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"
29 SIInstrInfo::SIInstrInfo(const AMDGPUSubtarget &st)
30 : AMDGPUInstrInfo(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::M0 == DestReg) {
329 // Check if M0 isn't already set to this value
330 for (MachineBasicBlock::reverse_iterator E = MBB.rend(),
331 I = MachineBasicBlock::reverse_iterator(MI); I != E; ++I) {
333 if (!I->definesRegister(AMDGPU::M0))
336 unsigned Opc = I->getOpcode();
337 if (Opc != TargetOpcode::COPY && Opc != AMDGPU::S_MOV_B32)
340 if (!I->readsRegister(SrcReg))
343 // The copy isn't necessary
348 if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
349 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
350 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
351 .addReg(SrcReg, getKillRegState(KillSrc));
354 } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
355 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
356 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
357 .addReg(SrcReg, getKillRegState(KillSrc));
360 } else if (AMDGPU::SReg_128RegClass.contains(DestReg)) {
361 assert(AMDGPU::SReg_128RegClass.contains(SrcReg));
362 Opcode = AMDGPU::S_MOV_B32;
365 } else if (AMDGPU::SReg_256RegClass.contains(DestReg)) {
366 assert(AMDGPU::SReg_256RegClass.contains(SrcReg));
367 Opcode = AMDGPU::S_MOV_B32;
370 } else if (AMDGPU::SReg_512RegClass.contains(DestReg)) {
371 assert(AMDGPU::SReg_512RegClass.contains(SrcReg));
372 Opcode = AMDGPU::S_MOV_B32;
373 SubIndices = Sub0_15;
375 } else if (AMDGPU::VReg_32RegClass.contains(DestReg)) {
376 assert(AMDGPU::VReg_32RegClass.contains(SrcReg) ||
377 AMDGPU::SReg_32RegClass.contains(SrcReg));
378 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
379 .addReg(SrcReg, getKillRegState(KillSrc));
382 } else if (AMDGPU::VReg_64RegClass.contains(DestReg)) {
383 assert(AMDGPU::VReg_64RegClass.contains(SrcReg) ||
384 AMDGPU::SReg_64RegClass.contains(SrcReg));
385 Opcode = AMDGPU::V_MOV_B32_e32;
388 } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
389 assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
390 Opcode = AMDGPU::V_MOV_B32_e32;
393 } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
394 assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
395 AMDGPU::SReg_128RegClass.contains(SrcReg));
396 Opcode = AMDGPU::V_MOV_B32_e32;
399 } else if (AMDGPU::VReg_256RegClass.contains(DestReg)) {
400 assert(AMDGPU::VReg_256RegClass.contains(SrcReg) ||
401 AMDGPU::SReg_256RegClass.contains(SrcReg));
402 Opcode = AMDGPU::V_MOV_B32_e32;
405 } else if (AMDGPU::VReg_512RegClass.contains(DestReg)) {
406 assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
407 AMDGPU::SReg_512RegClass.contains(SrcReg));
408 Opcode = AMDGPU::V_MOV_B32_e32;
409 SubIndices = Sub0_15;
412 llvm_unreachable("Can't copy register!");
415 while (unsigned SubIdx = *SubIndices++) {
416 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
417 get(Opcode), RI.getSubReg(DestReg, SubIdx));
419 Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
422 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
426 unsigned SIInstrInfo::commuteOpcode(unsigned Opcode) const {
429 // Try to map original to commuted opcode
430 if ((NewOpc = AMDGPU::getCommuteRev(Opcode)) != -1)
433 // Try to map commuted to original opcode
434 if ((NewOpc = AMDGPU::getCommuteOrig(Opcode)) != -1)
440 static bool shouldTryToSpillVGPRs(MachineFunction *MF) {
442 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
443 const TargetMachine &TM = MF->getTarget();
445 // FIXME: Even though it can cause problems, we need to enable
446 // spilling at -O0, since the fast register allocator always
447 // spills registers that are live at the end of blocks.
448 return MFI->getShaderType() == ShaderType::COMPUTE &&
449 TM.getOptLevel() == CodeGenOpt::None;
453 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
454 MachineBasicBlock::iterator MI,
455 unsigned SrcReg, bool isKill,
457 const TargetRegisterClass *RC,
458 const TargetRegisterInfo *TRI) const {
459 MachineFunction *MF = MBB.getParent();
460 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
461 DebugLoc DL = MBB.findDebugLoc(MI);
464 if (RI.isSGPRClass(RC)) {
465 // We are only allowed to create one new instruction when spilling
466 // registers, so we need to use pseudo instruction for spilling
468 switch (RC->getSize() * 8) {
469 case 32: Opcode = AMDGPU::SI_SPILL_S32_SAVE; break;
470 case 64: Opcode = AMDGPU::SI_SPILL_S64_SAVE; break;
471 case 128: Opcode = AMDGPU::SI_SPILL_S128_SAVE; break;
472 case 256: Opcode = AMDGPU::SI_SPILL_S256_SAVE; break;
473 case 512: Opcode = AMDGPU::SI_SPILL_S512_SAVE; break;
475 } else if(shouldTryToSpillVGPRs(MF) && RI.hasVGPRs(RC)) {
476 switch(RC->getSize() * 8) {
477 case 32: Opcode = AMDGPU::SI_SPILL_V32_SAVE; break;
478 case 64: Opcode = AMDGPU::SI_SPILL_V64_SAVE; break;
479 case 96: Opcode = AMDGPU::SI_SPILL_V96_SAVE; break;
480 case 128: Opcode = AMDGPU::SI_SPILL_V128_SAVE; break;
481 case 256: Opcode = AMDGPU::SI_SPILL_V256_SAVE; break;
482 case 512: Opcode = AMDGPU::SI_SPILL_V512_SAVE; break;
487 FrameInfo->setObjectAlignment(FrameIndex, 4);
488 BuildMI(MBB, MI, DL, get(Opcode))
490 .addFrameIndex(FrameIndex);
492 LLVMContext &Ctx = MF->getFunction()->getContext();
493 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Do not know how to"
495 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), AMDGPU::VGPR0)
500 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
501 MachineBasicBlock::iterator MI,
502 unsigned DestReg, int FrameIndex,
503 const TargetRegisterClass *RC,
504 const TargetRegisterInfo *TRI) const {
505 MachineFunction *MF = MBB.getParent();
506 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
507 DebugLoc DL = MBB.findDebugLoc(MI);
510 if (RI.isSGPRClass(RC)){
511 switch(RC->getSize() * 8) {
512 case 32: Opcode = AMDGPU::SI_SPILL_S32_RESTORE; break;
513 case 64: Opcode = AMDGPU::SI_SPILL_S64_RESTORE; break;
514 case 128: Opcode = AMDGPU::SI_SPILL_S128_RESTORE; break;
515 case 256: Opcode = AMDGPU::SI_SPILL_S256_RESTORE; break;
516 case 512: Opcode = AMDGPU::SI_SPILL_S512_RESTORE; break;
518 } else if(shouldTryToSpillVGPRs(MF) && RI.hasVGPRs(RC)) {
519 switch(RC->getSize() * 8) {
520 case 32: Opcode = AMDGPU::SI_SPILL_V32_RESTORE; break;
521 case 64: Opcode = AMDGPU::SI_SPILL_V64_RESTORE; break;
522 case 96: Opcode = AMDGPU::SI_SPILL_V96_RESTORE; break;
523 case 128: Opcode = AMDGPU::SI_SPILL_V128_RESTORE; break;
524 case 256: Opcode = AMDGPU::SI_SPILL_V256_RESTORE; break;
