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)) {
91 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
94 if (Load0->getOperand(1) != Load1->getOperand(1))
98 if (findChainOperand(Load0) != findChainOperand(Load1))
101 // Skip read2 / write2 variants for simplicity.
102 // TODO: We should report true if the used offsets are adjacent (excluded
104 if (AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::data1) != -1 ||
105 AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::data1) != -1)
108 Offset0 = cast<ConstantSDNode>(Load0->getOperand(2))->getZExtValue();
109 Offset1 = cast<ConstantSDNode>(Load1->getOperand(2))->getZExtValue();
113 if (isSMRD(Opc0) && isSMRD(Opc1)) {
114 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
117 if (Load0->getOperand(0) != Load1->getOperand(0))
121 if (findChainOperand(Load0) != findChainOperand(Load1))
124 Offset0 = cast<ConstantSDNode>(Load0->getOperand(1))->getZExtValue();
125 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue();
129 // MUBUF and MTBUF can access the same addresses.
130 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1))) {
132 // MUBUF and MTBUF have vaddr at different indices.
133 if (!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::soffset) ||
134 findChainOperand(Load0) != findChainOperand(Load1) ||
135 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::vaddr) ||
136 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::srsrc))
139 int OffIdx0 = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
140 int OffIdx1 = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
142 if (OffIdx0 == -1 || OffIdx1 == -1)
145 // getNamedOperandIdx returns the index for MachineInstrs. Since they
146 // inlcude the output in the operand list, but SDNodes don't, we need to
147 // subtract the index by one.
151 SDValue Off0 = Load0->getOperand(OffIdx0);
152 SDValue Off1 = Load1->getOperand(OffIdx1);
154 // The offset might be a FrameIndexSDNode.
155 if (!isa<ConstantSDNode>(Off0) || !isa<ConstantSDNode>(Off1))
158 Offset0 = cast<ConstantSDNode>(Off0)->getZExtValue();
159 Offset1 = cast<ConstantSDNode>(Off1)->getZExtValue();
166 static bool isStride64(unsigned Opc) {
168 case AMDGPU::DS_READ2ST64_B32:
169 case AMDGPU::DS_READ2ST64_B64:
170 case AMDGPU::DS_WRITE2ST64_B32:
171 case AMDGPU::DS_WRITE2ST64_B64:
178 bool SIInstrInfo::getLdStBaseRegImmOfs(MachineInstr *LdSt,
179 unsigned &BaseReg, unsigned &Offset,
180 const TargetRegisterInfo *TRI) const {
181 unsigned Opc = LdSt->getOpcode();
183 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
184 AMDGPU::OpName::offset);
186 // Normal, single offset LDS instruction.
187 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
188 AMDGPU::OpName::addr);
190 BaseReg = AddrReg->getReg();
191 Offset = OffsetImm->getImm();
195 // The 2 offset instructions use offset0 and offset1 instead. We can treat
196 // these as a load with a single offset if the 2 offsets are consecutive. We
197 // will use this for some partially aligned loads.
198 const MachineOperand *Offset0Imm = getNamedOperand(*LdSt,
199 AMDGPU::OpName::offset0);
200 const MachineOperand *Offset1Imm = getNamedOperand(*LdSt,
201 AMDGPU::OpName::offset1);
203 uint8_t Offset0 = Offset0Imm->getImm();
204 uint8_t Offset1 = Offset1Imm->getImm();
205 assert(Offset1 > Offset0);
207 if (Offset1 - Offset0 == 1) {
208 // Each of these offsets is in element sized units, so we need to convert
209 // to bytes of the individual reads.
213 EltSize = getOpRegClass(*LdSt, 0)->getSize() / 2;
215 assert(LdSt->mayStore());
216 int Data0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
217 EltSize = getOpRegClass(*LdSt, Data0Idx)->getSize();
223 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
224 AMDGPU::OpName::addr);
225 BaseReg = AddrReg->getReg();
226 Offset = EltSize * Offset0;
233 if (isMUBUF(Opc) || isMTBUF(Opc)) {
234 if (AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::soffset) != -1)
237 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
238 AMDGPU::OpName::vaddr);
242 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
243 AMDGPU::OpName::offset);
244 BaseReg = AddrReg->getReg();
245 Offset = OffsetImm->getImm();
250 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
251 AMDGPU::OpName::offset);
255 const MachineOperand *SBaseReg = getNamedOperand(*LdSt,
256 AMDGPU::OpName::sbase);
257 BaseReg = SBaseReg->getReg();
258 Offset = OffsetImm->getImm();
265 bool SIInstrInfo::shouldClusterLoads(MachineInstr *FirstLdSt,
266 MachineInstr *SecondLdSt,
267 unsigned NumLoads) const {
268 unsigned Opc0 = FirstLdSt->getOpcode();
269 unsigned Opc1 = SecondLdSt->getOpcode();
271 // TODO: This needs finer tuning
275 if (isDS(Opc0) && isDS(Opc1))
278 if (isSMRD(Opc0) && isSMRD(Opc1))
281 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1)))
288 SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
289 MachineBasicBlock::iterator MI, DebugLoc DL,
290 unsigned DestReg, unsigned SrcReg,
291 bool KillSrc) const {
293 // If we are trying to copy to or from SCC, there is a bug somewhere else in
294 // the backend. While it may be theoretically possible to do this, it should
295 // never be necessary.
296 assert(DestReg != AMDGPU::SCC && SrcReg != AMDGPU::SCC);
298 static const int16_t Sub0_15[] = {
299 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
300 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
301 AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
302 AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15, 0
305 static const int16_t Sub0_7[] = {
306 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
307 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7, 0
310 static const int16_t Sub0_3[] = {
311 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
314 static const int16_t Sub0_2[] = {
315 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
318 static const int16_t Sub0_1[] = {
319 AMDGPU::sub0, AMDGPU::sub1, 0
323 const int16_t *SubIndices;
325 if (AMDGPU::M0 == DestReg) {
326 // Check if M0 isn't already set to this value
327 for (MachineBasicBlock::reverse_iterator E = MBB.rend(),
328 I = MachineBasicBlock::reverse_iterator(MI); I != E; ++I) {
330 if (!I->definesRegister(AMDGPU::M0))
333 unsigned Opc = I->getOpcode();
334 if (Opc != TargetOpcode::COPY && Opc != AMDGPU::S_MOV_B32)
337 if (!I->readsRegister(SrcReg))
340 // The copy isn't necessary
345 if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
346 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
347 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
348 .addReg(SrcReg, getKillRegState(KillSrc));
351 } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
352 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
353 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
354 .addReg(SrcReg, getKillRegState(KillSrc));
357 } else if (AMDGPU::SReg_128RegClass.contains(DestReg)) {
358 assert(AMDGPU::SReg_128RegClass.contains(SrcReg));
359 Opcode = AMDGPU::S_MOV_B32;
362 } else if (AMDGPU::SReg_256RegClass.contains(DestReg)) {
363 assert(AMDGPU::SReg_256RegClass.contains(SrcReg));
364 Opcode = AMDGPU::S_MOV_B32;
367 } else if (AMDGPU::SReg_512RegClass.contains(DestReg)) {
368 assert(AMDGPU::SReg_512RegClass.contains(SrcReg));
369 Opcode = AMDGPU::S_MOV_B32;
370 SubIndices = Sub0_15;
372 } else if (AMDGPU::VReg_32RegClass.contains(DestReg)) {
373 assert(AMDGPU::VReg_32RegClass.contains(SrcReg) ||
374 AMDGPU::SReg_32RegClass.contains(SrcReg));
375 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
376 .addReg(SrcReg, getKillRegState(KillSrc));
379 } else if (AMDGPU::VReg_64RegClass.contains(DestReg)) {
380 assert(AMDGPU::VReg_64RegClass.contains(SrcReg) ||
381 AMDGPU::SReg_64RegClass.contains(SrcReg));
382 Opcode = AMDGPU::V_MOV_B32_e32;
385 } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
386 assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
387 Opcode = AMDGPU::V_MOV_B32_e32;
390 } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
391 assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
392 AMDGPU::SReg_128RegClass.contains(SrcReg));
393 Opcode = AMDGPU::V_MOV_B32_e32;
396 } else if (AMDGPU::VReg_256RegClass.contains(DestReg)) {
397 assert(AMDGPU::VReg_256RegClass.contains(SrcReg) ||
398 AMDGPU::SReg_256RegClass.contains(SrcReg));
399 Opcode = AMDGPU::V_MOV_B32_e32;
402 } else if (AMDGPU::VReg_512RegClass.contains(DestReg)) {
403 assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
404 AMDGPU::SReg_512RegClass.contains(SrcReg));
405 Opcode = AMDGPU::V_MOV_B32_e32;
406 SubIndices = Sub0_15;
409 llvm_unreachable("Can't copy register!");
412 while (unsigned SubIdx = *SubIndices++) {
413 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
414 get(Opcode), RI.getSubReg(DestReg, SubIdx));
416 Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
419 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
423 unsigned SIInstrInfo::commuteOpcode(unsigned Opcode) const {
426 // Try to map original to commuted opcode
427 if ((NewOpc = AMDGPU::getCommuteRev(Opcode)) != -1)
430 // Try to map commuted to original opcode
431 if ((NewOpc = AMDGPU::getCommuteOrig(Opcode)) != -1)
437 static bool shouldTryToSpillVGPRs(MachineFunction *MF) {
439 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
440 const TargetMachine &TM = MF->getTarget();
442 // FIXME: Even though it can cause problems, we need to enable
443 // spilling at -O0, since the fast register allocator always
444 // spills registers that are live at the end of blocks.
