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),
34 //===----------------------------------------------------------------------===//
35 // TargetInstrInfo callbacks
36 //===----------------------------------------------------------------------===//
38 static unsigned getNumOperandsNoGlue(SDNode *Node) {
39 unsigned N = Node->getNumOperands();
40 while (N && Node->getOperand(N - 1).getValueType() == MVT::Glue)
45 static SDValue findChainOperand(SDNode *Load) {
46 SDValue LastOp = Load->getOperand(getNumOperandsNoGlue(Load) - 1);
47 assert(LastOp.getValueType() == MVT::Other && "Chain missing from load node");
51 /// \brief Returns true if both nodes have the same value for the given
52 /// operand \p Op, or if both nodes do not have this operand.
53 static bool nodesHaveSameOperandValue(SDNode *N0, SDNode* N1, unsigned OpName) {
54 unsigned Opc0 = N0->getMachineOpcode();
55 unsigned Opc1 = N1->getMachineOpcode();
57 int Op0Idx = AMDGPU::getNamedOperandIdx(Opc0, OpName);
58 int Op1Idx = AMDGPU::getNamedOperandIdx(Opc1, OpName);
60 if (Op0Idx == -1 && Op1Idx == -1)
64 if ((Op0Idx == -1 && Op1Idx != -1) ||
65 (Op1Idx == -1 && Op0Idx != -1))
68 // getNamedOperandIdx returns the index for the MachineInstr's operands,
69 // which includes the result as the first operand. We are indexing into the
70 // MachineSDNode's operands, so we need to skip the result operand to get
75 return N0->getOperand(Op0Idx) == N1->getOperand(Op1Idx);
78 bool SIInstrInfo::areLoadsFromSameBasePtr(SDNode *Load0, SDNode *Load1,
80 int64_t &Offset1) const {
81 if (!Load0->isMachineOpcode() || !Load1->isMachineOpcode())
84 unsigned Opc0 = Load0->getMachineOpcode();
85 unsigned Opc1 = Load1->getMachineOpcode();
87 // Make sure both are actually loads.
88 if (!get(Opc0).mayLoad() || !get(Opc1).mayLoad())
91 if (isDS(Opc0) && isDS(Opc1)) {
93 // FIXME: Handle this case:
94 if (getNumOperandsNoGlue(Load0) != getNumOperandsNoGlue(Load1))
98 if (Load0->getOperand(1) != Load1->getOperand(1))
102 if (findChainOperand(Load0) != findChainOperand(Load1))
105 // Skip read2 / write2 variants for simplicity.
106 // TODO: We should report true if the used offsets are adjacent (excluded
108 if (AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::data1) != -1 ||
109 AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::data1) != -1)
112 Offset0 = cast<ConstantSDNode>(Load0->getOperand(2))->getZExtValue();
113 Offset1 = cast<ConstantSDNode>(Load1->getOperand(2))->getZExtValue();
117 if (isSMRD(Opc0) && isSMRD(Opc1)) {
118 assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
121 if (Load0->getOperand(0) != Load1->getOperand(0))
125 if (findChainOperand(Load0) != findChainOperand(Load1))
128 Offset0 = cast<ConstantSDNode>(Load0->getOperand(1))->getZExtValue();
129 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue();
133 // MUBUF and MTBUF can access the same addresses.
134 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1))) {
136 // MUBUF and MTBUF have vaddr at different indices.
137 if (!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::soffset) ||
138 findChainOperand(Load0) != findChainOperand(Load1) ||
139 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::vaddr) ||
140 !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::srsrc))
143 int OffIdx0 = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
144 int OffIdx1 = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
146 if (OffIdx0 == -1 || OffIdx1 == -1)
149 // getNamedOperandIdx returns the index for MachineInstrs. Since they
150 // inlcude the output in the operand list, but SDNodes don't, we need to
151 // subtract the index by one.
155 SDValue Off0 = Load0->getOperand(OffIdx0);
156 SDValue Off1 = Load1->getOperand(OffIdx1);
158 // The offset might be a FrameIndexSDNode.
159 if (!isa<ConstantSDNode>(Off0) || !isa<ConstantSDNode>(Off1))
162 Offset0 = cast<ConstantSDNode>(Off0)->getZExtValue();
163 Offset1 = cast<ConstantSDNode>(Off1)->getZExtValue();
170 static bool isStride64(unsigned Opc) {
172 case AMDGPU::DS_READ2ST64_B32:
173 case AMDGPU::DS_READ2ST64_B64:
174 case AMDGPU::DS_WRITE2ST64_B32:
175 case AMDGPU::DS_WRITE2ST64_B64:
182 bool SIInstrInfo::getLdStBaseRegImmOfs(MachineInstr *LdSt,
183 unsigned &BaseReg, unsigned &Offset,
184 const TargetRegisterInfo *TRI) const {
185 unsigned Opc = LdSt->getOpcode();
187 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
188 AMDGPU::OpName::offset);
190 // Normal, single offset LDS instruction.
191 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
192 AMDGPU::OpName::addr);
194 BaseReg = AddrReg->getReg();
195 Offset = OffsetImm->getImm();
199 // The 2 offset instructions use offset0 and offset1 instead. We can treat
200 // these as a load with a single offset if the 2 offsets are consecutive. We
201 // will use this for some partially aligned loads.
202 const MachineOperand *Offset0Imm = getNamedOperand(*LdSt,
203 AMDGPU::OpName::offset0);
204 const MachineOperand *Offset1Imm = getNamedOperand(*LdSt,
205 AMDGPU::OpName::offset1);
207 uint8_t Offset0 = Offset0Imm->getImm();
208 uint8_t Offset1 = Offset1Imm->getImm();
209 assert(Offset1 > Offset0);
211 if (Offset1 - Offset0 == 1) {
212 // Each of these offsets is in element sized units, so we need to convert
213 // to bytes of the individual reads.
217 EltSize = getOpRegClass(*LdSt, 0)->getSize() / 2;
219 assert(LdSt->mayStore());
220 int Data0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
221 EltSize = getOpRegClass(*LdSt, Data0Idx)->getSize();
227 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
228 AMDGPU::OpName::addr);
229 BaseReg = AddrReg->getReg();
230 Offset = EltSize * Offset0;
237 if (isMUBUF(Opc) || isMTBUF(Opc)) {
238 if (AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::soffset) != -1)
241 const MachineOperand *AddrReg = getNamedOperand(*LdSt,
242 AMDGPU::OpName::vaddr);
246 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
247 AMDGPU::OpName::offset);
248 BaseReg = AddrReg->getReg();
249 Offset = OffsetImm->getImm();
254 const MachineOperand *OffsetImm = getNamedOperand(*LdSt,
255 AMDGPU::OpName::offset);
259 const MachineOperand *SBaseReg = getNamedOperand(*LdSt,
260 AMDGPU::OpName::sbase);
261 BaseReg = SBaseReg->getReg();
262 Offset = OffsetImm->getImm();
269 bool SIInstrInfo::shouldClusterLoads(MachineInstr *FirstLdSt,
270 MachineInstr *SecondLdSt,
271 unsigned NumLoads) const {
272 unsigned Opc0 = FirstLdSt->getOpcode();
273 unsigned Opc1 = SecondLdSt->getOpcode();
275 // TODO: This needs finer tuning
279 if (isDS(Opc0) && isDS(Opc1))
282 if (isSMRD(Opc0) && isSMRD(Opc1))
285 if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1)))
292 SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
293 MachineBasicBlock::iterator MI, DebugLoc DL,
294 unsigned DestReg, unsigned SrcReg,
295 bool KillSrc) const {
297 // If we are trying to copy to or from SCC, there is a bug somewhere else in
298 // the backend. While it may be theoretically possible to do this, it should
299 // never be necessary.
