1 //===-- SIRegisterInfo.cpp - SI Register 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 the TargetRegisterInfo class.
13 //===----------------------------------------------------------------------===//
16 #include "SIRegisterInfo.h"
17 #include "AMDGPUSubtarget.h"
18 #include "SIInstrInfo.h"
19 #include "SIMachineFunctionInfo.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/RegisterScavenging.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/LLVMContext.h"
26 #include "llvm/Support/Debug.h"
29 SIRegisterInfo::SIRegisterInfo(const AMDGPUSubtarget &st)
30 : AMDGPURegisterInfo(st)
33 BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
34 BitVector Reserved(getNumRegs());
35 Reserved.set(AMDGPU::EXEC);
37 // EXEC_LO and EXEC_HI could be allocated and used as regular register,
38 // but this seems likely to result in bugs, so I'm marking them as reserved.
39 Reserved.set(AMDGPU::EXEC_LO);
40 Reserved.set(AMDGPU::EXEC_HI);
42 Reserved.set(AMDGPU::INDIRECT_BASE_ADDR);
43 Reserved.set(AMDGPU::FLAT_SCR);
44 Reserved.set(AMDGPU::FLAT_SCR_LO);
45 Reserved.set(AMDGPU::FLAT_SCR_HI);
47 // Reserve some VGPRs to use as temp registers in case we have to spill VGPRs
48 Reserved.set(AMDGPU::VGPR255);
49 Reserved.set(AMDGPU::VGPR254);
54 unsigned SIRegisterInfo::getRegPressureSetLimit(unsigned Idx) const {
56 // FIXME: We should adjust the max number of waves based on LDS size.
57 unsigned SGPRLimit = getNumSGPRsAllowed(ST.getMaxWavesPerCU());
58 unsigned VGPRLimit = getNumVGPRsAllowed(ST.getMaxWavesPerCU());
60 for (regclass_iterator I = regclass_begin(), E = regclass_end();
63 unsigned NumSubRegs = std::max((int)(*I)->getSize() / 4, 1);
66 if (isSGPRClass(*I)) {
67 Limit = SGPRLimit / NumSubRegs;
69 Limit = VGPRLimit / NumSubRegs;
72 const int *Sets = getRegClassPressureSets(*I);
74 for (unsigned i = 0; Sets[i] != -1; ++i) {
75 if (Sets[i] == (int)Idx)
82 bool SIRegisterInfo::requiresRegisterScavenging(const MachineFunction &Fn) const {
83 return Fn.getFrameInfo()->hasStackObjects();
86 static unsigned getNumSubRegsForSpillOp(unsigned Op) {
89 case AMDGPU::SI_SPILL_S512_SAVE:
90 case AMDGPU::SI_SPILL_S512_RESTORE:
91 case AMDGPU::SI_SPILL_V512_SAVE:
92 case AMDGPU::SI_SPILL_V512_RESTORE:
94 case AMDGPU::SI_SPILL_S256_SAVE:
95 case AMDGPU::SI_SPILL_S256_RESTORE:
96 case AMDGPU::SI_SPILL_V256_SAVE:
97 case AMDGPU::SI_SPILL_V256_RESTORE:
99 case AMDGPU::SI_SPILL_S128_SAVE:
100 case AMDGPU::SI_SPILL_S128_RESTORE:
101 case AMDGPU::SI_SPILL_V128_SAVE:
102 case AMDGPU::SI_SPILL_V128_RESTORE:
104 case AMDGPU::SI_SPILL_V96_SAVE:
105 case AMDGPU::SI_SPILL_V96_RESTORE:
107 case AMDGPU::SI_SPILL_S64_SAVE:
108 case AMDGPU::SI_SPILL_S64_RESTORE:
109 case AMDGPU::SI_SPILL_V64_SAVE:
110 case AMDGPU::SI_SPILL_V64_RESTORE:
112 case AMDGPU::SI_SPILL_S32_SAVE:
113 case AMDGPU::SI_SPILL_S32_RESTORE:
114 case AMDGPU::SI_SPILL_V32_SAVE:
115 case AMDGPU::SI_SPILL_V32_RESTORE:
117 default: llvm_unreachable("Invalid spill opcode");
121 void SIRegisterInfo::buildScratchLoadStore(MachineBasicBlock::iterator MI,
122 unsigned LoadStoreOp,
124 unsigned ScratchRsrcReg,
125 unsigned ScratchOffset,
127 RegScavenger *RS) const {
129 const SIInstrInfo *TII = static_cast<const SIInstrInfo*>(ST.getInstrInfo());
130 MachineBasicBlock *MBB = MI->getParent();
131 const MachineFunction *MF = MI->getParent()->getParent();
132 LLVMContext &Ctx = MF->getFunction()->getContext();
133 DebugLoc DL = MI->getDebugLoc();
134 bool IsLoad = TII->get(LoadStoreOp).mayLoad();
136 bool RanOutOfSGPRs = false;
137 unsigned SOffset = ScratchOffset;
139 unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
140 unsigned Size = NumSubRegs * 4;
142 if (!isUInt<12>(Offset + Size)) {
143 dbgs() << "Offset scavenge\n";
144 SOffset = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, MI, 0);
145 if (SOffset == AMDGPU::NoRegister) {
146 RanOutOfSGPRs = true;
147 SOffset = AMDGPU::SGPR0;
149 BuildMI(*MBB, MI, DL, TII->get(AMDGPU::S_ADD_U32), SOffset)
150 .addReg(ScratchOffset)
156 Ctx.emitError("Ran out of SGPRs for spilling VGPRS");
158 for (unsigned i = 0, e = NumSubRegs; i != e; ++i, Offset += 4) {
159 unsigned SubReg = NumSubRegs > 1 ?
