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 // XXX: Commute instructions with FPImm operands
716 if (NewMI || !Src1.isImm() ||
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();
736 Src0.ChangeToImmediate(Src1.getImm());
737 Src1.ChangeToRegister(Reg, false);
738 Src1.setSubReg(SubReg);
740 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
744 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
749 // This needs to be implemented because the source modifiers may be inserted
750 // between the true commutable operands, and the base
751 // TargetInstrInfo::commuteInstruction uses it.
752 bool SIInstrInfo::findCommutedOpIndices(MachineInstr *MI,
754 unsigned &SrcOpIdx2) const {
755 const MCInstrDesc &MCID = MI->getDesc();
756 if (!MCID.isCommutable())
759 unsigned Opc = MI->getOpcode();
760 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
764 // FIXME: Workaround TargetInstrInfo::commuteInstruction asserting on
766 if (!MI->getOperand(Src0Idx).isReg())
769 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
773 if (!MI->getOperand(Src1Idx).isReg())
781 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
782 MachineBasicBlock::iterator I,
784 unsigned SrcReg) const {
785 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
786 DstReg) .addReg(SrcReg);
789 bool SIInstrInfo::isMov(unsigned Opcode) const {
791 default: return false;
792 case AMDGPU::S_MOV_B32:
793 case AMDGPU::S_MOV_B64:
794 case AMDGPU::V_MOV_B32_e32:
795 case AMDGPU::V_MOV_B32_e64:
801 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
802 return RC != &AMDGPU::EXECRegRegClass;
806 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
807 AliasAnalysis *AA) const {
808 switch(MI->getOpcode()) {
809 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
810 case AMDGPU::S_MOV_B32:
811 case AMDGPU::S_MOV_B64:
812 case AMDGPU::V_MOV_B32_e32:
813 return MI->getOperand(1).isImm();
819 // Helper function generated by tablegen. We are wrapping this with
820 // an SIInstrInfo function that returns bool rather than int.
821 int isDS(uint16_t Opcode);
825 bool SIInstrInfo::isDS(uint16_t Opcode) const {
826 return ::AMDGPU::isDS(Opcode) != -1;
829 bool SIInstrInfo::isMIMG(uint16_t Opcode) const {
830 return get(Opcode).TSFlags & SIInstrFlags::MIMG;
833 bool SIInstrInfo::isSMRD(uint16_t Opcode) const {
834 return get(Opcode).TSFlags & SIInstrFlags::SMRD;
837 bool SIInstrInfo::isMUBUF(uint16_t Opcode) const {
838 return get(Opcode).TSFlags & SIInstrFlags::MUBUF;
841 bool SIInstrInfo::isMTBUF(uint16_t Opcode) const {
842 return get(Opcode).TSFlags & SIInstrFlags::MTBUF;
845 bool SIInstrInfo::isFLAT(uint16_t Opcode) const {
846 return get(Opcode).TSFlags & SIInstrFlags::FLAT;
849 bool SIInstrInfo::isVOP1(uint16_t Opcode) const {
850 return get(Opcode).TSFlags & SIInstrFlags::VOP1;
853 bool SIInstrInfo::isVOP2(uint16_t Opcode) const {
854 return get(Opcode).TSFlags & SIInstrFlags::VOP2;
857 bool SIInstrInfo::isVOP3(uint16_t Opcode) const {
858 return get(Opcode).TSFlags & SIInstrFlags::VOP3;
861 bool SIInstrInfo::isVOPC(uint16_t Opcode) const {
862 return get(Opcode).TSFlags & SIInstrFlags::VOPC;
865 bool SIInstrInfo::isSALUInstr(const MachineInstr &MI) const {
866 return get(MI.getOpcode()).TSFlags & SIInstrFlags::SALU;
869 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
870 int32_t Val = Imm.getSExtValue();
871 if (Val >= -16 && Val <= 64)
874 // The actual type of the operand does not seem to matter as long
875 // as the bits match one of the inline immediate values. For example:
877 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
878 // so it is a legal inline immediate.
880 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
881 // floating-point, so it is a legal inline immediate.
883 return (APInt::floatToBits(0.0f) == Imm) ||
884 (APInt::floatToBits(1.0f) == Imm) ||
885 (APInt::floatToBits(-1.0f) == Imm) ||
886 (APInt::floatToBits(0.5f) == Imm) ||
887 (APInt::floatToBits(-0.5f) == Imm) ||
888 (APInt::floatToBits(2.0f) == Imm) ||
889 (APInt::floatToBits(-2.0f) == Imm) ||
890 (APInt::floatToBits(4.0f) == Imm) ||
891 (APInt::floatToBits(-4.0f) == Imm);
894 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO) const {
896 return isInlineConstant(APInt(32, MO.getImm(), true));
899 APFloat FpImm = MO.getFPImm()->getValueAPF();
900 return isInlineConstant(FpImm.bitcastToAPInt());
906 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO) const {
907 return (MO.isImm() || MO.isFPImm()) && !isInlineConstant(MO);
910 static bool compareMachineOp(const MachineOperand &Op0,
911 const MachineOperand &Op1) {
912 if (Op0.getType() != Op1.getType())
915 switch (Op0.getType()) {
916 case MachineOperand::MO_Register:
917 return Op0.getReg() == Op1.getReg();
918 case MachineOperand::MO_Immediate:
919 return Op0.getImm() == Op1.getImm();
920 case MachineOperand::MO_FPImmediate:
921 return Op0.getFPImm() == Op1.getFPImm();
923 llvm_unreachable("Didn't expect to be comparing these operand types");
927 bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
928 const MachineOperand &MO) const {
929 const MCOperandInfo &OpInfo = get(MI->getOpcode()).OpInfo[OpNo];
931 assert(MO.isImm() || MO.isFPImm() || MO.isTargetIndex() || MO.isFI());
933 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
936 if (OpInfo.RegClass < 0)
939 if (isLiteralConstant(MO))
940 return RI.regClassCanUseLiteralConstant(OpInfo.RegClass);
942 return RI.regClassCanUseInlineConstant(OpInfo.RegClass);
945 bool SIInstrInfo::canFoldOffset(unsigned OffsetSize, unsigned AS) {
947 case AMDGPUAS::GLOBAL_ADDRESS: {
948 // MUBUF instructions a 12-bit offset in bytes.
