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/MC/MCInstrDesc.h"
28 SIInstrInfo::SIInstrInfo(const AMDGPUSubtarget &st)
29 : AMDGPUInstrInfo(st),
32 //===----------------------------------------------------------------------===//
33 // TargetInstrInfo callbacks
34 //===----------------------------------------------------------------------===//
36 static unsigned getNumOperandsNoGlue(SDNode *Node) {
37 unsigned N = Node->getNumOperands();
38 while (N && Node->getOperand(N - 1).getValueType() == MVT::Glue)
43 static SDValue findChainOperand(SDNode *Load) {
44 SDValue LastOp = Load->getOperand(getNumOperandsNoGlue(Load) - 1);
45 assert(LastOp.getValueType() == MVT::Other && "Chain missing from load node");
49 /// \brief Returns true if both nodes have the same value for the given
50 /// operand \p Op, or if both nodes do not have this operand.
51 static bool nodesHaveSameOperandValue(SDNode *N0, SDNode* N1, unsigned OpName) {
52 unsigned Opc0 = N0->getMachineOpcode();
53 unsigned Opc1 = N1->getMachineOpcode();
55 int Op0Idx = AMDGPU::getNamedOperandIdx(Opc0, OpName);
56 int Op1Idx = AMDGPU::getNamedOperandIdx(Opc1, OpName);
58 if (Op0Idx == -1 && Op1Idx == -1)
62 if ((Op0Idx == -1 && Op1Idx != -1) ||
63 (Op1Idx == -1 && Op0Idx != -1))
66 // getNamedOperandIdx returns the index for the MachineInstr's operands,
67 // which includes the result as the first operand. We are indexing into the
68 // MachineSDNode's operands, so we need to skip the result operand to get
73 return N0->getOperand(Op0Idx) == N1->getOperand(Op1Idx);
76 bool SIInstrInfo::areLoadsFromSameBasePtr(SDNode *Load0, SDNode *Load1,
78 int64_t &Offset1) const {
79 if (!Load0->isMachineOpcode() || !Load1->isMachineOpcode())
82 unsigned Opc0 = Load0->getMachineOpcode();
83 unsigned Opc1 = Load1->getMachineOpcode();
85 // Make sure both are actually loads.
86 if (!get(Opc0).mayLoad() || !get(Opc1).mayLoad())
89 if (isDS(Opc0) && isDS(Opc1)) {
90 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
93 if (Load0->getOperand(1) != Load1->getOperand(1))
97 if (findChainOperand(Load0) != findChainOperand(Load1))
100 // Skip read2 / write2 variants for simplicity.
101 // TODO: We should report true if the used offsets are adjacent (excluded
103 if (AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::data1) != -1 ||
104 AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::data1) != -1)
107 Offset0 = cast<ConstantSDNode>(Load0->getOperand(2))->getZExtValue();
108 Offset1 = cast<ConstantSDNode>(Load1->getOperand(2))->getZExtValue();
112 if (isSMRD(Opc0) && isSMRD(Opc1)) {
113 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
116 if (Load0->getOperand(0) != Load1->getOperand(0))
120 if (findChainOperand(Load0) != findChainOperand(Load1))
123 Offset0 = cast<ConstantSDNode>(Load0->getOperand(1))->getZExtValue();
124 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue();
128 // MUBUF and MTBUF can access the same addresses.
129 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1))) {
131 // MUBUF and MTBUF have vaddr at different indices.
132 if (!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::soffset) ||
133 findChainOperand(Load0) != findChainOperand(Load1) ||
134 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::vaddr) ||
135 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::srsrc))
138 int OffIdx0 = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
139 int OffIdx1 = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
141 if (OffIdx0 == -1 || OffIdx1 == -1)
144 // getNamedOperandIdx returns the index for MachineInstrs. Since they
145 // inlcude the output in the operand list, but SDNodes don't, we need to
146 // subtract the index by one.
150 SDValue Off0 = Load0->getOperand(OffIdx0);
151 SDValue Off1 = Load1->getOperand(OffIdx1);
153 // The offset might be a FrameIndexSDNode.
154 if (!isa<ConstantSDNode>(Off0) || !isa<ConstantSDNode>(Off1))
157 Offset0 = cast<ConstantSDNode>(Off0)->getZExtValue();
158 Offset1 = cast<ConstantSDNode>(Off1)->getZExtValue();
165 static bool isStride64(unsigned Opc) {
167 case AMDGPU::DS_READ2ST64_B32:
168 case AMDGPU::DS_READ2ST64_B64:
169 case AMDGPU::DS_WRITE2ST64_B32:
170 case AMDGPU::DS_WRITE2ST64_B64:
177 bool SIInstrInfo::getLdStBaseRegImmOfs(MachineInstr *LdSt,
178 unsigned &BaseReg, unsigned &Offset,
179 const TargetRegisterInfo *TRI) const {
180 unsigned Opc = LdSt->getOpcode();
182 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
183 AMDGPU::OpName::offset);
185 // Normal, single offset LDS instruction.
186 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
187 AMDGPU::OpName::addr);
189 BaseReg = AddrReg->getReg();
190 Offset = OffsetImm->getImm();
194 // The 2 offset instructions use offset0 and offset1 instead. We can treat
195 // these as a load with a single offset if the 2 offsets are consecutive. We
196 // will use this for some partially aligned loads.
197 const MachineOperand *Offset0Imm = getNamedOperand(*LdSt,
198 AMDGPU::OpName::offset0);
199 const MachineOperand *Offset1Imm = getNamedOperand(*LdSt,
200 AMDGPU::OpName::offset1);
202 uint8_t Offset0 = Offset0Imm->getImm();
203 uint8_t Offset1 = Offset1Imm->getImm();
204 assert(Offset1 > Offset0);
206 if (Offset1 - Offset0 == 1) {
207 // Each of these offsets is in element sized units, so we need to convert
208 // to bytes of the individual reads.
212 EltSize = getOpRegClass(*LdSt, 0)->getSize() / 2;
214 assert(LdSt->mayStore());
215 int Data0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
216 EltSize = getOpRegClass(*LdSt, Data0Idx)->getSize();
222 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
223 AMDGPU::OpName::addr);
224 BaseReg = AddrReg->getReg();
225 Offset = EltSize * Offset0;
232 if (isMUBUF(Opc) || isMTBUF(Opc)) {
233 if (AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::soffset) != -1)
236 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
237 AMDGPU::OpName::vaddr);
241 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
242 AMDGPU::OpName::offset);
243 BaseReg = AddrReg->getReg();
244 Offset = OffsetImm->getImm();
249 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
250 AMDGPU::OpName::offset);
254 const MachineOperand *SBaseReg = getNamedOperand(*LdSt,
255 AMDGPU::OpName::sbase);
256 BaseReg = SBaseReg->getReg();
257 Offset = OffsetImm->getImm();
264 bool SIInstrInfo::shouldClusterLoads(MachineInstr *FirstLdSt,
265 MachineInstr *SecondLdSt,
266 unsigned NumLoads) const {
267 unsigned Opc0 = FirstLdSt->getOpcode();
268 unsigned Opc1 = SecondLdSt->getOpcode();
270 // TODO: This needs finer tuning
274 if (isDS(Opc0) && isDS(Opc1))
277 if (isSMRD(Opc0) && isSMRD(Opc1))
280 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1)))
287 SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
288 MachineBasicBlock::iterator MI, DebugLoc DL,
289 unsigned DestReg, unsigned SrcReg,
290 bool KillSrc) const {
292 // If we are trying to copy to or from SCC, there is a bug somewhere else in
293 // the backend. While it may be theoretically possible to do this, it should
294 // never be necessary.