525 case 512: Opcode = AMDGPU::SI_SPILL_V512_RESTORE; break;
530 FrameInfo->setObjectAlignment(FrameIndex, 4);
531 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
532 .addFrameIndex(FrameIndex);
534 LLVMContext &Ctx = MF->getFunction()->getContext();
535 Ctx.emitError("SIInstrInfo::loadRegFromStackSlot - Do not know how to"
536 " restore register");
537 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
538 .addReg(AMDGPU::VGPR0);
542 /// \param @Offset Offset in bytes of the FrameIndex being spilled
543 unsigned SIInstrInfo::calculateLDSSpillAddress(MachineBasicBlock &MBB,
544 MachineBasicBlock::iterator MI,
545 RegScavenger *RS, unsigned TmpReg,
546 unsigned FrameOffset,
547 unsigned Size) const {
548 MachineFunction *MF = MBB.getParent();
549 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
550 const AMDGPUSubtarget &ST = MF->getTarget().getSubtarget<AMDGPUSubtarget>();
551 const SIRegisterInfo *TRI =
552 static_cast<const SIRegisterInfo*>(ST.getRegisterInfo());
553 DebugLoc DL = MBB.findDebugLoc(MI);
554 unsigned WorkGroupSize = MFI->getMaximumWorkGroupSize(*MF);
555 unsigned WavefrontSize = ST.getWavefrontSize();
557 unsigned TIDReg = MFI->getTIDReg();
558 if (!MFI->hasCalculatedTID()) {
559 MachineBasicBlock &Entry = MBB.getParent()->front();
560 MachineBasicBlock::iterator Insert = Entry.front();
561 DebugLoc DL = Insert->getDebugLoc();
563 TIDReg = RI.findUnusedVGPR(MF->getRegInfo());
564 if (TIDReg == AMDGPU::NoRegister)
568 if (MFI->getShaderType() == ShaderType::COMPUTE &&
569 WorkGroupSize > WavefrontSize) {
571 unsigned TIDIGXReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_X);
572 unsigned TIDIGYReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Y);
573 unsigned TIDIGZReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Z);
574 unsigned InputPtrReg =
575 TRI->getPreloadedValue(*MF, SIRegisterInfo::INPUT_PTR);
576 static const unsigned TIDIGRegs[3] = {
577 TIDIGXReg, TIDIGYReg, TIDIGZReg
579 for (unsigned Reg : TIDIGRegs) {
580 if (!Entry.isLiveIn(Reg))
581 Entry.addLiveIn(Reg);
584 RS->enterBasicBlock(&Entry);
585 unsigned STmp0 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
586 unsigned STmp1 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
587 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp0)
589 .addImm(SI::KernelInputOffsets::NGROUPS_Z);
590 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp1)
592 .addImm(SI::KernelInputOffsets::NGROUPS_Y);
594 // NGROUPS.X * NGROUPS.Y
595 BuildMI(Entry, Insert, DL, get(AMDGPU::S_MUL_I32), STmp1)
598 // (NGROUPS.X * NGROUPS.Y) * TIDIG.X
599 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MUL_U32_U24_e32), TIDReg)
602 // NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)
603 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MAD_U32_U24), TIDReg)
607 // (NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)) + TIDIG.Z
608 BuildMI(Entry, Insert, DL, get(AMDGPU::V_ADD_I32_e32), TIDReg)
613 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_LO_U32_B32_e64),
618 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_HI_U32_B32_e32),
624 BuildMI(Entry, Insert, DL, get(AMDGPU::V_LSHLREV_B32_e32),
628 MFI->setTIDReg(TIDReg);
631 // Add FrameIndex to LDS offset
632 unsigned LDSOffset = MFI->LDSSize + (FrameOffset * WorkGroupSize);
633 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_I32_e32), TmpReg)
640 void SIInstrInfo::insertNOPs(MachineBasicBlock::iterator MI,
649 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(AMDGPU::S_NOP))
654 bool SIInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
655 MachineBasicBlock &MBB = *MI->getParent();
656 DebugLoc DL = MBB.findDebugLoc(MI);
657 switch (MI->getOpcode()) {
658 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
660 case AMDGPU::SI_CONSTDATA_PTR: {
661 unsigned Reg = MI->getOperand(0).getReg();
662 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
663 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
665 BuildMI(MBB, MI, DL, get(AMDGPU::S_GETPC_B64), Reg);
667 // Add 32-bit offset from this instruction to the start of the constant data.
668 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADD_U32), RegLo)
670 .addTargetIndex(AMDGPU::TI_CONSTDATA_START)
671 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit);
672 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADDC_U32), RegHi)
675 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit)
676 .addReg(AMDGPU::SCC, RegState::Implicit);
677 MI->eraseFromParent();
680 case AMDGPU::SGPR_USE:
681 // This is just a placeholder for register allocation.
682 MI->eraseFromParent();
688 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
690 if (MI->getNumOperands() < 3)
693 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
694 AMDGPU::OpName::src0);
695 assert(Src0Idx != -1 && "Should always have src0 operand");
697 MachineOperand &Src0 = MI->getOperand(Src0Idx);
701 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
702 AMDGPU::OpName::src1);
706 MachineOperand &Src1 = MI->getOperand(Src1Idx);
708 // Make sure it s legal to commute operands for VOP2.
709 if (isVOP2(MI->getOpcode()) &&
710 (!isOperandLegal(MI, Src0Idx, &Src1) ||
711 !isOperandLegal(MI, Src1Idx, &Src0)))
715 // Allow commuting instructions with Imm or FPImm operands.
716 if (NewMI || (!Src1.isImm() && !Src1.isFPImm()) ||
717 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
721 // XXX: Commute VOP3 instructions with abs and neg set .
722 const MachineOperand *Src0Mods = getNamedOperand(*MI,
723 AMDGPU::OpName::src0_modifiers);
724 const MachineOperand *Src1Mods = getNamedOperand(*MI,
725 AMDGPU::OpName::src1_modifiers);
726 const MachineOperand *Src2Mods = getNamedOperand(*MI,
727 AMDGPU::OpName::src2_modifiers);
729 if ((Src0Mods && Src0Mods->getImm()) ||
730 (Src1Mods && Src1Mods->getImm()) ||
731 (Src2Mods && Src2Mods->getImm()))
734 unsigned Reg = Src0.getReg();
735 unsigned SubReg = Src0.getSubReg();
737 Src0.ChangeToImmediate(Src1.getImm());
738 else if (Src1.isFPImm())
739 Src0.ChangeToFPImmediate(Src1.getFPImm());
741 llvm_unreachable("Should only have immediates");
743 Src1.ChangeToRegister(Reg, false);
744 Src1.setSubReg(SubReg);
746 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
750 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
755 // This needs to be implemented because the source modifiers may be inserted
756 // between the true commutable operands, and the base
757 // TargetInstrInfo::commuteInstruction uses it.