949 return isUInt<12>(OffsetSize);
951 case AMDGPUAS::CONSTANT_ADDRESS: {
952 // SMRD instructions have an 8-bit offset in dwords.
953 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
955 case AMDGPUAS::LOCAL_ADDRESS:
956 case AMDGPUAS::REGION_ADDRESS: {
957 // The single offset versions have a 16-bit offset in bytes.
958 return isUInt<16>(OffsetSize);
960 case AMDGPUAS::PRIVATE_ADDRESS:
961 // Indirect register addressing does not use any offsets.
967 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
968 return AMDGPU::getVOPe32(Opcode) != -1;
971 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
972 // The src0_modifier operand is present on all instructions
973 // that have modifiers.
975 return AMDGPU::getNamedOperandIdx(Opcode,
976 AMDGPU::OpName::src0_modifiers) != -1;
979 bool SIInstrInfo::usesConstantBus(const MachineRegisterInfo &MRI,
980 const MachineOperand &MO) const {
981 // Literal constants use the constant bus.
982 if (isLiteralConstant(MO))
985 if (!MO.isReg() || !MO.isUse())
988 if (TargetRegisterInfo::isVirtualRegister(MO.getReg()))
989 return RI.isSGPRClass(MRI.getRegClass(MO.getReg()));
991 // FLAT_SCR is just an SGPR pair.
992 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
995 // EXEC register uses the constant bus.
996 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
999 // SGPRs use the constant bus
1000 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
1001 (!MO.isImplicit() &&
1002 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
1003 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
1010 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
1011 StringRef &ErrInfo) const {
1012 uint16_t Opcode = MI->getOpcode();
1013 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1014 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
1015 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
1016 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
1018 // Make sure the number of operands is correct.
1019 const MCInstrDesc &Desc = get(Opcode);
1020 if (!Desc.isVariadic() &&
1021 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
1022 ErrInfo = "Instruction has wrong number of operands.";
1026 // Make sure the register classes are correct
1027 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
1028 switch (Desc.OpInfo[i].OperandType) {
1029 case MCOI::OPERAND_REGISTER: {
1030 if ((MI->getOperand(i).isImm() || MI->getOperand(i).isFPImm()) &&
1031 !isImmOperandLegal(MI, i, MI->getOperand(i))) {
1032 ErrInfo = "Illegal immediate value for operand.";
1037 case MCOI::OPERAND_IMMEDIATE:
1038 // Check if this operand is an immediate.
1039 // FrameIndex operands will be replaced by immediates, so they are
1041 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFPImm() &&
1042 !MI->getOperand(i).isFI()) {
1043 ErrInfo = "Expected immediate, but got non-immediate";
1051 if (!MI->getOperand(i).isReg())
1054 int RegClass = Desc.OpInfo[i].RegClass;
1055 if (RegClass != -1) {
1056 unsigned Reg = MI->getOperand(i).getReg();
1057 if (TargetRegisterInfo::isVirtualRegister(Reg))
1060 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
1061 if (!RC->contains(Reg)) {
1062 ErrInfo = "Operand has incorrect register class.";
1070 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
1071 unsigned ConstantBusCount = 0;
1072 unsigned SGPRUsed = AMDGPU::NoRegister;
1073 for (int i = 0, e = MI->getNumOperands(); i != e; ++i) {
1074 const MachineOperand &MO = MI->getOperand(i);
1075 if (usesConstantBus(MRI, MO)) {
1077 if (MO.getReg() != SGPRUsed)
1079 SGPRUsed = MO.getReg();
1085 if (ConstantBusCount > 1) {
1086 ErrInfo = "VOP* instruction uses the constant bus more than once";
1091 // Verify SRC1 for VOP2 and VOPC
1092 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
1093 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1094 if (Src1.isImm() || Src1.isFPImm()) {
1095 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
1101 if (isVOP3(Opcode)) {
1102 if (Src0Idx != -1 && isLiteralConstant(MI->getOperand(Src0Idx))) {
1103 ErrInfo = "VOP3 src0 cannot be a literal constant.";
1106 if (Src1Idx != -1 && isLiteralConstant(MI->getOperand(Src1Idx))) {
1107 ErrInfo = "VOP3 src1 cannot be a literal constant.";
1110 if (Src2Idx != -1 && isLiteralConstant(MI->getOperand(Src2Idx))) {
1111 ErrInfo = "VOP3 src2 cannot be a literal constant.";
1116 // Verify misc. restrictions on specific instructions.
