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 NewOpc = AMDGPU::getCommuteRev(Opcode);
411 // Check if the commuted (REV) opcode exists on the target.
412 if (NewOpc != -1 && pseudoToMCOpcode(NewOpc) != -1)
415 // Try to map commuted to original opcode
416 NewOpc = AMDGPU::getCommuteOrig(Opcode);
417 // Check if the original (non-REV) opcode exists on the target.
418 if (NewOpc != -1 && pseudoToMCOpcode(NewOpc) != -1)
424 unsigned SIInstrInfo::getMovOpcode(const TargetRegisterClass *DstRC) const {
426 if (DstRC->getSize() == 4) {
427 return RI.isSGPRClass(DstRC) ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32;
428 } else if (DstRC->getSize() == 8 && RI.isSGPRClass(DstRC)) {
429 return AMDGPU::S_MOV_B64;
430 } else if (DstRC->getSize() == 8 && !RI.isSGPRClass(DstRC)) {
431 return AMDGPU::V_MOV_B64_PSEUDO;
436 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
437 MachineBasicBlock::iterator MI,
438 unsigned SrcReg, bool isKill,
440 const TargetRegisterClass *RC,
441 const TargetRegisterInfo *TRI) const {
442 MachineFunction *MF = MBB.getParent();
443 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
444 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
445 DebugLoc DL = MBB.findDebugLoc(MI);
448 if (RI.isSGPRClass(RC)) {
449 // We are only allowed to create one new instruction when spilling
450 // registers, so we need to use pseudo instruction for spilling
452 switch (RC->getSize() * 8) {
453 case 32: Opcode = AMDGPU::SI_SPILL_S32_SAVE; break;
454 case 64: Opcode = AMDGPU::SI_SPILL_S64_SAVE; break;
455 case 128: Opcode = AMDGPU::SI_SPILL_S128_SAVE; break;
456 case 256: Opcode = AMDGPU::SI_SPILL_S256_SAVE; break;
457 case 512: Opcode = AMDGPU::SI_SPILL_S512_SAVE; break;
459 } else if(RI.hasVGPRs(RC) && ST.isVGPRSpillingEnabled(MFI)) {
460 MFI->setHasSpilledVGPRs();
462 switch(RC->getSize() * 8) {
463 case 32: Opcode = AMDGPU::SI_SPILL_V32_SAVE; break;
464 case 64: Opcode = AMDGPU::SI_SPILL_V64_SAVE; break;
465 case 96: Opcode = AMDGPU::SI_SPILL_V96_SAVE; break;
466 case 128: Opcode = AMDGPU::SI_SPILL_V128_SAVE; break;
467 case 256: Opcode = AMDGPU::SI_SPILL_V256_SAVE; break;
468 case 512: Opcode = AMDGPU::SI_SPILL_V512_SAVE; break;
473 FrameInfo->setObjectAlignment(FrameIndex, 4);
474 BuildMI(MBB, MI, DL, get(Opcode))
476 .addFrameIndex(FrameIndex)
477 // Place-holder registers, these will be filled in by
478 // SIPrepareScratchRegs.
479 .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
480 .addReg(AMDGPU::SGPR0, RegState::Undef);
482 LLVMContext &Ctx = MF->getFunction()->getContext();
483 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Do not know how to"
485 BuildMI(MBB, MI, DL, get(AMDGPU::KILL))
490 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
491 MachineBasicBlock::iterator MI,
492 unsigned DestReg, int FrameIndex,
493 const TargetRegisterClass *RC,
494 const TargetRegisterInfo *TRI) const {
495 MachineFunction *MF = MBB.getParent();
496 const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
497 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
498 DebugLoc DL = MBB.findDebugLoc(MI);
501 if (RI.isSGPRClass(RC)){
502 switch(RC->getSize() * 8) {
503 case 32: Opcode = AMDGPU::SI_SPILL_S32_RESTORE; break;
504 case 64: Opcode = AMDGPU::SI_SPILL_S64_RESTORE; break;
505 case 128: Opcode = AMDGPU::SI_SPILL_S128_RESTORE; break;
506 case 256: Opcode = AMDGPU::SI_SPILL_S256_RESTORE; break;
507 case 512: Opcode = AMDGPU::SI_SPILL_S512_RESTORE; break;
509 } else if(RI.hasVGPRs(RC) && ST.isVGPRSpillingEnabled(MFI)) {
510 switch(RC->getSize() * 8) {
511 case 32: Opcode = AMDGPU::SI_SPILL_V32_RESTORE; break;
512 case 64: Opcode = AMDGPU::SI_SPILL_V64_RESTORE; break;
513 case 96: Opcode = AMDGPU::SI_SPILL_V96_RESTORE; break;
514 case 128: Opcode = AMDGPU::SI_SPILL_V128_RESTORE; break;
515 case 256: Opcode = AMDGPU::SI_SPILL_V256_RESTORE; break;
516 case 512: Opcode = AMDGPU::SI_SPILL_V512_RESTORE; break;
521 FrameInfo->setObjectAlignment(FrameIndex, 4);
522 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
523 .addFrameIndex(FrameIndex)
524 // Place-holder registers, these will be filled in by
525 // SIPrepareScratchRegs.
526 .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
527 .addReg(AMDGPU::SGPR0, RegState::Undef);
530 LLVMContext &Ctx = MF->getFunction()->getContext();
531 Ctx.emitError("SIInstrInfo::loadRegFromStackSlot - Do not know how to"
532 " restore register");
533 BuildMI(MBB, MI, DL, get(AMDGPU::IMPLICIT_DEF), DestReg);
537 /// \param @Offset Offset in bytes of the FrameIndex being spilled
538 unsigned SIInstrInfo::calculateLDSSpillAddress(MachineBasicBlock &MBB,
539 MachineBasicBlock::iterator MI,
540 RegScavenger *RS, unsigned TmpReg,
541 unsigned FrameOffset,
542 unsigned Size) const {
543 MachineFunction *MF = MBB.getParent();
544 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
545 const AMDGPUSubtarget &ST = MF->getSubtarget<AMDGPUSubtarget>();
546 const SIRegisterInfo *TRI =
547 static_cast<const SIRegisterInfo*>(ST.getRegisterInfo());
548 DebugLoc DL = MBB.findDebugLoc(MI);
549 unsigned WorkGroupSize = MFI->getMaximumWorkGroupSize(*MF);
550 unsigned WavefrontSize = ST.getWavefrontSize();
552 unsigned TIDReg = MFI->getTIDReg();
553 if (!MFI->hasCalculatedTID()) {
554 MachineBasicBlock &Entry = MBB.getParent()->front();
555 MachineBasicBlock::iterator Insert = Entry.front();
556 DebugLoc DL = Insert->getDebugLoc();
558 TIDReg = RI.findUnusedRegister(MF->getRegInfo(), &AMDGPU::VGPR_32RegClass);
559 if (TIDReg == AMDGPU::NoRegister)
563 if (MFI->getShaderType() == ShaderType::COMPUTE &&
564 WorkGroupSize > WavefrontSize) {
566 unsigned TIDIGXReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_X);
567 unsigned TIDIGYReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Y);
568 unsigned TIDIGZReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Z);
569 unsigned InputPtrReg =
570 TRI->getPreloadedValue(*MF, SIRegisterInfo::INPUT_PTR);
571 static const unsigned TIDIGRegs[3] = {
572 TIDIGXReg, TIDIGYReg, TIDIGZReg
574 for (unsigned Reg : TIDIGRegs) {
575 if (!Entry.isLiveIn(Reg))
576 Entry.addLiveIn(Reg);
579 RS->enterBasicBlock(&Entry);
580 unsigned STmp0 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
581 unsigned STmp1 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
582 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp0)
584 .addImm(SI::KernelInputOffsets::NGROUPS_Z);
585 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp1)
587 .addImm(SI::KernelInputOffsets::NGROUPS_Y);
589 // NGROUPS.X * NGROUPS.Y
590 BuildMI(Entry, Insert, DL, get(AMDGPU::S_MUL_I32), STmp1)
593 // (NGROUPS.X * NGROUPS.Y) * TIDIG.X
594 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MUL_U32_U24_e32), TIDReg)
597 // NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)
598 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MAD_U32_U24), TIDReg)
602 // (NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)) + TIDIG.Z
603 BuildMI(Entry, Insert, DL, get(AMDGPU::V_ADD_I32_e32), TIDReg)
608 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_LO_U32_B32_e64),
613 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_HI_U32_B32_e64),
619 BuildMI(Entry, Insert, DL, get(AMDGPU::V_LSHLREV_B32_e32),
623 MFI->setTIDReg(TIDReg);
626 // Add FrameIndex to LDS offset
627 unsigned LDSOffset = MFI->LDSSize + (FrameOffset * WorkGroupSize);
628 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_I32_e32), TmpReg)
635 void SIInstrInfo::insertNOPs(MachineBasicBlock::iterator MI,
644 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(AMDGPU::S_NOP))
649 bool SIInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
650 MachineBasicBlock &MBB = *MI->getParent();
651 DebugLoc DL = MBB.findDebugLoc(MI);
652 switch (MI->getOpcode()) {
653 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
655 case AMDGPU::SI_CONSTDATA_PTR: {
656 unsigned Reg = MI->getOperand(0).getReg();
657 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
658 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
660 BuildMI(MBB, MI, DL, get(AMDGPU::S_GETPC_B64), Reg);
662 // Add 32-bit offset from this instruction to the start of the constant data.
