1 //===-- SIInstrInfo.cpp - SI Instruction Information ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief SI Implementation of TargetInstrInfo.
13 //===----------------------------------------------------------------------===//
16 #include "SIInstrInfo.h"
17 #include "AMDGPUTargetMachine.h"
18 #include "SIDefines.h"
19 #include "SIMachineFunctionInfo.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/CodeGen/RegisterScavenging.h"
25 #include "llvm/MC/MCInstrDesc.h"
26 #include "llvm/Support/Debug.h"
30 SIInstrInfo::SIInstrInfo(const AMDGPUSubtarget &st)
31 : AMDGPUInstrInfo(st), RI() {}
33 //===----------------------------------------------------------------------===//
34 // TargetInstrInfo callbacks
35 //===----------------------------------------------------------------------===//
37 static unsigned getNumOperandsNoGlue(SDNode *Node) {
38 unsigned N = Node->getNumOperands();
39 while (N && Node->getOperand(N - 1).getValueType() == MVT::Glue)
44 static SDValue findChainOperand(SDNode *Load) {
45 SDValue LastOp = Load->getOperand(getNumOperandsNoGlue(Load) - 1);
46 assert(LastOp.getValueType() == MVT::Other && "Chain missing from load node");
50 /// \brief Returns true if both nodes have the same value for the given
51 /// operand \p Op, or if both nodes do not have this operand.
52 static bool nodesHaveSameOperandValue(SDNode *N0, SDNode* N1, unsigned OpName) {
53 unsigned Opc0 = N0->getMachineOpcode();
54 unsigned Opc1 = N1->getMachineOpcode();
56 int Op0Idx = AMDGPU::getNamedOperandIdx(Opc0, OpName);
57 int Op1Idx = AMDGPU::getNamedOperandIdx(Opc1, OpName);
59 if (Op0Idx == -1 && Op1Idx == -1)
63 if ((Op0Idx == -1 && Op1Idx != -1) ||
64 (Op1Idx == -1 && Op0Idx != -1))
67 // getNamedOperandIdx returns the index for the MachineInstr's operands,
68 // which includes the result as the first operand. We are indexing into the
69 // MachineSDNode's operands, so we need to skip the result operand to get
74 return N0->getOperand(Op0Idx) == N1->getOperand(Op1Idx);
77 bool SIInstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
78 AliasAnalysis *AA) const {
79 // TODO: The generic check fails for VALU instructions that should be
80 // rematerializable due to implicit reads of exec. We really want all of the
81 // generic logic for this except for this.
82 switch (MI->getOpcode()) {
83 case AMDGPU::V_MOV_B32_e32:
84 case AMDGPU::V_MOV_B32_e64:
91 bool SIInstrInfo::areLoadsFromSameBasePtr(SDNode *Load0, SDNode *Load1,
93 int64_t &Offset1) const {
94 if (!Load0->isMachineOpcode() || !Load1->isMachineOpcode())
97 unsigned Opc0 = Load0->getMachineOpcode();
98 unsigned Opc1 = Load1->getMachineOpcode();
100 // Make sure both are actually loads.
101 if (!get(Opc0).mayLoad() || !get(Opc1).mayLoad())
104 if (isDS(Opc0) && isDS(Opc1)) {
106 // FIXME: Handle this case:
107 if (getNumOperandsNoGlue(Load0) != getNumOperandsNoGlue(Load1))
111 if (Load0->getOperand(1) != Load1->getOperand(1))
115 if (findChainOperand(Load0) != findChainOperand(Load1))
118 // Skip read2 / write2 variants for simplicity.
119 // TODO: We should report true if the used offsets are adjacent (excluded
121 if (AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::data1) != -1 ||
122 AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::data1) != -1)
125 Offset0 = cast<ConstantSDNode>(Load0->getOperand(2))->getZExtValue();
126 Offset1 = cast<ConstantSDNode>(Load1->getOperand(2))->getZExtValue();
130 if (isSMRD(Opc0) && isSMRD(Opc1)) {
131 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
134 if (Load0->getOperand(0) != Load1->getOperand(0))
137 const ConstantSDNode *Load0Offset =
138 dyn_cast<ConstantSDNode>(Load0->getOperand(1));
139 const ConstantSDNode *Load1Offset =
140 dyn_cast<ConstantSDNode>(Load1->getOperand(1));
142 if (!Load0Offset || !Load1Offset)
146 if (findChainOperand(Load0) != findChainOperand(Load1))
149 Offset0 = Load0Offset->getZExtValue();
150 Offset1 = Load1Offset->getZExtValue();
154 // MUBUF and MTBUF can access the same addresses.
155 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1))) {
157 // MUBUF and MTBUF have vaddr at different indices.
158 if (!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::soffset) ||
159 findChainOperand(Load0) != findChainOperand(Load1) ||
160 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::vaddr) ||
161 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::srsrc))
164 int OffIdx0 = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
165 int OffIdx1 = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
167 if (OffIdx0 == -1 || OffIdx1 == -1)
170 // getNamedOperandIdx returns the index for MachineInstrs. Since they
171 // inlcude the output in the operand list, but SDNodes don't, we need to
172 // subtract the index by one.
176 SDValue Off0 = Load0->getOperand(OffIdx0);
177 SDValue Off1 = Load1->getOperand(OffIdx1);
179 // The offset might be a FrameIndexSDNode.
180 if (!isa<ConstantSDNode>(Off0) || !isa<ConstantSDNode>(Off1))
183 Offset0 = cast<ConstantSDNode>(Off0)->getZExtValue();
184 Offset1 = cast<ConstantSDNode>(Off1)->getZExtValue();
191 static bool isStride64(unsigned Opc) {
193 case AMDGPU::DS_READ2ST64_B32:
194 case AMDGPU::DS_READ2ST64_B64:
195 case AMDGPU::DS_WRITE2ST64_B32:
196 case AMDGPU::DS_WRITE2ST64_B64:
203 bool SIInstrInfo::getMemOpBaseRegImmOfs(MachineInstr *LdSt, unsigned &BaseReg,
205 const TargetRegisterInfo *TRI) const {
206 unsigned Opc = LdSt->getOpcode();
208 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
209 AMDGPU::OpName::offset);
211 // Normal, single offset LDS instruction.
212 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
213 AMDGPU::OpName::addr);
215 BaseReg = AddrReg->getReg();
216 Offset = OffsetImm->getImm();
220 // The 2 offset instructions use offset0 and offset1 instead. We can treat
221 // these as a load with a single offset if the 2 offsets are consecutive. We
222 // will use this for some partially aligned loads.
223 const MachineOperand *Offset0Imm = getNamedOperand(*LdSt,
224 AMDGPU::OpName::offset0);
225 const MachineOperand *Offset1Imm = getNamedOperand(*LdSt,
226 AMDGPU::OpName::offset1);
228 uint8_t Offset0 = Offset0Imm->getImm();
229 uint8_t Offset1 = Offset1Imm->getImm();
230 assert(Offset1 > Offset0);
232 if (Offset1 - Offset0 == 1) {
233 // Each of these offsets is in element sized units, so we need to convert
234 // to bytes of the individual reads.
238 EltSize = getOpRegClass(*LdSt, 0)->getSize() / 2;
240 assert(LdSt->mayStore());
241 int Data0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
242 EltSize = getOpRegClass(*LdSt, Data0Idx)->getSize();
248 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
249 AMDGPU::OpName::addr);
250 BaseReg = AddrReg->getReg();
251 Offset = EltSize * Offset0;
258 if (isMUBUF(Opc) || isMTBUF(Opc)) {
259 if (AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::soffset) != -1)
262 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
263 AMDGPU::OpName::vaddr);
267 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
268 AMDGPU::OpName::offset);
269 BaseReg = AddrReg->getReg();
270 Offset = OffsetImm->getImm();
275 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
276 AMDGPU::OpName::offset);
280 const MachineOperand *SBaseReg = getNamedOperand(*LdSt,
281 AMDGPU::OpName::sbase);
282 BaseReg = SBaseReg->getReg();
283 Offset = OffsetImm->getImm();
290 bool SIInstrInfo::shouldClusterLoads(MachineInstr *FirstLdSt,
291 MachineInstr *SecondLdSt,
292 unsigned NumLoads) const {
293 unsigned Opc0 = FirstLdSt->getOpcode();
294 unsigned Opc1 = SecondLdSt->getOpcode();
296 // TODO: This needs finer tuning
300 if (isDS(Opc0) && isDS(Opc1))
303 if (isSMRD(Opc0) && isSMRD(Opc1))
306 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1)))
313 SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
314 MachineBasicBlock::iterator MI, DebugLoc DL,
315 unsigned DestReg, unsigned SrcReg,
316 bool KillSrc) const {
318 // If we are trying to copy to or from SCC, there is a bug somewhere else in
319 // the backend. While it may be theoretically possible to do this, it should
320 // never be necessary.
321 assert(DestReg != AMDGPU::SCC && SrcReg != AMDGPU::SCC);
323 static const int16_t Sub0_15[] = {
324 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
325 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
326 AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
327 AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15, 0
330 static const int16_t Sub0_7[] = {
331 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
332 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7, 0
335 static const int16_t Sub0_3[] = {
336 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
339 static const int16_t Sub0_2[] = {
340 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
343 static const int16_t Sub0_1[] = {
344 AMDGPU::sub0, AMDGPU::sub1, 0
348 const int16_t *SubIndices;
350 if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
351 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
352 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
353 .addReg(SrcReg, getKillRegState(KillSrc));
356 } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
357 if (DestReg == AMDGPU::VCC) {
358 if (AMDGPU::SReg_64RegClass.contains(SrcReg)) {
359 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), AMDGPU::VCC)
360 .addReg(SrcReg, getKillRegState(KillSrc));
362 // FIXME: Hack until VReg_1 removed.
