X86RegisterInfo.cpp revision f6a9988ceab0ca660fa4f4e89d8d683f487118eb
1//===- X86RegisterInfo.cpp - X86 Register Information -----------*- C++ -*-===// 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// This file contains the X86 implementation of the TargetRegisterInfo class. 11// This file is responsible for the frame pointer elimination optimization 12// on X86. 13// 14//===----------------------------------------------------------------------===// 15 16#include "X86.h" 17#include "X86RegisterInfo.h" 18#include "X86InstrBuilder.h" 19#include "X86MachineFunctionInfo.h" 20#include "X86Subtarget.h" 21#include "X86TargetMachine.h" 22#include "llvm/Constants.h" 23#include "llvm/Function.h" 24#include "llvm/Type.h" 25#include "llvm/CodeGen/ValueTypes.h" 26#include "llvm/CodeGen/MachineInstrBuilder.h" 27#include "llvm/CodeGen/MachineFunction.h" 28#include "llvm/CodeGen/MachineFunctionPass.h" 29#include "llvm/CodeGen/MachineFrameInfo.h" 30#include "llvm/CodeGen/MachineLocation.h" 31#include "llvm/CodeGen/MachineModuleInfo.h" 32#include "llvm/CodeGen/MachineRegisterInfo.h" 33#include "llvm/Target/TargetAsmInfo.h" 34#include "llvm/Target/TargetFrameInfo.h" 35#include "llvm/Target/TargetInstrInfo.h" 36#include "llvm/Target/TargetMachine.h" 37#include "llvm/Target/TargetOptions.h" 38#include "llvm/ADT/BitVector.h" 39#include "llvm/ADT/STLExtras.h" 40#include "llvm/Support/Compiler.h" 41using namespace llvm; 42 43X86RegisterInfo::X86RegisterInfo(X86TargetMachine &tm, 44 const TargetInstrInfo &tii) 45 : X86GenRegisterInfo(tm.getSubtarget<X86Subtarget>().is64Bit() ? 46 X86::ADJCALLSTACKDOWN64 : 47 X86::ADJCALLSTACKDOWN32, 48 tm.getSubtarget<X86Subtarget>().is64Bit() ? 49 X86::ADJCALLSTACKUP64 : 50 X86::ADJCALLSTACKUP32), 51 TM(tm), TII(tii) { 52 // Cache some information. 53 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>(); 54 Is64Bit = Subtarget->is64Bit(); 55 IsWin64 = Subtarget->isTargetWin64(); 56 StackAlign = TM.getFrameInfo()->getStackAlignment(); 57 if (Is64Bit) { 58 SlotSize = 8; 59 StackPtr = X86::RSP; 60 FramePtr = X86::RBP; 61 } else { 62 SlotSize = 4; 63 StackPtr = X86::ESP; 64 FramePtr = X86::EBP; 65 } 66} 67 68// getDwarfRegNum - This function maps LLVM register identifiers to the 69// Dwarf specific numbering, used in debug info and exception tables. 70 71int X86RegisterInfo::getDwarfRegNum(unsigned RegNo, bool isEH) const { 72 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>(); 73 unsigned Flavour = DWARFFlavour::X86_64; 74 if (!Subtarget->is64Bit()) { 75 if (Subtarget->isTargetDarwin()) { 76 if (isEH) 77 Flavour = DWARFFlavour::X86_32_DarwinEH; 78 else 79 Flavour = DWARFFlavour::X86_32_Generic; 80 } else if (Subtarget->isTargetCygMing()) { 81 // Unsupported by now, just quick fallback 82 Flavour = DWARFFlavour::X86_32_Generic; 83 } else { 84 Flavour = DWARFFlavour::X86_32_Generic; 85 } 86 } 87 88 return X86GenRegisterInfo::getDwarfRegNumFull(RegNo, Flavour); 89} 90 91// getX86RegNum - This function maps LLVM register identifiers to their X86 92// specific numbering, which is used in various places encoding instructions. 93// 94unsigned X86RegisterInfo::getX86RegNum(unsigned RegNo) { 95 switch(RegNo) { 96 case X86::RAX: case X86::EAX: case X86::AX: case X86::AL: return N86::EAX; 97 case X86::RCX: case X86::ECX: case X86::CX: case X86::CL: return N86::ECX; 98 case X86::RDX: case X86::EDX: case X86::DX: case X86::DL: return N86::EDX; 99 case X86::RBX: case X86::EBX: case X86::BX: case X86::BL: return N86::EBX; 100 case X86::RSP: case X86::ESP: case X86::SP: case X86::SPL: case X86::AH: 101 return N86::ESP; 102 case X86::RBP: case X86::EBP: case X86::BP: case X86::BPL: case X86::CH: 103 return N86::EBP; 104 case X86::RSI: case X86::ESI: case X86::SI: case X86::SIL: case X86::DH: 105 return N86::ESI; 106 case X86::RDI: case X86::EDI: case X86::DI: case X86::DIL: case X86::BH: 107 return N86::EDI; 108 109 case X86::R8: case X86::R8D: case X86::R8W: case X86::R8B: 110 return N86::EAX; 111 case X86::R9: case X86::R9D: case X86::R9W: case X86::R9B: 112 return N86::ECX; 113 case X86::R10: case X86::R10D: case X86::R10W: case X86::R10B: 114 return N86::EDX; 115 case X86::R11: case X86::R11D: case X86::R11W: case X86::R11B: 116 return N86::EBX; 117 case X86::R12: case X86::R12D: case X86::R12W: case X86::R12B: 118 return N86::ESP; 119 case X86::R13: case X86::R13D: case X86::R13W: case X86::R13B: 120 return N86::EBP; 121 case X86::R14: case X86::R14D: case X86::R14W: case X86::R14B: 122 return N86::ESI; 123 case X86::R15: case X86::R15D: case X86::R15W: case X86::R15B: 124 return N86::EDI; 125 126 case X86::ST0: case X86::ST1: case X86::ST2: case X86::ST3: 127 case X86::ST4: case X86::ST5: case X86::ST6: case X86::ST7: 128 return RegNo-X86::ST0; 129 130 case X86::XMM0: case X86::XMM8: case X86::MM0: 131 return 0; 132 case X86::XMM1: case X86::XMM9: case X86::MM1: 133 return 1; 134 case X86::XMM2: case X86::XMM10: case X86::MM2: 135 return 2; 136 case X86::XMM3: case X86::XMM11: case X86::MM3: 137 return 3; 138 case X86::XMM4: case X86::XMM12: case X86::MM4: 139 return 4; 140 case X86::XMM5: case X86::XMM13: case X86::MM5: 141 return 5; 142 case X86::XMM6: case X86::XMM14: case X86::MM6: 143 return 6; 144 case X86::XMM7: case X86::XMM15: case X86::MM7: 145 return 7; 146 147 default: 148 assert(isVirtualRegister(RegNo) && "Unknown