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));
92 // TODO: Also shouldn't see read2st
93 assert(Opc0 != AMDGPU::DS_READ2_B32 &&
94 Opc0 != AMDGPU::DS_READ2_B64 &&
95 Opc1 != AMDGPU::DS_READ2_B32 &&
96 Opc1 != AMDGPU::DS_READ2_B64);
99 if (Load0->getOperand(1) != Load1->getOperand(1))
103 if (findChainOperand(Load0) != findChainOperand(Load1))
106 Offset0 = cast<ConstantSDNode>(Load0->getOperand(2))->getZExtValue();
107 Offset1 = cast<ConstantSDNode>(Load1->getOperand(2))->getZExtValue();
111 if (isSMRD(Opc0) && isSMRD(Opc1)) {
112 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
115 if (Load0->getOperand(0) != Load1->getOperand(0))
119 if (findChainOperand(Load0) != findChainOperand(Load1))
122 Offset0 = cast<ConstantSDNode>(Load0->getOperand(1))->getZExtValue();
123 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue();
127 // MUBUF and MTBUF can access the same addresses.
128 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1))) {
130 // MUBUF and MTBUF have vaddr at different indices.
131 if (!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::soffset) ||
132 findChainOperand(Load0) != findChainOperand(Load1) ||
133 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::vaddr) ||
134 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::srsrc))
137 int OffIdx0 = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
138 int OffIdx1 = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
140 if (OffIdx0 == -1 || OffIdx1 == -1)
143 // getNamedOperandIdx returns the index for MachineInstrs. Since they
144 // inlcude the output in the operand list, but SDNodes don't, we need to
145 // subtract the index by one.
149 SDValue Off0 = Load0->getOperand(OffIdx0);
150 SDValue Off1 = Load1->getOperand(OffIdx1);
152 // The offset might be a FrameIndexSDNode.
153 if (!isa<ConstantSDNode>(Off0) || !isa<ConstantSDNode>(Off1))
156 Offset0 = cast<ConstantSDNode>(Off0)->getZExtValue();
157 Offset1 = cast<ConstantSDNode>(Off1)->getZExtValue();
164 static bool isStride64(unsigned Opc) {
166 case AMDGPU::DS_READ2ST64_B32:
167 case AMDGPU::DS_READ2ST64_B64:
168 case AMDGPU::DS_WRITE2ST64_B32:
169 case AMDGPU::DS_WRITE2ST64_B64:
176 bool SIInstrInfo::getLdStBaseRegImmOfs(MachineInstr *LdSt,
177 unsigned &BaseReg, unsigned &Offset,
178 const TargetRegisterInfo *TRI) const {
179 unsigned Opc = LdSt->getOpcode();
181 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
182 AMDGPU::OpName::offset);
184 // Normal, single offset LDS instruction.
185 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
186 AMDGPU::OpName::addr);
188 BaseReg = AddrReg->getReg();
189 Offset = OffsetImm->getImm();
193 // The 2 offset instructions use offset0 and offset1 instead. We can treat
194 // these as a load with a single offset if the 2 offsets are consecutive. We
195 // will use this for some partially aligned loads.
196 const MachineOperand *Offset0Imm = getNamedOperand(*LdSt,
197 AMDGPU::OpName::offset0);
198 const MachineOperand *Offset1Imm = getNamedOperand(*LdSt,
199 AMDGPU::OpName::offset1);
201 uint8_t Offset0 = Offset0Imm->getImm();
202 uint8_t Offset1 = Offset1Imm->getImm();
203 assert(Offset1 > Offset0);
205 if (Offset1 - Offset0 == 1) {
206 // Each of these offsets is in element sized units, so we need to convert
207 // to bytes of the individual reads.
211 EltSize = getOpRegClass(*LdSt, 0)->getSize() / 2;
213 assert(LdSt->mayStore());
214 int Data0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
215 EltSize = getOpRegClass(*LdSt, Data0Idx)->getSize();
221 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
222 AMDGPU::OpName::addr);
223 BaseReg = AddrReg->getReg();
224 Offset = EltSize * Offset0;
231 if (isMUBUF(Opc) || isMTBUF(Opc)) {
232 if (AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::soffset) != -1)
235 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
236 AMDGPU::OpName::vaddr);
240 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
241 AMDGPU::OpName::offset);
242 BaseReg = AddrReg->getReg();
243 Offset = OffsetImm->getImm();
248 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
249 AMDGPU::OpName::offset);
253 const MachineOperand *SBaseReg = getNamedOperand(*LdSt,
254 AMDGPU::OpName::sbase);
255 BaseReg = SBaseReg->getReg();
256 Offset = OffsetImm->getImm();
264 SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
265 MachineBasicBlock::iterator MI, DebugLoc DL,
266 unsigned DestReg, unsigned SrcReg,
267 bool KillSrc) const {
269 // If we are trying to copy to or from SCC, there is a bug somewhere else in
270 // the backend. While it may be theoretically possible to do this, it should
271 // never be necessary.