300 assert(DestReg != AMDGPU::SCC && SrcReg != AMDGPU::SCC);
302 static const int16_t Sub0_15[] = {
303 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
304 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
305 AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
306 AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15, 0
309 static const int16_t Sub0_7[] = {
310 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
311 AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7, 0
314 static const int16_t Sub0_3[] = {
315 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
318 static const int16_t Sub0_2[] = {
319 AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
322 static const int16_t Sub0_1[] = {
323 AMDGPU::sub0, AMDGPU::sub1, 0
327 const int16_t *SubIndices;
329 if (AMDGPU::SReg_32RegClass.contains(DestReg)) {
330 assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
331 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
332 .addReg(SrcReg, getKillRegState(KillSrc));
335 } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
336 assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
337 BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
338 .addReg(SrcReg, getKillRegState(KillSrc));
341 } else if (AMDGPU::SReg_128RegClass.contains(DestReg)) {
342 assert(AMDGPU::SReg_128RegClass.contains(SrcReg));
343 Opcode = AMDGPU::S_MOV_B32;
346 } else if (AMDGPU::SReg_256RegClass.contains(DestReg)) {
347 assert(AMDGPU::SReg_256RegClass.contains(SrcReg));
348 Opcode = AMDGPU::S_MOV_B32;
351 } else if (AMDGPU::SReg_512RegClass.contains(DestReg)) {
352 assert(AMDGPU::SReg_512RegClass.contains(SrcReg));
353 Opcode = AMDGPU::S_MOV_B32;
354 SubIndices = Sub0_15;
356 } else if (AMDGPU::VReg_32RegClass.contains(DestReg)) {
357 assert(AMDGPU::VReg_32RegClass.contains(SrcReg) ||
358 AMDGPU::SReg_32RegClass.contains(SrcReg));
359 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
360 .addReg(SrcReg, getKillRegState(KillSrc));
363 } else if (AMDGPU::VReg_64RegClass.contains(DestReg)) {
364 assert(AMDGPU::VReg_64RegClass.contains(SrcReg) ||
365 AMDGPU::SReg_64RegClass.contains(SrcReg));
366 Opcode = AMDGPU::V_MOV_B32_e32;
369 } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
370 assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
371 Opcode = AMDGPU::V_MOV_B32_e32;
374 } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
375 assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
376 AMDGPU::SReg_128RegClass.contains(SrcReg));
377 Opcode = AMDGPU::V_MOV_B32_e32;
380 } else if (AMDGPU::VReg_256RegClass.contains(DestReg)) {
381 assert(AMDGPU::VReg_256RegClass.contains(SrcReg) ||
382 AMDGPU::SReg_256RegClass.contains(SrcReg));
383 Opcode = AMDGPU::V_MOV_B32_e32;
386 } else if (AMDGPU::VReg_512RegClass.contains(DestReg)) {
387 assert(AMDGPU::VReg_512RegClass.contains(SrcReg) ||
388 AMDGPU::SReg_512RegClass.contains(SrcReg));
389 Opcode = AMDGPU::V_MOV_B32_e32;
390 SubIndices = Sub0_15;
393 llvm_unreachable("Can't copy register!");
396 while (unsigned SubIdx = *SubIndices++) {
397 MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
398 get(Opcode), RI.getSubReg(DestReg, SubIdx));
400 Builder.addReg(RI.getSubReg(SrcReg, SubIdx), getKillRegState(KillSrc));
403 Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
407 unsigned SIInstrInfo::commuteOpcode(unsigned Opcode) const {
410 // Try to map original to commuted opcode
411 if ((NewOpc = AMDGPU::getCommuteRev(Opcode)) != -1)
414 // Try to map commuted to original opcode
415 if ((NewOpc = AMDGPU::getCommuteOrig(Opcode)) != -1)
421 static bool shouldTryToSpillVGPRs(MachineFunction *MF) {
423 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
424 const TargetMachine &TM = MF->getTarget();
426 // FIXME: Even though it can cause problems, we need to enable
427 // spilling at -O0, since the fast register allocator always
428 // spills registers that are live at the end of blocks.
429 return MFI->getShaderType() == ShaderType::COMPUTE &&
430 TM.getOptLevel() == CodeGenOpt::None;
434 void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
435 MachineBasicBlock::iterator MI,
436 unsigned SrcReg, bool isKill,
438 const TargetRegisterClass *RC,
439 const TargetRegisterInfo *TRI) const {
440 MachineFunction *MF = MBB.getParent();
441 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
442 DebugLoc DL = MBB.findDebugLoc(MI);
445 if (RI.isSGPRClass(RC)) {
446 // We are only allowed to create one new instruction when spilling
447 // registers, so we need to use pseudo instruction for spilling
449 switch (RC->getSize() * 8) {
450 case 32: Opcode = AMDGPU::SI_SPILL_S32_SAVE; break;
451 case 64: Opcode = AMDGPU::SI_SPILL_S64_SAVE; break;
452 case 128: Opcode = AMDGPU::SI_SPILL_S128_SAVE; break;
453 case 256: Opcode = AMDGPU::SI_SPILL_S256_SAVE; break;
454 case 512: Opcode = AMDGPU::SI_SPILL_S512_SAVE; break;
456 } else if(shouldTryToSpillVGPRs(MF) && RI.hasVGPRs(RC)) {
457 switch(RC->getSize() * 8) {
458 case 32: Opcode = AMDGPU::SI_SPILL_V32_SAVE; break;
459 case 64: Opcode = AMDGPU::SI_SPILL_V64_SAVE; break;
460 case 96: Opcode = AMDGPU::SI_SPILL_V96_SAVE; break;
461 case 128: Opcode = AMDGPU::SI_SPILL_V128_SAVE; break;
462 case 256: Opcode = AMDGPU::SI_SPILL_V256_SAVE; break;
463 case 512: Opcode = AMDGPU::SI_SPILL_V512_SAVE; break;
468 FrameInfo->setObjectAlignment(FrameIndex, 4);
469 BuildMI(MBB, MI, DL, get(Opcode))
471 .addFrameIndex(FrameIndex);
473 LLVMContext &Ctx = MF->getFunction()->getContext();
474 Ctx.emitError("SIInstrInfo::storeRegToStackSlot - Do not know how to"
476 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), AMDGPU::VGPR0)
481 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
482 MachineBasicBlock::iterator MI,
483 unsigned DestReg, int FrameIndex,
484 const TargetRegisterClass *RC,
485 const TargetRegisterInfo *TRI) const {
486 MachineFunction *MF = MBB.getParent();
487 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
488 DebugLoc DL = MBB.findDebugLoc(MI);
491 if (RI.isSGPRClass(RC)){
492 switch(RC->getSize() * 8) {
493 case 32: Opcode = AMDGPU::SI_SPILL_S32_RESTORE; break;
494 case 64: Opcode = AMDGPU::SI_SPILL_S64_RESTORE; break;
495 case 128: Opcode = AMDGPU::SI_SPILL_S128_RESTORE; break;
496 case 256: Opcode = AMDGPU::SI_SPILL_S256_RESTORE; break;
497 case 512: Opcode = AMDGPU::SI_SPILL_S512_RESTORE; break;
499 } else if(shouldTryToSpillVGPRs(MF) && RI.hasVGPRs(RC)) {
500 switch(RC->getSize() * 8) {
501 case 32: Opcode = AMDGPU::SI_SPILL_V32_RESTORE; break;
502 case 64: Opcode = AMDGPU::SI_SPILL_V64_RESTORE; break;
503 case 96: Opcode = AMDGPU::SI_SPILL_V96_RESTORE; break;
504 case 128: Opcode = AMDGPU::SI_SPILL_V128_RESTORE; break;
505 case 256: Opcode = AMDGPU::SI_SPILL_V256_RESTORE; break;
506 case 512: Opcode = AMDGPU::SI_SPILL_V512_RESTORE; break;
511 FrameInfo->setObjectAlignment(FrameIndex, 4);
512 BuildMI(MBB, MI, DL, get(Opcode), DestReg)
513 .addFrameIndex(FrameIndex);
515 LLVMContext &Ctx = MF->getFunction()->getContext();
516 Ctx.emitError("SIInstrInfo::loadRegFromStackSlot - Do not know how to"
517 " restore register");
518 BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
519 .addReg(AMDGPU::VGPR0);
523 /// \param @Offset Offset in bytes of the FrameIndex being spilled
524 unsigned SIInstrInfo::calculateLDSSpillAddress(MachineBasicBlock &MBB,
525 MachineBasicBlock::iterator MI,
526 RegScavenger *RS, unsigned TmpReg,
527 unsigned FrameOffset,
528 unsigned Size) const {
529 MachineFunction *MF = MBB.getParent();
530 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
531 const AMDGPUSubtarget &ST = MF->getTarget().getSubtarget<AMDGPUSubtarget>();
532 const SIRegisterInfo *TRI =
533 static_cast<const SIRegisterInfo*>(ST.getRegisterInfo());
534 DebugLoc DL = MBB.findDebugLoc(MI);
535 unsigned WorkGroupSize = MFI->getMaximumWorkGroupSize(*MF);
536 unsigned WavefrontSize = ST.getWavefrontSize();
538 unsigned TIDReg = MFI->getTIDReg();
539 if (!MFI->hasCalculatedTID()) {
540 MachineBasicBlock &Entry = MBB.getParent()->front();
541 MachineBasicBlock::iterator Insert = Entry.front();
542 DebugLoc DL = Insert->getDebugLoc();
544 TIDReg = RI.findUnusedVGPR(MF->getRegInfo());
545 if (TIDReg == AMDGPU::NoRegister)
549 if (MFI->getShaderType() == ShaderType::COMPUTE &&
550 WorkGroupSize > WavefrontSize) {
552 unsigned TIDIGXReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_X);
553 unsigned TIDIGYReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Y);
554 unsigned TIDIGZReg = TRI->getPreloadedValue(*MF, SIRegisterInfo::TIDIG_Z);
555 unsigned InputPtrReg =
556 TRI->getPreloadedValue(*MF, SIRegisterInfo::INPUT_PTR);
557 static const unsigned TIDIGRegs[3] = {
558 TIDIGXReg, TIDIGYReg, TIDIGZReg
560 for (unsigned Reg : TIDIGRegs) {
561 if (!Entry.isLiveIn(Reg))
562 Entry.addLiveIn(Reg);
565 RS->enterBasicBlock(&Entry);
566 unsigned STmp0 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
567 unsigned STmp1 = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
568 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp0)
570 .addImm(SI::KernelInputOffsets::NGROUPS_Z);
571 BuildMI(Entry, Insert, DL, get(AMDGPU::S_LOAD_DWORD_IMM), STmp1)
573 .addImm(SI::KernelInputOffsets::NGROUPS_Y);
575 // NGROUPS.X * NGROUPS.Y
576 BuildMI(Entry, Insert, DL, get(AMDGPU::S_MUL_I32), STmp1)
579 // (NGROUPS.X * NGROUPS.Y) * TIDIG.X
580 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MUL_U32_U24_e32), TIDReg)
583 // NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)
584 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MAD_U32_U24), TIDReg)
588 // (NGROUPS.Z * TIDIG.Y + (NGROUPS.X * NGROPUS.Y * TIDIG.X)) + TIDIG.Z
589 BuildMI(Entry, Insert, DL, get(AMDGPU::V_ADD_I32_e32), TIDReg)
594 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_LO_U32_B32_e64),
599 BuildMI(Entry, Insert, DL, get(AMDGPU::V_MBCNT_HI_U32_B32_e32),
605 BuildMI(Entry, Insert, DL, get(AMDGPU::V_LSHLREV_B32_e32),
609 MFI->setTIDReg(TIDReg);
612 // Add FrameIndex to LDS offset
613 unsigned LDSOffset = MFI->LDSSize + (FrameOffset * WorkGroupSize);
614 BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_I32_e32), TmpReg)
621 void SIInstrInfo::insertNOPs(MachineBasicBlock::iterator MI,
630 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(AMDGPU::S_NOP))
635 bool SIInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
636 MachineBasicBlock &MBB = *MI->getParent();
637 DebugLoc DL = MBB.findDebugLoc(MI);
638 switch (MI->getOpcode()) {
639 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
641 case AMDGPU::SI_CONSTDATA_PTR: {
642 unsigned Reg = MI->getOperand(0).getReg();
643 unsigned RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
644 unsigned RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
646 BuildMI(MBB, MI, DL, get(AMDGPU::S_GETPC_B64), Reg);
648 // Add 32-bit offset from this instruction to the start of the constant data.