physical register!"); 149 assert(0 && "Register allocator hasn't allocated reg correctly yet!"); 150 return 0; 151 } 152} 153 154const TargetRegisterClass * 155X86RegisterInfo::getCrossCopyRegClass(const TargetRegisterClass *RC) const { 156 if (RC == &X86::CCRRegClass) { 157 if (Is64Bit) 158 return &X86::GR64RegClass; 159 else 160 return &X86::GR32RegClass; 161 } 162 return NULL; 163} 164 165const unsigned * 166X86RegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { 167 bool callsEHReturn = false; 168 169 if (MF) { 170 const MachineFrameInfo *MFI = MF->getFrameInfo(); 171 const MachineModuleInfo *MMI = MFI->getMachineModuleInfo(); 172 callsEHReturn = (MMI ? MMI->callsEHReturn() : false); 173 } 174 175 static const unsigned CalleeSavedRegs32Bit[] = { 176 X86::ESI, X86::EDI, X86::EBX, X86::EBP, 0 177 }; 178 179 static const unsigned CalleeSavedRegs32EHRet[] = { 180 X86::EAX, X86::EDX, X86::ESI, X86::EDI, X86::EBX, X86::EBP, 0 181 }; 182 183 static const unsigned CalleeSavedRegs64Bit[] = { 184 X86::RBX, X86::R12, X86::R13, X86::R14, X86::R15, X86::RBP, 0 185 }; 186 187 static const unsigned CalleeSavedRegs64EHRet[] = { 188 X86::RAX, X86::RDX, X86::RBX, X86::R12, 189 X86::R13, X86::R14, X86::R15, X86::RBP, 0 190 }; 191 192 static const unsigned CalleeSavedRegsWin64[] = { 193 X86::RBX, X86::RBP, X86::RDI, X86::RSI, 194 X86::R12, X86::R13, X86::R14, X86::R15, 195 X86::XMM6, X86::XMM7, X86::XMM8, X86::XMM9, 196 X86::XMM10, X86::XMM11, X86::XMM12, X86::XMM13, 197 X86::XMM14, X86::XMM15, 0 198 }; 199 200 if (Is64Bit) { 201 if (IsWin64) 202 return CalleeSavedRegsWin64; 203 else 204 return (callsEHReturn ? CalleeSavedRegs64EHRet : CalleeSavedRegs64Bit); 205 } else { 206 return (callsEHReturn ? CalleeSavedRegs32EHRet : CalleeSavedRegs32Bit); 207 } 208} 209 210const TargetRegisterClass* const* 211X86RegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const { 212 bool callsEHReturn = false; 213 214 if (MF) { 215 const MachineFrameInfo *MFI = MF->getFrameInfo(); 216 const MachineModuleInfo *MMI = MFI->getMachineModuleInfo(); 217 callsEHReturn = (MMI ? MMI->callsEHReturn() : false); 218 } 219 220 static const TargetRegisterClass * const CalleeSavedRegClasses32Bit[] = { 221 &X86::GR32RegClass, &X86::GR32RegClass, 222 &X86::GR32RegClass, &X86::GR32RegClass, 0 223 }; 224 static const TargetRegisterClass * const CalleeSavedRegClasses32EHRet[] = { 225 &X86::GR32RegClass, &X86::GR32RegClass, 226 &X86::GR32RegClass, &X86::GR32RegClass, 227 &X86::GR32RegClass, &X86::GR32RegClass, 0 228 }; 229 static const TargetRegisterClass * const CalleeSavedRegClasses64Bit[] = { 230 &X86::GR64RegClass, &X86::GR64RegClass, 231 &X86::GR64RegClass, &X86::GR64RegClass, 232 &X86::GR64RegClass, &X86::GR64RegClass, 0 233 }; 234 static const TargetRegisterClass * const CalleeSavedRegClasses64EHRet[] = { 235 &X86::GR64RegClass, &X86::GR64RegClass, 236 &X86::GR64RegClass, &X86::GR64RegClass, 237 &X86::GR64RegClass, &X86::GR64RegClass, 238 &X86::GR64RegClass, &X86::GR64RegClass, 0 239 }; 240 static const TargetRegisterClass * const CalleeSavedRegClassesWin64[] = { 241 &X86::GR64RegClass, &X86::GR64RegClass, 242 &X86::GR64RegClass, &X86::GR64RegClass, 243 &X86::GR64RegClass, &X86::GR64RegClass, 244 &X86::GR64RegClass, &X86::GR64RegClass, 245 &X86::VR128RegClass, &X86::VR128RegClass, 246 &X86::VR128RegClass, &X86::VR128RegClass, 247 &X86::VR128RegClass, &X86::VR128RegClass, 248 &X86::VR128RegClass, &X86::VR128RegClass, 249 &X86::VR128RegClass, &X86::VR128RegClass, 0 250 }; 251 252 if (Is64Bit) { 253 if (IsWin64) 254 return CalleeSavedRegClassesWin64; 255 else 256 return (callsEHReturn ? 257 CalleeSavedRegClasses64EHRet : CalleeSavedRegClasses64Bit); 258 } else { 259 return (callsEHReturn ? 260 CalleeSavedRegClasses32EHRet : CalleeSavedRegClasses32Bit); 261 } 262} 263 264BitVector X86RegisterInfo::getReservedRegs(const MachineFunction &MF) const { 265 BitVector Reserved(getNumRegs()); 266 Reserved.set(X86::RSP); 267 Reserved.set(X86::ESP); 268 Reserved.set(X86::SP); 269 Reserved.set(X86::SPL); 270 if (hasFP(MF)) { 271 Reserved.set(X86::RBP); 272 Reserved.set(X86::EBP); 273 Reserved.set(X86::BP); 274 Reserved.set(X86::BPL); 275 } 276 return Reserved; 277} 278 279//===----------------------------------------------------------------------===// 280// Stack Frame Processing methods 281//===----------------------------------------------------------------------===// 282 283static unsigned calculateMaxStackAlignment(const MachineFrameInfo *FFI) { 284 unsigned MaxAlign = 0; 285 for (int i = FFI->getObjectIndexBegin(), 286 e = FFI->getObjectIndexEnd(); i != e; ++i) { 287 if (FFI->isDeadObjectIndex(i)) 288 continue; 289 unsigned Align = FFI->getObjectAlignment(i); 290 MaxAlign = std::max(MaxAlign, Align); 291 } 292 293 return MaxAlign; 294} 295 296// hasFP - Return true if the specified function should have a dedicated frame 297// pointer register. This is true if the function has variable sized allocas or 298// if frame pointer elimination is disabled. 