363 assert(AMDGPU::VGPR_32RegClass.contains(SrcReg));
364 BuildMI(MBB, MI, DL, get(AMDGPU::V_CMP_NE_I32_e32), AMDGPU::VCC)
366 .addReg(SrcReg, getKillRegState(KillSrc));
372 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
373 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
374 .addReg(SrcReg, getKillRegState(KillSrc));
377 } else if (AMDGPU::SReg_128RegClass.contains(DestReg)) {
378 assert(AMDGPU::SReg_128RegClass.contains(SrcReg));
379 Opcode = AMDGPU::S_MOV_B32;
382 } else if (AMDGPU::SReg_256RegClass.contains(DestReg)) {
383 assert(AMDGPU::SReg_256RegClass.contains(SrcReg));
384 Opcode = AMDGPU::S_MOV_B32;
387 } else if (AMDGPU::SReg_512RegClass.contains(DestReg)) {
388 assert(AMDGPU::SReg_512RegClass.contains(SrcReg));
389 Opcode = AMDGPU::S_MOV_B32;
390 SubIndices = Sub0_15;
392 } else if (AMDGPU::VGPR_32RegClass.contains(DestReg)) {
393 assert(AMDGPU::VGPR_32RegClass.contains(SrcReg) ||
394 AMDGPU::SReg_32RegClass.contains(SrcReg));
395 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
396 .addReg(SrcReg, getKillRegState(KillSrc));
399 } else if (AMDGPU::VReg_64RegClass.contains(DestReg)) {
400 assert(AMDGPU::VReg_64RegClass.contains(SrcReg) ||
401 AMDGPU::SReg_64RegClass.contains(SrcReg));
402 Opcode = AMDGPU::V_MOV_B32_e32;
405 } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
406 assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
407 Opcode = AMDGPU::V_MOV_B32_e32;
410 } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
411 assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
412 AMDGPU::SReg_128RegClass.contains(SrcReg));
413 Opcode = AMDGPU::V_MOV_B32_e32;
416 } else if (AMDGPU::VReg_256RegClass.contains(DestReg)) {
417 assert(AMDGPU::VReg_256RegClass.contains(SrcReg) ||
418 AMDGPU::SReg_256RegClass.contains(SrcReg));
419 Opcode = AMDGPU::V_MOV_B32_e32;
422 } else if (AMDGPU::VReg_512RegClass.contains(DestReg)) {
423 assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
424 AMDGPU::SReg_512RegClass.contains(SrcReg));
425 Opcode = AMDGPU::V_MOV_B32_e32;
426 SubIndices = Sub0_15;
429 llvm_unreachable("Can't copy register!");
432 while (unsigned SubIdx = *SubIndices++) {
433 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
434 get(Opcode), RI.getSubReg(DestReg, SubIdx));
436 Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
439 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
443 int SIInstrInfo::commuteOpcode(const MachineInstr &MI) const {
444 const unsigned Opcode = MI.getOpcode();
448 // Try to map original to commuted opcode
449 NewOpc = AMDGPU::getCommuteRev(Opcode);
451 // Check if the commuted (REV) opcode exists on the target.
452 return pseudoToMCOpcode(NewOpc) != -1 ? NewOpc : -1;
454 // Try to map commuted to original opcode
455 NewOpc = AMDGPU::getCommuteOrig(Opcode);
457 // Check if the original (non-REV) opcode exists on the target.
458 return pseudoToMCOpcode(NewOpc) != -1 ? NewOpc : -1;
463 unsigned SIInstrInfo::getMovOpcode(const TargetRegisterClass *DstRC) const {
465 if (DstRC->getSize() == 4) {
466 return RI.isSGPRClass(DstRC) ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32;
467 } else if (DstRC->getSize() == 8 && RI.isSGPRClass(DstRC)) {
468 return AMDGPU::S_MOV_B64;
469 } else if (DstRC->getSize() == 8 && !RI.isSGPRClass(DstRC)) {
470 return AMDGPU::V_MOV_B64_PSEUDO;
475 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
476 MachineBasicBlock::iterator MI,
477 unsigned SrcReg, bool isKill,
479 const TargetRegisterClass *RC,
480 const TargetRegisterInfo *TRI) const {
481 MachineFunction *MF = MBB.getParent();
482 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
483 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
484 DebugLoc DL = MBB.findDebugLoc(MI);
487 if (RI.isSGPRClass(RC)) {
488 // We are only allowed to create one new instruction when spilling
489 // registers, so we need to use pseudo instruction for spilling
491 switch (RC->getSize() * 8) {
492 case 32: Opcode = AMDGPU::SI_SPILL_S32_SAVE; break;
493 case 64: Opcode = AMDGPU::SI_SPILL_S64_SAVE; break;
494 case 128: Opcode = AMDGPU::SI_SPILL_S128_SAVE; break;
495 case 256: Opcode = AMDGPU::SI_SPILL_S256_SAVE; break;
496 case 512: Opcode = AMDGPU::SI_SPILL_S512_SAVE; break;
498 } else if(RI.hasVGPRs(RC) && ST.isVGPRSpillingEnabled(MFI)) {
499 MFI->setHasSpilledVGPRs();
501 switch(RC->getSize() * 8) {
502 case 32: Opcode = AMDGPU::SI_SPILL_V32_SAVE; break;
503 case 64: Opcode = AMDGPU::SI_SPILL_V64_SAVE; break;
504 case 96: Opcode = AMDGPU::SI_SPILL_V96_SAVE; break;
505 case 128: Opcode = AMDGPU::SI_SPILL_V128_SAVE; break;
506 case 256: Opcode = AMDGPU::SI_SPILL_V256_SAVE; break;
507 case 512: Opcode = AMDGPU::SI_SPILL_V512_SAVE; break;
512 FrameInfo->setObjectAlignment(FrameIndex, 4);
513 BuildMI(MBB, MI, DL, get(Opcode))
515 .addFrameIndex(FrameIndex)
516 // Place-holder registers, these will be filled in by
517 // SIPrepareScratchRegs.
518 .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
519 .addReg(AMDGPU::SGPR0, RegState::Undef);
521 LLVMContext &Ctx = MF->getFunction()->getContext();
522 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Do not know how to"
524 BuildMI(MBB, MI, DL, get(AMDGPU::KILL))
529 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
530 MachineBasicBlock::iterator MI,
531 unsigned DestReg, int FrameIndex,
532 const TargetRegisterClass *RC,
533 const TargetRegisterInfo *TRI) const {
534 MachineFunction *MF = MBB.getParent();
535 const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
536 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
537 DebugLoc DL = MBB.findDebugLoc(MI);
540 if (RI.isSGPRClass(RC)){
541 switch(RC->getSize() * 8) {
542 case 32: Opcode = AMDGPU::SI_SPILL_S32_RESTORE; break;
543 case 64: Opcode = AMDGPU::SI_SPILL_S64_RESTORE; break;
544 case 128: Opcode = AMDGPU::SI_SPILL_S128_RESTORE; break;
545 case 256: Opcode = AMDGPU::SI_SPILL_S256_RESTORE; break;
546 case 512: Opcode = AMDGPU::SI_SPILL_S512_RESTORE; break;
548 } else if(RI.hasVGPRs(RC) && ST.isVGPRSpillingEnabled(MFI)) {
549 switch(RC->getSize() * 8) {
550 case 32: Opcode = AMDGPU::SI_SPILL_V32_RESTORE; break;
551 case 64: Opcode = AMDGPU::SI_SPILL_V64_RESTORE; break;
552 case 96: Opcode = AMDGPU::SI_SPILL_V96_RESTORE; break;
553 case 128: Opcode = AMDGPU::SI_SPILL_V128_RESTORE; break;
554 case 256: Opcode = AMDGPU::SI_SPILL_V256_RESTORE; break;
555 case 512: Opcode = AMDGPU::SI_SPILL_V512_RESTORE; break;
560 FrameInfo->setObjectAlignment(FrameIndex, 4);
561 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
562 .addFrameIndex(FrameIndex)
563 // Place-holder registers, these will be filled in by
564 // SIPrepareScratchRegs.
565 .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
566 .addReg(AMDGPU::SGPR0, RegState::Undef);
569 LLVMContext &Ctx = MF->getFunction()->getContext();
570 Ctx.emitError("SIInstrInfo::loadRegFromStackSlot - Do not know how to"
571 " restore register");
572 BuildMI(MBB, MI, DL, get(AMDGPU::IMPLICIT_DEF), DestReg);
576 /// \param @Offset Offset in bytes of the FrameIndex being spilled
577 unsigned SIInstrInfo::calculateLDSSpillAddress(MachineBasicBlock &MBB,
578 MachineBasicBlock::iterator MI,
579 RegScavenger *RS, unsigned TmpReg,
580 unsigned FrameOffset,
581 unsigned Size) const {
582 MachineFunction *MF = MBB.getParent();
583 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
584 const AMDGPUSubtarget &ST = MF->getSubtarget<AMDGPUSubtarget>();
585 const SIRegisterInfo *TRI =
586 static_cast<const SIRegisterInfo*>(ST.getRegisterInfo());
587 DebugLoc DL = MBB.findDebugLoc(MI);
588 unsigned WorkGroupSize = MFI->getMaximumWorkGroupSize(*MF);
589 unsigned WavefrontSize = ST.getWavefrontSize();
591 unsigned TIDReg = MFI->getTIDReg();
592 if (!MFI->hasCalculatedTID()) {
593 MachineBasicBlock &Entry = MBB.getParent()->front();
594 MachineBasicBlock::iterator Insert = Entry.front();
595 DebugLoc DL = Insert->getDebugLoc();
597 TIDReg = RI.findUnusedRegister(MF->getRegInfo(), &AMDGPU::VGPR_32RegClass);
598 if (TIDReg == AMDGPU::NoRegister)
602 if (MFI->getShaderType() == ShaderType::COMPUTE &&
603 WorkGroupSize > WavefrontSize) {
605 unsigned TIDIGXReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_X);
606 unsigned TIDIGYReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Y);
607 unsigned TIDIGZReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Z);
608 unsigned InputPtrReg =
609 TRI->getPreloadedValue(*MF, SIRegisterInfo::INPUT_PTR);
610 for (unsigned Reg : {TIDIGXReg, TIDIGYReg, TIDIGZReg}) {
611 if (!Entry.isLiveIn(Reg))
612 Entry.addLiveIn(Reg);
615 RS->enterBasicBlock(&Entry);
616 unsigned STmp0 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
617 unsigned STmp1 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
618 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp0)
620 .addImm(SI::KernelInputOffsets::NGROUPS_Z);
621 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp1)
623 .addImm(SI::KernelInputOffsets::NGROUPS_Y);
625 // NGROUPS.X * NGROUPS.Y
626 BuildMI(Entry, Insert, DL, get(AMDGPU::S_MUL_I32), STmp1)
629 // (NGROUPS.X * NGROUPS.Y) * TIDIG.X
630 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MUL_U32_U24_e32), TIDReg)
633 // NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)
634 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MAD_U32_U24), TIDReg)
638 // (NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)) + TIDIG.Z
639 BuildMI(Entry, Insert, DL, get(AMDGPU::V_ADD_I32_e32), TIDReg)
644 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_LO_U32_B32_e64),
649 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_HI_U32_B32_e64),
655 BuildMI(Entry, Insert, DL, get(AMDGPU::V_LSHLREV_B32_e32),
659 MFI->setTIDReg(TIDReg);
662 // Add FrameIndex to LDS offset
663 unsigned LDSOffset = MFI->LDSSize + (FrameOffset * WorkGroupSize);
664 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_I32_e32), TmpReg)
671 void SIInstrInfo::insertNOPs(MachineBasicBlock::iterator MI,
680 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(AMDGPU::S_NOP))
685 bool SIInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
686 MachineBasicBlock &MBB = *MI->getParent();
687 DebugLoc DL = MBB.findDebugLoc(MI);
688 switch (MI->getOpcode()) {
689 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
691 case AMDGPU::SI_CONSTDATA_PTR: {
692 unsigned Reg = MI->getOperand(0).getReg();
693 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
694 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
696 BuildMI(MBB, MI, DL, get(AMDGPU::S_GETPC_B64), Reg);
698 // Add 32-bit offset from this instruction to the start of the constant data.
