X86RegisterInfo.cpp revision 3d2445f5d97ab615ecb8b26d8c5440a6d774c28d
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 *X86RegisterInfo::getPointerRegClass() const { 155 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>(); 156 if (Subtarget->is64Bit()) 157 return &X86::GR64RegClass; 158 else 159 return &X86::GR32RegClass; 160} 161 162const TargetRegisterClass * 163X86RegisterInfo::getCrossCopyRegClass(const TargetRegisterClass *RC) const { 164 if (RC == &X86::CCRRegClass) { 165 if (Is64Bit) 166 return &X86::GR64RegClass; 167 else 168 return &X86::GR32RegClass; 169 } 170 return NULL; 171} 172 173const unsigned * 174X86RegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { 175 bool callsEHReturn = false; 176 177 if (MF) { 178 const MachineFrameInfo *MFI = MF->getFrameInfo(); 179 const MachineModuleInfo *MMI = MFI->getMachineModuleInfo(); 180 callsEHReturn = (MMI ? MMI->callsEHReturn() : false); 181 } 182 183 static const unsigned CalleeSavedRegs32Bit[] = { 184 X86::ESI, X86::EDI, X86::EBX, X86::EBP, 0 185 }; 186 187 static const unsigned CalleeSavedRegs32EHRet[] = { 188 X86::EAX, X86::EDX, X86::ESI, X86::EDI, X86::EBX, X86::EBP, 0 189 }; 190 191 static const unsigned CalleeSavedRegs64Bit[] = { 192 X86::RBX, X86::R12, X86::R13, X86::R14, X86::R15, X86::RBP, 0 193 }; 194 195 static const unsigned CalleeSavedRegs64EHRet[] = { 196 X86::RAX, X86::RDX, X86::RBX, X86::R12, 197 X86::R13, X86::R14, X86::R15, X86::RBP, 0 198 }; 199 200 static const unsigned CalleeSavedRegsWin64[] = { 201 X86::RBX, X86::RBP, X86::RDI, X86::RSI, 202 X86::R12, X86::R13, X86::R14, X86::R15, 203 X86::XMM6, X86::XMM7, X86::XMM8, X86::XMM9, 204 X86::XMM10, X86::XMM11, X86::XMM12, X86::XMM13, 205 X86::XMM14, X86::XMM15, 0 206 }; 207 208 if (Is64Bit) { 209 if (IsWin64) 210 return CalleeSavedRegsWin64; 211 else 212 return (callsEHReturn ? CalleeSavedRegs64EHRet : CalleeSavedRegs64Bit); 213 } else { 214 return (callsEHReturn ? CalleeSavedRegs32EHRet : CalleeSavedRegs32Bit); 215 } 216} 217 218const TargetRegisterClass* const* 219X86RegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const { 220 bool callsEHReturn = false; 221 222 if (MF) { 223 const MachineFrameInfo *MFI = MF->getFrameInfo(); 224 const MachineModuleInfo *MMI = MFI->getMachineModuleInfo(); 225 callsEHReturn = (MMI ? MMI->callsEHReturn() : false); 226 } 227 228 static const TargetRegisterClass * const CalleeSavedRegClasses32Bit[] = { 229 &X86::GR32RegClass, &X86::GR32RegClass, 230 &X86::GR32RegClass, &X86::GR32RegClass, 0 231 }; 232 static const TargetRegisterClass * const CalleeSavedRegClasses32EHRet[] = { 233 &X86::GR32RegClass, &X86::GR32RegClass, 234 &X86::GR32RegClass, &X86::GR32RegClass, 235 &X86::GR32RegClass, &X86::GR32RegClass, 0 236 }; 237 static const TargetRegisterClass * const CalleeSavedRegClasses64Bit[] = { 238 &X86::GR64RegClass, &X86::GR64RegClass, 239 &X86::GR64RegClass, &X86::GR64RegClass, 240 &X86::GR64RegClass, &X86::GR64RegClass, 0 241 }; 242 static const TargetRegisterClass * const CalleeSavedRegClasses64EHRet[] = { 243 &X86::GR64RegClass, &X86::GR64RegClass, 244 &X86::GR64RegClass, &X86::GR64RegClass, 245 &X86::GR64RegClass, &X86::GR64RegClass, 246 &X86::GR64RegClass, &X86::GR64RegClass, 0 247 }; 248 static const TargetRegisterClass * const CalleeSavedRegClassesWin64[] = { 249 &X86::GR64RegClass, &X86::GR64RegClass, 250 &X86::GR64RegClass, &X86::GR64RegClass, 251 &X86::GR64RegClass, &X86::GR64RegClass, 252 &X86::GR64RegClass, &X86::GR64RegClass, 253 &X86::VR128RegClass, &X86::VR128RegClass, 254 &X86::VR128RegClass, &X86::VR128RegClass, 255 &X86::VR128RegClass, &X86::VR128RegClass, 256 &X86::VR128RegClass, &X86::VR128RegClass, 257 &X86::VR128RegClass, &X86::VR128RegClass, 0 258 }; 259 260 if (Is64Bit) { 261 if (IsWin64) 262 return CalleeSavedRegClassesWin64; 263 else 264 return (callsEHReturn ? 265 CalleeSavedRegClasses64EHRet : CalleeSavedRegClasses64Bit); 266 } else { 267 return (callsEHReturn ? 268 CalleeSavedRegClasses32EHRet : CalleeSavedRegClasses32Bit); 269 } 270} 271 272BitVector X86RegisterInfo::getReservedRegs(const MachineFunction &MF) const { 273 BitVector Reserved(getNumRegs()); 274 // Set the stack-pointer register and its aliases as reserved. 275 Reserved.set(X86::RSP); 276 Reserved.set(X86::ESP); 277 Reserved.set(X86::SP); 278 Reserved.set(X86::SPL); 279 // Set the frame-pointer register and its aliases as reserved if needed. 280 if (hasFP(MF)) { 281 Reserved.set(X86::RBP); 282 Reserved.set(X86::EBP); 283 Reserved.set(X86::BP); 284 Reserved.set(X86::BPL); 285 } 286 // Mark the x87 stack registers as reserved, since they don't 287 // behave normally with respect to liveness. We don't fully 288 // model the effects of x87 stack pushes and pops after 289 // stackification. 