649 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADD_U32), RegLo)
651 .addTargetIndex(AMDGPU::TI_CONSTDATA_START)
652 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit);
653 BuildMI(MBB, MI, DL, get(AMDGPU::S_ADDC_U32), RegHi)
656 .addReg(AMDGPU::SCC, RegState::Define | RegState::Implicit)
657 .addReg(AMDGPU::SCC, RegState::Implicit);
658 MI->eraseFromParent();
661 case AMDGPU::SGPR_USE:
662 // This is just a placeholder for register allocation.
663 MI->eraseFromParent();
669 MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
671 if (MI->getNumOperands() < 3)
674 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
675 AMDGPU::OpName::src0);
676 assert(Src0Idx != -1 && "Should always have src0 operand");
678 MachineOperand &Src0 = MI->getOperand(Src0Idx);
682 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
683 AMDGPU::OpName::src1);
687 MachineOperand &Src1 = MI->getOperand(Src1Idx);
689 // Make sure it's legal to commute operands for VOP2.
690 if (isVOP2(MI->getOpcode()) &&
691 (!isOperandLegal(MI, Src0Idx, &Src1) ||
692 !isOperandLegal(MI, Src1Idx, &Src0)))
696 // Allow commuting instructions with Imm or FPImm operands.
697 if (NewMI || (!Src1.isImm() && !Src1.isFPImm()) ||
698 (!isVOP2(MI->getOpcode()) && !isVOP3(MI->getOpcode()))) {
702 // Be sure to copy the source modifiers to the right place.
703 if (MachineOperand *Src0Mods
704 = getNamedOperand(*MI, AMDGPU::OpName::src0_modifiers)) {
705 MachineOperand *Src1Mods
706 = getNamedOperand(*MI, AMDGPU::OpName::src1_modifiers);
708 int Src0ModsVal = Src0Mods->getImm();
709 if (!Src1Mods && Src0ModsVal != 0)
712 // XXX - This assert might be a lie. It might be useful to have a neg
713 // modifier with 0.0.
714 int Src1ModsVal = Src1Mods->getImm();
715 assert((Src1ModsVal == 0) && "Not expecting modifiers with immediates");
717 Src1Mods->setImm(Src0ModsVal);
718 Src0Mods->setImm(Src1ModsVal);
721 unsigned Reg = Src0.getReg();
722 unsigned SubReg = Src0.getSubReg();
724 Src0.ChangeToImmediate(Src1.getImm());
725 else if (Src1.isFPImm())
726 Src0.ChangeToFPImmediate(Src1.getFPImm());
728 llvm_unreachable("Should only have immediates");
730 Src1.ChangeToRegister(Reg, false);
731 Src1.setSubReg(SubReg);
733 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
737 MI->setDesc(get(commuteOpcode(MI->getOpcode())));
742 // This needs to be implemented because the source modifiers may be inserted
743 // between the true commutable operands, and the base
744 // TargetInstrInfo::commuteInstruction uses it.
745 bool SIInstrInfo::findCommutedOpIndices(MachineInstr *MI,
747 unsigned &SrcOpIdx2) const {
748 const MCInstrDesc &MCID = MI->getDesc();
749 if (!MCID.isCommutable())
752 unsigned Opc = MI->getOpcode();
753 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
757 // FIXME: Workaround TargetInstrInfo::commuteInstruction asserting on
759 if (!MI->getOperand(Src0Idx).isReg())
762 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
766 if (!MI->getOperand(Src1Idx).isReg())
769 // If any source modifiers are set, the generic instruction commuting won't
770 // understand how to copy the source modifiers.
771 if (hasModifiersSet(*MI, AMDGPU::OpName::src0_modifiers) ||
772 hasModifiersSet(*MI, AMDGPU::OpName::src1_modifiers))
780 MachineInstr *SIInstrInfo::buildMovInstr(MachineBasicBlock *MBB,
781 MachineBasicBlock::iterator I,
783 unsigned SrcReg) const {
784 return BuildMI(*MBB, I, MBB->findDebugLoc(I), get(AMDGPU::V_MOV_B32_e32),
785 DstReg) .addReg(SrcReg);
788 bool SIInstrInfo::isMov(unsigned Opcode) const {
790 default: return false;
791 case AMDGPU::S_MOV_B32:
792 case AMDGPU::S_MOV_B64:
793 case AMDGPU::V_MOV_B32_e32:
794 case AMDGPU::V_MOV_B32_e64:
800 SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
801 return RC != &AMDGPU::EXECRegRegClass;
805 SIInstrInfo::isTriviallyReMaterializable(const MachineInstr *MI,
806 AliasAnalysis *AA) const {
807 switch(MI->getOpcode()) {
808 default: return AMDGPUInstrInfo::isTriviallyReMaterializable(MI, AA);
809 case AMDGPU::S_MOV_B32:
810 case AMDGPU::S_MOV_B64:
811 case AMDGPU::V_MOV_B32_e32:
812 return MI->getOperand(1).isImm();
816 static bool offsetsDoNotOverlap(int WidthA, int OffsetA,
817 int WidthB, int OffsetB) {
818 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB;
819 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA;
820 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
821 return LowOffset + LowWidth <= HighOffset;
824 bool SIInstrInfo::checkInstOffsetsDoNotOverlap(MachineInstr *MIa,
825 MachineInstr *MIb) const {
826 unsigned BaseReg0, Offset0;
827 unsigned BaseReg1, Offset1;
829 if (getLdStBaseRegImmOfs(MIa, BaseReg0, Offset0, &RI) &&
830 getLdStBaseRegImmOfs(MIb, BaseReg1, Offset1, &RI)) {
831 assert(MIa->hasOneMemOperand() && MIb->hasOneMemOperand() &&
832 "read2 / write2 not expected here yet");
833 unsigned Width0 = (*MIa->memoperands_begin())->getSize();
834 unsigned Width1 = (*MIb->memoperands_begin())->getSize();
835 if (BaseReg0 == BaseReg1 &&
836 offsetsDoNotOverlap(Width0, Offset0, Width1, Offset1)) {
844 bool SIInstrInfo::areMemAccessesTriviallyDisjoint(MachineInstr *MIa,
846 AliasAnalysis *AA) const {
847 unsigned Opc0 = MIa->getOpcode();
848 unsigned Opc1 = MIb->getOpcode();
850 assert(MIa && (MIa->mayLoad() || MIa->mayStore()) &&
851 "MIa must load from or modify a memory location");
852 assert(MIb && (MIb->mayLoad() || MIb->mayStore()) &&
853 "MIb must load from or modify a memory location");
855 if (MIa->hasUnmodeledSideEffects() || MIb->hasUnmodeledSideEffects())
858 // XXX - Can we relax this between address spaces?
859 if (MIa->hasOrderedMemoryRef() || MIb->hasOrderedMemoryRef())
862 // TODO: Should we check the address space from the MachineMemOperand? That
863 // would allow us to distinguish objects we know don't alias based on the
864 // underlying addres space, even if it was lowered to a different one,
865 // e.g. private accesses lowered to use MUBUF instructions on a scratch
869 return checkInstOffsetsDoNotOverlap(MIa, MIb);
871 return !isFLAT(Opc1);
874 if (isMUBUF(Opc0) || isMTBUF(Opc0)) {
875 if (isMUBUF(Opc1) || isMTBUF(Opc1))
876 return checkInstOffsetsDoNotOverlap(MIa, MIb);
878 return !isFLAT(Opc1) && !isSMRD(Opc1);
883 return checkInstOffsetsDoNotOverlap(MIa, MIb);
885 return !isFLAT(Opc1) && !isMUBUF(Opc0) && !isMTBUF(Opc0);
890 return checkInstOffsetsDoNotOverlap(MIa, MIb);
898 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
899 int32_t Val = Imm.getSExtValue();
900 if (Val >= -16 && Val <= 64)
903 // The actual type of the operand does not seem to matter as long
904 // as the bits match one of the inline immediate values. For example:
906 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
907 // so it is a legal inline immediate.
909 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
910 // floating-point, so it is a legal inline immediate.