272 assert(DestReg != AMDGPU::SCC && SrcReg != AMDGPU::SCC);
274 static const int16_t Sub0_15[] = {
275 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
276 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
277 AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
278 AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15, 0
281 static const int16_t Sub0_7[] = {
282 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
283 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7, 0
286 static const int16_t Sub0_3[] = {
287 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
290 static const int16_t Sub0_2[] = {
291 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
294 static const int16_t Sub0_1[] = {
295 AMDGPU::sub0, AMDGPU::sub1, 0
299 const int16_t *SubIndices;
301 if (AMDGPU::M0 == DestReg) {
302 // Check if M0 isn't already set to this value
303 for (MachineBasicBlock::reverse_iterator E = MBB.rend(),
304 I = MachineBasicBlock::reverse_iterator(MI); I != E; ++I) {
306 if (!I->definesRegister(AMDGPU::M0))
309 unsigned Opc = I->getOpcode();
310 if (Opc != TargetOpcode::COPY && Opc != AMDGPU::S_MOV_B32)
313 if (!I->readsRegister(SrcReg))
316 // The copy isn't necessary
321 if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
322 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
323 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
324 .addReg(SrcReg, getKillRegState(KillSrc));
327 } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
328 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
329 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
330 .addReg(SrcReg, getKillRegState(KillSrc));
333 } else if (AMDGPU::SReg_128RegClass.contains(DestReg)) {
334 assert(AMDGPU::SReg_128RegClass.contains(SrcReg));
335 Opcode = AMDGPU::S_MOV_B32;
338 } else if (AMDGPU::SReg_256RegClass.contains(DestReg)) {
339 assert(AMDGPU::SReg_256RegClass.contains(SrcReg));
340 Opcode = AMDGPU::S_MOV_B32;
343 } else if (AMDGPU::SReg_512RegClass.contains(DestReg)) {
344 assert(AMDGPU::SReg_512RegClass.contains(SrcReg));
345 Opcode = AMDGPU::S_MOV_B32;
346 SubIndices = Sub0_15;
348 } else if (AMDGPU::VReg_32RegClass.contains(DestReg)) {
349 assert(AMDGPU::VReg_32RegClass.contains(SrcReg) ||
350 AMDGPU::SReg_32RegClass.contains(SrcReg));
351 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
352 .addReg(SrcReg, getKillRegState(KillSrc));
355 } else if (AMDGPU::VReg_64RegClass.contains(DestReg)) {
356 assert(AMDGPU::VReg_64RegClass.contains(SrcReg) ||
357 AMDGPU::SReg_64RegClass.contains(SrcReg));
358 Opcode = AMDGPU::V_MOV_B32_e32;
361 } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
362 assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
363 Opcode = AMDGPU::V_MOV_B32_e32;
366 } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
367 assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
368 AMDGPU::SReg_128RegClass.contains(SrcReg));
369 Opcode = AMDGPU::V_MOV_B32_e32;
372 } else if (AMDGPU::VReg_256RegClass.contains(DestReg)) {
373 assert(AMDGPU::VReg_256RegClass.contains(SrcReg) ||
374 AMDGPU::SReg_256RegClass.contains(SrcReg));
375 Opcode = AMDGPU::V_MOV_B32_e32;
378 } else if (AMDGPU::VReg_512RegClass.contains(DestReg)) {
379 assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
380 AMDGPU::SReg_512RegClass.contains(SrcReg));
381 Opcode = AMDGPU::V_MOV_B32_e32;
382 SubIndices = Sub0_15;
385 llvm_unreachable("Can't copy register!");
388 while (unsigned SubIdx = *SubIndices++) {
389 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
390 get(Opcode), RI.getSubReg(DestReg, SubIdx));
392 Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
395 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
399 unsigned SIInstrInfo::commuteOpcode(unsigned Opcode) const {
402 // Try to map original to commuted opcode
403 if ((NewOpc = AMDGPU::getCommuteRev(Opcode)) != -1)
406 // Try to map commuted to original opcode
407 if ((NewOpc = AMDGPU::getCommuteOrig(Opcode)) != -1)
413 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
414 MachineBasicBlock::iterator MI,
415 unsigned SrcReg, bool isKill,
417 const TargetRegisterClass *RC,
418 const TargetRegisterInfo *TRI) const {
419 MachineFunction *MF = MBB.getParent();
420 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
421 DebugLoc DL = MBB.findDebugLoc(MI);
423 if (RI.hasVGPRs(RC)) {
424 LLVMContext &Ctx = MF->getFunction()->getContext();
425 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Can't spill VGPR!");
426 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), AMDGPU::VGPR0)
428 } else if (RI.isSGPRClass(RC)) {
429 // We are only allowed to create one new instruction when spilling
430 // registers, so we need to use pseudo instruction for spilling
433 switch (RC->getSize() * 8) {
434 case 32: Opcode = AMDGPU::SI_SPILL_S32_SAVE; break;
435 case 64: Opcode = AMDGPU::SI_SPILL_S64_SAVE; break;
436 case 128: Opcode = AMDGPU::SI_SPILL_S128_SAVE; break;
437 case 256: Opcode = AMDGPU::SI_SPILL_S256_SAVE; break;
438 case 512: Opcode = AMDGPU::SI_SPILL_S512_SAVE; break;
439 default: llvm_unreachable("Cannot spill register class");
442 FrameInfo->setObjectAlignment(FrameIndex, 4);
443 BuildMI(MBB, MI, DL, get(Opcode))
445 .addFrameIndex(FrameIndex);
447 llvm_unreachable("VGPR spilling not supported");
451 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
452 MachineBasicBlock::iterator MI,
453 unsigned DestReg, int FrameIndex,
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);
460 if (RI.hasVGPRs(RC)) {
461 LLVMContext &Ctx = MF->getFunction()->getContext();
462 Ctx.emitError("SIInstrInfo::loadRegToStackSlot - Can't retrieve spilled VGPR!");
463 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
465 } else if (RI.isSGPRClass(RC)){
467 switch(RC->getSize() * 8) {
468 case 32: Opcode = AMDGPU::SI_SPILL_S32_RESTORE; break;
469 case 64: Opcode = AMDGPU::SI_SPILL_S64_RESTORE; break;
470 case 128: Opcode = AMDGPU::SI_SPILL_S128_RESTORE; break;
471 case 256: Opcode = AMDGPU::SI_SPILL_S256_RESTORE; break;
472 case 512: Opcode = AMDGPU::SI_SPILL_S512_RESTORE; break;
473 default: llvm_unreachable("Cannot spill register class");
476 FrameInfo->setObjectAlignment(FrameIndex, 4);
477 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
478 .addFrameIndex(FrameIndex);
480 llvm_unreachable("VGPR spilling not supported");
484 void SIInstrInfo::insertNOPs(MachineBasicBlock::iterator MI,
493 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(AMDGPU::S_NOP))
498 bool SIInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
499 MachineBasicBlock &MBB = *MI->getParent();
500 DebugLoc DL = MBB.findDebugLoc(MI);
501 switch (MI->getOpcode()) {
502 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
504 case AMDGPU::SI_CONSTDATA_PTR: {
505 unsigned Reg = MI->getOperand(0).getReg();
506 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
507 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
509 BuildMI(MBB, MI, DL, get(AMDGPU::S_GETPC_B64), Reg);
511 // Add 32-bit offset from this instruction to the start of the constant data.
512 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADD_U32), RegLo)
514 .addTargetIndex(AMDGPU::TI_CONSTDATA_START)
515 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit);
516 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADDC_U32), RegHi)
519 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit)
520 .addReg(AMDGPU::SCC, RegState::Implicit);
521 MI->eraseFromParent();
528 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
531 if (MI->getNumOperands() < 3 || !MI->getOperand(1).isReg())
534 // Make sure it s legal to commute operands for VOP2.
535 if (isVOP2(MI->getOpcode()) &&
536 (!isOperandLegal(MI, 1, &MI->getOperand(2)) ||
537 !isOperandLegal(MI, 2, &MI->getOperand(1))))
540 if (!MI->getOperand(2).isReg()) {
541 // XXX: Commute instructions with FPImm operands
542 if (NewMI || MI->getOperand(2).isFPImm() ||
543 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
547 // XXX: Commute VOP3 instructions with abs and neg set .
