X86RegisterInfo.cpp revision 396618b43a85e12d290a90b181c6af5d7c0c5f11
1//===-- X86RegisterInfo.cpp - X86 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// 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 "X86RegisterInfo.h" 17#include "X86.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/MachineModuleInfo.h" 31#include "llvm/CodeGen/MachineRegisterInfo.h" 32#include "llvm/MC/MCAsmInfo.h" 33#include "llvm/Target/TargetFrameLowering.h" 34#include "llvm/Target/TargetInstrInfo.h" 35#include "llvm/Target/TargetMachine.h" 36#include "llvm/Target/TargetOptions.h" 37#include "llvm/ADT/BitVector.h" 38#include "llvm/ADT/STLExtras.h" 39#include "llvm/Support/ErrorHandling.h" 40#include "llvm/Support/CommandLine.h" 41 42#define GET_REGINFO_TARGET_DESC 43#include "X86GenRegisterInfo.inc" 44 45using namespace llvm; 46 47cl::opt<bool> 48ForceStackAlign("force-align-stack", 49 cl::desc("Force align the stack to the minimum alignment" 50 " needed for the function."), 51 cl::init(false), cl::Hidden); 52 53X86RegisterInfo::X86RegisterInfo(X86TargetMachine &tm, 54 const TargetInstrInfo &tii) 55 : X86GenRegisterInfo(tm.getSubtarget<X86Subtarget>().is64Bit() 56 ? X86::RIP : X86::EIP, 57 X86_MC::getDwarfRegFlavour(tm.getTargetTriple(), false), 58 X86_MC::getDwarfRegFlavour(tm.getTargetTriple(), true)), 59 TM(tm), TII(tii) { 60 X86_MC::InitLLVM2SEHRegisterMapping(this); 61 62 // Cache some information. 63 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>(); 64 Is64Bit = Subtarget->is64Bit(); 65 IsWin64 = Subtarget->isTargetWin64(); 66 67 if (Is64Bit) { 68 SlotSize = 8; 69 StackPtr = X86::RSP; 70 FramePtr = X86::RBP; 71 } else { 72 SlotSize = 4; 73 StackPtr = X86::ESP; 74 FramePtr = X86::EBP; 75 } 76} 77 78/// getCompactUnwindRegNum - This function maps the register to the number for 79/// compact unwind encoding. Return -1 if the register isn't valid. 80int X86RegisterInfo::getCompactUnwindRegNum(unsigned RegNum, bool isEH) const { 81 switch (getLLVMRegNum(RegNum, isEH)) { 82 case X86::EBX: case X86::RBX: return 1; 83 case X86::ECX: case X86::R12: return 2; 84 case X86::EDX: case X86::R13: return 3; 85 case X86::EDI: case X86::R14: return 4; 86 case X86::ESI: case X86::R15: return 5; 87 case X86::EBP: case X86::RBP: return 6; 88 } 89 90 return -1; 91} 92 93bool 94X86RegisterInfo::trackLivenessAfterRegAlloc(const MachineFunction &MF) const { 95 // Only enable when post-RA scheduling is enabled and this is needed. 96 return TM.getSubtargetImpl()->postRAScheduler(); 97} 98 99int 100X86RegisterInfo::getSEHRegNum(unsigned i) const { 101 int reg = X86_MC::getX86RegNum(i); 102 switch (i) { 103 case X86::R8: case X86::R8D: case X86::R8W: case X86::R8B: 104 case X86::R9: case X86::R9D: case X86::R9W: case X86::R9B: 105 case X86::R10: case X86::R10D: case X86::R10W: case X86::R10B: 106 case X86::R11: case X86::R11D: case X86::R11W: case X86::R11B: 107 case X86::R12: case X86::R12D: case X86::R12W: case X86::R12B: 108 case X86::R13: case X86::R13D: case X86::R13W: case X86::R13B: 109 case X86::R14: case X86::R14D: case X86::R14W: case X86::R14B: 110 case X86::R15: case X86::R15D: case X86::R15W: case X86::R15B: 111 case X86::XMM8: case X86::XMM9: case X86::XMM10: case X86::XMM11: 112 case X86::XMM12: case X86::XMM13: case X86::XMM14: case X86::XMM15: 113 case X86::YMM8: case X86::YMM9: case X86::YMM10: case X86::YMM11: 114 case X86::YMM12: case X86::YMM13: case X86::YMM14: case X86::YMM15: 115 reg += 8; 116 } 117 return reg; 118} 119 120const TargetRegisterClass * 121X86RegisterInfo::getSubClassWithSubReg(const TargetRegisterClass *RC, 122 unsigned Idx) const { 123 // The sub_8bit sub-register index is more constrained in 32-bit mode. 124 // It behaves just like the sub_8bit_hi index. 125 if (!Is64Bit && Idx == X86::sub_8bit) 126 Idx = X86::sub_8bit_hi; 127 128 // Forward to TableGen's default version. 