758 bool SIInstrInfo::findCommutedOpIndices(MachineInstr *MI,
760 unsigned &SrcOpIdx2) const {
761 const MCInstrDesc &MCID = MI->getDesc();
762 if (!MCID.isCommutable())
765 unsigned Opc = MI->getOpcode();
766 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
770 // FIXME: Workaround TargetInstrInfo::commuteInstruction asserting on
772 if (!MI->getOperand(Src0Idx).isReg())
775 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
779 if (!MI->getOperand(Src1Idx).isReg())
787 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
788 MachineBasicBlock::iterator I,
790 unsigned SrcReg) const {
791 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
792 DstReg) .addReg(SrcReg);
795 bool SIInstrInfo::isMov(unsigned Opcode) const {
797 default: return false;
798 case AMDGPU::S_MOV_B32:
799 case AMDGPU::S_MOV_B64:
800 case AMDGPU::V_MOV_B32_e32:
801 case AMDGPU::V_MOV_B32_e64:
807 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
808 return RC != &AMDGPU::EXECRegRegClass;
812 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
813 AliasAnalysis *AA) const {
814 switch(MI->getOpcode()) {
815 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
816 case AMDGPU::S_MOV_B32:
817 case AMDGPU::S_MOV_B64:
818 case AMDGPU::V_MOV_B32_e32:
819 return MI->getOperand(1).isImm();
825 // Helper function generated by tablegen. We are wrapping this with
826 // an SIInstrInfo function that returns bool rather than int.
827 int isDS(uint16_t Opcode);
831 bool SIInstrInfo::isDS(uint16_t Opcode) const {
832 return ::AMDGPU::isDS(Opcode) != -1;
835 bool SIInstrInfo::isMIMG(uint16_t Opcode) const {
836 return get(Opcode).TSFlags & SIInstrFlags::MIMG;
839 bool SIInstrInfo::isSMRD(uint16_t Opcode) const {
840 return get(Opcode).TSFlags & SIInstrFlags::SMRD;
843 bool SIInstrInfo::isMUBUF(uint16_t Opcode) const {
844 return get(Opcode).TSFlags & SIInstrFlags::MUBUF;
847 bool SIInstrInfo::isMTBUF(uint16_t Opcode) const {
848 return get(Opcode).TSFlags & SIInstrFlags::MTBUF;
851 bool SIInstrInfo::isFLAT(uint16_t Opcode) const {
852 return get(Opcode).TSFlags & SIInstrFlags::FLAT;
855 bool SIInstrInfo::isVOP1(uint16_t Opcode) const {
856 return get(Opcode).TSFlags & SIInstrFlags::VOP1;
859 bool SIInstrInfo::isVOP2(uint16_t Opcode) const {
860 return get(Opcode).TSFlags & SIInstrFlags::VOP2;
863 bool SIInstrInfo::isVOP3(uint16_t Opcode) const {
864 return get(Opcode).TSFlags & SIInstrFlags::VOP3;
867 bool SIInstrInfo::isVOPC(uint16_t Opcode) const {
868 return get(Opcode).TSFlags & SIInstrFlags::VOPC;
871 bool SIInstrInfo::isSALUInstr(const MachineInstr &MI) const {
872 return get(MI.getOpcode()).TSFlags & SIInstrFlags::SALU;
875 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
876 int32_t Val = Imm.getSExtValue();
877 if (Val >= -16 && Val <= 64)
880 // The actual type of the operand does not seem to matter as long
881 // as the bits match one of the inline immediate values. For example:
883 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
884 // so it is a legal inline immediate.
886 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
887 // floating-point, so it is a legal inline immediate.
889 return (APInt::floatToBits(0.0f) == Imm) ||
890 (APInt::floatToBits(1.0f) == Imm) ||
891 (APInt::floatToBits(-1.0f) == Imm) ||
892 (APInt::floatToBits(0.5f) == Imm) ||
893 (APInt::floatToBits(-0.5f) == Imm) ||
894 (APInt::floatToBits(2.0f) == Imm) ||
895 (APInt::floatToBits(-2.0f) == Imm) ||
896 (APInt::floatToBits(4.0f) == Imm) ||
897 (APInt::floatToBits(-4.0f) == Imm);
900 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO) const {
902 return isInlineConstant(APInt(32, MO.getImm(), true));
905 APFloat FpImm = MO.getFPImm()->getValueAPF();
906 return isInlineConstant(FpImm.bitcastToAPInt());
912 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO) const {
913 return (MO.isImm() || MO.isFPImm()) && !isInlineConstant(MO);
916 static bool compareMachineOp(const MachineOperand &Op0,
917 const MachineOperand &Op1) {
918 if (Op0.getType() != Op1.getType())
921 switch (Op0.getType()) {
922 case MachineOperand::MO_Register:
923 return Op0.getReg() == Op1.getReg();
924 case MachineOperand::MO_Immediate:
925 return Op0.getImm() == Op1.getImm();
926 case MachineOperand::MO_FPImmediate:
927 return Op0.getFPImm() == Op1.getFPImm();
929 llvm_unreachable("Didn't expect to be comparing these operand types");
933 bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
934 const MachineOperand &MO) const {
935 const MCOperandInfo &OpInfo = get(MI->getOpcode()).OpInfo[OpNo];
937 assert(MO.isImm() || MO.isFPImm() || MO.isTargetIndex() || MO.isFI());
939 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
942 if (OpInfo.RegClass < 0)
945 if (isLiteralConstant(MO))
946 return RI.regClassCanUseLiteralConstant(OpInfo.RegClass);
948 return RI.regClassCanUseInlineConstant(OpInfo.RegClass);
951 bool SIInstrInfo::canFoldOffset(unsigned OffsetSize, unsigned AS) {
953 case AMDGPUAS::GLOBAL_ADDRESS: {
954 // MUBUF instructions a 12-bit offset in bytes.
955 return isUInt<12>(OffsetSize);
957 case AMDGPUAS::CONSTANT_ADDRESS: {
958 // SMRD instructions have an 8-bit offset in dwords.
959 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
961 case AMDGPUAS::LOCAL_ADDRESS:
962 case AMDGPUAS::REGION_ADDRESS: {
963 // The single offset versions have a 16-bit offset in bytes.
964 return isUInt<16>(OffsetSize);
966 case AMDGPUAS::PRIVATE_ADDRESS:
967 // Indirect register addressing does not use any offsets.
973 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
974 return AMDGPU::getVOPe32(Opcode) != -1;
977 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
978 // The src0_modifier operand is present on all instructions
979 // that have modifiers.