160 getPhysRegSubReg(Value, &AMDGPU::VGPR_32RegClass, i) :
162 bool IsKill = (i == e - 1);
164 BuildMI(*MBB, MI, DL, TII->get(LoadStoreOp))
165 .addReg(SubReg, getDefRegState(IsLoad))
166 .addReg(ScratchRsrcReg, getKillRegState(IsKill))
172 .addReg(Value, RegState::Implicit | getDefRegState(IsLoad));
176 void SIRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator MI,
177 int SPAdj, unsigned FIOperandNum,
178 RegScavenger *RS) const {
179 MachineFunction *MF = MI->getParent()->getParent();
180 MachineBasicBlock *MBB = MI->getParent();
181 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
182 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
183 const SIInstrInfo *TII = static_cast<const SIInstrInfo*>(ST.getInstrInfo());
184 DebugLoc DL = MI->getDebugLoc();
186 MachineOperand &FIOp = MI->getOperand(FIOperandNum);
187 int Index = MI->getOperand(FIOperandNum).getIndex();
189 switch (MI->getOpcode()) {
190 // SGPR register spill
191 case AMDGPU::SI_SPILL_S512_SAVE:
192 case AMDGPU::SI_SPILL_S256_SAVE:
193 case AMDGPU::SI_SPILL_S128_SAVE:
194 case AMDGPU::SI_SPILL_S64_SAVE:
195 case AMDGPU::SI_SPILL_S32_SAVE: {
196 unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
198 for (unsigned i = 0, e = NumSubRegs; i < e; ++i) {
199 unsigned SubReg = getPhysRegSubReg(MI->getOperand(0).getReg(),
200 &AMDGPU::SGPR_32RegClass, i);
201 struct SIMachineFunctionInfo::SpilledReg Spill =
202 MFI->getSpilledReg(MF, Index, i);
204 if (Spill.VGPR == AMDGPU::NoRegister) {
205 LLVMContext &Ctx = MF->getFunction()->getContext();
206 Ctx.emitError("Ran out of VGPRs for spilling SGPR");
209 BuildMI(*MBB, MI, DL, TII->get(AMDGPU::V_WRITELANE_B32), Spill.VGPR)
214 MI->eraseFromParent();
218 // SGPR register restore
219 case AMDGPU::SI_SPILL_S512_RESTORE:
220 case AMDGPU::SI_SPILL_S256_RESTORE:
221 case AMDGPU::SI_SPILL_S128_RESTORE:
222 case AMDGPU::SI_SPILL_S64_RESTORE:
223 case AMDGPU::SI_SPILL_S32_RESTORE: {
224 unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
226 for (unsigned i = 0, e = NumSubRegs; i < e; ++i) {
227 unsigned SubReg = getPhysRegSubReg(MI->getOperand(0).getReg(),
228 &AMDGPU::SGPR_32RegClass, i);
229 bool isM0 = SubReg == AMDGPU::M0;
230 struct SIMachineFunctionInfo::SpilledReg Spill =
231 MFI->getSpilledReg(MF, Index, i);
233 if (Spill.VGPR == AMDGPU::NoRegister) {
234 LLVMContext &Ctx = MF->getFunction()->getContext();
235 Ctx.emitError("Ran out of VGPRs for spilling SGPR");
239 dbgs() << "Scavenge M0\n";
240 SubReg = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, MI, 0);
243 BuildMI(*MBB, MI, DL, TII->get(AMDGPU::V_READLANE_B32), SubReg)
246 .addReg(MI->getOperand(0).getReg(), RegState::ImplicitDefine);
248 BuildMI(*MBB, MI, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
252 TII->insertNOPs(MI, 3);
253 MI->eraseFromParent();
257 // VGPR register spill
258 case AMDGPU::SI_SPILL_V512_SAVE:
259 case AMDGPU::SI_SPILL_V256_SAVE:
260 case AMDGPU::SI_SPILL_V128_SAVE:
261 case AMDGPU::SI_SPILL_V96_SAVE:
262 case AMDGPU::SI_SPILL_V64_SAVE:
263 case AMDGPU::SI_SPILL_V32_SAVE:
264 buildScratchLoadStore(MI, AMDGPU::BUFFER_STORE_DWORD_OFFSET,
265 TII->getNamedOperand(*MI, AMDGPU::OpName::src)->getReg(),
266 TII->getNamedOperand(*MI, AMDGPU::OpName::scratch_rsrc)->getReg(),
267 TII->getNamedOperand(*MI, AMDGPU::OpName::scratch_offset)->getReg(),