445 return MFI->getShaderType() == ShaderType::COMPUTE &&
446 TM.getOptLevel() == CodeGenOpt::None;
450 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
451 MachineBasicBlock::iterator MI,
452 unsigned SrcReg, bool isKill,
454 const TargetRegisterClass *RC,
455 const TargetRegisterInfo *TRI) const {
456 MachineFunction *MF = MBB.getParent();
457 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
458 DebugLoc DL = MBB.findDebugLoc(MI);
461 if (RI.isSGPRClass(RC)) {
462 // We are only allowed to create one new instruction when spilling
463 // registers, so we need to use pseudo instruction for spilling
465 switch (RC->getSize() * 8) {
466 case 32: Opcode = AMDGPU::SI_SPILL_S32_SAVE; break;
467 case 64: Opcode = AMDGPU::SI_SPILL_S64_SAVE; break;
468 case 128: Opcode = AMDGPU::SI_SPILL_S128_SAVE; break;
469 case 256: Opcode = AMDGPU::SI_SPILL_S256_SAVE; break;
470 case 512: Opcode = AMDGPU::SI_SPILL_S512_SAVE; break;
472 } else if(shouldTryToSpillVGPRs(MF) && RI.hasVGPRs(RC)) {
473 switch(RC->getSize() * 8) {
474 case 32: Opcode = AMDGPU::SI_SPILL_V32_SAVE; break;
475 case 64: Opcode = AMDGPU::SI_SPILL_V64_SAVE; break;
476 case 96: Opcode = AMDGPU::SI_SPILL_V96_SAVE; break;
477 case 128: Opcode = AMDGPU::SI_SPILL_V128_SAVE; break;
478 case 256: Opcode = AMDGPU::SI_SPILL_V256_SAVE; break;
479 case 512: Opcode = AMDGPU::SI_SPILL_V512_SAVE; break;
484 FrameInfo->setObjectAlignment(FrameIndex, 4);
485 BuildMI(MBB, MI, DL, get(Opcode))
487 .addFrameIndex(FrameIndex);
489 LLVMContext &Ctx = MF->getFunction()->getContext();
490 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Do not know how to"
492 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), AMDGPU::VGPR0)
497 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
498 MachineBasicBlock::iterator MI,
499 unsigned DestReg, int FrameIndex,
500 const TargetRegisterClass *RC,
501 const TargetRegisterInfo *TRI) const {
502 MachineFunction *MF = MBB.getParent();
503 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
504 DebugLoc DL = MBB.findDebugLoc(MI);
507 if (RI.isSGPRClass(RC)){
508 switch(RC->getSize() * 8) {
509 case 32: Opcode = AMDGPU::SI_SPILL_S32_RESTORE; break;
510 case 64: Opcode = AMDGPU::SI_SPILL_S64_RESTORE; break;
511 case 128: Opcode = AMDGPU::SI_SPILL_S128_RESTORE; break;
512 case 256: Opcode = AMDGPU::SI_SPILL_S256_RESTORE; break;
513 case 512: Opcode = AMDGPU::SI_SPILL_S512_RESTORE; break;
515 } else if(shouldTryToSpillVGPRs(MF) && RI.hasVGPRs(RC)) {
516 switch(RC->getSize() * 8) {
517 case 32: Opcode = AMDGPU::SI_SPILL_V32_RESTORE; break;
518 case 64: Opcode = AMDGPU::SI_SPILL_V64_RESTORE; break;
519 case 96: Opcode = AMDGPU::SI_SPILL_V96_RESTORE; break;
520 case 128: Opcode = AMDGPU::SI_SPILL_V128_RESTORE; break;
521 case 256: Opcode = AMDGPU::SI_SPILL_V256_RESTORE; break;
522 case 512: Opcode = AMDGPU::SI_SPILL_V512_RESTORE; break;
527 FrameInfo->setObjectAlignment(FrameIndex, 4);
528 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
529 .addFrameIndex(FrameIndex);
531 LLVMContext &Ctx = MF->getFunction()->getContext();
532 Ctx.emitError("SIInstrInfo::loadRegFromStackSlot - Do not know how to"
533 " restore register");
534 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
535 .addReg(AMDGPU::VGPR0);
539 /// \param @Offset Offset in bytes of the FrameIndex being spilled
540 unsigned SIInstrInfo::calculateLDSSpillAddress(MachineBasicBlock &MBB,
541 MachineBasicBlock::iterator MI,
542 RegScavenger *RS, unsigned TmpReg,
543 unsigned FrameOffset,
544 unsigned Size) const {
545 MachineFunction *MF = MBB.getParent();
546 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
547 const AMDGPUSubtarget &ST = MF->getTarget().getSubtarget<AMDGPUSubtarget>();
548 const SIRegisterInfo *TRI =
549 static_cast<const SIRegisterInfo*>(ST.getRegisterInfo());
550 DebugLoc DL = MBB.findDebugLoc(MI);
551 unsigned WorkGroupSize = MFI->getMaximumWorkGroupSize(*MF);
552 unsigned WavefrontSize = ST.getWavefrontSize();
554 unsigned TIDReg = MFI->getTIDReg();
555 if (!MFI->hasCalculatedTID()) {
556 MachineBasicBlock &Entry = MBB.getParent()->front();
557 MachineBasicBlock::iterator Insert = Entry.front();
558 DebugLoc DL = Insert->getDebugLoc();
560 TIDReg = RI.findUnusedVGPR(MF->getRegInfo());
561 if (TIDReg == AMDGPU::NoRegister)
565 if (MFI->getShaderType() == ShaderType::COMPUTE &&
566 WorkGroupSize > WavefrontSize) {
568 unsigned TIDIGXReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_X);
569 unsigned TIDIGYReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Y);
570 unsigned TIDIGZReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Z);
571 unsigned InputPtrReg =
572 TRI->getPreloadedValue(*MF, SIRegisterInfo::INPUT_PTR);
573 static const unsigned TIDIGRegs[3] = {
574 TIDIGXReg, TIDIGYReg, TIDIGZReg
576 for (unsigned Reg : TIDIGRegs) {
577 if (!Entry.isLiveIn(Reg))
578 Entry.addLiveIn(Reg);
581 RS->enterBasicBlock(&Entry);
582 unsigned STmp0 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
583 unsigned STmp1 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
584 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp0)
586 .addImm(SI::KernelInputOffsets::NGROUPS_Z);
587 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp1)
589 .addImm(SI::KernelInputOffsets::NGROUPS_Y);
591 // NGROUPS.X * NGROUPS.Y
592 BuildMI(Entry, Insert, DL, get(AMDGPU::S_MUL_I32), STmp1)
595 // (NGROUPS.X * NGROUPS.Y) * TIDIG.X
596 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MUL_U32_U24_e32), TIDReg)
599 // NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)
600 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MAD_U32_U24), TIDReg)
604 // (NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)) + TIDIG.Z
605 BuildMI(Entry, Insert, DL, get(AMDGPU::V_ADD_I32_e32), TIDReg)
610 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_LO_U32_B32_e64),
615 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_HI_U32_B32_e32),
621 BuildMI(Entry, Insert, DL, get(AMDGPU::V_LSHLREV_B32_e32),
625 MFI->setTIDReg(TIDReg);
628 // Add FrameIndex to LDS offset
629 unsigned LDSOffset = MFI->LDSSize + (FrameOffset * WorkGroupSize);
630 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_I32_e32), TmpReg)