295 assert(DestReg != AMDGPU::SCC && SrcReg != AMDGPU::SCC);
297 static const int16_t Sub0_15[] = {
298 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
299 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
300 AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
301 AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15, 0
304 static const int16_t Sub0_7[] = {
305 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
306 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7, 0
309 static const int16_t Sub0_3[] = {
310 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
313 static const int16_t Sub0_2[] = {
314 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
317 static const int16_t Sub0_1[] = {
318 AMDGPU::sub0, AMDGPU::sub1, 0
322 const int16_t *SubIndices;
324 if (AMDGPU::M0 == DestReg) {
325 // Check if M0 isn't already set to this value
326 for (MachineBasicBlock::reverse_iterator E = MBB.rend(),
327 I = MachineBasicBlock::reverse_iterator(MI); I != E; ++I) {
329 if (!I->definesRegister(AMDGPU::M0))
332 unsigned Opc = I->getOpcode();
333 if (Opc != TargetOpcode::COPY && Opc != AMDGPU::S_MOV_B32)
336 if (!I->readsRegister(SrcReg))
339 // The copy isn't necessary
344 if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
345 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
346 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
347 .addReg(SrcReg, getKillRegState(KillSrc));
350 } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
351 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
352 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
353 .addReg(SrcReg, getKillRegState(KillSrc));
356 } else if (AMDGPU::SReg_128RegClass.contains(DestReg)) {
357 assert(AMDGPU::SReg_128RegClass.contains(SrcReg));
358 Opcode = AMDGPU::S_MOV_B32;
361 } else if (AMDGPU::SReg_256RegClass.contains(DestReg)) {
362 assert(AMDGPU::SReg_256RegClass.contains(SrcReg));
363 Opcode = AMDGPU::S_MOV_B32;
366 } else if (AMDGPU::SReg_512RegClass.contains(DestReg)) {
367 assert(AMDGPU::SReg_512RegClass.contains(SrcReg));
368 Opcode = AMDGPU::S_MOV_B32;
369 SubIndices = Sub0_15;
371 } else if (AMDGPU::VReg_32RegClass.contains(DestReg)) {
372 assert(AMDGPU::VReg_32RegClass.contains(SrcReg) ||
373 AMDGPU::SReg_32RegClass.contains(SrcReg));
374 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
375 .addReg(SrcReg, getKillRegState(KillSrc));
378 } else if (AMDGPU::VReg_64RegClass.contains(DestReg)) {
379 assert(AMDGPU::VReg_64RegClass.contains(SrcReg) ||
380 AMDGPU::SReg_64RegClass.contains(SrcReg));
381 Opcode = AMDGPU::V_MOV_B32_e32;
384 } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
385 assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
386 Opcode = AMDGPU::V_MOV_B32_e32;
389 } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
390 assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
391 AMDGPU::SReg_128RegClass.contains(SrcReg));
392 Opcode = AMDGPU::V_MOV_B32_e32;
395 } else if (AMDGPU::VReg_256RegClass.contains(DestReg)) {
396 assert(AMDGPU::VReg_256RegClass.contains(SrcReg) ||
397 AMDGPU::SReg_256RegClass.contains(SrcReg));
398 Opcode = AMDGPU::V_MOV_B32_e32;
401 } else if (AMDGPU::VReg_512RegClass.contains(DestReg)) {
402 assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
403 AMDGPU::SReg_512RegClass.contains(SrcReg));
404 Opcode = AMDGPU::V_MOV_B32_e32;
405 SubIndices = Sub0_15;
408 llvm_unreachable("Can't copy register!");
411 while (unsigned SubIdx = *SubIndices++) {
412 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
413 get(Opcode), RI.getSubReg(DestReg, SubIdx));
415 Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
418 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
422 unsigned SIInstrInfo::commuteOpcode(unsigned Opcode) const {
425 // Try to map original to commuted opcode
426 if ((NewOpc = AMDGPU::getCommuteRev(Opcode)) != -1)
429 // Try to map commuted to original opcode
430 if ((NewOpc = AMDGPU::getCommuteOrig(Opcode)) != -1)
436 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
437 MachineBasicBlock::iterator MI,
438 unsigned SrcReg, bool isKill,
440 const TargetRegisterClass *RC,
441 const TargetRegisterInfo *TRI) const {
442 MachineFunction *MF = MBB.getParent();
443 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
444 DebugLoc DL = MBB.findDebugLoc(MI);
446 if (RI.hasVGPRs(RC)) {
447 LLVMContext &Ctx = MF->getFunction()->getContext();
448 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Can't spill VGPR!");
449 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), AMDGPU::VGPR0)
451 } else if (RI.isSGPRClass(RC)) {
452 // We are only allowed to create one new instruction when spilling
453 // registers, so we need to use pseudo instruction for spilling
456 switch (RC->getSize() * 8) {
457 case 32: Opcode = AMDGPU::SI_SPILL_S32_SAVE; break;
458 case 64: Opcode = AMDGPU::SI_SPILL_S64_SAVE; break;
459 case 128: Opcode = AMDGPU::SI_SPILL_S128_SAVE; break;
460 case 256: Opcode = AMDGPU::SI_SPILL_S256_SAVE; break;
461 case 512: Opcode = AMDGPU::SI_SPILL_S512_SAVE; break;
462 default: llvm_unreachable("Cannot spill register class");
465 FrameInfo->setObjectAlignment(FrameIndex, 4);
466 BuildMI(MBB, MI, DL, get(Opcode))
468 .addFrameIndex(FrameIndex);
470 llvm_unreachable("VGPR spilling not supported");
474 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
475 MachineBasicBlock::iterator MI,
476 unsigned DestReg, int FrameIndex,
477 const TargetRegisterClass *RC,
478 const TargetRegisterInfo *TRI) const {
479 MachineFunction *MF = MBB.getParent();
480 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
481 DebugLoc DL = MBB.findDebugLoc(MI);
483 if (RI.hasVGPRs(RC)) {
484 LLVMContext &Ctx = MF->getFunction()->getContext();
485 Ctx.emitError("SIInstrInfo::loadRegToStackSlot - Can't retrieve spilled VGPR!");
486 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
488 } else if (RI.isSGPRClass(RC)){
490 switch(RC->getSize() * 8) {
491 case 32: Opcode = AMDGPU::SI_SPILL_S32_RESTORE; break;
492 case 64: Opcode = AMDGPU::SI_SPILL_S64_RESTORE; break;
493 case 128: Opcode = AMDGPU::SI_SPILL_S128_RESTORE; break;
494 case 256: Opcode = AMDGPU::SI_SPILL_S256_RESTORE; break;
495 case 512: Opcode = AMDGPU::SI_SPILL_S512_RESTORE; break;
496 default: llvm_unreachable("Cannot spill register class");
499 FrameInfo->setObjectAlignment(FrameIndex, 4);
500 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
501 .addFrameIndex(FrameIndex);
503 llvm_unreachable("VGPR spilling not supported");
507 void SIInstrInfo::insertNOPs(MachineBasicBlock::iterator MI,
516 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(AMDGPU::S_NOP))
521 bool SIInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
522 MachineBasicBlock &MBB = *MI->getParent();
523 DebugLoc DL = MBB.findDebugLoc(MI);
524 switch (MI->getOpcode()) {
525 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
527 case AMDGPU::SI_CONSTDATA_PTR: {
528 unsigned Reg = MI->getOperand(0).getReg();
529 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
530 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
532 BuildMI(MBB, MI, DL, get(AMDGPU::S_GETPC_B64), Reg);
534 // Add 32-bit offset from this instruction to the start of the constant data.