699 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADD_U32), RegLo)
701 .addTargetIndex(AMDGPU::TI_CONSTDATA_START)
702 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit);
703 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADDC_U32), RegHi)
706 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit)
707 .addReg(AMDGPU::SCC, RegState::Implicit);
708 MI->eraseFromParent();
711 case AMDGPU::SGPR_USE:
712 // This is just a placeholder for register allocation.
713 MI->eraseFromParent();
716 case AMDGPU::V_MOV_B64_PSEUDO: {
717 unsigned Dst = MI->getOperand(0).getReg();
718 unsigned DstLo = RI.getSubReg(Dst, AMDGPU::sub0);
719 unsigned DstHi = RI.getSubReg(Dst, AMDGPU::sub1);
721 const MachineOperand &SrcOp = MI->getOperand(1);
722 // FIXME: Will this work for 64-bit floating point immediates?
723 assert(!SrcOp.isFPImm());
725 APInt Imm(64, SrcOp.getImm());
726 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
727 .addImm(Imm.getLoBits(32).getZExtValue())
728 .addReg(Dst, RegState::Implicit);
729 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
730 .addImm(Imm.getHiBits(32).getZExtValue())
731 .addReg(Dst, RegState::Implicit);
733 assert(SrcOp.isReg());
734 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
735 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub0))
736 .addReg(Dst, RegState::Implicit);
737 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
738 .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub1))
739 .addReg(Dst, RegState::Implicit);
741 MI->eraseFromParent();
745 case AMDGPU::V_CNDMASK_B64_PSEUDO: {
746 unsigned Dst = MI->getOperand(0).getReg();
747 unsigned DstLo = RI.getSubReg(Dst, AMDGPU::sub0);
748 unsigned DstHi = RI.getSubReg(Dst, AMDGPU::sub1);
749 unsigned Src0 = MI->getOperand(1).getReg();
750 unsigned Src1 = MI->getOperand(2).getReg();
751 const MachineOperand &SrcCond = MI->getOperand(3);
753 BuildMI(MBB, MI, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstLo)
754 .addReg(RI.getSubReg(Src0, AMDGPU::sub0))
755 .addReg(RI.getSubReg(Src1, AMDGPU::sub0))
756 .addOperand(SrcCond);
757 BuildMI(MBB, MI, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstHi)
758 .addReg(RI.getSubReg(Src0, AMDGPU::sub1))
759 .addReg(RI.getSubReg(Src1, AMDGPU::sub1))
760 .addOperand(SrcCond);
761 MI->eraseFromParent();
768 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
771 if (MI->getNumOperands() < 3)
774 int CommutedOpcode = commuteOpcode(*MI);
775 if (CommutedOpcode == -1)
778 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
779 AMDGPU::OpName::src0);
780 assert(Src0Idx != -1 && "Should always have src0 operand");
782 MachineOperand &Src0 = MI->getOperand(Src0Idx);
786 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
787 AMDGPU::OpName::src1);
791 MachineOperand &Src1 = MI->getOperand(Src1Idx);
793 // Make sure it's legal to commute operands for VOP2.
794 if (isVOP2(MI->getOpcode()) &&
795 (!isOperandLegal(MI, Src0Idx, &Src1) ||
796 !isOperandLegal(MI, Src1Idx, &Src0))) {
801 // Allow commuting instructions with Imm operands.
802 if (NewMI || !Src1.isImm() ||
803 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
807 // Be sure to copy the source modifiers to the right place.
808 if (MachineOperand *Src0Mods
809 = getNamedOperand(*MI, AMDGPU::OpName::src0_modifiers)) {
810 MachineOperand *Src1Mods
811 = getNamedOperand(*MI, AMDGPU::OpName::src1_modifiers);
813 int Src0ModsVal = Src0Mods->getImm();
814 if (!Src1Mods && Src0ModsVal != 0)
817 // XXX - This assert might be a lie. It might be useful to have a neg
818 // modifier with 0.0.
819 int Src1ModsVal = Src1Mods->getImm();
820 assert((Src1ModsVal == 0) && "Not expecting modifiers with immediates");
822 Src1Mods->setImm(Src0ModsVal);
823 Src0Mods->setImm(Src1ModsVal);
826 unsigned Reg = Src0.getReg();
827 unsigned SubReg = Src0.getSubReg();
829 Src0.ChangeToImmediate(Src1.getImm());
831 llvm_unreachable("Should only have immediates");
833 Src1.ChangeToRegister(Reg, false);
834 Src1.setSubReg(SubReg);
836 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
840 MI->setDesc(get(CommutedOpcode));
845 // This needs to be implemented because the source modifiers may be inserted
846 // between the true commutable operands, and the base
847 // TargetInstrInfo::commuteInstruction uses it.
848 bool SIInstrInfo::findCommutedOpIndices(MachineInstr *MI,
850 unsigned &SrcOpIdx2) const {
851 const MCInstrDesc &MCID = MI->getDesc();
852 if (!MCID.isCommutable())
855 unsigned Opc = MI->getOpcode();
856 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
860 // FIXME: Workaround TargetInstrInfo::commuteInstruction asserting on
862 if (!MI->getOperand(Src0Idx).isReg())
865 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
869 if (!MI->getOperand(Src1Idx).isReg())
872 // If any source modifiers are set, the generic instruction commuting won't
873 // understand how to copy the source modifiers.
874 if (hasModifiersSet(*MI, AMDGPU::OpName::src0_modifiers) ||
875 hasModifiersSet(*MI, AMDGPU::OpName::src1_modifiers))
883 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
884 MachineBasicBlock::iterator I,
886 unsigned SrcReg) const {
887 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
888 DstReg) .addReg(SrcReg);
891 bool SIInstrInfo::isMov(unsigned Opcode) const {
893 default: return false;
894 case AMDGPU::S_MOV_B32:
895 case AMDGPU::S_MOV_B64:
896 case AMDGPU::V_MOV_B32_e32:
897 case AMDGPU::V_MOV_B32_e64:
903 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
904 return RC != &AMDGPU::EXECRegRegClass;
907 static void removeModOperands(MachineInstr &MI) {
908 unsigned Opc = MI.getOpcode();
909 int Src0ModIdx = AMDGPU::getNamedOperandIdx(Opc,
910 AMDGPU::OpName::src0_modifiers);
911 int Src1ModIdx = AMDGPU::getNamedOperandIdx(Opc,
912 AMDGPU::OpName::src1_modifiers);
913 int Src2ModIdx = AMDGPU::getNamedOperandIdx(Opc,
914 AMDGPU::OpName::src2_modifiers);
916 MI.RemoveOperand(Src2ModIdx);
917 MI.RemoveOperand(Src1ModIdx);
918 MI.RemoveOperand(Src0ModIdx);
921 bool SIInstrInfo::FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI,
922 unsigned Reg, MachineRegisterInfo *MRI) const {
923 if (!MRI->hasOneNonDBGUse(Reg))
926 unsigned Opc = UseMI->getOpcode();
927 if (Opc == AMDGPU::V_MAD_F32) {
928 // Don't fold if we are using source modifiers. The new VOP2 instructions
930 if (hasModifiersSet(*UseMI, AMDGPU::OpName::src0_modifiers) ||
931 hasModifiersSet(*UseMI, AMDGPU::OpName::src1_modifiers) ||
932 hasModifiersSet(*UseMI, AMDGPU::OpName::src2_modifiers)) {
936 MachineOperand *Src0 = getNamedOperand(*UseMI, AMDGPU::OpName::src0);
937 MachineOperand *Src1 = getNamedOperand(*UseMI, AMDGPU::OpName::src1);
938 MachineOperand *Src2 = getNamedOperand(*UseMI, AMDGPU::OpName::src2);
940 // Multiplied part is the constant: Use v_madmk_f32
941 // We should only expect these to be on src0 due to canonicalizations.
942 if (Src0->isReg() && Src0->getReg() == Reg) {
943 if (!Src1->isReg() ||
944 (Src1->isReg() && RI.isSGPRClass(MRI->getRegClass(Src1->getReg()))))
947 if (!Src2->isReg() ||
948 (Src2->isReg() && RI.isSGPRClass(MRI->getRegClass(Src2->getReg()))))
951 // We need to do some weird looking operand shuffling since the madmk
952 // operands are out of the normal expected order with the multiplied
953 // constant as the last operand.