290 Reserved.set(X86::ST0); 291 Reserved.set(X86::ST1); 292 Reserved.set(X86::ST2); 293 Reserved.set(X86::ST3); 294 Reserved.set(X86::ST4); 295 Reserved.set(X86::ST5); 296 Reserved.set(X86::ST6); 297 Reserved.set(X86::ST7); 298 return Reserved; 299} 300 301//===----------------------------------------------------------------------===// 302// Stack Frame Processing methods 303//===----------------------------------------------------------------------===// 304 305static unsigned calculateMaxStackAlignment(const MachineFrameInfo *FFI) { 306 unsigned MaxAlign = 0; 307 for (int i = FFI->getObjectIndexBegin(), 308 e = FFI->getObjectIndexEnd(); i != e; ++i) { 309 if (FFI->isDeadObjectIndex(i)) 310 continue; 311 unsigned Align = FFI->getObjectAlignment(i); 312 MaxAlign = std::max(MaxAlign, Align); 313 } 314 315 return MaxAlign; 316} 317 318// hasFP - Return true if the specified function should have a dedicated frame 319// pointer register. This is true if the function has variable sized allocas or 320// if frame pointer elimination is disabled. 321// 322bool X86RegisterInfo::hasFP(const MachineFunction &MF) const { 323 const MachineFrameInfo *MFI = MF.getFrameInfo(); 324 const MachineModuleInfo *MMI = MFI->getMachineModuleInfo(); 325 326 return (NoFramePointerElim || 327 needsStackRealignment(MF) || 328 MFI->hasVarSizedObjects() || 329 MFI->isFrameAddressTaken() || 330 MF.getInfo<X86MachineFunctionInfo>()->getForceFramePointer() || 331 (MMI && MMI->callsUnwindInit())); 332} 333 334bool X86RegisterInfo::needsStackRealignment(const MachineFunction &MF) const { 335 const MachineFrameInfo *MFI = MF.getFrameInfo();; 336 337 // FIXME: Currently we don't support stack realignment for functions with 338 // variable-sized allocas 339 return (RealignStack && 340 (MFI->getMaxAlignment() > StackAlign && 341 !MFI->hasVarSizedObjects())); 342} 343 344bool X86RegisterInfo::hasReservedCallFrame(MachineFunction &MF) const { 345 return !MF.getFrameInfo()->hasVarSizedObjects(); 346} 347 348int 349X86RegisterInfo::getFrameIndexOffset(MachineFunction &MF, int FI) const { 350 int Offset = MF.getFrameInfo()->getObjectOffset(FI) + SlotSize; 351 uint64_t StackSize = MF.getFrameInfo()->getStackSize(); 352 353 if (needsStackRealignment(MF)) { 354 if (FI < 0) 355 // Skip the saved EBP 356 Offset += SlotSize; 357 else { 358 unsigned Align = MF.getFrameInfo()->getObjectAlignment(FI); 359 assert( (-(Offset + StackSize)) % Align == 0); 360 Align = 0; 361 return Offset + StackSize; 362 } 363 364 // FIXME: Support tail calls 365 } else { 366 if (!hasFP(MF)) 367 return Offset + StackSize; 368 369 // Skip the saved EBP 370 Offset += SlotSize; 371 372 // Skip the RETADDR move area 373 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 374 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta(); 375 if (TailCallReturnAddrDelta < 0) Offset -= TailCallReturnAddrDelta; 376 } 377 378 return Offset; 379} 380 381void X86RegisterInfo:: 382eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, 383 MachineBasicBlock::iterator I) const { 384 if (!hasReservedCallFrame(MF)) { 385 // If the stack pointer can be changed after prologue, turn the 386 // adjcallstackup instruction into a 'sub ESP, <amt>' and the 387 // adjcallstackdown instruction into 'add ESP, <amt>' 388 // TODO: consider using push / pop instead of sub + store / add 389 MachineInstr *Old = I; 390 uint64_t Amount = Old->getOperand(0).getImm(); 391 if (Amount != 0) { 392 // We need to keep the stack aligned properly. To do this, we round the 393 // amount of space needed for the outgoing arguments up to the next 394 // alignment boundary. 395 Amount = (Amount+StackAlign-1)/StackAlign*StackAlign; 396 397 MachineInstr *New = 0; 398 if (Old->getOpcode() == getCallFrameSetupOpcode()) { 399 New = BuildMI(MF, Old->getDebugLoc(), 400 TII.get(Is64Bit ? X86::SUB64ri32 : X86::SUB32ri), 401 StackPtr).addReg(StackPtr).addImm(Amount); 402 } else { 403 assert(Old->getOpcode() == getCallFrameDestroyOpcode()); 404 // factor out the amount the callee already popped. 405 uint64_t CalleeAmt = Old->getOperand(1).getImm(); 406 Amount -= CalleeAmt; 407 if (Amount) { 408 unsigned Opc = (Amount < 128) ? 409 (Is64Bit ? X86::ADD64ri8 : X86::ADD32ri8) : 410 (Is64Bit ? X86::ADD64ri32 : X86::ADD32ri); 411 New = BuildMI(MF, Old->getDebugLoc(), TII.get(Opc), StackPtr) 412 .addReg(StackPtr).addImm(Amount); 413 } 414 } 415 416 if (New) { 417 // The EFLAGS implicit def is dead. 418 New->getOperand(3).setIsDead(); 419 420 // Replace the pseudo instruction with a new instruction... 421 MBB.insert(I, New); 422 } 423 } 424 } else if (I->getOpcode() == getCallFrameDestroyOpcode()) { 425 // If we are performing frame pointer elimination and if the callee pops 426 // something off the stack pointer, add it back. We do this until we have 427 // more advanced stack pointer tracking ability. 428 if (uint64_t CalleeAmt = I->getOperand(1).getImm()) { 429 unsigned Opc = (CalleeAmt < 128) ? 