is a register we can replace this with a
2010 if (Inst->getOperand(1).isReg()) {
2011 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
2012 .addOperand(Inst->getOperand(0))
2013 .addOperand(Inst->getOperand(1));
2014 Worklist.push_back(Copy);
2016 // Otherwise, we need to split this into two movs, because there is
2017 // no 64-bit VALU move instruction.
2018 unsigned Reg = Inst->getOperand(0).getReg();
2019 unsigned Dst = split64BitImm(Worklist,
2022 MRI.getRegClass(Reg),
2023 Inst->getOperand(1));
2024 MRI.replaceRegWith(Reg, Dst);
2026 Inst->eraseFromParent();
2029 case AMDGPU::S_AND_B64:
2030 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
2031 Inst->eraseFromParent();
2034 case AMDGPU::S_OR_B64:
2035 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
2036 Inst->eraseFromParent();
2039 case AMDGPU::S_XOR_B64:
2040 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
2041 Inst->eraseFromParent();
2044 case AMDGPU::S_NOT_B64:
2045 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
2046 Inst->eraseFromParent();
2049 case AMDGPU::S_BCNT1_I32_B64:
2050 splitScalar64BitBCNT(Worklist, Inst);
2051 Inst->eraseFromParent();
2054 case AMDGPU::S_BFE_I64: {
2055 splitScalar64BitBFE(Worklist, Inst);
2056 Inst->eraseFromParent();
2060 case AMDGPU::S_LSHL_B32:
2061 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2062 NewOpcode = AMDGPU::V_LSHLREV_B32_e64;
2066 case AMDGPU::S_ASHR_I32:
2067 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2068 NewOpcode = AMDGPU::V_ASHRREV_I32_e64;
2072 case AMDGPU::S_LSHR_B32:
2073 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2074 NewOpcode = AMDGPU::V_LSHRREV_B32_e64;
2078 case AMDGPU::S_LSHL_B64:
2079 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2080 NewOpcode = AMDGPU::V_LSHLREV_B64;
2084 case AMDGPU::S_ASHR_I64:
2085 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2086 NewOpcode = AMDGPU::V_ASHRREV_I64;
2090 case AMDGPU::S_LSHR_B64:
2091 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2092 NewOpcode = AMDGPU::V_LSHRREV_B64;
2097 case AMDGPU::S_BFE_U64:
2098 case AMDGPU::S_BFM_B64:
2099 llvm_unreachable("Moving this op to VALU not implemented");
2102 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
2103 // We cannot move this instruction to the VALU, so we should try to
2104 // legalize its operands instead.
2105 legalizeOperands(Inst);
2109 // Use the new VALU Opcode.
2110 const MCInstrDesc &NewDesc = get(NewOpcode);
2111 Inst->setDesc(NewDesc);
2113 // Remove any references to SCC. Vector instructions can't read from it, and
2114 // We're just about to add the implicit use / defs of VCC, and we don't want
2116 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
2117 MachineOperand &Op = Inst->getOperand(i);
2118 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
2119 Inst->RemoveOperand(i);
2122 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
2123 // We are converting these to a BFE, so we need to add the missing
2124 // operands for the size and offset.
2125 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
2126 Inst->addOperand(MachineOperand::CreateImm(0));
2127 Inst->addOperand(MachineOperand::CreateImm(Size));
2129 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
2130 // The VALU version adds the second operand to the result, so insert an
2132 Inst->addOperand(MachineOperand::CreateImm(0));
2135 addDescImplicitUseDef(NewDesc, Inst);
2137 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
2138 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
2139 // If we need to move this to VGPRs, we need to unpack the second operand
2140 // back into the 2 separate ones for bit offset and width.
2141 assert(OffsetWidthOp.isImm() &&
2142 "Scalar BFE is only implemented for constant width and offset");
2143 uint32_t Imm = OffsetWidthOp.getImm();
2145 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
2146 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
2147 Inst->RemoveOperand(2); // Remove old immediate.
2148 Inst->addOperand(MachineOperand::CreateImm(Offset));
2149 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
2152 // Update the destination register class.