535 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADD_U32), RegLo)
537 .addTargetIndex(AMDGPU::TI_CONSTDATA_START)
538 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit);
539 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADDC_U32), RegHi)
542 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit)
543 .addReg(AMDGPU::SCC, RegState::Implicit);
544 MI->eraseFromParent();
551 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
554 if (MI->getNumOperands() < 3 || !MI->getOperand(1).isReg())
557 // Make sure it s legal to commute operands for VOP2.
558 if (isVOP2(MI->getOpcode()) &&
559 (!isOperandLegal(MI, 1, &MI->getOperand(2)) ||
560 !isOperandLegal(MI, 2, &MI->getOperand(1))))
563 if (!MI->getOperand(2).isReg()) {
564 // XXX: Commute instructions with FPImm operands
565 if (NewMI || MI->getOperand(2).isFPImm() ||
566 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
570 // XXX: Commute VOP3 instructions with abs and neg set .
571 const MachineOperand *Abs = getNamedOperand(*MI, AMDGPU::OpName::abs);
572 const MachineOperand *Neg = getNamedOperand(*MI, AMDGPU::OpName::neg);
573 const MachineOperand *Src0Mods = getNamedOperand(*MI,
574 AMDGPU::OpName::src0_modifiers);
575 const MachineOperand *Src1Mods = getNamedOperand(*MI,
576 AMDGPU::OpName::src1_modifiers);
577 const MachineOperand *Src2Mods = getNamedOperand(*MI,
578 AMDGPU::OpName::src2_modifiers);
580 if ((Abs && Abs->getImm()) || (Neg && Neg->getImm()) ||
581 (Src0Mods && Src0Mods->getImm()) || (Src1Mods && Src1Mods->getImm()) ||
582 (Src2Mods && Src2Mods->getImm()))
585 unsigned Reg = MI->getOperand(1).getReg();
586 unsigned SubReg = MI->getOperand(1).getSubReg();
587 MI->getOperand(1).ChangeToImmediate(MI->getOperand(2).getImm());
588 MI->getOperand(2).ChangeToRegister(Reg, false);
589 MI->getOperand(2).setSubReg(SubReg);
591 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
595 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
600 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
601 MachineBasicBlock::iterator I,
603 unsigned SrcReg) const {
604 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
605 DstReg) .addReg(SrcReg);
608 bool SIInstrInfo::isMov(unsigned Opcode) const {
610 default: return false;
611 case AMDGPU::S_MOV_B32:
612 case AMDGPU::S_MOV_B64:
613 case AMDGPU::V_MOV_B32_e32:
614 case AMDGPU::V_MOV_B32_e64:
620 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
621 return RC != &AMDGPU::EXECRegRegClass;
625 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
626 AliasAnalysis *AA) const {
627 switch(MI->getOpcode()) {
628 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
629 case AMDGPU::S_MOV_B32:
630 case AMDGPU::S_MOV_B64:
631 case AMDGPU::V_MOV_B32_e32:
632 return MI->getOperand(1).isImm();
638 // Helper function generated by tablegen. We are wrapping this with
639 // an SIInstrInfo function that returns bool rather than int.
640 int isDS(uint16_t Opcode);
644 bool SIInstrInfo::isDS(uint16_t Opcode) const {
645 return ::AMDGPU::isDS(Opcode) != -1;
648 bool SIInstrInfo::isMIMG(uint16_t Opcode) const {
649 return get(Opcode).TSFlags & SIInstrFlags::MIMG;
652 bool SIInstrInfo::isSMRD(uint16_t Opcode) const {
653 return get(Opcode).TSFlags & SIInstrFlags::SMRD;
656 bool SIInstrInfo::isMUBUF(uint16_t Opcode) const {
657 return get(Opcode).TSFlags & SIInstrFlags::MUBUF;
660 bool SIInstrInfo::isMTBUF(uint16_t Opcode) const {
661 return get(Opcode).TSFlags & SIInstrFlags::MTBUF;
664 bool SIInstrInfo::isFLAT(uint16_t Opcode) const {
665 return get(Opcode).TSFlags & SIInstrFlags::FLAT;
668 bool SIInstrInfo::isVOP1(uint16_t Opcode) const {
669 return get(Opcode).TSFlags & SIInstrFlags::VOP1;
672 bool SIInstrInfo::isVOP2(uint16_t Opcode) const {
673 return get(Opcode).TSFlags & SIInstrFlags::VOP2;
676 bool SIInstrInfo::isVOP3(uint16_t Opcode) const {
677 return get(Opcode).TSFlags & SIInstrFlags::VOP3;
680 bool SIInstrInfo::isVOPC(uint16_t Opcode) const {
681 return get(Opcode).TSFlags & SIInstrFlags::VOPC;
684 bool SIInstrInfo::isSALUInstr(const MachineInstr &MI) const {
685 return get(MI.getOpcode()).TSFlags & SIInstrFlags::SALU;
688 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
689 int32_t Val = Imm.getSExtValue();
690 if (Val >= -16 && Val <= 64)
693 // The actual type of the operand does not seem to matter as long
694 // as the bits match one of the inline immediate values. For example:
696 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
697 // so it is a legal inline immediate.
699 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
700 // floating-point, so it is a legal inline immediate.