268 FrameInfo->getObjectOffset(Index), RS);
269 MI->eraseFromParent();
271 case AMDGPU::SI_SPILL_V32_RESTORE:
272 case AMDGPU::SI_SPILL_V64_RESTORE:
273 case AMDGPU::SI_SPILL_V128_RESTORE:
274 case AMDGPU::SI_SPILL_V256_RESTORE:
275 case AMDGPU::SI_SPILL_V512_RESTORE: {
276 buildScratchLoadStore(MI, AMDGPU::BUFFER_LOAD_DWORD_OFFSET,
277 TII->getNamedOperand(*MI, AMDGPU::OpName::dst)->getReg(),
278 TII->getNamedOperand(*MI, AMDGPU::OpName::scratch_rsrc)->getReg(),
279 TII->getNamedOperand(*MI, AMDGPU::OpName::scratch_offset)->getReg(),
280 FrameInfo->getObjectOffset(Index), RS);
281 MI->eraseFromParent();
286 int64_t Offset = FrameInfo->getObjectOffset(Index);
287 FIOp.ChangeToImmediate(Offset);
288 if (!TII->isImmOperandLegal(MI, FIOperandNum, FIOp)) {
289 unsigned TmpReg = RS->scavengeRegister(&AMDGPU::VGPR_32RegClass, MI, SPAdj);
290 BuildMI(*MBB, MI, MI->getDebugLoc(),
291 TII->get(AMDGPU::V_MOV_B32_e32), TmpReg)
293 FIOp.ChangeToRegister(TmpReg, false, false, true);
299 const TargetRegisterClass * SIRegisterInfo::getCFGStructurizerRegClass(
301 switch(VT.SimpleTy) {
303 case MVT::i32: return &AMDGPU::VGPR_32RegClass;
307 unsigned SIRegisterInfo::getHWRegIndex(unsigned Reg) const {
308 return getEncodingValue(Reg) & 0xff;
311 const TargetRegisterClass *SIRegisterInfo::getPhysRegClass(unsigned Reg) const {
312 assert(!TargetRegisterInfo::isVirtualRegister(Reg));
314 static const TargetRegisterClass *BaseClasses[] = {
315 &AMDGPU::VGPR_32RegClass,
316 &AMDGPU::SReg_32RegClass,
317 &AMDGPU::VReg_64RegClass,
318 &AMDGPU::SReg_64RegClass,
319 &AMDGPU::VReg_96RegClass,
320 &AMDGPU::VReg_128RegClass,
321 &AMDGPU::SReg_128RegClass,
322 &AMDGPU::VReg_256RegClass,
323 &AMDGPU::SReg_256RegClass,
324 &AMDGPU::VReg_512RegClass
327 for (const TargetRegisterClass *BaseClass : BaseClasses) {
328 if (BaseClass->contains(Reg)) {
335 bool SIRegisterInfo::hasVGPRs(const TargetRegisterClass *RC) const {
336 return getCommonSubClass(&AMDGPU::VGPR_32RegClass, RC) ||
337 getCommonSubClass(&AMDGPU::VReg_64RegClass, RC) ||
338 getCommonSubClass(&AMDGPU::VReg_96RegClass, RC) ||
339 getCommonSubClass(&AMDGPU::VReg_128RegClass, RC) ||
340 getCommonSubClass(&AMDGPU::VReg_256RegClass, RC) ||
341 getCommonSubClass(&AMDGPU::VReg_512RegClass, RC);
344 const TargetRegisterClass *SIRegisterInfo::getEquivalentVGPRClass(
345 const TargetRegisterClass *SRC) const {
348 } else if (SRC == &AMDGPU::SCCRegRegClass) {
349 return &AMDGPU::VCCRegRegClass;
350 } else if (getCommonSubClass(SRC, &AMDGPU::SGPR_32RegClass)) {
351 return &AMDGPU::VGPR_32RegClass;
352 } else if (getCommonSubClass(SRC, &AMDGPU::SGPR_64RegClass)) {
353 return &AMDGPU::VReg_64RegClass;
354 } else if (getCommonSubClass(SRC, &AMDGPU::SReg_128RegClass)) {
355 return &AMDGPU::VReg_128RegClass;
356 } else if (getCommonSubClass(SRC, &AMDGPU::SReg_256RegClass)) {
357 return &AMDGPU::VReg_256RegClass;
358 } else if (getCommonSubClass(SRC, &AMDGPU::SReg_512RegClass)) {
359 return &AMDGPU::VReg_512RegClass;
364 const TargetRegisterClass *SIRegisterInfo::getSubRegClass(
365 const TargetRegisterClass *RC, unsigned SubIdx) const {
366 if (SubIdx == AMDGPU::NoSubRegister)
369 // If this register has a sub-register, we can safely assume it is a 32-bit
370 // register, because all of SI's sub-registers are 32-bit.
