1 //===-- SIRegisterInfo.cpp - SI Register Information ---------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// \brief SI implementation of the TargetRegisterInfo class.
12 //
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 using namespace llvm;
28 SIRegisterInfo::SIRegisterInfo(const AMDGPUSubtarget &st)
29 : AMDGPURegisterInfo(st)
30 { }
32 BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
33 BitVector Reserved(getNumRegs());
34 Reserved.set(AMDGPU::EXEC);
36 // EXEC_LO and EXEC_HI could be allocated and used as regular register,
37 // but this seems likely to result in bugs, so I'm marking them as reserved.
38 Reserved.set(AMDGPU::EXEC_LO);
39 Reserved.set(AMDGPU::EXEC_HI);
41 Reserved.set(AMDGPU::INDIRECT_BASE_ADDR);
42 Reserved.set(AMDGPU::FLAT_SCR);
44 // Reserve some VGPRs to use as temp registers in case we have to spill VGPRs
45 Reserved.set(AMDGPU::VGPR255);
46 Reserved.set(AMDGPU::VGPR254);
48 return Reserved;
49 }
51 unsigned SIRegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC,
52 MachineFunction &MF) const {
53 return RC->getNumRegs();
54 }
56 bool SIRegisterInfo::requiresRegisterScavenging(const MachineFunction &Fn) const {
57 return Fn.getFrameInfo()->hasStackObjects();
58 }
60 static unsigned getNumSubRegsForSpillOp(unsigned Op) {
62 switch (Op) {
63 case AMDGPU::SI_SPILL_S512_SAVE:
64 case AMDGPU::SI_SPILL_S512_RESTORE:
65 case AMDGPU::SI_SPILL_V512_SAVE:
66 case AMDGPU::SI_SPILL_V512_RESTORE:
67 return 16;
68 case AMDGPU::SI_SPILL_S256_SAVE:
69 case AMDGPU::SI_SPILL_S256_RESTORE:
70 case AMDGPU::SI_SPILL_V256_SAVE:
71 case AMDGPU::SI_SPILL_V256_RESTORE:
72 return 8;
73 case AMDGPU::SI_SPILL_S128_SAVE:
74 case AMDGPU::SI_SPILL_S128_RESTORE:
75 case AMDGPU::SI_SPILL_V128_SAVE:
76 case AMDGPU::SI_SPILL_V128_RESTORE:
77 return 4;
78 case AMDGPU::SI_SPILL_V96_SAVE:
79 case AMDGPU::SI_SPILL_V96_RESTORE:
80 return 3;
81 case AMDGPU::SI_SPILL_S64_SAVE:
82 case AMDGPU::SI_SPILL_S64_RESTORE:
83 case AMDGPU::SI_SPILL_V64_SAVE:
84 case AMDGPU::SI_SPILL_V64_RESTORE:
85 return 2;
86 case AMDGPU::SI_SPILL_S32_SAVE:
87 case AMDGPU::SI_SPILL_S32_RESTORE:
88 case AMDGPU::SI_SPILL_V32_SAVE:
89 case AMDGPU::SI_SPILL_V32_RESTORE:
90 return 1;
91 default: llvm_unreachable("Invalid spill opcode");
92 }
93 }
95 void SIRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator MI,
96 int SPAdj, unsigned FIOperandNum,
97 RegScavenger *RS) const {
98 MachineFunction *MF = MI->getParent()->getParent();
99 MachineBasicBlock *MBB = MI->getParent();
100 SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
101 MachineFrameInfo *FrameInfo = MF->getFrameInfo();
102 const SIInstrInfo *TII = static_cast<const SIInstrInfo*>(ST.getInstrInfo());
103 DebugLoc DL = MI->getDebugLoc();
105 MachineOperand &FIOp = MI->getOperand(FIOperandNum);
106 int Index = MI->getOperand(FIOperandNum).getIndex();
108 switch (MI->getOpcode()) {
109 // SGPR register spill
110 case AMDGPU::SI_SPILL_S512_SAVE:
111 case AMDGPU::SI_SPILL_S256_SAVE:
112 case AMDGPU::SI_SPILL_S128_SAVE:
113 case AMDGPU::SI_SPILL_S64_SAVE:
114 case AMDGPU::SI_SPILL_S32_SAVE: {
115 unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
117 for (unsigned i = 0, e = NumSubRegs; i < e; ++i) {
118 unsigned SubReg = getPhysRegSubReg(MI->getOperand(0).getReg(),
119 &AMDGPU::SGPR_32RegClass, i);