299// 300bool X86RegisterInfo::hasFP(const MachineFunction &MF) const { 301 const MachineFrameInfo *MFI = MF.getFrameInfo(); 302 const MachineModuleInfo *MMI = MFI->getMachineModuleInfo(); 303 304 return (NoFramePointerElim || 305 needsStackRealignment(MF) || 306 MFI->hasVarSizedObjects() || 307 MFI->isFrameAddressTaken() || 308 MF.getInfo<X86MachineFunctionInfo>()->getForceFramePointer() || 309 (MMI && MMI->callsUnwindInit())); 310} 311 312bool X86RegisterInfo::needsStackRealignment(const MachineFunction &MF) const { 313 const MachineFrameInfo *MFI = MF.getFrameInfo();; 314 315 // FIXME: Currently we don't support stack realignment for functions with 316 // variable-sized allocas 317 return (RealignStack && 318 (MFI->getMaxAlignment() > StackAlign && 319 !MFI->hasVarSizedObjects())); 320} 321 322bool X86RegisterInfo::hasReservedCallFrame(MachineFunction &MF) const { 323 return !MF.getFrameInfo()->hasVarSizedObjects(); 324} 325 326int 327X86RegisterInfo::getFrameIndexOffset(MachineFunction &MF, int FI) const { 328 int Offset = MF.getFrameInfo()->getObjectOffset(FI) + SlotSize; 329 uint64_t StackSize = MF.getFrameInfo()->getStackSize(); 330 331 if (needsStackRealignment(MF)) { 332 if (FI < 0) 333 // Skip the saved EBP 334 Offset += SlotSize; 335 else { 336 unsigned Align = MF.getFrameInfo()->getObjectAlignment(FI); 337 assert( (-(Offset + StackSize)) % Align == 0); 338 return Offset + StackSize; 339 } 340 341 // FIXME: Support tail calls 342 } else { 343 if (!hasFP(MF)) 344 return Offset + StackSize; 345 346 // Skip the saved EBP 347 Offset += SlotSize; 348 349 // Skip the RETADDR move area 350 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 351 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta(); 352 if (TailCallReturnAddrDelta < 0) Offset -= TailCallReturnAddrDelta; 353 } 354 355 return Offset; 356} 357 358void X86RegisterInfo:: 359eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, 360 MachineBasicBlock::iterator I) const { 361 if (!hasReservedCallFrame(MF)) { 362 // If the stack pointer can be changed after prologue, turn the 363 // adjcallstackup instruction into a 'sub ESP, <amt>' and the 364 // adjcallstackdown instruction into 'add ESP, <amt>' 365 // TODO: consider using push / pop instead of sub + store / add 366 MachineInstr *Old = I; 367 uint64_t Amount = Old->getOperand(0).getImm(); 368 if (Amount != 0) { 369 // We need to keep the stack aligned properly. To do this, we round the 370 // amount of space needed for the outgoing arguments up to the next 371 // alignment boundary. 372 Amount = (Amount+StackAlign-1)/StackAlign*StackAlign; 373 374 MachineInstr *New = 0; 375 if (Old->getOpcode() == getCallFrameSetupOpcode()) { 376 New = BuildMI(MF, TII.get(Is64Bit ? X86::SUB64ri32 : X86::SUB32ri), 377 StackPtr).addReg(StackPtr).addImm(Amount); 378 } else { 379 assert(Old->getOpcode() == getCallFrameDestroyOpcode()); 380 // factor out the amount the callee already popped. 381 uint64_t CalleeAmt = Old->getOperand(1).getImm(); 382 Amount -= CalleeAmt; 383 if (Amount) { 384 unsigned Opc = (Amount < 128) ? 385 (Is64Bit ? X86::ADD64ri8 : X86::ADD32ri8) : 386 (Is64Bit ? X86::ADD64ri32 : X86::ADD32ri); 387 New = BuildMI(MF, TII.get(Opc), StackPtr) 388 .addReg(StackPtr).addImm(Amount); 389 } 390 } 391 392 // Replace the pseudo instruction with a new instruction... 393 if (New) MBB.insert(I, New); 394 } 395 } else if (I->getOpcode() == getCallFrameDestroyOpcode()) { 396 // If we are performing frame pointer elimination and if the callee pops 397 // something off the stack pointer, add it back. We do this until we have 398 // more advanced stack pointer tracking ability. 399 if (uint64_t CalleeAmt = I->getOperand(1).getImm()) { 400 unsigned Opc = (CalleeAmt < 128) ? 401 (Is64Bit ? X86::SUB64ri8 : X86::SUB32ri8) : 402 (Is64Bit ? X86::SUB64ri32 : X86::SUB32ri); 403 MachineInstr *New = 404 BuildMI(MF, TII.get(Opc), StackPtr).addReg(StackPtr).addImm(CalleeAmt); 405 MBB.insert(I, New); 406 } 407 } 408 409 MBB.erase(I); 410} 411 412void X86RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, 413 int SPAdj, RegScavenger *RS) const{ 414 assert(SPAdj == 0 && "Unexpected"); 415 416 unsigned i = 0; 417 MachineInstr &MI = *II; 418 MachineFunction &MF = *MI.getParent()->getParent(); 419 while (!MI.getOperand(i).isFI()) { 420 ++i; 421 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!"); 422 } 423 424 int FrameIndex = MI.getOperand(i).getIndex(); 425 426 unsigned BasePtr; 427 if (needsStackRealignment(MF)) 428 BasePtr = (FrameIndex < 0 ? FramePtr : StackPtr); 429 else 430 BasePtr = (hasFP(MF) ? FramePtr : StackPtr); 431 432 // This must be part of a four operand memory reference. Replace the 433 // FrameIndex with base register with EBP. Add an offset to the offset. 434 MI.getOperand(i).ChangeToRegister(BasePtr, false); 435 436 // Now add the frame object offset to the offset from EBP. Offset is a 437 // 32-bit integer. 438 int Offset = getFrameIndexOffset(MF, FrameIndex) + 439 (int)(MI.getOperand(i+3).getImm()); 440 441 MI.getOperand(i+3).ChangeToImmediate(Offset); 442} 443 444void 445X86RegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF, 446 RegScavenger *RS) const { 447 MachineFrameInfo *FFI = MF.getFrameInfo(); 448 449 // Calculate and set max stack object alignment early, so we can decide 450 // whether we will need stack realignment (and thus FP). 451 unsigned MaxAlign = std::max(FFI->getMaxAlignment(), 452 calculateMaxStackAlignment(FFI)); 453 454 FFI->setMaxAlignment(MaxAlign); 455} 456 457void 458X86RegisterInfo::processFunctionBeforeFrameFinalized(MachineFunction &MF) const{ 459 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 460 int32_t TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta(); 461 if (TailCallReturnAddrDelta < 0) { 462 // create RETURNADDR area 463 // arg 464 // arg 465 // RETADDR 466 // { ... 