129 return X86GenRegisterInfo::getSubClassWithSubReg(RC, Idx); 130} 131 132const TargetRegisterClass * 133X86RegisterInfo::getMatchingSuperRegClass(const TargetRegisterClass *A, 134 const TargetRegisterClass *B, 135 unsigned SubIdx) const { 136 // The sub_8bit sub-register index is more constrained in 32-bit mode. 137 if (!Is64Bit && SubIdx == X86::sub_8bit) { 138 A = X86GenRegisterInfo::getSubClassWithSubReg(A, X86::sub_8bit_hi); 139 if (!A) 140 return 0; 141 } 142 return X86GenRegisterInfo::getMatchingSuperRegClass(A, B, SubIdx); 143} 144 145const TargetRegisterClass* 146X86RegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC) const{ 147 // Don't allow super-classes of GR8_NOREX. This class is only used after 148 // extrating sub_8bit_hi sub-registers. The H sub-registers cannot be copied 149 // to the full GR8 register class in 64-bit mode, so we cannot allow the 150 // reigster class inflation. 151 // 152 // The GR8_NOREX class is always used in a way that won't be constrained to a 153 // sub-class, so sub-classes like GR8_ABCD_L are allowed to expand to the 154 // full GR8 class. 155 if (RC == &X86::GR8_NOREXRegClass) 156 return RC; 157 158 const TargetRegisterClass *Super = RC; 159 TargetRegisterClass::sc_iterator I = RC->getSuperClasses(); 160 do { 161 switch (Super->getID()) { 162 case X86::GR8RegClassID: 163 case X86::GR16RegClassID: 164 case X86::GR32RegClassID: 165 case X86::GR64RegClassID: 166 case X86::FR32RegClassID: 167 case X86::FR64RegClassID: 168 case X86::RFP32RegClassID: 169 case X86::RFP64RegClassID: 170 case X86::RFP80RegClassID: 171 case X86::VR128RegClassID: 172 case X86::VR256RegClassID: 173 // Don't return a super-class that would shrink the spill size. 174 // That can happen with the vector and float classes. 175 if (Super->getSize() == RC->getSize()) 176 return Super; 177 } 178 Super = *I++; 179 } while (Super); 180 return RC; 181} 182 183const TargetRegisterClass * 184X86RegisterInfo::getPointerRegClass(const MachineFunction &MF, unsigned Kind) 185 const { 186 switch (Kind) { 187 default: llvm_unreachable("Unexpected Kind in getPointerRegClass!"); 188 case 0: // Normal GPRs. 189 if (TM.getSubtarget<X86Subtarget>().is64Bit()) 190 return &X86::GR64RegClass; 191 return &X86::GR32RegClass; 192 case 1: // Normal GPRs except the stack pointer (for encoding reasons). 193 if (TM.getSubtarget<X86Subtarget>().is64Bit()) 194 return &X86::GR64_NOSPRegClass; 195 return &X86::GR32_NOSPRegClass; 196 case 2: // Available for tailcall (not callee-saved GPRs). 197 if (TM.getSubtarget<X86Subtarget>().isTargetWin64()) 198 return &X86::GR64_TCW64RegClass; 199 if (TM.getSubtarget<X86Subtarget>().is64Bit()) 200 return &X86::GR64_TCRegClass; 201 return &X86::GR32_TCRegClass; 202 } 203} 204 205const TargetRegisterClass * 206X86RegisterInfo::getCrossCopyRegClass(const TargetRegisterClass *RC) const { 207 if (RC == &X86::CCRRegClass) { 208 if (Is64Bit) 209 return &X86::GR64RegClass; 210 else 211 return &X86::GR32RegClass; 212 } 213 return RC; 214} 215 216unsigned 217X86RegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC, 218 MachineFunction &MF) const { 219 const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering(); 220 221 unsigned FPDiff = TFI->hasFP(MF) ? 1 : 0; 222 switch (RC->getID()) { 223 default: 224 return 0; 225 case X86::GR32RegClassID: 226 return 4 - FPDiff; 227 case X86::GR64RegClassID: 228 return 12 - FPDiff; 229 case X86::VR128RegClassID: 230 return TM.getSubtarget<X86Subtarget>().is64Bit() ? 