1117 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
1118 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
1119 const MachineOperand &Src0 = MI->getOperand(Src0Idx);
1120 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1121 const MachineOperand &Src2 = MI->getOperand(Src2Idx);
1122 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
1123 if (!compareMachineOp(Src0, Src1) &&
1124 !compareMachineOp(Src0, Src2)) {
1125 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
1134 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
1135 switch (MI.getOpcode()) {
1136 default: return AMDGPU::INSTRUCTION_LIST_END;
1137 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
1138 case AMDGPU::COPY: return AMDGPU::COPY;
1139 case AMDGPU::PHI: return AMDGPU::PHI;
1140 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
1141 case AMDGPU::S_MOV_B32:
1142 return MI.getOperand(1).isReg() ?
1143 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
1144 case AMDGPU::S_ADD_I32:
1145 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
1146 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
1147 case AMDGPU::S_SUB_I32:
1148 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
1149 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
1150 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
1151 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
1152 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
1153 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
1154 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
1155 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
1156 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
1157 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
1158 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
1159 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
1160 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
1161 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
1162 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
1163 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
1164 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
1165 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
1166 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
1167 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
1168 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
1169 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
1170 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
1171 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
1172 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
1173 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
1174 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
1175 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
1176 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
1177 case AMDGPU::S_LOAD_DWORD_IMM:
1178 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1179 case AMDGPU::S_LOAD_DWORDX2_IMM:
1180 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1181 case AMDGPU::S_LOAD_DWORDX4_IMM:
1182 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1183 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e32;
1184 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
1185 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
1189 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
1190 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
1193 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
1194 unsigned OpNo) const {
1195 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1196 const MCInstrDesc &Desc = get(MI.getOpcode());
1197 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
1198 Desc.OpInfo[OpNo].RegClass == -1)
1199 return MRI.getRegClass(MI.getOperand(OpNo).getReg());
1201 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
1202 return RI.getRegClass(RCID);
1205 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
1206 switch (MI.getOpcode()) {
1208 case AMDGPU::REG_SEQUENCE:
1210 case AMDGPU::INSERT_SUBREG:
1211 return RI.hasVGPRs(getOpRegClass(MI, 0));
1213 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
1217 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
1218 MachineBasicBlock::iterator I = MI;
1219 MachineBasicBlock *MBB = MI->getParent();
1220 MachineOperand &MO = MI->getOperand(OpIdx);
1221 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1222 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
1223 const TargetRegisterClass *RC = RI.getRegClass(RCID);
1224 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1226 Opcode = AMDGPU::COPY;
1227 else if (RI.isSGPRClass(RC))
1228 Opcode = AMDGPU::S_MOV_B32;
1231 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
1232 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC))
1233 VRC = &AMDGPU::VReg_64RegClass;
1235 VRC = &AMDGPU::VReg_32RegClass;
1237 unsigned Reg = MRI.createVirtualRegister(VRC);
1238 DebugLoc DL = MBB->findDebugLoc(I);
1239 BuildMI(*MI->getParent(), I, DL, get(Opcode), Reg)
1241 MO.ChangeToRegister(Reg, false);
1244 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
1245 MachineRegisterInfo &MRI,
1246 MachineOperand &SuperReg,
1247 const TargetRegisterClass *SuperRC,
1249 const TargetRegisterClass *SubRC)
1251 assert(SuperReg.isReg());
1253 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
1254 unsigned SubReg = MRI.createVirtualRegister(SubRC);
1256 // Just in case the super register is itself a sub-register, copy it to a new
1257 // value so we don't need to worry about merging its subreg index with the
1258 // SubIdx passed to this function. The register coalescer should be able to
1259 // eliminate this extra copy.
1260 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1262 .addOperand(SuperReg);
1264 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1266 .addReg(NewSuperReg, 0, SubIdx);
1270 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
1271 MachineBasicBlock::iterator MII,
1272 MachineRegisterInfo &MRI,
1274 const TargetRegisterClass *SuperRC,
1276 const TargetRegisterClass *SubRC) const {
1278 // XXX - Is there a better way to do this?
1279 if (SubIdx == AMDGPU::sub0)
1280 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
1281 if (SubIdx == AMDGPU::sub1)
1282 return MachineOperand::CreateImm(Op.getImm() >> 32);
1284 llvm_unreachable("Unhandled register index for immediate");
1287 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
1289 return MachineOperand::CreateReg(SubReg, false);
1292 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
1293 MachineBasicBlock::iterator MI,
1294 MachineRegisterInfo &MRI,
1295 const TargetRegisterClass *RC,
1296 const MachineOperand &Op) const {
1297 MachineBasicBlock *MBB = MI->getParent();
1298 DebugLoc DL = MI->getDebugLoc();
1299 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1300 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1301 unsigned Dst = MRI.createVirtualRegister(RC);
1303 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1305 .addImm(Op.getImm() & 0xFFFFFFFF);
1306 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1308 .addImm(Op.getImm() >> 32);
1310 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
1312 .addImm(AMDGPU::sub0)
1314 .addImm(AMDGPU::sub1);
1316 Worklist.push_back(Lo);
1317 Worklist.push_back(Hi);
1322 bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
1323 const MachineOperand *MO) const {
1324 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1325 const MCInstrDesc &InstDesc = get(MI->getOpcode());
1326 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
1327 const TargetRegisterClass *DefinedRC =
1328 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
1330 MO = &MI->getOperand(OpIdx);
1332 if (usesConstantBus(MRI, *MO)) {
1334 MO->isReg() ? MO->getReg() : (unsigned)AMDGPU::NoRegister;
1335 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1338 if (usesConstantBus(MRI, MI->getOperand(i)) &&
1339 MI->getOperand(i).isReg() && MI->getOperand(i).getReg() != SGPRUsed) {
1347 const TargetRegisterClass *RC = MRI.getRegClass(MO->getReg());
1348 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass));
1352 // Handle non-register types that are treated like immediates.