663 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADD_U32), RegLo)
665 .addTargetIndex(AMDGPU::TI_CONSTDATA_START)
666 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit);
667 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADDC_U32), RegHi)
670 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit)
671 .addReg(AMDGPU::SCC, RegState::Implicit);
672 MI->eraseFromParent();
675 case AMDGPU::SGPR_USE:
676 // This is just a placeholder for register allocation.
677 MI->eraseFromParent();
680 case AMDGPU::V_MOV_B64_PSEUDO: {
681 unsigned Dst = MI->getOperand(0).getReg();
682 unsigned DstLo = RI.getSubReg(Dst, AMDGPU::sub0);
683 unsigned DstHi = RI.getSubReg(Dst, AMDGPU::sub1);
685 const MachineOperand &SrcOp = MI->getOperand(1);
686 // FIXME: Will this work for 64-bit floating point immediates?
687 assert(!SrcOp.isFPImm());
689 APInt Imm(64, SrcOp.getImm());
690 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
691 .addImm(Imm.getLoBits(32).getZExtValue())
692 .addReg(Dst, RegState::Implicit);
693 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
694 .addImm(Imm.getHiBits(32).getZExtValue())
695 .addReg(Dst, RegState::Implicit);
697 assert(SrcOp.isReg());
698 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
699 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub0))
700 .addReg(Dst, RegState::Implicit);
701 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
702 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub1))
703 .addReg(Dst, RegState::Implicit);
705 MI->eraseFromParent();
712 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
715 if (MI->getNumOperands() < 3)
718 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
719 AMDGPU::OpName::src0);
720 assert(Src0Idx != -1 && "Should always have src0 operand");
722 MachineOperand &Src0 = MI->getOperand(Src0Idx);
726 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
727 AMDGPU::OpName::src1);
731 MachineOperand &Src1 = MI->getOperand(Src1Idx);
733 // Make sure it's legal to commute operands for VOP2.
734 if (isVOP2(MI->getOpcode()) &&
735 (!isOperandLegal(MI, Src0Idx, &Src1) ||
736 !isOperandLegal(MI, Src1Idx, &Src0))) {
741 // Allow commuting instructions with Imm operands.
742 if (NewMI || !Src1.isImm() ||
743 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
747 // Be sure to copy the source modifiers to the right place.
748 if (MachineOperand *Src0Mods
749 = getNamedOperand(*MI, AMDGPU::OpName::src0_modifiers)) {
750 MachineOperand *Src1Mods
751 = getNamedOperand(*MI, AMDGPU::OpName::src1_modifiers);
753 int Src0ModsVal = Src0Mods->getImm();
754 if (!Src1Mods && Src0ModsVal != 0)
757 // XXX - This assert might be a lie. It might be useful to have a neg
758 // modifier with 0.0.
759 int Src1ModsVal = Src1Mods->getImm();
760 assert((Src1ModsVal == 0) && "Not expecting modifiers with immediates");
762 Src1Mods->setImm(Src0ModsVal);
763 Src0Mods->setImm(Src1ModsVal);
766 unsigned Reg = Src0.getReg();
767 unsigned SubReg = Src0.getSubReg();
769 Src0.ChangeToImmediate(Src1.getImm());
771 llvm_unreachable("Should only have immediates");
773 Src1.ChangeToRegister(Reg, false);
774 Src1.setSubReg(SubReg);
776 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
780 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
785 // This needs to be implemented because the source modifiers may be inserted
786 // between the true commutable operands, and the base
787 // TargetInstrInfo::commuteInstruction uses it.
788 bool SIInstrInfo::findCommutedOpIndices(MachineInstr *MI,
790 unsigned &SrcOpIdx2) const {
791 const MCInstrDesc &MCID = MI->getDesc();
792 if (!MCID.isCommutable())
795 unsigned Opc = MI->getOpcode();
796 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
800 // FIXME: Workaround TargetInstrInfo::commuteInstruction asserting on
802 if (!MI->getOperand(Src0Idx).isReg())
805 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
809 if (!MI->getOperand(Src1Idx).isReg())
812 // If any source modifiers are set, the generic instruction commuting won't
813 // understand how to copy the source modifiers.
814 if (hasModifiersSet(*MI, AMDGPU::OpName::src0_modifiers) ||
815 hasModifiersSet(*MI, AMDGPU::OpName::src1_modifiers))
823 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
824 MachineBasicBlock::iterator I,
826 unsigned SrcReg) const {
827 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
828 DstReg) .addReg(SrcReg);
831 bool SIInstrInfo::isMov(unsigned Opcode) const {
833 default: return false;
834 case AMDGPU::S_MOV_B32:
835 case AMDGPU::S_MOV_B64:
836 case AMDGPU::V_MOV_B32_e32:
837 case AMDGPU::V_MOV_B32_e64:
843 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
844 return RC != &AMDGPU::EXECRegRegClass;
848 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
849 AliasAnalysis *AA) const {
850 switch(MI->getOpcode()) {
851 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
852 case AMDGPU::S_MOV_B32:
853 case AMDGPU::S_MOV_B64:
854 case AMDGPU::V_MOV_B32_e32:
855 return MI->getOperand(1).isImm();
859 static bool offsetsDoNotOverlap(int WidthA, int OffsetA,
860 int WidthB, int OffsetB) {
861 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB;
862 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA;
863 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
864 return LowOffset + LowWidth <= HighOffset;
867 bool SIInstrInfo::checkInstOffsetsDoNotOverlap(MachineInstr *MIa,
868 MachineInstr *MIb) const {
869 unsigned BaseReg0, Offset0;
870 unsigned BaseReg1, Offset1;
872 if (getLdStBaseRegImmOfs(MIa, BaseReg0, Offset0, &RI) &&
873 getLdStBaseRegImmOfs(MIb, BaseReg1, Offset1, &RI)) {
874 assert(MIa->hasOneMemOperand() && MIb->hasOneMemOperand() &&
875 "read2 / write2 not expected here yet");
876 unsigned Width0 = (*MIa->memoperands_begin())->getSize();
877 unsigned Width1 = (*MIb->memoperands_begin())->getSize();
878 if (BaseReg0 == BaseReg1 &&
879 offsetsDoNotOverlap(Width0, Offset0, Width1, Offset1)) {
887 bool SIInstrInfo::areMemAccessesTriviallyDisjoint(MachineInstr *MIa,
889 AliasAnalysis *AA) const {
890 unsigned Opc0 = MIa->getOpcode();
891 unsigned Opc1 = MIb->getOpcode();
893 assert(MIa && (MIa->mayLoad() || MIa->mayStore()) &&
894 "MIa must load from or modify a memory location");
895 assert(MIb && (MIb->mayLoad() || MIb->mayStore()) &&
896 "MIb must load from or modify a memory location");
898 if (MIa->hasUnmodeledSideEffects() || MIb->hasUnmodeledSideEffects())
901 // XXX - Can we relax this between address spaces?
