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 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
336 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
337 .addReg(SrcReg, getKillRegState(KillSrc));
340 } else if (AMDGPU::SReg_128RegClass.contains(DestReg)) {
341 assert(AMDGPU::SReg_128RegClass.contains(SrcReg));
342 Opcode = AMDGPU::S_MOV_B32;
345 } else if (AMDGPU::SReg_256RegClass.contains(DestReg)) {
346 assert(AMDGPU::SReg_256RegClass.contains(SrcReg));
347 Opcode = AMDGPU::S_MOV_B32;
350 } else if (AMDGPU::SReg_512RegClass.contains(DestReg)) {
351 assert(AMDGPU::SReg_512RegClass.contains(SrcReg));
352 Opcode = AMDGPU::S_MOV_B32;
353 SubIndices = Sub0_15;
355 } else if (AMDGPU::VGPR_32RegClass.contains(DestReg)) {
356 assert(AMDGPU::VGPR_32RegClass.contains(SrcReg) ||
357 AMDGPU::SReg_32RegClass.contains(SrcReg));
358 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
359 .addReg(SrcReg, getKillRegState(KillSrc));
362 } else if (AMDGPU::VReg_64RegClass.contains(DestReg)) {
363 assert(AMDGPU::VReg_64RegClass.contains(SrcReg) ||
364 AMDGPU::SReg_64RegClass.contains(SrcReg));
365 Opcode = AMDGPU::V_MOV_B32_e32;
368 } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
369 assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
370 Opcode = AMDGPU::V_MOV_B32_e32;
373 } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
374 assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
375 AMDGPU::SReg_128RegClass.contains(SrcReg));
376 Opcode = AMDGPU::V_MOV_B32_e32;
379 } else if (AMDGPU::VReg_256RegClass.contains(DestReg)) {
380 assert(AMDGPU::VReg_256RegClass.contains(SrcReg) ||
381 AMDGPU::SReg_256RegClass.contains(SrcReg));
382 Opcode = AMDGPU::V_MOV_B32_e32;
385 } else if (AMDGPU::VReg_512RegClass.contains(DestReg)) {
386 assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
387 AMDGPU::SReg_512RegClass.contains(SrcReg));
388 Opcode = AMDGPU::V_MOV_B32_e32;
389 SubIndices = Sub0_15;
392 llvm_unreachable("Can't copy register!");
395 while (unsigned SubIdx = *SubIndices++) {
396 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
397 get(Opcode), RI.getSubReg(DestReg, SubIdx));
399 Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
402 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
406 unsigned SIInstrInfo::commuteOpcode(unsigned Opcode) const {
409 // Try to map original to commuted opcode
410 if ((NewOpc = AMDGPU::getCommuteRev(Opcode)) != -1)
413 // Try to map commuted to original opcode
414 if ((NewOpc = AMDGPU::getCommuteOrig(Opcode)) != -1)
420 unsigned SIInstrInfo::getMovOpcode(const TargetRegisterClass *DstRC) const {
422 if (DstRC->getSize() == 4) {
423 return RI.isSGPRClass(DstRC) ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32;
424 } else if (DstRC->getSize() == 8 && RI.isSGPRClass(DstRC)) {
425 return AMDGPU::S_MOV_B64;
426 } else if (DstRC->getSize() == 8 && !RI.isSGPRClass(DstRC)) {
427 return AMDGPU::V_MOV_B64_PSEUDO;
432 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
433 MachineBasicBlock::iterator MI,
434 unsigned SrcReg, bool isKill,
436 const TargetRegisterClass *RC,
437 const TargetRegisterInfo *TRI) const {
438 MachineFunction *MF = MBB.getParent();
439 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
440 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
441 DebugLoc DL = MBB.findDebugLoc(MI);
444 if (RI.isSGPRClass(RC)) {
445 // We are only allowed to create one new instruction when spilling
446 // registers, so we need to use pseudo instruction for spilling
448 switch (RC->getSize() * 8) {
449 case 32: Opcode = AMDGPU::SI_SPILL_S32_SAVE; break;
450 case 64: Opcode = AMDGPU::SI_SPILL_S64_SAVE; break;
451 case 128: Opcode = AMDGPU::SI_SPILL_S128_SAVE; break;
452 case 256: Opcode = AMDGPU::SI_SPILL_S256_SAVE; break;
453 case 512: Opcode = AMDGPU::SI_SPILL_S512_SAVE; break;
455 } else if(RI.hasVGPRs(RC) && ST.isVGPRSpillingEnabled(MFI)) {
456 MFI->setHasSpilledVGPRs();
458 switch(RC->getSize() * 8) {
459 case 32: Opcode = AMDGPU::SI_SPILL_V32_SAVE; break;
460 case 64: Opcode = AMDGPU::SI_SPILL_V64_SAVE; break;
461 case 96: Opcode = AMDGPU::SI_SPILL_V96_SAVE; break;
462 case 128: Opcode = AMDGPU::SI_SPILL_V128_SAVE; break;
463 case 256: Opcode = AMDGPU::SI_SPILL_V256_SAVE; break;
464 case 512: Opcode = AMDGPU::SI_SPILL_V512_SAVE; break;
469 FrameInfo->setObjectAlignment(FrameIndex, 4);
470 BuildMI(MBB, MI, DL, get(Opcode))
472 .addFrameIndex(FrameIndex)
473 // Place-holder registers, these will be filled in by
474 // SIPrepareScratchRegs.
475 .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
476 .addReg(AMDGPU::SGPR0, RegState::Undef);
478 LLVMContext &Ctx = MF->getFunction()->getContext();
479 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Do not know how to"
481 BuildMI(MBB, MI, DL, get(AMDGPU::KILL))
486 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
487 MachineBasicBlock::iterator MI,
488 unsigned DestReg, int FrameIndex,
489 const TargetRegisterClass *RC,
490 const TargetRegisterInfo *TRI) const {
491 MachineFunction *MF = MBB.getParent();
492 const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
493 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
494 DebugLoc DL = MBB.findDebugLoc(MI);
497 if (RI.isSGPRClass(RC)){
498 switch(RC->getSize() * 8) {
499 case 32: Opcode = AMDGPU::SI_SPILL_S32_RESTORE; break;
500 case 64: Opcode = AMDGPU::SI_SPILL_S64_RESTORE; break;
501 case 128: Opcode = AMDGPU::SI_SPILL_S128_RESTORE; break;
502 case 256: Opcode = AMDGPU::SI_SPILL_S256_RESTORE; break;
503 case 512: Opcode = AMDGPU::SI_SPILL_S512_RESTORE; break;
505 } else if(RI.hasVGPRs(RC) && ST.isVGPRSpillingEnabled(MFI)) {
506 switch(RC->getSize() * 8) {
507 case 32: Opcode = AMDGPU::SI_SPILL_V32_RESTORE; break;
508 case 64: Opcode = AMDGPU::SI_SPILL_V64_RESTORE; break;
509 case 96: Opcode = AMDGPU::SI_SPILL_V96_RESTORE; break;
510 case 128: Opcode = AMDGPU::SI_SPILL_V128_RESTORE; break;
511 case 256: Opcode = AMDGPU::SI_SPILL_V256_RESTORE; break;
512 case 512: Opcode = AMDGPU::SI_SPILL_V512_RESTORE; break;
517 FrameInfo->setObjectAlignment(FrameIndex, 4);
518 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
519 .addFrameIndex(FrameIndex)
520 // Place-holder registers, these will be filled in by
521 // SIPrepareScratchRegs.