1353 assert(MO->isImm() || MO->isFPImm() || MO->isTargetIndex() || MO->isFI());
1356 // This operand expects an immediate.
1360 return isImmOperandLegal(MI, OpIdx, *MO);
1363 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
1364 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1366 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1367 AMDGPU::OpName::src0);
1368 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1369 AMDGPU::OpName::src1);
1370 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1371 AMDGPU::OpName::src2);
1374 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
1376 if (!isOperandLegal(MI, Src0Idx))
1377 legalizeOpWithMove(MI, Src0Idx);
1380 if (isOperandLegal(MI, Src1Idx))
1383 // Usually src0 of VOP2 instructions allow more types of inputs
1384 // than src1, so try to commute the instruction to decrease our
1385 // chances of having to insert a MOV instruction to legalize src1.
1386 if (MI->isCommutable()) {
1387 if (commuteInstruction(MI))
1388 // If we are successful in commuting, then we know MI is legal, so
1393 legalizeOpWithMove(MI, Src1Idx);
1397 // XXX - Do any VOP3 instructions read VCC?
1399 if (isVOP3(MI->getOpcode())) {
1400 int VOP3Idx[3] = { Src0Idx, Src1Idx, Src2Idx };
1402 // Find the one SGPR operand we are allowed to use.
1403 unsigned SGPRReg = findUsedSGPR(MI, VOP3Idx);
1405 for (unsigned i = 0; i < 3; ++i) {
1406 int Idx = VOP3Idx[i];
1409 MachineOperand &MO = MI->getOperand(Idx);
1412 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
1413 continue; // VGPRs are legal
1415 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
1417 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
1418 SGPRReg = MO.getReg();
1419 // We can use one SGPR in each VOP3 instruction.
1422 } else if (!isLiteralConstant(MO)) {
1423 // If it is not a register and not a literal constant, then it must be
1424 // an inline constant which is always legal.
1427 // If we make it this far, then the operand is not legal and we must
1429 legalizeOpWithMove(MI, Idx);
1433 // Legalize REG_SEQUENCE and PHI
1434 // The register class of the operands much be the same type as the register
1435 // class of the output.
1436 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
1437 MI->getOpcode() == AMDGPU::PHI) {
1438 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
1439 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1440 if (!MI->getOperand(i).isReg() ||
1441 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1443 const TargetRegisterClass *OpRC =
1444 MRI.getRegClass(MI->getOperand(i).getReg());
1445 if (RI.hasVGPRs(OpRC)) {
1452 // If any of the operands are VGPR registers, then they all most be
1453 // otherwise we will create illegal VGPR->SGPR copies when legalizing
1455 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
1458 VRC = RI.getEquivalentVGPRClass(SRC);
1465 // Update all the operands so they have the same type.
1466 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1467 if (!MI->getOperand(i).isReg() ||
1468 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1470 unsigned DstReg = MRI.createVirtualRegister(RC);
1471 MachineBasicBlock *InsertBB;
1472 MachineBasicBlock::iterator Insert;
1473 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
1474 InsertBB = MI->getParent();
1477 // MI is a PHI instruction.
1478 InsertBB = MI->getOperand(i + 1).getMBB();
1479 Insert = InsertBB->getFirstTerminator();
1481 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
1482 get(AMDGPU::COPY), DstReg)
1483 .addOperand(MI->getOperand(i));
1484 MI->getOperand(i).setReg(DstReg);
1488 // Legalize INSERT_SUBREG
1489 // src0 must have the same register class as dst
1490 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
1491 unsigned Dst = MI->getOperand(0).getReg();
1492 unsigned Src0 = MI->getOperand(1).getReg();
1493 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
1494 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
1495 if (DstRC != Src0RC) {
1496 MachineBasicBlock &MBB = *MI->getParent();
1497 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
1498 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
1500 MI->getOperand(1).setReg(NewSrc0);
1505 // Legalize MUBUF* instructions
1506 // FIXME: If we start using the non-addr64 instructions for compute, we
1507 // may need to legalize them here.
1509 AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::srsrc);
1510 if (SRsrcIdx != -1) {
1511 // We have an MUBUF instruction
1512 MachineOperand *SRsrc = &MI->getOperand(SRsrcIdx);
1513 unsigned SRsrcRC = get(MI->getOpcode()).OpInfo[SRsrcIdx].RegClass;
1514 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
1515 RI.getRegClass(SRsrcRC))) {
1516 // The operands are legal.