10 : 4; 231 case X86::VR64RegClassID: 232 return 4; 233 } 234} 235 236const uint16_t * 237X86RegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { 238 bool callsEHReturn = false; 239 bool ghcCall = false; 240 241 if (MF) { 242 callsEHReturn = MF->getMMI().callsEHReturn(); 243 const Function *F = MF->getFunction(); 244 ghcCall = (F ? F->getCallingConv() == CallingConv::GHC : false); 245 } 246 247 if (ghcCall) 248 return CSR_NoRegs_SaveList; 249 if (Is64Bit) { 250 if (IsWin64) 251 return CSR_Win64_SaveList; 252 if (callsEHReturn) 253 return CSR_64EHRet_SaveList; 254 return CSR_64_SaveList; 255 } 256 if (callsEHReturn) 257 return CSR_32EHRet_SaveList; 258 return CSR_32_SaveList; 259} 260 261const uint32_t* 262X86RegisterInfo::getCallPreservedMask(CallingConv::ID CC) const { 263 if (CC == CallingConv::GHC) 264 return CSR_NoRegs_RegMask; 265 if (!Is64Bit) 266 return CSR_32_RegMask; 267 if (IsWin64) 268 return CSR_Win64_RegMask; 269 return CSR_64_RegMask; 270} 271 272BitVector X86RegisterInfo::getReservedRegs(const MachineFunction &MF) const { 273 BitVector Reserved(getNumRegs()); 274 const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering(); 275 276 // Set the stack-pointer register and its aliases as reserved. 277 Reserved.set(X86::RSP); 278 for (MCSubRegIterator I(X86::RSP, this); I.isValid(); ++I) 279 Reserved.set(*I); 280 281 // Set the instruction pointer register and its aliases as reserved. 282 Reserved.set(X86::RIP); 283 for (MCSubRegIterator I(X86::RIP, this); I.isValid(); ++I) 284 Reserved.set(*I); 285 286 // Set the frame-pointer register and its aliases as reserved if needed. 287 if (TFI->hasFP(MF)) { 288 Reserved.set(X86::RBP); 289 for (MCSubRegIterator I(X86::RBP, this); I.isValid(); ++I) 290 Reserved.set(*I); 291 } 292 293 // Mark the segment registers as reserved. 294 Reserved.set(X86::CS); 295 Reserved.set(X86::SS); 296 Reserved.set(X86::DS); 297 Reserved.set(X86::ES); 298 Reserved.set(X86::FS); 299 Reserved.set(X86::GS); 300 301 // Mark the floating point stack registers as reserved. 302 Reserved.set(X86::ST0); 303 Reserved.set(X86::ST1); 304 Reserved.set(X86::ST2); 305 Reserved.set(X86::ST3); 306 Reserved.set(X86::ST4); 307 Reserved.set(X86::ST5); 308 Reserved.set(X86::ST6); 309 Reserved.set(X86::ST7); 310 311 // Reserve the registers that only exist in 64-bit mode. 312 if (!Is64Bit) { 313 // These 8-bit registers are part of the x86-64 extension even though their 314 // super-registers are old 32-bits. 315 Reserved.set(X86::SIL); 316 Reserved.set(X86::DIL); 317 Reserved.set(X86::BPL); 318 Reserved.set(X86::SPL); 319 320 for (unsigned n = 0; n != 8; ++n) { 321 // R8, R9, ... 322 static const uint16_t GPR64[] = { 323 X86::R8, X86::R9, X86::R10, X86::R11, 324 X86::R12, X86::R13, X86::R14, X86::R15 325 }; 326 for (MCRegAliasIterator AI(GPR64[n], this, true); AI.isValid(); ++AI) 327 Reserved.set(*AI); 328 329 // XMM8, XMM9, ... 330 assert(X86::XMM15 == X86::XMM8+7); 331 for (MCRegAliasIterator AI(X86::XMM8 + n, this, true); AI.isValid(); ++AI) 332 Reserved.set(*AI); 333 } 334 } 335 336 return Reserved; 337} 338 339//===----------------------------------------------------------------------===// 340// Stack Frame Processing methods 341//===----------------------------------------------------------------------===// 342 343bool X86RegisterInfo::canRealignStack(const MachineFunction &MF) const { 344 const MachineFrameInfo *MFI = MF.getFrameInfo(); 345 return (MF.getTarget().Options.RealignStack && 346 !MFI->hasVarSizedObjects()); 347} 348 349bool X86RegisterInfo::needsStackRealignment(const MachineFunction &MF) const { 350 const MachineFrameInfo *MFI = MF.getFrameInfo(); 351 const Function *F = MF.getFunction(); 352 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment(); 353 bool requiresRealignment = ((MFI->getMaxAlignment() > StackAlign) || 354 F->hasFnAttr(Attribute::StackAlignment)); 355 356 // FIXME: Currently we don't support stack realignment for functions with 357 // variable-sized allocas. 