702 return (APInt::floatToBits(0.0f) == Imm) ||
703 (APInt::floatToBits(1.0f) == Imm) ||
704 (APInt::floatToBits(-1.0f) == Imm) ||
705 (APInt::floatToBits(0.5f) == Imm) ||
706 (APInt::floatToBits(-0.5f) == Imm) ||
707 (APInt::floatToBits(2.0f) == Imm) ||
708 (APInt::floatToBits(-2.0f) == Imm) ||
709 (APInt::floatToBits(4.0f) == Imm) ||
710 (APInt::floatToBits(-4.0f) == Imm);
713 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO) const {
715 return isInlineConstant(APInt(32, MO.getImm(), true));
718 APFloat FpImm = MO.getFPImm()->getValueAPF();
719 return isInlineConstant(FpImm.bitcastToAPInt());
725 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO) const {
726 return (MO.isImm() || MO.isFPImm()) && !isInlineConstant(MO);
729 static bool compareMachineOp(const MachineOperand &Op0,
730 const MachineOperand &Op1) {
731 if (Op0.getType() != Op1.getType())
734 switch (Op0.getType()) {
735 case MachineOperand::MO_Register:
736 return Op0.getReg() == Op1.getReg();
737 case MachineOperand::MO_Immediate:
738 return Op0.getImm() == Op1.getImm();
739 case MachineOperand::MO_FPImmediate:
740 return Op0.getFPImm() == Op1.getFPImm();
742 llvm_unreachable("Didn't expect to be comparing these operand types");
746 bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
747 const MachineOperand &MO) const {
748 const MCOperandInfo &OpInfo = get(MI->getOpcode()).OpInfo[OpNo];
750 assert(MO.isImm() || MO.isFPImm());
752 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
755 if (OpInfo.RegClass < 0)
758 return RI.regClassCanUseImmediate(OpInfo.RegClass);
761 bool SIInstrInfo::canFoldOffset(unsigned OffsetSize, unsigned AS) {
763 case AMDGPUAS::GLOBAL_ADDRESS: {
764 // MUBUF instructions a 12-bit offset in bytes.
765 return isUInt<12>(OffsetSize);
767 case AMDGPUAS::CONSTANT_ADDRESS: {
768 // SMRD instructions have an 8-bit offset in dwords.
769 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
771 case AMDGPUAS::LOCAL_ADDRESS:
772 case AMDGPUAS::REGION_ADDRESS: {
773 // The single offset versions have a 16-bit offset in bytes.
774 return isUInt<16>(OffsetSize);
776 case AMDGPUAS::PRIVATE_ADDRESS:
777 // Indirect register addressing does not use any offsets.
783 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
784 return AMDGPU::getVOPe32(Opcode) != -1;
787 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
788 // The src0_modifier operand is present on all instructions
789 // that have modifiers.
791 return AMDGPU::getNamedOperandIdx(Opcode,
792 AMDGPU::OpName::src0_modifiers) != -1;
795 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
796 StringRef &ErrInfo) const {
797 uint16_t Opcode = MI->getOpcode();
798 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
799 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
800 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
802 // Make sure the number of operands is correct.
803 const MCInstrDesc &Desc = get(Opcode);
804 if (!Desc.isVariadic() &&
805 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
806 ErrInfo = "Instruction has wrong number of operands.";
810 // Make sure the register classes are correct
811 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
812 switch (Desc.OpInfo[i].OperandType) {
813 case MCOI::OPERAND_REGISTER: {
814 int RegClass = Desc.OpInfo[i].RegClass;
815 if (!RI.regClassCanUseImmediate(RegClass) &&
816 (MI->getOperand(i).isImm() || MI->getOperand(i).isFPImm())) {
817 // Handle some special cases:
818 // Src0 can of VOP1, VOP2, VOPC can be an immediate no matter what
819 // the register class.
820 if (i != Src0Idx || (!isVOP1(Opcode) && !isVOP2(Opcode) &&
822 ErrInfo = "Expected register, but got immediate";
828 case MCOI::OPERAND_IMMEDIATE:
829 // Check if this operand is an immediate.
830 // FrameIndex operands will be replaced by immediates, so they are
832 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFPImm() &&
833 !MI->getOperand(i).isFI()) {
834 ErrInfo = "Expected immediate, but got non-immediate";
842 if (!MI->getOperand(i).isReg())
845 int RegClass = Desc.OpInfo[i].RegClass;
846 if (RegClass != -1) {
847 unsigned Reg = MI->getOperand(i).getReg();
848 if (TargetRegisterInfo::isVirtualRegister(Reg))
851 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
852 if (!RC->contains(Reg)) {
853 ErrInfo = "Operand has incorrect register class.";
861 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
862 unsigned ConstantBusCount = 0;
863 unsigned SGPRUsed = AMDGPU::NoRegister;
864 for (int i = 0, e = MI->getNumOperands(); i != e; ++i) {
865 const MachineOperand &MO = MI->getOperand(i);
866 if (MO.isReg() && MO.isUse() &&
867 !TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
869 // EXEC register uses the constant bus.
870 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
873 // FLAT_SCR is just an SGPR pair.
874 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
877 // SGPRs use the constant bus
878 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
880 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
881 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
882 if (SGPRUsed != MO.getReg()) {
884 SGPRUsed = MO.getReg();
888 // Literal constants use the constant bus.
889 if (isLiteralConstant(MO))
892 if (ConstantBusCount > 1) {
893 ErrInfo = "VOP* instruction uses the constant bus more than once";
898 // Verify SRC1 for VOP2 and VOPC
899 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
900 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
901 if (Src1.isImm() || Src1.isFPImm()) {
902 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
908 if (isVOP3(Opcode)) {
909 if (Src0Idx != -1 && isLiteralConstant(MI->getOperand(Src0Idx))) {
910 ErrInfo = "VOP3 src0 cannot be a literal constant.";
913 if (Src1Idx != -1 && isLiteralConstant(MI->getOperand(Src1Idx))) {
914 ErrInfo = "VOP3 src1 cannot be a literal constant.";
917 if (Src2Idx != -1 && isLiteralConstant(MI->getOperand(Src2Idx))) {
918 ErrInfo = "VOP3 src2 cannot be a literal constant.";
923 // Verify misc. restrictions on specific instructions.
924 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
925 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
928 const MachineOperand &Src0 = MI->getOperand(2);
929 const MachineOperand &Src1 = MI->getOperand(3);
930 const MachineOperand &Src2 = MI->getOperand(4);
931 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
932 if (!compareMachineOp(Src0, Src1) &&
933 !compareMachineOp(Src0, Src2)) {
934 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
943 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
944 switch (MI.getOpcode()) {
945 default: return AMDGPU::INSTRUCTION_LIST_END;
946 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
947 case AMDGPU::COPY: return AMDGPU::COPY;
948 case AMDGPU::PHI: return AMDGPU::PHI;
949 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
950 case AMDGPU::S_MOV_B32:
951 return MI.getOperand(1).isReg() ?