120 struct SIMachineFunctionInfo::SpilledReg Spill =
121 MFI->getSpilledReg(MF, Index, i);
123 if (Spill.VGPR == AMDGPU::NoRegister) {
124 LLVMContext &Ctx = MF->getFunction()->getContext();
125 Ctx.emitError("Ran out of VGPRs for spilling SGPR");
126 }
128 BuildMI(*MBB, MI, DL, TII->get(AMDGPU::V_WRITELANE_B32), Spill.VGPR)
129 .addReg(SubReg)
130 .addImm(Spill.Lane);
132 }
133 MI->eraseFromParent();
134 break;
135 }
137 // SGPR register restore
138 case AMDGPU::SI_SPILL_S512_RESTORE:
139 case AMDGPU::SI_SPILL_S256_RESTORE:
140 case AMDGPU::SI_SPILL_S128_RESTORE:
141 case AMDGPU::SI_SPILL_S64_RESTORE:
142 case AMDGPU::SI_SPILL_S32_RESTORE: {
143 unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
145 for (unsigned i = 0, e = NumSubRegs; i < e; ++i) {
146 unsigned SubReg = getPhysRegSubReg(MI->getOperand(0).getReg(),
147 &AMDGPU::SGPR_32RegClass, i);
148 struct SIMachineFunctionInfo::SpilledReg Spill =
149 MFI->getSpilledReg(MF, Index, i);
151 if (Spill.VGPR == AMDGPU::NoRegister) {
152 LLVMContext &Ctx = MF->getFunction()->getContext();
153 Ctx.emitError("Ran out of VGPRs for spilling SGPR");
154 }
156 BuildMI(*MBB, MI, DL, TII->get(AMDGPU::V_READLANE_B32), SubReg)
157 .addReg(Spill.VGPR)
158 .addImm(Spill.Lane);
160 }
161 TII->insertNOPs(MI, 3);
162 MI->eraseFromParent();
163 break;
164 }
166 // VGPR register spill
167 case AMDGPU::SI_SPILL_V512_SAVE:
168 case AMDGPU::SI_SPILL_V256_SAVE:
169 case AMDGPU::SI_SPILL_V128_SAVE:
170 case AMDGPU::SI_SPILL_V96_SAVE:
171 case AMDGPU::SI_SPILL_V64_SAVE:
172 case AMDGPU::SI_SPILL_V32_SAVE: {
173 unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
174 unsigned SrcReg = MI->getOperand(0).getReg();
175 int64_t Offset = FrameInfo->getObjectOffset(Index);
176 unsigned Size = NumSubRegs * 4;
177 unsigned TmpReg = RS->scavengeRegister(&AMDGPU::VGPR_32RegClass, MI, 0);
179 for (unsigned i = 0, e = NumSubRegs; i != e; ++i) {
180 unsigned SubReg = NumSubRegs > 1 ?
181 getPhysRegSubReg(SrcReg, &AMDGPU::VGPR_32RegClass, i) :
182 SrcReg;
183 Offset += (i * 4);
184 MFI->LDSWaveSpillSize = std::max((unsigned)Offset + 4, (unsigned)MFI->LDSWaveSpillSize);
186 unsigned AddrReg = TII->calculateLDSSpillAddress(*MBB, MI, RS, TmpReg,
187 Offset, Size);
189 if (AddrReg == AMDGPU::NoRegister) {
190 LLVMContext &Ctx = MF->getFunction()->getContext();
191 Ctx.emitError("Ran out of VGPRs for spilling VGPRS");
192 AddrReg = AMDGPU::VGPR0;
193 }
195 // Store the value in LDS
196 BuildMI(*MBB, MI, DL, TII->get(AMDGPU::DS_WRITE_B32))
197 .addImm(0) // gds
198 .addReg(AddrReg, RegState::Kill) // addr
199 .addReg(SubReg) // data0
200 .addImm(0); // offset
201 }
203 MI->eraseFromParent();
204 break;
205 }
206 case AMDGPU::SI_SPILL_V32_RESTORE:
207 case AMDGPU::SI_SPILL_V64_RESTORE:
208 case AMDGPU::SI_SPILL_V128_RESTORE:
209 case AMDGPU::SI_SPILL_V256_RESTORE:
210 case AMDGPU::SI_SPILL_V512_RESTORE: {
211 unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
212 unsigned DstReg = MI->getOperand(0).getReg();
213 int64_t Offset = FrameInfo->getObjectOffset(Index);
214 unsigned Size = NumSubRegs * 4;
215 unsigned TmpReg = RS->scavengeRegister(&AMDGPU::VGPR_32RegClass, MI, 0);
217 // FIXME: We could use DS_READ_B64 here to optimize for larger registers.
218 for (unsigned i = 0, e = NumSubRegs; i != e; ++i) {
219 unsigned SubReg = NumSubRegs > 1 ?