430 (Is64Bit ? X86::SUB64ri8 : X86::SUB32ri8) : 431 (Is64Bit ? X86::SUB64ri32 : X86::SUB32ri); 432 MachineInstr *Old = I; 433 MachineInstr *New = 434 BuildMI(MF, Old->getDebugLoc(), TII.get(Opc), 435 StackPtr).addReg(StackPtr).addImm(CalleeAmt); 436 // The EFLAGS implicit def is dead. 437 New->getOperand(3).setIsDead(); 438 439 MBB.insert(I, New); 440 } 441 } 442 443 MBB.erase(I); 444} 445 446void X86RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, 447 int SPAdj, RegScavenger *RS) const{ 448 assert(SPAdj == 0 && "Unexpected"); 449 450 unsigned i = 0; 451 MachineInstr &MI = *II; 452 MachineFunction &MF = *MI.getParent()->getParent(); 453 while (!MI.getOperand(i).isFI()) { 454 ++i; 455 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!"); 456 } 457 458 int FrameIndex = MI.getOperand(i).getIndex(); 459 460 unsigned BasePtr; 461 if (needsStackRealignment(MF)) 462 BasePtr = (FrameIndex < 0 ? FramePtr : StackPtr); 463 else 464 BasePtr = (hasFP(MF) ? FramePtr : StackPtr); 465 466 // This must be part of a four operand memory reference. Replace the 467 // FrameIndex with base register with EBP. Add an offset to the offset. 468 MI.getOperand(i).ChangeToRegister(BasePtr, false); 469 470 // Now add the frame object offset to the offset from EBP. 471 if (MI.getOperand(i+3).isImm()) { 472 // Offset is a 32-bit integer. 473 int Offset = getFrameIndexOffset(MF, FrameIndex) + 474 (int)(MI.getOperand(i+3).getImm()); 475 476 MI.getOperand(i+3).ChangeToImmediate(Offset); 477 } else { 478 // Offset is symbolic. This is extremely rare. 479 uint64_t Offset = getFrameIndexOffset(MF, FrameIndex) + 480 (uint64_t)MI.getOperand(i+3).getOffset(); 481 MI.getOperand(i+3).setOffset(Offset); 482 } 483} 484 485void 486X86RegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF, 487 RegScavenger *RS) const { 488 MachineFrameInfo *FFI = MF.getFrameInfo(); 489 490 // Calculate and set max stack object alignment early, so we can decide 491 // whether we will need stack realignment (and thus FP). 492 unsigned MaxAlign = std::max(FFI->getMaxAlignment(), 493 calculateMaxStackAlignment(FFI)); 494 495 FFI->setMaxAlignment(MaxAlign); 496} 497 498void 499X86RegisterInfo::processFunctionBeforeFrameFinalized(MachineFunction &MF) const{ 500 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 501 int32_t TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta(); 502 if (TailCallReturnAddrDelta < 0) { 503 // create RETURNADDR area 504 // arg 505 // arg 506 // RETADDR 507 // { ... 508 // RETADDR area 509 // ... 510 // } 511 // [EBP] 512 MF.getFrameInfo()-> 513 CreateFixedObject(-TailCallReturnAddrDelta, 514 (-1*SlotSize)+TailCallReturnAddrDelta); 515 } 516 if (hasFP(MF)) { 517 assert((TailCallReturnAddrDelta <= 0) && 518 "The Delta should always be zero or negative"); 519 // Create a frame entry for the EBP register that must be saved. 520 int FrameIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, 521 (int)SlotSize * -2+ 522 TailCallReturnAddrDelta); 523 assert(FrameIdx == MF.getFrameInfo()->getObjectIndexBegin() && 524 "Slot for EBP register must be last in order to be found!"); 525 FrameIdx = 0; 526 } 527} 528 529/// emitSPUpdate - Emit a series of instructions to increment / decrement the 530/// stack pointer by a constant value. 531static 532void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI, 533 unsigned StackPtr, int64_t NumBytes, bool Is64Bit, 534 const TargetInstrInfo &TII) { 535 bool isSub = NumBytes < 0; 536 uint64_t Offset = isSub ? -NumBytes : NumBytes; 537 unsigned Opc = isSub 538 ? ((Offset < 128) ? 539 (Is64Bit ? X86::SUB64ri8 : X86::SUB32ri8) : 540 (Is64Bit ? X86::SUB64ri32 : X86::SUB32ri)) 541 : ((Offset < 128) ? 542 (Is64Bit ? X86::ADD64ri8 : X86::ADD32ri8) : 543 (Is64Bit ? X86::ADD64ri32 : X86::ADD32ri)); 544 uint64_t Chunk = (1LL << 31) - 1; 545 DebugLoc DL = MBBI->getDebugLoc(); 546 547 while (Offset) { 548 uint64_t ThisVal = (Offset > Chunk) ? Chunk : Offset; 549 MachineInstr *MI = 550 BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr) 551 .addReg(StackPtr).addImm(ThisVal); 552 // The EFLAGS implicit def is dead. 553 MI->getOperand(3).setIsDead(); 554 Offset -= ThisVal; 555 } 556} 557 558// mergeSPUpdatesUp - Merge two stack-manipulating instructions upper iterator. 559static 560void mergeSPUpdatesUp(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI, 561 unsigned StackPtr, uint64_t *NumBytes = NULL) { 562 if (MBBI == MBB.begin()) return; 563 564 MachineBasicBlock::iterator PI = prior(MBBI); 565 unsigned Opc = PI->getOpcode(); 566 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 || 567 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) && 568 PI->getOperand(0).getReg() == StackPtr) { 569 if (NumBytes) 570 *NumBytes += PI->getOperand(2).getImm(); 571 MBB.erase(PI); 572 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 || 573 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) && 574 PI->getOperand(0).getReg() == StackPtr) { 575 if (NumBytes) 576 *NumBytes -= PI->getOperand(2).getImm(); 577 MBB.erase(PI); 578 } 579} 580 581// mergeSPUpdatesUp - Merge two stack-manipulating instructions lower iterator. 