637 void SIInstrInfo::insertNOPs(MachineBasicBlock::iterator MI,
646 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(AMDGPU::S_NOP))
651 bool SIInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
652 MachineBasicBlock &MBB = *MI->getParent();
653 DebugLoc DL = MBB.findDebugLoc(MI);
654 switch (MI->getOpcode()) {
655 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
657 case AMDGPU::SI_CONSTDATA_PTR: {
658 unsigned Reg = MI->getOperand(0).getReg();
659 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
660 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
662 BuildMI(MBB, MI, DL, get(AMDGPU::S_GETPC_B64), Reg);
664 // Add 32-bit offset from this instruction to the start of the constant data.
665 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADD_U32), RegLo)
667 .addTargetIndex(AMDGPU::TI_CONSTDATA_START)
668 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit);
669 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADDC_U32), RegHi)
672 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit)
673 .addReg(AMDGPU::SCC, RegState::Implicit);
674 MI->eraseFromParent();
677 case AMDGPU::SGPR_USE:
678 // This is just a placeholder for register allocation.
679 MI->eraseFromParent();
685 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
687 if (MI->getNumOperands() < 3)
690 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
691 AMDGPU::OpName::src0);
692 assert(Src0Idx != -1 && "Should always have src0 operand");
694 if (!MI->getOperand(Src0Idx).isReg())
697 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
698 AMDGPU::OpName::src1);
700 // Make sure it s legal to commute operands for VOP2.
701 if ((Src1Idx != -1) && isVOP2(MI->getOpcode()) &&
702 (!isOperandLegal(MI, Src0Idx, &MI->getOperand(Src1Idx)) ||
703 !isOperandLegal(MI, Src1Idx, &MI->getOperand(Src0Idx))))
706 if (Src1Idx != -1 && !MI->getOperand(Src1Idx).isReg()) {
707 // XXX: Commute instructions with FPImm operands
708 if (NewMI || !MI->getOperand(Src1Idx).isImm() ||
709 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
713 // XXX: Commute VOP3 instructions with abs and neg set .
714 const MachineOperand *Src0Mods = getNamedOperand(*MI,
715 AMDGPU::OpName::src0_modifiers);
716 const MachineOperand *Src1Mods = getNamedOperand(*MI,
717 AMDGPU::OpName::src1_modifiers);
718 const MachineOperand *Src2Mods = getNamedOperand(*MI,
719 AMDGPU::OpName::src2_modifiers);
721 if ((Src0Mods && Src0Mods->getImm()) ||
722 (Src1Mods && Src1Mods->getImm()) ||
723 (Src2Mods && Src2Mods->getImm()))
726 unsigned Reg = MI->getOperand(Src0Idx).getReg();
727 unsigned SubReg = MI->getOperand(Src0Idx).getSubReg();
728 MI->getOperand(Src0Idx).ChangeToImmediate(MI->getOperand(Src1Idx).getImm());
729 MI->getOperand(Src1Idx).ChangeToRegister(Reg, false);
730 MI->getOperand(Src1Idx).setSubReg(SubReg);
732 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
736 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
741 // This needs to be implemented because the source modifiers may be inserted
742 // between the true commutable operands, and the base
743 // TargetInstrInfo::commuteInstruction uses it.
744 bool SIInstrInfo::findCommutedOpIndices(MachineInstr *MI,
746 unsigned &SrcOpIdx2) const {
747 const MCInstrDesc &MCID = MI->getDesc();
748 if (!MCID.isCommutable())
751 unsigned Opc = MI->getOpcode();
752 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
756 // FIXME: Workaround TargetInstrInfo::commuteInstruction asserting on
758 if (!MI->getOperand(Src0Idx).isReg())
761 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
765 if (!MI->getOperand(Src1Idx).isReg())
773 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
774 MachineBasicBlock::iterator I,
776 unsigned SrcReg) const {
777 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
778 DstReg) .addReg(SrcReg);
781 bool SIInstrInfo::isMov(unsigned Opcode) const {
783 default: return false;
784 case AMDGPU::S_MOV_B32:
785 case AMDGPU::S_MOV_B64:
786 case AMDGPU::V_MOV_B32_e32:
787 case AMDGPU::V_MOV_B32_e64:
793 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
794 return RC != &AMDGPU::EXECRegRegClass;
798 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
799 AliasAnalysis *AA) const {
800 switch(MI->getOpcode()) {
801 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
802 case AMDGPU::S_MOV_B32:
803 case AMDGPU::S_MOV_B64:
804 case AMDGPU::V_MOV_B32_e32:
805 return MI->getOperand(1).isImm();
811 // Helper function generated by tablegen. We are wrapping this with
812 // an SIInstrInfo function that returns bool rather than int.
813 int isDS(uint16_t Opcode);
817 bool SIInstrInfo::isDS(uint16_t Opcode) const {
818 return ::AMDGPU::isDS(Opcode) != -1;
821 bool SIInstrInfo::isMIMG(uint16_t Opcode) const {
822 return get(Opcode).TSFlags & SIInstrFlags::MIMG;
825 bool SIInstrInfo::isSMRD(uint16_t Opcode) const {
826 return get(Opcode).TSFlags & SIInstrFlags::SMRD;
829 bool SIInstrInfo::isMUBUF(uint16_t Opcode) const {
830 return get(Opcode).TSFlags & SIInstrFlags::MUBUF;
833 bool SIInstrInfo::isMTBUF(uint16_t Opcode) const {
834 return get(Opcode).TSFlags & SIInstrFlags::MTBUF;
837 bool SIInstrInfo::isFLAT(uint16_t Opcode) const {
838 return get(Opcode).TSFlags & SIInstrFlags::FLAT;
841 bool SIInstrInfo::isVOP1(uint16_t Opcode) const {
842 return get(Opcode).TSFlags & SIInstrFlags::VOP1;
845 bool SIInstrInfo::isVOP2(uint16_t Opcode) const {
846 return get(Opcode).TSFlags & SIInstrFlags::VOP2;
849 bool SIInstrInfo::isVOP3(uint16_t Opcode) const {
850 return get(Opcode).TSFlags & SIInstrFlags::VOP3;
853 bool SIInstrInfo::isVOPC(uint16_t Opcode) const {
854 return get(Opcode).TSFlags & SIInstrFlags::VOPC;
857 bool SIInstrInfo::isSALUInstr(const MachineInstr &MI) const {
858 return get(MI.getOpcode()).TSFlags & SIInstrFlags::SALU;
861 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
862 int32_t Val = Imm.getSExtValue();
863 if (Val >= -16 && Val <= 64)
866 // The actual type of the operand does not seem to matter as long
867 // as the bits match one of the inline immediate values. For example:
869 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
870 // so it is a legal inline immediate.
