1 //===----------------------- SIFrameLowering.cpp --------------------------===//
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 //==-----------------------------------------------------------------------===//
10 #include "SIFrameLowering.h"
11 #include "AMDGPUSubtarget.h"
12 #include "SIInstrInfo.h"
13 #include "SIMachineFunctionInfo.h"
14 #include "SIRegisterInfo.h"
16 #include "llvm/CodeGen/MachineFrameInfo.h"
17 #include "llvm/CodeGen/MachineFunction.h"
18 #include "llvm/CodeGen/MachineInstrBuilder.h"
19 #include "llvm/CodeGen/RegisterScavenging.h"
24 static ArrayRef<MCPhysReg> getAllSGPR128(const SISubtarget &ST,
25 const MachineFunction &MF) {
26 return makeArrayRef(AMDGPU::SGPR_128RegClass.begin(),
27 ST.getMaxNumSGPRs(MF) / 4);
30 static ArrayRef<MCPhysReg> getAllSGPRs(const SISubtarget &ST,
31 const MachineFunction &MF) {
32 return makeArrayRef(AMDGPU::SGPR_32RegClass.begin(),
33 ST.getMaxNumSGPRs(MF));
36 void SIFrameLowering::emitFlatScratchInit(const SISubtarget &ST,
38 MachineBasicBlock &MBB) const {
39 const SIInstrInfo *TII = ST.getInstrInfo();
40 const SIRegisterInfo* TRI = &TII->getRegisterInfo();
42 // We don't need this if we only have spills since there is no user facing
45 // TODO: If we know we don't have flat instructions earlier, we can omit
46 // this from the input registers.
48 // TODO: We only need to know if we access scratch space through a flat
49 // pointer. Because we only detect if flat instructions are used at all,
50 // this will be used more often than necessary on VI.
52 // Debug location must be unknown since the first debug location is used to
53 // determine the end of the prologue.
55 MachineBasicBlock::iterator I = MBB.begin();
57 unsigned FlatScratchInitReg
58 = TRI->getPreloadedValue(MF, SIRegisterInfo::FLAT_SCRATCH_INIT);
60 MachineRegisterInfo &MRI = MF.getRegInfo();
61 MRI.addLiveIn(FlatScratchInitReg);
62 MBB.addLiveIn(FlatScratchInitReg);
64 unsigned FlatScrInitLo = TRI->getSubReg(FlatScratchInitReg, AMDGPU::sub0);
65 unsigned FlatScrInitHi = TRI->getSubReg(FlatScratchInitReg, AMDGPU::sub1);
67 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
68 unsigned ScratchWaveOffsetReg = MFI->getScratchWaveOffsetReg();
70 // Do a 64-bit pointer add.
71 if (ST.flatScratchIsPointer()) {
72 BuildMI(MBB, I, DL, TII->get(AMDGPU::S_ADD_U32), AMDGPU::FLAT_SCR_LO)
73 .addReg(FlatScrInitLo)
74 .addReg(ScratchWaveOffsetReg);
75 BuildMI(MBB, I, DL, TII->get(AMDGPU::S_ADDC_U32), AMDGPU::FLAT_SCR_HI)
76 .addReg(FlatScrInitHi)
82 // Copy the size in bytes.
83 BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), AMDGPU::FLAT_SCR_LO)
84 .addReg(FlatScrInitHi, RegState::Kill);
86 // Add wave offset in bytes to private base offset.
87 // See comment in AMDKernelCodeT.h for enable_sgpr_flat_scratch_init.
88 BuildMI(MBB, I, DL, TII->get(AMDGPU::S_ADD_U32), FlatScrInitLo)
89 .addReg(FlatScrInitLo)
90 .addReg(ScratchWaveOffsetReg);
92 // Convert offset to 256-byte units.