522 .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
523 .addReg(AMDGPU::SGPR0, RegState::Undef);
526 LLVMContext &Ctx = MF->getFunction()->getContext();
527 Ctx.emitError("SIInstrInfo::loadRegFromStackSlot - Do not know how to"
528 " restore register");
529 BuildMI(MBB, MI, DL, get(AMDGPU::IMPLICIT_DEF), DestReg);
533 /// \param @Offset Offset in bytes of the FrameIndex being spilled
534 unsigned SIInstrInfo::calculateLDSSpillAddress(MachineBasicBlock &MBB,
535 MachineBasicBlock::iterator MI,
536 RegScavenger *RS, unsigned TmpReg,
537 unsigned FrameOffset,
538 unsigned Size) const {
539 MachineFunction *MF = MBB.getParent();
540 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
541 const AMDGPUSubtarget &ST = MF->getSubtarget<AMDGPUSubtarget>();
542 const SIRegisterInfo *TRI =
543 static_cast<const SIRegisterInfo*>(ST.getRegisterInfo());
544 DebugLoc DL = MBB.findDebugLoc(MI);
545 unsigned WorkGroupSize = MFI->getMaximumWorkGroupSize(*MF);
546 unsigned WavefrontSize = ST.getWavefrontSize();
548 unsigned TIDReg = MFI->getTIDReg();
549 if (!MFI->hasCalculatedTID()) {
550 MachineBasicBlock &Entry = MBB.getParent()->front();
551 MachineBasicBlock::iterator Insert = Entry.front();
552 DebugLoc DL = Insert->getDebugLoc();
554 TIDReg = RI.findUnusedRegister(MF->getRegInfo(), &AMDGPU::VGPR_32RegClass);
555 if (TIDReg == AMDGPU::NoRegister)
559 if (MFI->getShaderType() == ShaderType::COMPUTE &&
560 WorkGroupSize > WavefrontSize) {
562 unsigned TIDIGXReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_X);
563 unsigned TIDIGYReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Y);
564 unsigned TIDIGZReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Z);
565 unsigned InputPtrReg =
566 TRI->getPreloadedValue(*MF, SIRegisterInfo::INPUT_PTR);
567 static const unsigned TIDIGRegs[3] = {
568 TIDIGXReg, TIDIGYReg, TIDIGZReg
570 for (unsigned Reg : TIDIGRegs) {
571 if (!Entry.isLiveIn(Reg))
572 Entry.addLiveIn(Reg);
575 RS->enterBasicBlock(&Entry);
576 unsigned STmp0 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
577 unsigned STmp1 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
578 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp0)
580 .addImm(SI::KernelInputOffsets::NGROUPS_Z);
581 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp1)
583 .addImm(SI::KernelInputOffsets::NGROUPS_Y);
585 // NGROUPS.X * NGROUPS.Y
586 BuildMI(Entry, Insert, DL, get(AMDGPU::S_MUL_I32), STmp1)
589 // (NGROUPS.X * NGROUPS.Y) * TIDIG.X
590 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MUL_U32_U24_e32), TIDReg)
593 // NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)
594 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MAD_U32_U24), TIDReg)
598 // (NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)) + TIDIG.Z
599 BuildMI(Entry, Insert, DL, get(AMDGPU::V_ADD_I32_e32), TIDReg)
604 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_LO_U32_B32_e64),
609 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_HI_U32_B32_e64),
615 BuildMI(Entry, Insert, DL, get(AMDGPU::V_LSHLREV_B32_e32),
619 MFI->setTIDReg(TIDReg);
622 // Add FrameIndex to LDS offset
623 unsigned LDSOffset = MFI->LDSSize + (FrameOffset * WorkGroupSize);
624 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_I32_e32), TmpReg)
631 void SIInstrInfo::insertNOPs(MachineBasicBlock::iterator MI,
640 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(AMDGPU::S_NOP))
645 bool SIInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
646 MachineBasicBlock &MBB = *MI->getParent();
647 DebugLoc DL = MBB.findDebugLoc(MI);
648 switch (MI->getOpcode()) {
649 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
651 case AMDGPU::SI_CONSTDATA_PTR: {
652 unsigned Reg = MI->getOperand(0).getReg();
653 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
654 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
656 BuildMI(MBB, MI, DL, get(AMDGPU::S_GETPC_B64), Reg);
658 // Add 32-bit offset from this instruction to the start of the constant data.
659 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADD_U32), RegLo)
661 .addTargetIndex(AMDGPU::TI_CONSTDATA_START)
662 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit);
663 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADDC_U32), RegHi)
666 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit)
667 .addReg(AMDGPU::SCC, RegState::Implicit);
668 MI->eraseFromParent();
671 case AMDGPU::SGPR_USE:
672 // This is just a placeholder for register allocation.
673 MI->eraseFromParent();
676 case AMDGPU::V_MOV_B64_PSEUDO: {
677 unsigned Dst = MI->getOperand(0).getReg();
678 unsigned DstLo = RI.getSubReg(Dst, AMDGPU::sub0);
679 unsigned DstHi = RI.getSubReg(Dst, AMDGPU::sub1);
681 const MachineOperand &SrcOp = MI->getOperand(1);
682 // FIXME: Will this work for 64-bit floating point immediates?
683 assert(!SrcOp.isFPImm());
685 APInt Imm(64, SrcOp.getImm());
686 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
687 .addImm(Imm.getLoBits(32).getZExtValue())
688 .addReg(Dst, RegState::Implicit);
689 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
690 .addImm(Imm.getHiBits(32).getZExtValue())
691 .addReg(Dst, RegState::Implicit);
693 assert(SrcOp.isReg());
694 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
695 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub0))
696 .addReg(Dst, RegState::Implicit);
697 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
698 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub1))
699 .addReg(Dst, RegState::Implicit);
701 MI->eraseFromParent();
708 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
711 if (MI->getNumOperands() < 3)
714 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
715 AMDGPU::OpName::src0);
716 assert(Src0Idx != -1 && "Should always have src0 operand");
718 MachineOperand &Src0 = MI->getOperand(Src0Idx);
722 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
723 AMDGPU::OpName::src1);
727 MachineOperand &Src1 = MI->getOperand(Src1Idx);
729 // Make sure it's legal to commute operands for VOP2.
730 if (isVOP2(MI->getOpcode()) &&
731 (!isOperandLegal(MI, Src0Idx, &Src1) ||
732 !isOperandLegal(MI, Src1Idx, &Src0))) {
737 // Allow commuting instructions with Imm operands.
738 if (NewMI || !Src1.isImm() ||
739 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
743 // Be sure to copy the source modifiers to the right place.
744 if (MachineOperand *Src0Mods
745 = getNamedOperand(*MI, AMDGPU::OpName::src0_modifiers)) {
746 MachineOperand *Src1Mods
747 = getNamedOperand(*MI, AMDGPU::OpName::src1_modifiers);
749 int Src0ModsVal = Src0Mods->getImm();
750 if (!Src1Mods && Src0ModsVal != 0)
753 // XXX - This assert might be a lie. It might be useful to have a neg
754 // modifier with 0.0.
755 int Src1ModsVal = Src1Mods->getImm();
756 assert((Src1ModsVal == 0) && "Not expecting modifiers with immediates");
758 Src1Mods->setImm(Src0ModsVal);
759 Src0Mods->setImm(Src1ModsVal);
762 unsigned Reg = Src0.getReg();
763 unsigned SubReg = Src0.getSubReg();
765 Src0.ChangeToImmediate(Src1.getImm());
767 llvm_unreachable("Should only have immediates");
769 Src1.ChangeToRegister(Reg, false);
770 Src1.setSubReg(SubReg);
772 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
776 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
781 // This needs to be implemented because the source modifiers may be inserted
782 // between the true commutable operands, and the base
783 // TargetInstrInfo::commuteInstruction uses it.
784 bool SIInstrInfo::findCommutedOpIndices(MachineInstr *MI,
786 unsigned &SrcOpIdx2) const {
787 const MCInstrDesc &MCID = MI->getDesc();
788 if (!MCID.isCommutable())
791 unsigned Opc = MI->getOpcode();
792 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
796 // FIXME: Workaround TargetInstrInfo::commuteInstruction asserting on
798 if (!MI->getOperand(Src0Idx).isReg())
801 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
805 if (!MI->getOperand(Src1Idx).isReg())
808 // If any source modifiers are set, the generic instruction commuting won't
809 // understand how to copy the source modifiers.