952 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
953 case AMDGPU::S_ADD_I32:
954 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
955 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
956 case AMDGPU::S_SUB_I32:
957 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
958 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
959 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
960 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
961 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
962 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
963 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
964 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
965 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
966 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
967 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
968 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
969 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
970 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
971 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
972 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
973 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
974 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
975 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
976 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
977 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
978 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
979 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
980 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
981 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
982 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
983 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
984 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
985 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
986 case AMDGPU::S_LOAD_DWORD_IMM:
987 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
988 case AMDGPU::S_LOAD_DWORDX2_IMM:
989 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
990 case AMDGPU::S_LOAD_DWORDX4_IMM:
991 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
992 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e32;
993 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
994 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
998 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
999 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
1002 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
1003 unsigned OpNo) const {
1004 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1005 const MCInstrDesc &Desc = get(MI.getOpcode());
1006 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
1007 Desc.OpInfo[OpNo].RegClass == -1)
1008 return MRI.getRegClass(MI.getOperand(OpNo).getReg());
1010 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
1011 return RI.getRegClass(RCID);
1014 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
1015 switch (MI.getOpcode()) {
1017 case AMDGPU::REG_SEQUENCE:
1019 case AMDGPU::INSERT_SUBREG:
1020 return RI.hasVGPRs(getOpRegClass(MI, 0));
1022 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
1026 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
1027 MachineBasicBlock::iterator I = MI;
1028 MachineOperand &MO = MI->getOperand(OpIdx);
1029 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1030 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
1031 const TargetRegisterClass *RC = RI.getRegClass(RCID);
1032 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1034 Opcode = AMDGPU::COPY;
1035 } else if (RI.isSGPRClass(RC)) {
1036 Opcode = AMDGPU::S_MOV_B32;
1039 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
1040 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC)) {
1041 VRC = &AMDGPU::VReg_64RegClass;
1043 VRC = &AMDGPU::VReg_32RegClass;
1045 unsigned Reg = MRI.createVirtualRegister(VRC);
1046 BuildMI(*MI->getParent(), I, MI->getParent()->findDebugLoc(I), get(Opcode),
1047 Reg).addOperand(MO);
1048 MO.ChangeToRegister(Reg, false);
1051 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
1052 MachineRegisterInfo &MRI,
1053 MachineOperand &SuperReg,
1054 const TargetRegisterClass *SuperRC,
1056 const TargetRegisterClass *SubRC)
1058 assert(SuperReg.isReg());
1060 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
1061 unsigned SubReg = MRI.createVirtualRegister(SubRC);
1063 // Just in case the super register is itself a sub-register, copy it to a new
1064 // value so we don't need to worry about merging its subreg index with the
1065 // SubIdx passed to this function. The register coalescer should be able to
1066 // eliminate this extra copy.
1067 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1069 .addOperand(SuperReg);
1071 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1073 .addReg(NewSuperReg, 0, SubIdx);
1077 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
1078 MachineBasicBlock::iterator MII,
1079 MachineRegisterInfo &MRI,
1081 const TargetRegisterClass *SuperRC,
1083 const TargetRegisterClass *SubRC) const {
1085 // XXX - Is there a better way to do this?
1086 if (SubIdx == AMDGPU::sub0)
1087 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
1088 if (SubIdx == AMDGPU::sub1)
1089 return MachineOperand::CreateImm(Op.getImm() >> 32);
1091 llvm_unreachable("Unhandled register index for immediate");
1094 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
1096 return MachineOperand::CreateReg(SubReg, false);
1099 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
1100 MachineBasicBlock::iterator MI,
1101 MachineRegisterInfo &MRI,
1102 const TargetRegisterClass *RC,
1103 const MachineOperand &Op) const {
1104 MachineBasicBlock *MBB = MI->getParent();
1105 DebugLoc DL = MI->getDebugLoc();
1106 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1107 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1108 unsigned Dst = MRI.createVirtualRegister(RC);
1110 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1112 .addImm(Op.getImm() & 0xFFFFFFFF);
1113 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1115 .addImm(Op.getImm() >> 32);
1117 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
1119 .addImm(AMDGPU::sub0)
1121 .addImm(AMDGPU::sub1);
1123 Worklist.push_back(Lo);
1124 Worklist.push_back(Hi);
1129 bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
1130 const MachineOperand *MO) const {
1131 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1132 const MCInstrDesc &InstDesc = get(MI->getOpcode());
1133 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
1134 const TargetRegisterClass *DefinedRC =
1135 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
1137 MO = &MI->getOperand(OpIdx);
1141 const TargetRegisterClass *RC = MRI.getRegClass(MO->getReg());
1142 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass));
1146 // Handle non-register types that are treated like immediates.
1147 assert(MO->isImm() || MO->isFPImm() || MO->isTargetIndex() || MO->isFI());
1150 // This opperand expects an immediate
1153 return RI.regClassCanUseImmediate(DefinedRC);
1156 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
1157 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1159 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1160 AMDGPU::OpName::src0);
1161 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1162 AMDGPU::OpName::src1);
1163 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1164 AMDGPU::OpName::src2);
1167 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
1169 if (!isOperandLegal(MI, Src0Idx))
1170 legalizeOpWithMove(MI, Src0Idx);
1173 if (isOperandLegal(MI, Src1Idx))
1176 // Usually src0 of VOP2 instructions allow more types of inputs
1177 // than src1, so try to commute the instruction to decrease our
1178 // chances of having to insert a MOV instruction to legalize src1.
1179 if (MI->isCommutable()) {
1180 if (commuteInstruction(MI))
1181 // If we are successful in commuting, then we know MI is legal, so
1186 legalizeOpWithMove(MI, Src1Idx);
1190 // XXX - Do any VOP3 instructions read VCC?
1192 if (isVOP3(MI->getOpcode())) {
1193 int VOP3Idx[3] = {Src0Idx, Src1Idx, Src2Idx};
1194 unsigned SGPRReg = AMDGPU::NoRegister;
1195 for (unsigned i = 0; i < 3; ++i) {
1196 int Idx = VOP3Idx[i];
1199 MachineOperand &MO = MI->getOperand(Idx);
1202 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
1203 continue; // VGPRs are legal
1205 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
1207 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
1208 SGPRReg = MO.getReg();
1209 // We can use one SGPR in each VOP3 instruction.
1212 } else if (!isLiteralConstant(MO)) {
1213 // If it is not a register and not a literal constant, then it must be
1214 // an inline constant which is always legal.
1217 // If we make it this far, then the operand is not legal and we must
1219 legalizeOpWithMove(MI, Idx);
1223 // Legalize REG_SEQUENCE and PHI
1224 // The register class of the operands much be the same type as the register
1225 // class of the output.
1226 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
1227 MI->getOpcode() == AMDGPU::PHI) {
1228 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
1229 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1230 if (!MI->getOperand(i).isReg() ||
1231 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1233 const TargetRegisterClass *OpRC =
1234 MRI.getRegClass(MI->getOperand(i).getReg());
1235 if (RI.hasVGPRs(OpRC)) {
1242 // If any of the operands are VGPR registers, then they all most be
1243 // otherwise we will create illegal VGPR->SGPR copies when legalizing
1245 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
1248 VRC = RI.getEquivalentVGPRClass(SRC);
1255 // Update all the operands so they have the same type.