93 BuildMI(MBB, I, DL, TII->get(AMDGPU::S_LSHR_B32), AMDGPU::FLAT_SCR_HI)
94 .addReg(FlatScrInitLo, RegState::Kill)
98 unsigned SIFrameLowering::getReservedPrivateSegmentBufferReg(
99 const SISubtarget &ST,
100 const SIInstrInfo *TII,
101 const SIRegisterInfo *TRI,
102 SIMachineFunctionInfo *MFI,
103 MachineFunction &MF) const {
104 MachineRegisterInfo &MRI = MF.getRegInfo();
106 // We need to insert initialization of the scratch resource descriptor.
107 unsigned ScratchRsrcReg = MFI->getScratchRSrcReg();
108 if (ScratchRsrcReg == AMDGPU::NoRegister ||
109 !MRI.isPhysRegUsed(ScratchRsrcReg))
110 return AMDGPU::NoRegister;
112 if (ST.hasSGPRInitBug() ||
113 ScratchRsrcReg != TRI->reservedPrivateSegmentBufferReg(MF))
114 return ScratchRsrcReg;
116 // We reserved the last registers for this. Shift it down to the end of those
117 // which were actually used.
119 // FIXME: It might be safer to use a pseudoregister before replacement.
121 // FIXME: We should be able to eliminate unused input registers. We only
122 // cannot do this for the resources required for scratch access. For now we
123 // skip over user SGPRs and may leave unused holes.
125 // We find the resource first because it has an alignment requirement.
127 unsigned NumPreloaded = (MFI->getNumPreloadedSGPRs() + 3) / 4;
128 ArrayRef<MCPhysReg> AllSGPR128s = getAllSGPR128(ST, MF);
129 AllSGPR128s = AllSGPR128s.slice(std::min(static_cast<unsigned>(AllSGPR128s.size()), NumPreloaded));
131 // Skip the last N reserved elements because they should have already been
132 // reserved for VCC etc.
133 for (MCPhysReg Reg : AllSGPR128s) {
134 // Pick the first unallocated one. Make sure we don't clobber the other
135 // reserved input we needed.
136 if (!MRI.isPhysRegUsed(Reg) && MRI.isAllocatable(Reg)) {
137 MRI.replaceRegWith(ScratchRsrcReg, Reg);
138 MFI->setScratchRSrcReg(Reg);
143 return ScratchRsrcReg;
146 // Shift down registers reserved for the scratch wave offset and stack pointer
148 std::pair<unsigned, unsigned>
149 SIFrameLowering::getReservedPrivateSegmentWaveByteOffsetReg(
150 const SISubtarget &ST,
151 const SIInstrInfo *TII,
152 const SIRegisterInfo *TRI,
153 SIMachineFunctionInfo *MFI,
154 MachineFunction &MF) const {
155 MachineRegisterInfo &MRI = MF.getRegInfo();
156 unsigned ScratchWaveOffsetReg = MFI->getScratchWaveOffsetReg();
158 // No replacement necessary.
159 if (ScratchWaveOffsetReg == AMDGPU::NoRegister ||
160 !MRI.isPhysRegUsed(ScratchWaveOffsetReg)) {
161 assert(MFI->getStackPtrOffsetReg() == AMDGPU::NoRegister);
162 return std::make_pair(AMDGPU::NoRegister, AMDGPU::NoRegister);
165 unsigned SPReg = MFI->getStackPtrOffsetReg();
166 if (ST.hasSGPRInitBug())
167 return std::make_pair(ScratchWaveOffsetReg, SPReg);
169 unsigned NumPreloaded = MFI->getNumPreloadedSGPRs();
171 ArrayRef<MCPhysReg> AllSGPRs = getAllSGPRs(ST, MF);
172 if (NumPreloaded > AllSGPRs.size())
173 return std::make_pair(ScratchWaveOffsetReg, SPReg);
175 AllSGPRs = AllSGPRs.slice(NumPreloaded);
177 // We need to drop register from the end of the list that we cannot use
178 // for the scratch wave offset.