955 // v_mad_f32 src0, src1, src2 -> v_madmk_f32 src0 * src2K + src1
960 const int64_t Imm = DefMI->getOperand(1).getImm();
962 // FIXME: This would be a lot easier if we could return a new instruction
963 // instead of having to modify in place.
965 // Remove these first since they are at the end.
966 UseMI->RemoveOperand(AMDGPU::getNamedOperandIdx(AMDGPU::V_MAD_F32,
967 AMDGPU::OpName::omod));
968 UseMI->RemoveOperand(AMDGPU::getNamedOperandIdx(AMDGPU::V_MAD_F32,
969 AMDGPU::OpName::clamp));
971 unsigned Src1Reg = Src1->getReg();
972 unsigned Src1SubReg = Src1->getSubReg();
973 unsigned Src2Reg = Src2->getReg();
974 unsigned Src2SubReg = Src2->getSubReg();
975 Src0->setReg(Src1Reg);
976 Src0->setSubReg(Src1SubReg);
977 Src0->setIsKill(Src1->isKill());
979 Src1->setReg(Src2Reg);
980 Src1->setSubReg(Src2SubReg);
981 Src1->setIsKill(Src2->isKill());
983 Src2->ChangeToImmediate(Imm);
985 removeModOperands(*UseMI);
986 UseMI->setDesc(get(AMDGPU::V_MADMK_F32));
988 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
990 DefMI->eraseFromParent();
995 // Added part is the constant: Use v_madak_f32
996 if (Src2->isReg() && Src2->getReg() == Reg) {
997 // Not allowed to use constant bus for another operand.
998 // We can however allow an inline immediate as src0.
999 if (!Src0->isImm() &&
1000 (Src0->isReg() && RI.isSGPRClass(MRI->getRegClass(Src0->getReg()))))
1003 if (!Src1->isReg() ||
1004 (Src1->isReg() && RI.isSGPRClass(MRI->getRegClass(Src1->getReg()))))
1007 const int64_t Imm = DefMI->getOperand(1).getImm();
1009 // FIXME: This would be a lot easier if we could return a new instruction
1010 // instead of having to modify in place.
1012 // Remove these first since they are at the end.
1013 UseMI->RemoveOperand(AMDGPU::getNamedOperandIdx(AMDGPU::V_MAD_F32,
1014 AMDGPU::OpName::omod));
1015 UseMI->RemoveOperand(AMDGPU::getNamedOperandIdx(AMDGPU::V_MAD_F32,
1016 AMDGPU::OpName::clamp));
1018 Src2->ChangeToImmediate(Imm);
1020 // These come before src2.
1021 removeModOperands(*UseMI);
1022 UseMI->setDesc(get(AMDGPU::V_MADAK_F32));
1024 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
1026 DefMI->eraseFromParent();
1036 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
1037 AliasAnalysis *AA) const {
1038 switch(MI->getOpcode()) {
1039 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
1040 case AMDGPU::S_MOV_B32:
1041 case AMDGPU::S_MOV_B64:
1042 case AMDGPU::V_MOV_B32_e32:
1043 return MI->getOperand(1).isImm();
1047 static bool offsetsDoNotOverlap(int WidthA, int OffsetA,
1048 int WidthB, int OffsetB) {
1049 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB;
1050 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA;
1051 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
1052 return LowOffset + LowWidth <= HighOffset;
1055 bool SIInstrInfo::checkInstOffsetsDoNotOverlap(MachineInstr *MIa,
1056 MachineInstr *MIb) const {
1057 unsigned BaseReg0, Offset0;
1058 unsigned BaseReg1, Offset1;
1060 if (getMemOpBaseRegImmOfs(MIa, BaseReg0, Offset0, &RI) &&
1061 getMemOpBaseRegImmOfs(MIb, BaseReg1, Offset1, &RI)) {
1062 assert(MIa->hasOneMemOperand() && MIb->hasOneMemOperand() &&
1063 "read2 / write2 not expected here yet");
1064 unsigned Width0 = (*MIa->memoperands_begin())->getSize();
1065 unsigned Width1 = (*MIb->memoperands_begin())->getSize();
1066 if (BaseReg0 == BaseReg1 &&
1067 offsetsDoNotOverlap(Width0, Offset0, Width1, Offset1)) {
1075 bool SIInstrInfo::areMemAccessesTriviallyDisjoint(MachineInstr *MIa,
1077 AliasAnalysis *AA) const {
1078 unsigned Opc0 = MIa->getOpcode();
1079 unsigned Opc1 = MIb->getOpcode();
1081 assert(MIa && (MIa->mayLoad() || MIa->mayStore()) &&
1082 "MIa must load from or modify a memory location");
1083 assert(MIb && (MIb->mayLoad() || MIb->mayStore()) &&
1084 "MIb must load from or modify a memory location");
1086 if (MIa->hasUnmodeledSideEffects() || MIb->hasUnmodeledSideEffects())
1089 // XXX - Can we relax this between address spaces?
1090 if (MIa->hasOrderedMemoryRef() || MIb->hasOrderedMemoryRef())
1093 // TODO: Should we check the address space from the MachineMemOperand? That
1094 // would allow us to distinguish objects we know don't alias based on the
1095 // underlying addres space, even if it was lowered to a different one,
1096 // e.g. private accesses lowered to use MUBUF instructions on a scratch
1100 return checkInstOffsetsDoNotOverlap(MIa, MIb);
1102 return !isFLAT(Opc1);
1105 if (isMUBUF(Opc0) || isMTBUF(Opc0)) {
1106 if (isMUBUF(Opc1) || isMTBUF(Opc1))
1107 return checkInstOffsetsDoNotOverlap(MIa, MIb);
1109 return !isFLAT(Opc1) && !isSMRD(Opc1);
1114 return checkInstOffsetsDoNotOverlap(MIa, MIb);
1116 return !isFLAT(Opc1) && !isMUBUF(Opc0) && !isMTBUF(Opc0);
1121 return checkInstOffsetsDoNotOverlap(MIa, MIb);
1129 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
1130 int64_t SVal = Imm.getSExtValue();
1131 if (SVal >= -16 && SVal <= 64)
1134 if (Imm.getBitWidth() == 64) {
1135 uint64_t Val = Imm.getZExtValue();
1136 return (DoubleToBits(0.0) == Val) ||
1137 (DoubleToBits(1.0) == Val) ||
1138 (DoubleToBits(-1.0) == Val) ||
1139 (DoubleToBits(0.5) == Val) ||
1140 (DoubleToBits(-0.5) == Val) ||
1141 (DoubleToBits(2.0) == Val) ||
1142 (DoubleToBits(-2.0) == Val) ||
1143 (DoubleToBits(4.0) == Val) ||
1144 (DoubleToBits(-4.0) == Val);
1147 // The actual type of the operand does not seem to matter as long
1148 // as the bits match one of the inline immediate values. For example:
1150 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
1151 // so it is a legal inline immediate.
1153 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
1154 // floating-point, so it is a legal inline immediate.
1155 uint32_t Val = Imm.getZExtValue();
1157 return (FloatToBits(0.0f) == Val) ||
1158 (FloatToBits(1.0f) == Val) ||
1159 (FloatToBits(-1.0f) == Val) ||
1160 (FloatToBits(0.5f) == Val) ||
1161 (FloatToBits(-0.5f) == Val) ||
1162 (FloatToBits(2.0f) == Val) ||
1163 (FloatToBits(-2.0f) == Val) ||
1164 (FloatToBits(4.0f) == Val) ||
1165 (FloatToBits(-4.0f) == Val);
1168 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO,
1169 unsigned OpSize) const {
1171 // MachineOperand provides no way to tell the true operand size, since it
1172 // only records a 64-bit value. We need to know the size to determine if a
1173 // 32-bit floating point immediate bit pattern is legal for an integer
1174 // immediate. It would be for any 32-bit integer operand, but would not be
1175 // for a 64-bit one.
1177 unsigned BitSize = 8 * OpSize;
1178 return isInlineConstant(APInt(BitSize, MO.getImm(), true));
1184 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO,
1185 unsigned OpSize) const {
1186 return MO.isImm() && !isInlineConstant(MO, OpSize);
1189 static bool compareMachineOp(const MachineOperand &Op0,
1190 const MachineOperand &Op1) {
1191 if (Op0.getType() != Op1.getType())
1194 switch (Op0.getType()) {
1195 case MachineOperand::MO_Register:
1196 return Op0.getReg() == Op1.getReg();
1197 case MachineOperand::MO_Immediate:
1198 return Op0.getImm() == Op1.getImm();
1200 llvm_unreachable("Didn't expect to be comparing these operand types");
1204 bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
1205 const MachineOperand &MO) const {
1206 const MCOperandInfo &OpInfo = get(MI->getOpcode()).OpInfo[OpNo];
1208 assert(MO.isImm() || MO.isTargetIndex() || MO.isFI());
1210 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
1213 if (OpInfo.RegClass < 0)
1216 unsigned OpSize = RI.getRegClass(OpInfo.RegClass)->getSize();
1217 if (isLiteralConstant(MO, OpSize))
1218 return RI.opCanUseLiteralConstant(OpInfo.OperandType);
1220 return RI.opCanUseInlineConstant(OpInfo.OperandType);
1223 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
1224 int Op32 = AMDGPU::getVOPe32(Opcode);
1228 return pseudoToMCOpcode(Op32) != -1;
1231 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
1232 // The src0_modifier operand is present on all instructions
1233 // that have modifiers.
1235 return AMDGPU::getNamedOperandIdx(Opcode,
1236 AMDGPU::OpName::src0_modifiers) != -1;
1239 bool SIInstrInfo::hasModifiersSet(const MachineInstr &MI,
1240 unsigned OpName) const {
1241 const MachineOperand *Mods = getNamedOperand(MI, OpName);
1242 return Mods && Mods->getImm();
1245 bool SIInstrInfo::usesConstantBus(const MachineRegisterInfo &MRI,
1246 const MachineOperand &MO,
1247 unsigned OpSize) const {
1248 // Literal constants use the constant bus.