582static 583void mergeSPUpdatesDown(MachineBasicBlock &MBB, 584 MachineBasicBlock::iterator &MBBI, 585 unsigned StackPtr, uint64_t *NumBytes = NULL) { 586 return; 587 588 if (MBBI == MBB.end()) return; 589 590 MachineBasicBlock::iterator NI = next(MBBI); 591 if (NI == MBB.end()) return; 592 593 unsigned Opc = NI->getOpcode(); 594 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 || 595 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) && 596 NI->getOperand(0).getReg() == StackPtr) { 597 if (NumBytes) 598 *NumBytes -= NI->getOperand(2).getImm(); 599 MBB.erase(NI); 600 MBBI = NI; 601 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 || 602 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) && 603 NI->getOperand(0).getReg() == StackPtr) { 604 if (NumBytes) 605 *NumBytes += NI->getOperand(2).getImm(); 606 MBB.erase(NI); 607 MBBI = NI; 608 } 609} 610 611/// mergeSPUpdates - Checks the instruction before/after the passed 612/// instruction. If it is an ADD/SUB instruction it is deleted 613/// argument and the stack adjustment is returned as a positive value for ADD 614/// and a negative for SUB. 615static int mergeSPUpdates(MachineBasicBlock &MBB, 616 MachineBasicBlock::iterator &MBBI, 617 unsigned StackPtr, 618 bool doMergeWithPrevious) { 619 620 if ((doMergeWithPrevious && MBBI == MBB.begin()) || 621 (!doMergeWithPrevious && MBBI == MBB.end())) 622 return 0; 623 624 int Offset = 0; 625 626 MachineBasicBlock::iterator PI = doMergeWithPrevious ? prior(MBBI) : MBBI; 627 MachineBasicBlock::iterator NI = doMergeWithPrevious ? 0 : next(MBBI); 628 unsigned Opc = PI->getOpcode(); 629 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 || 630 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) && 631 PI->getOperand(0).getReg() == StackPtr){ 632 Offset += PI->getOperand(2).getImm(); 633 MBB.erase(PI); 634 if (!doMergeWithPrevious) MBBI = NI; 635 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 || 636 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) && 637 PI->getOperand(0).getReg() == StackPtr) { 638 Offset -= PI->getOperand(2).getImm(); 639 MBB.erase(PI); 640 if (!doMergeWithPrevious) MBBI = NI; 641 } 642 643 return Offset; 644} 645 646void X86RegisterInfo::emitFrameMoves(MachineFunction &MF, 647 unsigned FrameLabelId, 648 unsigned ReadyLabelId) const { 649 MachineFrameInfo *MFI = MF.getFrameInfo(); 650 MachineModuleInfo *MMI = MFI->getMachineModuleInfo(); 651 if (!MMI) 652 return; 653 654 uint64_t StackSize = MFI->getStackSize(); 655 std::vector<MachineMove> &Moves = MMI->getFrameMoves(); 656 const TargetData *TD = MF.getTarget().getTargetData(); 657 658 // Calculate amount of bytes used for return address storing 659 int stackGrowth = 660 (MF.getTarget().getFrameInfo()->getStackGrowthDirection() == 661 TargetFrameInfo::StackGrowsUp ? 662 TD->getPointerSize() : -TD->getPointerSize()); 663 664 if (StackSize) { 665 // Show update of SP. 666 if (hasFP(MF)) { 667 // Adjust SP 668 MachineLocation SPDst(MachineLocation::VirtualFP); 669 MachineLocation SPSrc(MachineLocation::VirtualFP, 2*stackGrowth); 670 Moves.push_back(MachineMove(FrameLabelId, SPDst, SPSrc)); 671 } else { 672 MachineLocation SPDst(MachineLocation::VirtualFP); 673 MachineLocation SPSrc(MachineLocation::VirtualFP, 674 -StackSize+stackGrowth); 675 Moves.push_back(MachineMove(FrameLabelId, SPDst, SPSrc)); 676 } 677 } else { 678 //FIXME: Verify & implement for FP 679 MachineLocation SPDst(StackPtr); 680 MachineLocation SPSrc(StackPtr, stackGrowth); 681 Moves.push_back(MachineMove(FrameLabelId, SPDst, SPSrc)); 682 } 683 684 // Add callee saved registers to move list. 685 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo(); 686 687 // FIXME: This is dirty hack. The code itself is pretty mess right now. 688 // It should be rewritten from scratch and generalized sometimes. 689 690 // Determine maximum offset (minumum due to stack growth) 691 int64_t MaxOffset = 0; 692 for (unsigned I = 0, E = CSI.size(); I!=E; ++I) 693 MaxOffset = std::min(MaxOffset, 694 MFI->getObjectOffset(CSI[I].getFrameIdx())); 695 696 // Calculate offsets 697 int64_t saveAreaOffset = (hasFP(MF) ? 3 : 2)*stackGrowth; 698 for (unsigned I = 0, E = CSI.size(); I!=E; ++I) { 699 int64_t Offset = MFI->getObjectOffset(CSI[I].getFrameIdx()); 700 unsigned Reg = CSI[I].getReg(); 701 Offset = (MaxOffset-Offset+saveAreaOffset); 702 MachineLocation CSDst(MachineLocation::VirtualFP, Offset); 703 MachineLocation CSSrc(Reg); 704 Moves.push_back(MachineMove(FrameLabelId, CSDst, CSSrc)); 705 } 706 707 if (hasFP(MF)) { 708 // Save FP 709 MachineLocation FPDst(MachineLocation::VirtualFP, 2*stackGrowth); 710 MachineLocation FPSrc(FramePtr); 711 Moves.push_back(MachineMove(ReadyLabelId, FPDst, FPSrc)); 712 } 713 714 MachineLocation FPDst(hasFP(MF) ? FramePtr : StackPtr); 715 MachineLocation FPSrc(MachineLocation::VirtualFP); 716 Moves.push_back(MachineMove(ReadyLabelId, FPDst, FPSrc)); 717} 718 719 720void X86RegisterInfo::emitPrologue(MachineFunction &MF) const { 721 MachineBasicBlock &MBB = MF.front(); // Prolog goes in entry BB 722 MachineFrameInfo *MFI = MF.getFrameInfo(); 723 const Function* Fn = MF.getFunction(); 724 const X86Subtarget* Subtarget = &MF.getTarget().getSubtarget<X86Subtarget>(); 725 MachineModuleInfo *MMI = MFI->getMachineModuleInfo(); 726 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 727 MachineBasicBlock::iterator MBBI = MBB.begin(); 728 bool needsFrameMoves = (MMI && MMI->hasDebugInfo()) || 729 !Fn->doesNotThrow() || 730 UnwindTablesMandatory; 731 DebugLoc DL = MBBI->getDebugLoc(); 732 733 // Prepare for frame info. 734 unsigned FrameLabelId = 0; 735 736 // Get the number of bytes to allocate from the FrameInfo. 