902 if (MIa->hasOrderedMemoryRef() || MIb->hasOrderedMemoryRef())
905 // TODO: Should we check the address space from the MachineMemOperand? That
906 // would allow us to distinguish objects we know don't alias based on the
907 // underlying addres space, even if it was lowered to a different one,
908 // e.g. private accesses lowered to use MUBUF instructions on a scratch
912 return checkInstOffsetsDoNotOverlap(MIa, MIb);
914 return !isFLAT(Opc1);
917 if (isMUBUF(Opc0) || isMTBUF(Opc0)) {
918 if (isMUBUF(Opc1) || isMTBUF(Opc1))
919 return checkInstOffsetsDoNotOverlap(MIa, MIb);
921 return !isFLAT(Opc1) && !isSMRD(Opc1);
926 return checkInstOffsetsDoNotOverlap(MIa, MIb);
928 return !isFLAT(Opc1) && !isMUBUF(Opc0) && !isMTBUF(Opc0);
933 return checkInstOffsetsDoNotOverlap(MIa, MIb);
941 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
942 int64_t SVal = Imm.getSExtValue();
943 if (SVal >= -16 && SVal <= 64)
946 if (Imm.getBitWidth() == 64) {
947 uint64_t Val = Imm.getZExtValue();
948 return (DoubleToBits(0.0) == Val) ||
949 (DoubleToBits(1.0) == Val) ||
950 (DoubleToBits(-1.0) == Val) ||
951 (DoubleToBits(0.5) == Val) ||
952 (DoubleToBits(-0.5) == Val) ||
953 (DoubleToBits(2.0) == Val) ||
954 (DoubleToBits(-2.0) == Val) ||
955 (DoubleToBits(4.0) == Val) ||
956 (DoubleToBits(-4.0) == Val);
959 // The actual type of the operand does not seem to matter as long
960 // as the bits match one of the inline immediate values. For example:
962 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
963 // so it is a legal inline immediate.
965 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
966 // floating-point, so it is a legal inline immediate.
967 uint32_t Val = Imm.getZExtValue();
969 return (FloatToBits(0.0f) == Val) ||
970 (FloatToBits(1.0f) == Val) ||
971 (FloatToBits(-1.0f) == Val) ||
972 (FloatToBits(0.5f) == Val) ||
973 (FloatToBits(-0.5f) == Val) ||
974 (FloatToBits(2.0f) == Val) ||
975 (FloatToBits(-2.0f) == Val) ||
976 (FloatToBits(4.0f) == Val) ||
977 (FloatToBits(-4.0f) == Val);
980 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO,
981 unsigned OpSize) const {
983 // MachineOperand provides no way to tell the true operand size, since it
984 // only records a 64-bit value. We need to know the size to determine if a
985 // 32-bit floating point immediate bit pattern is legal for an integer
986 // immediate. It would be for any 32-bit integer operand, but would not be
989 unsigned BitSize = 8 * OpSize;
990 return isInlineConstant(APInt(BitSize, MO.getImm(), true));
996 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO,
997 unsigned OpSize) const {
998 return MO.isImm() && !isInlineConstant(MO, OpSize);
1001 static bool compareMachineOp(const MachineOperand &Op0,
1002 const MachineOperand &Op1) {
1003 if (Op0.getType() != Op1.getType())
1006 switch (Op0.getType()) {
1007 case MachineOperand::MO_Register:
1008 return Op0.getReg() == Op1.getReg();
1009 case MachineOperand::MO_Immediate:
1010 return Op0.getImm() == Op1.getImm();
1012 llvm_unreachable("Didn't expect to be comparing these operand types");
1016 bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
1017 const MachineOperand &MO) const {
1018 const MCOperandInfo &OpInfo = get(MI->getOpcode()).OpInfo[OpNo];
1020 assert(MO.isImm() || MO.isTargetIndex() || MO.isFI());
1022 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
1025 if (OpInfo.RegClass < 0)
1028 unsigned OpSize = RI.getRegClass(OpInfo.RegClass)->getSize();
1029 if (isLiteralConstant(MO, OpSize))
1030 return RI.opCanUseLiteralConstant(OpInfo.OperandType);
1032 return RI.opCanUseInlineConstant(OpInfo.OperandType);
1035 bool SIInstrInfo::canFoldOffset(unsigned OffsetSize, unsigned AS) const {
1037 case AMDGPUAS::GLOBAL_ADDRESS: {
1038 // MUBUF instructions a 12-bit offset in bytes.
1039 return isUInt<12>(OffsetSize);
1041 case AMDGPUAS::CONSTANT_ADDRESS: {
1042 // SMRD instructions have an 8-bit offset in dwords on SI and
1043 // a 20-bit offset in bytes on VI.
1044 if (RI.ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1045 return isUInt<20>(OffsetSize);
1047 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
1049 case AMDGPUAS::LOCAL_ADDRESS:
1050 case AMDGPUAS::REGION_ADDRESS: {
1051 // The single offset versions have a 16-bit offset in bytes.
1052 return isUInt<16>(OffsetSize);
1054 case AMDGPUAS::PRIVATE_ADDRESS:
1055 // Indirect register addressing does not use any offsets.
1061 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
1062 int Op32 = AMDGPU::getVOPe32(Opcode);
1066 return pseudoToMCOpcode(Op32) != -1;
1069 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
1070 // The src0_modifier operand is present on all instructions
1071 // that have modifiers.
1073 return AMDGPU::getNamedOperandIdx(Opcode,
1074 AMDGPU::OpName::src0_modifiers) != -1;
1077 bool SIInstrInfo::hasModifiersSet(const MachineInstr &MI,
1078 unsigned OpName) const {
1079 const MachineOperand *Mods = getNamedOperand(MI, OpName);
1080 return Mods && Mods->getImm();
1083 bool SIInstrInfo::usesConstantBus(const MachineRegisterInfo &MRI,
1084 const MachineOperand &MO,
1085 unsigned OpSize) const {
1086 // Literal constants use the constant bus.
1087 if (isLiteralConstant(MO, OpSize))
1090 if (!MO.isReg() || !MO.isUse())
1093 if (TargetRegisterInfo::isVirtualRegister(MO.getReg()))
1094 return RI.isSGPRClass(MRI.getRegClass(MO.getReg()));
1096 // FLAT_SCR is just an SGPR pair.
1097 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
1100 // EXEC register uses the constant bus.
1101 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
1104 // SGPRs use the constant bus
1105 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
1106 (!MO.isImplicit() &&
1107 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
1108 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
1115 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
1116 StringRef &ErrInfo) const {
1117 uint16_t Opcode = MI->getOpcode();
1118 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1119 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
1120 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
1121 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
1123 // Make sure the number of operands is correct.