981 return AMDGPU::getNamedOperandIdx(Opcode,
982 AMDGPU::OpName::src0_modifiers) != -1;
985 bool SIInstrInfo::usesConstantBus(const MachineRegisterInfo &MRI,
986 const MachineOperand &MO) const {
987 // Literal constants use the constant bus.
988 if (isLiteralConstant(MO))
991 if (!MO.isReg() || !MO.isUse())
994 if (TargetRegisterInfo::isVirtualRegister(MO.getReg()))
995 return RI.isSGPRClass(MRI.getRegClass(MO.getReg()));
997 // FLAT_SCR is just an SGPR pair.
998 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
1001 // EXEC register uses the constant bus.
1002 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
1005 // SGPRs use the constant bus
1006 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
1007 (!MO.isImplicit() &&
1008 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
1009 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
1016 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
1017 StringRef &ErrInfo) const {
1018 uint16_t Opcode = MI->getOpcode();
1019 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1020 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
1021 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
1022 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
1024 // Make sure the number of operands is correct.
1025 const MCInstrDesc &Desc = get(Opcode);
1026 if (!Desc.isVariadic() &&
1027 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
1028 ErrInfo = "Instruction has wrong number of operands.";
1032 // Make sure the register classes are correct
1033 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
1034 switch (Desc.OpInfo[i].OperandType) {
1035 case MCOI::OPERAND_REGISTER: {
1036 if ((MI->getOperand(i).isImm() || MI->getOperand(i).isFPImm()) &&
1037 !isImmOperandLegal(MI, i, MI->getOperand(i))) {
1038 ErrInfo = "Illegal immediate value for operand.";
1043 case MCOI::OPERAND_IMMEDIATE:
1044 // Check if this operand is an immediate.
1045 // FrameIndex operands will be replaced by immediates, so they are
1047 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFPImm() &&
1048 !MI->getOperand(i).isFI()) {
1049 ErrInfo = "Expected immediate, but got non-immediate";
1057 if (!MI->getOperand(i).isReg())
1060 int RegClass = Desc.OpInfo[i].RegClass;
1061 if (RegClass != -1) {
1062 unsigned Reg = MI->getOperand(i).getReg();
1063 if (TargetRegisterInfo::isVirtualRegister(Reg))
1066 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
1067 if (!RC->contains(Reg)) {
1068 ErrInfo = "Operand has incorrect register class.";
1076 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
1077 unsigned ConstantBusCount = 0;
1078 unsigned SGPRUsed = AMDGPU::NoRegister;
1079 for (int i = 0, e = MI->getNumOperands(); i != e; ++i) {
1080 const MachineOperand &MO = MI->getOperand(i);
1081 if (usesConstantBus(MRI, MO)) {
1083 if (MO.getReg() != SGPRUsed)
1085 SGPRUsed = MO.getReg();
1091 if (ConstantBusCount > 1) {
1092 ErrInfo = "VOP* instruction uses the constant bus more than once";
1097 // Verify SRC1 for VOP2 and VOPC
1098 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
1099 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1100 if (Src1.isImm() || Src1.isFPImm()) {
1101 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
1107 if (isVOP3(Opcode)) {
1108 if (Src0Idx != -1 && isLiteralConstant(MI->getOperand(Src0Idx))) {
1109 ErrInfo = "VOP3 src0 cannot be a literal constant.";
1112 if (Src1Idx != -1 && isLiteralConstant(MI->getOperand(Src1Idx))) {
1113 ErrInfo = "VOP3 src1 cannot be a literal constant.";
1116 if (Src2Idx != -1 && isLiteralConstant(MI->getOperand(Src2Idx))) {
1117 ErrInfo = "VOP3 src2 cannot be a literal constant.";
1122 // Verify misc. restrictions on specific instructions.
1123 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
1124 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
1125 const MachineOperand &Src0 = MI->getOperand(Src0Idx);
1126 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1127 const MachineOperand &Src2 = MI->getOperand(Src2Idx);
1128 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
1129 if (!compareMachineOp(Src0, Src1) &&
1130 !compareMachineOp(Src0, Src2)) {
1131 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
1140 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
1141 switch (MI.getOpcode()) {
1142 default: return AMDGPU::INSTRUCTION_LIST_END;
1143 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
1144 case AMDGPU::COPY: return AMDGPU::COPY;
1145 case AMDGPU::PHI: return AMDGPU::PHI;
1146 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
1147 case AMDGPU::S_MOV_B32:
1148 return MI.getOperand(1).isReg() ?
1149 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
1150 case AMDGPU::S_ADD_I32:
1151 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
1152 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
1153 case AMDGPU::S_SUB_I32:
1154 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
1155 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
1156 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
1157 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
1158 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
1159 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
1160 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
1161 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
1162 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
1163 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
1164 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
1165 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
1166 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
1167 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
1168 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
1169 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
1170 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
1171 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
1172 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
1173 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
1174 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
1175 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
1176 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
1177 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
1178 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
1179 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
1180 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
1181 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
1182 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
1183 case AMDGPU::S_LOAD_DWORD_IMM:
1184 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1185 case AMDGPU::S_LOAD_DWORDX2_IMM:
1186 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1187 case AMDGPU::S_LOAD_DWORDX4_IMM:
1188 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1189 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e32;
1190 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
1191 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
1195 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
1196 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
1199 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
1200 unsigned OpNo) const {
1201 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1202 const MCInstrDesc &Desc = get(MI.getOpcode());
1203 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
1204 Desc.OpInfo[OpNo].RegClass == -1)
1205 return MRI.getRegClass(MI.getOperand(OpNo).getReg());
1207 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
1208 return RI.getRegClass(RCID);
1211 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
1212 switch (MI.getOpcode()) {
1214 case AMDGPU::REG_SEQUENCE:
1216 case AMDGPU::INSERT_SUBREG:
1217 return RI.hasVGPRs(getOpRegClass(MI, 0));
1219 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
1223 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
1224 MachineBasicBlock::iterator I = MI;
1225 MachineBasicBlock *MBB = MI->getParent();
1226 MachineOperand &MO = MI->getOperand(OpIdx);
1227 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1228 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
1229 const TargetRegisterClass *RC = RI.getRegClass(RCID);
1230 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1232 Opcode = AMDGPU::COPY;
1233 else if (RI.isSGPRClass(RC))
1234 Opcode = AMDGPU::S_MOV_B32;
1237 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
1238 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC))
1239 VRC = &AMDGPU::VReg_64RegClass;
1241 VRC = &AMDGPU::VReg_32RegClass;
1243 unsigned Reg = MRI.createVirtualRegister(VRC);
1244 DebugLoc DL = MBB->findDebugLoc(I);
1245 BuildMI(*MI->getParent(), I, DL, get(Opcode), Reg)
1247 MO.ChangeToRegister(Reg, false);
1250 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
1251 MachineRegisterInfo &MRI,
1252 MachineOperand &SuperReg,
1253 const TargetRegisterClass *SuperRC,
1255 const TargetRegisterClass *SubRC)
1257 assert(SuperReg.isReg());
1259 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
1260 unsigned SubReg = MRI.createVirtualRegister(SubRC);
1262 // Just in case the super register is itself a sub-register, copy it to a new
1263 // value so we don't need to worry about merging its subreg index with the
1264 // SubIdx passed to this function. The register coalescer should be able to
1265 // eliminate this extra copy.