912 return (APInt::floatToBits(0.0f) == Imm) ||
913 (APInt::floatToBits(1.0f) == Imm) ||
914 (APInt::floatToBits(-1.0f) == Imm) ||
915 (APInt::floatToBits(0.5f) == Imm) ||
916 (APInt::floatToBits(-0.5f) == Imm) ||
917 (APInt::floatToBits(2.0f) == Imm) ||
918 (APInt::floatToBits(-2.0f) == Imm) ||
919 (APInt::floatToBits(4.0f) == Imm) ||
920 (APInt::floatToBits(-4.0f) == Imm);
923 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO) const {
925 return isInlineConstant(APInt(32, MO.getImm(), true));
928 APFloat FpImm = MO.getFPImm()->getValueAPF();
929 return isInlineConstant(FpImm.bitcastToAPInt());
935 bool SIInstrInfo::isLiteralConstant(const MachineOperand &MO) const {
936 return (MO.isImm() || MO.isFPImm()) && !isInlineConstant(MO);
939 static bool compareMachineOp(const MachineOperand &Op0,
940 const MachineOperand &Op1) {
941 if (Op0.getType() != Op1.getType())
944 switch (Op0.getType()) {
945 case MachineOperand::MO_Register:
946 return Op0.getReg() == Op1.getReg();
947 case MachineOperand::MO_Immediate:
948 return Op0.getImm() == Op1.getImm();
949 case MachineOperand::MO_FPImmediate:
950 return Op0.getFPImm() == Op1.getFPImm();
952 llvm_unreachable("Didn't expect to be comparing these operand types");
956 bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
957 const MachineOperand &MO) const {
958 const MCOperandInfo &OpInfo = get(MI->getOpcode()).OpInfo[OpNo];
960 assert(MO.isImm() || MO.isFPImm() || MO.isTargetIndex() || MO.isFI());
962 if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
965 if (OpInfo.RegClass < 0)
968 if (isLiteralConstant(MO))
969 return RI.regClassCanUseLiteralConstant(OpInfo.RegClass);
971 return RI.regClassCanUseInlineConstant(OpInfo.RegClass);
974 bool SIInstrInfo::canFoldOffset(unsigned OffsetSize, unsigned AS) const {
976 case AMDGPUAS::GLOBAL_ADDRESS: {
977 // MUBUF instructions a 12-bit offset in bytes.
978 return isUInt<12>(OffsetSize);
980 case AMDGPUAS::CONSTANT_ADDRESS: {
981 // SMRD instructions have an 8-bit offset in dwords on SI and
982 // a 20-bit offset in bytes on VI.
983 if (RI.ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
984 return isUInt<20>(OffsetSize);
986 return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4);
988 case AMDGPUAS::LOCAL_ADDRESS:
989 case AMDGPUAS::REGION_ADDRESS: {
990 // The single offset versions have a 16-bit offset in bytes.
991 return isUInt<16>(OffsetSize);
993 case AMDGPUAS::PRIVATE_ADDRESS:
994 // Indirect register addressing does not use any offsets.
1000 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
1001 return AMDGPU::getVOPe32(Opcode) != -1;
1004 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
1005 // The src0_modifier operand is present on all instructions
1006 // that have modifiers.
1008 return AMDGPU::getNamedOperandIdx(Opcode,
1009 AMDGPU::OpName::src0_modifiers) != -1;
1012 bool SIInstrInfo::hasModifiersSet(const MachineInstr &MI,
1013 unsigned OpName) const {
1014 const MachineOperand *Mods = getNamedOperand(MI, OpName);
1015 return Mods && Mods->getImm();
1018 bool SIInstrInfo::usesConstantBus(const MachineRegisterInfo &MRI,
1019 const MachineOperand &MO) const {
1020 // Literal constants use the constant bus.
1021 if (isLiteralConstant(MO))
1024 if (!MO.isReg() || !MO.isUse())
1027 if (TargetRegisterInfo::isVirtualRegister(MO.getReg()))
1028 return RI.isSGPRClass(MRI.getRegClass(MO.getReg()));
1030 // FLAT_SCR is just an SGPR pair.
1031 if (!MO.isImplicit() && (MO.getReg() == AMDGPU::FLAT_SCR))
1034 // EXEC register uses the constant bus.
1035 if (!MO.isImplicit() && MO.getReg() == AMDGPU::EXEC)
1038 // SGPRs use the constant bus
1039 if (MO.getReg() == AMDGPU::M0 || MO.getReg() == AMDGPU::VCC ||
1040 (!MO.isImplicit() &&
1041 (AMDGPU::SGPR_32RegClass.contains(MO.getReg()) ||
1042 AMDGPU::SGPR_64RegClass.contains(MO.getReg())))) {
1049 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
1050 StringRef &ErrInfo) const {
1051 uint16_t Opcode = MI->getOpcode();
1052 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1053 int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
1054 int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
1055 int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
1057 // Make sure the number of operands is correct.
1058 const MCInstrDesc &Desc = get(Opcode);
1059 if (!Desc.isVariadic() &&
1060 Desc.getNumOperands() != MI->getNumExplicitOperands()) {
1061 ErrInfo = "Instruction has wrong number of operands.";
1065 // Make sure the register classes are correct
1066 for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
1067 switch (Desc.OpInfo[i].OperandType) {
1068 case MCOI::OPERAND_REGISTER: {
1069 if ((MI->getOperand(i).isImm() || MI->getOperand(i).isFPImm()) &&
1070 !isImmOperandLegal(MI, i, MI->getOperand(i))) {
1071 ErrInfo = "Illegal immediate value for operand.";
1076 case MCOI::OPERAND_IMMEDIATE:
1077 // Check if this operand is an immediate.
1078 // FrameIndex operands will be replaced by immediates, so they are
1080 if (!MI->getOperand(i).isImm() && !MI->getOperand(i).isFPImm() &&
1081 !MI->getOperand(i).isFI()) {
1082 ErrInfo = "Expected immediate, but got non-immediate";
1090 if (!MI->getOperand(i).isReg())
1093 int RegClass = Desc.OpInfo[i].RegClass;
1094 if (RegClass != -1) {
1095 unsigned Reg = MI->getOperand(i).getReg();
1096 if (TargetRegisterInfo::isVirtualRegister(Reg))
1099 const TargetRegisterClass *RC = RI.getRegClass(RegClass);
1100 if (!RC->contains(Reg)) {
1101 ErrInfo = "Operand has incorrect register class.";
1109 if (isVOP1(Opcode) || isVOP2(Opcode) || isVOP3(Opcode) || isVOPC(Opcode)) {
1110 // Only look at the true operands. Only a real operand can use the constant
1111 // bus, and we don't want to check pseudo-operands like the source modifier
1113 const int OpIndices[] = { Src0Idx, Src1Idx, Src2Idx };
1115 unsigned ConstantBusCount = 0;
1116 unsigned SGPRUsed = AMDGPU::NoRegister;
1117 for (int OpIdx : OpIndices) {
1121 const MachineOperand &MO = MI->getOperand(OpIdx);
1122 if (usesConstantBus(MRI, MO)) {
1124 if (MO.getReg() != SGPRUsed)
1126 SGPRUsed = MO.getReg();
1132 if (ConstantBusCount > 1) {
1133 ErrInfo = "VOP* instruction uses the constant bus more than once";
1138 // Verify SRC1 for VOP2 and VOPC
1139 if (Src1Idx != -1 && (isVOP2(Opcode) || isVOPC(Opcode))) {
1140 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1141 if (Src1.isImm() || Src1.isFPImm()) {
1142 ErrInfo = "VOP[2C] src1 cannot be an immediate.";
1148 if (isVOP3(Opcode)) {
1149 if (Src0Idx != -1 && isLiteralConstant(MI->getOperand(Src0Idx))) {
1150 ErrInfo = "VOP3 src0 cannot be a literal constant.";
1153 if (Src1Idx != -1 && isLiteralConstant(MI->getOperand(Src1Idx))) {
1154 ErrInfo = "VOP3 src1 cannot be a literal constant.";
1157 if (Src2Idx != -1 && isLiteralConstant(MI->getOperand(Src2Idx))) {
1158 ErrInfo = "VOP3 src2 cannot be a literal constant.";
1163 // Verify misc. restrictions on specific instructions.
1164 if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32 ||
1165 Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64) {
1166 const MachineOperand &Src0 = MI->getOperand(Src0Idx);
1167 const MachineOperand &Src1 = MI->getOperand(Src1Idx);
1168 const MachineOperand &Src2 = MI->getOperand(Src2Idx);
1169 if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
1170 if (!compareMachineOp(Src0, Src1) &&
1171 !compareMachineOp(Src0, Src2)) {
1172 ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
1181 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
1182 switch (MI.getOpcode()) {
1183 default: return AMDGPU::INSTRUCTION_LIST_END;
1184 case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
1185 case AMDGPU::COPY: return AMDGPU::COPY;
1186 case AMDGPU::PHI: return AMDGPU::PHI;
1187 case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
1188 case AMDGPU::S_MOV_B32:
1189 return MI.getOperand(1).isReg() ?