872 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
873 // floating-point, so it is a legal inline immediate.
875 return (APInt::floatToBits(0.0f) == Imm) ||
876 (APInt::floatToBits(1.0f) == Imm) ||
877 (APInt::floatToBits(-1.0f) == Imm) ||
878 (APInt::floatToBits(0.5f) == Imm) ||
879 (APInt::floatToBits(-0.5f) == Imm) ||
880 (APInt::floatToBits(2.0f) == Imm) ||
881 (APInt::floatToBits(-2.0f) == Imm) ||
882 (APInt::floatToBits(4.0f) == Imm) ||
883 (APInt::floatToBits(-4.0f) == Imm);
886 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO) const {
888 return isInlineConstant(APInt(32, MO.getImm(), true));
891 APFloat FpImm = MO.getFPImm()->getValueAPF();
892 return isInlineConstant(FpImm.bitcastToAPInt());
898 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO) const {
899 return (MO.isImm() || MO.isFPImm()) && !isInlineConstant(MO);
902 static bool compareMachineOp(const MachineOperand &Op0,
903 const MachineOperand &Op1) {
904 if (Op0.getType() != Op1.getType())
907 switch (Op0.getType()) {
908 case MachineOperand::MO_Register:
909 return Op0.getReg() == Op1.getReg();
910 case MachineOperand::MO_Immediate:
911 return Op0.getImm() == Op1.getImm();
912 case MachineOperand::MO_FPImmediate:
913 return Op0.getFPImm() == Op1.getFPImm();
915 llvm_unreachable("Didn't expect to be comparing these operand types");
919 bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
920 const MachineOperand &MO) const {
921 const MCOperandInfo &OpInfo = get(MI->getOpcode()).OpInfo[OpNo];
923 assert(MO.isImm() || MO.isFPImm() || MO.isTargetIndex() || MO.isFI());
925 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
928 if (OpInfo.RegClass < 0)
931 if (isLiteralConstant(MO))
932 return RI.regClassCanUseLiteralConstant(OpInfo.RegClass);
934 return RI.regClassCanUseInlineConstant(OpInfo.RegClass);
937 bool SIInstrInfo::canFoldOffset(unsigned OffsetSize, unsigned AS) {
939 case AMDGPUAS::GLOBAL_ADDRESS: {
940 // MUBUF instructions a 12-bit offset in bytes.
941 return isUInt<12>(OffsetSize);
943 case AMDGPUAS::CONSTANT_ADDRESS: {
944 // SMRD instructions have an 8-bit offset in dwords.
945 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
947 case AMDGPUAS::LOCAL_ADDRESS:
948 case AMDGPUAS::REGION_ADDRESS: {
949 // The single offset versions have a 16-bit offset in bytes.
950 return isUInt<16>(OffsetSize);
952 case AMDGPUAS::PRIVATE_ADDRESS:
953 // Indirect register addressing does not use any offsets.
959 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
960 return AMDGPU::getVOPe32(Opcode) != -1;
963 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
964 // The src0_modifier operand is present on all instructions
965 // that have modifiers.
967 return AMDGPU::getNamedOperandIdx(Opcode,
968 AMDGPU::OpName::src0_modifiers) != -1;
971 bool SIInstrInfo::usesConstantBus(const MachineRegisterInfo &MRI,
972 const MachineOperand &MO) const {
973 // Literal constants use the constant bus.
974 if (isLiteralConstant(MO))
977 if (!MO.isReg() || !MO.isUse())
980 if (TargetRegisterInfo::isVirtualRegister(MO.getReg()))
981 return RI.isSGPRClass(MRI.getRegClass(MO.getReg()));
983 // FLAT_SCR is just an SGPR pair.
984 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
987 // EXEC register uses the constant bus.
988 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
991 // SGPRs use the constant bus
992 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
994 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
995 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
1002 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
1003 StringRef &ErrInfo) const {
1004 uint16_t Opcode = MI->getOpcode();
1005 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1006 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
1007 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
1008 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
1010 // Make sure the number of operands is correct.
1011 const MCInstrDesc &Desc = get(Opcode);
1012 if (!Desc.isVariadic() &&
1013 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
1014 ErrInfo = "Instruction has wrong number of operands.";
1018 // Make sure the register classes are correct
1019 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
1020 switch (Desc.OpInfo[i].OperandType) {
1021 case MCOI::OPERAND_REGISTER: {
1022 if ((MI->getOperand(i).isImm() || MI->getOperand(i).isFPImm()) &&
1023 !isImmOperandLegal(MI, i, MI->getOperand(i))) {
1024 ErrInfo = "Illegal immediate value for operand.";
1029 case MCOI::OPERAND_IMMEDIATE:
1030 // Check if this operand is an immediate.
1031 // FrameIndex operands will be replaced by immediates, so they are
1033 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFPImm() &&
1034 !MI->getOperand(i).isFI()) {
1035 ErrInfo = "Expected immediate, but got non-immediate";
1043 if (!MI->getOperand(i).isReg())
1046 int RegClass = Desc.OpInfo[i].RegClass;
1047 if (RegClass != -1) {
1048 unsigned Reg = MI->getOperand(i).getReg();
1049 if (TargetRegisterInfo::isVirtualRegister(Reg))
1052 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
1053 if (!RC->contains(Reg)) {
1054 ErrInfo = "Operand has incorrect register class.";
1062 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
1063 unsigned ConstantBusCount = 0;
1064 unsigned SGPRUsed = AMDGPU::NoRegister;
1065 for (int i = 0, e = MI->getNumOperands(); i != e; ++i) {
1066 const MachineOperand &MO = MI->getOperand(i);
1067 if (usesConstantBus(MRI, MO)) {
1069 if (MO.getReg() != SGPRUsed)
1071 SGPRUsed = MO.getReg();
1077 if (ConstantBusCount > 1) {
1078 ErrInfo = "VOP* instruction uses the constant bus more than once";
1083 // Verify SRC1 for VOP2 and VOPC
1084 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
1085 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1086 if (Src1.isImm() || Src1.isFPImm()) {
1087 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
1093 if (isVOP3(Opcode)) {
1094 if (Src0Idx != -1 && isLiteralConstant(MI->getOperand(Src0Idx))) {
1095 ErrInfo = "VOP3 src0 cannot be a literal constant.";
1098 if (Src1Idx != -1 && isLiteralConstant(MI->getOperand(Src1Idx))) {
1099 ErrInfo = "VOP3 src1 cannot be a literal constant.";
1102 if (Src2Idx != -1 && isLiteralConstant(MI->getOperand(Src2Idx))) {
1103 ErrInfo = "VOP3 src2 cannot be a literal constant.";
1108 // Verify misc. restrictions on specific instructions.
1109 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
1110 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
1111 const MachineOperand &Src0 = MI->getOperand(Src0Idx);
1112 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1113 const MachineOperand &Src2 = MI->getOperand(Src2Idx);
1114 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
1115 if (!compareMachineOp(Src0, Src1) &&
1116 !compareMachineOp(Src0, Src2)) {
1117 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
1126 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
1127 switch (MI.getOpcode()) {
1128 default: return AMDGPU::INSTRUCTION_LIST_END;
1129 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
1130 case AMDGPU::COPY: return AMDGPU::COPY;
1131 case AMDGPU::PHI: return AMDGPU::PHI;
1132 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
1133 case AMDGPU::S_MOV_B32:
1134 return MI.getOperand(1).isReg() ?