1266 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1268 .addOperand(SuperReg);
1270 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1272 .addReg(NewSuperReg, 0, SubIdx);
1276 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
1277 MachineBasicBlock::iterator MII,
1278 MachineRegisterInfo &MRI,
1280 const TargetRegisterClass *SuperRC,
1282 const TargetRegisterClass *SubRC) const {
1284 // XXX - Is there a better way to do this?
1285 if (SubIdx == AMDGPU::sub0)
1286 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
1287 if (SubIdx == AMDGPU::sub1)
1288 return MachineOperand::CreateImm(Op.getImm() >> 32);
1290 llvm_unreachable("Unhandled register index for immediate");
1293 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
1295 return MachineOperand::CreateReg(SubReg, false);
1298 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
1299 MachineBasicBlock::iterator MI,
1300 MachineRegisterInfo &MRI,
1301 const TargetRegisterClass *RC,
1302 const MachineOperand &Op) const {
1303 MachineBasicBlock *MBB = MI->getParent();
1304 DebugLoc DL = MI->getDebugLoc();
1305 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1306 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1307 unsigned Dst = MRI.createVirtualRegister(RC);
1309 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1311 .addImm(Op.getImm() & 0xFFFFFFFF);
1312 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1314 .addImm(Op.getImm() >> 32);
1316 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
1318 .addImm(AMDGPU::sub0)
1320 .addImm(AMDGPU::sub1);
1322 Worklist.push_back(Lo);
1323 Worklist.push_back(Hi);
1328 bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
1329 const MachineOperand *MO) const {
1330 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1331 const MCInstrDesc &InstDesc = get(MI->getOpcode());
1332 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
1333 const TargetRegisterClass *DefinedRC =
1334 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
1336 MO = &MI->getOperand(OpIdx);
1338 if (usesConstantBus(MRI, *MO)) {
1340 MO->isReg() ? MO->getReg() : (unsigned)AMDGPU::NoRegister;
1341 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1344 if (usesConstantBus(MRI, MI->getOperand(i)) &&
1345 MI->getOperand(i).isReg() && MI->getOperand(i).getReg() != SGPRUsed) {
1353 const TargetRegisterClass *RC = MRI.getRegClass(MO->getReg());
1354 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass));
1358 // Handle non-register types that are treated like immediates.
1359 assert(MO->isImm() || MO->isFPImm() || MO->isTargetIndex() || MO->isFI());
1362 // This operand expects an immediate.
1366 return isImmOperandLegal(MI, OpIdx, *MO);
1369 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
1370 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1372 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1373 AMDGPU::OpName::src0);
1374 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1375 AMDGPU::OpName::src1);
1376 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1377 AMDGPU::OpName::src2);
1380 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
1382 if (!isOperandLegal(MI, Src0Idx))
1383 legalizeOpWithMove(MI, Src0Idx);
1386 if (isOperandLegal(MI, Src1Idx))
1389 // Usually src0 of VOP2 instructions allow more types of inputs
1390 // than src1, so try to commute the instruction to decrease our
1391 // chances of having to insert a MOV instruction to legalize src1.
1392 if (MI->isCommutable()) {
1393 if (commuteInstruction(MI))
1394 // If we are successful in commuting, then we know MI is legal, so
1399 legalizeOpWithMove(MI, Src1Idx);
1403 // XXX - Do any VOP3 instructions read VCC?
1405 if (isVOP3(MI->getOpcode())) {
1406 int VOP3Idx[3] = { Src0Idx, Src1Idx, Src2Idx };
1408 // Find the one SGPR operand we are allowed to use.
1409 unsigned SGPRReg = findUsedSGPR(MI, VOP3Idx);
1411 for (unsigned i = 0; i < 3; ++i) {
1412 int Idx = VOP3Idx[i];
1415 MachineOperand &MO = MI->getOperand(Idx);
1418 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
1419 continue; // VGPRs are legal
1421 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
1423 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
1424 SGPRReg = MO.getReg();
1425 // We can use one SGPR in each VOP3 instruction.
1428 } else if (!isLiteralConstant(MO)) {
1429 // If it is not a register and not a literal constant, then it must be
1430 // an inline constant which is always legal.
1433 // If we make it this far, then the operand is not legal and we must
1435 legalizeOpWithMove(MI, Idx);
1439 // Legalize REG_SEQUENCE and PHI
1440 // The register class of the operands much be the same type as the register
1441 // class of the output.
1442 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
1443 MI->getOpcode() == AMDGPU::PHI) {
1444 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
1445 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1446 if (!MI->getOperand(i).isReg() ||
1447 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1449 const TargetRegisterClass *OpRC =
1450 MRI.getRegClass(MI->getOperand(i).getReg());
1451 if (RI.hasVGPRs(OpRC)) {
1458 // If any of the operands are VGPR registers, then they all most be
1459 // otherwise we will create illegal VGPR->SGPR copies when legalizing
1461 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
1464 VRC = RI.getEquivalentVGPRClass(SRC);
1471 // Update all the operands so they have the same type.
1472 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1473 if (!MI->getOperand(i).isReg() ||
1474 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1476 unsigned DstReg = MRI.createVirtualRegister(RC);
1477 MachineBasicBlock *InsertBB;
1478 MachineBasicBlock::iterator Insert;
1479 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
1480 InsertBB = MI->getParent();
1483 // MI is a PHI instruction.
1484 InsertBB = MI->getOperand(i + 1).getMBB();
1485 Insert = InsertBB->getFirstTerminator();
1487 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
1488 get(AMDGPU::COPY), DstReg)
1489 .addOperand(MI->getOperand(i));
1490 MI->getOperand(i).setReg(DstReg);
1494 // Legalize INSERT_SUBREG
1495 // src0 must have the same register class as dst
1496 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
1497 unsigned Dst = MI->getOperand(0).getReg();
1498 unsigned Src0 = MI->getOperand(1).getReg();
1499 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
1500 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
1501 if (DstRC != Src0RC) {
1502 MachineBasicBlock &MBB = *MI->getParent();
1503 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
1504 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
1506 MI->getOperand(1).setReg(NewSrc0);
1511 // Legalize MUBUF* instructions
1512 // FIXME: If we start using the non-addr64 instructions for compute, we
1513 // may need to legalize them here.
1515 AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::srsrc);
1516 if (SRsrcIdx != -1) {
1517 // We have an MUBUF instruction
1518 MachineOperand *SRsrc = &MI->getOperand(SRsrcIdx);
1519 unsigned SRsrcRC = get(MI->getOpcode()).OpInfo[SRsrcIdx].RegClass;
1520 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
1521 RI.getRegClass(SRsrcRC))) {
1522 // The operands are legal.