1124 const MCInstrDesc &Desc = get(Opcode);
1125 if (!Desc.isVariadic() &&
1126 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
1127 ErrInfo = "Instruction has wrong number of operands.";
1131 // Make sure the register classes are correct
1132 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
1133 if (MI->getOperand(i).isFPImm()) {
1134 ErrInfo = "FPImm Machine Operands are not supported. ISel should bitcast "
1135 "all fp values to integers.";
1139 switch (Desc.OpInfo[i].OperandType) {
1140 case MCOI::OPERAND_REGISTER:
1141 if (MI->getOperand(i).isImm()) {
1142 ErrInfo = "Illegal immediate value for operand.";
1146 case AMDGPU::OPERAND_REG_IMM32:
1148 case AMDGPU::OPERAND_REG_INLINE_C:
1149 if (MI->getOperand(i).isImm()) {
1150 int RegClass = Desc.OpInfo[i].RegClass;
1151 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
1152 if (!isInlineConstant(MI->getOperand(i), RC->getSize())) {
1153 ErrInfo = "Illegal immediate value for operand.";
1158 case MCOI::OPERAND_IMMEDIATE:
1159 // Check if this operand is an immediate.
1160 // FrameIndex operands will be replaced by immediates, so they are
1162 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFI()) {
1163 ErrInfo = "Expected immediate, but got non-immediate";
1171 if (!MI->getOperand(i).isReg())
1174 int RegClass = Desc.OpInfo[i].RegClass;
1175 if (RegClass != -1) {
1176 unsigned Reg = MI->getOperand(i).getReg();
1177 if (TargetRegisterInfo::isVirtualRegister(Reg))
1180 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
1181 if (!RC->contains(Reg)) {
1182 ErrInfo = "Operand has incorrect register class.";
1190 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
1191 // Only look at the true operands. Only a real operand can use the constant
1192 // bus, and we don't want to check pseudo-operands like the source modifier
1194 const int OpIndices[] = { Src0Idx, Src1Idx, Src2Idx };
1196 unsigned ConstantBusCount = 0;
1197 unsigned SGPRUsed = AMDGPU::NoRegister;
1198 for (int OpIdx : OpIndices) {
1201 const MachineOperand &MO = MI->getOperand(OpIdx);
1202 if (usesConstantBus(MRI, MO, getOpSize(Opcode, OpIdx))) {
1204 if (MO.getReg() != SGPRUsed)
1206 SGPRUsed = MO.getReg();
1212 if (ConstantBusCount > 1) {
1213 ErrInfo = "VOP* instruction uses the constant bus more than once";
1218 // Verify SRC1 for VOP2 and VOPC
1219 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
1220 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1222 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
1228 if (isVOP3(Opcode)) {
1229 if (Src0Idx != -1 &&
1230 isLiteralConstant(MI->getOperand(Src0Idx), getOpSize(Opcode, Src0Idx))) {
1231 ErrInfo = "VOP3 src0 cannot be a literal constant.";
1234 if (Src1Idx != -1 &&
1235 isLiteralConstant(MI->getOperand(Src1Idx), getOpSize(Opcode, Src1Idx))) {
1236 ErrInfo = "VOP3 src1 cannot be a literal constant.";
1239 if (Src2Idx != -1 &&
1240 isLiteralConstant(MI->getOperand(Src2Idx), getOpSize(Opcode, Src2Idx))) {
1241 ErrInfo = "VOP3 src2 cannot be a literal constant.";
1246 // Verify misc. restrictions on specific instructions.
1247 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
1248 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
1249 const MachineOperand &Src0 = MI->getOperand(Src0Idx);
1250 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1251 const MachineOperand &Src2 = MI->getOperand(Src2Idx);
1252 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
1253 if (!compareMachineOp(Src0, Src1) &&
1254 !compareMachineOp(Src0, Src2)) {
1255 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
1264 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
1265 switch (MI.getOpcode()) {
1266 default: return AMDGPU::INSTRUCTION_LIST_END;
1267 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
1268 case AMDGPU::COPY: return AMDGPU::COPY;
1269 case AMDGPU::PHI: return AMDGPU::PHI;
1270 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
1271 case AMDGPU::S_MOV_B32:
1272 return MI.getOperand(1).isReg() ?
1273 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
1274 case AMDGPU::S_ADD_I32:
1275 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
1276 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
1277 case AMDGPU::S_SUB_I32:
1278 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
1279 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
1280 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
1281 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
1282 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
1283 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
1284 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
1285 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
1286 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
1287 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
1288 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
1289 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
1290 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
1291 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
1292 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
1293 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
1294 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
1295 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
1296 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
1297 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
1298 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
1299 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
1300 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
1301 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
1302 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
1303 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
1304 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
1305 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
1306 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
1307 case AMDGPU::S_LOAD_DWORD_IMM:
1308 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1309 case AMDGPU::S_LOAD_DWORDX2_IMM:
1310 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1311 case AMDGPU::S_LOAD_DWORDX4_IMM:
1312 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1313 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e64;
1314 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
1315 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
1319 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
1320 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
1323 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
1324 unsigned OpNo) const {
1325 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1326 const MCInstrDesc &Desc = get(MI.getOpcode());
1327 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
1328 Desc.OpInfo[OpNo].RegClass == -1) {
1329 unsigned Reg = MI.getOperand(OpNo).getReg();
1331 if (TargetRegisterInfo::isVirtualRegister(Reg))
1332 return MRI.getRegClass(Reg);
1333 return RI.getPhysRegClass(Reg);
1336 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
1337 return RI.getRegClass(RCID);
1340 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
1341 switch (MI.getOpcode()) {
1343 case AMDGPU::REG_SEQUENCE:
1345 case AMDGPU::INSERT_SUBREG:
1346 return RI.hasVGPRs(getOpRegClass(MI, 0));
1348 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
1352 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
1353 MachineBasicBlock::iterator I = MI;
1354 MachineBasicBlock *MBB = MI->getParent();
1355 MachineOperand &MO = MI->getOperand(OpIdx);
1356 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1357 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
1358 const TargetRegisterClass *RC = RI.getRegClass(RCID);
1359 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1361 Opcode = AMDGPU::COPY;
1362 else if (RI.isSGPRClass(RC))
1363 Opcode = AMDGPU::S_MOV_B32;
1366 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
1367 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC))
1368 VRC = &AMDGPU::VReg_64RegClass;
1370 VRC = &AMDGPU::VGPR_32RegClass;
1372 unsigned Reg = MRI.createVirtualRegister(VRC);
1373 DebugLoc DL = MBB->findDebugLoc(I);
1374 BuildMI(*MI->getParent(), I, DL, get(Opcode), Reg)
1376 MO.ChangeToRegister(Reg, false);
1379 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
1380 MachineRegisterInfo &MRI,
1381 MachineOperand &SuperReg,
1382 const TargetRegisterClass *SuperRC,
1384 const TargetRegisterClass *SubRC)
1386 assert(SuperReg.isReg());
1388 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
1389 unsigned SubReg = MRI.createVirtualRegister(SubRC);
1391 // Just in case the super register is itself a sub-register, copy it to a new
1392 // value so we don't need to worry about merging its subreg index with the
1393 // SubIdx passed to this function. The register coalescer should be able to
1394 // eliminate this extra copy.
1395 MachineBasicBlock *MBB = MI->getParent();
1396 DebugLoc DL = MI->getDebugLoc();
1398 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), NewSuperReg)
1399 .addReg(SuperReg.getReg(), 0, SuperReg.getSubReg());
1401 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
1402 .addReg(NewSuperReg, 0, SubIdx);
1407 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
1408 MachineBasicBlock::iterator MII,
1409 MachineRegisterInfo &MRI,
1411 const TargetRegisterClass *SuperRC,
1413 const TargetRegisterClass *SubRC) const {
1415 // XXX - Is there a better way to do this?