1249 if (isLiteralConstant(MO, OpSize))
1252 if (!MO.isReg() || !MO.isUse())
1255 if (TargetRegisterInfo::isVirtualRegister(MO.getReg()))
1256 return RI.isSGPRClass(MRI.getRegClass(MO.getReg()));
1258 // FLAT_SCR is just an SGPR pair.
1259 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
1262 // EXEC register uses the constant bus.
1263 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
1266 // SGPRs use the constant bus
1267 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
1268 (!MO.isImplicit() &&
1269 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
1270 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
1277 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
1278 StringRef &ErrInfo) const {
1279 uint16_t Opcode = MI->getOpcode();
1280 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1281 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
1282 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
1283 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
1285 // Make sure the number of operands is correct.
1286 const MCInstrDesc &Desc = get(Opcode);
1287 if (!Desc.isVariadic() &&
1288 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
1289 ErrInfo = "Instruction has wrong number of operands.";
1293 // Make sure the register classes are correct
1294 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
1295 if (MI->getOperand(i).isFPImm()) {
1296 ErrInfo = "FPImm Machine Operands are not supported. ISel should bitcast "
1297 "all fp values to integers.";
1301 int RegClass = Desc.OpInfo[i].RegClass;
1303 switch (Desc.OpInfo[i].OperandType) {
1304 case MCOI::OPERAND_REGISTER:
1305 if (MI->getOperand(i).isImm()) {
1306 ErrInfo = "Illegal immediate value for operand.";
1310 case AMDGPU::OPERAND_REG_IMM32:
1312 case AMDGPU::OPERAND_REG_INLINE_C:
1313 if (isLiteralConstant(MI->getOperand(i),
1314 RI.getRegClass(RegClass)->getSize())) {
1315 ErrInfo = "Illegal immediate value for operand.";
1319 case MCOI::OPERAND_IMMEDIATE:
1320 // Check if this operand is an immediate.
1321 // FrameIndex operands will be replaced by immediates, so they are
1323 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFI()) {
1324 ErrInfo = "Expected immediate, but got non-immediate";
1332 if (!MI->getOperand(i).isReg())
1335 if (RegClass != -1) {
1336 unsigned Reg = MI->getOperand(i).getReg();
1337 if (TargetRegisterInfo::isVirtualRegister(Reg))
1340 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
1341 if (!RC->contains(Reg)) {
1342 ErrInfo = "Operand has incorrect register class.";
1350 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
1351 // Only look at the true operands. Only a real operand can use the constant
1352 // bus, and we don't want to check pseudo-operands like the source modifier
1354 const int OpIndices[] = { Src0Idx, Src1Idx, Src2Idx };
1356 unsigned ConstantBusCount = 0;
1357 unsigned SGPRUsed = AMDGPU::NoRegister;
1358 for (int OpIdx : OpIndices) {
1361 const MachineOperand &MO = MI->getOperand(OpIdx);
1362 if (usesConstantBus(MRI, MO, getOpSize(Opcode, OpIdx))) {
1364 if (MO.getReg() != SGPRUsed)
1366 SGPRUsed = MO.getReg();
1372 if (ConstantBusCount > 1) {
1373 ErrInfo = "VOP* instruction uses the constant bus more than once";
1378 // Verify misc. restrictions on specific instructions.
1379 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
1380 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
1381 const MachineOperand &Src0 = MI->getOperand(Src0Idx);
1382 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1383 const MachineOperand &Src2 = MI->getOperand(Src2Idx);
1384 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
1385 if (!compareMachineOp(Src0, Src1) &&
1386 !compareMachineOp(Src0, Src2)) {
1387 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
1396 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
1397 switch (MI.getOpcode()) {
1398 default: return AMDGPU::INSTRUCTION_LIST_END;
1399 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
1400 case AMDGPU::COPY: return AMDGPU::COPY;
1401 case AMDGPU::PHI: return AMDGPU::PHI;
1402 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
1403 case AMDGPU::S_MOV_B32:
1404 return MI.getOperand(1).isReg() ?
1405 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
1406 case AMDGPU::S_ADD_I32:
1407 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
1408 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
1409 case AMDGPU::S_SUB_I32:
1410 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
1411 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
1412 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
1413 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
1414 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
1415 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
1416 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
1417 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
1418 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
1419 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
1420 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
1421 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
1422 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
1423 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
1424 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
1425 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
1426 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
1427 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
1428 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
1429 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
1430 case AMDGPU::S_BFM_B32: return AMDGPU::V_BFM_B32_e64;
1431 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
1432 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
1433 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
1434 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
1435 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
1436 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
1437 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
1438 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
1439 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
1440 case AMDGPU::S_LOAD_DWORD_IMM:
1441 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1442 case AMDGPU::S_LOAD_DWORDX2_IMM:
1443 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1444 case AMDGPU::S_LOAD_DWORDX4_IMM:
1445 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1446 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e64;
1447 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
1448 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
1449 case AMDGPU::S_FLBIT_I32: return AMDGPU::V_FFBH_I32_e64;
1453 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
1454 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
1457 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
1458 unsigned OpNo) const {
1459 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1460 const MCInstrDesc &Desc = get(MI.getOpcode());
1461 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
1462 Desc.OpInfo[OpNo].RegClass == -1) {
1463 unsigned Reg = MI.getOperand(OpNo).getReg();
1465 if (TargetRegisterInfo::isVirtualRegister(Reg))
1466 return MRI.getRegClass(Reg);
1467 return RI.getPhysRegClass(Reg);
1470 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
1471 return RI.getRegClass(RCID);
1474 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
1475 switch (MI.getOpcode()) {
1477 case AMDGPU::REG_SEQUENCE:
1479 case AMDGPU::INSERT_SUBREG:
1480 return RI.hasVGPRs(getOpRegClass(MI, 0));
1482 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
1486 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
1487 MachineBasicBlock::iterator I = MI;
1488 MachineBasicBlock *MBB = MI->getParent();
1489 MachineOperand &MO = MI->getOperand(OpIdx);
1490 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1491 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
1492 const TargetRegisterClass *RC = RI.getRegClass(RCID);
1493 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1495 Opcode = AMDGPU::COPY;
1496 else if (RI.isSGPRClass(RC))
1497 Opcode = AMDGPU::S_MOV_B32;
1500 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
1501 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC))
1502 VRC = &AMDGPU::VReg_64RegClass;
1504 VRC = &AMDGPU::VGPR_32RegClass;
1506 unsigned Reg = MRI.createVirtualRegister(VRC);
1507 DebugLoc DL = MBB->findDebugLoc(I);
1508 BuildMI(*MI->getParent(), I, DL, get(Opcode), Reg)
1510 MO.ChangeToRegister(Reg, false);
1513 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
1514 MachineRegisterInfo &MRI,
1515 MachineOperand &SuperReg,
1516 const TargetRegisterClass *SuperRC,
1518 const TargetRegisterClass *SubRC)
1520 assert(SuperReg.isReg());
1522 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
1523 unsigned SubReg = MRI.createVirtualRegister(SubRC);
1525 // Just in case the super register is itself a sub-register, copy it to a new
1526 // value so we don't need to worry about merging its subreg index with the
1527 // SubIdx passed to this function. The register coalescer should be able to
1528 // eliminate this extra copy.
1529 MachineBasicBlock *MBB = MI->getParent();
1530 DebugLoc DL = MI->getDebugLoc();
1532 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), NewSuperReg)
1533 .addReg(SuperReg.getReg(), 0, SuperReg.getSubReg());
1535 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
1536 .addReg(NewSuperReg, 0, SubIdx);
1541 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
1542 MachineBasicBlock::iterator MII,
1543 MachineRegisterInfo &MRI,
1545 const TargetRegisterClass *SuperRC,
1547 const TargetRegisterClass *SubRC) const {
1549 // XXX - Is there a better way to do this?
1550 if (SubIdx == AMDGPU::sub0)
1551 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
1552 if (SubIdx == AMDGPU::sub1)
1553 return MachineOperand::CreateImm(Op.getImm() >> 32);
1555 llvm_unreachable("Unhandled register index for immediate");
1558 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
1560 return MachineOperand::CreateReg(SubReg, false);
1563 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
1564 MachineBasicBlock::iterator MI,
1565 MachineRegisterInfo &MRI,
1566 const TargetRegisterClass *RC,
1567 const MachineOperand &Op) const {
1568 MachineBasicBlock *MBB = MI->getParent();
1569 DebugLoc DL = MI->getDebugLoc();
1570 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1571 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1572 unsigned Dst = MRI.createVirtualRegister(RC);
1574 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1576 .addImm(Op.getImm() & 0xFFFFFFFF);
1577 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1579 .addImm(Op.getImm() >> 32);
1581 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
1583 .addImm(AMDGPU::sub0)
1585 .addImm(AMDGPU::sub1);
1587 Worklist.push_back(Lo);
1588 Worklist.push_back(Hi);
1593 // Change the order of operands from (0, 1, 2) to (0, 2, 1)
1594 void SIInstrInfo::swapOperands(MachineBasicBlock::iterator Inst) const {
1595 assert(Inst->getNumExplicitOperands() == 3);
1596 MachineOperand Op1 = Inst->getOperand(1);
1597 Inst->RemoveOperand(1);
1598 Inst->addOperand(Op1);
1601 bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
1602 const MachineOperand *MO) const {
1603 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1604 const MCInstrDesc &InstDesc = get(MI->getOpcode());
1605 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
1606 const TargetRegisterClass *DefinedRC =
1607 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
1609 MO = &MI->getOperand(OpIdx);
1611 if (isVALU(InstDesc.Opcode) &&
1612 usesConstantBus(MRI, *MO, DefinedRC->getSize())) {
1614 MO->isReg() ? MO->getReg() : (unsigned)AMDGPU::NoRegister;
1615 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1618 const MachineOperand &Op = MI->getOperand(i);
1619 if (Op.isReg() && Op.getReg() != SGPRUsed &&
1620 usesConstantBus(MRI, Op, getOpSize(*MI, i))) {
1628 const TargetRegisterClass *RC =
1629 TargetRegisterInfo::isVirtualRegister(MO->getReg()) ?
1630 MRI.getRegClass(MO->getReg()) :
1631 RI.getPhysRegClass(MO->getReg());
1633 // In order to be legal, the common sub-class must be equal to the
1634 // class of the current operand. For example:
1636 // v_mov_b32 s0 ; Operand defined as vsrc_32
1637 // ; RI.getCommonSubClass(s0,vsrc_32) = sgpr ; LEGAL
1639 // s_sendmsg 0, s0 ; Operand defined as m0reg
1640 // ; RI.getCommonSubClass(s0,m0reg) = m0reg ; NOT LEGAL
1642 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass)) == RC;
1646 // Handle non-register types that are treated like immediates.