548 const MachineOperand *Abs = getNamedOperand(*MI, AMDGPU::OpName::abs);
549 const MachineOperand *Neg = getNamedOperand(*MI, AMDGPU::OpName::neg);
550 const MachineOperand *Src0Mods = getNamedOperand(*MI,
551 AMDGPU::OpName::src0_modifiers);
552 const MachineOperand *Src1Mods = getNamedOperand(*MI,
553 AMDGPU::OpName::src1_modifiers);
554 const MachineOperand *Src2Mods = getNamedOperand(*MI,
555 AMDGPU::OpName::src2_modifiers);
557 if ((Abs && Abs->getImm()) || (Neg && Neg->getImm()) ||
558 (Src0Mods && Src0Mods->getImm()) || (Src1Mods && Src1Mods->getImm()) ||
559 (Src2Mods && Src2Mods->getImm()))
562 unsigned Reg = MI->getOperand(1).getReg();
563 unsigned SubReg = MI->getOperand(1).getSubReg();
564 MI->getOperand(1).ChangeToImmediate(MI->getOperand(2).getImm());
565 MI->getOperand(2).ChangeToRegister(Reg, false);
566 MI->getOperand(2).setSubReg(SubReg);
568 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
572 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
577 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
578 MachineBasicBlock::iterator I,
580 unsigned SrcReg) const {
581 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
582 DstReg) .addReg(SrcReg);
585 bool SIInstrInfo::isMov(unsigned Opcode) const {
587 default: return false;
588 case AMDGPU::S_MOV_B32:
589 case AMDGPU::S_MOV_B64:
590 case AMDGPU::V_MOV_B32_e32:
591 case AMDGPU::V_MOV_B32_e64:
597 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
598 return RC != &AMDGPU::EXECRegRegClass;
602 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
603 AliasAnalysis *AA) const {
604 switch(MI->getOpcode()) {
605 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
606 case AMDGPU::S_MOV_B32:
607 case AMDGPU::S_MOV_B64:
608 case AMDGPU::V_MOV_B32_e32:
609 return MI->getOperand(1).isImm();
615 // Helper function generated by tablegen. We are wrapping this with
616 // an SIInstrInfo function that returns bool rather than int.
617 int isDS(uint16_t Opcode);
621 bool SIInstrInfo::isDS(uint16_t Opcode) const {
622 return ::AMDGPU::isDS(Opcode) != -1;
625 bool SIInstrInfo::isMIMG(uint16_t Opcode) const {
626 return get(Opcode).TSFlags & SIInstrFlags::MIMG;
629 bool SIInstrInfo::isSMRD(uint16_t Opcode) const {
630 return get(Opcode).TSFlags & SIInstrFlags::SMRD;
633 bool SIInstrInfo::isMUBUF(uint16_t Opcode) const {
634 return get(Opcode).TSFlags & SIInstrFlags::MUBUF;
637 bool SIInstrInfo::isMTBUF(uint16_t Opcode) const {
638 return get(Opcode).TSFlags & SIInstrFlags::MTBUF;
641 bool SIInstrInfo::isFLAT(uint16_t Opcode) const {
642 return get(Opcode).TSFlags & SIInstrFlags::FLAT;
645 bool SIInstrInfo::isVOP1(uint16_t Opcode) const {
646 return get(Opcode).TSFlags & SIInstrFlags::VOP1;
649 bool SIInstrInfo::isVOP2(uint16_t Opcode) const {
650 return get(Opcode).TSFlags & SIInstrFlags::VOP2;
653 bool SIInstrInfo::isVOP3(uint16_t Opcode) const {
654 return get(Opcode).TSFlags & SIInstrFlags::VOP3;
657 bool SIInstrInfo::isVOPC(uint16_t Opcode) const {
658 return get(Opcode).TSFlags & SIInstrFlags::VOPC;
661 bool SIInstrInfo::isSALUInstr(const MachineInstr &MI) const {
662 return get(MI.getOpcode()).TSFlags & SIInstrFlags::SALU;
665 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
666 int32_t Val = Imm.getSExtValue();
667 if (Val >= -16 && Val <= 64)
670 // The actual type of the operand does not seem to matter as long
671 // as the bits match one of the inline immediate values. For example:
673 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
674 // so it is a legal inline immediate.
676 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
677 // floating-point, so it is a legal inline immediate.
679 return (APInt::floatToBits(0.0f) == Imm) ||
680 (APInt::floatToBits(1.0f) == Imm) ||
681 (APInt::floatToBits(-1.0f) == Imm) ||
682 (APInt::floatToBits(0.5f) == Imm) ||
683 (APInt::floatToBits(-0.5f) == Imm) ||
684 (APInt::floatToBits(2.0f) == Imm) ||
685 (APInt::floatToBits(-2.0f) == Imm) ||
686 (APInt::floatToBits(4.0f) == Imm) ||
687 (APInt::floatToBits(-4.0f) == Imm);
690 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO) const {
692 return isInlineConstant(APInt(32, MO.getImm(), true));
695 APFloat FpImm = MO.getFPImm()->getValueAPF();
696 return isInlineConstant(FpImm.bitcastToAPInt());
702 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO) const {
703 return (MO.isImm() || MO.isFPImm()) && !isInlineConstant(MO);
706 static bool compareMachineOp(const MachineOperand &Op0,
707 const MachineOperand &Op1) {
708 if (Op0.getType() != Op1.getType())
711 switch (Op0.getType()) {
712 case MachineOperand::MO_Register:
713 return Op0.getReg() == Op1.getReg();
714 case MachineOperand::MO_Immediate:
715 return Op0.getImm() == Op1.getImm();
716 case MachineOperand::MO_FPImmediate:
717 return Op0.getFPImm() == Op1.getFPImm();
719 llvm_unreachable("Didn't expect to be comparing these operand types");
723 bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
724 const MachineOperand &MO) const {
725 const MCOperandInfo &OpInfo = get(MI->getOpcode()).OpInfo[OpNo];
727 assert(MO.isImm() || MO.isFPImm());
729 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
732 if (OpInfo.RegClass < 0)
735 return RI.regClassCanUseImmediate(OpInfo.RegClass);
738 bool SIInstrInfo::canFoldOffset(unsigned OffsetSize, unsigned AS) {
740 case AMDGPUAS::GLOBAL_ADDRESS: {
741 // MUBUF instructions a 12-bit offset in bytes.