1416 if (SubIdx == AMDGPU::sub0)
1417 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
1418 if (SubIdx == AMDGPU::sub1)
1419 return MachineOperand::CreateImm(Op.getImm() >> 32);
1421 llvm_unreachable("Unhandled register index for immediate");
1424 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
1426 return MachineOperand::CreateReg(SubReg, false);
1429 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
1430 MachineBasicBlock::iterator MI,
1431 MachineRegisterInfo &MRI,
1432 const TargetRegisterClass *RC,
1433 const MachineOperand &Op) const {
1434 MachineBasicBlock *MBB = MI->getParent();
1435 DebugLoc DL = MI->getDebugLoc();
1436 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1437 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1438 unsigned Dst = MRI.createVirtualRegister(RC);
1440 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1442 .addImm(Op.getImm() & 0xFFFFFFFF);
1443 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1445 .addImm(Op.getImm() >> 32);
1447 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
1449 .addImm(AMDGPU::sub0)
1451 .addImm(AMDGPU::sub1);
1453 Worklist.push_back(Lo);
1454 Worklist.push_back(Hi);
1459 // Change the order of operands from (0, 1, 2) to (0, 2, 1)
1460 void SIInstrInfo::swapOperands(MachineBasicBlock::iterator Inst) const {
1461 assert(Inst->getNumExplicitOperands() == 3);
1462 MachineOperand Op1 = Inst->getOperand(1);
1463 Inst->RemoveOperand(1);
1464 Inst->addOperand(Op1);
1467 bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
1468 const MachineOperand *MO) const {
1469 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1470 const MCInstrDesc &InstDesc = get(MI->getOpcode());
1471 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
1472 const TargetRegisterClass *DefinedRC =
1473 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
1475 MO = &MI->getOperand(OpIdx);
1477 if (isVALU(InstDesc.Opcode) &&
1478 usesConstantBus(MRI, *MO, DefinedRC->getSize())) {
1480 MO->isReg() ? MO->getReg() : (unsigned)AMDGPU::NoRegister;
1481 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1484 const MachineOperand &Op = MI->getOperand(i);
1485 if (Op.isReg() && Op.getReg() != SGPRUsed &&
1486 usesConstantBus(MRI, Op, getOpSize(*MI, i))) {
1494 const TargetRegisterClass *RC = MRI.getRegClass(MO->getReg());
1496 // In order to be legal, the common sub-class must be equal to the
1497 // class of the current operand. For example:
1499 // v_mov_b32 s0 ; Operand defined as vsrc_32
1500 // ; RI.getCommonSubClass(s0,vsrc_32) = sgpr ; LEGAL
1502 // s_sendmsg 0, s0 ; Operand defined as m0reg
1503 // ; RI.getCommonSubClass(s0,m0reg) = m0reg ; NOT LEGAL
1505 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass)) == RC;
1509 // Handle non-register types that are treated like immediates.
1510 assert(MO->isImm() || MO->isTargetIndex() || MO->isFI());
1513 // This operand expects an immediate.
1517 return isImmOperandLegal(MI, OpIdx, *MO);
1520 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
1521 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1523 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1524 AMDGPU::OpName::src0);
1525 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1526 AMDGPU::OpName::src1);
1527 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1528 AMDGPU::OpName::src2);
1531 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
1533 if (!isOperandLegal(MI, Src0Idx))
1534 legalizeOpWithMove(MI, Src0Idx);
1537 if (isOperandLegal(MI, Src1Idx))
1540 // Usually src0 of VOP2 instructions allow more types of inputs
1541 // than src1, so try to commute the instruction to decrease our
1542 // chances of having to insert a MOV instruction to legalize src1.
1543 if (MI->isCommutable()) {
1544 if (commuteInstruction(MI))
1545 // If we are successful in commuting, then we know MI is legal, so
1550 legalizeOpWithMove(MI, Src1Idx);
1554 // XXX - Do any VOP3 instructions read VCC?
1556 if (isVOP3(MI->getOpcode())) {
1557 int VOP3Idx[3] = { Src0Idx, Src1Idx, Src2Idx };
1559 // Find the one SGPR operand we are allowed to use.
1560 unsigned SGPRReg = findUsedSGPR(MI, VOP3Idx);
1562 for (unsigned i = 0; i < 3; ++i) {
1563 int Idx = VOP3Idx[i];
1566 MachineOperand &MO = MI->getOperand(Idx);
1569 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
1570 continue; // VGPRs are legal
1572 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
1574 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
1575 SGPRReg = MO.getReg();
1576 // We can use one SGPR in each VOP3 instruction.
1579 } else if (!isLiteralConstant(MO, getOpSize(MI->getOpcode(), Idx))) {
1580 // If it is not a register and not a literal constant, then it must be
1581 // an inline constant which is always legal.
1584 // If we make it this far, then the operand is not legal and we must
1586 legalizeOpWithMove(MI, Idx);
1590 // Legalize REG_SEQUENCE and PHI
1591 // The register class of the operands much be the same type as the register
1592 // class of the output.
1593 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
1594 MI->getOpcode() == AMDGPU::PHI) {
1595 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
1596 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1597 if (!MI->getOperand(i).isReg() ||
1598 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1600 const TargetRegisterClass *OpRC =
1601 MRI.getRegClass(MI->getOperand(i).getReg());
1602 if (RI.hasVGPRs(OpRC)) {
1609 // If any of the operands are VGPR registers, then they all most be
1610 // otherwise we will create illegal VGPR->SGPR copies when legalizing
1612 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
1615 VRC = RI.getEquivalentVGPRClass(SRC);
1622 // Update all the operands so they have the same type.
1623 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1624 if (!MI->getOperand(i).isReg() ||
1625 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1627 unsigned DstReg = MRI.createVirtualRegister(RC);
1628 MachineBasicBlock *InsertBB;
1629 MachineBasicBlock::iterator Insert;
1630 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
1631 InsertBB = MI->getParent();
1634 // MI is a PHI instruction.
1635 InsertBB = MI->getOperand(i + 1).getMBB();
1636 Insert = InsertBB->getFirstTerminator();
1638 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
1639 get(AMDGPU::COPY), DstReg)
1640 .addOperand(MI->getOperand(i));
1641 MI->getOperand(i).setReg(DstReg);
1645 // Legalize INSERT_SUBREG
1646 // src0 must have the same register class as dst
1647 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
1648 unsigned Dst = MI->getOperand(0).getReg();
1649 unsigned Src0 = MI->getOperand(1).getReg();
1650 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
1651 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
1652 if (DstRC != Src0RC) {
1653 MachineBasicBlock &MBB = *MI->getParent();
1654 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
1655 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
1657 MI->getOperand(1).setReg(NewSrc0);
1662 // Legalize MUBUF* instructions
1663 // FIXME: If we start using the non-addr64 instructions for compute, we
1664 // may need to legalize them here.
1666 AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::srsrc);
1667 if (SRsrcIdx != -1) {
1668 // We have an MUBUF instruction
1669 MachineOperand *SRsrc = &MI->getOperand(SRsrcIdx);
1670 unsigned SRsrcRC = get(MI->getOpcode()).OpInfo[SRsrcIdx].RegClass;
1671 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
1672 RI.getRegClass(SRsrcRC))) {
1673 // The operands are legal.
1674 // FIXME: We may need to legalize operands besided srsrc.
1678 MachineBasicBlock &MBB = *MI->getParent();
1679 // Extract the the ptr from the resource descriptor.
1681 // SRsrcPtrLo = srsrc:sub0
1682 unsigned SRsrcPtrLo = buildExtractSubReg(MI, MRI, *SRsrc,
1683 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VGPR_32RegClass);
1685 // SRsrcPtrHi = srsrc:sub1
1686 unsigned SRsrcPtrHi = buildExtractSubReg(MI, MRI, *SRsrc,
1687 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VGPR_32RegClass);
1689 // Create an empty resource descriptor
1690 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1691 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1692 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1693 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1694 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
1697 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1701 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1702 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1704 .addImm(RsrcDataFormat & 0xFFFFFFFF);
1706 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1707 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1709 .addImm(RsrcDataFormat >> 32);
1711 // NewSRsrc = {Zero64, SRsrcFormat}
1712 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1715 .addImm(AMDGPU::sub0_sub1)
1716 .addReg(SRsrcFormatLo)
1717 .addImm(AMDGPU::sub2)
1718 .addReg(SRsrcFormatHi)
1719 .addImm(AMDGPU::sub3);
1721 MachineOperand *VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1722 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
1723 unsigned NewVAddrLo;
1724 unsigned NewVAddrHi;
1726 // This is already an ADDR64 instruction so we need to add the pointer
1727 // extracted from the resource descriptor to the current value of VAddr.