2154 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
2157 // For target instructions, getOpRegClass just returns the virtual
2158 // register class associated with the operand, so we need to find an
2159 // equivalent VGPR register class in order to move the instruction to the
2163 case AMDGPU::REG_SEQUENCE:
2164 case AMDGPU::INSERT_SUBREG:
2165 if (RI.hasVGPRs(NewDstRC))
2167 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
2175 unsigned DstReg = Inst->getOperand(0).getReg();
2176 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
2177 MRI.replaceRegWith(DstReg, NewDstReg);
2179 // Legalize the operands
2180 legalizeOperands(Inst);
2182 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
2183 E = MRI.use_end(); I != E; ++I) {
2184 MachineInstr &UseMI = *I->getParent();
2185 if (!canReadVGPR(UseMI, I.getOperandNo())) {
2186 Worklist.push_back(&UseMI);
2192 //===----------------------------------------------------------------------===//
2193 // Indirect addressing callbacks
2194 //===----------------------------------------------------------------------===//
2196 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
2197 unsigned Channel) const {
2198 assert(Channel == 0);
2202 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
2203 return &AMDGPU::VGPR_32RegClass;
2206 void SIInstrInfo::splitScalar64BitUnaryOp(
2207 SmallVectorImpl<MachineInstr *> &Worklist,
2209 unsigned Opcode) const {
2210 MachineBasicBlock &MBB = *Inst->getParent();
2211 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2213 MachineOperand &Dest = Inst->getOperand(0);
2214 MachineOperand &Src0 = Inst->getOperand(1);
2215 DebugLoc DL = Inst->getDebugLoc();
2217 MachineBasicBlock::iterator MII = Inst;
2219 const MCInstrDesc &InstDesc = get(Opcode);
2220 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2221 MRI.getRegClass(Src0.getReg()) :
2222 &AMDGPU::SGPR_32RegClass;
2224 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2226 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2227 AMDGPU::sub0, Src0SubRC);
2229 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2230 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2232 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2233 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2234 .addOperand(SrcReg0Sub0);
2236 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2237 AMDGPU::sub1, Src0SubRC);
2239 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2240 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2241 .addOperand(SrcReg0Sub1);
2243 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2244 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2246 .addImm(AMDGPU::sub0)
2248 .addImm(AMDGPU::sub1);
2250 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2252 // Try to legalize the operands in case we need to swap the order to keep it
2254 Worklist.push_back(LoHalf);
2255 Worklist.push_back(HiHalf);
2258 void SIInstrInfo::splitScalar64BitBinaryOp(
2259 SmallVectorImpl<MachineInstr *> &Worklist,
2261 unsigned Opcode) const {
2262 MachineBasicBlock &MBB = *Inst->getParent();
2263 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2265 MachineOperand &Dest = Inst->getOperand(0);
2266 MachineOperand &Src0 = Inst->getOperand(1);
2267 MachineOperand &Src1 = Inst->getOperand(2);
2268 DebugLoc DL = Inst->getDebugLoc();
2270 MachineBasicBlock::iterator MII = Inst;
2272 const MCInstrDesc &InstDesc = get(Opcode);
2273 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2274 MRI.getRegClass(Src0.getReg()) :
2275 &AMDGPU::SGPR_32RegClass;
2277 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2278 const TargetRegisterClass *Src1RC = Src1.isReg() ?