467 // RETADDR area 468 // ... 469 // } 470 // [EBP] 471 MF.getFrameInfo()-> 472 CreateFixedObject(-TailCallReturnAddrDelta, 473 (-1*SlotSize)+TailCallReturnAddrDelta); 474 } 475 if (hasFP(MF)) { 476 assert((TailCallReturnAddrDelta <= 0) && 477 "The Delta should always be zero or negative"); 478 // Create a frame entry for the EBP register that must be saved. 479 int FrameIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, 480 (int)SlotSize * -2+ 481 TailCallReturnAddrDelta); 482 assert(FrameIdx == MF.getFrameInfo()->getObjectIndexBegin() && 483 "Slot for EBP register must be last in order to be found!"); 484 } 485} 486 487/// emitSPUpdate - Emit a series of instructions to increment / decrement the 488/// stack pointer by a constant value. 489static 490void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI, 491 unsigned StackPtr, int64_t NumBytes, bool Is64Bit, 492 const TargetInstrInfo &TII) { 493 bool isSub = NumBytes < 0; 494 uint64_t Offset = isSub ? -NumBytes : NumBytes; 495 unsigned Opc = isSub 496 ? ((Offset < 128) ? 497 (Is64Bit ? X86::SUB64ri8 : X86::SUB32ri8) : 498 (Is64Bit ? X86::SUB64ri32 : X86::SUB32ri)) 499 : ((Offset < 128) ? 500 (Is64Bit ? X86::ADD64ri8 : X86::ADD32ri8) : 501 (Is64Bit ? X86::ADD64ri32 : X86::ADD32ri)); 502 uint64_t Chunk = (1LL << 31) - 1; 503 504 while (Offset) { 505 uint64_t ThisVal = (Offset > Chunk) ? Chunk : Offset; 506 BuildMI(MBB, MBBI, TII.get(Opc), StackPtr).addReg(StackPtr).addImm(ThisVal); 507 Offset -= ThisVal; 508 } 509} 510 511// mergeSPUpdatesUp - Merge two stack-manipulating instructions upper iterator. 512static 513void mergeSPUpdatesUp(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI, 514 unsigned StackPtr, uint64_t *NumBytes = NULL) { 515 if (MBBI == MBB.begin()) return; 516 517 MachineBasicBlock::iterator PI = prior(MBBI); 518 unsigned Opc = PI->getOpcode(); 519 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 || 520 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) && 521 PI->getOperand(0).getReg() == StackPtr) { 522 if (NumBytes) 523 *NumBytes += PI->getOperand(2).getImm(); 524 MBB.erase(PI); 525 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 || 526 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) && 527 PI->getOperand(0).getReg() == StackPtr) { 528 if (NumBytes) 529 *NumBytes -= PI->getOperand(2).getImm(); 530 MBB.erase(PI); 531 } 532} 533 534// mergeSPUpdatesUp - Merge two stack-manipulating instructions lower iterator. 535static 536void mergeSPUpdatesDown(MachineBasicBlock &MBB, 537 MachineBasicBlock::iterator &MBBI, 538 unsigned StackPtr, uint64_t *NumBytes = NULL) { 539 return; 540 541 if (MBBI == MBB.end()) return; 542 543 MachineBasicBlock::iterator NI = next(MBBI); 544 if (NI == MBB.end()) return; 545 546 unsigned Opc = NI->getOpcode(); 547 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 || 548 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) && 549 NI->getOperand(0).getReg() == StackPtr) { 550 if (NumBytes) 551 *NumBytes -= NI->getOperand(2).getImm(); 552 MBB.erase(NI); 553 MBBI = NI; 554 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 || 555 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) && 556 NI->getOperand(0).getReg() == StackPtr) { 557 if (NumBytes) 558 *NumBytes += NI->getOperand(2).getImm(); 559 MBB.erase(NI); 560 MBBI = NI; 561 } 562} 563 564/// mergeSPUpdates - Checks the instruction before/after the passed 565/// instruction. If it is an ADD/SUB instruction it is deleted 566/// argument and the stack adjustment is returned as a positive value for ADD 567/// and a negative for SUB. 568static int mergeSPUpdates(MachineBasicBlock &MBB, 569 MachineBasicBlock::iterator &MBBI, 570 unsigned StackPtr, 571 bool doMergeWithPrevious) { 572 573 if ((doMergeWithPrevious && MBBI == MBB.begin()) || 574 (!doMergeWithPrevious && MBBI == MBB.end())) 575 return 0; 576 577 int Offset = 0; 578 579 MachineBasicBlock::iterator PI = doMergeWithPrevious ? prior(MBBI) : MBBI; 580 MachineBasicBlock::iterator NI = doMergeWithPrevious ? 0 : next(MBBI); 581 unsigned Opc = PI->getOpcode(); 582 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 || 583 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) && 584 PI->getOperand(0).getReg() == StackPtr){ 585 Offset += PI->getOperand(2).getImm(); 586 MBB.erase(PI); 587 if (!doMergeWithPrevious) MBBI = NI; 588 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 || 589 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) && 590 PI->getOperand(0).getReg() == StackPtr) { 591 Offset -= PI->getOperand(2).getImm(); 592 MBB.erase(PI); 593 if (!doMergeWithPrevious) MBBI = NI; 594 } 595 596 return Offset; 597} 598 599void X86RegisterInfo::emitFrameMoves(MachineFunction &MF, 600 unsigned FrameLabelId, 601 unsigned ReadyLabelId) const { 602 MachineFrameInfo *MFI = MF.getFrameInfo(); 603 MachineModuleInfo *MMI = MFI->getMachineModuleInfo(); 604 if (!MMI) 605 return; 606 607 uint64_t StackSize = MFI->getStackSize(); 608 std::vector<MachineMove> &Moves = MMI->getFrameMoves(); 609 const TargetData *TD = MF.getTarget().getTargetData(); 610 611 // Calculate amount of bytes used for return address storing 612 int stackGrowth = 613 (MF.getTarget().getFrameInfo()->getStackGrowthDirection() == 614 TargetFrameInfo::StackGrowsUp ? 