179 // + 2 s102 and s103 do not exist on VI.
181 // + 2 for xnack_mask
182 // + 2 for flat_scratch
183 // + 4 for registers reserved for scratch resource register
184 // + 1 for register reserved for scratch wave offset. (By exluding this
185 // register from the list to consider, it means that when this
186 // register is being used for the scratch wave offset and there
187 // are no other free SGPRs, then the value will stay in this register.
188 // + 1 if stack pointer is used.
191 unsigned ReservedRegCount = 13;
193 if (AllSGPRs.size() < ReservedRegCount)
194 return std::make_pair(ScratchWaveOffsetReg, SPReg);
196 bool HandledScratchWaveOffsetReg =
197 ScratchWaveOffsetReg != TRI->reservedPrivateSegmentWaveByteOffsetReg(MF);
199 for (MCPhysReg Reg : AllSGPRs.drop_back(ReservedRegCount)) {
200 // Pick the first unallocated SGPR. Be careful not to pick an alias of the
201 // scratch descriptor, since we haven’t added its uses yet.
202 if (!MRI.isPhysRegUsed(Reg) && MRI.isAllocatable(Reg)) {
203 if (!HandledScratchWaveOffsetReg) {
204 HandledScratchWaveOffsetReg = true;
206 MRI.replaceRegWith(ScratchWaveOffsetReg, Reg);
207 MFI->setScratchWaveOffsetReg(Reg);
208 ScratchWaveOffsetReg = Reg;
214 return std::make_pair(ScratchWaveOffsetReg, SPReg);
217 void SIFrameLowering::emitEntryFunctionPrologue(MachineFunction &MF,
218 MachineBasicBlock &MBB) const {
219 // Emit debugger prologue if "amdgpu-debugger-emit-prologue" attribute was
221 const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
222 auto AMDGPUASI = ST.getAMDGPUAS();
223 if (ST.debuggerEmitPrologue())
224 emitDebuggerPrologue(MF, MBB);
226 assert(&MF.front() == &MBB && "Shrink-wrapping not yet supported");
228 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
230 // If we only have SGPR spills, we won't actually be using scratch memory
231 // since these spill to VGPRs.
233 // FIXME: We should be cleaning up these unused SGPR spill frame indices
236 const SIInstrInfo *TII = ST.getInstrInfo();
237 const SIRegisterInfo *TRI = &TII->getRegisterInfo();
238 MachineRegisterInfo &MRI = MF.getRegInfo();
240 // We need to do the replacement of the private segment buffer and wave offset
241 // register even if there are no stack objects. There could be stores to undef
242 // or a constant without an associated object.
244 // FIXME: We still have implicit uses on SGPR spill instructions in case they
245 // need to spill to vector memory. It's likely that will not happen, but at
246 // this point it appears we need the setup. This part of the prolog should be
247 // emitted after frame indices are eliminated.
249 if (MF.getFrameInfo().hasStackObjects() && MFI->hasFlatScratchInit())
250 emitFlatScratchInit(ST, MF, MBB);
252 unsigned SPReg = MFI->getStackPtrOffsetReg();
253 if (SPReg != AMDGPU::NoRegister) {
255 int64_t StackSize = MF.getFrameInfo().getStackSize();
257 if (StackSize == 0) {
258 BuildMI(MBB, MBB.begin(), DL, TII->get(AMDGPU::COPY), SPReg)
259 .addReg(MFI->getScratchWaveOffsetReg());
261 BuildMI(MBB, MBB.begin(), DL, TII->get(AMDGPU::S_ADD_U32), SPReg)
262 .addReg(MFI->getScratchWaveOffsetReg())
263 .addImm(StackSize * ST.getWavefrontSize());
267 unsigned ScratchRsrcReg
268 = getReservedPrivateSegmentBufferReg(ST, TII, TRI, MFI, MF);
270 unsigned ScratchWaveOffsetReg;
271 std::tie(ScratchWaveOffsetReg, SPReg)
272 = getReservedPrivateSegmentWaveByteOffsetReg(ST, TII, TRI, MFI, MF);
274 // It's possible to have uses of only ScratchWaveOffsetReg without
275 // ScratchRsrcReg if it's only used for the initialization of flat_scratch,
276 // but the inverse is not true.