1135 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
1136 case AMDGPU::S_ADD_I32:
1137 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
1138 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
1139 case AMDGPU::S_SUB_I32:
1140 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
1141 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
1142 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
1143 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
1144 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
1145 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
1146 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
1147 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
1148 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
1149 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
1150 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
1151 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
1152 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
1153 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
1154 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
1155 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
1156 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
1157 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
1158 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
1159 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
1160 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
1161 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
1162 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
1163 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
1164 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
1165 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
1166 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
1167 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
1168 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
1169 case AMDGPU::S_LOAD_DWORD_IMM:
1170 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1171 case AMDGPU::S_LOAD_DWORDX2_IMM:
1172 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1173 case AMDGPU::S_LOAD_DWORDX4_IMM:
1174 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1175 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e32;
1176 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
1177 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
1181 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
1182 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
1185 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
1186 unsigned OpNo) const {
1187 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1188 const MCInstrDesc &Desc = get(MI.getOpcode());
1189 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
1190 Desc.OpInfo[OpNo].RegClass == -1)
1191 return MRI.getRegClass(MI.getOperand(OpNo).getReg());
1193 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
1194 return RI.getRegClass(RCID);
1197 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
1198 switch (MI.getOpcode()) {
1200 case AMDGPU::REG_SEQUENCE:
1202 case AMDGPU::INSERT_SUBREG:
1203 return RI.hasVGPRs(getOpRegClass(MI, 0));
1205 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
1209 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
1210 MachineBasicBlock::iterator I = MI;
1211 MachineOperand &MO = MI->getOperand(OpIdx);
1212 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1213 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
1214 const TargetRegisterClass *RC = RI.getRegClass(RCID);
1215 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1217 Opcode = AMDGPU::COPY;
1218 } else if (RI.isSGPRClass(RC)) {
1219 Opcode = AMDGPU::S_MOV_B32;
1222 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
1223 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC)) {
1224 VRC = &AMDGPU::VReg_64RegClass;
1226 VRC = &AMDGPU::VReg_32RegClass;
1228 unsigned Reg = MRI.createVirtualRegister(VRC);
1229 BuildMI(*MI->getParent(), I, MI->getParent()->findDebugLoc(I), get(Opcode),
1230 Reg).addOperand(MO);
1231 MO.ChangeToRegister(Reg, false);
1234 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
1235 MachineRegisterInfo &MRI,
1236 MachineOperand &SuperReg,
1237 const TargetRegisterClass *SuperRC,
1239 const TargetRegisterClass *SubRC)
1241 assert(SuperReg.isReg());
1243 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
1244 unsigned SubReg = MRI.createVirtualRegister(SubRC);
1246 // Just in case the super register is itself a sub-register, copy it to a new
1247 // value so we don't need to worry about merging its subreg index with the
1248 // SubIdx passed to this function. The register coalescer should be able to
1249 // eliminate this extra copy.
1250 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1252 .addOperand(SuperReg);
1254 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1256 .addReg(NewSuperReg, 0, SubIdx);
1260 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
1261 MachineBasicBlock::iterator MII,
1262 MachineRegisterInfo &MRI,
1264 const TargetRegisterClass *SuperRC,
1266 const TargetRegisterClass *SubRC) const {
1268 // XXX - Is there a better way to do this?
1269 if (SubIdx == AMDGPU::sub0)
1270 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
1271 if (SubIdx == AMDGPU::sub1)
1272 return MachineOperand::CreateImm(Op.getImm() >> 32);
1274 llvm_unreachable("Unhandled register index for immediate");
1277 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
1279 return MachineOperand::CreateReg(SubReg, false);
1282 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
1283 MachineBasicBlock::iterator MI,
1284 MachineRegisterInfo &MRI,
1285 const TargetRegisterClass *RC,
1286 const MachineOperand &Op) const {
1287 MachineBasicBlock *MBB = MI->getParent();
1288 DebugLoc DL = MI->getDebugLoc();
1289 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1290 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1291 unsigned Dst = MRI.createVirtualRegister(RC);
1293 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1295 .addImm(Op.getImm() & 0xFFFFFFFF);
1296 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1298 .addImm(Op.getImm() >> 32);
1300 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
1302 .addImm(AMDGPU::sub0)
1304 .addImm(AMDGPU::sub1);
1306 Worklist.push_back(Lo);
1307 Worklist.push_back(Hi);
1312 bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
1313 const MachineOperand *MO) const {
1314 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1315 const MCInstrDesc &InstDesc = get(MI->getOpcode());
1316 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
1317 const TargetRegisterClass *DefinedRC =
1318 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
1320 MO = &MI->getOperand(OpIdx);
1322 if (usesConstantBus(MRI, *MO)) {
1324 MO->isReg() ? MO->getReg() : (unsigned)AMDGPU::NoRegister;
1325 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1328 if (usesConstantBus(MRI, MI->getOperand(i)) &&
1329 MI->getOperand(i).isReg() && MI->getOperand(i).getReg() != SGPRUsed) {
1337 const TargetRegisterClass *RC = MRI.getRegClass(MO->getReg());
1338 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass));
1342 // Handle non-register types that are treated like immediates.
1343 assert(MO->isImm() || MO->isFPImm() || MO->isTargetIndex() || MO->isFI());
1346 // This operand expects an immediate.
1350 return isImmOperandLegal(MI, OpIdx, *MO);
1353 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
1354 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1356 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1357 AMDGPU::OpName::src0);
1358 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1359 AMDGPU::OpName::src1);
1360 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1361 AMDGPU::OpName::src2);
1364 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
1366 if (!isOperandLegal(MI, Src0Idx))
1367 legalizeOpWithMove(MI, Src0Idx);
1370 if (isOperandLegal(MI, Src1Idx))
1373 // Usually src0 of VOP2 instructions allow more types of inputs
1374 // than src1, so try to commute the instruction to decrease our
1375 // chances of having to insert a MOV instruction to legalize src1.
1376 if (MI->isCommutable()) {
1377 if (commuteInstruction(MI))
1378 // If we are successful in commuting, then we know MI is legal, so
1383 legalizeOpWithMove(MI, Src1Idx);
1387 // XXX - Do any VOP3 instructions read VCC?
1389 if (isVOP3(MI->getOpcode())) {
1390 int VOP3Idx[3] = { Src0Idx, Src1Idx, Src2Idx };
1392 // Find the one SGPR operand we are allowed to use.
1393 unsigned SGPRReg = findUsedSGPR(MI, VOP3Idx);
1395 for (unsigned i = 0; i < 3; ++i) {
1396 int Idx = VOP3Idx[i];
1399 MachineOperand &MO = MI->getOperand(Idx);
1402 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
1403 continue; // VGPRs are legal
1405 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
1407 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
1408 SGPRReg = MO.getReg();
1409 // We can use one SGPR in each VOP3 instruction.
1412 } else if (!isLiteralConstant(MO)) {
1413 // If it is not a register and not a literal constant, then it must be
1414 // an inline constant which is always legal.
1417 // If we make it this far, then the operand is not legal and we must
1419 legalizeOpWithMove(MI, Idx);
1423 // Legalize REG_SEQUENCE and PHI
1424 // The register class of the operands much be the same type as the register
1425 // class of the output.