220 getPhysRegSubReg(DstReg, &AMDGPU::VGPR_32RegClass, i) :
221 DstReg;
223 Offset += (i * 4);
224 unsigned AddrReg = TII->calculateLDSSpillAddress(*MBB, MI, RS, TmpReg,
225 Offset, Size);
226 if (AddrReg == AMDGPU::NoRegister) {
227 LLVMContext &Ctx = MF->getFunction()->getContext();
228 Ctx.emitError("Ran out of VGPRs for spilling VGPRs");
229 AddrReg = AMDGPU::VGPR0;
230 }
232 BuildMI(*MBB, MI, DL, TII->get(AMDGPU::DS_READ_B32), SubReg)
233 .addImm(0) // gds
234 .addReg(AddrReg, RegState::Kill) // addr
235 .addImm(0); //offset
236 }
237 MI->eraseFromParent();
238 break;
239 }
241 default: {
242 int64_t Offset = FrameInfo->getObjectOffset(Index);
243 FIOp.ChangeToImmediate(Offset);
244 if (!TII->isImmOperandLegal(MI, FIOperandNum, FIOp)) {
245 unsigned TmpReg = RS->scavengeRegister(&AMDGPU::VReg_32RegClass, MI, SPAdj);
246 BuildMI(*MBB, MI, MI->getDebugLoc(),
247 TII->get(AMDGPU::V_MOV_B32_e32), TmpReg)
248 .addImm(Offset);
249 FIOp.ChangeToRegister(TmpReg, false);
250 }
251 }
252 }
253 }
255 const TargetRegisterClass * SIRegisterInfo::getCFGStructurizerRegClass(
256 MVT VT) const {
257 switch(VT.SimpleTy) {
258 default:
259 case MVT::i32: return &AMDGPU::VReg_32RegClass;
260 }
261 }
263 unsigned SIRegisterInfo::getHWRegIndex(unsigned Reg) const {
264 return getEncodingValue(Reg) & 0xff;
265 }
267 const TargetRegisterClass *SIRegisterInfo::getPhysRegClass(unsigned Reg) const {
268 assert(!TargetRegisterInfo::isVirtualRegister(Reg));
270 static const TargetRegisterClass *BaseClasses[] = {
271 &AMDGPU::VReg_32RegClass,
272 &AMDGPU::SReg_32RegClass,
273 &AMDGPU::VReg_64RegClass,
274 &AMDGPU::SReg_64RegClass,
275 &AMDGPU::VReg_96RegClass,
276 &AMDGPU::VReg_128RegClass,
277 &AMDGPU::SReg_128RegClass,
278 &AMDGPU::VReg_256RegClass,
279 &AMDGPU::SReg_256RegClass,
280 &AMDGPU::VReg_512RegClass
281 };
283 for (const TargetRegisterClass *BaseClass : BaseClasses) {
284 if (BaseClass->contains(Reg)) {
285 return BaseClass;
286 }
287 }
288 return nullptr;
289 }
291 bool SIRegisterInfo::hasVGPRs(const TargetRegisterClass *RC) const {
292 return getCommonSubClass(&AMDGPU::VReg_32RegClass, RC) ||
293 getCommonSubClass(&AMDGPU::VReg_64RegClass, RC) ||
294 getCommonSubClass(&AMDGPU::VReg_96RegClass, RC) ||
295 getCommonSubClass(&AMDGPU::VReg_128RegClass, RC) ||
296 getCommonSubClass(&AMDGPU::VReg_256RegClass, RC) ||
297 getCommonSubClass(&AMDGPU::VReg_512RegClass, RC);
298 }
300 const TargetRegisterClass *SIRegisterInfo::getEquivalentVGPRClass(
301 const TargetRegisterClass *SRC) const {
302 if (hasVGPRs(SRC)) {
303 return SRC;
304 } else if (SRC == &AMDGPU::SCCRegRegClass) {
305 return &AMDGPU::VCCRegRegClass;
306 } else if (getCommonSubClass(SRC, &AMDGPU::SGPR_32RegClass)) {
307 return &AMDGPU::VReg_32RegClass;
308 } else if (getCommonSubClass(SRC, &AMDGPU::SGPR_64RegClass)) {
309 return &AMDGPU::VReg_64RegClass;
310 } else if (getCommonSubClass(SRC, &AMDGPU::SReg_128RegClass)) {
311 return &AMDGPU::VReg_128RegClass;
312 } else if (getCommonSubClass(SRC, &AMDGPU::SReg_256RegClass)) {
313 return &AMDGPU::VReg_256RegClass;
314 } else if (getCommonSubClass(SRC, &AMDGPU::SReg_512RegClass)) {
315 return &AMDGPU::VReg_512RegClass;
316 }
317 return nullptr;
318 }
320 const TargetRegisterClass *SIRegisterInfo::getSubRegClass(
321 const TargetRegisterClass *RC, unsigned SubIdx) const {
322 if (SubIdx == AMDGPU::NoSubRegister)
323 return RC;
325 // If this register has a sub-register, we can safely assume it is a 32-bit
326 // register, because all of SI's sub-registers are 32-bit.