737 uint64_t StackSize = MFI->getStackSize(); 738 739 // Get desired stack alignment 740 uint64_t MaxAlign = MFI->getMaxAlignment(); 741 742 // Add RETADDR move area to callee saved frame size. 743 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta(); 744 if (TailCallReturnAddrDelta < 0) 745 X86FI->setCalleeSavedFrameSize( 746 X86FI->getCalleeSavedFrameSize() +(-TailCallReturnAddrDelta)); 747 748 // If this is x86-64 and the Red Zone is not disabled, if we are a leaf 749 // function, and use up to 128 bytes of stack space, don't have a frame 750 // pointer, calls, or dynamic alloca then we do not need to adjust the 751 // stack pointer (we fit in the Red Zone). 752 if (Is64Bit && !DisableRedZone && 753 !needsStackRealignment(MF) && 754 !MFI->hasVarSizedObjects() && // No dynamic alloca. 755 !MFI->hasCalls()) { // No calls. 756 uint64_t MinSize = X86FI->getCalleeSavedFrameSize(); 757 if (hasFP(MF)) MinSize += SlotSize; 758 StackSize = std::max(MinSize, 759 StackSize > 128 ? StackSize - 128 : 0); 760 MFI->setStackSize(StackSize); 761 } 762 763 // Insert stack pointer adjustment for later moving of return addr. Only 764 // applies to tail call optimized functions where the callee argument stack 765 // size is bigger than the callers. 766 if (TailCallReturnAddrDelta < 0) { 767 MachineInstr *MI = 768 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit? X86::SUB64ri32 : X86::SUB32ri), 769 StackPtr).addReg(StackPtr).addImm(-TailCallReturnAddrDelta); 770 // The EFLAGS implicit def is dead. 771 MI->getOperand(3).setIsDead(); 772 } 773 774 uint64_t NumBytes = 0; 775 if (hasFP(MF)) { 776 // Calculate required stack adjustment 777 uint64_t FrameSize = StackSize - SlotSize; 778 if (needsStackRealignment(MF)) 779 FrameSize = (FrameSize + MaxAlign - 1)/MaxAlign*MaxAlign; 780 781 NumBytes = FrameSize - X86FI->getCalleeSavedFrameSize(); 782 783 // Get the offset of the stack slot for the EBP register... which is 784 // guaranteed to be the last slot by processFunctionBeforeFrameFinalized. 785 // Update the frame offset adjustment. 786 MFI->setOffsetAdjustment(-NumBytes); 787 788 // Save EBP into the appropriate stack slot... 789 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::PUSH64r : X86::PUSH32r)) 790 .addReg(FramePtr, /*isDef=*/false, /*isImp=*/false, /*isKill=*/true); 791 792 if (needsFrameMoves) { 793 // Mark effective beginning of when frame pointer becomes valid. 794 FrameLabelId = MMI->NextLabelID(); 795 BuildMI(MBB, MBBI, DL, TII.get(X86::DBG_LABEL)).addImm(FrameLabelId); 796 } 797 798 // Update EBP with the new base value... 799 BuildMI(MBB, MBBI, DL, 800 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), FramePtr) 801 .addReg(StackPtr); 802 803 // Mark the FramePtr as live-in in every block except the entry. 804 for (MachineFunction::iterator I = next(MF.begin()), E = MF.end(); 805 I != E; ++I) 806 I->addLiveIn(FramePtr); 807 808 // Realign stack 809 if (needsStackRealignment(MF)) { 810 MachineInstr *MI = 811 BuildMI(MBB, MBBI, DL, 812 TII.get(Is64Bit ? X86::AND64ri32 : X86::AND32ri), 813 StackPtr).addReg(StackPtr).addImm(-MaxAlign); 814 // The EFLAGS implicit def is dead. 815 MI->getOperand(3).setIsDead(); 816 } 817 } else { 818 NumBytes = StackSize - X86FI->getCalleeSavedFrameSize(); 819 } 820 821 unsigned ReadyLabelId = 0; 822 if (needsFrameMoves) { 823 // Mark effective beginning of when frame pointer is ready. 824 ReadyLabelId = MMI->NextLabelID(); 825 BuildMI(MBB, MBBI, DL, TII.get(X86::DBG_LABEL)).addImm(ReadyLabelId); 826 } 827 828 // Skip the callee-saved push instructions. 829 while (MBBI != MBB.end() && 830 (MBBI->getOpcode() == X86::PUSH32r || 831 MBBI->getOpcode() == X86::PUSH64r)) 832 ++MBBI; 833 834 if (MBBI != MBB.end()) 835 DL = MBBI->getDebugLoc(); 836 837 if (NumBytes) { // adjust stack pointer: ESP -= numbytes 838 if (NumBytes >= 4096 && Subtarget->isTargetCygMing()) { 839 // Check, whether EAX is livein for this function 840 bool isEAXAlive = false; 841 for (MachineRegisterInfo::livein_iterator 842 II = MF.getRegInfo().livein_begin(), 843 EE = MF.getRegInfo().livein_end(); (II != EE) && !isEAXAlive; ++II) { 844 unsigned Reg = II->first; 845 isEAXAlive = (Reg == X86::EAX || Reg == X86::AX || 846 Reg == X86::AH || Reg == X86::AL); 847 } 848 849 // Function prologue calls _alloca to probe the stack when allocating 850 // more than 4k bytes in one go. Touching the stack at 4K increments is 851 // necessary to ensure that the guard pages used by the OS virtual memory 852 // manager are allocated in correct sequence. 