1517 // FIXME: We may need to legalize operands besided srsrc.
1521 MachineBasicBlock &MBB = *MI->getParent();
1522 // Extract the the ptr from the resource descriptor.
1524 // SRsrcPtrLo = srsrc:sub0
1525 unsigned SRsrcPtrLo = buildExtractSubReg(MI, MRI, *SRsrc,
1526 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VReg_32RegClass);
1528 // SRsrcPtrHi = srsrc:sub1
1529 unsigned SRsrcPtrHi = buildExtractSubReg(MI, MRI, *SRsrc,
1530 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VReg_32RegClass);
1532 // Create an empty resource descriptor
1533 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1534 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1535 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1536 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1539 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1543 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1544 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1546 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1548 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1549 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1551 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1553 // NewSRsrc = {Zero64, SRsrcFormat}
1554 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1557 .addImm(AMDGPU::sub0_sub1)
1558 .addReg(SRsrcFormatLo)
1559 .addImm(AMDGPU::sub2)
1560 .addReg(SRsrcFormatHi)
1561 .addImm(AMDGPU::sub3);
1563 MachineOperand *VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1564 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
1565 unsigned NewVAddrLo;
1566 unsigned NewVAddrHi;
1568 // This is already an ADDR64 instruction so we need to add the pointer
1569 // extracted from the resource descriptor to the current value of VAddr.
1570 NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1571 NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1573 // NewVaddrLo = SRsrcPtrLo + VAddr:sub0
1574 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1577 .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
1578 .addReg(AMDGPU::VCC, RegState::ImplicitDefine);
1580 // NewVaddrHi = SRsrcPtrHi + VAddr:sub1
1581 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1584 .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
1585 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1586 .addReg(AMDGPU::VCC, RegState::Implicit);
1589 // This instructions is the _OFFSET variant, so we need to convert it to
1591 MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
1592 MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
1593 MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
1594 assert(SOffset->isImm() && SOffset->getImm() == 0 && "Legalizing MUBUF "
1595 "with non-zero soffset is not implemented");
1598 // Create the new instruction.
1599 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
1600 MachineInstr *Addr64 =
1601 BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
1604 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
1605 // This will be replaced later
1606 // with the new value of vaddr.
1607 .addOperand(*Offset);
1609 MI->removeFromParent();
1612 NewVAddrLo = SRsrcPtrLo;
1613 NewVAddrHi = SRsrcPtrHi;
1614 VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1615 SRsrc = getNamedOperand(*MI, AMDGPU::OpName::srsrc);
1618 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1619 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1622 .addImm(AMDGPU::sub0)
1624 .addImm(AMDGPU::sub1);
1627 // Update the instruction to use NewVaddr
1628 VAddr->setReg(NewVAddr);
1629 // Update the instruction to use NewSRsrc
1630 SRsrc->setReg(NewSRsrc);
1634 void SIInstrInfo::splitSMRD(MachineInstr *MI,
1635 const TargetRegisterClass *HalfRC,
1636 unsigned HalfImmOp, unsigned HalfSGPROp,
1637 MachineInstr *&Lo, MachineInstr *&Hi) const {
1639 DebugLoc DL = MI->getDebugLoc();
1640 MachineBasicBlock *MBB = MI->getParent();
1641 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1642 unsigned RegLo = MRI.createVirtualRegister(HalfRC);
1643 unsigned RegHi = MRI.createVirtualRegister(HalfRC);
1644 unsigned HalfSize = HalfRC->getSize();
1645 const MachineOperand *OffOp =
1646 getNamedOperand(*MI, AMDGPU::OpName::offset);
1647 const MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
1650 // Handle the _IMM variant
1651 unsigned LoOffset = OffOp->getImm();
1652 unsigned HiOffset = LoOffset + (HalfSize / 4);
1653 Lo = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegLo)
1657 if (!isUInt<8>(HiOffset)) {
1658 unsigned OffsetSGPR =
1659 MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1660 BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32), OffsetSGPR)
1661 .addImm(HiOffset << 2); // The immediate offset is in dwords,
1662 // but offset in register is in bytes.
1663 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
1665 .addReg(OffsetSGPR);
1667 Hi = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegHi)
1672 // Handle the _SGPR variant
1673 MachineOperand *SOff = getNamedOperand(*MI, AMDGPU::OpName::soff);
1674 Lo = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegLo)
1677 unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1678 BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
1681 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp))
1683 .addReg(OffsetSGPR);
1686 unsigned SubLo, SubHi;
1689 SubLo = AMDGPU::sub0;
1690 SubHi = AMDGPU::sub1;
1693 SubLo = AMDGPU::sub0_sub1;
1694 SubHi = AMDGPU::sub2_sub3;
1697 SubLo = AMDGPU::sub0_sub1_sub2_sub3;
1698 SubHi = AMDGPU::sub4_sub5_sub6_sub7;
1701 SubLo = AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
1702 SubHi = AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15;
1705 llvm_unreachable("Unhandled HalfSize");
1708 BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE))
1709 .addOperand(MI->getOperand(0))
1716 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
1717 MachineBasicBlock *MBB = MI->getParent();
1718 switch (MI->getOpcode()) {
1719 case AMDGPU::S_LOAD_DWORD_IMM:
1720 case AMDGPU::S_LOAD_DWORD_SGPR:
1721 case AMDGPU::S_LOAD_DWORDX2_IMM:
1722 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1723 case AMDGPU::S_LOAD_DWORDX4_IMM:
1724 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1725 unsigned NewOpcode = getVALUOp(*MI);
1729 if (MI->getOperand(2).isReg()) {
1730 RegOffset = MI->getOperand(2).getReg();
1733 assert(MI->getOperand(2).isImm());
1734 // SMRD instructions take a dword offsets and MUBUF instructions
1735 // take a byte offset.