615 TD->getPointerSize() : -TD->getPointerSize()); 616 617 if (StackSize) { 618 // Show update of SP. 619 if (hasFP(MF)) { 620 // Adjust SP 621 MachineLocation SPDst(MachineLocation::VirtualFP); 622 MachineLocation SPSrc(MachineLocation::VirtualFP, 2*stackGrowth); 623 Moves.push_back(MachineMove(FrameLabelId, SPDst, SPSrc)); 624 } else { 625 MachineLocation SPDst(MachineLocation::VirtualFP); 626 MachineLocation SPSrc(MachineLocation::VirtualFP, 627 -StackSize+stackGrowth); 628 Moves.push_back(MachineMove(FrameLabelId, SPDst, SPSrc)); 629 } 630 } else { 631 //FIXME: Verify & implement for FP 632 MachineLocation SPDst(StackPtr); 633 MachineLocation SPSrc(StackPtr, stackGrowth); 634 Moves.push_back(MachineMove(FrameLabelId, SPDst, SPSrc)); 635 } 636 637 // Add callee saved registers to move list. 638 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo(); 639 640 // FIXME: This is dirty hack. The code itself is pretty mess right now. 641 // It should be rewritten from scratch and generalized sometimes. 642 643 // Determine maximum offset (minumum due to stack growth) 644 int64_t MaxOffset = 0; 645 for (unsigned I = 0, E = CSI.size(); I!=E; ++I) 646 MaxOffset = std::min(MaxOffset, 647 MFI->getObjectOffset(CSI[I].getFrameIdx())); 648 649 // Calculate offsets 650 int64_t saveAreaOffset = (hasFP(MF) ? 3 : 2)*stackGrowth; 651 for (unsigned I = 0, E = CSI.size(); I!=E; ++I) { 652 int64_t Offset = MFI->getObjectOffset(CSI[I].getFrameIdx()); 653 unsigned Reg = CSI[I].getReg(); 654 Offset = (MaxOffset-Offset+saveAreaOffset); 655 MachineLocation CSDst(MachineLocation::VirtualFP, Offset); 656 MachineLocation CSSrc(Reg); 657 Moves.push_back(MachineMove(FrameLabelId, CSDst, CSSrc)); 658 } 659 660 if (hasFP(MF)) { 661 // Save FP 662 MachineLocation FPDst(MachineLocation::VirtualFP, 2*stackGrowth); 663 MachineLocation FPSrc(FramePtr); 664 Moves.push_back(MachineMove(ReadyLabelId, FPDst, FPSrc)); 665 } 666 667 MachineLocation FPDst(hasFP(MF) ? FramePtr : StackPtr); 668 MachineLocation FPSrc(MachineLocation::VirtualFP); 669 Moves.push_back(MachineMove(ReadyLabelId, FPDst, FPSrc)); 670} 671 672 673void X86RegisterInfo::emitPrologue(MachineFunction &MF) const { 674 MachineBasicBlock &MBB = MF.front(); // Prolog goes in entry BB 675 MachineFrameInfo *MFI = MF.getFrameInfo(); 676 const Function* Fn = MF.getFunction(); 677 const X86Subtarget* Subtarget = &MF.getTarget().getSubtarget<X86Subtarget>(); 678 MachineModuleInfo *MMI = MFI->getMachineModuleInfo(); 679 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 680 MachineBasicBlock::iterator MBBI = MBB.begin(); 681 bool needsFrameMoves = (MMI && MMI->hasDebugInfo()) || 682 !Fn->doesNotThrow() || 683 UnwindTablesMandatory; 684 // Prepare for frame info. 685 unsigned FrameLabelId = 0; 686 687 // Get the number of bytes to allocate from the FrameInfo. 688 uint64_t StackSize = MFI->getStackSize(); 689 // Get desired stack alignment 690 uint64_t MaxAlign = MFI->getMaxAlignment(); 691 692 // Add RETADDR move area to callee saved frame size. 693 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta(); 694 if (TailCallReturnAddrDelta < 0) 695 X86FI->setCalleeSavedFrameSize( 696 X86FI->getCalleeSavedFrameSize() +(-TailCallReturnAddrDelta)); 697 698 // Insert stack pointer adjustment for later moving of return addr. Only 699 // applies to tail call optimized functions where the callee argument stack 700 // size is bigger than the callers. 701 if (TailCallReturnAddrDelta < 0) { 702 BuildMI(MBB, MBBI, TII.get(Is64Bit? X86::SUB64ri32 : X86::SUB32ri), 703 StackPtr).addReg(StackPtr).addImm(-TailCallReturnAddrDelta); 704 } 705 706 uint64_t NumBytes = 0; 707 if (hasFP(MF)) { 708 // Calculate required stack adjustment 709 uint64_t FrameSize = StackSize - SlotSize; 710 if (needsStackRealignment(MF)) 711 FrameSize = (FrameSize + MaxAlign - 1)/MaxAlign*MaxAlign; 712 713 NumBytes = FrameSize - X86FI->getCalleeSavedFrameSize(); 714 715 // Get the offset of the stack slot for the EBP register... which is 716 // guaranteed to be the last slot by processFunctionBeforeFrameFinalized. 717 // Update the frame offset adjustment. 718 MFI->setOffsetAdjustment(-NumBytes); 719 720 // Save EBP into the appropriate stack slot... 721 BuildMI(MBB, MBBI, TII.get(Is64Bit ? X86::PUSH64r : X86::PUSH32r)) 722 .addReg(FramePtr); 723 724 if (needsFrameMoves) { 725 // Mark effective beginning of when frame pointer becomes valid. 726 FrameLabelId = MMI->NextLabelID(); 727 BuildMI(MBB, MBBI, TII.get(X86::DBG_LABEL)).addImm(FrameLabelId); 728 } 729 730 // Update EBP with the new base value... 731 BuildMI(MBB, MBBI, TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), FramePtr) 732 .addReg(StackPtr); 733 734 // Realign stack 735 if (needsStackRealignment(MF)) 736 BuildMI(MBB, MBBI, 737 TII.get(Is64Bit ? X86::AND64ri32 : X86::AND32ri), 738 StackPtr).addReg(StackPtr).addImm(-MaxAlign); 739 } else 740 NumBytes = StackSize - X86FI->getCalleeSavedFrameSize(); 741 742 unsigned ReadyLabelId = 0; 743 if (needsFrameMoves) { 744 // Mark effective beginning of when frame pointer is ready. 745 ReadyLabelId = MMI->NextLabelID(); 746 BuildMI(MBB, MBBI, TII.get(X86::DBG_LABEL)).addImm(ReadyLabelId); 747 } 748 749 // Skip the callee-saved push instructions. 