853 if (!isEAXAlive) { 854 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX) 855 .addImm(NumBytes); 856 BuildMI(MBB, MBBI, DL, TII.get(X86::CALLpcrel32)) 857 .addExternalSymbol("_alloca"); 858 } else { 859 // Save EAX 860 BuildMI(MBB, MBBI, DL, TII.get(X86::PUSH32r)) 861 .addReg(X86::EAX, /*isDef=*/false, /*isImp=*/false, /*isKill=*/true); 862 // Allocate NumBytes-4 bytes on stack. We'll also use 4 already 863 // allocated bytes for EAX. 864 BuildMI(MBB, MBBI, DL, 865 TII.get(X86::MOV32ri), X86::EAX).addImm(NumBytes-4); 866 BuildMI(MBB, MBBI, DL, TII.get(X86::CALLpcrel32)) 867 .addExternalSymbol("_alloca"); 868 // Restore EAX 869 MachineInstr *MI = addRegOffset(BuildMI(MF, DL, TII.get(X86::MOV32rm), 870 X86::EAX), 871 StackPtr, false, NumBytes-4); 872 MBB.insert(MBBI, MI); 873 } 874 } else { 875 // If there is an SUB32ri of ESP immediately before this instruction, 876 // merge the two. This can be the case when tail call elimination is 877 // enabled and the callee has more arguments then the caller. 878 NumBytes -= mergeSPUpdates(MBB, MBBI, StackPtr, true); 879 // If there is an ADD32ri or SUB32ri of ESP immediately after this 880 // instruction, merge the two instructions. 881 mergeSPUpdatesDown(MBB, MBBI, StackPtr, &NumBytes); 882 883 if (NumBytes) 884 emitSPUpdate(MBB, MBBI, StackPtr, -(int64_t)NumBytes, Is64Bit, TII); 885 } 886 } 887 888 if (needsFrameMoves) 889 emitFrameMoves(MF, FrameLabelId, ReadyLabelId); 890} 891 892void X86RegisterInfo::emitEpilogue(MachineFunction &MF, 893 MachineBasicBlock &MBB) const { 894 const MachineFrameInfo *MFI = MF.getFrameInfo(); 895 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 896 MachineBasicBlock::iterator MBBI = prior(MBB.end()); 897 unsigned RetOpcode = MBBI->getOpcode(); 898 DebugLoc DL = MBBI->getDebugLoc(); 899 900 switch (RetOpcode) { 901 case X86::RET: 902 case X86::RETI: 903 case X86::TCRETURNdi: 904 case X86::TCRETURNri: 905 case X86::TCRETURNri64: 906 case X86::TCRETURNdi64: 907 case X86::EH_RETURN: 908 case X86::EH_RETURN64: 909 case X86::TAILJMPd: 910 case X86::TAILJMPr: 911 case X86::TAILJMPm: break; // These are ok 912 default: 913 assert(0 && "Can only insert epilog into returning blocks"); 914 } 915 916 // Get the number of bytes to allocate from the FrameInfo 917 uint64_t StackSize = MFI->getStackSize(); 918 uint64_t MaxAlign = MFI->getMaxAlignment(); 919 unsigned CSSize = X86FI->getCalleeSavedFrameSize(); 920 uint64_t NumBytes = 0; 921 922 if (hasFP(MF)) { 923 // Calculate required stack adjustment 924 uint64_t FrameSize = StackSize - SlotSize; 925 if (needsStackRealignment(MF)) 926 FrameSize = (FrameSize + MaxAlign - 1)/MaxAlign*MaxAlign; 927 928 NumBytes = FrameSize - CSSize; 929 930 // pop EBP. 931 BuildMI(MBB, MBBI, DL, 932 TII.get(Is64Bit ? X86::POP64r : X86::POP32r), FramePtr); 933 } else { 934 NumBytes = StackSize - CSSize; 935 } 936 937 // Skip the callee-saved pop instructions. 938 MachineBasicBlock::iterator LastCSPop = MBBI; 939 while (MBBI != MBB.begin()) { 940 MachineBasicBlock::iterator PI = prior(MBBI); 941 unsigned Opc = PI->getOpcode(); 942 if (Opc != X86::POP32r && Opc != X86::POP64r && 943 !PI->getDesc().isTerminator()) 944 break; 945 --MBBI; 946 } 947 948 DL = MBBI->getDebugLoc(); 949 950 // If there is an ADD32ri or SUB32ri of ESP immediately before this 951 // instruction, merge the two instructions. 952 if (NumBytes || MFI->hasVarSizedObjects()) 953 mergeSPUpdatesUp(MBB, MBBI, StackPtr, &NumBytes); 954 955 // If dynamic alloca is used, then reset esp to point to the last callee-saved 956 // slot before popping them off! Same applies for the case, when stack was 957 // realigned 958 if (needsStackRealignment(MF)) { 959 // We cannot use LEA here, because stack pointer was realigned. We need to 960 // deallocate local frame back 961 if (CSSize) { 962 emitSPUpdate(MBB, MBBI, StackPtr, NumBytes, Is64Bit, TII); 963 MBBI = prior(LastCSPop); 964 } 965 966 BuildMI(MBB, MBBI, DL, 967 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), 968 StackPtr).addReg(FramePtr); 969 } else if (MFI->hasVarSizedObjects()) { 970 if (CSSize) { 971 unsigned Opc = Is64Bit ? X86::LEA64r : X86::LEA32r; 972 MachineInstr *MI = addRegOffset(BuildMI(MF, DL, TII.get(Opc), StackPtr), 973 FramePtr, false, -CSSize); 974 MBB.insert(MBBI, MI); 975 } else 976 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), 977 StackPtr).addReg(FramePtr); 978 979 } else { 980 // adjust stack pointer back: ESP += numbytes 981 if (NumBytes) 982 emitSPUpdate(MBB, MBBI, StackPtr, NumBytes, Is64Bit, TII); 983 } 984 985 // We're returning from function via eh_return. 986 if (RetOpcode == X86::EH_RETURN || RetOpcode == X86::EH_RETURN64) { 987 MBBI = prior(MBB.end()); 988 MachineOperand &DestAddr = MBBI->getOperand(0); 989 assert(DestAddr.isReg() && "Offset should be in register!"); 990 BuildMI(MBB, MBBI, DL, 991 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), 992 StackPtr).addReg(DestAddr.getReg()); 993 // Tail call return: adjust the stack pointer and jump to callee 994 } else if (RetOpcode == X86::TCRETURNri || RetOpcode == X86::TCRETURNdi || 995 RetOpcode== X86::TCRETURNri64 || RetOpcode == X86::TCRETURNdi64) { 996 MBBI = prior(MBB.end()); 997 MachineOperand &JumpTarget = MBBI->getOperand(0); 998 MachineOperand &StackAdjust = MBBI->getOperand(1); 999 assert(StackAdjust.isImm() && "Expecting immediate value."); 1000 1001 // Adjust stack pointer. 