2279 MRI.getRegClass(Src1.getReg()) :
2280 &AMDGPU::SGPR_32RegClass;
2282 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
2284 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2285 AMDGPU::sub0, Src0SubRC);
2286 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2287 AMDGPU::sub0, Src1SubRC);
2289 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2290 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2292 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2293 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2294 .addOperand(SrcReg0Sub0)
2295 .addOperand(SrcReg1Sub0);
2297 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2298 AMDGPU::sub1, Src0SubRC);
2299 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2300 AMDGPU::sub1, Src1SubRC);
2302 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2303 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2304 .addOperand(SrcReg0Sub1)
2305 .addOperand(SrcReg1Sub1);
2307 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2308 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2310 .addImm(AMDGPU::sub0)
2312 .addImm(AMDGPU::sub1);
2314 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2316 // Try to legalize the operands in case we need to swap the order to keep it
2318 Worklist.push_back(LoHalf);
2319 Worklist.push_back(HiHalf);
2322 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
2323 MachineInstr *Inst) const {
2324 MachineBasicBlock &MBB = *Inst->getParent();
2325 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2327 MachineBasicBlock::iterator MII = Inst;
2328 DebugLoc DL = Inst->getDebugLoc();
2330 MachineOperand &Dest = Inst->getOperand(0);
2331 MachineOperand &Src = Inst->getOperand(1);
2333 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e64);
2334 const TargetRegisterClass *SrcRC = Src.isReg() ?
2335 MRI.getRegClass(Src.getReg()) :
2336 &AMDGPU::SGPR_32RegClass;
2338 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2339 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2341 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
2343 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2344 AMDGPU::sub0, SrcSubRC);
2345 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2346 AMDGPU::sub1, SrcSubRC);
2348 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
2349 .addOperand(SrcRegSub0)
2352 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
2353 .addOperand(SrcRegSub1)
2356 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2358 Worklist.push_back(First);
2359 Worklist.push_back(Second);
2362 void SIInstrInfo::splitScalar64BitBFE(SmallVectorImpl<MachineInstr *> &Worklist,
2363 MachineInstr *Inst) const {
2364 MachineBasicBlock &MBB = *Inst->getParent();
2365 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2366 MachineBasicBlock::iterator MII = Inst;
2367 DebugLoc DL = Inst->getDebugLoc();
2369 MachineOperand &Dest = Inst->getOperand(0);
2370 uint32_t Imm = Inst->getOperand(2).getImm();
2371 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
2372 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
2376 // Only sext_inreg cases handled.
2377 assert(Inst->getOpcode() == AMDGPU::S_BFE_I64 &&
2382 if (BitWidth < 32) {
2383 unsigned MidRegLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2384 unsigned MidRegHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2385 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
2387 BuildMI(MBB, MII, DL, get(AMDGPU::V_BFE_I32), MidRegLo)
2388 .addReg(Inst->getOperand(1).getReg(), 0, AMDGPU::sub0)
2392 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e32), MidRegHi)
2396 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
2398 .addImm(AMDGPU::sub0)
2400 .addImm(AMDGPU::sub1);
2402 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2406 MachineOperand &Src = Inst->getOperand(1);
2407 unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2408 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
2410 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e64), TmpReg)
2412 .addReg(Src.getReg(), 0, AMDGPU::sub0);
2414 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
2415 .addReg(Src.getReg(), 0, AMDGPU::sub0)
2416 .addImm(AMDGPU::sub0)
2418 .addImm(AMDGPU::sub1);
2420 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2423 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
2424 MachineInstr *Inst) const {
2425 // Add the implict and explicit register definitions.
2426 if (NewDesc.ImplicitUses) {
2427 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
2428 unsigned Reg = NewDesc.ImplicitUses[i];
2429 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
2433 if (NewDesc.ImplicitDefs) {
2434 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
2435 unsigned Reg = NewDesc.ImplicitDefs[i];
2436 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
2441 unsigned SIInstrInfo::findUsedSGPR(const MachineInstr *MI,
2442 int OpIndices[3]) const {
2443 const MCInstrDesc &Desc = get(MI->getOpcode());
2445 // Find the one SGPR operand we are allowed to use.