1647 assert(MO->isImm() || MO->isTargetIndex() || MO->isFI());
1650 // This operand expects an immediate.
1654 return isImmOperandLegal(MI, OpIdx, *MO);
1657 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
1658 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1660 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1661 AMDGPU::OpName::src0);
1662 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1663 AMDGPU::OpName::src1);
1664 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1665 AMDGPU::OpName::src2);
1668 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
1670 if (!isOperandLegal(MI, Src0Idx))
1671 legalizeOpWithMove(MI, Src0Idx);
1674 if (isOperandLegal(MI, Src1Idx))
1677 // Usually src0 of VOP2 instructions allow more types of inputs
1678 // than src1, so try to commute the instruction to decrease our
1679 // chances of having to insert a MOV instruction to legalize src1.
1680 if (MI->isCommutable()) {
1681 if (commuteInstruction(MI))
1682 // If we are successful in commuting, then we know MI is legal, so
1687 legalizeOpWithMove(MI, Src1Idx);
1691 // XXX - Do any VOP3 instructions read VCC?
1693 if (isVOP3(MI->getOpcode())) {
1694 int VOP3Idx[3] = { Src0Idx, Src1Idx, Src2Idx };
1696 // Find the one SGPR operand we are allowed to use.
1697 unsigned SGPRReg = findUsedSGPR(MI, VOP3Idx);
1699 for (unsigned i = 0; i < 3; ++i) {
1700 int Idx = VOP3Idx[i];
1703 MachineOperand &MO = MI->getOperand(Idx);
1706 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
1707 continue; // VGPRs are legal
1709 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
1711 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
1712 SGPRReg = MO.getReg();
1713 // We can use one SGPR in each VOP3 instruction.
1716 } else if (!isLiteralConstant(MO, getOpSize(MI->getOpcode(), Idx))) {
1717 // If it is not a register and not a literal constant, then it must be
1718 // an inline constant which is always legal.
1721 // If we make it this far, then the operand is not legal and we must
1723 legalizeOpWithMove(MI, Idx);
1727 // Legalize REG_SEQUENCE and PHI
1728 // The register class of the operands much be the same type as the register
1729 // class of the output.
1730 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
1731 MI->getOpcode() == AMDGPU::PHI) {
1732 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
1733 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1734 if (!MI->getOperand(i).isReg() ||
1735 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1737 const TargetRegisterClass *OpRC =
1738 MRI.getRegClass(MI->getOperand(i).getReg());
1739 if (RI.hasVGPRs(OpRC)) {
1746 // If any of the operands are VGPR registers, then they all most be
1747 // otherwise we will create illegal VGPR->SGPR copies when legalizing
1749 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
1752 VRC = RI.getEquivalentVGPRClass(SRC);
1759 // Update all the operands so they have the same type.
1760 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1761 if (!MI->getOperand(i).isReg() ||
1762 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1764 unsigned DstReg = MRI.createVirtualRegister(RC);
1765 MachineBasicBlock *InsertBB;
1766 MachineBasicBlock::iterator Insert;
1767 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
1768 InsertBB = MI->getParent();
1771 // MI is a PHI instruction.
1772 InsertBB = MI->getOperand(i + 1).getMBB();
1773 Insert = InsertBB->getFirstTerminator();
1775 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
1776 get(AMDGPU::COPY), DstReg)
1777 .addOperand(MI->getOperand(i));
1778 MI->getOperand(i).setReg(DstReg);
1782 // Legalize INSERT_SUBREG
1783 // src0 must have the same register class as dst
1784 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
1785 unsigned Dst = MI->getOperand(0).getReg();
1786 unsigned Src0 = MI->getOperand(1).getReg();
1787 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
1788 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
1789 if (DstRC != Src0RC) {
1790 MachineBasicBlock &MBB = *MI->getParent();
1791 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
1792 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
1794 MI->getOperand(1).setReg(NewSrc0);
1799 // Legalize MUBUF* instructions
1800 // FIXME: If we start using the non-addr64 instructions for compute, we
1801 // may need to legalize them here.
1803 AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::srsrc);
1804 if (SRsrcIdx != -1) {
1805 // We have an MUBUF instruction
1806 MachineOperand *SRsrc = &MI->getOperand(SRsrcIdx);
1807 unsigned SRsrcRC = get(MI->getOpcode()).OpInfo[SRsrcIdx].RegClass;
1808 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
1809 RI.getRegClass(SRsrcRC))) {
1810 // The operands are legal.
1811 // FIXME: We may need to legalize operands besided srsrc.
1815 MachineBasicBlock &MBB = *MI->getParent();
1816 // Extract the ptr from the resource descriptor.
1818 // SRsrcPtrLo = srsrc:sub0
1819 unsigned SRsrcPtrLo = buildExtractSubReg(MI, MRI, *SRsrc,
1820 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VGPR_32RegClass);
1822 // SRsrcPtrHi = srsrc:sub1
1823 unsigned SRsrcPtrHi = buildExtractSubReg(MI, MRI, *SRsrc,
1824 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VGPR_32RegClass);
1826 // Create an empty resource descriptor
1827 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1828 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1829 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1830 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1831 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
1834 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1838 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1839 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1841 .addImm(RsrcDataFormat & 0xFFFFFFFF);
1843 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1844 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1846 .addImm(RsrcDataFormat >> 32);
1848 // NewSRsrc = {Zero64, SRsrcFormat}
1849 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1852 .addImm(AMDGPU::sub0_sub1)
1853 .addReg(SRsrcFormatLo)
1854 .addImm(AMDGPU::sub2)
1855 .addReg(SRsrcFormatHi)
1856 .addImm(AMDGPU::sub3);
1858 MachineOperand *VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1859 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
1860 unsigned NewVAddrLo;
1861 unsigned NewVAddrHi;
1863 // This is already an ADDR64 instruction so we need to add the pointer
1864 // extracted from the resource descriptor to the current value of VAddr.
1865 NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1866 NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1868 // NewVaddrLo = SRsrcPtrLo + VAddr:sub0
1869 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1872 .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
1873 .addReg(AMDGPU::VCC, RegState::ImplicitDefine);
1875 // NewVaddrHi = SRsrcPtrHi + VAddr:sub1
1876 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1879 .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
1880 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1881 .addReg(AMDGPU::VCC, RegState::Implicit);
1884 // This instructions is the _OFFSET variant, so we need to convert it to
1886 MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
1887 MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
1888 MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
1890 // Create the new instruction.
1891 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
1892 MachineInstr *Addr64 =
1893 BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
1895 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
1896 // This will be replaced later
1897 // with the new value of vaddr.
1899 .addOperand(*SOffset)
1900 .addOperand(*Offset)
1905 MI->removeFromParent();
1908 NewVAddrLo = SRsrcPtrLo;
1909 NewVAddrHi = SRsrcPtrHi;
1910 VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1911 SRsrc = getNamedOperand(*MI, AMDGPU::OpName::srsrc);
1914 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1915 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1918 .addImm(AMDGPU::sub0)
1920 .addImm(AMDGPU::sub1);
1923 // Update the instruction to use NewVaddr
1924 VAddr->setReg(NewVAddr);
1925 // Update the instruction to use NewSRsrc
1926 SRsrc->setReg(NewSRsrc);
1930 void SIInstrInfo::splitSMRD(MachineInstr *MI,
1931 const TargetRegisterClass *HalfRC,
1932 unsigned HalfImmOp, unsigned HalfSGPROp,
1933 MachineInstr *&Lo, MachineInstr *&Hi) const {
1935 DebugLoc DL = MI->getDebugLoc();
1936 MachineBasicBlock *MBB = MI->getParent();
1937 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1938 unsigned RegLo = MRI.createVirtualRegister(HalfRC);
1939 unsigned RegHi = MRI.createVirtualRegister(HalfRC);
1940 unsigned HalfSize = HalfRC->getSize();
1941 const MachineOperand *OffOp =
1942 getNamedOperand(*MI, AMDGPU::OpName::offset);
1943 const MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
1945 // The SMRD has an 8-bit offset in dwords on SI and a 20-bit offset in bytes
1948 bool IsKill = SBase->isKill();
1951 MBB->getParent()->getSubtarget<AMDGPUSubtarget>().getGeneration() >=
1952 AMDGPUSubtarget::VOLCANIC_ISLANDS;
1953 unsigned OffScale = isVI ? 1 : 4;
1954 // Handle the _IMM variant
1955 unsigned LoOffset = OffOp->getImm() * OffScale;
1956 unsigned HiOffset = LoOffset + HalfSize;
1957 Lo = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegLo)
1958 // Use addReg instead of addOperand
1959 // to make sure kill flag is cleared.
1960 .addReg(SBase->getReg(), 0, SBase->getSubReg())
1961 .addImm(LoOffset / OffScale);
1963 if (!isUInt<20>(HiOffset) || (!isVI && !isUInt<8>(HiOffset / OffScale))) {
1964 unsigned OffsetSGPR =
1965 MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1966 BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32), OffsetSGPR)
1967 .addImm(HiOffset); // The offset in register is in bytes.