327 if (isSGPRClass(RC)) {
328 return &AMDGPU::SGPR_32RegClass;
329 } else {
330 return &AMDGPU::VGPR_32RegClass;
331 }
332 }
334 unsigned SIRegisterInfo::getPhysRegSubReg(unsigned Reg,
335 const TargetRegisterClass *SubRC,
336 unsigned Channel) const {
338 switch (Reg) {
339 case AMDGPU::VCC:
340 switch(Channel) {
341 case 0: return AMDGPU::VCC_LO;
342 case 1: return AMDGPU::VCC_HI;
343 default: llvm_unreachable("Invalid SubIdx for VCC");
344 }
346 case AMDGPU::FLAT_SCR:
347 switch (Channel) {
348 case 0:
349 return AMDGPU::FLAT_SCR_LO;
350 case 1:
351 return AMDGPU::FLAT_SCR_HI;
352 default:
353 llvm_unreachable("Invalid SubIdx for FLAT_SCR");
354 }
355 break;
357 case AMDGPU::EXEC:
358 switch (Channel) {
359 case 0:
360 return AMDGPU::EXEC_LO;
361 case 1:
362 return AMDGPU::EXEC_HI;
363 default:
364 llvm_unreachable("Invalid SubIdx for EXEC");
365 }
366 break;
367 }
369 const TargetRegisterClass *RC = getPhysRegClass(Reg);
370 // 32-bit registers don't have sub-registers, so we can just return the
371 // Reg. We need to have this check here, because the calculation below
372 // using getHWRegIndex() will fail with special 32-bit registers like
373 // VCC_LO, VCC_HI, EXEC_LO, EXEC_HI and M0.
374 if (RC->getSize() == 4) {
375 assert(Channel == 0);
376 return Reg;
377 }
379 unsigned Index = getHWRegIndex(Reg);
380 return SubRC->getRegister(Index + Channel);
381 }
383 bool SIRegisterInfo::regClassCanUseLiteralConstant(int RCID) const {
384 switch (RCID) {
385 default: return false;
386 case AMDGPU::SSrc_32RegClassID:
387 case AMDGPU::SSrc_64RegClassID:
388 case AMDGPU::VSrc_32RegClassID:
389 case AMDGPU::VSrc_64RegClassID:
390 return true;
391 }
392 }
394 bool SIRegisterInfo::regClassCanUseLiteralConstant(
395 const TargetRegisterClass *RC) const {
396 return regClassCanUseLiteralConstant(RC->getID());
397 }
399 bool SIRegisterInfo::regClassCanUseInlineConstant(int RCID) const {
400 if (regClassCanUseLiteralConstant(RCID))
401 return true;
403 switch (RCID) {
404 default: return false;
405 case AMDGPU::VCSrc_32RegClassID:
406 case AMDGPU::VCSrc_64RegClassID:
407 return true;
408 }
409 }
411 bool SIRegisterInfo::regClassCanUseInlineConstant(
412 const TargetRegisterClass *RC) const {
413 return regClassCanUseInlineConstant(RC->getID());
414 }
417 unsigned SIRegisterInfo::getPreloadedValue(const MachineFunction &MF,
418 enum PreloadedValue Value) const {
420 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
421 switch (Value) {
422 case SIRegisterInfo::TGID_X:
423 return AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 0);
424 case SIRegisterInfo::TGID_Y:
425 return AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 1);
426 case SIRegisterInfo::TGID_Z:
427 return AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 2);
428 case SIRegisterInfo::SCRATCH_WAVE_OFFSET:
429 return AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 4);
430 case SIRegisterInfo::SCRATCH_PTR:
431 return AMDGPU::SGPR2_SGPR3;
432 case SIRegisterInfo::INPUT_PTR:
433 return AMDGPU::SGPR0_SGPR1;
434 case SIRegisterInfo::TIDIG_X:
435 return AMDGPU::VGPR0;
436 case SIRegisterInfo::TIDIG_Y:
437 return AMDGPU::VGPR1;
438 case SIRegisterInfo::TIDIG_Z:
439 return AMDGPU::VGPR2;
440 }
441 llvm_unreachable("unexpected preloaded value type");
442 }
444 /// \brief Returns a register that is not used at any point in the function.
445 /// If all registers are used, then this function will return
446 // AMDGPU::NoRegister.
447 unsigned SIRegisterInfo::findUnusedVGPR(const MachineRegisterInfo &MRI) const {
449 const TargetRegisterClass *RC = &AMDGPU::VGPR_32RegClass;
451 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
452 I != E; ++I) {
453 if (!MRI.isPhysRegUsed(*I))
454 return *I;
455 }
456 return AMDGPU::NoRegister;
457 }