1728 NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1729 NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1731 // NewVaddrLo = SRsrcPtrLo + VAddr:sub0
1732 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1735 .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
1736 .addReg(AMDGPU::VCC, RegState::ImplicitDefine);
1738 // NewVaddrHi = SRsrcPtrHi + VAddr:sub1
1739 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1742 .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
1743 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1744 .addReg(AMDGPU::VCC, RegState::Implicit);
1747 // This instructions is the _OFFSET variant, so we need to convert it to
1749 MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
1750 MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
1751 MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
1753 // Create the new instruction.
1754 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
1755 MachineInstr *Addr64 =
1756 BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
1759 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
1760 // This will be replaced later
1761 // with the new value of vaddr.
1762 .addOperand(*SOffset)
1763 .addOperand(*Offset);
1765 MI->removeFromParent();
1768 NewVAddrLo = SRsrcPtrLo;
1769 NewVAddrHi = SRsrcPtrHi;
1770 VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1771 SRsrc = getNamedOperand(*MI, AMDGPU::OpName::srsrc);
1774 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1775 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1778 .addImm(AMDGPU::sub0)
1780 .addImm(AMDGPU::sub1);
1783 // Update the instruction to use NewVaddr
1784 VAddr->setReg(NewVAddr);
1785 // Update the instruction to use NewSRsrc
1786 SRsrc->setReg(NewSRsrc);
1790 void SIInstrInfo::splitSMRD(MachineInstr *MI,
1791 const TargetRegisterClass *HalfRC,
1792 unsigned HalfImmOp, unsigned HalfSGPROp,
1793 MachineInstr *&Lo, MachineInstr *&Hi) const {
1795 DebugLoc DL = MI->getDebugLoc();
1796 MachineBasicBlock *MBB = MI->getParent();
1797 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1798 unsigned RegLo = MRI.createVirtualRegister(HalfRC);
1799 unsigned RegHi = MRI.createVirtualRegister(HalfRC);
1800 unsigned HalfSize = HalfRC->getSize();
1801 const MachineOperand *OffOp =
1802 getNamedOperand(*MI, AMDGPU::OpName::offset);
1803 const MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
1805 // The SMRD has an 8-bit offset in dwords on SI and a 20-bit offset in bytes
1808 bool isVI = RI.ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS;
1809 unsigned OffScale = isVI ? 1 : 4;
1810 // Handle the _IMM variant
1811 unsigned LoOffset = OffOp->getImm() * OffScale;
1812 unsigned HiOffset = LoOffset + HalfSize;
1813 Lo = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegLo)
1815 .addImm(LoOffset / OffScale);
1817 if (!isUInt<20>(HiOffset) || (!isVI && !isUInt<8>(HiOffset / OffScale))) {
1818 unsigned OffsetSGPR =
1819 MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1820 BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32), OffsetSGPR)
1821 .addImm(HiOffset); // The offset in register is in bytes.
1822 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
1824 .addReg(OffsetSGPR);
1826 Hi = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegHi)
1828 .addImm(HiOffset / OffScale);
1831 // Handle the _SGPR variant
1832 MachineOperand *SOff = getNamedOperand(*MI, AMDGPU::OpName::soff);
1833 Lo = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegLo)
1836 unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1837 BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
1840 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp))
1842 .addReg(OffsetSGPR);
1845 unsigned SubLo, SubHi;
1848 SubLo = AMDGPU::sub0;
1849 SubHi = AMDGPU::sub1;
1852 SubLo = AMDGPU::sub0_sub1;
1853 SubHi = AMDGPU::sub2_sub3;
1856 SubLo = AMDGPU::sub0_sub1_sub2_sub3;
1857 SubHi = AMDGPU::sub4_sub5_sub6_sub7;
1860 SubLo = AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
1861 SubHi = AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15;
1864 llvm_unreachable("Unhandled HalfSize");
1867 BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE))
1868 .addOperand(MI->getOperand(0))
1875 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
1876 MachineBasicBlock *MBB = MI->getParent();
1877 switch (MI->getOpcode()) {
1878 case AMDGPU::S_LOAD_DWORD_IMM:
1879 case AMDGPU::S_LOAD_DWORD_SGPR:
1880 case AMDGPU::S_LOAD_DWORDX2_IMM:
1881 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1882 case AMDGPU::S_LOAD_DWORDX4_IMM:
1883 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1884 unsigned NewOpcode = getVALUOp(*MI);
1888 if (MI->getOperand(2).isReg()) {
1889 RegOffset = MI->getOperand(2).getReg();
1892 assert(MI->getOperand(2).isImm());
1893 // SMRD instructions take a dword offsets on SI and byte offset on VI
1894 // and MUBUF instructions always take a byte offset.
1895 ImmOffset = MI->getOperand(2).getImm();
1896 if (RI.ST.getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS)
1898 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1900 if (isUInt<12>(ImmOffset)) {
1901 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1905 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1912 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1913 unsigned DWord0 = RegOffset;
1914 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1915 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1916 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1917 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
1919 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
1921 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
1922 .addImm(RsrcDataFormat & 0xFFFFFFFF);
1923 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
1924 .addImm(RsrcDataFormat >> 32);
1925 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
1927 .addImm(AMDGPU::sub0)
1929 .addImm(AMDGPU::sub1)
1931 .addImm(AMDGPU::sub2)
1933 .addImm(AMDGPU::sub3);
1934 MI->setDesc(get(NewOpcode));
1935 if (MI->getOperand(2).isReg()) {
1936 MI->getOperand(2).setReg(MI->getOperand(1).getReg());
1938 MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
1940 MI->getOperand(1).setReg(SRsrc);
1941 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(0));
1942 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
1944 const TargetRegisterClass *NewDstRC =
1945 RI.getRegClass(get(NewOpcode).OpInfo[0].RegClass);
1947 unsigned DstReg = MI->getOperand(0).getReg();
1948 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1949 MRI.replaceRegWith(DstReg, NewDstReg);
1952 case AMDGPU::S_LOAD_DWORDX8_IMM:
1953 case AMDGPU::S_LOAD_DWORDX8_SGPR: {
1954 MachineInstr *Lo, *Hi;
1955 splitSMRD(MI, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM,
1956 AMDGPU::S_LOAD_DWORDX4_SGPR, Lo, Hi);
1957 MI->eraseFromParent();
1958 moveSMRDToVALU(Lo, MRI);
1959 moveSMRDToVALU(Hi, MRI);
1963 case AMDGPU::S_LOAD_DWORDX16_IMM:
1964 case AMDGPU::S_LOAD_DWORDX16_SGPR: {
1965 MachineInstr *Lo, *Hi;
1966 splitSMRD(MI, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM,
1967 AMDGPU::S_LOAD_DWORDX8_SGPR, Lo, Hi);
1968 MI->eraseFromParent();
1969 moveSMRDToVALU(Lo, MRI);
1970 moveSMRDToVALU(Hi, MRI);
1976 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
1977 SmallVector<MachineInstr *, 128> Worklist;
1978 Worklist.push_back(&TopInst);
1980 while (!Worklist.empty()) {
1981 MachineInstr *Inst = Worklist.pop_back_val();
1982 MachineBasicBlock *MBB = Inst->getParent();
1983 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1985 unsigned Opcode = Inst->getOpcode();
1986 unsigned NewOpcode = getVALUOp(*Inst);
1988 // Handle some special cases
1991 if (isSMRD(Inst->getOpcode())) {
1992 moveSMRDToVALU(Inst, MRI);
1995 case AMDGPU::S_MOV_B64: {
1996 DebugLoc DL = Inst->getDebugLoc();
1998 // If the source operand is a register we can replace this with a
2000 if (Inst->getOperand(1).isReg()) {
2001 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
2002 .addOperand(Inst->getOperand(0))
2003 .addOperand(Inst->getOperand(1));
2004 Worklist.push_back(Copy);
2006 // Otherwise, we need to split this into two movs, because there is
2007 // no 64-bit VALU move instruction.