1736 ImmOffset = MI->getOperand(2).getImm() << 2;
1737 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1738 if (isUInt<12>(ImmOffset)) {
1739 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1743 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1750 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1751 unsigned DWord0 = RegOffset;
1752 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1753 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1754 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1756 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
1758 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
1759 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1760 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
1761 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1762 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
1764 .addImm(AMDGPU::sub0)
1766 .addImm(AMDGPU::sub1)
1768 .addImm(AMDGPU::sub2)
1770 .addImm(AMDGPU::sub3);
1771 MI->setDesc(get(NewOpcode));
1772 if (MI->getOperand(2).isReg()) {
1773 MI->getOperand(2).setReg(MI->getOperand(1).getReg());
1775 MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
1777 MI->getOperand(1).setReg(SRsrc);
1778 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
1780 const TargetRegisterClass *NewDstRC =
1781 RI.getRegClass(get(NewOpcode).OpInfo[0].RegClass);
1783 unsigned DstReg = MI->getOperand(0).getReg();
1784 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1785 MRI.replaceRegWith(DstReg, NewDstReg);
1788 case AMDGPU::S_LOAD_DWORDX8_IMM:
1789 case AMDGPU::S_LOAD_DWORDX8_SGPR: {
1790 MachineInstr *Lo, *Hi;
1791 splitSMRD(MI, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM,
1792 AMDGPU::S_LOAD_DWORDX4_SGPR, Lo, Hi);
1793 MI->eraseFromParent();
1794 moveSMRDToVALU(Lo, MRI);
1795 moveSMRDToVALU(Hi, MRI);
1799 case AMDGPU::S_LOAD_DWORDX16_IMM:
1800 case AMDGPU::S_LOAD_DWORDX16_SGPR: {
1801 MachineInstr *Lo, *Hi;
1802 splitSMRD(MI, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM,
1803 AMDGPU::S_LOAD_DWORDX8_SGPR, Lo, Hi);
1804 MI->eraseFromParent();
1805 moveSMRDToVALU(Lo, MRI);
1806 moveSMRDToVALU(Hi, MRI);
1812 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
1813 SmallVector<MachineInstr *, 128> Worklist;
1814 Worklist.push_back(&TopInst);
1816 while (!Worklist.empty()) {
1817 MachineInstr *Inst = Worklist.pop_back_val();
1818 MachineBasicBlock *MBB = Inst->getParent();
1819 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1821 unsigned Opcode = Inst->getOpcode();
1822 unsigned NewOpcode = getVALUOp(*Inst);
1824 // Handle some special cases
1827 if (isSMRD(Inst->getOpcode())) {
1828 moveSMRDToVALU(Inst, MRI);
1831 case AMDGPU::S_MOV_B64: {
1832 DebugLoc DL = Inst->getDebugLoc();
1834 // If the source operand is a register we can replace this with a
1836 if (Inst->getOperand(1).isReg()) {
1837 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
1838 .addOperand(Inst->getOperand(0))
1839 .addOperand(Inst->getOperand(1));
1840 Worklist.push_back(Copy);
1842 // Otherwise, we need to split this into two movs, because there is
1843 // no 64-bit VALU move instruction.
1844 unsigned Reg = Inst->getOperand(0).getReg();
1845 unsigned Dst = split64BitImm(Worklist,
1848 MRI.getRegClass(Reg),
1849 Inst->getOperand(1));
1850 MRI.replaceRegWith(Reg, Dst);
1852 Inst->eraseFromParent();
1855 case AMDGPU::S_AND_B64:
1856 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
1857 Inst->eraseFromParent();
1860 case AMDGPU::S_OR_B64:
1861 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
1862 Inst->eraseFromParent();
1865 case AMDGPU::S_XOR_B64:
1866 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
1867 Inst->eraseFromParent();
1870 case AMDGPU::S_NOT_B64:
1871 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
1872 Inst->eraseFromParent();
1875 case AMDGPU::S_BCNT1_I32_B64:
1876 splitScalar64BitBCNT(Worklist, Inst);
1877 Inst->eraseFromParent();
1880 case AMDGPU::S_BFE_U64:
1881 case AMDGPU::S_BFE_I64:
1882 case AMDGPU::S_BFM_B64:
1883 llvm_unreachable("Moving this op to VALU not implemented");
1886 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
1887 // We cannot move this instruction to the VALU, so we should try to
1888 // legalize its operands instead.
1889 legalizeOperands(Inst);
1893 // Use the new VALU Opcode.
1894 const MCInstrDesc &NewDesc = get(NewOpcode);
1895 Inst->setDesc(NewDesc);
1897 // Remove any references to SCC. Vector instructions can't read from it, and
1898 // We're just about to add the implicit use / defs of VCC, and we don't want
1900 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
1901 MachineOperand &Op = Inst->getOperand(i);
1902 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
1903 Inst->RemoveOperand(i);
1906 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
1907 // We are converting these to a BFE, so we need to add the missing
1908 // operands for the size and offset.