810 if (hasModifiersSet(*MI, AMDGPU::OpName::src0_modifiers) ||
811 hasModifiersSet(*MI, AMDGPU::OpName::src1_modifiers))
819 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
820 MachineBasicBlock::iterator I,
822 unsigned SrcReg) const {
823 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
824 DstReg) .addReg(SrcReg);
827 bool SIInstrInfo::isMov(unsigned Opcode) const {
829 default: return false;
830 case AMDGPU::S_MOV_B32:
831 case AMDGPU::S_MOV_B64:
832 case AMDGPU::V_MOV_B32_e32:
833 case AMDGPU::V_MOV_B32_e64:
839 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
840 return RC != &AMDGPU::EXECRegRegClass;
844 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
845 AliasAnalysis *AA) const {
846 switch(MI->getOpcode()) {
847 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
848 case AMDGPU::S_MOV_B32:
849 case AMDGPU::S_MOV_B64:
850 case AMDGPU::V_MOV_B32_e32:
851 return MI->getOperand(1).isImm();
855 static bool offsetsDoNotOverlap(int WidthA, int OffsetA,
856 int WidthB, int OffsetB) {
857 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB;
858 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA;
859 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
860 return LowOffset + LowWidth <= HighOffset;
863 bool SIInstrInfo::checkInstOffsetsDoNotOverlap(MachineInstr *MIa,
864 MachineInstr *MIb) const {
865 unsigned BaseReg0, Offset0;
866 unsigned BaseReg1, Offset1;
868 if (getLdStBaseRegImmOfs(MIa, BaseReg0, Offset0, &RI) &&
869 getLdStBaseRegImmOfs(MIb, BaseReg1, Offset1, &RI)) {
870 assert(MIa->hasOneMemOperand() && MIb->hasOneMemOperand() &&
871 "read2 / write2 not expected here yet");
872 unsigned Width0 = (*MIa->memoperands_begin())->getSize();
873 unsigned Width1 = (*MIb->memoperands_begin())->getSize();
874 if (BaseReg0 == BaseReg1 &&
875 offsetsDoNotOverlap(Width0, Offset0, Width1, Offset1)) {
883 bool SIInstrInfo::areMemAccessesTriviallyDisjoint(MachineInstr *MIa,
885 AliasAnalysis *AA) const {
886 unsigned Opc0 = MIa->getOpcode();
887 unsigned Opc1 = MIb->getOpcode();
889 assert(MIa && (MIa->mayLoad() || MIa->mayStore()) &&
890 "MIa must load from or modify a memory location");
891 assert(MIb && (MIb->mayLoad() || MIb->mayStore()) &&
892 "MIb must load from or modify a memory location");
894 if (MIa->hasUnmodeledSideEffects() || MIb->hasUnmodeledSideEffects())
897 // XXX - Can we relax this between address spaces?
898 if (MIa->hasOrderedMemoryRef() || MIb->hasOrderedMemoryRef())
901 // TODO: Should we check the address space from the MachineMemOperand? That
902 // would allow us to distinguish objects we know don't alias based on the
903 // underlying addres space, even if it was lowered to a different one,
904 // e.g. private accesses lowered to use MUBUF instructions on a scratch
908 return checkInstOffsetsDoNotOverlap(MIa, MIb);
910 return !isFLAT(Opc1);
913 if (isMUBUF(Opc0) || isMTBUF(Opc0)) {
914 if (isMUBUF(Opc1) || isMTBUF(Opc1))
915 return checkInstOffsetsDoNotOverlap(MIa, MIb);
917 return !isFLAT(Opc1) && !isSMRD(Opc1);
922 return checkInstOffsetsDoNotOverlap(MIa, MIb);
924 return !isFLAT(Opc1) && !isMUBUF(Opc0) && !isMTBUF(Opc0);
929 return checkInstOffsetsDoNotOverlap(MIa, MIb);
937 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
938 int64_t SVal = Imm.getSExtValue();
939 if (SVal >= -16 && SVal <= 64)
942 if (Imm.getBitWidth() == 64) {
943 uint64_t Val = Imm.getZExtValue();
944 return (DoubleToBits(0.0) == Val) ||
945 (DoubleToBits(1.0) == Val) ||
946 (DoubleToBits(-1.0) == Val) ||
947 (DoubleToBits(0.5) == Val) ||
948 (DoubleToBits(-0.5) == Val) ||
949 (DoubleToBits(2.0) == Val) ||
950 (DoubleToBits(-2.0) == Val) ||
951 (DoubleToBits(4.0) == Val) ||
952 (DoubleToBits(-4.0) == Val);
955 // The actual type of the operand does not seem to matter as long
956 // as the bits match one of the inline immediate values. For example:
958 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
959 // so it is a legal inline immediate.
961 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
962 // floating-point, so it is a legal inline immediate.
963 uint32_t Val = Imm.getZExtValue();
965 return (FloatToBits(0.0f) == Val) ||
966 (FloatToBits(1.0f) == Val) ||
967 (FloatToBits(-1.0f) == Val) ||
968 (FloatToBits(0.5f) == Val) ||
969 (FloatToBits(-0.5f) == Val) ||
970 (FloatToBits(2.0f) == Val) ||
971 (FloatToBits(-2.0f) == Val) ||
972 (FloatToBits(4.0f) == Val) ||
973 (FloatToBits(-4.0f) == Val);
976 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO) const {
978 return isInlineConstant(APInt(32, MO.getImm(), true));
983 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO) const {
984 return MO.isImm() && !isInlineConstant(MO);
987 static bool compareMachineOp(const MachineOperand &Op0,
988 const MachineOperand &Op1) {
989 if (Op0.getType() != Op1.getType())
992 switch (Op0.getType()) {
993 case MachineOperand::MO_Register:
994 return Op0.getReg() == Op1.getReg();
995 case MachineOperand::MO_Immediate:
996 return Op0.getImm() == Op1.getImm();
998 llvm_unreachable("Didn't expect to be comparing these operand types");
1002 bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
1003 const MachineOperand &MO) const {
1004 const MCOperandInfo &OpInfo = get(MI->getOpcode()).OpInfo[OpNo];
1006 assert(MO.isImm() || MO.isTargetIndex() || MO.isFI());
1008 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
1011 if (OpInfo.RegClass < 0)
1014 if (isLiteralConstant(MO))
1015 return RI.opCanUseLiteralConstant(OpInfo.OperandType);
1017 return RI.opCanUseInlineConstant(OpInfo.OperandType);
1020 bool SIInstrInfo::canFoldOffset(unsigned OffsetSize, unsigned AS) const {
1022 case AMDGPUAS::GLOBAL_ADDRESS: {
1023 // MUBUF instructions a 12-bit offset in bytes.
1024 return isUInt<12>(OffsetSize);
1026 case AMDGPUAS::CONSTANT_ADDRESS: {
1027 // SMRD instructions have an 8-bit offset in dwords on SI and
1028 // a 20-bit offset in bytes on VI.
1029 if (RI.ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1030 return isUInt<20>(OffsetSize);
1032 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
1034 case AMDGPUAS::LOCAL_ADDRESS:
1035 case AMDGPUAS::REGION_ADDRESS: {
1036 // The single offset versions have a 16-bit offset in bytes.
1037 return isUInt<16>(OffsetSize);
1039 case AMDGPUAS::PRIVATE_ADDRESS:
1040 // Indirect register addressing does not use any offsets.
1046 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
1047 int Op32 = AMDGPU::getVOPe32(Opcode);
1051 return pseudoToMCOpcode(Op32) != -1;
1054 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
1055 // The src0_modifier operand is present on all instructions
1056 // that have modifiers.
1058 return AMDGPU::getNamedOperandIdx(Opcode,
1059 AMDGPU::OpName::src0_modifiers) != -1;
1062 bool SIInstrInfo::hasModifiersSet(const MachineInstr &MI,
1063 unsigned OpName) const {
1064 const MachineOperand *Mods = getNamedOperand(MI, OpName);
1065 return Mods && Mods->getImm();
1068 bool SIInstrInfo::usesConstantBus(const MachineRegisterInfo &MRI,
1069 const MachineOperand &MO) const {
1070 // Literal constants use the constant bus.