1426 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
1427 MI->getOpcode() == AMDGPU::PHI) {
1428 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
1429 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1430 if (!MI->getOperand(i).isReg() ||
1431 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1433 const TargetRegisterClass *OpRC =
1434 MRI.getRegClass(MI->getOperand(i).getReg());
1435 if (RI.hasVGPRs(OpRC)) {
1442 // If any of the operands are VGPR registers, then they all most be
1443 // otherwise we will create illegal VGPR->SGPR copies when legalizing
1445 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
1448 VRC = RI.getEquivalentVGPRClass(SRC);
1455 // Update all the operands so they have the same type.
1456 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1457 if (!MI->getOperand(i).isReg() ||
1458 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1460 unsigned DstReg = MRI.createVirtualRegister(RC);
1461 MachineBasicBlock *InsertBB;
1462 MachineBasicBlock::iterator Insert;
1463 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
1464 InsertBB = MI->getParent();
1467 // MI is a PHI instruction.
1468 InsertBB = MI->getOperand(i + 1).getMBB();
1469 Insert = InsertBB->getFirstTerminator();
1471 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
1472 get(AMDGPU::COPY), DstReg)
1473 .addOperand(MI->getOperand(i));
1474 MI->getOperand(i).setReg(DstReg);
1478 // Legalize INSERT_SUBREG
1479 // src0 must have the same register class as dst
1480 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
1481 unsigned Dst = MI->getOperand(0).getReg();
1482 unsigned Src0 = MI->getOperand(1).getReg();
1483 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
1484 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
1485 if (DstRC != Src0RC) {
1486 MachineBasicBlock &MBB = *MI->getParent();
1487 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
1488 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
1490 MI->getOperand(1).setReg(NewSrc0);
1495 // Legalize MUBUF* instructions
1496 // FIXME: If we start using the non-addr64 instructions for compute, we
1497 // may need to legalize them here.
1499 AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::srsrc);
1500 if (SRsrcIdx != -1) {
1501 // We have an MUBUF instruction
1502 MachineOperand *SRsrc = &MI->getOperand(SRsrcIdx);
1503 unsigned SRsrcRC = get(MI->getOpcode()).OpInfo[SRsrcIdx].RegClass;
1504 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
1505 RI.getRegClass(SRsrcRC))) {
1506 // The operands are legal.
1507 // FIXME: We may need to legalize operands besided srsrc.
1511 MachineBasicBlock &MBB = *MI->getParent();
1512 // Extract the the ptr from the resource descriptor.
1514 // SRsrcPtrLo = srsrc:sub0
1515 unsigned SRsrcPtrLo = buildExtractSubReg(MI, MRI, *SRsrc,
1516 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VReg_32RegClass);
1518 // SRsrcPtrHi = srsrc:sub1
1519 unsigned SRsrcPtrHi = buildExtractSubReg(MI, MRI, *SRsrc,
1520 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VReg_32RegClass);
1522 // Create an empty resource descriptor
1523 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1524 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1525 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1526 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1529 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1533 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1534 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1536 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1538 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1539 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1541 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1543 // NewSRsrc = {Zero64, SRsrcFormat}
1544 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1547 .addImm(AMDGPU::sub0_sub1)
1548 .addReg(SRsrcFormatLo)
1549 .addImm(AMDGPU::sub2)
1550 .addReg(SRsrcFormatHi)
1551 .addImm(AMDGPU::sub3);
1553 MachineOperand *VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1554 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
1555 unsigned NewVAddrLo;
1556 unsigned NewVAddrHi;
1558 // This is already an ADDR64 instruction so we need to add the pointer
1559 // extracted from the resource descriptor to the current value of VAddr.
1560 NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1561 NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1563 // NewVaddrLo = SRsrcPtrLo + VAddr:sub0
1564 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1567 .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
1568 .addReg(AMDGPU::VCC, RegState::ImplicitDefine);
1570 // NewVaddrHi = SRsrcPtrHi + VAddr:sub1
1571 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1574 .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
1575 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1576 .addReg(AMDGPU::VCC, RegState::Implicit);
1579 // This instructions is the _OFFSET variant, so we need to convert it to
1581 MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
1582 MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
1583 MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
1584 assert(SOffset->isImm() && SOffset->getImm() == 0 && "Legalizing MUBUF "
1585 "with non-zero soffset is not implemented");
1588 // Create the new instruction.
1589 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
1590 MachineInstr *Addr64 =
1591 BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
1594 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
1595 // This will be replaced later
1596 // with the new value of vaddr.
1597 .addOperand(*Offset);
1599 MI->removeFromParent();
1602 NewVAddrLo = SRsrcPtrLo;
1603 NewVAddrHi = SRsrcPtrHi;
1604 VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1605 SRsrc = getNamedOperand(*MI, AMDGPU::OpName::srsrc);
1608 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1609 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1612 .addImm(AMDGPU::sub0)
1614 .addImm(AMDGPU::sub1);
1617 // Update the instruction to use NewVaddr
1618 VAddr->setReg(NewVAddr);
1619 // Update the instruction to use NewSRsrc
1620 SRsrc->setReg(NewSRsrc);
1624 void SIInstrInfo::splitSMRD(MachineInstr *MI,
1625 const TargetRegisterClass *HalfRC,
1626 unsigned HalfImmOp, unsigned HalfSGPROp,
1627 MachineInstr *&Lo, MachineInstr *&Hi) const {
1629 DebugLoc DL = MI->getDebugLoc();
1630 MachineBasicBlock *MBB = MI->getParent();
1631 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1632 unsigned RegLo = MRI.createVirtualRegister(HalfRC);
1633 unsigned RegHi = MRI.createVirtualRegister(HalfRC);
1634 unsigned HalfSize = HalfRC->getSize();
1635 const MachineOperand *OffOp =
1636 getNamedOperand(*MI, AMDGPU::OpName::offset);
1637 const MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
1640 // Handle the _IMM variant
1641 unsigned LoOffset = OffOp->getImm();
1642 unsigned HiOffset = LoOffset + (HalfSize / 4);
1643 Lo = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegLo)
1647 if (!isUInt<8>(HiOffset)) {
1648 unsigned OffsetSGPR =
1649 MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1650 BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32), OffsetSGPR)
1651 .addImm(HiOffset << 2); // The immediate offset is in dwords,
1652 // but offset in register is in bytes.
1653 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
1655 .addReg(OffsetSGPR);
1657 Hi = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegHi)
1662 // Handle the _SGPR variant
1663 MachineOperand *SOff = getNamedOperand(*MI, AMDGPU::OpName::soff);
1664 Lo = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegLo)
1667 unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1668 BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
1671 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp))
1673 .addReg(OffsetSGPR);
1676 unsigned SubLo, SubHi;
1679 SubLo = AMDGPU::sub0;
1680 SubHi = AMDGPU::sub1;
1683 SubLo = AMDGPU::sub0_sub1;
1684 SubHi = AMDGPU::sub2_sub3;
1687 SubLo = AMDGPU::sub0_sub1_sub2_sub3;
1688 SubHi = AMDGPU::sub4_sub5_sub6_sub7;
1691 SubLo = AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
1692 SubHi = AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15;
1695 llvm_unreachable("Unhandled HalfSize");
1698 BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE))
1699 .addOperand(MI->getOperand(0))
1706 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
1707 MachineBasicBlock *MBB = MI->getParent();
1708 switch (MI->getOpcode()) {
1709 case AMDGPU::S_LOAD_DWORD_IMM:
1710 case AMDGPU::S_LOAD_DWORD_SGPR:
1711 case AMDGPU::S_LOAD_DWORDX2_IMM:
1712 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1713 case AMDGPU::S_LOAD_DWORDX4_IMM:
1714 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1715 unsigned NewOpcode = getVALUOp(*MI);
1719 if (MI->getOperand(2).isReg()) {
1720 RegOffset = MI->getOperand(2).getReg();
1723 assert(MI->getOperand(2).isImm());
1724 // SMRD instructions take a dword offsets and MUBUF instructions
1725 // take a byte offset.