1523 // FIXME: We may need to legalize operands besided srsrc.
1527 MachineBasicBlock &MBB = *MI->getParent();
1528 // Extract the the ptr from the resource descriptor.
1530 // SRsrcPtrLo = srsrc:sub0
1531 unsigned SRsrcPtrLo = buildExtractSubReg(MI, MRI, *SRsrc,
1532 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VReg_32RegClass);
1534 // SRsrcPtrHi = srsrc:sub1
1535 unsigned SRsrcPtrHi = buildExtractSubReg(MI, MRI, *SRsrc,
1536 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VReg_32RegClass);
1538 // Create an empty resource descriptor
1539 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1540 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1541 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1542 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1545 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1549 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1550 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1552 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1554 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1555 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1557 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1559 // NewSRsrc = {Zero64, SRsrcFormat}
1560 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1563 .addImm(AMDGPU::sub0_sub1)
1564 .addReg(SRsrcFormatLo)
1565 .addImm(AMDGPU::sub2)
1566 .addReg(SRsrcFormatHi)
1567 .addImm(AMDGPU::sub3);
1569 MachineOperand *VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1570 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
1571 unsigned NewVAddrLo;
1572 unsigned NewVAddrHi;
1574 // This is already an ADDR64 instruction so we need to add the pointer
1575 // extracted from the resource descriptor to the current value of VAddr.
1576 NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1577 NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1579 // NewVaddrLo = SRsrcPtrLo + VAddr:sub0
1580 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1583 .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
1584 .addReg(AMDGPU::VCC, RegState::ImplicitDefine);
1586 // NewVaddrHi = SRsrcPtrHi + VAddr:sub1
1587 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1590 .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
1591 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1592 .addReg(AMDGPU::VCC, RegState::Implicit);
1595 // This instructions is the _OFFSET variant, so we need to convert it to
1597 MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
1598 MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
1599 MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
1600 assert(SOffset->isImm() && SOffset->getImm() == 0 && "Legalizing MUBUF "
1601 "with non-zero soffset is not implemented");
1604 // Create the new instruction.
1605 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
1606 MachineInstr *Addr64 =
1607 BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
1610 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
1611 // This will be replaced later
1612 // with the new value of vaddr.
1613 .addOperand(*Offset);
1615 MI->removeFromParent();
1618 NewVAddrLo = SRsrcPtrLo;
1619 NewVAddrHi = SRsrcPtrHi;
1620 VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1621 SRsrc = getNamedOperand(*MI, AMDGPU::OpName::srsrc);
1624 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1625 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1628 .addImm(AMDGPU::sub0)
1630 .addImm(AMDGPU::sub1);
1633 // Update the instruction to use NewVaddr
1634 VAddr->setReg(NewVAddr);
1635 // Update the instruction to use NewSRsrc
1636 SRsrc->setReg(NewSRsrc);
1640 void SIInstrInfo::splitSMRD(MachineInstr *MI,
1641 const TargetRegisterClass *HalfRC,
1642 unsigned HalfImmOp, unsigned HalfSGPROp,
1643 MachineInstr *&Lo, MachineInstr *&Hi) const {
1645 DebugLoc DL = MI->getDebugLoc();
1646 MachineBasicBlock *MBB = MI->getParent();
1647 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1648 unsigned RegLo = MRI.createVirtualRegister(HalfRC);
1649 unsigned RegHi = MRI.createVirtualRegister(HalfRC);
1650 unsigned HalfSize = HalfRC->getSize();
1651 const MachineOperand *OffOp =
1652 getNamedOperand(*MI, AMDGPU::OpName::offset);
1653 const MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
1656 // Handle the _IMM variant
1657 unsigned LoOffset = OffOp->getImm();
1658 unsigned HiOffset = LoOffset + (HalfSize / 4);
1659 Lo = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegLo)
1663 if (!isUInt<8>(HiOffset)) {
1664 unsigned OffsetSGPR =
1665 MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1666 BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32), OffsetSGPR)
1667 .addImm(HiOffset << 2); // The immediate offset is in dwords,
1668 // but offset in register is in bytes.
1669 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
1671 .addReg(OffsetSGPR);
1673 Hi = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegHi)
1678 // Handle the _SGPR variant
1679 MachineOperand *SOff = getNamedOperand(*MI, AMDGPU::OpName::soff);
1680 Lo = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegLo)
1683 unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1684 BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
1687 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp))
1689 .addReg(OffsetSGPR);
1692 unsigned SubLo, SubHi;
1695 SubLo = AMDGPU::sub0;
1696 SubHi = AMDGPU::sub1;
1699 SubLo = AMDGPU::sub0_sub1;
1700 SubHi = AMDGPU::sub2_sub3;
1703 SubLo = AMDGPU::sub0_sub1_sub2_sub3;
1704 SubHi = AMDGPU::sub4_sub5_sub6_sub7;
1707 SubLo = AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
1708 SubHi = AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15;
1711 llvm_unreachable("Unhandled HalfSize");
1714 BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE))
1715 .addOperand(MI->getOperand(0))
1722 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
1723 MachineBasicBlock *MBB = MI->getParent();
1724 switch (MI->getOpcode()) {
1725 case AMDGPU::S_LOAD_DWORD_IMM:
1726 case AMDGPU::S_LOAD_DWORD_SGPR:
1727 case AMDGPU::S_LOAD_DWORDX2_IMM:
1728 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1729 case AMDGPU::S_LOAD_DWORDX4_IMM:
1730 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1731 unsigned NewOpcode = getVALUOp(*MI);
1735 if (MI->getOperand(2).isReg()) {
1736 RegOffset = MI->getOperand(2).getReg();
1739 assert(MI->getOperand(2).isImm());
1740 // SMRD instructions take a dword offsets and MUBUF instructions
1741 // take a byte offset.