1002 int StackAdj = StackAdjust.getImm(); 1003 int MaxTCDelta = X86FI->getTCReturnAddrDelta(); 1004 int Offset = 0; 1005 assert(MaxTCDelta <= 0 && "MaxTCDelta should never be positive"); 1006 // Incoporate the retaddr area. 1007 Offset = StackAdj-MaxTCDelta; 1008 assert(Offset >= 0 && "Offset should never be negative"); 1009 1010 if (Offset) { 1011 // Check for possible merge with preceeding ADD instruction. 1012 Offset += mergeSPUpdates(MBB, MBBI, StackPtr, true); 1013 emitSPUpdate(MBB, MBBI, StackPtr, Offset, Is64Bit, TII); 1014 } 1015 1016 // Jump to label or value in register. 1017 if (RetOpcode == X86::TCRETURNdi|| RetOpcode == X86::TCRETURNdi64) 1018 BuildMI(MBB, MBBI, DL, TII.get(X86::TAILJMPd)). 1019 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset()); 1020 else if (RetOpcode== X86::TCRETURNri64) 1021 BuildMI(MBB, MBBI, DL, TII.get(X86::TAILJMPr64), JumpTarget.getReg()); 1022 else 1023 BuildMI(MBB, MBBI, DL, TII.get(X86::TAILJMPr), JumpTarget.getReg()); 1024 1025 // Delete the pseudo instruction TCRETURN. 1026 MBB.erase(MBBI); 1027 } else if ((RetOpcode == X86::RET || RetOpcode == X86::RETI) && 1028 (X86FI->getTCReturnAddrDelta() < 0)) { 1029 // Add the return addr area delta back since we are not tail calling. 1030 int delta = -1*X86FI->getTCReturnAddrDelta(); 1031 MBBI = prior(MBB.end()); 1032 // Check for possible merge with preceeding ADD instruction. 1033 delta += mergeSPUpdates(MBB, MBBI, StackPtr, true); 1034 emitSPUpdate(MBB, MBBI, StackPtr, delta, Is64Bit, TII); 1035 } 1036} 1037 1038unsigned X86RegisterInfo::getRARegister() const { 1039 if (Is64Bit) 1040 return X86::RIP; // Should have dwarf #16 1041 else 1042 return X86::EIP; // Should have dwarf #8 1043} 1044 1045unsigned X86RegisterInfo::getFrameRegister(MachineFunction &MF) const { 1046 return hasFP(MF) ? FramePtr : StackPtr; 1047} 1048 1049void X86RegisterInfo::getInitialFrameState(std::vector<MachineMove> &Moves) 1050 const { 1051 // Calculate amount of bytes used for return address storing 1052 int stackGrowth = (Is64Bit ? -8 : -4); 1053 1054 // Initial state of the frame pointer is esp+4. 1055 MachineLocation Dst(MachineLocation::VirtualFP); 1056 MachineLocation Src(StackPtr, stackGrowth); 1057 Moves.push_back(MachineMove(0, Dst, Src)); 1058 1059 // Add return address to move list 1060 MachineLocation CSDst(StackPtr, stackGrowth); 1061 MachineLocation CSSrc(getRARegister()); 1062 Moves.push_back(MachineMove(0, CSDst, CSSrc)); 1063} 1064 1065unsigned X86RegisterInfo::getEHExceptionRegister() const { 1066 assert(0 && "What is the exception register"); 1067 return 0; 1068} 1069 1070unsigned X86RegisterInfo::getEHHandlerRegister() const { 1071 assert(0 && "What is the exception handler register"); 1072 return 0; 1073} 1074 1075namespace llvm { 1076unsigned getX86SubSuperRegister(unsigned Reg, MVT VT, bool High) { 1077 switch (VT.getSimpleVT()) { 1078 default: return Reg; 1079 case MVT::i8: 1080 if (High) { 1081 switch (Reg) { 1082 default: return 0; 1083 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 1084 return X86::AH; 1085 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 1086 return X86::DH; 1087 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 1088 return X86::CH; 1089 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 1090 return X86::BH; 1091 } 1092 } else { 1093 switch (Reg) { 1094 default: return 0; 1095 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 1096 return X86::AL; 1097 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 1098 return X86::DL; 1099 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 1100 return X86::CL; 1101 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 1102 return X86::BL; 1103 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 1104 return X86::SIL; 1105 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 1106 return X86::DIL; 1107 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 1108 return X86::BPL; 1109 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 1110 return X86::SPL; 1111 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 1112 return X86::R8B; 1113 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 1114 return X86::R9B; 1115 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 1116 return X86::R10B; 1117 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 1118 return X86::R11B; 1119 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 1120 return X86::R12B; 1121 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 1122 return X86::R13B; 1123 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 1124 return X86::R14B; 1125 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 1126 return X86::R15B; 1127 } 1128 } 1129 case MVT::i16: 1130 switch (Reg) { 1131 default: return Reg; 1132 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 1133 return X86::AX; 1134 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 1135 return X86::DX; 1136 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 1137 return X86::CX; 1138 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 