1726 ImmOffset = MI->getOperand(2).getImm() << 2;
1727 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1728 if (isUInt<12>(ImmOffset)) {
1729 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1733 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1740 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1741 unsigned DWord0 = RegOffset;
1742 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1743 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1744 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1746 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
1748 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
1749 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1750 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
1751 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1752 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
1754 .addImm(AMDGPU::sub0)
1756 .addImm(AMDGPU::sub1)
1758 .addImm(AMDGPU::sub2)
1760 .addImm(AMDGPU::sub3);
1761 MI->setDesc(get(NewOpcode));
1762 if (MI->getOperand(2).isReg()) {
1763 MI->getOperand(2).setReg(MI->getOperand(1).getReg());
1765 MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
1767 MI->getOperand(1).setReg(SRsrc);
1768 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
1770 const TargetRegisterClass *NewDstRC =
1771 RI.getRegClass(get(NewOpcode).OpInfo[0].RegClass);
1773 unsigned DstReg = MI->getOperand(0).getReg();
1774 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1775 MRI.replaceRegWith(DstReg, NewDstReg);
1778 case AMDGPU::S_LOAD_DWORDX8_IMM:
1779 case AMDGPU::S_LOAD_DWORDX8_SGPR: {
1780 MachineInstr *Lo, *Hi;
1781 splitSMRD(MI, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM,
1782 AMDGPU::S_LOAD_DWORDX4_SGPR, Lo, Hi);
1783 MI->eraseFromParent();
1784 moveSMRDToVALU(Lo, MRI);
1785 moveSMRDToVALU(Hi, MRI);
1789 case AMDGPU::S_LOAD_DWORDX16_IMM:
1790 case AMDGPU::S_LOAD_DWORDX16_SGPR: {
1791 MachineInstr *Lo, *Hi;
1792 splitSMRD(MI, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM,
1793 AMDGPU::S_LOAD_DWORDX8_SGPR, Lo, Hi);
1794 MI->eraseFromParent();
1795 moveSMRDToVALU(Lo, MRI);
1796 moveSMRDToVALU(Hi, MRI);
1802 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
1803 SmallVector<MachineInstr *, 128> Worklist;
1804 Worklist.push_back(&TopInst);
1806 while (!Worklist.empty()) {
1807 MachineInstr *Inst = Worklist.pop_back_val();
1808 MachineBasicBlock *MBB = Inst->getParent();
1809 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1811 unsigned Opcode = Inst->getOpcode();
1812 unsigned NewOpcode = getVALUOp(*Inst);
1814 // Handle some special cases
1817 if (isSMRD(Inst->getOpcode())) {
1818 moveSMRDToVALU(Inst, MRI);
1821 case AMDGPU::S_MOV_B64: {
1822 DebugLoc DL = Inst->getDebugLoc();
1824 // If the source operand is a register we can replace this with a
1826 if (Inst->getOperand(1).isReg()) {
1827 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
1828 .addOperand(Inst->getOperand(0))
1829 .addOperand(Inst->getOperand(1));
1830 Worklist.push_back(Copy);
1832 // Otherwise, we need to split this into two movs, because there is
1833 // no 64-bit VALU move instruction.
1834 unsigned Reg = Inst->getOperand(0).getReg();
1835 unsigned Dst = split64BitImm(Worklist,
1838 MRI.getRegClass(Reg),
1839 Inst->getOperand(1));
1840 MRI.replaceRegWith(Reg, Dst);
1842 Inst->eraseFromParent();
1845 case AMDGPU::S_AND_B64:
1846 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
1847 Inst->eraseFromParent();
1850 case AMDGPU::S_OR_B64:
1851 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
1852 Inst->eraseFromParent();
1855 case AMDGPU::S_XOR_B64:
1856 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
1857 Inst->eraseFromParent();
1860 case AMDGPU::S_NOT_B64:
1861 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
1862 Inst->eraseFromParent();
1865 case AMDGPU::S_BCNT1_I32_B64:
1866 splitScalar64BitBCNT(Worklist, Inst);
1867 Inst->eraseFromParent();
1870 case AMDGPU::S_BFE_U64:
1871 case AMDGPU::S_BFE_I64:
1872 case AMDGPU::S_BFM_B64:
1873 llvm_unreachable("Moving this op to VALU not implemented");
1876 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
1877 // We cannot move this instruction to the VALU, so we should try to
1878 // legalize its operands instead.
1879 legalizeOperands(Inst);
1883 // Use the new VALU Opcode.
1884 const MCInstrDesc &NewDesc = get(NewOpcode);
1885 Inst->setDesc(NewDesc);
1887 // Remove any references to SCC. Vector instructions can't read from it, and
1888 // We're just about to add the implicit use / defs of VCC, and we don't want
1890 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
1891 MachineOperand &Op = Inst->getOperand(i);
1892 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
1893 Inst->RemoveOperand(i);
1896 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
1897 // We are converting these to a BFE, so we need to add the missing
1898 // operands for the size and offset.
1899 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
1900 Inst->addOperand(MachineOperand::CreateImm(0));
1901 Inst->addOperand(MachineOperand::CreateImm(Size));
1903 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
1904 // The VALU version adds the second operand to the result, so insert an
1906 Inst->addOperand(MachineOperand::CreateImm(0));
1909 addDescImplicitUseDef(NewDesc, Inst);
1911 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
1912 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
1913 // If we need to move this to VGPRs, we need to unpack the second operand
1914 // back into the 2 separate ones for bit offset and width.
1915 assert(OffsetWidthOp.isImm() &&
1916 "Scalar BFE is only implemented for constant width and offset");
1917 uint32_t Imm = OffsetWidthOp.getImm();
1919 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
1920 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
1921 Inst->RemoveOperand(2); // Remove old immediate.
1922 Inst->addOperand(MachineOperand::CreateImm(Offset));
1923 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
1926 // Update the destination register class.
1928 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
1931 // For target instructions, getOpRegClass just returns the virtual
1932 // register class associated with the operand, so we need to find an
1933 // equivalent VGPR register class in order to move the instruction to the
1937 case AMDGPU::REG_SEQUENCE:
1938 case AMDGPU::INSERT_SUBREG:
1939 if (RI.hasVGPRs(NewDstRC))
1941 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
1949 unsigned DstReg = Inst->getOperand(0).getReg();
1950 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1951 MRI.replaceRegWith(DstReg, NewDstReg);
1953 // Legalize the operands
1954 legalizeOperands(Inst);
1956 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
1957 E = MRI.use_end(); I != E; ++I) {
1958 MachineInstr &UseMI = *I->getParent();
1959 if (!canReadVGPR(UseMI, I.getOperandNo())) {
1960 Worklist.push_back(&UseMI);
1966 //===----------------------------------------------------------------------===//
1967 // Indirect addressing callbacks
1968 //===----------------------------------------------------------------------===//
1970 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
1971 unsigned Channel) const {
1972 assert(Channel == 0);
1976 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
1977 return &AMDGPU::VReg_32RegClass;
1980 void SIInstrInfo::splitScalar64BitUnaryOp(
1981 SmallVectorImpl<MachineInstr *> &Worklist,
1983 unsigned Opcode) const {
1984 MachineBasicBlock &MBB = *Inst->getParent();
1985 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1987 MachineOperand &Dest = Inst->getOperand(0);
1988 MachineOperand &Src0 = Inst->getOperand(1);
1989 DebugLoc DL = Inst->getDebugLoc();
1991 MachineBasicBlock::iterator MII = Inst;
1993 const MCInstrDesc &InstDesc = get(Opcode);
1994 const TargetRegisterClass *Src0RC = Src0.isReg() ?