2008 unsigned Reg = Inst->getOperand(0).getReg();
2009 unsigned Dst = split64BitImm(Worklist,
2012 MRI.getRegClass(Reg),
2013 Inst->getOperand(1));
2014 MRI.replaceRegWith(Reg, Dst);
2016 Inst->eraseFromParent();
2019 case AMDGPU::S_AND_B64:
2020 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
2021 Inst->eraseFromParent();
2024 case AMDGPU::S_OR_B64:
2025 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
2026 Inst->eraseFromParent();
2029 case AMDGPU::S_XOR_B64:
2030 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
2031 Inst->eraseFromParent();
2034 case AMDGPU::S_NOT_B64:
2035 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
2036 Inst->eraseFromParent();
2039 case AMDGPU::S_BCNT1_I32_B64:
2040 splitScalar64BitBCNT(Worklist, Inst);
2041 Inst->eraseFromParent();
2044 case AMDGPU::S_BFE_I64: {
2045 splitScalar64BitBFE(Worklist, Inst);
2046 Inst->eraseFromParent();
2050 case AMDGPU::S_LSHL_B32:
2051 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2052 NewOpcode = AMDGPU::V_LSHLREV_B32_e64;
2056 case AMDGPU::S_ASHR_I32:
2057 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2058 NewOpcode = AMDGPU::V_ASHRREV_I32_e64;
2062 case AMDGPU::S_LSHR_B32:
2063 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2064 NewOpcode = AMDGPU::V_LSHRREV_B32_e64;
2068 case AMDGPU::S_LSHL_B64:
2069 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2070 NewOpcode = AMDGPU::V_LSHLREV_B64;
2074 case AMDGPU::S_ASHR_I64:
2075 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2076 NewOpcode = AMDGPU::V_ASHRREV_I64;
2080 case AMDGPU::S_LSHR_B64:
2081 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2082 NewOpcode = AMDGPU::V_LSHRREV_B64;
2087 case AMDGPU::S_BFE_U64:
2088 case AMDGPU::S_BFM_B64:
2089 llvm_unreachable("Moving this op to VALU not implemented");
2092 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
2093 // We cannot move this instruction to the VALU, so we should try to
2094 // legalize its operands instead.
2095 legalizeOperands(Inst);
2099 // Use the new VALU Opcode.
2100 const MCInstrDesc &NewDesc = get(NewOpcode);
2101 Inst->setDesc(NewDesc);
2103 // Remove any references to SCC. Vector instructions can't read from it, and
2104 // We're just about to add the implicit use / defs of VCC, and we don't want
2106 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
2107 MachineOperand &Op = Inst->getOperand(i);
2108 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
2109 Inst->RemoveOperand(i);
2112 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
2113 // We are converting these to a BFE, so we need to add the missing
2114 // operands for the size and offset.
2115 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
2116 Inst->addOperand(MachineOperand::CreateImm(0));
2117 Inst->addOperand(MachineOperand::CreateImm(Size));
2119 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
2120 // The VALU version adds the second operand to the result, so insert an
2122 Inst->addOperand(MachineOperand::CreateImm(0));
2125 addDescImplicitUseDef(NewDesc, Inst);
2127 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
2128 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
2129 // If we need to move this to VGPRs, we need to unpack the second operand
2130 // back into the 2 separate ones for bit offset and width.
2131 assert(OffsetWidthOp.isImm() &&
2132 "Scalar BFE is only implemented for constant width and offset");
2133 uint32_t Imm = OffsetWidthOp.getImm();
2135 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
2136 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
2137 Inst->RemoveOperand(2); // Remove old immediate.
2138 Inst->addOperand(MachineOperand::CreateImm(Offset));
2139 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
2142 // Update the destination register class.
2144 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
2147 // For target instructions, getOpRegClass just returns the virtual
2148 // register class associated with the operand, so we need to find an
2149 // equivalent VGPR register class in order to move the instruction to the
2153 case AMDGPU::REG_SEQUENCE:
2154 case AMDGPU::INSERT_SUBREG:
2155 if (RI.hasVGPRs(NewDstRC))
2157 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
2165 unsigned DstReg = Inst->getOperand(0).getReg();
2166 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
2167 MRI.replaceRegWith(DstReg, NewDstReg);
2169 // Legalize the operands
2170 legalizeOperands(Inst);
2172 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
2173 E = MRI.use_end(); I != E; ++I) {
2174 MachineInstr &UseMI = *I->getParent();
2175 if (!canReadVGPR(UseMI, I.getOperandNo())) {
2176 Worklist.push_back(&UseMI);
2182 //===----------------------------------------------------------------------===//
2183 // Indirect addressing callbacks
2184 //===----------------------------------------------------------------------===//
2186 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
2187 unsigned Channel) const {
2188 assert(Channel == 0);
2192 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
2193 return &AMDGPU::VGPR_32RegClass;
2196 void SIInstrInfo::splitScalar64BitUnaryOp(
2197 SmallVectorImpl<MachineInstr *> &Worklist,
2199 unsigned Opcode) const {
2200 MachineBasicBlock &MBB = *Inst->getParent();
2201 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2203 MachineOperand &Dest = Inst->getOperand(0);
2204 MachineOperand &Src0 = Inst->getOperand(1);
2205 DebugLoc DL = Inst->getDebugLoc();
2207 MachineBasicBlock::iterator MII = Inst;
2209 const MCInstrDesc &InstDesc = get(Opcode);
2210 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2211 MRI.getRegClass(Src0.getReg()) :
2212 &AMDGPU::SGPR_32RegClass;
2214 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2216 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2217 AMDGPU::sub0, Src0SubRC);
2219 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2220 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2222 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2223 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2224 .addOperand(SrcReg0Sub0);
2226 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2227 AMDGPU::sub1, Src0SubRC);
2229 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2230 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2231 .addOperand(SrcReg0Sub1);
2233 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2234 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2236 .addImm(AMDGPU::sub0)
2238 .addImm(AMDGPU::sub1);
2240 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2242 // Try to legalize the operands in case we need to swap the order to keep it
2244 Worklist.push_back(LoHalf);
2245 Worklist.push_back(HiHalf);
2248 void SIInstrInfo::splitScalar64BitBinaryOp(
2249 SmallVectorImpl<MachineInstr *> &Worklist,
2251 unsigned Opcode) const {
2252 MachineBasicBlock &MBB = *Inst->getParent();
2253 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2255 MachineOperand &Dest = Inst->getOperand(0);
2256 MachineOperand &Src0 = Inst->getOperand(1);
2257 MachineOperand &Src1 = Inst->getOperand(2);
2258 DebugLoc DL = Inst->getDebugLoc();
2260 MachineBasicBlock::iterator MII = Inst;
2262 const MCInstrDesc &InstDesc = get(Opcode);
2263 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2264 MRI.getRegClass(Src0.getReg()) :
2265 &AMDGPU::SGPR_32RegClass;
2267 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2268 const TargetRegisterClass *Src1RC = Src1.isReg() ?