750 while (MBBI != MBB.end() && 751 (MBBI->getOpcode() == X86::PUSH32r || 752 MBBI->getOpcode() == X86::PUSH64r)) 753 ++MBBI; 754 755 if (NumBytes) { // adjust stack pointer: ESP -= numbytes 756 if (NumBytes >= 4096 && Subtarget->isTargetCygMing()) { 757 // Check, whether EAX is livein for this function 758 bool isEAXAlive = false; 759 for (MachineRegisterInfo::livein_iterator 760 II = MF.getRegInfo().livein_begin(), 761 EE = MF.getRegInfo().livein_end(); (II != EE) && !isEAXAlive; ++II) { 762 unsigned Reg = II->first; 763 isEAXAlive = (Reg == X86::EAX || Reg == X86::AX || 764 Reg == X86::AH || Reg == X86::AL); 765 } 766 767 // Function prologue calls _alloca to probe the stack when allocating 768 // more than 4k bytes in one go. Touching the stack at 4K increments is 769 // necessary to ensure that the guard pages used by the OS virtual memory 770 // manager are allocated in correct sequence. 771 if (!isEAXAlive) { 772 BuildMI(MBB, MBBI, TII.get(X86::MOV32ri), X86::EAX).addImm(NumBytes); 773 BuildMI(MBB, MBBI, TII.get(X86::CALLpcrel32)) 774 .addExternalSymbol("_alloca"); 775 } else { 776 // Save EAX 777 BuildMI(MBB, MBBI, TII.get(X86::PUSH32r), X86::EAX); 778 // Allocate NumBytes-4 bytes on stack. We'll also use 4 already 779 // allocated bytes for EAX. 780 BuildMI(MBB, MBBI, TII.get(X86::MOV32ri), X86::EAX).addImm(NumBytes-4); 781 BuildMI(MBB, MBBI, TII.get(X86::CALLpcrel32)) 782 .addExternalSymbol("_alloca"); 783 // Restore EAX 784 MachineInstr *MI = addRegOffset(BuildMI(MF, TII.get(X86::MOV32rm),X86::EAX), 785 StackPtr, false, NumBytes-4); 786 MBB.insert(MBBI, MI); 787 } 788 } else { 789 // If there is an SUB32ri of ESP immediately before this instruction, 790 // merge the two. This can be the case when tail call elimination is 791 // enabled and the callee has more arguments then the caller. 792 NumBytes -= mergeSPUpdates(MBB, MBBI, StackPtr, true); 793 // If there is an ADD32ri or SUB32ri of ESP immediately after this 794 // instruction, merge the two instructions. 795 mergeSPUpdatesDown(MBB, MBBI, StackPtr, &NumBytes); 796 797 if (NumBytes) 798 emitSPUpdate(MBB, MBBI, StackPtr, -(int64_t)NumBytes, Is64Bit, TII); 799 } 800 } 801 802 if (needsFrameMoves) 803 emitFrameMoves(MF, FrameLabelId, ReadyLabelId); 804} 805 806void X86RegisterInfo::emitEpilogue(MachineFunction &MF, 807 MachineBasicBlock &MBB) const { 808 const MachineFrameInfo *MFI = MF.getFrameInfo(); 809 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 810 MachineBasicBlock::iterator MBBI = prior(MBB.end()); 811 unsigned RetOpcode = MBBI->getOpcode(); 812 813 switch (RetOpcode) { 814 case X86::RET: 815 case X86::RETI: 816 case X86::TCRETURNdi: 817 case X86::TCRETURNri: 818 case X86::TCRETURNri64: 819 case X86::TCRETURNdi64: 820 case X86::EH_RETURN: 821 case X86::EH_RETURN64: 822 case X86::TAILJMPd: 823 case X86::TAILJMPr: 824 case X86::TAILJMPm: break; // These are ok 825 default: 826 assert(0 && "Can only insert epilog into returning blocks"); 827 } 828 829 // Get the number of bytes to allocate from the FrameInfo 830 uint64_t StackSize = MFI->getStackSize(); 831 uint64_t MaxAlign = MFI->getMaxAlignment(); 832 unsigned CSSize = X86FI->getCalleeSavedFrameSize(); 833 uint64_t NumBytes = 0; 834 835 if (hasFP(MF)) { 836 // Calculate required stack adjustment 837 uint64_t FrameSize = StackSize - SlotSize; 838 if (needsStackRealignment(MF)) 839 FrameSize = (FrameSize + MaxAlign - 1)/MaxAlign*MaxAlign; 840 841 NumBytes = FrameSize - CSSize; 842 843 // pop EBP. 844 BuildMI(MBB, MBBI, TII.get(Is64Bit ? X86::POP64r : X86::POP32r), FramePtr); 845 } else 846 NumBytes = StackSize - CSSize; 847 848 // Skip the callee-saved pop instructions. 849 MachineBasicBlock::iterator LastCSPop = MBBI; 850 while (MBBI != MBB.begin()) { 851 MachineBasicBlock::iterator PI = prior(MBBI); 852 unsigned Opc = PI->getOpcode(); 853 if (Opc != X86::POP32r && Opc != X86::POP64r && !PI->getDesc().isReturn()) 854 break; 855 --MBBI; 856 } 857 858 // If there is an ADD32ri or SUB32ri of ESP immediately before this 859 // instruction, merge the two instructions. 860 if (NumBytes || MFI->hasVarSizedObjects()) 861 mergeSPUpdatesUp(MBB, MBBI, StackPtr, &NumBytes); 862 863 // If dynamic alloca is used, then reset esp to point to the last callee-saved 864 // slot before popping them off! Same applies for the case, when stack was 865 // realigned 866 if (needsStackRealignment(MF)) { 867 // We cannot use LEA here, because stack pointer was realigned. We need to 868 // deallocate local frame back 869 if (CSSize) { 870 emitSPUpdate(MBB, MBBI, StackPtr, NumBytes, Is64Bit, TII); 871 MBBI = prior(LastCSPop); 872 } 873 874 BuildMI(MBB, MBBI, 875 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), 876 StackPtr).addReg(FramePtr); 877 } else if (MFI->hasVarSizedObjects()) { 878 if (CSSize) { 879 unsigned Opc = Is64Bit ? X86::LEA64r : X86::LEA32r; 880 MachineInstr *MI = addRegOffset(BuildMI(MF, TII.get(Opc), StackPtr), 881 FramePtr, false, -CSSize); 882 MBB.insert(MBBI, MI); 883 } else 884 BuildMI(MBB, MBBI, TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), 885 StackPtr).addReg(FramePtr); 886 887 } else { 888 // adjust stack pointer back: ESP += numbytes 889 if (NumBytes) 890 emitSPUpdate(MBB, MBBI, StackPtr, NumBytes, Is64Bit, TII); 891 } 892 893 // We're returning from function via eh_return. 