371 if (isSGPRClass(RC)) {
372 return &AMDGPU::SGPR_32RegClass;
374 return &AMDGPU::VGPR_32RegClass;
378 unsigned SIRegisterInfo::getPhysRegSubReg(unsigned Reg,
379 const TargetRegisterClass *SubRC,
380 unsigned Channel) const {
385 case 0: return AMDGPU::VCC_LO;
386 case 1: return AMDGPU::VCC_HI;
387 default: llvm_unreachable("Invalid SubIdx for VCC");
390 case AMDGPU::FLAT_SCR:
393 return AMDGPU::FLAT_SCR_LO;
395 return AMDGPU::FLAT_SCR_HI;
397 llvm_unreachable("Invalid SubIdx for FLAT_SCR");
404 return AMDGPU::EXEC_LO;
406 return AMDGPU::EXEC_HI;
408 llvm_unreachable("Invalid SubIdx for EXEC");
413 const TargetRegisterClass *RC = getPhysRegClass(Reg);
414 // 32-bit registers don't have sub-registers, so we can just return the
415 // Reg. We need to have this check here, because the calculation below
416 // using getHWRegIndex() will fail with special 32-bit registers like
417 // VCC_LO, VCC_HI, EXEC_LO, EXEC_HI and M0.
418 if (RC->getSize() == 4) {
419 assert(Channel == 0);
423 unsigned Index = getHWRegIndex(Reg);
424 return SubRC->getRegister(Index + Channel);
427 bool SIRegisterInfo::opCanUseLiteralConstant(unsigned OpType) const {
428 return OpType == AMDGPU::OPERAND_REG_IMM32;
431 bool SIRegisterInfo::opCanUseInlineConstant(unsigned OpType) const {
432 if (opCanUseLiteralConstant(OpType))
435 return OpType == AMDGPU::OPERAND_REG_INLINE_C;
438 unsigned SIRegisterInfo::getPreloadedValue(const MachineFunction &MF,
439 enum PreloadedValue Value) const {
441 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
443 case SIRegisterInfo::TGID_X:
444 return AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 0);
445 case SIRegisterInfo::TGID_Y:
446 return AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 1);
447 case SIRegisterInfo::TGID_Z:
448 return AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 2);
449 case SIRegisterInfo::SCRATCH_WAVE_OFFSET:
450 if (MFI->getShaderType() != ShaderType::COMPUTE)
451 return MFI->ScratchOffsetReg;
452 return AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 4);
453 case SIRegisterInfo::SCRATCH_PTR:
454 return AMDGPU::SGPR2_SGPR3;
455 case SIRegisterInfo::INPUT_PTR:
456 return AMDGPU::SGPR0_SGPR1;
457 case SIRegisterInfo::TIDIG_X:
458 return AMDGPU::VGPR0;
459 case SIRegisterInfo::TIDIG_Y:
460 return AMDGPU::VGPR1;
461 case SIRegisterInfo::TIDIG_Z:
462 return AMDGPU::VGPR2;
464 llvm_unreachable("unexpected preloaded value type");
467 /// \brief Returns a register that is not used at any point in the function.
468 /// If all registers are used, then this function will return
469 // AMDGPU::NoRegister.
470 unsigned SIRegisterInfo::findUnusedRegister(const MachineRegisterInfo &MRI,
471 const TargetRegisterClass *RC) const {
473 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
475 if (!MRI.isPhysRegUsed(*I))
478 return AMDGPU::NoRegister;
481 unsigned SIRegisterInfo::getNumVGPRsAllowed(unsigned WaveCount) const {
496 unsigned SIRegisterInfo::getNumSGPRsAllowed(unsigned WaveCount) const {