277 if (ScratchWaveOffsetReg == AMDGPU::NoRegister) {
278 assert(ScratchRsrcReg == AMDGPU::NoRegister);
282 // We need to insert initialization of the scratch resource descriptor.
283 unsigned PreloadedScratchWaveOffsetReg = TRI->getPreloadedValue(
284 MF, SIRegisterInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET);
286 unsigned PreloadedPrivateBufferReg = AMDGPU::NoRegister;
287 if (ST.isAmdCodeObjectV2(MF)) {
288 PreloadedPrivateBufferReg = TRI->getPreloadedValue(
289 MF, SIRegisterInfo::PRIVATE_SEGMENT_BUFFER);
292 bool OffsetRegUsed = MRI.isPhysRegUsed(ScratchWaveOffsetReg);
293 bool ResourceRegUsed = ScratchRsrcReg != AMDGPU::NoRegister &&
294 MRI.isPhysRegUsed(ScratchRsrcReg);
296 // We added live-ins during argument lowering, but since they were not used
297 // they were deleted. We're adding the uses now, so add them back.
299 assert(PreloadedScratchWaveOffsetReg != AMDGPU::NoRegister &&
300 "scratch wave offset input is required");
301 MRI.addLiveIn(PreloadedScratchWaveOffsetReg);
302 MBB.addLiveIn(PreloadedScratchWaveOffsetReg);
305 if (ResourceRegUsed && PreloadedPrivateBufferReg != AMDGPU::NoRegister) {
306 assert(ST.isAmdCodeObjectV2(MF) || ST.isMesaGfxShader(MF));
307 MRI.addLiveIn(PreloadedPrivateBufferReg);
308 MBB.addLiveIn(PreloadedPrivateBufferReg);
311 // Make the register selected live throughout the function.
312 for (MachineBasicBlock &OtherBB : MF) {
313 if (&OtherBB == &MBB)
317 OtherBB.addLiveIn(ScratchWaveOffsetReg);
320 OtherBB.addLiveIn(ScratchRsrcReg);
324 MachineBasicBlock::iterator I = MBB.begin();
326 // If we reserved the original input registers, we don't need to copy to the
327 // reserved registers.
329 bool CopyBuffer = ResourceRegUsed &&
330 PreloadedPrivateBufferReg != AMDGPU::NoRegister &&
331 ST.isAmdCodeObjectV2(MF) &&
332 ScratchRsrcReg != PreloadedPrivateBufferReg;
334 // This needs to be careful of the copying order to avoid overwriting one of
335 // the input registers before it's been copied to it's final
336 // destination. Usually the offset should be copied first.
337 bool CopyBufferFirst = TRI->isSubRegisterEq(PreloadedPrivateBufferReg,
338 ScratchWaveOffsetReg);
339 if (CopyBuffer && CopyBufferFirst) {
340 BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), ScratchRsrcReg)
341 .addReg(PreloadedPrivateBufferReg, RegState::Kill);
345 PreloadedScratchWaveOffsetReg != ScratchWaveOffsetReg) {
346 BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), ScratchWaveOffsetReg)
347 .addReg(PreloadedScratchWaveOffsetReg,
348 MRI.isPhysRegUsed(ScratchWaveOffsetReg) ? 0 : RegState::Kill);
351 if (CopyBuffer && !CopyBufferFirst) {
352 BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), ScratchRsrcReg)
353 .addReg(PreloadedPrivateBufferReg, RegState::Kill);
356 if (ResourceRegUsed && (ST.isMesaGfxShader(MF) || (PreloadedPrivateBufferReg == AMDGPU::NoRegister))) {
357 assert(!ST.isAmdCodeObjectV2(MF));
358 const MCInstrDesc &SMovB32 = TII->get(AMDGPU::S_MOV_B32);
360 unsigned Rsrc2 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub2);
361 unsigned Rsrc3 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub3);
363 // Use relocations to get the pointer, and setup the other bits manually.