1071 if (isLiteralConstant(MO))
1074 if (!MO.isReg() || !MO.isUse())
1077 if (TargetRegisterInfo::isVirtualRegister(MO.getReg()))
1078 return RI.isSGPRClass(MRI.getRegClass(MO.getReg()));
1080 // FLAT_SCR is just an SGPR pair.
1081 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
1084 // EXEC register uses the constant bus.
1085 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
1088 // SGPRs use the constant bus
1089 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
1090 (!MO.isImplicit() &&
1091 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
1092 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
1099 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
1100 StringRef &ErrInfo) const {
1101 uint16_t Opcode = MI->getOpcode();
1102 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1103 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
1104 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
1105 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
1107 // Make sure the number of operands is correct.
1108 const MCInstrDesc &Desc = get(Opcode);
1109 if (!Desc.isVariadic() &&
1110 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
1111 ErrInfo = "Instruction has wrong number of operands.";
1115 // Make sure the register classes are correct
1116 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
1117 if (MI->getOperand(i).isFPImm()) {
1118 ErrInfo = "FPImm Machine Operands are not supported. ISel should bitcast "
1119 "all fp values to integers.";
1123 switch (Desc.OpInfo[i].OperandType) {
1124 case MCOI::OPERAND_REGISTER:
1125 if (MI->getOperand(i).isImm() || MI->getOperand(i).isFPImm()) {
1126 ErrInfo = "Illegal immediate value for operand.";
1130 case AMDGPU::OPERAND_REG_IMM32:
1132 case AMDGPU::OPERAND_REG_INLINE_C:
1133 if (MI->getOperand(i).isImm() && !isInlineConstant(MI->getOperand(i))) {
1134 ErrInfo = "Illegal immediate value for operand.";
1138 case MCOI::OPERAND_IMMEDIATE:
1139 // Check if this operand is an immediate.
1140 // FrameIndex operands will be replaced by immediates, so they are
1142 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFI()) {
1143 ErrInfo = "Expected immediate, but got non-immediate";
1151 if (!MI->getOperand(i).isReg())
1154 int RegClass = Desc.OpInfo[i].RegClass;
1155 if (RegClass != -1) {
1156 unsigned Reg = MI->getOperand(i).getReg();
1157 if (TargetRegisterInfo::isVirtualRegister(Reg))
1160 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
1161 if (!RC->contains(Reg)) {
1162 ErrInfo = "Operand has incorrect register class.";
1170 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
1171 // Only look at the true operands. Only a real operand can use the constant
1172 // bus, and we don't want to check pseudo-operands like the source modifier
1174 const int OpIndices[] = { Src0Idx, Src1Idx, Src2Idx };
1176 unsigned ConstantBusCount = 0;
1177 unsigned SGPRUsed = AMDGPU::NoRegister;
1178 for (int OpIdx : OpIndices) {
1182 const MachineOperand &MO = MI->getOperand(OpIdx);
1183 if (usesConstantBus(MRI, MO)) {
1185 if (MO.getReg() != SGPRUsed)
1187 SGPRUsed = MO.getReg();
1193 if (ConstantBusCount > 1) {
1194 ErrInfo = "VOP* instruction uses the constant bus more than once";
1199 // Verify SRC1 for VOP2 and VOPC
1200 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
1201 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1203 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
1209 if (isVOP3(Opcode)) {
1210 if (Src0Idx != -1 && isLiteralConstant(MI->getOperand(Src0Idx))) {
1211 ErrInfo = "VOP3 src0 cannot be a literal constant.";
1214 if (Src1Idx != -1 && isLiteralConstant(MI->getOperand(Src1Idx))) {
1215 ErrInfo = "VOP3 src1 cannot be a literal constant.";
1218 if (Src2Idx != -1 && isLiteralConstant(MI->getOperand(Src2Idx))) {
1219 ErrInfo = "VOP3 src2 cannot be a literal constant.";
1224 // Verify misc. restrictions on specific instructions.
1225 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
1226 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
1227 const MachineOperand &Src0 = MI->getOperand(Src0Idx);
1228 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1229 const MachineOperand &Src2 = MI->getOperand(Src2Idx);
1230 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
1231 if (!compareMachineOp(Src0, Src1) &&
1232 !compareMachineOp(Src0, Src2)) {
1233 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
1242 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
1243 switch (MI.getOpcode()) {
1244 default: return AMDGPU::INSTRUCTION_LIST_END;
1245 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
1246 case AMDGPU::COPY: return AMDGPU::COPY;
1247 case AMDGPU::PHI: return AMDGPU::PHI;
1248 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
1249 case AMDGPU::S_MOV_B32:
1250 return MI.getOperand(1).isReg() ?
1251 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
1252 case AMDGPU::S_ADD_I32:
1253 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
1254 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
1255 case AMDGPU::S_SUB_I32:
1256 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
1257 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
1258 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
1259 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
1260 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
1261 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
1262 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
1263 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
1264 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
1265 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
1266 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
1267 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
1268 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
1269 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
1270 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
1271 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
1272 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
1273 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
1274 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
1275 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
1276 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
1277 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
1278 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
1279 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
1280 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
1281 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
1282 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
1283 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
1284 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
1285 case AMDGPU::S_LOAD_DWORD_IMM:
1286 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1287 case AMDGPU::S_LOAD_DWORDX2_IMM:
1288 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1289 case AMDGPU::S_LOAD_DWORDX4_IMM:
1290 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1291 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e64;
1292 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
1293 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
1297 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
1298 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
1301 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
1302 unsigned OpNo) const {
1303 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1304 const MCInstrDesc &Desc = get(MI.getOpcode());
1305 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
1306 Desc.OpInfo[OpNo].RegClass == -1) {
1307 unsigned Reg = MI.getOperand(OpNo).getReg();
1309 if (TargetRegisterInfo::isVirtualRegister(Reg))
1310 return MRI.getRegClass(Reg);
1311 return RI.getRegClass(Reg);
1314 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
1315 return RI.getRegClass(RCID);
1318 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
1319 switch (MI.getOpcode()) {
1321 case AMDGPU::REG_SEQUENCE:
1323 case AMDGPU::INSERT_SUBREG:
1324 return RI.hasVGPRs(getOpRegClass(MI, 0));
1326 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
1330 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
1331 MachineBasicBlock::iterator I = MI;
1332 MachineBasicBlock *MBB = MI->getParent();
1333 MachineOperand &MO = MI->getOperand(OpIdx);
1334 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1335 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
1336 const TargetRegisterClass *RC = RI.getRegClass(RCID);
1337 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1339 Opcode = AMDGPU::COPY;
1340 else if (RI.isSGPRClass(RC))
1341 Opcode = AMDGPU::S_MOV_B32;
1344 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
1345 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC))
1346 VRC = &AMDGPU::VReg_64RegClass;
1348 VRC = &AMDGPU::VGPR_32RegClass;
1350 unsigned Reg = MRI.createVirtualRegister(VRC);
1351 DebugLoc DL = MBB->findDebugLoc(I);
1352 BuildMI(*MI->getParent(), I, DL, get(Opcode), Reg)
1354 MO.ChangeToRegister(Reg, false);
1357 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
1358 MachineRegisterInfo &MRI,
1359 MachineOperand &SuperReg,
1360 const TargetRegisterClass *SuperRC,
1362 const TargetRegisterClass *SubRC)
1364 assert(SuperReg.isReg());
1366 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
1367 unsigned SubReg = MRI.createVirtualRegister(SubRC);
1369 // Just in case the super register is itself a sub-register, copy it to a new
1370 // value so we don't need to worry about merging its subreg index with the
1371 // SubIdx passed to this function. The register coalescer should be able to
1372 // eliminate this extra copy.