742 return isUInt<12>(OffsetSize);
744 case AMDGPUAS::CONSTANT_ADDRESS: {
745 // SMRD instructions have an 8-bit offset in dwords.
746 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
748 case AMDGPUAS::LOCAL_ADDRESS:
749 case AMDGPUAS::REGION_ADDRESS: {
750 // The single offset versions have a 16-bit offset in bytes.
751 return isUInt<16>(OffsetSize);
753 case AMDGPUAS::PRIVATE_ADDRESS:
754 // Indirect register addressing does not use any offsets.
760 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
761 return AMDGPU::getVOPe32(Opcode) != -1;
764 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
765 // The src0_modifier operand is present on all instructions
766 // that have modifiers.
768 return AMDGPU::getNamedOperandIdx(Opcode,
769 AMDGPU::OpName::src0_modifiers) != -1;
772 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
773 StringRef &ErrInfo) const {
774 uint16_t Opcode = MI->getOpcode();
775 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
776 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
777 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
779 // Make sure the number of operands is correct.
780 const MCInstrDesc &Desc = get(Opcode);
781 if (!Desc.isVariadic() &&
782 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
783 ErrInfo = "Instruction has wrong number of operands.";
787 // Make sure the register classes are correct
788 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
789 switch (Desc.OpInfo[i].OperandType) {
790 case MCOI::OPERAND_REGISTER: {
791 int RegClass = Desc.OpInfo[i].RegClass;
792 if (!RI.regClassCanUseImmediate(RegClass) &&
793 (MI->getOperand(i).isImm() || MI->getOperand(i).isFPImm())) {
794 // Handle some special cases:
795 // Src0 can of VOP1, VOP2, VOPC can be an immediate no matter what
796 // the register class.
797 if (i != Src0Idx || (!isVOP1(Opcode) && !isVOP2(Opcode) &&
799 ErrInfo = "Expected register, but got immediate";
805 case MCOI::OPERAND_IMMEDIATE:
806 // Check if this operand is an immediate.
807 // FrameIndex operands will be replaced by immediates, so they are
809 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFPImm() &&
810 !MI->getOperand(i).isFI()) {
811 ErrInfo = "Expected immediate, but got non-immediate";
819 if (!MI->getOperand(i).isReg())
822 int RegClass = Desc.OpInfo[i].RegClass;
823 if (RegClass != -1) {
824 unsigned Reg = MI->getOperand(i).getReg();
825 if (TargetRegisterInfo::isVirtualRegister(Reg))
828 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
829 if (!RC->contains(Reg)) {
830 ErrInfo = "Operand has incorrect register class.";
838 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
839 unsigned ConstantBusCount = 0;
840 unsigned SGPRUsed = AMDGPU::NoRegister;
841 for (int i = 0, e = MI->getNumOperands(); i != e; ++i) {
842 const MachineOperand &MO = MI->getOperand(i);
843 if (MO.isReg() && MO.isUse() &&
844 !TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
846 // EXEC register uses the constant bus.
847 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
850 // FLAT_SCR is just an SGPR pair.
851 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
854 // SGPRs use the constant bus
855 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
857 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
858 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
859 if (SGPRUsed != MO.getReg()) {
861 SGPRUsed = MO.getReg();
865 // Literal constants use the constant bus.
866 if (isLiteralConstant(MO))
869 if (ConstantBusCount > 1) {
870 ErrInfo = "VOP* instruction uses the constant bus more than once";
875 // Verify SRC1 for VOP2 and VOPC
876 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
877 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
878 if (Src1.isImm() || Src1.isFPImm()) {
879 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
885 if (isVOP3(Opcode)) {
886 if (Src0Idx != -1 && isLiteralConstant(MI->getOperand(Src0Idx))) {
887 ErrInfo = "VOP3 src0 cannot be a literal constant.";
890 if (Src1Idx != -1 && isLiteralConstant(MI->getOperand(Src1Idx))) {
891 ErrInfo = "VOP3 src1 cannot be a literal constant.";
894 if (Src2Idx != -1 && isLiteralConstant(MI->getOperand(Src2Idx))) {
895 ErrInfo = "VOP3 src2 cannot be a literal constant.";
900 // Verify misc. restrictions on specific instructions.
901 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
902 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
905 const MachineOperand &Src0 = MI->getOperand(2);
906 const MachineOperand &Src1 = MI->getOperand(3);
907 const MachineOperand &Src2 = MI->getOperand(4);
908 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
909 if (!compareMachineOp(Src0, Src1) &&
910 !compareMachineOp(Src0, Src2)) {
911 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
920 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
921 switch (MI.getOpcode()) {
922 default: return AMDGPU::INSTRUCTION_LIST_END;
923 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
924 case AMDGPU::COPY: return AMDGPU::COPY;
925 case AMDGPU::PHI: return AMDGPU::PHI;
926 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
927 case AMDGPU::S_MOV_B32:
928 return MI.getOperand(1).isReg() ?
929 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
930 case AMDGPU::S_ADD_I32:
931 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
932 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
933 case AMDGPU::S_SUB_I32:
934 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
935 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
936 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
937 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
938 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
939 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
940 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
941 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
942 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
943 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
944 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
945 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
946 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
947 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
948 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
949 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
950 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
951 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
952 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
953 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
954 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
955 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
956 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
957 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
958 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
959 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
960 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
961 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
962 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
963 case AMDGPU::S_LOAD_DWORD_IMM:
964 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
965 case AMDGPU::S_LOAD_DWORDX2_IMM:
966 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
967 case AMDGPU::S_LOAD_DWORDX4_IMM:
968 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
969 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e32;
970 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
971 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
975 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
976 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
979 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
980 unsigned OpNo) const {
981 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
982 const MCInstrDesc &Desc = get(MI.getOpcode());
983 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
984 Desc.OpInfo[OpNo].RegClass == -1)
985 return MRI.getRegClass(MI.getOperand(OpNo).getReg());
987 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
988 return RI.getRegClass(RCID);
991 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
992 switch (MI.getOpcode()) {
994 case AMDGPU::REG_SEQUENCE:
996 case AMDGPU::INSERT_SUBREG:
997 return RI.hasVGPRs(getOpRegClass(MI, 0));
999 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
1003 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
1004 MachineBasicBlock::iterator I = MI;
1005 MachineOperand &MO = MI->getOperand(OpIdx);
1006 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1007 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
1008 const TargetRegisterClass *RC = RI.getRegClass(RCID);
1009 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1011 Opcode = AMDGPU::COPY;
1012 } else if (RI.isSGPRClass(RC)) {
1013 Opcode = AMDGPU::S_MOV_B32;
1016 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
1017 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC)) {
1018 VRC = &AMDGPU::VReg_64RegClass;
1020 VRC = &AMDGPU::VReg_32RegClass;
1022 unsigned Reg = MRI.createVirtualRegister(VRC);
1023 BuildMI(*MI->getParent(), I, MI->getParent()->findDebugLoc(I), get(Opcode),
1024 Reg).addOperand(MO);
1025 MO.ChangeToRegister(Reg, false);
1028 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
1029 MachineRegisterInfo &MRI,
1030 MachineOperand &SuperReg,
1031 const TargetRegisterClass *SuperRC,
1033 const TargetRegisterClass *SubRC)
1035 assert(SuperReg.isReg());
1037 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
1038 unsigned SubReg = MRI.createVirtualRegister(SubRC);
1040 // Just in case the super register is itself a sub-register, copy it to a new
1041 // value so we don't need to worry about merging its subreg index with the
1042 // SubIdx passed to this function. The register coalescer should be able to
1043 // eliminate this extra copy.