1968 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
1969 .addReg(SBase->getReg(), getKillRegState(IsKill),
1971 .addReg(OffsetSGPR);
1973 Hi = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegHi)
1974 .addReg(SBase->getReg(), getKillRegState(IsKill),
1976 .addImm(HiOffset / OffScale);
1979 // Handle the _SGPR variant
1980 MachineOperand *SOff = getNamedOperand(*MI, AMDGPU::OpName::soff);
1981 Lo = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegLo)
1982 .addReg(SBase->getReg(), 0, SBase->getSubReg())
1984 unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1985 BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
1988 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp))
1989 .addReg(SBase->getReg(), getKillRegState(IsKill),
1991 .addReg(OffsetSGPR);
1994 unsigned SubLo, SubHi;
1997 SubLo = AMDGPU::sub0;
1998 SubHi = AMDGPU::sub1;
2001 SubLo = AMDGPU::sub0_sub1;
2002 SubHi = AMDGPU::sub2_sub3;
2005 SubLo = AMDGPU::sub0_sub1_sub2_sub3;
2006 SubHi = AMDGPU::sub4_sub5_sub6_sub7;
2009 SubLo = AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
2010 SubHi = AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15;
2013 llvm_unreachable("Unhandled HalfSize");
2016 BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE))
2017 .addOperand(MI->getOperand(0))
2024 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
2025 MachineBasicBlock *MBB = MI->getParent();
2026 switch (MI->getOpcode()) {
2027 case AMDGPU::S_LOAD_DWORD_IMM:
2028 case AMDGPU::S_LOAD_DWORD_SGPR:
2029 case AMDGPU::S_LOAD_DWORDX2_IMM:
2030 case AMDGPU::S_LOAD_DWORDX2_SGPR:
2031 case AMDGPU::S_LOAD_DWORDX4_IMM:
2032 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
2033 unsigned NewOpcode = getVALUOp(*MI);
2037 if (MI->getOperand(2).isReg()) {
2038 RegOffset = MI->getOperand(2).getReg();
2041 assert(MI->getOperand(2).isImm());
2042 // SMRD instructions take a dword offsets on SI and byte offset on VI
2043 // and MUBUF instructions always take a byte offset.
2044 ImmOffset = MI->getOperand(2).getImm();
2045 if (MBB->getParent()->getSubtarget<AMDGPUSubtarget>().getGeneration() <=
2046 AMDGPUSubtarget::SEA_ISLANDS)
2048 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
2050 if (isUInt<12>(ImmOffset)) {
2051 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
2055 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
2062 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
2063 unsigned DWord0 = RegOffset;
2064 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
2065 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
2066 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
2067 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
2069 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
2071 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
2072 .addImm(RsrcDataFormat & 0xFFFFFFFF);
2073 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
2074 .addImm(RsrcDataFormat >> 32);
2075 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
2077 .addImm(AMDGPU::sub0)
2079 .addImm(AMDGPU::sub1)
2081 .addImm(AMDGPU::sub2)
2083 .addImm(AMDGPU::sub3);
2084 MI->setDesc(get(NewOpcode));
2085 if (MI->getOperand(2).isReg()) {
2086 MI->getOperand(2).setReg(SRsrc);
2088 MI->getOperand(2).ChangeToRegister(SRsrc, false);
2090 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(0));
2091 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
2092 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(0)); // glc
2093 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(0)); // slc
2094 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(0)); // tfe
2096 const TargetRegisterClass *NewDstRC =
2097 RI.getRegClass(get(NewOpcode).OpInfo[0].RegClass);
2099 unsigned DstReg = MI->getOperand(0).getReg();
2100 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
2101 MRI.replaceRegWith(DstReg, NewDstReg);
2104 case AMDGPU::S_LOAD_DWORDX8_IMM:
2105 case AMDGPU::S_LOAD_DWORDX8_SGPR: {
2106 MachineInstr *Lo, *Hi;
2107 splitSMRD(MI, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM,
2108 AMDGPU::S_LOAD_DWORDX4_SGPR, Lo, Hi);
2109 MI->eraseFromParent();
2110 moveSMRDToVALU(Lo, MRI);
2111 moveSMRDToVALU(Hi, MRI);
2115 case AMDGPU::S_LOAD_DWORDX16_IMM:
2116 case AMDGPU::S_LOAD_DWORDX16_SGPR: {
2117 MachineInstr *Lo, *Hi;
2118 splitSMRD(MI, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM,
2119 AMDGPU::S_LOAD_DWORDX8_SGPR, Lo, Hi);
2120 MI->eraseFromParent();
2121 moveSMRDToVALU(Lo, MRI);
2122 moveSMRDToVALU(Hi, MRI);
2128 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
2129 SmallVector<MachineInstr *, 128> Worklist;
2130 Worklist.push_back(&TopInst);
2132 while (!Worklist.empty()) {
2133 MachineInstr *Inst = Worklist.pop_back_val();
2134 MachineBasicBlock *MBB = Inst->getParent();
2135 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
2137 unsigned Opcode = Inst->getOpcode();
2138 unsigned NewOpcode = getVALUOp(*Inst);
2140 // Handle some special cases
2143 if (isSMRD(Inst->getOpcode())) {
2144 moveSMRDToVALU(Inst, MRI);
2147 case AMDGPU::S_MOV_B64: {
2148 DebugLoc DL = Inst->getDebugLoc();
2150 // If the source operand is a register we can replace this with a
2152 if (Inst->getOperand(1).isReg()) {
2153 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
2154 .addOperand(Inst->getOperand(0))
2155 .addOperand(Inst->getOperand(1));
2156 Worklist.push_back(Copy);
2158 // Otherwise, we need to split this into two movs, because there is
2159 // no 64-bit VALU move instruction.
2160 unsigned Reg = Inst->getOperand(0).getReg();
2161 unsigned Dst = split64BitImm(Worklist,
2164 MRI.getRegClass(Reg),
2165 Inst->getOperand(1));
2166 MRI.replaceRegWith(Reg, Dst);
2168 Inst->eraseFromParent();
2171 case AMDGPU::S_AND_B64:
2172 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
2173 Inst->eraseFromParent();
2176 case AMDGPU::S_OR_B64:
2177 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
2178 Inst->eraseFromParent();
2181 case AMDGPU::S_XOR_B64:
2182 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
2183 Inst->eraseFromParent();
2186 case AMDGPU::S_NOT_B64:
2187 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
2188 Inst->eraseFromParent();
2191 case AMDGPU::S_BCNT1_I32_B64:
2192 splitScalar64BitBCNT(Worklist, Inst);
2193 Inst->eraseFromParent();
2196 case AMDGPU::S_BFE_I64: {
2197 splitScalar64BitBFE(Worklist, Inst);
2198 Inst->eraseFromParent();
2202 case AMDGPU::S_LSHL_B32:
2203 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2204 NewOpcode = AMDGPU::V_LSHLREV_B32_e64;
2208 case AMDGPU::S_ASHR_I32:
2209 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2210 NewOpcode = AMDGPU::V_ASHRREV_I32_e64;
2214 case AMDGPU::S_LSHR_B32:
2215 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2216 NewOpcode = AMDGPU::V_LSHRREV_B32_e64;
2220 case AMDGPU::S_LSHL_B64:
2221 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2222 NewOpcode = AMDGPU::V_LSHLREV_B64;
2226 case AMDGPU::S_ASHR_I64:
2227 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2228 NewOpcode = AMDGPU::V_ASHRREV_I64;
2232 case AMDGPU::S_LSHR_B64:
2233 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
2234 NewOpcode = AMDGPU::V_LSHRREV_B64;
2239 case AMDGPU::S_BFE_U64:
2240 case AMDGPU::S_BFM_B64:
2241 llvm_unreachable("Moving this op to VALU not implemented");
2244 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
2245 // We cannot move this instruction to the VALU, so we should try to
2246 // legalize its operands instead.
2247 legalizeOperands(Inst);
2251 // Use the new VALU Opcode.
2252 const MCInstrDesc &NewDesc = get(NewOpcode);
2253 Inst->setDesc(NewDesc);
2255 // Remove any references to SCC. Vector instructions can't read from it, and
2256 // We're just about to add the implicit use / defs of VCC, and we don't want
2258 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
2259 MachineOperand &Op = Inst->getOperand(i);
2260 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
2261 Inst->RemoveOperand(i);
2264 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
2265 // We are converting these to a BFE, so we need to add the missing
2266 // operands for the size and offset.
2267 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
2268 Inst->addOperand(MachineOperand::CreateImm(0));
2269 Inst->addOperand(MachineOperand::CreateImm(Size));
2271 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
2272 // The VALU version adds the second operand to the result, so insert an
2274 Inst->addOperand(MachineOperand::CreateImm(0));
2277 addDescImplicitUseDef(NewDesc, Inst);
2279 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
2280 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
2281 // If we need to move this to VGPRs, we need to unpack the second operand
2282 // back into the 2 separate ones for bit offset and width.
2283 assert(OffsetWidthOp.isImm() &&
2284 "Scalar BFE is only implemented for constant width and offset");
2285 uint32_t Imm = OffsetWidthOp.getImm();
2287 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
2288 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
2289 Inst->RemoveOperand(2); // Remove old immediate.
2290 Inst->addOperand(MachineOperand::CreateImm(Offset));
2291 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
2294 // Update the destination register class.