364 uint64_t Rsrc23 = TII->getScratchRsrcWords23();
366 if (MFI->hasImplicitBufferPtr()) {
367 unsigned Rsrc01 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub0_sub1);
369 if (AMDGPU::isCompute(MF.getFunction()->getCallingConv())) {
370 const MCInstrDesc &Mov64 = TII->get(AMDGPU::S_MOV_B64);
372 BuildMI(MBB, I, DL, Mov64, Rsrc01)
373 .addReg(MFI->getImplicitBufferPtrUserSGPR())
374 .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
376 const MCInstrDesc &LoadDwordX2 = TII->get(AMDGPU::S_LOAD_DWORDX2_IMM);
379 PointerType::get(Type::getInt64Ty(MF.getFunction()->getContext()),
380 AMDGPUASI.CONSTANT_ADDRESS);
381 MachinePointerInfo PtrInfo(UndefValue::get(PtrTy));
382 auto MMO = MF.getMachineMemOperand(PtrInfo,
383 MachineMemOperand::MOLoad |
384 MachineMemOperand::MOInvariant |
385 MachineMemOperand::MODereferenceable,
387 BuildMI(MBB, I, DL, LoadDwordX2, Rsrc01)
388 .addReg(MFI->getImplicitBufferPtrUserSGPR())
392 .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
395 unsigned Rsrc0 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub0);
396 unsigned Rsrc1 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub1);
398 BuildMI(MBB, I, DL, SMovB32, Rsrc0)
399 .addExternalSymbol("SCRATCH_RSRC_DWORD0")
400 .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
402 BuildMI(MBB, I, DL, SMovB32, Rsrc1)
403 .addExternalSymbol("SCRATCH_RSRC_DWORD1")
404 .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
408 BuildMI(MBB, I, DL, SMovB32, Rsrc2)
409 .addImm(Rsrc23 & 0xffffffff)
410 .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
412 BuildMI(MBB, I, DL, SMovB32, Rsrc3)
413 .addImm(Rsrc23 >> 32)
414 .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
418 void SIFrameLowering::emitPrologue(MachineFunction &MF,
419 MachineBasicBlock &MBB) const {
420 const SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>();
421 if (FuncInfo->isEntryFunction()) {
422 emitEntryFunctionPrologue(MF, MBB);
426 const MachineFrameInfo &MFI = MF.getFrameInfo();
427 const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
428 const SIInstrInfo *TII = ST.getInstrInfo();
430 unsigned StackPtrReg = FuncInfo->getStackPtrOffsetReg();
431 unsigned FramePtrReg = FuncInfo->getFrameOffsetReg();
433 MachineBasicBlock::iterator MBBI = MBB.begin();
436 bool NeedFP = hasFP(MF);
438 // If we need a base pointer, set it up here. It's whatever the value of
439 // the stack pointer is at this point. Any variable size objects will be
440 // allocated after this, so we can still use the base pointer to reference
442 BuildMI(MBB, MBBI, DL, TII->get(AMDGPU::COPY), FramePtrReg)
444 .setMIFlag(MachineInstr::FrameSetup);
447 uint32_t NumBytes = MFI.getStackSize();
448 if (NumBytes != 0 && hasSP(MF)) {
449 BuildMI(MBB, MBBI, DL, TII->get(AMDGPU::S_ADD_U32), StackPtrReg)
451 .addImm(NumBytes * ST.getWavefrontSize())
452 .setMIFlag(MachineInstr::FrameSetup);
456 void SIFrameLowering::emitEpilogue(MachineFunction &MF,
457 MachineBasicBlock &MBB) const {
458 const SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>();
459 if (FuncInfo->isEntryFunction())
462 unsigned StackPtrReg = FuncInfo->getStackPtrOffsetReg();
463 if (StackPtrReg == AMDGPU::NoRegister)
466 const MachineFrameInfo &MFI = MF.getFrameInfo();
467 uint32_t NumBytes = MFI.getStackSize();
469 const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
470 const SIInstrInfo *TII = ST.getInstrInfo();
471 MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator();
474 // FIXME: Clarify distinction between no set SP and SP. For callee functions,
475 // it's really whether we need SP to be accurate or not.