1190 AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
1191 case AMDGPU::S_ADD_I32:
1192 case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
1193 case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
1194 case AMDGPU::S_SUB_I32:
1195 case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
1196 case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
1197 case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
1198 case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
1199 case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
1200 case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;
1201 case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e32;
1202 case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e32;
1203 case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e32;
1204 case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e32;
1205 case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
1206 case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64;
1207 case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
1208 case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
1209 case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
1210 case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
1211 case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
1212 case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
1213 case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
1214 case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
1215 case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
1216 case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
1217 case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
1218 case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
1219 case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
1220 case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;
1221 case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
1222 case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
1223 case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
1224 case AMDGPU::S_LOAD_DWORD_IMM:
1225 case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
1226 case AMDGPU::S_LOAD_DWORDX2_IMM:
1227 case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
1228 case AMDGPU::S_LOAD_DWORDX4_IMM:
1229 case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
1230 case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e32;
1231 case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
1232 case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
1236 bool SIInstrInfo::isSALUOpSupportedOnVALU(const MachineInstr &MI) const {
1237 return getVALUOp(MI) != AMDGPU::INSTRUCTION_LIST_END;
1240 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
1241 unsigned OpNo) const {
1242 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1243 const MCInstrDesc &Desc = get(MI.getOpcode());
1244 if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
1245 Desc.OpInfo[OpNo].RegClass == -1) {
1246 unsigned Reg = MI.getOperand(OpNo).getReg();
1248 if (TargetRegisterInfo::isVirtualRegister(Reg))
1249 return MRI.getRegClass(Reg);
1250 return RI.getRegClass(Reg);
1253 unsigned RCID = Desc.OpInfo[OpNo].RegClass;
1254 return RI.getRegClass(RCID);
1257 bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
1258 switch (MI.getOpcode()) {
1260 case AMDGPU::REG_SEQUENCE:
1262 case AMDGPU::INSERT_SUBREG:
1263 return RI.hasVGPRs(getOpRegClass(MI, 0));
1265 return RI.hasVGPRs(getOpRegClass(MI, OpNo));
1269 void SIInstrInfo::legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const {
1270 MachineBasicBlock::iterator I = MI;
1271 MachineBasicBlock *MBB = MI->getParent();
1272 MachineOperand &MO = MI->getOperand(OpIdx);
1273 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1274 unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
1275 const TargetRegisterClass *RC = RI.getRegClass(RCID);
1276 unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1278 Opcode = AMDGPU::COPY;
1279 else if (RI.isSGPRClass(RC))
1280 Opcode = AMDGPU::S_MOV_B32;
1283 const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
1284 if (RI.getCommonSubClass(&AMDGPU::VReg_64RegClass, VRC))
1285 VRC = &AMDGPU::VReg_64RegClass;
1287 VRC = &AMDGPU::VReg_32RegClass;
1289 unsigned Reg = MRI.createVirtualRegister(VRC);
1290 DebugLoc DL = MBB->findDebugLoc(I);
1291 BuildMI(*MI->getParent(), I, DL, get(Opcode), Reg)
1293 MO.ChangeToRegister(Reg, false);
1296 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
1297 MachineRegisterInfo &MRI,
1298 MachineOperand &SuperReg,
1299 const TargetRegisterClass *SuperRC,
1301 const TargetRegisterClass *SubRC)
1303 assert(SuperReg.isReg());
1305 unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
1306 unsigned SubReg = MRI.createVirtualRegister(SubRC);
1308 // Just in case the super register is itself a sub-register, copy it to a new
1309 // value so we don't need to worry about merging its subreg index with the
1310 // SubIdx passed to this function. The register coalescer should be able to
1311 // eliminate this extra copy.
1312 MachineBasicBlock *MBB = MI->getParent();
1313 DebugLoc DL = MI->getDebugLoc();
1315 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), NewSuperReg)
1316 .addReg(SuperReg.getReg(), 0, SuperReg.getSubReg());
1318 BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
1319 .addReg(NewSuperReg, 0, SubIdx);
1324 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
1325 MachineBasicBlock::iterator MII,
1326 MachineRegisterInfo &MRI,
1328 const TargetRegisterClass *SuperRC,
1330 const TargetRegisterClass *SubRC) const {
1332 // XXX - Is there a better way to do this?
1333 if (SubIdx == AMDGPU::sub0)
1334 return MachineOperand::CreateImm(Op.getImm() & 0xFFFFFFFF);
1335 if (SubIdx == AMDGPU::sub1)
1336 return MachineOperand::CreateImm(Op.getImm() >> 32);
1338 llvm_unreachable("Unhandled register index for immediate");
1341 unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
1343 return MachineOperand::CreateReg(SubReg, false);
1346 unsigned SIInstrInfo::split64BitImm(SmallVectorImpl<MachineInstr *> &Worklist,
1347 MachineBasicBlock::iterator MI,
1348 MachineRegisterInfo &MRI,
1349 const TargetRegisterClass *RC,
1350 const MachineOperand &Op) const {
1351 MachineBasicBlock *MBB = MI->getParent();
1352 DebugLoc DL = MI->getDebugLoc();
1353 unsigned LoDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1354 unsigned HiDst = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1355 unsigned Dst = MRI.createVirtualRegister(RC);
1357 MachineInstr *Lo = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1359 .addImm(Op.getImm() & 0xFFFFFFFF);
1360 MachineInstr *Hi = BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32),
1362 .addImm(Op.getImm() >> 32);
1364 BuildMI(*MBB, MI, DL, get(TargetOpcode::REG_SEQUENCE), Dst)
1366 .addImm(AMDGPU::sub0)
1368 .addImm(AMDGPU::sub1);
1370 Worklist.push_back(Lo);
1371 Worklist.push_back(Hi);
1376 // Change the order of operands from (0, 1, 2) to (0, 2, 1)
1377 void SIInstrInfo::swapOperands(MachineBasicBlock::iterator Inst) const {
1378 assert(Inst->getNumExplicitOperands() == 3);
1379 MachineOperand Op1 = Inst->getOperand(1);
1380 Inst->RemoveOperand(1);
1381 Inst->addOperand(Op1);
1384 bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
1385 const MachineOperand *MO) const {
1386 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1387 const MCInstrDesc &InstDesc = get(MI->getOpcode());
1388 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx];
1389 const TargetRegisterClass *DefinedRC =
1390 OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
1392 MO = &MI->getOperand(OpIdx);
1394 if (usesConstantBus(MRI, *MO)) {
1396 MO->isReg() ? MO->getReg() : (unsigned)AMDGPU::NoRegister;
1397 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1400 if (usesConstantBus(MRI, MI->getOperand(i)) &&
1401 MI->getOperand(i).isReg() && MI->getOperand(i).getReg() != SGPRUsed) {
1409 const TargetRegisterClass *RC = MRI.getRegClass(MO->getReg());
1411 // In order to be legal, the common sub-class must be equal to the
1412 // class of the current operand. For example:
1414 // v_mov_b32 s0 ; Operand defined as vsrc_32
1415 // ; RI.getCommonSubClass(s0,vsrc_32) = sgpr ; LEGAL
1417 // s_sendmsg 0, s0 ; Operand defined as m0reg
1418 // ; RI.getCommonSubClass(s0,m0reg) = m0reg ; NOT LEGAL
1419 return RI.getCommonSubClass(RC, RI.getRegClass(OpInfo.RegClass)) == RC;
1423 // Handle non-register types that are treated like immediates.
1424 assert(MO->isImm() || MO->isFPImm() || MO->isTargetIndex() || MO->isFI());
1427 // This operand expects an immediate.
1431 return isImmOperandLegal(MI, OpIdx, *MO);
1434 void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
1435 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1437 int Src0Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1438 AMDGPU::OpName::src0);
1439 int Src1Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1440 AMDGPU::OpName::src1);
1441 int Src2Idx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1442 AMDGPU::OpName::src2);
1445 if (isVOP2(MI->getOpcode()) && Src1Idx != -1) {
1447 if (!isOperandLegal(MI, Src0Idx))
1448 legalizeOpWithMove(MI, Src0Idx);
1451 if (isOperandLegal(MI, Src1Idx))
1454 // Usually src0 of VOP2 instructions allow more types of inputs
1455 // than src1, so try to commute the instruction to decrease our
1456 // chances of having to insert a MOV instruction to legalize src1.
1457 if (MI->isCommutable()) {
1458 if (commuteInstruction(MI))
1459 // If we are successful in commuting, then we know MI is legal, so
1464 legalizeOpWithMove(MI, Src1Idx);
1468 // XXX - Do any VOP3 instructions read VCC?
1470 if (isVOP3(MI->getOpcode())) {
1471 int VOP3Idx[3] = { Src0Idx, Src1Idx, Src2Idx };
1473 // Find the one SGPR operand we are allowed to use.
1474 unsigned SGPRReg = findUsedSGPR(MI, VOP3Idx);
1476 for (unsigned i = 0; i < 3; ++i) {
1477 int Idx = VOP3Idx[i];
1480 MachineOperand &MO = MI->getOperand(Idx);
1483 if (!RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
1484 continue; // VGPRs are legal
1486 assert(MO.getReg() != AMDGPU::SCC && "SCC operand to VOP3 instruction");
1488 if (SGPRReg == AMDGPU::NoRegister || SGPRReg == MO.getReg()) {
1489 SGPRReg = MO.getReg();
1490 // We can use one SGPR in each VOP3 instruction.
1493 } else if (!isLiteralConstant(MO)) {
1494 // If it is not a register and not a literal constant, then it must be
1495 // an inline constant which is always legal.
1498 // If we make it this far, then the operand is not legal and we must
1500 legalizeOpWithMove(MI, Idx);
1504 // Legalize REG_SEQUENCE and PHI
1505 // The register class of the operands much be the same type as the register
1506 // class of the output.
1507 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE ||
1508 MI->getOpcode() == AMDGPU::PHI) {
1509 const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
1510 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1511 if (!MI->getOperand(i).isReg() ||
1512 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1514 const TargetRegisterClass *OpRC =
1515 MRI.getRegClass(MI->getOperand(i).getReg());
1516 if (RI.hasVGPRs(OpRC)) {
1523 // If any of the operands are VGPR registers, then they all most be
1524 // otherwise we will create illegal VGPR->SGPR copies when legalizing
1526 if (VRC || !RI.isSGPRClass(getOpRegClass(*MI, 0))) {
1529 VRC = RI.getEquivalentVGPRClass(SRC);
1536 // Update all the operands so they have the same type.
1537 for (unsigned i = 1, e = MI->getNumOperands(); i != e; i+=2) {
1538 if (!MI->getOperand(i).isReg() ||
1539 !TargetRegisterInfo::isVirtualRegister(MI->getOperand(i).getReg()))
1541 unsigned DstReg = MRI.createVirtualRegister(RC);
1542 MachineBasicBlock *InsertBB;
1543 MachineBasicBlock::iterator Insert;
1544 if (MI->getOpcode() == AMDGPU::REG_SEQUENCE) {
1545 InsertBB = MI->getParent();
1548 // MI is a PHI instruction.