1742 ImmOffset = MI->getOperand(2).getImm() << 2;
1743 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1744 if (isUInt<12>(ImmOffset)) {
1745 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1749 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1756 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1757 unsigned DWord0 = RegOffset;
1758 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1759 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1760 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1762 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
1764 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
1765 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1766 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
1767 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1768 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
1770 .addImm(AMDGPU::sub0)
1772 .addImm(AMDGPU::sub1)
1774 .addImm(AMDGPU::sub2)
1776 .addImm(AMDGPU::sub3);
1777 MI->setDesc(get(NewOpcode));
1778 if (MI->getOperand(2).isReg()) {
1779 MI->getOperand(2).setReg(MI->getOperand(1).getReg());
1781 MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
1783 MI->getOperand(1).setReg(SRsrc);
1784 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
1786 const TargetRegisterClass *NewDstRC =
1787 RI.getRegClass(get(NewOpcode).OpInfo[0].RegClass);
1789 unsigned DstReg = MI->getOperand(0).getReg();
1790 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1791 MRI.replaceRegWith(DstReg, NewDstReg);
1794 case AMDGPU::S_LOAD_DWORDX8_IMM:
1795 case AMDGPU::S_LOAD_DWORDX8_SGPR: {
1796 MachineInstr *Lo, *Hi;
1797 splitSMRD(MI, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM,
1798 AMDGPU::S_LOAD_DWORDX4_SGPR, Lo, Hi);
1799 MI->eraseFromParent();
1800 moveSMRDToVALU(Lo, MRI);
1801 moveSMRDToVALU(Hi, MRI);
1805 case AMDGPU::S_LOAD_DWORDX16_IMM:
1806 case AMDGPU::S_LOAD_DWORDX16_SGPR: {
1807 MachineInstr *Lo, *Hi;
1808 splitSMRD(MI, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM,
1809 AMDGPU::S_LOAD_DWORDX8_SGPR, Lo, Hi);
1810 MI->eraseFromParent();
1811 moveSMRDToVALU(Lo, MRI);
1812 moveSMRDToVALU(Hi, MRI);
1818 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
1819 SmallVector<MachineInstr *, 128> Worklist;
1820 Worklist.push_back(&TopInst);
1822 while (!Worklist.empty()) {
1823 MachineInstr *Inst = Worklist.pop_back_val();
1824 MachineBasicBlock *MBB = Inst->getParent();
1825 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1827 unsigned Opcode = Inst->getOpcode();
1828 unsigned NewOpcode = getVALUOp(*Inst);
1830 // Handle some special cases
1833 if (isSMRD(Inst->getOpcode())) {
1834 moveSMRDToVALU(Inst, MRI);
1837 case AMDGPU::S_MOV_B64: {
1838 DebugLoc DL = Inst->getDebugLoc();
1840 // If the source operand is a register we can replace this with a
1842 if (Inst->getOperand(1).isReg()) {
1843 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
1844 .addOperand(Inst->getOperand(0))
1845 .addOperand(Inst->getOperand(1));
1846 Worklist.push_back(Copy);
1848 // Otherwise, we need to split this into two movs, because there is
1849 // no 64-bit VALU move instruction.
1850 unsigned Reg = Inst->getOperand(0).getReg();
1851 unsigned Dst = split64BitImm(Worklist,
1854 MRI.getRegClass(Reg),
1855 Inst->getOperand(1));
1856 MRI.replaceRegWith(Reg, Dst);
1858 Inst->eraseFromParent();
1861 case AMDGPU::S_AND_B64:
1862 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
1863 Inst->eraseFromParent();
1866 case AMDGPU::S_OR_B64:
1867 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
1868 Inst->eraseFromParent();
1871 case AMDGPU::S_XOR_B64:
1872 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
1873 Inst->eraseFromParent();
1876 case AMDGPU::S_NOT_B64:
1877 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
1878 Inst->eraseFromParent();
1881 case AMDGPU::S_BCNT1_I32_B64:
1882 splitScalar64BitBCNT(Worklist, Inst);
1883 Inst->eraseFromParent();
1886 case AMDGPU::S_BFE_U64:
1887 case AMDGPU::S_BFE_I64:
1888 case AMDGPU::S_BFM_B64:
1889 llvm_unreachable("Moving this op to VALU not implemented");
1892 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
1893 // We cannot move this instruction to the VALU, so we should try to
1894 // legalize its operands instead.
1895 legalizeOperands(Inst);
1899 // Use the new VALU Opcode.
1900 const MCInstrDesc &NewDesc = get(NewOpcode);
1901 Inst->setDesc(NewDesc);
1903 // Remove any references to SCC. Vector instructions can't read from it, and
1904 // We're just about to add the implicit use / defs of VCC, and we don't want
1906 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
1907 MachineOperand &Op = Inst->getOperand(i);
1908 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
1909 Inst->RemoveOperand(i);
1912 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
1913 // We are converting these to a BFE, so we need to add the missing
1914 // operands for the size and offset.
1915 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
1916 Inst->addOperand(MachineOperand::CreateImm(0));
1917 Inst->addOperand(MachineOperand::CreateImm(Size));
1919 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
1920 // The VALU version adds the second operand to the result, so insert an
1922 Inst->addOperand(MachineOperand::CreateImm(0));
1925 addDescImplicitUseDef(NewDesc, Inst);
1927 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
1928 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
1929 // If we need to move this to VGPRs, we need to unpack the second operand
1930 // back into the 2 separate ones for bit offset and width.
1931 assert(OffsetWidthOp.isImm() &&
1932 "Scalar BFE is only implemented for constant width and offset");
1933 uint32_t Imm = OffsetWidthOp.getImm();
1935 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
1936 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
1937 Inst->RemoveOperand(2); // Remove old immediate.
1938 Inst->addOperand(MachineOperand::CreateImm(Offset));
1939 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
1942 // Update the destination register class.
1944 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
1947 // For target instructions, getOpRegClass just returns the virtual
1948 // register class associated with the operand, so we need to find an
1949 // equivalent VGPR register class in order to move the instruction to the
1953 case AMDGPU::REG_SEQUENCE:
1954 case AMDGPU::INSERT_SUBREG:
1955 if (RI.hasVGPRs(NewDstRC))
1957 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
1965 unsigned DstReg = Inst->getOperand(0).getReg();
1966 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1967 MRI.replaceRegWith(DstReg, NewDstReg);
1969 // Legalize the operands
1970 legalizeOperands(Inst);
1972 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
1973 E = MRI.use_end(); I != E; ++I) {
1974 MachineInstr &UseMI = *I->getParent();
1975 if (!canReadVGPR(UseMI, I.getOperandNo())) {
1976 Worklist.push_back(&UseMI);
1982 //===----------------------------------------------------------------------===//
1983 // Indirect addressing callbacks
1984 //===----------------------------------------------------------------------===//
1986 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
1987 unsigned Channel) const {
1988 assert(Channel == 0);
1992 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
1993 return &AMDGPU::VReg_32RegClass;
1996 void SIInstrInfo::splitScalar64BitUnaryOp(
1997 SmallVectorImpl<MachineInstr *> &Worklist,
1999 unsigned Opcode) const {
2000 MachineBasicBlock &MBB = *Inst->getParent();
2001 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2003 MachineOperand &Dest = Inst->getOperand(0);
2004 MachineOperand &Src0 = Inst->getOperand(1);
2005 DebugLoc DL = Inst->getDebugLoc();
2007 MachineBasicBlock::iterator MII = Inst;
2009 const MCInstrDesc &InstDesc = get(Opcode);
2010 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2011 MRI.getRegClass(Src0.getReg()) :
2012 &AMDGPU::SGPR_32RegClass;
2014 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2016 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2017 AMDGPU::sub0, Src0SubRC);
2019 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2020 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2022 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2023 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2024 .addOperand(SrcReg0Sub0);
2026 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2027 AMDGPU::sub1, Src0SubRC);
2029 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2030 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2031 .addOperand(SrcReg0Sub1);
2033 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2034 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2036 .addImm(AMDGPU::sub0)
2038 .addImm(AMDGPU::sub1);
2040 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2042 // Try to legalize the operands in case we need to swap the order to keep it
2044 Worklist.push_back(LoHalf);
2045 Worklist.push_back(HiHalf);
2048 void SIInstrInfo::splitScalar64BitBinaryOp(
2049 SmallVectorImpl<MachineInstr *> &Worklist,
2051 unsigned Opcode) const {
2052 MachineBasicBlock &MBB = *Inst->getParent();
2053 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2055 MachineOperand &Dest = Inst->getOperand(0);
2056 MachineOperand &Src0 = Inst->getOperand(1);
2057 MachineOperand &Src1 = Inst->getOperand(2);
2058 DebugLoc DL = Inst->getDebugLoc();
2060 MachineBasicBlock::iterator MII = Inst;
2062 const MCInstrDesc &InstDesc = get(Opcode);
2063 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2064 MRI.getRegClass(Src0.getReg()) :
2065 &AMDGPU::SGPR_32RegClass;
2067 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2068 const TargetRegisterClass *Src1RC = Src1.isReg() ?