1044 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1046 .addOperand(SuperReg);
1048 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY),
1050 .addReg(NewSuperReg, 0, SubIdx);
1054 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
1055 MachineBasicBlock::iterator MII,
1056 MachineRegisterInfo &MRI,
1058 const TargetRegisterClass *SuperRC,
1060 const TargetRegisterClass *SubRC) const {
1062 // XXX - Is there a better way to do this?
1063 if (SubIdx == AMDGPU::sub0)
1064 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
1065 if (SubIdx == AMDGPU::sub1)
1066 return MachineOperand::CreateImm(Op.getImm() >> 32);
1068 llvm_unreachable("Unhandled register index for immediate");
1071 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
1073 return MachineOperand::CreateReg(SubReg, false);
1076 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
1077 MachineBasicBlock::iterator MI,
1078 MachineRegisterInfo &MRI,
1079 const TargetRegisterClass *RC,
1080 const MachineOperand &Op) const {
1081 MachineBasicBlock *MBB = MI->getParent();
1082 DebugLoc DL = MI->getDebugLoc();
1083 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1084 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1085 unsigned Dst = MRI.createVirtualRegister(RC);
1087 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1089 .addImm(Op.getImm() & 0xFFFFFFFF);
1090 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1092 .addImm(Op.getImm() >> 32);
1094 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
1096 .addImm(AMDGPU::sub0)
1098 .addImm(AMDGPU::sub1);
1100 Worklist.push_back(Lo);
1101 Worklist.push_back(Hi);
1106 bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
1107 const MachineOperand *MO) const {
1108 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1109 const MCInstrDesc &InstDesc = get(MI->getOpcode());
1110 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
1111 const TargetRegisterClass *DefinedRC =
1112 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
1114 MO = &MI->getOperand(OpIdx);
1118 const TargetRegisterClass *RC = MRI.getRegClass(MO->getReg());
1119 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass));
1123 // Handle non-register types that are treated like immediates.
1124 assert(MO->isImm() || MO->isFPImm() || MO->isTargetIndex() || MO->isFI());
1127 // This opperand expects an immediate
1130 return RI.regClassCanUseImmediate(DefinedRC);
1133 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
1134 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1136 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1137 AMDGPU::OpName::src0);
1138 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1139 AMDGPU::OpName::src1);
1140 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1141 AMDGPU::OpName::src2);
1144 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
1146 if (!isOperandLegal(MI, Src0Idx))
1147 legalizeOpWithMove(MI, Src0Idx);
1150 if (isOperandLegal(MI, Src1Idx))
1153 // Usually src0 of VOP2 instructions allow more types of inputs
1154 // than src1, so try to commute the instruction to decrease our
1155 // chances of having to insert a MOV instruction to legalize src1.
1156 if (MI->isCommutable()) {
1157 if (commuteInstruction(MI))
1158 // If we are successful in commuting, then we know MI is legal, so
1163 legalizeOpWithMove(MI, Src1Idx);
1167 // XXX - Do any VOP3 instructions read VCC?
1169 if (isVOP3(MI->getOpcode())) {
1170 int VOP3Idx[3] = {Src0Idx, Src1Idx, Src2Idx};
1171 unsigned SGPRReg = AMDGPU::NoRegister;
1172 for (unsigned i = 0; i < 3; ++i) {
1173 int Idx = VOP3Idx[i];
1176 MachineOperand &MO = MI->getOperand(Idx);
1179 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
1180 continue; // VGPRs are legal
1182 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
1184 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
1185 SGPRReg = MO.getReg();
1186 // We can use one SGPR in each VOP3 instruction.
1189 } else if (!isLiteralConstant(MO)) {
1190 // If it is not a register and not a literal constant, then it must be
1191 // an inline constant which is always legal.
1194 // If we make it this far, then the operand is not legal and we must
1196 legalizeOpWithMove(MI, Idx);
1200 // Legalize REG_SEQUENCE and PHI
1201 // The register class of the operands much be the same type as the register
1202 // class of the output.
1203 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
1204 MI->getOpcode() == AMDGPU::PHI) {
1205 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
1206 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1207 if (!MI->getOperand(i).isReg() ||
1208 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1210 const TargetRegisterClass *OpRC =
1211 MRI.getRegClass(MI->getOperand(i).getReg());
1212 if (RI.hasVGPRs(OpRC)) {
1219 // If any of the operands are VGPR registers, then they all most be
1220 // otherwise we will create illegal VGPR->SGPR copies when legalizing
1222 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
1225 VRC = RI.getEquivalentVGPRClass(SRC);
1232 // Update all the operands so they have the same type.
1233 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1234 if (!MI->getOperand(i).isReg() ||
1235 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1237 unsigned DstReg = MRI.createVirtualRegister(RC);
1238 MachineBasicBlock *InsertBB;
1239 MachineBasicBlock::iterator Insert;
1240 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
1241 InsertBB = MI->getParent();
1244 // MI is a PHI instruction.
1245 InsertBB = MI->getOperand(i + 1).getMBB();
1246 Insert = InsertBB->getFirstTerminator();
1248 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
1249 get(AMDGPU::COPY), DstReg)
1250 .addOperand(MI->getOperand(i));
1251 MI->getOperand(i).setReg(DstReg);
1255 // Legalize INSERT_SUBREG
1256 // src0 must have the same register class as dst
1257 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
1258 unsigned Dst = MI->getOperand(0).getReg();
1259 unsigned Src0 = MI->getOperand(1).getReg();
1260 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
1261 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
1262 if (DstRC != Src0RC) {
1263 MachineBasicBlock &MBB = *MI->getParent();
1264 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
1265 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
1267 MI->getOperand(1).setReg(NewSrc0);
1272 // Legalize MUBUF* instructions
1273 // FIXME: If we start using the non-addr64 instructions for compute, we
1274 // may need to legalize them here.