2269 MRI.getRegClass(Src1.getReg()) :
2270 &AMDGPU::SGPR_32RegClass;
2272 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
2274 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2275 AMDGPU::sub0, Src0SubRC);
2276 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2277 AMDGPU::sub0, Src1SubRC);
2279 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2280 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2282 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2283 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2284 .addOperand(SrcReg0Sub0)
2285 .addOperand(SrcReg1Sub0);
2287 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2288 AMDGPU::sub1, Src0SubRC);
2289 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2290 AMDGPU::sub1, Src1SubRC);
2292 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2293 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2294 .addOperand(SrcReg0Sub1)
2295 .addOperand(SrcReg1Sub1);
2297 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2298 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2300 .addImm(AMDGPU::sub0)
2302 .addImm(AMDGPU::sub1);
2304 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2306 // Try to legalize the operands in case we need to swap the order to keep it
2308 Worklist.push_back(LoHalf);
2309 Worklist.push_back(HiHalf);
2312 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
2313 MachineInstr *Inst) const {
2314 MachineBasicBlock &MBB = *Inst->getParent();
2315 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2317 MachineBasicBlock::iterator MII = Inst;
2318 DebugLoc DL = Inst->getDebugLoc();
2320 MachineOperand &Dest = Inst->getOperand(0);
2321 MachineOperand &Src = Inst->getOperand(1);
2323 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e64);
2324 const TargetRegisterClass *SrcRC = Src.isReg() ?
2325 MRI.getRegClass(Src.getReg()) :
2326 &AMDGPU::SGPR_32RegClass;
2328 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2329 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2331 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
2333 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2334 AMDGPU::sub0, SrcSubRC);
2335 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2336 AMDGPU::sub1, SrcSubRC);
2338 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
2339 .addOperand(SrcRegSub0)
2342 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
2343 .addOperand(SrcRegSub1)
2346 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2348 Worklist.push_back(First);
2349 Worklist.push_back(Second);
2352 void SIInstrInfo::splitScalar64BitBFE(SmallVectorImpl<MachineInstr *> &Worklist,
2353 MachineInstr *Inst) const {
2354 MachineBasicBlock &MBB = *Inst->getParent();
2355 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2356 MachineBasicBlock::iterator MII = Inst;
2357 DebugLoc DL = Inst->getDebugLoc();
2359 MachineOperand &Dest = Inst->getOperand(0);
2360 uint32_t Imm = Inst->getOperand(2).getImm();
2361 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
2362 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
2366 // Only sext_inreg cases handled.
2367 assert(Inst->getOpcode() == AMDGPU::S_BFE_I64 &&
2372 if (BitWidth < 32) {
2373 unsigned MidRegLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2374 unsigned MidRegHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2375 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
2377 BuildMI(MBB, MII, DL, get(AMDGPU::V_BFE_I32), MidRegLo)
2378 .addReg(Inst->getOperand(1).getReg(), 0, AMDGPU::sub0)
2382 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e32), MidRegHi)
2386 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
2388 .addImm(AMDGPU::sub0)
2390 .addImm(AMDGPU::sub1);
2392 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2396 MachineOperand &Src = Inst->getOperand(1);
2397 unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2398 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
2400 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e64), TmpReg)
2402 .addReg(Src.getReg(), 0, AMDGPU::sub0);
2404 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
2405 .addReg(Src.getReg(), 0, AMDGPU::sub0)
2406 .addImm(AMDGPU::sub0)
2408 .addImm(AMDGPU::sub1);
2410 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2413 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
2414 MachineInstr *Inst) const {
2415 // Add the implict and explicit register definitions.
2416 if (NewDesc.ImplicitUses) {
2417 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
2418 unsigned Reg = NewDesc.ImplicitUses[i];
2419 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
2423 if (NewDesc.ImplicitDefs) {
2424 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
2425 unsigned Reg = NewDesc.ImplicitDefs[i];
2426 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
2431 unsigned SIInstrInfo::findUsedSGPR(const MachineInstr *MI,
2432 int OpIndices[3]) const {
2433 const MCInstrDesc &Desc = get(MI->getOpcode());
2435 // Find the one SGPR operand we are allowed to use.
2436 unsigned SGPRReg = AMDGPU::NoRegister;
2438 // First we need to consider the instruction's operand requirements before
2439 // legalizing. Some operands are required to be SGPRs, such as implicit uses
2440 // of VCC, but we are still bound by the constant bus requirement to only use
2443 // If the operand's class is an SGPR, we can never move it.
2445 for (const MachineOperand &MO : MI->implicit_operands()) {
2446 // We only care about reads.
2450 if (MO.getReg() == AMDGPU::VCC)
2453 if (MO.getReg() == AMDGPU::FLAT_SCR)
2454 return AMDGPU::FLAT_SCR;
2457 unsigned UsedSGPRs[3] = { AMDGPU::NoRegister };
2458 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
2460 for (unsigned i = 0; i < 3; ++i) {
2461 int Idx = OpIndices[i];
2465 const MachineOperand &MO = MI->getOperand(Idx);
2466 if (RI.isSGPRClassID(Desc.OpInfo[Idx].RegClass))
2467 SGPRReg = MO.getReg();
2469 if (MO.isReg() && RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
2470 UsedSGPRs[i] = MO.getReg();
2473 if (SGPRReg != AMDGPU::NoRegister)
2476 // We don't have a required SGPR operand, so we have a bit more freedom in
2477 // selecting operands to move.
2479 // Try to select the most used SGPR. If an SGPR is equal to one of the
2480 // others, we choose that.
2483 // V_FMA_F32 v0, s0, s0, s0 -> No moves
2484 // V_FMA_F32 v0, s0, s1, s0 -> Move s1
2486 if (UsedSGPRs[0] != AMDGPU::NoRegister) {
2487 if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
2488 SGPRReg = UsedSGPRs[0];
2491 if (SGPRReg == AMDGPU::NoRegister && UsedSGPRs[1] != AMDGPU::NoRegister) {
2492 if (UsedSGPRs[1] == UsedSGPRs[2])
2493 SGPRReg = UsedSGPRs[1];
2499 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
2500 MachineBasicBlock *MBB,
2501 MachineBasicBlock::iterator I,
2503 unsigned Address, unsigned OffsetReg) const {
2504 const DebugLoc &DL = MBB->findDebugLoc(I);
2505 unsigned IndirectBaseReg = AMDGPU::VGPR_32RegClass.getRegister(
2506 getIndirectIndexBegin(*MBB->getParent()));
2508 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
2509 .addReg(IndirectBaseReg, RegState::Define)
2510 .addOperand(I->getOperand(0))
2511 .addReg(IndirectBaseReg)
2517 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
2518 MachineBasicBlock *MBB,
2519 MachineBasicBlock::iterator I,
2521 unsigned Address, unsigned OffsetReg) const {
2522 const DebugLoc &DL = MBB->findDebugLoc(I);
2523 unsigned IndirectBaseReg = AMDGPU::VGPR_32RegClass.getRegister(
2524 getIndirectIndexBegin(*MBB->getParent()));
2526 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
2527 .addOperand(I->getOperand(0))
2528 .addOperand(I->getOperand(1))
2529 .addReg(IndirectBaseReg)
2535 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
2536 const MachineFunction &MF) const {
2537 int End = getIndirectIndexEnd(MF);
2538 int Begin = getIndirectIndexBegin(MF);
2544 for (int Index = Begin; Index <= End; ++Index)
2545 Reserved.set(AMDGPU::VGPR_32RegClass.getRegister(Index));
2547 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
2548 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
2550 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
2551 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
2553 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
2554 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
2556 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
2557 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
2559 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
2560 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));
2563 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
2564 unsigned OperandName) const {
2565 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
2569 return &MI.getOperand(Idx);
2572 uint64_t SIInstrInfo::getDefaultRsrcDataFormat() const {
2573 uint64_t RsrcDataFormat = AMDGPU::RSRC_DATA_FORMAT;
2574 if (ST.isAmdHsaOS())
2575 RsrcDataFormat |= (1ULL << 56);
2577 return RsrcDataFormat;