2069 MRI.getRegClass(Src1.getReg()) :
2070 &AMDGPU::SGPR_32RegClass;
2072 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
2074 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2075 AMDGPU::sub0, Src0SubRC);
2076 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2077 AMDGPU::sub0, Src1SubRC);
2079 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2080 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2082 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2083 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2084 .addOperand(SrcReg0Sub0)
2085 .addOperand(SrcReg1Sub0);
2087 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2088 AMDGPU::sub1, Src0SubRC);
2089 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2090 AMDGPU::sub1, Src1SubRC);
2092 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2093 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2094 .addOperand(SrcReg0Sub1)
2095 .addOperand(SrcReg1Sub1);
2097 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2098 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2100 .addImm(AMDGPU::sub0)
2102 .addImm(AMDGPU::sub1);
2104 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2106 // Try to legalize the operands in case we need to swap the order to keep it
2108 Worklist.push_back(LoHalf);
2109 Worklist.push_back(HiHalf);
2112 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
2113 MachineInstr *Inst) const {
2114 MachineBasicBlock &MBB = *Inst->getParent();
2115 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2117 MachineBasicBlock::iterator MII = Inst;
2118 DebugLoc DL = Inst->getDebugLoc();
2120 MachineOperand &Dest = Inst->getOperand(0);
2121 MachineOperand &Src = Inst->getOperand(1);
2123 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e32);
2124 const TargetRegisterClass *SrcRC = Src.isReg() ?
2125 MRI.getRegClass(Src.getReg()) :
2126 &AMDGPU::SGPR_32RegClass;
2128 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2129 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2131 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
2133 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2134 AMDGPU::sub0, SrcSubRC);
2135 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2136 AMDGPU::sub1, SrcSubRC);
2138 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
2139 .addOperand(SrcRegSub0)
2142 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
2143 .addOperand(SrcRegSub1)
2146 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2148 Worklist.push_back(First);
2149 Worklist.push_back(Second);
2152 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
2153 MachineInstr *Inst) const {
2154 // Add the implict and explicit register definitions.
2155 if (NewDesc.ImplicitUses) {
2156 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
2157 unsigned Reg = NewDesc.ImplicitUses[i];
2158 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
2162 if (NewDesc.ImplicitDefs) {
2163 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
2164 unsigned Reg = NewDesc.ImplicitDefs[i];
2165 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
2170 unsigned SIInstrInfo::findUsedSGPR(const MachineInstr *MI,
2171 int OpIndices[3]) const {
2172 const MCInstrDesc &Desc = get(MI->getOpcode());
2174 // Find the one SGPR operand we are allowed to use.
2175 unsigned SGPRReg = AMDGPU::NoRegister;
2177 // First we need to consider the instruction's operand requirements before
2178 // legalizing. Some operands are required to be SGPRs, such as implicit uses
2179 // of VCC, but we are still bound by the constant bus requirement to only use
2182 // If the operand's class is an SGPR, we can never move it.
2184 for (const MachineOperand &MO : MI->implicit_operands()) {
2185 // We only care about reads.
2189 if (MO.getReg() == AMDGPU::VCC)
2192 if (MO.getReg() == AMDGPU::FLAT_SCR)
2193 return AMDGPU::FLAT_SCR;
2196 unsigned UsedSGPRs[3] = { AMDGPU::NoRegister };
2197 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
2199 for (unsigned i = 0; i < 3; ++i) {
2200 int Idx = OpIndices[i];
2204 const MachineOperand &MO = MI->getOperand(Idx);
2205 if (RI.isSGPRClassID(Desc.OpInfo[Idx].RegClass))
2206 SGPRReg = MO.getReg();
2208 if (MO.isReg() && RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
2209 UsedSGPRs[i] = MO.getReg();
2212 if (SGPRReg != AMDGPU::NoRegister)
2215 // We don't have a required SGPR operand, so we have a bit more freedom in
2216 // selecting operands to move.
2218 // Try to select the most used SGPR. If an SGPR is equal to one of the
2219 // others, we choose that.
2222 // V_FMA_F32 v0, s0, s0, s0 -> No moves
2223 // V_FMA_F32 v0, s0, s1, s0 -> Move s1
2225 if (UsedSGPRs[0] != AMDGPU::NoRegister) {
2226 if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
2227 SGPRReg = UsedSGPRs[0];
2230 if (SGPRReg == AMDGPU::NoRegister && UsedSGPRs[1] != AMDGPU::NoRegister) {
2231 if (UsedSGPRs[1] == UsedSGPRs[2])
2232 SGPRReg = UsedSGPRs[1];
2238 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
2239 MachineBasicBlock *MBB,
2240 MachineBasicBlock::iterator I,
2242 unsigned Address, unsigned OffsetReg) const {
2243 const DebugLoc &DL = MBB->findDebugLoc(I);
2244 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
2245 getIndirectIndexBegin(*MBB->getParent()));
2247 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
2248 .addReg(IndirectBaseReg, RegState::Define)
2249 .addOperand(I->getOperand(0))
2250 .addReg(IndirectBaseReg)
2256 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
2257 MachineBasicBlock *MBB,
2258 MachineBasicBlock::iterator I,
2260 unsigned Address, unsigned OffsetReg) const {
2261 const DebugLoc &DL = MBB->findDebugLoc(I);
2262 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
2263 getIndirectIndexBegin(*MBB->getParent()));
2265 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
2266 .addOperand(I->getOperand(0))
2267 .addOperand(I->getOperand(1))
2268 .addReg(IndirectBaseReg)
2274 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
2275 const MachineFunction &MF) const {
2276 int End = getIndirectIndexEnd(MF);
2277 int Begin = getIndirectIndexBegin(MF);
2283 for (int Index = Begin; Index <= End; ++Index)
2284 Reserved.set(AMDGPU::VReg_32RegClass.getRegister(Index));
2286 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
2287 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
2289 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
2290 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
2292 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
2293 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
2295 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
2296 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
2298 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
2299 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));
2302 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
2303 unsigned OperandName) const {
2304 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
2308 return &MI.getOperand(Idx);