1909 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
1910 Inst->addOperand(MachineOperand::CreateImm(0));
1911 Inst->addOperand(MachineOperand::CreateImm(Size));
1913 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
1914 // The VALU version adds the second operand to the result, so insert an
1916 Inst->addOperand(MachineOperand::CreateImm(0));
1919 addDescImplicitUseDef(NewDesc, Inst);
1921 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
1922 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
1923 // If we need to move this to VGPRs, we need to unpack the second operand
1924 // back into the 2 separate ones for bit offset and width.
1925 assert(OffsetWidthOp.isImm() &&
1926 "Scalar BFE is only implemented for constant width and offset");
1927 uint32_t Imm = OffsetWidthOp.getImm();
1929 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
1930 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
1931 Inst->RemoveOperand(2); // Remove old immediate.
1932 Inst->addOperand(MachineOperand::CreateImm(Offset));
1933 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
1936 // Update the destination register class.
1938 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
1941 // For target instructions, getOpRegClass just returns the virtual
1942 // register class associated with the operand, so we need to find an
1943 // equivalent VGPR register class in order to move the instruction to the
1947 case AMDGPU::REG_SEQUENCE:
1948 case AMDGPU::INSERT_SUBREG:
1949 if (RI.hasVGPRs(NewDstRC))
1951 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
1959 unsigned DstReg = Inst->getOperand(0).getReg();
1960 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1961 MRI.replaceRegWith(DstReg, NewDstReg);
1963 // Legalize the operands
1964 legalizeOperands(Inst);
1966 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
1967 E = MRI.use_end(); I != E; ++I) {
1968 MachineInstr &UseMI = *I->getParent();
1969 if (!canReadVGPR(UseMI, I.getOperandNo())) {
1970 Worklist.push_back(&UseMI);
1976 //===----------------------------------------------------------------------===//
1977 // Indirect addressing callbacks
1978 //===----------------------------------------------------------------------===//
1980 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
1981 unsigned Channel) const {
1982 assert(Channel == 0);
1986 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
1987 return &AMDGPU::VReg_32RegClass;
1990 void SIInstrInfo::splitScalar64BitUnaryOp(
1991 SmallVectorImpl<MachineInstr *> &Worklist,
1993 unsigned Opcode) const {
1994 MachineBasicBlock &MBB = *Inst->getParent();
1995 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1997 MachineOperand &Dest = Inst->getOperand(0);
1998 MachineOperand &Src0 = Inst->getOperand(1);
1999 DebugLoc DL = Inst->getDebugLoc();
2001 MachineBasicBlock::iterator MII = Inst;
2003 const MCInstrDesc &InstDesc = get(Opcode);
2004 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2005 MRI.getRegClass(Src0.getReg()) :
2006 &AMDGPU::SGPR_32RegClass;
2008 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2010 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2011 AMDGPU::sub0, Src0SubRC);
2013 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2014 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2016 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2017 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2018 .addOperand(SrcReg0Sub0);
2020 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2021 AMDGPU::sub1, Src0SubRC);
2023 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2024 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2025 .addOperand(SrcReg0Sub1);
2027 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2028 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2030 .addImm(AMDGPU::sub0)
2032 .addImm(AMDGPU::sub1);
2034 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2036 // Try to legalize the operands in case we need to swap the order to keep it
2038 Worklist.push_back(LoHalf);
2039 Worklist.push_back(HiHalf);
2042 void SIInstrInfo::splitScalar64BitBinaryOp(
2043 SmallVectorImpl<MachineInstr *> &Worklist,
2045 unsigned Opcode) const {
2046 MachineBasicBlock &MBB = *Inst->getParent();
2047 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2049 MachineOperand &Dest = Inst->getOperand(0);
2050 MachineOperand &Src0 = Inst->getOperand(1);
2051 MachineOperand &Src1 = Inst->getOperand(2);
2052 DebugLoc DL = Inst->getDebugLoc();
2054 MachineBasicBlock::iterator MII = Inst;
2056 const MCInstrDesc &InstDesc = get(Opcode);
2057 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2058 MRI.getRegClass(Src0.getReg()) :
2059 &AMDGPU::SGPR_32RegClass;
2061 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2062 const TargetRegisterClass *Src1RC = Src1.isReg() ?
2063 MRI.getRegClass(Src1.getReg()) :
2064 &AMDGPU::SGPR_32RegClass;
2066 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
2068 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2069 AMDGPU::sub0, Src0SubRC);
2070 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2071 AMDGPU::sub0, Src1SubRC);
2073 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2074 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2076 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2077 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2078 .addOperand(SrcReg0Sub0)
2079 .addOperand(SrcReg1Sub0);
2081 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2082 AMDGPU::sub1, Src0SubRC);
2083 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2084 AMDGPU::sub1, Src1SubRC);
2086 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2087 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2088 .addOperand(SrcReg0Sub1)
2089 .addOperand(SrcReg1Sub1);
2091 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2092 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2094 .addImm(AMDGPU::sub0)
2096 .addImm(AMDGPU::sub1);
2098 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2100 // Try to legalize the operands in case we need to swap the order to keep it
2102 Worklist.push_back(LoHalf);
2103 Worklist.push_back(HiHalf);
2106 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
2107 MachineInstr *Inst) const {
2108 MachineBasicBlock &MBB = *Inst->getParent();
2109 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2111 MachineBasicBlock::iterator MII = Inst;
2112 DebugLoc DL = Inst->getDebugLoc();
2114 MachineOperand &Dest = Inst->getOperand(0);
2115 MachineOperand &Src = Inst->getOperand(1);
2117 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e32);
2118 const TargetRegisterClass *SrcRC = Src.isReg() ?