1256 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1257 if (!MI->getOperand(i).isReg() ||
1258 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1260 unsigned DstReg = MRI.createVirtualRegister(RC);
1261 MachineBasicBlock *InsertBB;
1262 MachineBasicBlock::iterator Insert;
1263 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
1264 InsertBB = MI->getParent();
1267 // MI is a PHI instruction.
1268 InsertBB = MI->getOperand(i + 1).getMBB();
1269 Insert = InsertBB->getFirstTerminator();
1271 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
1272 get(AMDGPU::COPY), DstReg)
1273 .addOperand(MI->getOperand(i));
1274 MI->getOperand(i).setReg(DstReg);
1278 // Legalize INSERT_SUBREG
1279 // src0 must have the same register class as dst
1280 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
1281 unsigned Dst = MI->getOperand(0).getReg();
1282 unsigned Src0 = MI->getOperand(1).getReg();
1283 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
1284 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
1285 if (DstRC != Src0RC) {
1286 MachineBasicBlock &MBB = *MI->getParent();
1287 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
1288 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
1290 MI->getOperand(1).setReg(NewSrc0);
1295 // Legalize MUBUF* instructions
1296 // FIXME: If we start using the non-addr64 instructions for compute, we
1297 // may need to legalize them here.
1299 AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::srsrc);
1300 if (SRsrcIdx != -1) {
1301 // We have an MUBUF instruction
1302 MachineOperand *SRsrc = &MI->getOperand(SRsrcIdx);
1303 unsigned SRsrcRC = get(MI->getOpcode()).OpInfo[SRsrcIdx].RegClass;
1304 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
1305 RI.getRegClass(SRsrcRC))) {
1306 // The operands are legal.
1307 // FIXME: We may need to legalize operands besided srsrc.
1311 MachineBasicBlock &MBB = *MI->getParent();
1312 // Extract the the ptr from the resource descriptor.
1314 // SRsrcPtrLo = srsrc:sub0
1315 unsigned SRsrcPtrLo = buildExtractSubReg(MI, MRI, *SRsrc,
1316 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VReg_32RegClass);
1318 // SRsrcPtrHi = srsrc:sub1
1319 unsigned SRsrcPtrHi = buildExtractSubReg(MI, MRI, *SRsrc,
1320 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VReg_32RegClass);
1322 // Create an empty resource descriptor
1323 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1324 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1325 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1326 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1329 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1333 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1334 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1336 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1338 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1339 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1341 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1343 // NewSRsrc = {Zero64, SRsrcFormat}
1344 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1347 .addImm(AMDGPU::sub0_sub1)
1348 .addReg(SRsrcFormatLo)
1349 .addImm(AMDGPU::sub2)
1350 .addReg(SRsrcFormatHi)
1351 .addImm(AMDGPU::sub3);
1353 MachineOperand *VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1354 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
1355 unsigned NewVAddrLo;
1356 unsigned NewVAddrHi;
1358 // This is already an ADDR64 instruction so we need to add the pointer
1359 // extracted from the resource descriptor to the current value of VAddr.
1360 NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1361 NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1363 // NewVaddrLo = SRsrcPtrLo + VAddr:sub0
1364 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1367 .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
1368 .addReg(AMDGPU::VCC, RegState::ImplicitDefine);
1370 // NewVaddrHi = SRsrcPtrHi + VAddr:sub1
1371 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1374 .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
1375 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1376 .addReg(AMDGPU::VCC, RegState::Implicit);
1379 // This instructions is the _OFFSET variant, so we need to convert it to
1381 MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
1382 MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
1383 MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
1384 assert(SOffset->isImm() && SOffset->getImm() == 0 && "Legalizing MUBUF "
1385 "with non-zero soffset is not implemented");
1388 // Create the new instruction.
1389 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
1390 MachineInstr *Addr64 =
1391 BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
1394 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
1395 // This will be replaced later
1396 // with the new value of vaddr.
1397 .addOperand(*Offset);
1399 MI->removeFromParent();
1402 NewVAddrLo = SRsrcPtrLo;
1403 NewVAddrHi = SRsrcPtrHi;
1404 VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1405 SRsrc = getNamedOperand(*MI, AMDGPU::OpName::srsrc);
1408 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1409 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1412 .addImm(AMDGPU::sub0)
1414 .addImm(AMDGPU::sub1);
1417 // Update the instruction to use NewVaddr
1418 VAddr->setReg(NewVAddr);
1419 // Update the instruction to use NewSRsrc
1420 SRsrc->setReg(NewSRsrc);
1424 void SIInstrInfo::splitSMRD(MachineInstr *MI,
1425 const TargetRegisterClass *HalfRC,
1426 unsigned HalfImmOp, unsigned HalfSGPROp,
1427 MachineInstr *&Lo, MachineInstr *&Hi) const {
1429 DebugLoc DL = MI->getDebugLoc();
1430 MachineBasicBlock *MBB = MI->getParent();
1431 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1432 unsigned RegLo = MRI.createVirtualRegister(HalfRC);
1433 unsigned RegHi = MRI.createVirtualRegister(HalfRC);
1434 unsigned HalfSize = HalfRC->getSize();
1435 const MachineOperand *OffOp =
1436 getNamedOperand(*MI, AMDGPU::OpName::offset);
1437 const MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
1440 // Handle the _IMM variant
1441 unsigned LoOffset = OffOp->getImm();
1442 unsigned HiOffset = LoOffset + (HalfSize / 4);
1443 Lo = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegLo)
1447 if (!isUInt<8>(HiOffset)) {
1448 unsigned OffsetSGPR =
1449 MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1450 BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32), OffsetSGPR)
1451 .addImm(HiOffset << 2); // The immediate offset is in dwords,
1452 // but offset in register is in bytes.
1453 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
1455 .addReg(OffsetSGPR);
1457 Hi = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegHi)
1462 // Handle the _SGPR variant
1463 MachineOperand *SOff = getNamedOperand(*MI, AMDGPU::OpName::soff);
1464 Lo = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegLo)
1467 unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1468 BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
1471 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp))
1473 .addReg(OffsetSGPR);
1476 unsigned SubLo, SubHi;
1479 SubLo = AMDGPU::sub0;
1480 SubHi = AMDGPU::sub1;
1483 SubLo = AMDGPU::sub0_sub1;
1484 SubHi = AMDGPU::sub2_sub3;
1487 SubLo = AMDGPU::sub0_sub1_sub2_sub3;
1488 SubHi = AMDGPU::sub4_sub5_sub6_sub7;
1491 SubLo = AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
1492 SubHi = AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15;
1495 llvm_unreachable("Unhandled HalfSize");
1498 BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE))
1499 .addOperand(MI->getOperand(0))
1506 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
1507 MachineBasicBlock *MBB = MI->getParent();
1508 switch (MI->getOpcode()) {
1509 case AMDGPU::S_LOAD_DWORD_IMM:
1510 case AMDGPU::S_LOAD_DWORD_SGPR:
1511 case AMDGPU::S_LOAD_DWORDX2_IMM:
1512 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1513 case AMDGPU::S_LOAD_DWORDX4_IMM:
1514 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1515 unsigned NewOpcode = getVALUOp(*MI);
1519 if (MI->getOperand(2).isReg()) {
1520 RegOffset = MI->getOperand(2).getReg();
1523 assert(MI->getOperand(2).isImm());
1524 // SMRD instructions take a dword offsets and MUBUF instructions
1525 // take a byte offset.