1276 AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::srsrc);
1277 if (SRsrcIdx != -1) {
1278 // We have an MUBUF instruction
1279 MachineOperand *SRsrc = &MI->getOperand(SRsrcIdx);
1280 unsigned SRsrcRC = get(MI->getOpcode()).OpInfo[SRsrcIdx].RegClass;
1281 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
1282 RI.getRegClass(SRsrcRC))) {
1283 // The operands are legal.
1284 // FIXME: We may need to legalize operands besided srsrc.
1288 MachineBasicBlock &MBB = *MI->getParent();
1289 // Extract the the ptr from the resource descriptor.
1291 // SRsrcPtrLo = srsrc:sub0
1292 unsigned SRsrcPtrLo = buildExtractSubReg(MI, MRI, *SRsrc,
1293 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VReg_32RegClass);
1295 // SRsrcPtrHi = srsrc:sub1
1296 unsigned SRsrcPtrHi = buildExtractSubReg(MI, MRI, *SRsrc,
1297 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VReg_32RegClass);
1299 // Create an empty resource descriptor
1300 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1301 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1302 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1303 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1306 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1310 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1311 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1313 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1315 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1316 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1318 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1320 // NewSRsrc = {Zero64, SRsrcFormat}
1321 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1324 .addImm(AMDGPU::sub0_sub1)
1325 .addReg(SRsrcFormatLo)
1326 .addImm(AMDGPU::sub2)
1327 .addReg(SRsrcFormatHi)
1328 .addImm(AMDGPU::sub3);
1330 MachineOperand *VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1331 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
1332 unsigned NewVAddrLo;
1333 unsigned NewVAddrHi;
1335 // This is already an ADDR64 instruction so we need to add the pointer
1336 // extracted from the resource descriptor to the current value of VAddr.
1337 NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1338 NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1340 // NewVaddrLo = SRsrcPtrLo + VAddr:sub0
1341 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1344 .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
1345 .addReg(AMDGPU::VCC, RegState::ImplicitDefine);
1347 // NewVaddrHi = SRsrcPtrHi + VAddr:sub1
1348 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1351 .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
1352 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1353 .addReg(AMDGPU::VCC, RegState::Implicit);
1356 // This instructions is the _OFFSET variant, so we need to convert it to
1358 MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
1359 MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
1360 MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
1361 assert(SOffset->isImm() && SOffset->getImm() == 0 && "Legalizing MUBUF "
1362 "with non-zero soffset is not implemented");
1365 // Create the new instruction.
1366 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
1367 MachineInstr *Addr64 =
1368 BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
1371 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
1372 // This will be replaced later
1373 // with the new value of vaddr.
1374 .addOperand(*Offset);
1376 MI->removeFromParent();
1379 NewVAddrLo = SRsrcPtrLo;
1380 NewVAddrHi = SRsrcPtrHi;
1381 VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1382 SRsrc = getNamedOperand(*MI, AMDGPU::OpName::srsrc);
1385 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1386 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1389 .addImm(AMDGPU::sub0)
1391 .addImm(AMDGPU::sub1);
1394 // Update the instruction to use NewVaddr
1395 VAddr->setReg(NewVAddr);
1396 // Update the instruction to use NewSRsrc
1397 SRsrc->setReg(NewSRsrc);
1401 void SIInstrInfo::splitSMRD(MachineInstr *MI,
1402 const TargetRegisterClass *HalfRC,
1403 unsigned HalfImmOp, unsigned HalfSGPROp,
1404 MachineInstr *&Lo, MachineInstr *&Hi) const {
1406 DebugLoc DL = MI->getDebugLoc();
1407 MachineBasicBlock *MBB = MI->getParent();
1408 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1409 unsigned RegLo = MRI.createVirtualRegister(HalfRC);
1410 unsigned RegHi = MRI.createVirtualRegister(HalfRC);
1411 unsigned HalfSize = HalfRC->getSize();
1412 const MachineOperand *OffOp =
1413 getNamedOperand(*MI, AMDGPU::OpName::offset);
1414 const MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
1417 // Handle the _IMM variant
1418 unsigned LoOffset = OffOp->getImm();
1419 unsigned HiOffset = LoOffset + (HalfSize / 4);
1420 Lo = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegLo)
1424 if (!isUInt<8>(HiOffset)) {
1425 unsigned OffsetSGPR =
1426 MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1427 BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32), OffsetSGPR)
1428 .addImm(HiOffset << 2); // The immediate offset is in dwords,
1429 // but offset in register is in bytes.
1430 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
1432 .addReg(OffsetSGPR);
1434 Hi = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegHi)
1439 // Handle the _SGPR variant
1440 MachineOperand *SOff = getNamedOperand(*MI, AMDGPU::OpName::soff);
1441 Lo = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegLo)
1444 unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1445 BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
1448 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp))
1450 .addReg(OffsetSGPR);
1453 unsigned SubLo, SubHi;
1456 SubLo = AMDGPU::sub0;
1457 SubHi = AMDGPU::sub1;
1460 SubLo = AMDGPU::sub0_sub1;
1461 SubHi = AMDGPU::sub2_sub3;
1464 SubLo = AMDGPU::sub0_sub1_sub2_sub3;
1465 SubHi = AMDGPU::sub4_sub5_sub6_sub7;
1468 SubLo = AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
1469 SubHi = AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15;
1472 llvm_unreachable("Unhandled HalfSize");
1475 BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE))
1476 .addOperand(MI->getOperand(0))
1483 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
1484 MachineBasicBlock *MBB = MI->getParent();
1485 switch (MI->getOpcode()) {
1486 case AMDGPU::S_LOAD_DWORD_IMM:
1487 case AMDGPU::S_LOAD_DWORD_SGPR:
1488 case AMDGPU::S_LOAD_DWORDX2_IMM:
1489 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1490 case AMDGPU::S_LOAD_DWORDX4_IMM:
1491 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1492 unsigned NewOpcode = getVALUOp(*MI);
1496 if (MI->getOperand(2).isReg()) {
1497 RegOffset = MI->getOperand(2).getReg();
1500 assert(MI->getOperand(2).isImm());
1501 // SMRD instructions take a dword offsets and MUBUF instructions
1502 // take a byte offset.