2119 MRI.getRegClass(Src.getReg()) :
2120 &AMDGPU::SGPR_32RegClass;
2122 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2123 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2125 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
2127 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2128 AMDGPU::sub0, SrcSubRC);
2129 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2130 AMDGPU::sub1, SrcSubRC);
2132 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
2133 .addOperand(SrcRegSub0)
2136 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
2137 .addOperand(SrcRegSub1)
2140 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2142 Worklist.push_back(First);
2143 Worklist.push_back(Second);
2146 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
2147 MachineInstr *Inst) const {
2148 // Add the implict and explicit register definitions.
2149 if (NewDesc.ImplicitUses) {
2150 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
2151 unsigned Reg = NewDesc.ImplicitUses[i];
2152 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
2156 if (NewDesc.ImplicitDefs) {
2157 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
2158 unsigned Reg = NewDesc.ImplicitDefs[i];
2159 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
2164 unsigned SIInstrInfo::findUsedSGPR(const MachineInstr *MI,
2165 int OpIndices[3]) const {
2166 const MCInstrDesc &Desc = get(MI->getOpcode());
2168 // Find the one SGPR operand we are allowed to use.
2169 unsigned SGPRReg = AMDGPU::NoRegister;
2171 // First we need to consider the instruction's operand requirements before
2172 // legalizing. Some operands are required to be SGPRs, such as implicit uses
2173 // of VCC, but we are still bound by the constant bus requirement to only use
2176 // If the operand's class is an SGPR, we can never move it.
2178 for (const MachineOperand &MO : MI->implicit_operands()) {
2179 // We only care about reads.
2183 if (MO.getReg() == AMDGPU::VCC)
2186 if (MO.getReg() == AMDGPU::FLAT_SCR)
2187 return AMDGPU::FLAT_SCR;
2190 unsigned UsedSGPRs[3] = { AMDGPU::NoRegister };
2191 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
2193 for (unsigned i = 0; i < 3; ++i) {
2194 int Idx = OpIndices[i];
2198 const MachineOperand &MO = MI->getOperand(Idx);
2199 if (RI.isSGPRClassID(Desc.OpInfo[Idx].RegClass))
2200 SGPRReg = MO.getReg();
2202 if (MO.isReg() && RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
2203 UsedSGPRs[i] = MO.getReg();
2206 if (SGPRReg != AMDGPU::NoRegister)
2209 // We don't have a required SGPR operand, so we have a bit more freedom in
2210 // selecting operands to move.
2212 // Try to select the most used SGPR. If an SGPR is equal to one of the
2213 // others, we choose that.
2216 // V_FMA_F32 v0, s0, s0, s0 -> No moves
2217 // V_FMA_F32 v0, s0, s1, s0 -> Move s1
2219 if (UsedSGPRs[0] != AMDGPU::NoRegister) {
2220 if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
2221 SGPRReg = UsedSGPRs[0];
2224 if (SGPRReg == AMDGPU::NoRegister && UsedSGPRs[1] != AMDGPU::NoRegister) {
2225 if (UsedSGPRs[1] == UsedSGPRs[2])
2226 SGPRReg = UsedSGPRs[1];
2232 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
2233 MachineBasicBlock *MBB,
2234 MachineBasicBlock::iterator I,
2236 unsigned Address, unsigned OffsetReg) const {
2237 const DebugLoc &DL = MBB->findDebugLoc(I);
2238 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
2239 getIndirectIndexBegin(*MBB->getParent()));
2241 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
2242 .addReg(IndirectBaseReg, RegState::Define)
2243 .addOperand(I->getOperand(0))
2244 .addReg(IndirectBaseReg)
2250 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
2251 MachineBasicBlock *MBB,
2252 MachineBasicBlock::iterator I,
2254 unsigned Address, unsigned OffsetReg) const {
2255 const DebugLoc &DL = MBB->findDebugLoc(I);
2256 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
2257 getIndirectIndexBegin(*MBB->getParent()));
2259 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
2260 .addOperand(I->getOperand(0))
2261 .addOperand(I->getOperand(1))
2262 .addReg(IndirectBaseReg)
2268 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
2269 const MachineFunction &MF) const {
2270 int End = getIndirectIndexEnd(MF);
2271 int Begin = getIndirectIndexBegin(MF);
2277 for (int Index = Begin; Index <= End; ++Index)
2278 Reserved.set(AMDGPU::VReg_32RegClass.getRegister(Index));
2280 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
2281 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
2283 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
2284 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
2286 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
2287 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
2289 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
2290 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
2292 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
2293 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));
2296 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
2297 unsigned OperandName) const {
2298 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
2302 return &MI.getOperand(Idx);