477 if (NumBytes != 0 && hasSP(MF)) {
478 BuildMI(MBB, MBBI, DL, TII->get(AMDGPU::S_SUB_U32), StackPtrReg)
480 .addImm(NumBytes * ST.getWavefrontSize())
481 .setMIFlag(MachineInstr::FrameDestroy);
485 static bool allStackObjectsAreDead(const MachineFrameInfo &MFI) {
486 for (int I = MFI.getObjectIndexBegin(), E = MFI.getObjectIndexEnd();
488 if (!MFI.isDeadObjectIndex(I))
495 int SIFrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI,
496 unsigned &FrameReg) const {
497 const SIRegisterInfo *RI = MF.getSubtarget<SISubtarget>().getRegisterInfo();
499 FrameReg = RI->getFrameRegister(MF);
500 return MF.getFrameInfo().getObjectOffset(FI);
503 void SIFrameLowering::processFunctionBeforeFrameFinalized(
505 RegScavenger *RS) const {
506 MachineFrameInfo &MFI = MF.getFrameInfo();
508 if (!MFI.hasStackObjects())
511 const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
512 const SIInstrInfo *TII = ST.getInstrInfo();
513 const SIRegisterInfo &TRI = TII->getRegisterInfo();
514 SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>();
515 bool AllSGPRSpilledToVGPRs = false;
517 if (TRI.spillSGPRToVGPR() && FuncInfo->hasSpilledSGPRs()) {
518 AllSGPRSpilledToVGPRs = true;
520 // Process all SGPR spills before frame offsets are finalized. Ideally SGPRs
521 // are spilled to VGPRs, in which case we can eliminate the stack usage.
523 // XXX - This operates under the assumption that only other SGPR spills are
524 // users of the frame index. I'm not 100% sure this is correct. The
525 // StackColoring pass has a comment saying a future improvement would be to
526 // merging of allocas with spill slots, but for now according to
527 // MachineFrameInfo isSpillSlot can't alias any other object.
528 for (MachineBasicBlock &MBB : MF) {
529 MachineBasicBlock::iterator Next;
530 for (auto I = MBB.begin(), E = MBB.end(); I != E; I = Next) {
531 MachineInstr &MI = *I;
534 if (TII->isSGPRSpill(MI)) {
535 int FI = TII->getNamedOperand(MI, AMDGPU::OpName::addr)->getIndex();
536 if (FuncInfo->allocateSGPRSpillToVGPR(MF, FI)) {
537 bool Spilled = TRI.eliminateSGPRToVGPRSpillFrameIndex(MI, FI, RS);
539 assert(Spilled && "failed to spill SGPR to VGPR when allocated");
541 AllSGPRSpilledToVGPRs = false;
546 FuncInfo->removeSGPRToVGPRFrameIndices(MFI);
549 // FIXME: The other checks should be redundant with allStackObjectsAreDead,
550 // but currently hasNonSpillStackObjects is set only from source
551 // allocas. Stack temps produced from legalization are not counted currently.