1373 MachineBasicBlock *MBB = MI->getParent();
1374 DebugLoc DL = MI->getDebugLoc();
1376 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), NewSuperReg)
1377 .addReg(SuperReg.getReg(), 0, SuperReg.getSubReg());
1379 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
1380 .addReg(NewSuperReg, 0, SubIdx);
1385 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
1386 MachineBasicBlock::iterator MII,
1387 MachineRegisterInfo &MRI,
1389 const TargetRegisterClass *SuperRC,
1391 const TargetRegisterClass *SubRC) const {
1393 // XXX - Is there a better way to do this?
1394 if (SubIdx == AMDGPU::sub0)
1395 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
1396 if (SubIdx == AMDGPU::sub1)
1397 return MachineOperand::CreateImm(Op.getImm() >> 32);
1399 llvm_unreachable("Unhandled register index for immediate");
1402 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
1404 return MachineOperand::CreateReg(SubReg, false);
1407 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
1408 MachineBasicBlock::iterator MI,
1409 MachineRegisterInfo &MRI,
1410 const TargetRegisterClass *RC,
1411 const MachineOperand &Op) const {
1412 MachineBasicBlock *MBB = MI->getParent();
1413 DebugLoc DL = MI->getDebugLoc();
1414 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1415 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1416 unsigned Dst = MRI.createVirtualRegister(RC);
1418 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1420 .addImm(Op.getImm() & 0xFFFFFFFF);
1421 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1423 .addImm(Op.getImm() >> 32);
1425 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
1427 .addImm(AMDGPU::sub0)
1429 .addImm(AMDGPU::sub1);
1431 Worklist.push_back(Lo);
1432 Worklist.push_back(Hi);
1437 // Change the order of operands from (0, 1, 2) to (0, 2, 1)
1438 void SIInstrInfo::swapOperands(MachineBasicBlock::iterator Inst) const {
1439 assert(Inst->getNumExplicitOperands() == 3);
1440 MachineOperand Op1 = Inst->getOperand(1);
1441 Inst->RemoveOperand(1);
1442 Inst->addOperand(Op1);
1445 bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
1446 const MachineOperand *MO) const {
1447 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1448 const MCInstrDesc &InstDesc = get(MI->getOpcode());
1449 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
1450 const TargetRegisterClass *DefinedRC =
1451 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
1453 MO = &MI->getOperand(OpIdx);
1455 if (isVALU(InstDesc.Opcode) && usesConstantBus(MRI, *MO)) {
1457 MO->isReg() ? MO->getReg() : (unsigned)AMDGPU::NoRegister;
1458 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1461 if (usesConstantBus(MRI, MI->getOperand(i)) &&
1462 MI->getOperand(i).isReg() && MI->getOperand(i).getReg() != SGPRUsed) {
1470 const TargetRegisterClass *RC = MRI.getRegClass(MO->getReg());
1472 // In order to be legal, the common sub-class must be equal to the
1473 // class of the current operand. For example:
1475 // v_mov_b32 s0 ; Operand defined as vsrc_32
1476 // ; RI.getCommonSubClass(s0,vsrc_32) = sgpr ; LEGAL
1478 // s_sendmsg 0, s0 ; Operand defined as m0reg
1479 // ; RI.getCommonSubClass(s0,m0reg) = m0reg ; NOT LEGAL
1481 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass)) == RC;
1485 // Handle non-register types that are treated like immediates.
1486 assert(MO->isImm() || MO->isTargetIndex() || MO->isFI());
1489 // This operand expects an immediate.
1493 return isImmOperandLegal(MI, OpIdx, *MO);
1496 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
1497 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1499 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1500 AMDGPU::OpName::src0);
1501 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1502 AMDGPU::OpName::src1);
1503 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1504 AMDGPU::OpName::src2);
1507 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
1509 if (!isOperandLegal(MI, Src0Idx))
1510 legalizeOpWithMove(MI, Src0Idx);
1513 if (isOperandLegal(MI, Src1Idx))
1516 // Usually src0 of VOP2 instructions allow more types of inputs
1517 // than src1, so try to commute the instruction to decrease our
1518 // chances of having to insert a MOV instruction to legalize src1.
1519 if (MI->isCommutable()) {
1520 if (commuteInstruction(MI))
1521 // If we are successful in commuting, then we know MI is legal, so
1526 legalizeOpWithMove(MI, Src1Idx);
1530 // XXX - Do any VOP3 instructions read VCC?
1532 if (isVOP3(MI->getOpcode())) {
1533 int VOP3Idx[3] = { Src0Idx, Src1Idx, Src2Idx };
1535 // Find the one SGPR operand we are allowed to use.
1536 unsigned SGPRReg = findUsedSGPR(MI, VOP3Idx);
1538 for (unsigned i = 0; i < 3; ++i) {
1539 int Idx = VOP3Idx[i];
1542 MachineOperand &MO = MI->getOperand(Idx);
1545 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
1546 continue; // VGPRs are legal
1548 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
1550 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
1551 SGPRReg = MO.getReg();
1552 // We can use one SGPR in each VOP3 instruction.
1555 } else if (!isLiteralConstant(MO)) {
1556 // If it is not a register and not a literal constant, then it must be
1557 // an inline constant which is always legal.
1560 // If we make it this far, then the operand is not legal and we must
1562 legalizeOpWithMove(MI, Idx);
1566 // Legalize REG_SEQUENCE and PHI
1567 // The register class of the operands much be the same type as the register
1568 // class of the output.
1569 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
1570 MI->getOpcode() == AMDGPU::PHI) {
1571 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
1572 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1573 if (!MI->getOperand(i).isReg() ||
1574 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1576 const TargetRegisterClass *OpRC =
1577 MRI.getRegClass(MI->getOperand(i).getReg());
1578 if (RI.hasVGPRs(OpRC)) {
1585 // If any of the operands are VGPR registers, then they all most be
1586 // otherwise we will create illegal VGPR->SGPR copies when legalizing
1588 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
1591 VRC = RI.getEquivalentVGPRClass(SRC);
1598 // Update all the operands so they have the same type.
1599 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1600 if (!MI->getOperand(i).isReg() ||
1601 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1603 unsigned DstReg = MRI.createVirtualRegister(RC);
1604 MachineBasicBlock *InsertBB;
1605 MachineBasicBlock::iterator Insert;
1606 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
1607 InsertBB = MI->getParent();
1610 // MI is a PHI instruction.
1611 InsertBB = MI->getOperand(i + 1).getMBB();
1612 Insert = InsertBB->getFirstTerminator();
1614 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
1615 get(AMDGPU::COPY), DstReg)
1616 .addOperand(MI->getOperand(i));
1617 MI->getOperand(i).setReg(DstReg);
1621 // Legalize INSERT_SUBREG
1622 // src0 must have the same register class as dst
1623 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
1624 unsigned Dst = MI->getOperand(0).getReg();
1625 unsigned Src0 = MI->getOperand(1).getReg();
1626 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
1627 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
1628 if (DstRC != Src0RC) {
1629 MachineBasicBlock &MBB = *MI->getParent();
1630 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
1631 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
1633 MI->getOperand(1).setReg(NewSrc0);
1638 // Legalize MUBUF* instructions
1639 // FIXME: If we start using the non-addr64 instructions for compute, we
1640 // may need to legalize them here.
1642 AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::srsrc);
1643 if (SRsrcIdx != -1) {
1644 // We have an MUBUF instruction
1645 MachineOperand *SRsrc = &MI->getOperand(SRsrcIdx);
1646 unsigned SRsrcRC = get(MI->getOpcode()).OpInfo[SRsrcIdx].RegClass;
1647 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
1648 RI.getRegClass(SRsrcRC))) {
1649 // The operands are legal.
1650 // FIXME: We may need to legalize operands besided srsrc.
1654 MachineBasicBlock &MBB = *MI->getParent();
1655 // Extract the the ptr from the resource descriptor.