1526 ImmOffset = MI->getOperand(2).getImm() << 2;
1527 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1528 if (isUInt<12>(ImmOffset)) {
1529 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1533 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1540 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1541 unsigned DWord0 = RegOffset;
1542 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1543 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1544 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1546 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
1548 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
1549 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1550 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
1551 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1552 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
1554 .addImm(AMDGPU::sub0)
1556 .addImm(AMDGPU::sub1)
1558 .addImm(AMDGPU::sub2)
1560 .addImm(AMDGPU::sub3);
1561 MI->setDesc(get(NewOpcode));
1562 if (MI->getOperand(2).isReg()) {
1563 MI->getOperand(2).setReg(MI->getOperand(1).getReg());
1565 MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
1567 MI->getOperand(1).setReg(SRsrc);
1568 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
1570 const TargetRegisterClass *NewDstRC =
1571 RI.getRegClass(get(NewOpcode).OpInfo[0].RegClass);
1573 unsigned DstReg = MI->getOperand(0).getReg();
1574 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1575 MRI.replaceRegWith(DstReg, NewDstReg);
1578 case AMDGPU::S_LOAD_DWORDX8_IMM:
1579 case AMDGPU::S_LOAD_DWORDX8_SGPR: {
1580 MachineInstr *Lo, *Hi;
1581 splitSMRD(MI, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM,
1582 AMDGPU::S_LOAD_DWORDX4_SGPR, Lo, Hi);
1583 MI->eraseFromParent();
1584 moveSMRDToVALU(Lo, MRI);
1585 moveSMRDToVALU(Hi, MRI);
1589 case AMDGPU::S_LOAD_DWORDX16_IMM:
1590 case AMDGPU::S_LOAD_DWORDX16_SGPR: {
1591 MachineInstr *Lo, *Hi;
1592 splitSMRD(MI, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM,
1593 AMDGPU::S_LOAD_DWORDX8_SGPR, Lo, Hi);
1594 MI->eraseFromParent();
1595 moveSMRDToVALU(Lo, MRI);
1596 moveSMRDToVALU(Hi, MRI);
1602 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
1603 SmallVector<MachineInstr *, 128> Worklist;
1604 Worklist.push_back(&TopInst);
1606 while (!Worklist.empty()) {
1607 MachineInstr *Inst = Worklist.pop_back_val();
1608 MachineBasicBlock *MBB = Inst->getParent();
1609 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1611 unsigned Opcode = Inst->getOpcode();
1612 unsigned NewOpcode = getVALUOp(*Inst);
1614 // Handle some special cases
1617 if (isSMRD(Inst->getOpcode())) {
1618 moveSMRDToVALU(Inst, MRI);
1621 case AMDGPU::S_MOV_B64: {
1622 DebugLoc DL = Inst->getDebugLoc();
1624 // If the source operand is a register we can replace this with a
1626 if (Inst->getOperand(1).isReg()) {
1627 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
1628 .addOperand(Inst->getOperand(0))
1629 .addOperand(Inst->getOperand(1));
1630 Worklist.push_back(Copy);
1632 // Otherwise, we need to split this into two movs, because there is
1633 // no 64-bit VALU move instruction.
1634 unsigned Reg = Inst->getOperand(0).getReg();
1635 unsigned Dst = split64BitImm(Worklist,
1638 MRI.getRegClass(Reg),
1639 Inst->getOperand(1));
1640 MRI.replaceRegWith(Reg, Dst);
1642 Inst->eraseFromParent();
1645 case AMDGPU::S_AND_B64:
1646 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
1647 Inst->eraseFromParent();
1650 case AMDGPU::S_OR_B64:
1651 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
1652 Inst->eraseFromParent();
1655 case AMDGPU::S_XOR_B64:
1656 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
1657 Inst->eraseFromParent();
1660 case AMDGPU::S_NOT_B64:
1661 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
1662 Inst->eraseFromParent();
1665 case AMDGPU::S_BCNT1_I32_B64:
1666 splitScalar64BitBCNT(Worklist, Inst);
1667 Inst->eraseFromParent();
1670 case AMDGPU::S_BFE_U64:
1671 case AMDGPU::S_BFE_I64:
1672 case AMDGPU::S_BFM_B64:
1673 llvm_unreachable("Moving this op to VALU not implemented");
1676 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
1677 // We cannot move this instruction to the VALU, so we should try to
1678 // legalize its operands instead.
1679 legalizeOperands(Inst);
1683 // Use the new VALU Opcode.
1684 const MCInstrDesc &NewDesc = get(NewOpcode);
1685 Inst->setDesc(NewDesc);
1687 // Remove any references to SCC. Vector instructions can't read from it, and
1688 // We're just about to add the implicit use / defs of VCC, and we don't want
1690 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
1691 MachineOperand &Op = Inst->getOperand(i);
1692 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
1693 Inst->RemoveOperand(i);
1696 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
1697 // We are converting these to a BFE, so we need to add the missing
1698 // operands for the size and offset.
1699 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
1700 Inst->addOperand(MachineOperand::CreateImm(0));
1701 Inst->addOperand(MachineOperand::CreateImm(Size));
1703 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
1704 // The VALU version adds the second operand to the result, so insert an
1706 Inst->addOperand(MachineOperand::CreateImm(0));
1709 addDescImplicitUseDef(NewDesc, Inst);
1711 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
1712 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
1713 // If we need to move this to VGPRs, we need to unpack the second operand
1714 // back into the 2 separate ones for bit offset and width.
1715 assert(OffsetWidthOp.isImm() &&
1716 "Scalar BFE is only implemented for constant width and offset");
1717 uint32_t Imm = OffsetWidthOp.getImm();
1719 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
1720 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
1721 Inst->RemoveOperand(2); // Remove old immediate.
1722 Inst->addOperand(MachineOperand::CreateImm(Offset));
1723 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
1726 // Update the destination register class.