2296 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
2299 // For target instructions, getOpRegClass just returns the virtual
2300 // register class associated with the operand, so we need to find an
2301 // equivalent VGPR register class in order to move the instruction to the
2305 case AMDGPU::REG_SEQUENCE:
2306 case AMDGPU::INSERT_SUBREG:
2307 if (RI.hasVGPRs(NewDstRC))
2309 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
2317 unsigned DstReg = Inst->getOperand(0).getReg();
2318 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
2319 MRI.replaceRegWith(DstReg, NewDstReg);
2321 // Legalize the operands
2322 legalizeOperands(Inst);
2324 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
2325 E = MRI.use_end(); I != E; ++I) {
2326 MachineInstr &UseMI = *I->getParent();
2327 if (!canReadVGPR(UseMI, I.getOperandNo())) {
2328 Worklist.push_back(&UseMI);
2334 //===----------------------------------------------------------------------===//
2335 // Indirect addressing callbacks
2336 //===----------------------------------------------------------------------===//
2338 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
2339 unsigned Channel) const {
2340 assert(Channel == 0);
2344 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
2345 return &AMDGPU::VGPR_32RegClass;
2348 void SIInstrInfo::splitScalar64BitUnaryOp(
2349 SmallVectorImpl<MachineInstr *> &Worklist,
2351 unsigned Opcode) const {
2352 MachineBasicBlock &MBB = *Inst->getParent();
2353 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2355 MachineOperand &Dest = Inst->getOperand(0);
2356 MachineOperand &Src0 = Inst->getOperand(1);
2357 DebugLoc DL = Inst->getDebugLoc();
2359 MachineBasicBlock::iterator MII = Inst;
2361 const MCInstrDesc &InstDesc = get(Opcode);
2362 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2363 MRI.getRegClass(Src0.getReg()) :
2364 &AMDGPU::SGPR_32RegClass;
2366 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2368 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2369 AMDGPU::sub0, Src0SubRC);
2371 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2372 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2374 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2375 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2376 .addOperand(SrcReg0Sub0);
2378 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2379 AMDGPU::sub1, Src0SubRC);
2381 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2382 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2383 .addOperand(SrcReg0Sub1);
2385 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2386 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2388 .addImm(AMDGPU::sub0)
2390 .addImm(AMDGPU::sub1);
2392 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2394 // Try to legalize the operands in case we need to swap the order to keep it
2396 Worklist.push_back(LoHalf);
2397 Worklist.push_back(HiHalf);
2400 void SIInstrInfo::splitScalar64BitBinaryOp(
2401 SmallVectorImpl<MachineInstr *> &Worklist,
2403 unsigned Opcode) const {
2404 MachineBasicBlock &MBB = *Inst->getParent();
2405 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2407 MachineOperand &Dest = Inst->getOperand(0);
2408 MachineOperand &Src0 = Inst->getOperand(1);
2409 MachineOperand &Src1 = Inst->getOperand(2);
2410 DebugLoc DL = Inst->getDebugLoc();
2412 MachineBasicBlock::iterator MII = Inst;
2414 const MCInstrDesc &InstDesc = get(Opcode);
2415 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2416 MRI.getRegClass(Src0.getReg()) :
2417 &AMDGPU::SGPR_32RegClass;
2419 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2420 const TargetRegisterClass *Src1RC = Src1.isReg() ?
2421 MRI.getRegClass(Src1.getReg()) :
2422 &AMDGPU::SGPR_32RegClass;
2424 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
2426 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2427 AMDGPU::sub0, Src0SubRC);
2428 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2429 AMDGPU::sub0, Src1SubRC);
2431 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2432 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2434 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2435 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2436 .addOperand(SrcReg0Sub0)
2437 .addOperand(SrcReg1Sub0);
2439 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2440 AMDGPU::sub1, Src0SubRC);
2441 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2442 AMDGPU::sub1, Src1SubRC);
2444 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2445 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2446 .addOperand(SrcReg0Sub1)
2447 .addOperand(SrcReg1Sub1);
2449 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2450 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2452 .addImm(AMDGPU::sub0)
2454 .addImm(AMDGPU::sub1);
2456 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2458 // Try to legalize the operands in case we need to swap the order to keep it
2460 Worklist.push_back(LoHalf);
2461 Worklist.push_back(HiHalf);
2464 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
2465 MachineInstr *Inst) const {
2466 MachineBasicBlock &MBB = *Inst->getParent();
2467 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2469 MachineBasicBlock::iterator MII = Inst;
2470 DebugLoc DL = Inst->getDebugLoc();
2472 MachineOperand &Dest = Inst->getOperand(0);
2473 MachineOperand &Src = Inst->getOperand(1);
2475 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e64);
2476 const TargetRegisterClass *SrcRC = Src.isReg() ?
2477 MRI.getRegClass(Src.getReg()) :
2478 &AMDGPU::SGPR_32RegClass;
2480 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2481 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2483 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
2485 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2486 AMDGPU::sub0, SrcSubRC);
2487 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2488 AMDGPU::sub1, SrcSubRC);
2490 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
2491 .addOperand(SrcRegSub0)
2494 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
2495 .addOperand(SrcRegSub1)
2498 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2500 Worklist.push_back(First);
2501 Worklist.push_back(Second);
2504 void SIInstrInfo::splitScalar64BitBFE(SmallVectorImpl<MachineInstr *> &Worklist,
2505 MachineInstr *Inst) const {
2506 MachineBasicBlock &MBB = *Inst->getParent();
2507 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2508 MachineBasicBlock::iterator MII = Inst;
2509 DebugLoc DL = Inst->getDebugLoc();
2511 MachineOperand &Dest = Inst->getOperand(0);
2512 uint32_t Imm = Inst->getOperand(2).getImm();
2513 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
2514 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
2518 // Only sext_inreg cases handled.
2519 assert(Inst->getOpcode() == AMDGPU::S_BFE_I64 &&
2524 if (BitWidth < 32) {
2525 unsigned MidRegLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2526 unsigned MidRegHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2527 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
2529 BuildMI(MBB, MII, DL, get(AMDGPU::V_BFE_I32), MidRegLo)
2530 .addReg(Inst->getOperand(1).getReg(), 0, AMDGPU::sub0)
2534 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e32), MidRegHi)
2538 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
2540 .addImm(AMDGPU::sub0)
2542 .addImm(AMDGPU::sub1);
2544 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2548 MachineOperand &Src = Inst->getOperand(1);
2549 unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2550 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
2552 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e64), TmpReg)
2554 .addReg(Src.getReg(), 0, AMDGPU::sub0);
2556 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
2557 .addReg(Src.getReg(), 0, AMDGPU::sub0)
2558 .addImm(AMDGPU::sub0)
2560 .addImm(AMDGPU::sub1);
2562 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2565 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
2566 MachineInstr *Inst) const {
2567 // Add the implict and explicit register definitions.
2568 if (NewDesc.ImplicitUses) {
2569 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
2570 unsigned Reg = NewDesc.ImplicitUses[i];
2571 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
2575 if (NewDesc.ImplicitDefs) {
2576 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
2577 unsigned Reg = NewDesc.ImplicitDefs[i];
2578 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
2583 unsigned SIInstrInfo::findUsedSGPR(const MachineInstr *MI,
2584 int OpIndices[3]) const {
2585 const MCInstrDesc &Desc = get(MI->getOpcode());
2587 // Find the one SGPR operand we are allowed to use.
2588 unsigned SGPRReg = AMDGPU::NoRegister;
2590 // First we need to consider the instruction's operand requirements before
2591 // legalizing. Some operands are required to be SGPRs, such as implicit uses
2592 // of VCC, but we are still bound by the constant bus requirement to only use
2595 // If the operand's class is an SGPR, we can never move it.
2597 for (const MachineOperand &MO : MI->implicit_operands()) {
2598 // We only care about reads.
2602 if (MO.getReg() == AMDGPU::VCC)
2605 if (MO.getReg() == AMDGPU::FLAT_SCR)
2606 return AMDGPU::FLAT_SCR;
2609 unsigned UsedSGPRs[3] = { AMDGPU::NoRegister };
2610 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
2612 for (unsigned i = 0; i < 3; ++i) {
2613 int Idx = OpIndices[i];
2617 const MachineOperand &MO = MI->getOperand(Idx);
2618 if (RI.isSGPRClassID(Desc.OpInfo[Idx].RegClass))
2619 SGPRReg = MO.getReg();
2621 if (MO.isReg() && RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
2622 UsedSGPRs[i] = MO.getReg();
2625 if (SGPRReg != AMDGPU::NoRegister)
2628 // We don't have a required SGPR operand, so we have a bit more freedom in
2629 // selecting operands to move.
2631 // Try to select the most used SGPR. If an SGPR is equal to one of the
2632 // others, we choose that.
2635 // V_FMA_F32 v0, s0, s0, s0 -> No moves
2636 // V_FMA_F32 v0, s0, s1, s0 -> Move s1
2638 if (UsedSGPRs[0] != AMDGPU::NoRegister) {
2639 if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
2640 SGPRReg = UsedSGPRs[0];
2643 if (SGPRReg == AMDGPU::NoRegister && UsedSGPRs[1] != AMDGPU::NoRegister) {
2644 if (UsedSGPRs[1] == UsedSGPRs[2])
2645 SGPRReg = UsedSGPRs[1];
2651 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
2652 MachineBasicBlock *MBB,
2653 MachineBasicBlock::iterator I,
2655 unsigned Address, unsigned OffsetReg) const {
2656 const DebugLoc &DL = MBB->findDebugLoc(I);
2657 unsigned IndirectBaseReg = AMDGPU::VGPR_32RegClass.getRegister(
2658 getIndirectIndexBegin(*MBB->getParent()));
2660 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
2661 .addReg(IndirectBaseReg, RegState::Define)
2662 .addOperand(I->getOperand(0))
2663 .addReg(IndirectBaseReg)
2669 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
2670 MachineBasicBlock *MBB,
2671 MachineBasicBlock::iterator I,
2673 unsigned Address, unsigned OffsetReg) const {
2674 const DebugLoc &DL = MBB->findDebugLoc(I);
2675 unsigned IndirectBaseReg = AMDGPU::VGPR_32RegClass.getRegister(
2676 getIndirectIndexBegin(*MBB->getParent()));
2678 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
2679 .addOperand(I->getOperand(0))
2680 .addOperand(I->getOperand(1))
2681 .addReg(IndirectBaseReg)
2687 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
2688 const MachineFunction &MF) const {
2689 int End = getIndirectIndexEnd(MF);
2690 int Begin = getIndirectIndexBegin(MF);
2696 for (int Index = Begin; Index <= End; ++Index)
2697 Reserved.set(AMDGPU::VGPR_32RegClass.getRegister(Index));
2699 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
2700 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
2702 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
2703 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
2705 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
2706 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
2708 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
2709 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
2711 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
2712 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));
2715 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
2716 unsigned OperandName) const {
2717 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
2721 return &MI.getOperand(Idx);
2724 uint64_t SIInstrInfo::getDefaultRsrcDataFormat() const {
2725 uint64_t RsrcDataFormat = AMDGPU::RSRC_DATA_FORMAT;
2726 if (ST.isAmdHsaOS()) {
2727 RsrcDataFormat |= (1ULL << 56);
2729 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
2731 RsrcDataFormat |= (2ULL << 59);
2734 return RsrcDataFormat;