1657 // SRsrcPtrLo = srsrc:sub0
1658 unsigned SRsrcPtrLo = buildExtractSubReg(MI, MRI, *SRsrc,
1659 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VGPR_32RegClass);
1661 // SRsrcPtrHi = srsrc:sub1
1662 unsigned SRsrcPtrHi = buildExtractSubReg(MI, MRI, *SRsrc,
1663 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VGPR_32RegClass);
1665 // Create an empty resource descriptor
1666 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1667 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1668 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1669 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1670 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
1673 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1677 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1678 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1680 .addImm(RsrcDataFormat & 0xFFFFFFFF);
1682 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1683 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1685 .addImm(RsrcDataFormat >> 32);
1687 // NewSRsrc = {Zero64, SRsrcFormat}
1688 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1691 .addImm(AMDGPU::sub0_sub1)
1692 .addReg(SRsrcFormatLo)
1693 .addImm(AMDGPU::sub2)
1694 .addReg(SRsrcFormatHi)
1695 .addImm(AMDGPU::sub3);
1697 MachineOperand *VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1698 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
1699 unsigned NewVAddrLo;
1700 unsigned NewVAddrHi;
1702 // This is already an ADDR64 instruction so we need to add the pointer
1703 // extracted from the resource descriptor to the current value of VAddr.
1704 NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1705 NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1707 // NewVaddrLo = SRsrcPtrLo + VAddr:sub0
1708 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1711 .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
1712 .addReg(AMDGPU::VCC, RegState::ImplicitDefine);
1714 // NewVaddrHi = SRsrcPtrHi + VAddr:sub1
1715 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1718 .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
1719 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1720 .addReg(AMDGPU::VCC, RegState::Implicit);
1723 // This instructions is the _OFFSET variant, so we need to convert it to
1725 MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
1726 MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
1727 MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
1728 assert(SOffset->isImm() && SOffset->getImm() == 0 && "Legalizing MUBUF "
1729 "with non-zero soffset is not implemented");
1732 // Create the new instruction.
1733 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
1734 MachineInstr *Addr64 =
1735 BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
1738 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
1739 // This will be replaced later
1740 // with the new value of vaddr.
1741 .addOperand(*Offset);
1743 MI->removeFromParent();
1746 NewVAddrLo = SRsrcPtrLo;
1747 NewVAddrHi = SRsrcPtrHi;
1748 VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1749 SRsrc = getNamedOperand(*MI, AMDGPU::OpName::srsrc);
1752 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1753 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1756 .addImm(AMDGPU::sub0)
1758 .addImm(AMDGPU::sub1);
1761 // Update the instruction to use NewVaddr
1762 VAddr->setReg(NewVAddr);
1763 // Update the instruction to use NewSRsrc
1764 SRsrc->setReg(NewSRsrc);
1768 void SIInstrInfo::splitSMRD(MachineInstr *MI,
1769 const TargetRegisterClass *HalfRC,
1770 unsigned HalfImmOp, unsigned HalfSGPROp,
1771 MachineInstr *&Lo, MachineInstr *&Hi) const {
1773 DebugLoc DL = MI->getDebugLoc();
1774 MachineBasicBlock *MBB = MI->getParent();
1775 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1776 unsigned RegLo = MRI.createVirtualRegister(HalfRC);
1777 unsigned RegHi = MRI.createVirtualRegister(HalfRC);
1778 unsigned HalfSize = HalfRC->getSize();
1779 const MachineOperand *OffOp =
1780 getNamedOperand(*MI, AMDGPU::OpName::offset);
1781 const MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
1783 // The SMRD has an 8-bit offset in dwords on SI and a 20-bit offset in bytes
1786 bool isVI = RI.ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS;
1787 unsigned OffScale = isVI ? 1 : 4;
1788 // Handle the _IMM variant
1789 unsigned LoOffset = OffOp->getImm() * OffScale;
1790 unsigned HiOffset = LoOffset + HalfSize;
1791 Lo = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegLo)
1793 .addImm(LoOffset / OffScale);
1795 if (!isUInt<20>(HiOffset) || (!isVI && !isUInt<8>(HiOffset / OffScale))) {
1796 unsigned OffsetSGPR =
1797 MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1798 BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32), OffsetSGPR)
1799 .addImm(HiOffset); // The offset in register is in bytes.
1800 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
1802 .addReg(OffsetSGPR);
1804 Hi = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegHi)
1806 .addImm(HiOffset / OffScale);
1809 // Handle the _SGPR variant
1810 MachineOperand *SOff = getNamedOperand(*MI, AMDGPU::OpName::soff);
1811 Lo = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegLo)
1814 unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1815 BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
1818 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp))
1820 .addReg(OffsetSGPR);
1823 unsigned SubLo, SubHi;
1826 SubLo = AMDGPU::sub0;
1827 SubHi = AMDGPU::sub1;
1830 SubLo = AMDGPU::sub0_sub1;
1831 SubHi = AMDGPU::sub2_sub3;
1834 SubLo = AMDGPU::sub0_sub1_sub2_sub3;
1835 SubHi = AMDGPU::sub4_sub5_sub6_sub7;
1838 SubLo = AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
1839 SubHi = AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15;
1842 llvm_unreachable("Unhandled HalfSize");
1845 BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE))
1846 .addOperand(MI->getOperand(0))
1853 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
1854 MachineBasicBlock *MBB = MI->getParent();
1855 switch (MI->getOpcode()) {
1856 case AMDGPU::S_LOAD_DWORD_IMM:
1857 case AMDGPU::S_LOAD_DWORD_SGPR:
1858 case AMDGPU::S_LOAD_DWORDX2_IMM:
1859 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1860 case AMDGPU::S_LOAD_DWORDX4_IMM:
1861 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1862 unsigned NewOpcode = getVALUOp(*MI);
1866 if (MI->getOperand(2).isReg()) {
1867 RegOffset = MI->getOperand(2).getReg();
1870 assert(MI->getOperand(2).isImm());
1871 // SMRD instructions take a dword offsets on SI and byte offset on VI
1872 // and MUBUF instructions always take a byte offset.
1873 ImmOffset = MI->getOperand(2).getImm();
1874 if (RI.ST.getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS)
1876 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1878 if (isUInt<12>(ImmOffset)) {
1879 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1883 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1890 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1891 unsigned DWord0 = RegOffset;
1892 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1893 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1894 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1895 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
1897 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
1899 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
1900 .addImm(RsrcDataFormat & 0xFFFFFFFF);
1901 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
1902 .addImm(RsrcDataFormat >> 32);
1903 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
1905 .addImm(AMDGPU::sub0)
1907 .addImm(AMDGPU::sub1)
1909 .addImm(AMDGPU::sub2)
1911 .addImm(AMDGPU::sub3);
1912 MI->setDesc(get(NewOpcode));
1913 if (MI->getOperand(2).isReg()) {
1914 MI->getOperand(2).setReg(MI->getOperand(1).getReg());
1916 MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
1918 MI->getOperand(1).setReg(SRsrc);
1919 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
1921 const TargetRegisterClass *NewDstRC =
1922 RI.getRegClass(get(NewOpcode).OpInfo[0].RegClass);
1924 unsigned DstReg = MI->getOperand(0).getReg();
1925 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1926 MRI.replaceRegWith(DstReg, NewDstReg);
1929 case AMDGPU::S_LOAD_DWORDX8_IMM:
1930 case AMDGPU::S_LOAD_DWORDX8_SGPR: {
1931 MachineInstr *Lo, *Hi;
1932 splitSMRD(MI, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM,
1933 AMDGPU::S_LOAD_DWORDX4_SGPR, Lo, Hi);
1934 MI->eraseFromParent();
1935 moveSMRDToVALU(Lo, MRI);
1936 moveSMRDToVALU(Hi, MRI);
1940 case AMDGPU::S_LOAD_DWORDX16_IMM:
1941 case AMDGPU::S_LOAD_DWORDX16_SGPR: {
1942 MachineInstr *Lo, *Hi;
1943 splitSMRD(MI, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM,
1944 AMDGPU::S_LOAD_DWORDX8_SGPR, Lo, Hi);
1945 MI->eraseFromParent();
1946 moveSMRDToVALU(Lo, MRI);
1947 moveSMRDToVALU(Hi, MRI);
1953 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
1954 SmallVector<MachineInstr *, 128> Worklist;
1955 Worklist.push_back(&TopInst);
1957 while (!Worklist.empty()) {
1958 MachineInstr *Inst = Worklist.pop_back_val();
1959 MachineBasicBlock *MBB = Inst->getParent();
1960 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1962 unsigned Opcode = Inst->getOpcode();
1963 unsigned NewOpcode = getVALUOp(*Inst);
1965 // Handle some special cases
1968 if (isSMRD(Inst->getOpcode())) {
1969 moveSMRDToVALU(Inst, MRI);
1972 case AMDGPU::S_MOV_B64: {
1973 DebugLoc DL = Inst->getDebugLoc();
1975 // If the source operand is a register we can replace this with a
1977 if (Inst->getOperand(1).isReg()) {
1978 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
1979 .addOperand(Inst->getOperand(0))
1980 .addOperand(Inst->getOperand(1));
1981 Worklist.push_back(Copy);
1983 // Otherwise, we need to split this into two movs, because there is
1984 // no 64-bit VALU move instruction.