1549 InsertBB = MI->getOperand(i + 1).getMBB();
1550 Insert = InsertBB->getFirstTerminator();
1552 BuildMI(*InsertBB, Insert, MI->getDebugLoc(),
1553 get(AMDGPU::COPY), DstReg)
1554 .addOperand(MI->getOperand(i));
1555 MI->getOperand(i).setReg(DstReg);
1559 // Legalize INSERT_SUBREG
1560 // src0 must have the same register class as dst
1561 if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
1562 unsigned Dst = MI->getOperand(0).getReg();
1563 unsigned Src0 = MI->getOperand(1).getReg();
1564 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
1565 const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
1566 if (DstRC != Src0RC) {
1567 MachineBasicBlock &MBB = *MI->getParent();
1568 unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
1569 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
1571 MI->getOperand(1).setReg(NewSrc0);
1576 // Legalize MUBUF* instructions
1577 // FIXME: If we start using the non-addr64 instructions for compute, we
1578 // may need to legalize them here.
1580 AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::srsrc);
1581 if (SRsrcIdx != -1) {
1582 // We have an MUBUF instruction
1583 MachineOperand *SRsrc = &MI->getOperand(SRsrcIdx);
1584 unsigned SRsrcRC = get(MI->getOpcode()).OpInfo[SRsrcIdx].RegClass;
1585 if (RI.getCommonSubClass(MRI.getRegClass(SRsrc->getReg()),
1586 RI.getRegClass(SRsrcRC))) {
1587 // The operands are legal.
1588 // FIXME: We may need to legalize operands besided srsrc.
1592 MachineBasicBlock &MBB = *MI->getParent();
1593 // Extract the the ptr from the resource descriptor.
1595 // SRsrcPtrLo = srsrc:sub0
1596 unsigned SRsrcPtrLo = buildExtractSubReg(MI, MRI, *SRsrc,
1597 &AMDGPU::VReg_128RegClass, AMDGPU::sub0, &AMDGPU::VReg_32RegClass);
1599 // SRsrcPtrHi = srsrc:sub1
1600 unsigned SRsrcPtrHi = buildExtractSubReg(MI, MRI, *SRsrc,
1601 &AMDGPU::VReg_128RegClass, AMDGPU::sub1, &AMDGPU::VReg_32RegClass);
1603 // Create an empty resource descriptor
1604 unsigned Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
1605 unsigned SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1606 unsigned SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1607 unsigned NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1608 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
1611 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B64),
1615 // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
1616 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1618 .addImm(RsrcDataFormat & 0xFFFFFFFF);
1620 // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
1621 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1623 .addImm(RsrcDataFormat >> 32);
1625 // NewSRsrc = {Zero64, SRsrcFormat}
1626 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1629 .addImm(AMDGPU::sub0_sub1)
1630 .addReg(SRsrcFormatLo)
1631 .addImm(AMDGPU::sub2)
1632 .addReg(SRsrcFormatHi)
1633 .addImm(AMDGPU::sub3);
1635 MachineOperand *VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1636 unsigned NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
1637 unsigned NewVAddrLo;
1638 unsigned NewVAddrHi;
1640 // This is already an ADDR64 instruction so we need to add the pointer
1641 // extracted from the resource descriptor to the current value of VAddr.
1642 NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1643 NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
1645 // NewVaddrLo = SRsrcPtrLo + VAddr:sub0
1646 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADD_I32_e32),
1649 .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
1650 .addReg(AMDGPU::VCC, RegState::ImplicitDefine);
1652 // NewVaddrHi = SRsrcPtrHi + VAddr:sub1
1653 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::V_ADDC_U32_e32),
1656 .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
1657 .addReg(AMDGPU::VCC, RegState::ImplicitDefine)
1658 .addReg(AMDGPU::VCC, RegState::Implicit);
1661 // This instructions is the _OFFSET variant, so we need to convert it to
1663 MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
1664 MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
1665 MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
1666 assert(SOffset->isImm() && SOffset->getImm() == 0 && "Legalizing MUBUF "
1667 "with non-zero soffset is not implemented");
1670 // Create the new instruction.
1671 unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
1672 MachineInstr *Addr64 =
1673 BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
1676 .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
1677 // This will be replaced later
1678 // with the new value of vaddr.
1679 .addOperand(*Offset);
1681 MI->removeFromParent();
1684 NewVAddrLo = SRsrcPtrLo;
1685 NewVAddrHi = SRsrcPtrHi;
1686 VAddr = getNamedOperand(*MI, AMDGPU::OpName::vaddr);
1687 SRsrc = getNamedOperand(*MI, AMDGPU::OpName::srsrc);
1690 // NewVaddr = {NewVaddrHi, NewVaddrLo}
1691 BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
1694 .addImm(AMDGPU::sub0)
1696 .addImm(AMDGPU::sub1);
1699 // Update the instruction to use NewVaddr
1700 VAddr->setReg(NewVAddr);
1701 // Update the instruction to use NewSRsrc
1702 SRsrc->setReg(NewSRsrc);
1706 void SIInstrInfo::splitSMRD(MachineInstr *MI,
1707 const TargetRegisterClass *HalfRC,
1708 unsigned HalfImmOp, unsigned HalfSGPROp,
1709 MachineInstr *&Lo, MachineInstr *&Hi) const {
1711 DebugLoc DL = MI->getDebugLoc();
1712 MachineBasicBlock *MBB = MI->getParent();
1713 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1714 unsigned RegLo = MRI.createVirtualRegister(HalfRC);
1715 unsigned RegHi = MRI.createVirtualRegister(HalfRC);
1716 unsigned HalfSize = HalfRC->getSize();
1717 const MachineOperand *OffOp =
1718 getNamedOperand(*MI, AMDGPU::OpName::offset);
1719 const MachineOperand *SBase = getNamedOperand(*MI, AMDGPU::OpName::sbase);
1721 // The SMRD has an 8-bit offset in dwords on SI and a 20-bit offset in bytes
1724 bool isVI = RI.ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS;
1725 unsigned OffScale = isVI ? 1 : 4;
1726 // Handle the _IMM variant
1727 unsigned LoOffset = OffOp->getImm() * OffScale;
1728 unsigned HiOffset = LoOffset + HalfSize;
1729 Lo = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegLo)
1731 .addImm(LoOffset / OffScale);
1733 if (!isUInt<20>(HiOffset) || (!isVI && !isUInt<8>(HiOffset / OffScale))) {
1734 unsigned OffsetSGPR =
1735 MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1736 BuildMI(*MBB, MI, DL, get(AMDGPU::S_MOV_B32), OffsetSGPR)
1737 .addImm(HiOffset); // The offset in register is in bytes.
1738 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegHi)
1740 .addReg(OffsetSGPR);
1742 Hi = BuildMI(*MBB, MI, DL, get(HalfImmOp), RegHi)
1744 .addImm(HiOffset / OffScale);
1747 // Handle the _SGPR variant
1748 MachineOperand *SOff = getNamedOperand(*MI, AMDGPU::OpName::soff);
1749 Lo = BuildMI(*MBB, MI, DL, get(HalfSGPROp), RegLo)
1752 unsigned OffsetSGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
1753 BuildMI(*MBB, MI, DL, get(AMDGPU::S_ADD_I32), OffsetSGPR)
1756 Hi = BuildMI(*MBB, MI, DL, get(HalfSGPROp))
1758 .addReg(OffsetSGPR);
1761 unsigned SubLo, SubHi;
1764 SubLo = AMDGPU::sub0;
1765 SubHi = AMDGPU::sub1;
1768 SubLo = AMDGPU::sub0_sub1;
1769 SubHi = AMDGPU::sub2_sub3;
1772 SubLo = AMDGPU::sub0_sub1_sub2_sub3;
1773 SubHi = AMDGPU::sub4_sub5_sub6_sub7;
1776 SubLo = AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
1777 SubHi = AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15;
1780 llvm_unreachable("Unhandled HalfSize");
1783 BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE))
1784 .addOperand(MI->getOperand(0))
1791 void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
1792 MachineBasicBlock *MBB = MI->getParent();
1793 switch (MI->getOpcode()) {
1794 case AMDGPU::S_LOAD_DWORD_IMM:
1795 case AMDGPU::S_LOAD_DWORD_SGPR:
1796 case AMDGPU::S_LOAD_DWORDX2_IMM:
1797 case AMDGPU::S_LOAD_DWORDX2_SGPR:
1798 case AMDGPU::S_LOAD_DWORDX4_IMM:
1799 case AMDGPU::S_LOAD_DWORDX4_SGPR: {
1800 unsigned NewOpcode = getVALUOp(*MI);
1804 if (MI->getOperand(2).isReg()) {
1805 RegOffset = MI->getOperand(2).getReg();
1808 assert(MI->getOperand(2).isImm());
1809 // SMRD instructions take a dword offsets on SI and byte offset on VI
1810 // and MUBUF instructions always take a byte offset.