1139 return X86::BX; 1140 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 1141 return X86::SI; 1142 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 1143 return X86::DI; 1144 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 1145 return X86::BP; 1146 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 1147 return X86::SP; 1148 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 1149 return X86::R8W; 1150 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 1151 return X86::R9W; 1152 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 1153 return X86::R10W; 1154 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 1155 return X86::R11W; 1156 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 1157 return X86::R12W; 1158 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 1159 return X86::R13W; 1160 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 1161 return X86::R14W; 1162 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 1163 return X86::R15W; 1164 } 1165 case MVT::i32: 1166 switch (Reg) { 1167 default: return Reg; 1168 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 1169 return X86::EAX; 1170 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 1171 return X86::EDX; 1172 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 1173 return X86::ECX; 1174 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 1175 return X86::EBX; 1176 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 1177 return X86::ESI; 1178 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 1179 return X86::EDI; 1180 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 1181 return X86::EBP; 1182 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 1183 return X86::ESP; 1184 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 1185 return X86::R8D; 1186 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 1187 return X86::R9D; 1188 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 1189 return X86::R10D; 1190 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 1191 return X86::R11D; 1192 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 1193 return X86::R12D; 1194 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 1195 return X86::R13D; 1196 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 1197 return X86::R14D; 1198 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 1199 return X86::R15D; 1200 } 1201 case MVT::i64: 1202 switch (Reg) { 1203 default: return Reg; 1204 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 1205 return X86::RAX; 1206 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 1207 return X86::RDX; 1208 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 1209 return X86::RCX; 1210 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 1211 return X86::RBX; 1212 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 1213 return X86::RSI; 1214 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 1215 return X86::RDI; 1216 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 1217 return X86::RBP; 1218 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 1219 return X86::RSP; 1220 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 1221 return X86::R8; 1222 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 1223 return X86::R9; 1224 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 1225 return X86::R10; 1226 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 1227 return X86::R11; 1228 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 1229 return X86::R12; 1230 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 1231 return X86::R13; 1232 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 1233 return X86::R14; 1234 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 1235 return X86::R15; 1236 } 1237 } 1238 1239 return Reg; 1240} 1241} 1242 1243#include "X86GenRegisterInfo.inc" 1244 1245namespace { 1246 struct VISIBILITY_HIDDEN MSAC : public MachineFunctionPass { 1247 static char ID; 1248 MSAC() : MachineFunctionPass(&ID) {} 1249 1250 virtual bool runOnMachineFunction(MachineFunction &MF) { 1251 MachineFrameInfo *FFI = MF.getFrameInfo(); 1252 MachineRegisterInfo &RI = MF.getRegInfo(); 1253 1254 // Calculate max stack alignment of all already allocated stack objects. 1255 unsigned MaxAlign = calculateMaxStackAlignment(FFI); 1256 1257 // Be over-conservative: scan over all vreg defs and find, whether vector 1258 // registers are used. If yes - there is probability, that vector register 1259 // will be spilled and thus stack needs to be aligned properly. 1260 for (unsigned RegNum = TargetRegisterInfo::FirstVirtualRegister; 1261 RegNum < RI.getLastVirtReg(); ++RegNum) 1262 MaxAlign = std::max(MaxAlign, RI.getRegClass(RegNum)->getAlignment()); 1263 1264 FFI->setMaxAlignment(MaxAlign); 1265 1266 return false; 1267 } 1268 1269 virtual const char *getPassName() const { 1270 return "X86 Maximal Stack Alignment Calculator"; 1271 } 1272 }; 1273 1274 char MSAC::ID = 0; 1275} 1276 1277FunctionPass* 1278llvm::createX86MaxStackAlignmentCalculatorPass() { return new MSAC(); } 1279