1985 unsigned Reg = Inst->getOperand(0).getReg();
1986 unsigned Dst = split64BitImm(Worklist,
1989 MRI.getRegClass(Reg),
1990 Inst->getOperand(1));
1991 MRI.replaceRegWith(Reg, Dst);
1993 Inst->eraseFromParent();
1996 case AMDGPU::S_AND_B64:
1997 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
1998 Inst->eraseFromParent();
2001 case AMDGPU::S_OR_B64:
2002 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
2003 Inst->eraseFromParent();
2006 case AMDGPU::S_XOR_B64:
2007 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
2008 Inst->eraseFromParent();
2011 case AMDGPU::S_NOT_B64:
2012 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
2013 Inst->eraseFromParent();
2016 case AMDGPU::S_BCNT1_I32_B64:
2017 splitScalar64BitBCNT(Worklist, Inst);
2018 Inst->eraseFromParent();
2021 case AMDGPU::S_BFE_I64: {
2022 splitScalar64BitBFE(Worklist, Inst);
2023 Inst->eraseFromParent();
2027 case AMDGPU::S_LSHL_B32:
2028 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2029 NewOpcode = AMDGPU::V_LSHLREV_B32_e64;
2033 case AMDGPU::S_ASHR_I32:
2034 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2035 NewOpcode = AMDGPU::V_ASHRREV_I32_e64;
2039 case AMDGPU::S_LSHR_B32:
2040 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2041 NewOpcode = AMDGPU::V_LSHRREV_B32_e64;
2046 case AMDGPU::S_BFE_U64:
2047 case AMDGPU::S_BFM_B64:
2048 llvm_unreachable("Moving this op to VALU not implemented");
2051 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
2052 // We cannot move this instruction to the VALU, so we should try to
2053 // legalize its operands instead.
2054 legalizeOperands(Inst);
2058 // Use the new VALU Opcode.
2059 const MCInstrDesc &NewDesc = get(NewOpcode);
2060 Inst->setDesc(NewDesc);
2062 // Remove any references to SCC. Vector instructions can't read from it, and
2063 // We're just about to add the implicit use / defs of VCC, and we don't want
2065 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
2066 MachineOperand &Op = Inst->getOperand(i);
2067 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
2068 Inst->RemoveOperand(i);
2071 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
2072 // We are converting these to a BFE, so we need to add the missing
2073 // operands for the size and offset.
2074 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
2075 Inst->addOperand(MachineOperand::CreateImm(0));
2076 Inst->addOperand(MachineOperand::CreateImm(Size));
2078 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
2079 // The VALU version adds the second operand to the result, so insert an
2081 Inst->addOperand(MachineOperand::CreateImm(0));
2084 addDescImplicitUseDef(NewDesc, Inst);
2086 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
2087 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
2088 // If we need to move this to VGPRs, we need to unpack the second operand
2089 // back into the 2 separate ones for bit offset and width.
2090 assert(OffsetWidthOp.isImm() &&
2091 "Scalar BFE is only implemented for constant width and offset");
2092 uint32_t Imm = OffsetWidthOp.getImm();
2094 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
2095 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
2096 Inst->RemoveOperand(2); // Remove old immediate.
2097 Inst->addOperand(MachineOperand::CreateImm(Offset));
2098 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
2101 // Update the destination register class.
2103 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
2106 // For target instructions, getOpRegClass just returns the virtual
2107 // register class associated with the operand, so we need to find an
2108 // equivalent VGPR register class in order to move the instruction to the
2112 case AMDGPU::REG_SEQUENCE:
2113 case AMDGPU::INSERT_SUBREG:
2114 if (RI.hasVGPRs(NewDstRC))
2116 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
2124 unsigned DstReg = Inst->getOperand(0).getReg();
2125 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
2126 MRI.replaceRegWith(DstReg, NewDstReg);
2128 // Legalize the operands
2129 legalizeOperands(Inst);
2131 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
2132 E = MRI.use_end(); I != E; ++I) {
2133 MachineInstr &UseMI = *I->getParent();
2134 if (!canReadVGPR(UseMI, I.getOperandNo())) {
2135 Worklist.push_back(&UseMI);
2141 //===----------------------------------------------------------------------===//
2142 // Indirect addressing callbacks
2143 //===----------------------------------------------------------------------===//
2145 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
2146 unsigned Channel) const {
2147 assert(Channel == 0);
2151 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
2152 return &AMDGPU::VGPR_32RegClass;
2155 void SIInstrInfo::splitScalar64BitUnaryOp(
2156 SmallVectorImpl<MachineInstr *> &Worklist,
2158 unsigned Opcode) const {
2159 MachineBasicBlock &MBB = *Inst->getParent();
2160 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2162 MachineOperand &Dest = Inst->getOperand(0);
2163 MachineOperand &Src0 = Inst->getOperand(1);
2164 DebugLoc DL = Inst->getDebugLoc();
2166 MachineBasicBlock::iterator MII = Inst;
2168 const MCInstrDesc &InstDesc = get(Opcode);
2169 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2170 MRI.getRegClass(Src0.getReg()) :
2171 &AMDGPU::SGPR_32RegClass;
2173 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2175 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2176 AMDGPU::sub0, Src0SubRC);
2178 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2179 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2181 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2182 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2183 .addOperand(SrcReg0Sub0);
2185 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2186 AMDGPU::sub1, Src0SubRC);
2188 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2189 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2190 .addOperand(SrcReg0Sub1);
2192 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2193 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2195 .addImm(AMDGPU::sub0)
2197 .addImm(AMDGPU::sub1);
2199 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2201 // Try to legalize the operands in case we need to swap the order to keep it
2203 Worklist.push_back(LoHalf);
2204 Worklist.push_back(HiHalf);
2207 void SIInstrInfo::splitScalar64BitBinaryOp(
2208 SmallVectorImpl<MachineInstr *> &Worklist,
2210 unsigned Opcode) const {
2211 MachineBasicBlock &MBB = *Inst->getParent();
2212 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2214 MachineOperand &Dest = Inst->getOperand(0);
2215 MachineOperand &Src0 = Inst->getOperand(1);
2216 MachineOperand &Src1 = Inst->getOperand(2);
2217 DebugLoc DL = Inst->getDebugLoc();
2219 MachineBasicBlock::iterator MII = Inst;
2221 const MCInstrDesc &InstDesc = get(Opcode);
2222 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2223 MRI.getRegClass(Src0.getReg()) :
2224 &AMDGPU::SGPR_32RegClass;
2226 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2227 const TargetRegisterClass *Src1RC = Src1.isReg() ?