894 if (RetOpcode == X86::EH_RETURN || RetOpcode == X86::EH_RETURN64) { 895 MBBI = prior(MBB.end()); 896 MachineOperand &DestAddr = MBBI->getOperand(0); 897 assert(DestAddr.isReg() && "Offset should be in register!"); 898 BuildMI(MBB, MBBI, 899 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), 900 StackPtr).addReg(DestAddr.getReg()); 901 // Tail call return: adjust the stack pointer and jump to callee 902 } else if (RetOpcode == X86::TCRETURNri || RetOpcode == X86::TCRETURNdi || 903 RetOpcode== X86::TCRETURNri64 || RetOpcode == X86::TCRETURNdi64) { 904 MBBI = prior(MBB.end()); 905 MachineOperand &JumpTarget = MBBI->getOperand(0); 906 MachineOperand &StackAdjust = MBBI->getOperand(1); 907 assert(StackAdjust.isImm() && "Expecting immediate value."); 908 909 // Adjust stack pointer. 910 int StackAdj = StackAdjust.getImm(); 911 int MaxTCDelta = X86FI->getTCReturnAddrDelta(); 912 int Offset = 0; 913 assert(MaxTCDelta <= 0 && "MaxTCDelta should never be positive"); 914 // Incoporate the retaddr area. 915 Offset = StackAdj-MaxTCDelta; 916 assert(Offset >= 0 && "Offset should never be negative"); 917 if (Offset) { 918 // Check for possible merge with preceeding ADD instruction. 919 Offset += mergeSPUpdates(MBB, MBBI, StackPtr, true); 920 emitSPUpdate(MBB, MBBI, StackPtr, Offset, Is64Bit, TII); 921 } 922 // Jump to label or value in register. 923 if (RetOpcode == X86::TCRETURNdi|| RetOpcode == X86::TCRETURNdi64) 924 BuildMI(MBB, MBBI, TII.get(X86::TAILJMPd)). 925 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset()); 926 else if (RetOpcode== X86::TCRETURNri64) { 927 BuildMI(MBB, MBBI, TII.get(X86::TAILJMPr64), JumpTarget.getReg()); 928 } else 929 BuildMI(MBB, MBBI, TII.get(X86::TAILJMPr), JumpTarget.getReg()); 930 // Delete the pseudo instruction TCRETURN. 931 MBB.erase(MBBI); 932 } else if ((RetOpcode == X86::RET || RetOpcode == X86::RETI) && 933 (X86FI->getTCReturnAddrDelta() < 0)) { 934 // Add the return addr area delta back since we are not tail calling. 935 int delta = -1*X86FI->getTCReturnAddrDelta(); 936 MBBI = prior(MBB.end()); 937 // Check for possible merge with preceeding ADD instruction. 938 delta += mergeSPUpdates(MBB, MBBI, StackPtr, true); 939 emitSPUpdate(MBB, MBBI, StackPtr, delta, Is64Bit, TII); 940 } 941} 942 943unsigned X86RegisterInfo::getRARegister() const { 944 if (Is64Bit) 945 return X86::RIP; // Should have dwarf #16 946 else 947 return X86::EIP; // Should have dwarf #8 948} 949 950unsigned X86RegisterInfo::getFrameRegister(MachineFunction &MF) const { 951 return hasFP(MF) ? FramePtr : StackPtr; 952} 953 954void X86RegisterInfo::getInitialFrameState(std::vector<MachineMove> &Moves) 955 const { 956 // Calculate amount of bytes used for return address storing 957 int stackGrowth = (Is64Bit ? -8 : -4); 958 959 // Initial state of the frame pointer is esp+4. 960 MachineLocation Dst(MachineLocation::VirtualFP); 961 MachineLocation Src(StackPtr, stackGrowth); 962 Moves.push_back(MachineMove(0, Dst, Src)); 963 964 // Add return address to move list 965 MachineLocation CSDst(StackPtr, stackGrowth); 966 MachineLocation CSSrc(getRARegister()); 967 Moves.push_back(MachineMove(0, CSDst, CSSrc)); 968} 969 970unsigned X86RegisterInfo::getEHExceptionRegister() const { 971 assert(0 && "What is the exception register"); 972 return 0; 973} 974 975unsigned X86RegisterInfo::getEHHandlerRegister() const { 976 assert(0 && "What is the exception handler register"); 977 return 0; 978} 979 980namespace llvm { 981unsigned getX86SubSuperRegister(unsigned Reg, MVT VT, bool High) { 982 switch (VT.getSimpleVT()) { 983 default: return Reg; 984 case MVT::i8: 985 if (High) { 986 switch (Reg) { 987 default: return 0; 988 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 989 return X86::AH; 990 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 991 return X86::DH; 992 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 993 return X86::CH; 994 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 995 return X86::BH; 996 } 997 } else { 998 switch (Reg) { 999 default: return 0; 1000 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 1001 return X86::AL; 1002 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 1003 return X86::DL; 1004 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 1005 return X86::CL; 1006 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 1007 return X86::BL; 1008 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 1009 return X86::SIL; 1010 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 1011 return X86::DIL; 1012 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 1013 return X86::BPL; 1014 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 1015 return X86::SPL; 1016 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 1017 return X86::R8B; 1018 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 1019 return X86::R9B; 1020 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 1021 return X86::R10B; 1022 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 1023 return X86::R11B; 1024 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 1025 return X86::R12B; 1026 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 1027 return X86::R13B; 1028 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 1029 return X86::R14B; 1030 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 1031 return X86::R15B; 1032 } 1033 } 1034 case MVT::i16: 1035 switch (Reg) { 1036 default: return