552 if (FuncInfo->hasNonSpillStackObjects() || FuncInfo->hasSpilledVGPRs() ||
553 !AllSGPRSpilledToVGPRs || !allStackObjectsAreDead(MFI)) {
554 assert(RS && "RegScavenger required if spilling");
556 // We force this to be at offset 0 so no user object ever has 0 as an
557 // address, so we may use 0 as an invalid pointer value. This is because
558 // LLVM assumes 0 is an invalid pointer in address space 0. Because alloca
559 // is required to be address space 0, we are forced to accept this for
560 // now. Ideally we could have the stack in another address space with 0 as a
561 // valid pointer, and -1 as the null value.
563 // This will also waste additional space when user stack objects require > 4
566 // The main cost here is losing the offset for addressing modes. However
567 // this also ensures we shouldn't need a register for the offset when
568 // emergency scavenging.
569 int ScavengeFI = MFI.CreateFixedObject(
570 TRI.getSpillSize(AMDGPU::SGPR_32RegClass), 0, false);
571 RS->addScavengingFrameIndex(ScavengeFI);
575 void SIFrameLowering::emitDebuggerPrologue(MachineFunction &MF,
576 MachineBasicBlock &MBB) const {
577 const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
578 const SIInstrInfo *TII = ST.getInstrInfo();
579 const SIRegisterInfo *TRI = &TII->getRegisterInfo();
580 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
582 MachineBasicBlock::iterator I = MBB.begin();
585 // For each dimension:
586 for (unsigned i = 0; i < 3; ++i) {
587 // Get work group ID SGPR, and make it live-in again.
588 unsigned WorkGroupIDSGPR = MFI->getWorkGroupIDSGPR(i);
589 MF.getRegInfo().addLiveIn(WorkGroupIDSGPR);
590 MBB.addLiveIn(WorkGroupIDSGPR);
592 // Since SGPRs are spilled into VGPRs, copy work group ID SGPR to VGPR in
593 // order to spill it to scratch.
594 unsigned WorkGroupIDVGPR =
595 MF.getRegInfo().createVirtualRegister(&AMDGPU::VGPR_32RegClass);
596 BuildMI(MBB, I, DL, TII->get(AMDGPU::V_MOV_B32_e32), WorkGroupIDVGPR)
597 .addReg(WorkGroupIDSGPR);
599 // Spill work group ID.
600 int WorkGroupIDObjectIdx = MFI->getDebuggerWorkGroupIDStackObjectIndex(i);
601 TII->storeRegToStackSlot(MBB, I, WorkGroupIDVGPR, false,
602 WorkGroupIDObjectIdx, &AMDGPU::VGPR_32RegClass, TRI);
604 // Get work item ID VGPR, and make it live-in again.
605 unsigned WorkItemIDVGPR = MFI->getWorkItemIDVGPR(i);
606 MF.getRegInfo().addLiveIn(WorkItemIDVGPR);
607 MBB.addLiveIn(WorkItemIDVGPR);
609 // Spill work item ID.
610 int WorkItemIDObjectIdx = MFI->getDebuggerWorkItemIDStackObjectIndex(i);
611 TII->storeRegToStackSlot(MBB, I, WorkItemIDVGPR, false,
612 WorkItemIDObjectIdx, &AMDGPU::VGPR_32RegClass, TRI);
616 bool SIFrameLowering::hasFP(const MachineFunction &MF) const {
617 // All stack operations are relative to the frame offset SGPR.
618 // TODO: Still want to eliminate sometimes.
619 const MachineFrameInfo &MFI = MF.getFrameInfo();
621 // XXX - Is this only called after frame is finalized? Should be able to check
623 return MFI.hasStackObjects() && !allStackObjectsAreDead(MFI);
626 bool SIFrameLowering::hasSP(const MachineFunction &MF) const {
627 // All stack operations are relative to the frame offset SGPR.
628 const MachineFrameInfo &MFI = MF.getFrameInfo();
629 return MFI.hasCalls() || MFI.hasVarSizedObjects();