1503 ImmOffset = MI->getOperand(2).getImm() << 2;
1504 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1505 if (isUInt<12>(ImmOffset)) {
1506 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1510 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1517 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1518 unsigned DWord0 = RegOffset;
1519 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1520 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1521 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1523 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
1525 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
1526 .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
1527 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
1528 .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
1529 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
1531 .addImm(AMDGPU::sub0)
1533 .addImm(AMDGPU::sub1)
1535 .addImm(AMDGPU::sub2)
1537 .addImm(AMDGPU::sub3);
1538 MI->setDesc(get(NewOpcode));
1539 if (MI->getOperand(2).isReg()) {
1540 MI->getOperand(2).setReg(MI->getOperand(1).getReg());
1542 MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
1544 MI->getOperand(1).setReg(SRsrc);
1545 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
1547 const TargetRegisterClass *NewDstRC =
1548 RI.getRegClass(get(NewOpcode).OpInfo[0].RegClass);
1550 unsigned DstReg = MI->getOperand(0).getReg();
1551 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1552 MRI.replaceRegWith(DstReg, NewDstReg);
1555 case AMDGPU::S_LOAD_DWORDX8_IMM:
1556 case AMDGPU::S_LOAD_DWORDX8_SGPR: {
1557 MachineInstr *Lo, *Hi;
1558 splitSMRD(MI, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM,
1559 AMDGPU::S_LOAD_DWORDX4_SGPR, Lo, Hi);
1560 MI->eraseFromParent();
1561 moveSMRDToVALU(Lo, MRI);
1562 moveSMRDToVALU(Hi, MRI);
1566 case AMDGPU::S_LOAD_DWORDX16_IMM:
1567 case AMDGPU::S_LOAD_DWORDX16_SGPR: {
1568 MachineInstr *Lo, *Hi;
1569 splitSMRD(MI, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM,
1570 AMDGPU::S_LOAD_DWORDX8_SGPR, Lo, Hi);
1571 MI->eraseFromParent();
1572 moveSMRDToVALU(Lo, MRI);
1573 moveSMRDToVALU(Hi, MRI);
1579 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
1580 SmallVector<MachineInstr *, 128> Worklist;
1581 Worklist.push_back(&TopInst);
1583 while (!Worklist.empty()) {
1584 MachineInstr *Inst = Worklist.pop_back_val();
1585 MachineBasicBlock *MBB = Inst->getParent();
1586 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1588 unsigned Opcode = Inst->getOpcode();
1589 unsigned NewOpcode = getVALUOp(*Inst);
1591 // Handle some special cases
1594 if (isSMRD(Inst->getOpcode())) {
1595 moveSMRDToVALU(Inst, MRI);
1598 case AMDGPU::S_MOV_B64: {
1599 DebugLoc DL = Inst->getDebugLoc();
1601 // If the source operand is a register we can replace this with a
1603 if (Inst->getOperand(1).isReg()) {
1604 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
1605 .addOperand(Inst->getOperand(0))
1606 .addOperand(Inst->getOperand(1));
1607 Worklist.push_back(Copy);
1609 // Otherwise, we need to split this into two movs, because there is
1610 // no 64-bit VALU move instruction.
1611 unsigned Reg = Inst->getOperand(0).getReg();
1612 unsigned Dst = split64BitImm(Worklist,
1615 MRI.getRegClass(Reg),
1616 Inst->getOperand(1));
1617 MRI.replaceRegWith(Reg, Dst);
1619 Inst->eraseFromParent();
1622 case AMDGPU::S_AND_B64:
1623 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
1624 Inst->eraseFromParent();
1627 case AMDGPU::S_OR_B64:
1628 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
1629 Inst->eraseFromParent();
1632 case AMDGPU::S_XOR_B64:
1633 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
1634 Inst->eraseFromParent();
1637 case AMDGPU::S_NOT_B64:
1638 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
1639 Inst->eraseFromParent();
1642 case AMDGPU::S_BCNT1_I32_B64:
1643 splitScalar64BitBCNT(Worklist, Inst);
1644 Inst->eraseFromParent();
1647 case AMDGPU::S_BFE_U64:
1648 case AMDGPU::S_BFE_I64:
1649 case AMDGPU::S_BFM_B64:
1650 llvm_unreachable("Moving this op to VALU not implemented");
1653 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
1654 // We cannot move this instruction to the VALU, so we should try to
1655 // legalize its operands instead.
1656 legalizeOperands(Inst);
1660 // Use the new VALU Opcode.
1661 const MCInstrDesc &NewDesc = get(NewOpcode);
1662 Inst->setDesc(NewDesc);
1664 // Remove any references to SCC. Vector instructions can't read from it, and
1665 // We're just about to add the implicit use / defs of VCC, and we don't want
1667 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
1668 MachineOperand &Op = Inst->getOperand(i);
1669 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
1670 Inst->RemoveOperand(i);
1673 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
1674 // We are converting these to a BFE, so we need to add the missing
1675 // operands for the size and offset.
1676 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
1677 Inst->addOperand(MachineOperand::CreateImm(0));
1678 Inst->addOperand(MachineOperand::CreateImm(Size));
1680 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
1681 // The VALU version adds the second operand to the result, so insert an
1683 Inst->addOperand(MachineOperand::CreateImm(0));
1686 addDescImplicitUseDef(NewDesc, Inst);
1688 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
1689 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
1690 // If we need to move this to VGPRs, we need to unpack the second operand
1691 // back into the 2 separate ones for bit offset and width.
1692 assert(OffsetWidthOp.isImm() &&
1693 "Scalar BFE is only implemented for constant width and offset");
1694 uint32_t Imm = OffsetWidthOp.getImm();
1696 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
1697 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
1698 Inst->RemoveOperand(2); // Remove old immediate.
1699 Inst->addOperand(MachineOperand::CreateImm(Offset));
1700 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
1703 // Update the destination register class.