2228 MRI.getRegClass(Src1.getReg()) :
2229 &AMDGPU::SGPR_32RegClass;
2231 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
2233 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2234 AMDGPU::sub0, Src0SubRC);
2235 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2236 AMDGPU::sub0, Src1SubRC);
2238 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2239 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2241 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2242 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2243 .addOperand(SrcReg0Sub0)
2244 .addOperand(SrcReg1Sub0);
2246 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2247 AMDGPU::sub1, Src0SubRC);
2248 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2249 AMDGPU::sub1, Src1SubRC);
2251 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2252 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2253 .addOperand(SrcReg0Sub1)
2254 .addOperand(SrcReg1Sub1);
2256 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2257 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2259 .addImm(AMDGPU::sub0)
2261 .addImm(AMDGPU::sub1);
2263 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2265 // Try to legalize the operands in case we need to swap the order to keep it
2267 Worklist.push_back(LoHalf);
2268 Worklist.push_back(HiHalf);
2271 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
2272 MachineInstr *Inst) const {
2273 MachineBasicBlock &MBB = *Inst->getParent();
2274 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2276 MachineBasicBlock::iterator MII = Inst;
2277 DebugLoc DL = Inst->getDebugLoc();
2279 MachineOperand &Dest = Inst->getOperand(0);
2280 MachineOperand &Src = Inst->getOperand(1);
2282 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e64);
2283 const TargetRegisterClass *SrcRC = Src.isReg() ?
2284 MRI.getRegClass(Src.getReg()) :
2285 &AMDGPU::SGPR_32RegClass;
2287 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2288 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2290 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
2292 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2293 AMDGPU::sub0, SrcSubRC);
2294 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2295 AMDGPU::sub1, SrcSubRC);
2297 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
2298 .addOperand(SrcRegSub0)
2301 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
2302 .addOperand(SrcRegSub1)
2305 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2307 Worklist.push_back(First);
2308 Worklist.push_back(Second);
2311 void SIInstrInfo::splitScalar64BitBFE(SmallVectorImpl<MachineInstr *> &Worklist,
2312 MachineInstr *Inst) const {
2313 MachineBasicBlock &MBB = *Inst->getParent();
2314 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2315 MachineBasicBlock::iterator MII = Inst;
2316 DebugLoc DL = Inst->getDebugLoc();
2318 MachineOperand &Dest = Inst->getOperand(0);
2319 uint32_t Imm = Inst->getOperand(2).getImm();
2320 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
2321 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
2325 // Only sext_inreg cases handled.
2326 assert(Inst->getOpcode() == AMDGPU::S_BFE_I64 &&
2331 if (BitWidth < 32) {
2332 unsigned MidRegLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2333 unsigned MidRegHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2334 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
2336 BuildMI(MBB, MII, DL, get(AMDGPU::V_BFE_I32), MidRegLo)
2337 .addReg(Inst->getOperand(1).getReg(), 0, AMDGPU::sub0)
2341 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e32), MidRegHi)
2345 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
2347 .addImm(AMDGPU::sub0)
2349 .addImm(AMDGPU::sub1);
2351 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2355 MachineOperand &Src = Inst->getOperand(1);
2356 unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2357 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
2359 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e64), TmpReg)
2361 .addReg(Src.getReg(), 0, AMDGPU::sub0);
2363 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
2364 .addReg(Src.getReg(), 0, AMDGPU::sub0)
2365 .addImm(AMDGPU::sub0)
2367 .addImm(AMDGPU::sub1);
2369 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2372 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
2373 MachineInstr *Inst) const {
2374 // Add the implict and explicit register definitions.
2375 if (NewDesc.ImplicitUses) {
2376 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
2377 unsigned Reg = NewDesc.ImplicitUses[i];
2378 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
2382 if (NewDesc.ImplicitDefs) {
2383 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
2384 unsigned Reg = NewDesc.ImplicitDefs[i];
2385 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
2390 unsigned SIInstrInfo::findUsedSGPR(const MachineInstr *MI,
2391 int OpIndices[3]) const {
2392 const MCInstrDesc &Desc = get(MI->getOpcode());
2394 // Find the one SGPR operand we are allowed to use.
2395 unsigned SGPRReg = AMDGPU::NoRegister;
2397 // First we need to consider the instruction's operand requirements before
2398 // legalizing. Some operands are required to be SGPRs, such as implicit uses
2399 // of VCC, but we are still bound by the constant bus requirement to only use
2402 // If the operand's class is an SGPR, we can never move it.
2404 for (const MachineOperand &MO : MI->implicit_operands()) {
2405 // We only care about reads.
2409 if (MO.getReg() == AMDGPU::VCC)
2412 if (MO.getReg() == AMDGPU::FLAT_SCR)
2413 return AMDGPU::FLAT_SCR;
2416 unsigned UsedSGPRs[3] = { AMDGPU::NoRegister };
2417 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
2419 for (unsigned i = 0; i < 3; ++i) {
2420 int Idx = OpIndices[i];
2424 const MachineOperand &MO = MI->getOperand(Idx);
2425 if (RI.isSGPRClassID(Desc.OpInfo[Idx].RegClass))
2426 SGPRReg = MO.getReg();
2428 if (MO.isReg() && RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
2429 UsedSGPRs[i] = MO.getReg();
2432 if (SGPRReg != AMDGPU::NoRegister)
2435 // We don't have a required SGPR operand, so we have a bit more freedom in
2436 // selecting operands to move.
2438 // Try to select the most used SGPR. If an SGPR is equal to one of the
2439 // others, we choose that.
2442 // V_FMA_F32 v0, s0, s0, s0 -> No moves
2443 // V_FMA_F32 v0, s0, s1, s0 -> Move s1
2445 if (UsedSGPRs[0] != AMDGPU::NoRegister) {
2446 if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
2447 SGPRReg = UsedSGPRs[0];
2450 if (SGPRReg == AMDGPU::NoRegister && UsedSGPRs[1] != AMDGPU::NoRegister) {
2451 if (UsedSGPRs[1] == UsedSGPRs[2])
2452 SGPRReg = UsedSGPRs[1];
2458 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
2459 MachineBasicBlock *MBB,
2460 MachineBasicBlock::iterator I,
2462 unsigned Address, unsigned OffsetReg) const {
2463 const DebugLoc &DL = MBB->findDebugLoc(I);
2464 unsigned IndirectBaseReg = AMDGPU::VGPR_32RegClass.getRegister(
2465 getIndirectIndexBegin(*MBB->getParent()));
2467 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
2468 .addReg(IndirectBaseReg, RegState::Define)
2469 .addOperand(I->getOperand(0))
2470 .addReg(IndirectBaseReg)
2476 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
2477 MachineBasicBlock *MBB,
2478 MachineBasicBlock::iterator I,
2480 unsigned Address, unsigned OffsetReg) const {
2481 const DebugLoc &DL = MBB->findDebugLoc(I);
2482 unsigned IndirectBaseReg = AMDGPU::VGPR_32RegClass.getRegister(
2483 getIndirectIndexBegin(*MBB->getParent()));
2485 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
2486 .addOperand(I->getOperand(0))
2487 .addOperand(I->getOperand(1))
2488 .addReg(IndirectBaseReg)
2494 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
2495 const MachineFunction &MF) const {
2496 int End = getIndirectIndexEnd(MF);
2497 int Begin = getIndirectIndexBegin(MF);
2503 for (int Index = Begin; Index <= End; ++Index)
2504 Reserved.set(AMDGPU::VGPR_32RegClass.getRegister(Index));
2506 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
2507 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
2509 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
2510 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
2512 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
2513 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
2515 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
2516 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
2518 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
2519 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));
2522 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
2523 unsigned OperandName) const {
2524 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
2528 return &MI.getOperand(Idx);
2531 uint64_t SIInstrInfo::getDefaultRsrcDataFormat() const {
2532 uint64_t RsrcDataFormat = AMDGPU::RSRC_DATA_FORMAT;
2533 if (ST.isAmdHsaOS())
2534 RsrcDataFormat |= (1ULL << 56);
2536 return RsrcDataFormat;