1995 MRI.getRegClass(Src0.getReg()) :
1996 &AMDGPU::SGPR_32RegClass;
1998 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2000 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2001 AMDGPU::sub0, Src0SubRC);
2003 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2004 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2006 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2007 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2008 .addOperand(SrcReg0Sub0);
2010 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2011 AMDGPU::sub1, Src0SubRC);
2013 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2014 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2015 .addOperand(SrcReg0Sub1);
2017 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2018 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2020 .addImm(AMDGPU::sub0)
2022 .addImm(AMDGPU::sub1);
2024 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2026 // Try to legalize the operands in case we need to swap the order to keep it
2028 Worklist.push_back(LoHalf);
2029 Worklist.push_back(HiHalf);
2032 void SIInstrInfo::splitScalar64BitBinaryOp(
2033 SmallVectorImpl<MachineInstr *> &Worklist,
2035 unsigned Opcode) const {
2036 MachineBasicBlock &MBB = *Inst->getParent();
2037 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2039 MachineOperand &Dest = Inst->getOperand(0);
2040 MachineOperand &Src0 = Inst->getOperand(1);
2041 MachineOperand &Src1 = Inst->getOperand(2);
2042 DebugLoc DL = Inst->getDebugLoc();
2044 MachineBasicBlock::iterator MII = Inst;
2046 const MCInstrDesc &InstDesc = get(Opcode);
2047 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2048 MRI.getRegClass(Src0.getReg()) :
2049 &AMDGPU::SGPR_32RegClass;
2051 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2052 const TargetRegisterClass *Src1RC = Src1.isReg() ?
2053 MRI.getRegClass(Src1.getReg()) :
2054 &AMDGPU::SGPR_32RegClass;
2056 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
2058 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2059 AMDGPU::sub0, Src0SubRC);
2060 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2061 AMDGPU::sub0, Src1SubRC);
2063 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2064 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2066 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2067 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2068 .addOperand(SrcReg0Sub0)
2069 .addOperand(SrcReg1Sub0);
2071 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2072 AMDGPU::sub1, Src0SubRC);
2073 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2074 AMDGPU::sub1, Src1SubRC);
2076 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2077 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2078 .addOperand(SrcReg0Sub1)
2079 .addOperand(SrcReg1Sub1);
2081 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2082 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2084 .addImm(AMDGPU::sub0)
2086 .addImm(AMDGPU::sub1);
2088 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2090 // Try to legalize the operands in case we need to swap the order to keep it
2092 Worklist.push_back(LoHalf);
2093 Worklist.push_back(HiHalf);
2096 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
2097 MachineInstr *Inst) const {
2098 MachineBasicBlock &MBB = *Inst->getParent();
2099 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2101 MachineBasicBlock::iterator MII = Inst;
2102 DebugLoc DL = Inst->getDebugLoc();
2104 MachineOperand &Dest = Inst->getOperand(0);
2105 MachineOperand &Src = Inst->getOperand(1);
2107 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e32);
2108 const TargetRegisterClass *SrcRC = Src.isReg() ?
2109 MRI.getRegClass(Src.getReg()) :
2110 &AMDGPU::SGPR_32RegClass;
2112 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2113 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2115 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
2117 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2118 AMDGPU::sub0, SrcSubRC);
2119 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2120 AMDGPU::sub1, SrcSubRC);
2122 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
2123 .addOperand(SrcRegSub0)
2126 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
2127 .addOperand(SrcRegSub1)
2130 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2132 Worklist.push_back(First);
2133 Worklist.push_back(Second);
2136 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
2137 MachineInstr *Inst) const {
2138 // Add the implict and explicit register definitions.
2139 if (NewDesc.ImplicitUses) {
2140 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
2141 unsigned Reg = NewDesc.ImplicitUses[i];
2142 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
2146 if (NewDesc.ImplicitDefs) {
2147 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
2148 unsigned Reg = NewDesc.ImplicitDefs[i];
2149 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
2154 unsigned SIInstrInfo::findUsedSGPR(const MachineInstr *MI,
2155 int OpIndices[3]) const {
2156 const MCInstrDesc &Desc = get(MI->getOpcode());
2158 // Find the one SGPR operand we are allowed to use.
2159 unsigned SGPRReg = AMDGPU::NoRegister;
2161 // First we need to consider the instruction's operand requirements before
2162 // legalizing. Some operands are required to be SGPRs, such as implicit uses
2163 // of VCC, but we are still bound by the constant bus requirement to only use
2166 // If the operand's class is an SGPR, we can never move it.
2168 for (const MachineOperand &MO : MI->implicit_operands()) {
2169 // We only care about reads.
2173 if (MO.getReg() == AMDGPU::VCC)
2176 if (MO.getReg() == AMDGPU::FLAT_SCR)
2177 return AMDGPU::FLAT_SCR;
2180 unsigned UsedSGPRs[3] = { AMDGPU::NoRegister };
2181 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
2183 for (unsigned i = 0; i < 3; ++i) {
2184 int Idx = OpIndices[i];
2188 const MachineOperand &MO = MI->getOperand(Idx);
2189 if (RI.isSGPRClassID(Desc.OpInfo[Idx].RegClass))
2190 SGPRReg = MO.getReg();
2192 if (MO.isReg() && RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
2193 UsedSGPRs[i] = MO.getReg();
2196 if (SGPRReg != AMDGPU::NoRegister)
2199 // We don't have a required SGPR operand, so we have a bit more freedom in
2200 // selecting operands to move.
2202 // Try to select the most used SGPR. If an SGPR is equal to one of the
2203 // others, we choose that.
2206 // V_FMA_F32 v0, s0, s0, s0 -> No moves
2207 // V_FMA_F32 v0, s0, s1, s0 -> Move s1
2209 if (UsedSGPRs[0] != AMDGPU::NoRegister) {
2210 if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
2211 SGPRReg = UsedSGPRs[0];
2214 if (SGPRReg == AMDGPU::NoRegister && UsedSGPRs[1] != AMDGPU::NoRegister) {
2215 if (UsedSGPRs[1] == UsedSGPRs[2])
2216 SGPRReg = UsedSGPRs[1];
2222 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
2223 MachineBasicBlock *MBB,
2224 MachineBasicBlock::iterator I,
2226 unsigned Address, unsigned OffsetReg) const {
2227 const DebugLoc &DL = MBB->findDebugLoc(I);
2228 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
2229 getIndirectIndexBegin(*MBB->getParent()));
2231 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
2232 .addReg(IndirectBaseReg, RegState::Define)
2233 .addOperand(I->getOperand(0))
2234 .addReg(IndirectBaseReg)
2240 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
2241 MachineBasicBlock *MBB,
2242 MachineBasicBlock::iterator I,
2244 unsigned Address, unsigned OffsetReg) const {
2245 const DebugLoc &DL = MBB->findDebugLoc(I);
2246 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
2247 getIndirectIndexBegin(*MBB->getParent()));
2249 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
2250 .addOperand(I->getOperand(0))
2251 .addOperand(I->getOperand(1))
2252 .addReg(IndirectBaseReg)
2258 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
2259 const MachineFunction &MF) const {
2260 int End = getIndirectIndexEnd(MF);
2261 int Begin = getIndirectIndexBegin(MF);
2267 for (int Index = Begin; Index <= End; ++Index)
2268 Reserved.set(AMDGPU::VReg_32RegClass.getRegister(Index));
2270 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
2271 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
2273 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
2274 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
2276 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
2277 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
2279 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
2280 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
2282 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
2283 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));
2286 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
2287 unsigned OperandName) const {
2288 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
2292 return &MI.getOperand(Idx);