1705 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
1708 // For target instructions, getOpRegClass just returns the virtual
1709 // register class associated with the operand, so we need to find an
1710 // equivalent VGPR register class in order to move the instruction to the
1714 case AMDGPU::REG_SEQUENCE:
1715 case AMDGPU::INSERT_SUBREG:
1716 if (RI.hasVGPRs(NewDstRC))
1718 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
1726 unsigned DstReg = Inst->getOperand(0).getReg();
1727 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1728 MRI.replaceRegWith(DstReg, NewDstReg);
1730 // Legalize the operands
1731 legalizeOperands(Inst);
1733 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
1734 E = MRI.use_end(); I != E; ++I) {
1735 MachineInstr &UseMI = *I->getParent();
1736 if (!canReadVGPR(UseMI, I.getOperandNo())) {
1737 Worklist.push_back(&UseMI);
1743 //===----------------------------------------------------------------------===//
1744 // Indirect addressing callbacks
1745 //===----------------------------------------------------------------------===//
1747 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
1748 unsigned Channel) const {
1749 assert(Channel == 0);
1753 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
1754 return &AMDGPU::VReg_32RegClass;
1757 void SIInstrInfo::splitScalar64BitUnaryOp(
1758 SmallVectorImpl<MachineInstr *> &Worklist,
1760 unsigned Opcode) const {
1761 MachineBasicBlock &MBB = *Inst->getParent();
1762 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1764 MachineOperand &Dest = Inst->getOperand(0);
1765 MachineOperand &Src0 = Inst->getOperand(1);
1766 DebugLoc DL = Inst->getDebugLoc();
1768 MachineBasicBlock::iterator MII = Inst;
1770 const MCInstrDesc &InstDesc = get(Opcode);
1771 const TargetRegisterClass *Src0RC = Src0.isReg() ?
1772 MRI.getRegClass(Src0.getReg()) :
1773 &AMDGPU::SGPR_32RegClass;
1775 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
1777 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1778 AMDGPU::sub0, Src0SubRC);
1780 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
1781 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
1783 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
1784 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
1785 .addOperand(SrcReg0Sub0);
1787 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1788 AMDGPU::sub1, Src0SubRC);
1790 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
1791 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
1792 .addOperand(SrcReg0Sub1);
1794 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
1795 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
1797 .addImm(AMDGPU::sub0)
1799 .addImm(AMDGPU::sub1);
1801 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
1803 // Try to legalize the operands in case we need to swap the order to keep it
1805 Worklist.push_back(LoHalf);
1806 Worklist.push_back(HiHalf);
1809 void SIInstrInfo::splitScalar64BitBinaryOp(
1810 SmallVectorImpl<MachineInstr *> &Worklist,
1812 unsigned Opcode) const {
1813 MachineBasicBlock &MBB = *Inst->getParent();
1814 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1816 MachineOperand &Dest = Inst->getOperand(0);
1817 MachineOperand &Src0 = Inst->getOperand(1);
1818 MachineOperand &Src1 = Inst->getOperand(2);
1819 DebugLoc DL = Inst->getDebugLoc();
1821 MachineBasicBlock::iterator MII = Inst;
1823 const MCInstrDesc &InstDesc = get(Opcode);
1824 const TargetRegisterClass *Src0RC = Src0.isReg() ?
1825 MRI.getRegClass(Src0.getReg()) :
1826 &AMDGPU::SGPR_32RegClass;
1828 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
1829 const TargetRegisterClass *Src1RC = Src1.isReg() ?
1830 MRI.getRegClass(Src1.getReg()) :
1831 &AMDGPU::SGPR_32RegClass;
1833 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
1835 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1836 AMDGPU::sub0, Src0SubRC);
1837 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
1838 AMDGPU::sub0, Src1SubRC);
1840 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
1841 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
1843 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
1844 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
1845 .addOperand(SrcReg0Sub0)
1846 .addOperand(SrcReg1Sub0);
1848 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
1849 AMDGPU::sub1, Src0SubRC);
1850 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
1851 AMDGPU::sub1, Src1SubRC);
1853 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
1854 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
1855 .addOperand(SrcReg0Sub1)
1856 .addOperand(SrcReg1Sub1);
1858 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
1859 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
1861 .addImm(AMDGPU::sub0)
1863 .addImm(AMDGPU::sub1);
1865 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
1867 // Try to legalize the operands in case we need to swap the order to keep it
1869 Worklist.push_back(LoHalf);
1870 Worklist.push_back(HiHalf);
1873 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
1874 MachineInstr *Inst) const {
1875 MachineBasicBlock &MBB = *Inst->getParent();
1876 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1878 MachineBasicBlock::iterator MII = Inst;
1879 DebugLoc DL = Inst->getDebugLoc();
1881 MachineOperand &Dest = Inst->getOperand(0);
1882 MachineOperand &Src = Inst->getOperand(1);
1884 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e32);
1885 const TargetRegisterClass *SrcRC = Src.isReg() ?
1886 MRI.getRegClass(Src.getReg()) :
1887 &AMDGPU::SGPR_32RegClass;
1889 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1890 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1892 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
1894 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
1895 AMDGPU::sub0, SrcSubRC);
1896 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
1897 AMDGPU::sub1, SrcSubRC);
1899 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
1900 .addOperand(SrcRegSub0)
1903 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
1904 .addOperand(SrcRegSub1)
1907 MRI.replaceRegWith(Dest.getReg(), ResultReg);
1909 Worklist.push_back(First);
1910 Worklist.push_back(Second);
1913 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
1914 MachineInstr *Inst) const {
1915 // Add the implict and explicit register definitions.
1916 if (NewDesc.ImplicitUses) {
1917 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
1918 unsigned Reg = NewDesc.ImplicitUses[i];
1919 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
1923 if (NewDesc.ImplicitDefs) {
1924 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
1925 unsigned Reg = NewDesc.ImplicitDefs[i];
1926 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
1931 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
1932 MachineBasicBlock *MBB,
1933 MachineBasicBlock::iterator I,
1935 unsigned Address, unsigned OffsetReg) const {
1936 const DebugLoc &DL = MBB->findDebugLoc(I);
1937 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
1938 getIndirectIndexBegin(*MBB->getParent()));
1940 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
1941 .addReg(IndirectBaseReg, RegState::Define)
1942 .addOperand(I->getOperand(0))
1943 .addReg(IndirectBaseReg)
1949 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
1950 MachineBasicBlock *MBB,
1951 MachineBasicBlock::iterator I,
1953 unsigned Address, unsigned OffsetReg) const {
1954 const DebugLoc &DL = MBB->findDebugLoc(I);
1955 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
1956 getIndirectIndexBegin(*MBB->getParent()));
1958 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
1959 .addOperand(I->getOperand(0))
1960 .addOperand(I->getOperand(1))
1961 .addReg(IndirectBaseReg)
1967 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
1968 const MachineFunction &MF) const {
1969 int End = getIndirectIndexEnd(MF);
1970 int Begin = getIndirectIndexBegin(MF);
1976 for (int Index = Begin; Index <= End; ++Index)
1977 Reserved.set(AMDGPU::VReg_32RegClass.getRegister(Index));
1979 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
1980 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
1982 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
1983 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
1985 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
1986 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
1988 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
1989 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
1991 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
1992 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));
1995 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
1996 unsigned OperandName) const {
1997 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
2001 return &MI.getOperand(Idx);