1811 ImmOffset = MI->getOperand(2).getImm();
1812 if (RI.ST.getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS)
1814 RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1816 if (isUInt<12>(ImmOffset)) {
1817 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1821 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
1828 unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
1829 unsigned DWord0 = RegOffset;
1830 unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1831 unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1832 unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
1833 uint64_t RsrcDataFormat = getDefaultRsrcDataFormat();
1835 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
1837 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
1838 .addImm(RsrcDataFormat & 0xFFFFFFFF);
1839 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
1840 .addImm(RsrcDataFormat >> 32);
1841 BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
1843 .addImm(AMDGPU::sub0)
1845 .addImm(AMDGPU::sub1)
1847 .addImm(AMDGPU::sub2)
1849 .addImm(AMDGPU::sub3);
1850 MI->setDesc(get(NewOpcode));
1851 if (MI->getOperand(2).isReg()) {
1852 MI->getOperand(2).setReg(MI->getOperand(1).getReg());
1854 MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
1856 MI->getOperand(1).setReg(SRsrc);
1857 MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
1859 const TargetRegisterClass *NewDstRC =
1860 RI.getRegClass(get(NewOpcode).OpInfo[0].RegClass);
1862 unsigned DstReg = MI->getOperand(0).getReg();
1863 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
1864 MRI.replaceRegWith(DstReg, NewDstReg);
1867 case AMDGPU::S_LOAD_DWORDX8_IMM:
1868 case AMDGPU::S_LOAD_DWORDX8_SGPR: {
1869 MachineInstr *Lo, *Hi;
1870 splitSMRD(MI, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM,
1871 AMDGPU::S_LOAD_DWORDX4_SGPR, Lo, Hi);
1872 MI->eraseFromParent();
1873 moveSMRDToVALU(Lo, MRI);
1874 moveSMRDToVALU(Hi, MRI);
1878 case AMDGPU::S_LOAD_DWORDX16_IMM:
1879 case AMDGPU::S_LOAD_DWORDX16_SGPR: {
1880 MachineInstr *Lo, *Hi;
1881 splitSMRD(MI, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM,
1882 AMDGPU::S_LOAD_DWORDX8_SGPR, Lo, Hi);
1883 MI->eraseFromParent();
1884 moveSMRDToVALU(Lo, MRI);
1885 moveSMRDToVALU(Hi, MRI);
1891 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
1892 SmallVector<MachineInstr *, 128> Worklist;
1893 Worklist.push_back(&TopInst);
1895 while (!Worklist.empty()) {
1896 MachineInstr *Inst = Worklist.pop_back_val();
1897 MachineBasicBlock *MBB = Inst->getParent();
1898 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1900 unsigned Opcode = Inst->getOpcode();
1901 unsigned NewOpcode = getVALUOp(*Inst);
1903 // Handle some special cases
1906 if (isSMRD(Inst->getOpcode())) {
1907 moveSMRDToVALU(Inst, MRI);
1910 case AMDGPU::S_MOV_B64: {
1911 DebugLoc DL = Inst->getDebugLoc();
1913 // If the source operand is a register we can replace this with a
1915 if (Inst->getOperand(1).isReg()) {
1916 MachineInstr *Copy = BuildMI(*MBB, Inst, DL, get(TargetOpcode::COPY))
1917 .addOperand(Inst->getOperand(0))
1918 .addOperand(Inst->getOperand(1));
1919 Worklist.push_back(Copy);
1921 // Otherwise, we need to split this into two movs, because there is
1922 // no 64-bit VALU move instruction.
1923 unsigned Reg = Inst->getOperand(0).getReg();
1924 unsigned Dst = split64BitImm(Worklist,
1927 MRI.getRegClass(Reg),
1928 Inst->getOperand(1));
1929 MRI.replaceRegWith(Reg, Dst);
1931 Inst->eraseFromParent();
1934 case AMDGPU::S_AND_B64:
1935 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
1936 Inst->eraseFromParent();
1939 case AMDGPU::S_OR_B64:
1940 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
1941 Inst->eraseFromParent();
1944 case AMDGPU::S_XOR_B64:
1945 splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
1946 Inst->eraseFromParent();
1949 case AMDGPU::S_NOT_B64:
1950 splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
1951 Inst->eraseFromParent();
1954 case AMDGPU::S_BCNT1_I32_B64:
1955 splitScalar64BitBCNT(Worklist, Inst);
1956 Inst->eraseFromParent();
1959 case AMDGPU::S_BFE_I64: {
1960 splitScalar64BitBFE(Worklist, Inst);
1961 Inst->eraseFromParent();
1965 case AMDGPU::S_LSHL_B32:
1966 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
1967 NewOpcode = AMDGPU::V_LSHLREV_B32_e64;
1971 case AMDGPU::S_ASHR_I32:
1972 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
1973 NewOpcode = AMDGPU::V_ASHRREV_I32_e64;
1977 case AMDGPU::S_LSHR_B32:
1978 if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
1979 NewOpcode = AMDGPU::V_LSHRREV_B32_e64;
1984 case AMDGPU::S_BFE_U64:
1985 case AMDGPU::S_BFM_B64:
1986 llvm_unreachable("Moving this op to VALU not implemented");
1989 if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
1990 // We cannot move this instruction to the VALU, so we should try to
1991 // legalize its operands instead.
1992 legalizeOperands(Inst);
1996 // Use the new VALU Opcode.
1997 const MCInstrDesc &NewDesc = get(NewOpcode);
1998 Inst->setDesc(NewDesc);
2000 // Remove any references to SCC. Vector instructions can't read from it, and
2001 // We're just about to add the implicit use / defs of VCC, and we don't want
2003 for (unsigned i = Inst->getNumOperands() - 1; i > 0; --i) {
2004 MachineOperand &Op = Inst->getOperand(i);
2005 if (Op.isReg() && Op.getReg() == AMDGPU::SCC)
2006 Inst->RemoveOperand(i);
2009 if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
2010 // We are converting these to a BFE, so we need to add the missing
2011 // operands for the size and offset.
2012 unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
2013 Inst->addOperand(MachineOperand::CreateImm(0));
2014 Inst->addOperand(MachineOperand::CreateImm(Size));
2016 } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
2017 // The VALU version adds the second operand to the result, so insert an
2019 Inst->addOperand(MachineOperand::CreateImm(0));
2022 addDescImplicitUseDef(NewDesc, Inst);
2024 if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
2025 const MachineOperand &OffsetWidthOp = Inst->getOperand(2);
2026 // If we need to move this to VGPRs, we need to unpack the second operand
2027 // back into the 2 separate ones for bit offset and width.
2028 assert(OffsetWidthOp.isImm() &&
2029 "Scalar BFE is only implemented for constant width and offset");
2030 uint32_t Imm = OffsetWidthOp.getImm();
2032 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
2033 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
2034 Inst->RemoveOperand(2); // Remove old immediate.
2035 Inst->addOperand(MachineOperand::CreateImm(Offset));
2036 Inst->addOperand(MachineOperand::CreateImm(BitWidth));
2039 // Update the destination register class.
2041 const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
2044 // For target instructions, getOpRegClass just returns the virtual
2045 // register class associated with the operand, so we need to find an
2046 // equivalent VGPR register class in order to move the instruction to the
2050 case AMDGPU::REG_SEQUENCE:
2051 case AMDGPU::INSERT_SUBREG:
2052 if (RI.hasVGPRs(NewDstRC))
2054 NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
2062 unsigned DstReg = Inst->getOperand(0).getReg();
2063 unsigned NewDstReg = MRI.createVirtualRegister(NewDstRC);
2064 MRI.replaceRegWith(DstReg, NewDstReg);
2066 // Legalize the operands
2067 legalizeOperands(Inst);
2069 for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
2070 E = MRI.use_end(); I != E; ++I) {
2071 MachineInstr &UseMI = *I->getParent();
2072 if (!canReadVGPR(UseMI, I.getOperandNo())) {
2073 Worklist.push_back(&UseMI);
2079 //===----------------------------------------------------------------------===//
2080 // Indirect addressing callbacks
2081 //===----------------------------------------------------------------------===//
2083 unsigned SIInstrInfo::calculateIndirectAddress(unsigned RegIndex,
2084 unsigned Channel) const {
2085 assert(Channel == 0);
2089 const TargetRegisterClass *SIInstrInfo::getIndirectAddrRegClass() const {
2090 return &AMDGPU::VReg_32RegClass;
2093 void SIInstrInfo::splitScalar64BitUnaryOp(
2094 SmallVectorImpl<MachineInstr *> &Worklist,
2096 unsigned Opcode) const {
2097 MachineBasicBlock &MBB = *Inst->getParent();
2098 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2100 MachineOperand &Dest = Inst->getOperand(0);
2101 MachineOperand &Src0 = Inst->getOperand(1);
2102 DebugLoc DL = Inst->getDebugLoc();
2104 MachineBasicBlock::iterator MII = Inst;
2106 const MCInstrDesc &InstDesc = get(Opcode);
2107 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2108 MRI.getRegClass(Src0.getReg()) :
2109 &AMDGPU::SGPR_32RegClass;
2111 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2113 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2114 AMDGPU::sub0, Src0SubRC);
2116 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2117 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2119 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2120 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2121 .addOperand(SrcReg0Sub0);
2123 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2124 AMDGPU::sub1, Src0SubRC);
2126 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2127 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2128 .addOperand(SrcReg0Sub1);
2130 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2131 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2133 .addImm(AMDGPU::sub0)
2135 .addImm(AMDGPU::sub1);
2137 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2139 // Try to legalize the operands in case we need to swap the order to keep it
2141 Worklist.push_back(LoHalf);
2142 Worklist.push_back(HiHalf);
2145 void SIInstrInfo::splitScalar64BitBinaryOp(
2146 SmallVectorImpl<MachineInstr *> &Worklist,
2148 unsigned Opcode) const {
2149 MachineBasicBlock &MBB = *Inst->getParent();
2150 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2152 MachineOperand &Dest = Inst->getOperand(0);
2153 MachineOperand &Src0 = Inst->getOperand(1);
2154 MachineOperand &Src1 = Inst->getOperand(2);
2155 DebugLoc DL = Inst->getDebugLoc();
2157 MachineBasicBlock::iterator MII = Inst;
2159 const MCInstrDesc &InstDesc = get(Opcode);
2160 const TargetRegisterClass *Src0RC = Src0.isReg() ?
2161 MRI.getRegClass(Src0.getReg()) :
2162 &AMDGPU::SGPR_32RegClass;
2164 const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
2165 const TargetRegisterClass *Src1RC = Src1.isReg() ?