Reg; 1037 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 1038 return X86::AX; 1039 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 1040 return X86::DX; 1041 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 1042 return X86::CX; 1043 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 1044 return X86::BX; 1045 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 1046 return X86::SI; 1047 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 1048 return X86::DI; 1049 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 1050 return X86::BP; 1051 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 1052 return X86::SP; 1053 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 1054 return X86::R8W; 1055 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 1056 return X86::R9W; 1057 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 1058 return X86::R10W; 1059 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 1060 return X86::R11W; 1061 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 1062 return X86::R12W; 1063 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 1064 return X86::R13W; 1065 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 1066 return X86::R14W; 1067 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 1068 return X86::R15W; 1069 } 1070 case MVT::i32: 1071 switch (Reg) { 1072 default: return Reg; 1073 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 1074 return X86::EAX; 1075 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 1076 return X86::EDX; 1077 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 1078 return X86::ECX; 1079 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 1080 return X86::EBX; 1081 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 1082 return X86::ESI; 1083 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 1084 return X86::EDI; 1085 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 1086 return X86::EBP; 1087 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 1088 return X86::ESP; 1089 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 1090 return X86::R8D; 1091 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 1092 return X86::R9D; 1093 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 1094 return X86::R10D; 1095 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 1096 return X86::R11D; 1097 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 1098 return X86::R12D; 1099 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 1100 return X86::R13D; 1101 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 1102 return X86::R14D; 1103 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 1104 return X86::R15D; 1105 } 1106 case MVT::i64: 1107 switch (Reg) { 1108 default: return Reg; 1109 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 1110 return X86::RAX; 1111 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 1112 return X86::RDX; 1113 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 1114 return X86::RCX; 1115 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 1116 return X86::RBX; 1117 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 1118 return X86::RSI; 1119 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 1120 return X86::RDI; 1121 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 1122 return X86::RBP; 1123 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 1124 return X86::RSP; 1125 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 1126 return X86::R8; 1127 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 1128 return X86::R9; 1129 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 1130 return X86::R10; 1131 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 1132 return X86::R11; 1133 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 1134 return X86::R12; 1135 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 1136 return X86::R13; 1137 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 1138 return X86::R14; 1139 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 1140 return X86::R15; 1141 } 1142 } 1143 1144 return Reg; 1145} 1146} 1147 1148#include "X86GenRegisterInfo.inc" 1149 1150namespace { 1151 struct VISIBILITY_HIDDEN MSAC : public MachineFunctionPass { 1152 static char ID; 1153 MSAC() : MachineFunctionPass(&ID) {} 1154 1155 virtual bool runOnMachineFunction(MachineFunction &MF) { 1156 MachineFrameInfo *FFI = MF.getFrameInfo(); 1157 MachineRegisterInfo &RI = MF.getRegInfo(); 1158 1159 // Calculate max stack alignment of all already allocated stack objects. 1160 unsigned MaxAlign = calculateMaxStackAlignment(FFI); 1161 1162 // Be over-conservative: scan over all vreg defs and find, whether vector 1163 // registers are used. If yes - there is probability, that vector register 1164 // will be spilled and thus stack needs to be aligned properly. 1165 for (unsigned RegNum = TargetRegisterInfo::FirstVirtualRegister; 1166 RegNum < RI.getLastVirtReg(); ++RegNum) 1167 MaxAlign = std::max(MaxAlign, RI.getRegClass(RegNum)->getAlignment()); 1168 1169 FFI->setMaxAlignment(MaxAlign); 1170 1171 return false; 1172 } 1173 1174 virtual const char *getPassName() const { 1175 return "X86 Maximal Stack Alignment Calculator"; 1176 } 1177 }; 1178 1179 char MSAC::ID = 0; 1180} 1181 1182FunctionPass* 1183llvm::createX86MaxStackAlignmentCalculatorPass() { return new MSAC(); } 1184