1728 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
1731 // For target instructions, getOpRegClass just returns the virtual
1732 // register class associated with the operand, so we need to find an
1733 // equivalent VGPR register class in order to move the instruction to the
1737 case AMDGPU::REG_SEQUENCE:
1738 case AMDGPU::INSERT_SUBREG:
1739 if (RI.hasVGPRs(NewDstRC))
1741 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
1749 unsigned DstReg = Inst->getOperand(0).getReg();
1750 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1751 MRI.replaceRegWith(DstReg, NewDstReg);
1753 // Legalize the operands
1754 legalizeOperands(Inst);
1756 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
1757 E = MRI.use_end(); I != E; ++I) {
1758 MachineInstr &UseMI = *I->getParent();
1759 if (!canReadVGPR(UseMI, I.getOperandNo())) {
1760 Worklist.push_back(&UseMI);
1766 //===----------------------------------------------------------------------===//
1767 // Indirect addressing callbacks
1768 //===----------------------------------------------------------------------===//
1770 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
1771 unsigned Channel) const {
1772 assert(Channel == 0);
1776 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
1777 return &AMDGPU::VReg_32RegClass;
1780 void SIInstrInfo::splitScalar64BitUnaryOp(
1781 SmallVectorImpl<MachineInstr *> &Worklist,
1783 unsigned Opcode) const {
1784 MachineBasicBlock &MBB = *Inst->getParent();
1785 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1787 MachineOperand &Dest = Inst->getOperand(0);
1788 MachineOperand &Src0 = Inst->getOperand(1);
1789 DebugLoc DL = Inst->getDebugLoc();
1791 MachineBasicBlock::iterator MII = Inst;
1793 const MCInstrDesc &InstDesc = get(Opcode);
1794 const TargetRegisterClass *Src0RC = Src0.isReg() ?
1795 MRI.getRegClass(Src0.getReg()) :
1796 &AMDGPU::SGPR_32RegClass;
1798 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
1800 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1801 AMDGPU::sub0, Src0SubRC);
1803 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
1804 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
1806 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
1807 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
1808 .addOperand(SrcReg0Sub0);
1810 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1811 AMDGPU::sub1, Src0SubRC);
1813 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
1814 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
1815 .addOperand(SrcReg0Sub1);
1817 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
1818 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
1820 .addImm(AMDGPU::sub0)
1822 .addImm(AMDGPU::sub1);
1824 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
1826 // Try to legalize the operands in case we need to swap the order to keep it
1828 Worklist.push_back(LoHalf);
1829 Worklist.push_back(HiHalf);
1832 void SIInstrInfo::splitScalar64BitBinaryOp(
1833 SmallVectorImpl<MachineInstr *> &Worklist,
1835 unsigned Opcode) const {
1836 MachineBasicBlock &MBB = *Inst->getParent();
1837 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1839 MachineOperand &Dest = Inst->getOperand(0);
1840 MachineOperand &Src0 = Inst->getOperand(1);
1841 MachineOperand &Src1 = Inst->getOperand(2);
1842 DebugLoc DL = Inst->getDebugLoc();
1844 MachineBasicBlock::iterator MII = Inst;
1846 const MCInstrDesc &InstDesc = get(Opcode);
1847 const TargetRegisterClass *Src0RC = Src0.isReg() ?
1848 MRI.getRegClass(Src0.getReg()) :
1849 &AMDGPU::SGPR_32RegClass;
1851 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
1852 const TargetRegisterClass *Src1RC = Src1.isReg() ?
1853 MRI.getRegClass(Src1.getReg()) :
1854 &AMDGPU::SGPR_32RegClass;
1856 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
1858 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1859 AMDGPU::sub0, Src0SubRC);
1860 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
1861 AMDGPU::sub0, Src1SubRC);
1863 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
1864 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
1866 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
1867 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
1868 .addOperand(SrcReg0Sub0)
1869 .addOperand(SrcReg1Sub0);
1871 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1872 AMDGPU::sub1, Src0SubRC);
1873 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
1874 AMDGPU::sub1, Src1SubRC);
1876 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
1877 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
1878 .addOperand(SrcReg0Sub1)
1879 .addOperand(SrcReg1Sub1);
1881 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
1882 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
1884 .addImm(AMDGPU::sub0)
1886 .addImm(AMDGPU::sub1);
1888 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
1890 // Try to legalize the operands in case we need to swap the order to keep it
1892 Worklist.push_back(LoHalf);
1893 Worklist.push_back(HiHalf);
1896 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
1897 MachineInstr *Inst) const {
1898 MachineBasicBlock &MBB = *Inst->getParent();
1899 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1901 MachineBasicBlock::iterator MII = Inst;
1902 DebugLoc DL = Inst->getDebugLoc();
1904 MachineOperand &Dest = Inst->getOperand(0);
1905 MachineOperand &Src = Inst->getOperand(1);
1907 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e32);
1908 const TargetRegisterClass *SrcRC = Src.isReg() ?
1909 MRI.getRegClass(Src.getReg()) :
1910 &AMDGPU::SGPR_32RegClass;
1912 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1913 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1915 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
1917 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
1918 AMDGPU::sub0, SrcSubRC);
1919 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
1920 AMDGPU::sub1, SrcSubRC);
1922 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
1923 .addOperand(SrcRegSub0)
1926 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
1927 .addOperand(SrcRegSub1)
1930 MRI.replaceRegWith(Dest.getReg(), ResultReg);
1932 Worklist.push_back(First);
1933 Worklist.push_back(Second);
1936 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
1937 MachineInstr *Inst) const {
1938 // Add the implict and explicit register definitions.
1939 if (NewDesc.ImplicitUses) {
1940 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
1941 unsigned Reg = NewDesc.ImplicitUses[i];
1942 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
1946 if (NewDesc.ImplicitDefs) {
1947 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
1948 unsigned Reg = NewDesc.ImplicitDefs[i];
1949 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
1954 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
1955 MachineBasicBlock *MBB,
1956 MachineBasicBlock::iterator I,
1958 unsigned Address, unsigned OffsetReg) const {
1959 const DebugLoc &DL = MBB->findDebugLoc(I);
1960 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
1961 getIndirectIndexBegin(*MBB->getParent()));
1963 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
1964 .addReg(IndirectBaseReg, RegState::Define)
1965 .addOperand(I->getOperand(0))
1966 .addReg(IndirectBaseReg)
1972 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
1973 MachineBasicBlock *MBB,
1974 MachineBasicBlock::iterator I,
1976 unsigned Address, unsigned OffsetReg) const {
1977 const DebugLoc &DL = MBB->findDebugLoc(I);
1978 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
1979 getIndirectIndexBegin(*MBB->getParent()));
1981 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
1982 .addOperand(I->getOperand(0))
1983 .addOperand(I->getOperand(1))
1984 .addReg(IndirectBaseReg)
1990 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
1991 const MachineFunction &MF) const {
1992 int End = getIndirectIndexEnd(MF);
1993 int Begin = getIndirectIndexBegin(MF);
1999 for (int Index = Begin; Index <= End; ++Index)
2000 Reserved.set(AMDGPU::VReg_32RegClass.getRegister(Index));
2002 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
2003 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
2005 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
2006 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
2008 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
2009 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
2011 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
2012 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
2014 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
2015 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));
2018 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
2019 unsigned OperandName) const {
2020 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
2024 return &MI.getOperand(Idx);