2446 unsigned SGPRReg = AMDGPU::NoRegister;
2448 // First we need to consider the instruction's operand requirements before
2449 // legalizing. Some operands are required to be SGPRs, such as implicit uses
2450 // of VCC, but we are still bound by the constant bus requirement to only use
2453 // If the operand's class is an SGPR, we can never move it.
2455 for (const MachineOperand &MO : MI->implicit_operands()) {
2456 // We only care about reads.
2460 if (MO.getReg() == AMDGPU::VCC)
2463 if (MO.getReg() == AMDGPU::FLAT_SCR)
2464 return AMDGPU::FLAT_SCR;
2467 unsigned UsedSGPRs[3] = { AMDGPU::NoRegister };
2468 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
2470 for (unsigned i = 0; i < 3; ++i) {
2471 int Idx = OpIndices[i];
2475 const MachineOperand &MO = MI->getOperand(Idx);
2476 if (RI.isSGPRClassID(Desc.OpInfo[Idx].RegClass))
2477 SGPRReg = MO.getReg();
2479 if (MO.isReg() && RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
2480 UsedSGPRs[i] = MO.getReg();
2483 if (SGPRReg != AMDGPU::NoRegister)
2486 // We don't have a required SGPR operand, so we have a bit more freedom in
2487 // selecting operands to move.
2489 // Try to select the most used SGPR. If an SGPR is equal to one of the
2490 // others, we choose that.
2493 // V_FMA_F32 v0, s0, s0, s0 -> No moves
2494 // V_FMA_F32 v0, s0, s1, s0 -> Move s1
2496 if (UsedSGPRs[0] != AMDGPU::NoRegister) {
2497 if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
2498 SGPRReg = UsedSGPRs[0];
2501 if (SGPRReg == AMDGPU::NoRegister && UsedSGPRs[1] != AMDGPU::NoRegister) {
2502 if (UsedSGPRs[1] == UsedSGPRs[2])
2503 SGPRReg = UsedSGPRs[1];
2509 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
2510 MachineBasicBlock *MBB,
2511 MachineBasicBlock::iterator I,
2513 unsigned Address, unsigned OffsetReg) const {
2514 const DebugLoc &DL = MBB->findDebugLoc(I);
2515 unsigned IndirectBaseReg = AMDGPU::VGPR_32RegClass.getRegister(
2516 getIndirectIndexBegin(*MBB->getParent()));
2518 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
2519 .addReg(IndirectBaseReg, RegState::Define)
2520 .addOperand(I->getOperand(0))
2521 .addReg(IndirectBaseReg)
2527 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
2528 MachineBasicBlock *MBB,
2529 MachineBasicBlock::iterator I,
2531 unsigned Address, unsigned OffsetReg) const {
2532 const DebugLoc &DL = MBB->findDebugLoc(I);
2533 unsigned IndirectBaseReg = AMDGPU::VGPR_32RegClass.getRegister(
2534 getIndirectIndexBegin(*MBB->getParent()));
2536 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
2537 .addOperand(I->getOperand(0))
2538 .addOperand(I->getOperand(1))
2539 .addReg(IndirectBaseReg)
2545 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
2546 const MachineFunction &MF) const {
2547 int End = getIndirectIndexEnd(MF);
2548 int Begin = getIndirectIndexBegin(MF);
2554 for (int Index = Begin; Index <= End; ++Index)
2555 Reserved.set(AMDGPU::VGPR_32RegClass.getRegister(Index));
2557 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
2558 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
2560 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
2561 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
2563 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
2564 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
2566 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
2567 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
2569 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
2570 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));
2573 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
2574 unsigned OperandName) const {
2575 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
2579 return &MI.getOperand(Idx);
2582 uint64_t SIInstrInfo::getDefaultRsrcDataFormat() const {
2583 uint64_t RsrcDataFormat = AMDGPU::RSRC_DATA_FORMAT;
2584 if (ST.isAmdHsaOS())
2585 RsrcDataFormat |= (1ULL << 56);
2587 return RsrcDataFormat;