2166 MRI.getRegClass(Src1.getReg()) :
2167 &AMDGPU::SGPR_32RegClass;
2169 const TargetRegisterClass *Src1SubRC = RI.getSubRegClass(Src1RC, AMDGPU::sub0);
2171 MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2172 AMDGPU::sub0, Src0SubRC);
2173 MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2174 AMDGPU::sub0, Src1SubRC);
2176 const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
2177 const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
2179 unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
2180 MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
2181 .addOperand(SrcReg0Sub0)
2182 .addOperand(SrcReg1Sub0);
2184 MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
2185 AMDGPU::sub1, Src0SubRC);
2186 MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
2187 AMDGPU::sub1, Src1SubRC);
2189 unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
2190 MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
2191 .addOperand(SrcReg0Sub1)
2192 .addOperand(SrcReg1Sub1);
2194 unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
2195 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
2197 .addImm(AMDGPU::sub0)
2199 .addImm(AMDGPU::sub1);
2201 MRI.replaceRegWith(Dest.getReg(), FullDestReg);
2203 // Try to legalize the operands in case we need to swap the order to keep it
2205 Worklist.push_back(LoHalf);
2206 Worklist.push_back(HiHalf);
2209 void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
2210 MachineInstr *Inst) const {
2211 MachineBasicBlock &MBB = *Inst->getParent();
2212 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2214 MachineBasicBlock::iterator MII = Inst;
2215 DebugLoc DL = Inst->getDebugLoc();
2217 MachineOperand &Dest = Inst->getOperand(0);
2218 MachineOperand &Src = Inst->getOperand(1);
2220 const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e32);
2221 const TargetRegisterClass *SrcRC = Src.isReg() ?
2222 MRI.getRegClass(Src.getReg()) :
2223 &AMDGPU::SGPR_32RegClass;
2225 unsigned MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2226 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2228 const TargetRegisterClass *SrcSubRC = RI.getSubRegClass(SrcRC, AMDGPU::sub0);
2230 MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2231 AMDGPU::sub0, SrcSubRC);
2232 MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
2233 AMDGPU::sub1, SrcSubRC);
2235 MachineInstr *First = BuildMI(MBB, MII, DL, InstDesc, MidReg)
2236 .addOperand(SrcRegSub0)
2239 MachineInstr *Second = BuildMI(MBB, MII, DL, InstDesc, ResultReg)
2240 .addOperand(SrcRegSub1)
2243 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2245 Worklist.push_back(First);
2246 Worklist.push_back(Second);
2249 void SIInstrInfo::splitScalar64BitBFE(SmallVectorImpl<MachineInstr *> &Worklist,
2250 MachineInstr *Inst) const {
2251 MachineBasicBlock &MBB = *Inst->getParent();
2252 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2253 MachineBasicBlock::iterator MII = Inst;
2254 DebugLoc DL = Inst->getDebugLoc();
2256 MachineOperand &Dest = Inst->getOperand(0);
2257 uint32_t Imm = Inst->getOperand(2).getImm();
2258 uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
2259 uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
2263 // Only sext_inreg cases handled.
2264 assert(Inst->getOpcode() == AMDGPU::S_BFE_I64 &&
2269 if (BitWidth < 32) {
2270 unsigned MidRegLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2271 unsigned MidRegHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2272 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
2274 BuildMI(MBB, MII, DL, get(AMDGPU::V_BFE_I32), MidRegLo)
2275 .addReg(Inst->getOperand(1).getReg(), 0, AMDGPU::sub0)
2279 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e32), MidRegHi)
2283 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
2285 .addImm(AMDGPU::sub0)
2287 .addImm(AMDGPU::sub1);
2289 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2293 MachineOperand &Src = Inst->getOperand(1);
2294 unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2295 unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
2297 BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e64), TmpReg)
2299 .addReg(Src.getReg(), 0, AMDGPU::sub0);
2301 BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
2302 .addReg(Src.getReg(), 0, AMDGPU::sub0)
2303 .addImm(AMDGPU::sub0)
2305 .addImm(AMDGPU::sub1);
2307 MRI.replaceRegWith(Dest.getReg(), ResultReg);
2310 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
2311 MachineInstr *Inst) const {
2312 // Add the implict and explicit register definitions.
2313 if (NewDesc.ImplicitUses) {
2314 for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
2315 unsigned Reg = NewDesc.ImplicitUses[i];
2316 Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
2320 if (NewDesc.ImplicitDefs) {
2321 for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
2322 unsigned Reg = NewDesc.ImplicitDefs[i];
2323 Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
2328 unsigned SIInstrInfo::findUsedSGPR(const MachineInstr *MI,
2329 int OpIndices[3]) const {
2330 const MCInstrDesc &Desc = get(MI->getOpcode());
2332 // Find the one SGPR operand we are allowed to use.
2333 unsigned SGPRReg = AMDGPU::NoRegister;
2335 // First we need to consider the instruction's operand requirements before
2336 // legalizing. Some operands are required to be SGPRs, such as implicit uses
2337 // of VCC, but we are still bound by the constant bus requirement to only use
2340 // If the operand's class is an SGPR, we can never move it.
2342 for (const MachineOperand &MO : MI->implicit_operands()) {
2343 // We only care about reads.
2347 if (MO.getReg() == AMDGPU::VCC)
2350 if (MO.getReg() == AMDGPU::FLAT_SCR)
2351 return AMDGPU::FLAT_SCR;
2354 unsigned UsedSGPRs[3] = { AMDGPU::NoRegister };
2355 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
2357 for (unsigned i = 0; i < 3; ++i) {
2358 int Idx = OpIndices[i];
2362 const MachineOperand &MO = MI->getOperand(Idx);
2363 if (RI.isSGPRClassID(Desc.OpInfo[Idx].RegClass))
2364 SGPRReg = MO.getReg();
2366 if (MO.isReg() && RI.isSGPRClass(MRI.getRegClass(MO.getReg())))
2367 UsedSGPRs[i] = MO.getReg();
2370 if (SGPRReg != AMDGPU::NoRegister)
2373 // We don't have a required SGPR operand, so we have a bit more freedom in
2374 // selecting operands to move.
2376 // Try to select the most used SGPR. If an SGPR is equal to one of the
2377 // others, we choose that.
2380 // V_FMA_F32 v0, s0, s0, s0 -> No moves
2381 // V_FMA_F32 v0, s0, s1, s0 -> Move s1
2383 if (UsedSGPRs[0] != AMDGPU::NoRegister) {
2384 if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
2385 SGPRReg = UsedSGPRs[0];
2388 if (SGPRReg == AMDGPU::NoRegister && UsedSGPRs[1] != AMDGPU::NoRegister) {
2389 if (UsedSGPRs[1] == UsedSGPRs[2])
2390 SGPRReg = UsedSGPRs[1];
2396 MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
2397 MachineBasicBlock *MBB,
2398 MachineBasicBlock::iterator I,
2400 unsigned Address, unsigned OffsetReg) const {
2401 const DebugLoc &DL = MBB->findDebugLoc(I);
2402 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
2403 getIndirectIndexBegin(*MBB->getParent()));
2405 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_DST_V1))
2406 .addReg(IndirectBaseReg, RegState::Define)
2407 .addOperand(I->getOperand(0))
2408 .addReg(IndirectBaseReg)
2414 MachineInstrBuilder SIInstrInfo::buildIndirectRead(
2415 MachineBasicBlock *MBB,
2416 MachineBasicBlock::iterator I,
2418 unsigned Address, unsigned OffsetReg) const {
2419 const DebugLoc &DL = MBB->findDebugLoc(I);
2420 unsigned IndirectBaseReg = AMDGPU::VReg_32RegClass.getRegister(
2421 getIndirectIndexBegin(*MBB->getParent()));
2423 return BuildMI(*MBB, I, DL, get(AMDGPU::SI_INDIRECT_SRC))
2424 .addOperand(I->getOperand(0))
2425 .addOperand(I->getOperand(1))
2426 .addReg(IndirectBaseReg)
2432 void SIInstrInfo::reserveIndirectRegisters(BitVector &Reserved,
2433 const MachineFunction &MF) const {
2434 int End = getIndirectIndexEnd(MF);
2435 int Begin = getIndirectIndexBegin(MF);
2441 for (int Index = Begin; Index <= End; ++Index)
2442 Reserved.set(AMDGPU::VReg_32RegClass.getRegister(Index));
2444 for (int Index = std::max(0, Begin - 1); Index <= End; ++Index)
2445 Reserved.set(AMDGPU::VReg_64RegClass.getRegister(Index));
2447 for (int Index = std::max(0, Begin - 2); Index <= End; ++Index)
2448 Reserved.set(AMDGPU::VReg_96RegClass.getRegister(Index));
2450 for (int Index = std::max(0, Begin - 3); Index <= End; ++Index)
2451 Reserved.set(AMDGPU::VReg_128RegClass.getRegister(Index));
2453 for (int Index = std::max(0, Begin - 7); Index <= End; ++Index)
2454 Reserved.set(AMDGPU::VReg_256RegClass.getRegister(Index));
2456 for (int Index = std::max(0, Begin - 15); Index <= End; ++Index)
2457 Reserved.set(AMDGPU::VReg_512RegClass.getRegister(Index));
2460 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
2461 unsigned OperandName) const {
2462 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
2466 return &MI.getOperand(Idx);
2469 uint64_t SIInstrInfo::getDefaultRsrcDataFormat() const {
2470 uint64_t RsrcDataFormat = AMDGPU::RSRC_DATA_FORMAT;
2471 if (ST.isAmdHsaOS())
2472 RsrcDataFormat |= (1ULL << 56);
2474 return RsrcDataFormat;