358 // FIXME: It's more complicated than this... 359 if (0 && requiresRealignment && MFI->hasVarSizedObjects()) 360 report_fatal_error( 361 "Stack realignment in presence of dynamic allocas is not supported"); 362 363 // If we've requested that we force align the stack do so now. 364 if (ForceStackAlign) 365 return canRealignStack(MF); 366 367 return requiresRealignment && canRealignStack(MF); 368} 369 370bool X86RegisterInfo::hasReservedSpillSlot(const MachineFunction &MF, 371 unsigned Reg, int &FrameIdx) const { 372 const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering(); 373 374 if (Reg == FramePtr && TFI->hasFP(MF)) { 375 FrameIdx = MF.getFrameInfo()->getObjectIndexBegin(); 376 return true; 377 } 378 return false; 379} 380 381static unsigned getSUBriOpcode(unsigned is64Bit, int64_t Imm) { 382 if (is64Bit) { 383 if (isInt<8>(Imm)) 384 return X86::SUB64ri8; 385 return X86::SUB64ri32; 386 } else { 387 if (isInt<8>(Imm)) 388 return X86::SUB32ri8; 389 return X86::SUB32ri; 390 } 391} 392 393static unsigned getADDriOpcode(unsigned is64Bit, int64_t Imm) { 394 if (is64Bit) { 395 if (isInt<8>(Imm)) 396 return X86::ADD64ri8; 397 return X86::ADD64ri32; 398 } else { 399 if (isInt<8>(Imm)) 400 return X86::ADD32ri8; 401 return X86::ADD32ri; 402 } 403} 404 405void X86RegisterInfo:: 406eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, 407 MachineBasicBlock::iterator I) const { 408 const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering(); 409 bool reseveCallFrame = TFI->hasReservedCallFrame(MF); 410 int Opcode = I->getOpcode(); 411 bool isDestroy = Opcode == TII.getCallFrameDestroyOpcode(); 412 DebugLoc DL = I->getDebugLoc(); 413 uint64_t Amount = !reseveCallFrame ? I->getOperand(0).getImm() : 0; 414 uint64_t CalleeAmt = isDestroy ? I->getOperand(1).getImm() : 0; 415 I = MBB.erase(I); 416 417 if (!reseveCallFrame) { 418 // If the stack pointer can be changed after prologue, turn the 419 // adjcallstackup instruction into a 'sub ESP, <amt>' and the 420 // adjcallstackdown instruction into 'add ESP, <amt>' 421 // TODO: consider using push / pop instead of sub + store / add 422 if (Amount == 0) 423 return; 424 425 // We need to keep the stack aligned properly. To do this, we round the 426 // amount of space needed for the outgoing arguments up to the next 427 // alignment boundary. 428 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment(); 429 Amount = (Amount + StackAlign - 1) / StackAlign * StackAlign; 430 431 MachineInstr *New = 0; 432 if (Opcode == TII.getCallFrameSetupOpcode()) { 433 New = BuildMI(MF, DL, TII.get(getSUBriOpcode(Is64Bit, Amount)), 434 StackPtr) 435 .addReg(StackPtr) 436 .addImm(Amount); 437 } else { 438 assert(Opcode == TII.getCallFrameDestroyOpcode()); 439 440 // Factor out the amount the callee already popped. 441 Amount -= CalleeAmt; 442 443 if (Amount) { 444 unsigned Opc = getADDriOpcode(Is64Bit, Amount); 445 New = BuildMI(MF, DL, TII.get(Opc), StackPtr) 446 .addReg(StackPtr).addImm(Amount); 447 } 448 } 449 450 if (New) { 451 // The EFLAGS implicit def is dead. 452 New->getOperand(3).setIsDead(); 453 454 // Replace the pseudo instruction with a new instruction. 455 MBB.insert(I, New); 456 } 457 458 return; 459 } 460 461 if (Opcode == TII.getCallFrameDestroyOpcode() && CalleeAmt) { 462 // If we are performing frame pointer elimination and if the callee pops 463 // something off the stack pointer, add it back. We do this until we have 464 // more advanced stack pointer tracking ability. 465 unsigned Opc = getSUBriOpcode(Is64Bit, CalleeAmt); 466 MachineInstr *New = BuildMI(MF, DL, TII.get(Opc), StackPtr) 467 .addReg(StackPtr).addImm(CalleeAmt); 468 469 // The EFLAGS implicit def is dead. 470 New->getOperand(3).setIsDead(); 471 472 // We are not tracking the stack pointer adjustment by the callee, so make 473 // sure we restore the stack pointer immediately after the call, there may 474 // be spill code inserted between the CALL and ADJCALLSTACKUP instructions. 475 MachineBasicBlock::iterator B = MBB.begin(); 476 while (I != B && !llvm::prior(I)->isCall()) 477 --I; 478 MBB.insert(I, New); 479 } 480} 481 482void 483X86RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, 484 int SPAdj, RegScavenger *RS) const{ 485 assert(SPAdj == 0 && "Unexpected"); 486 487 unsigned i = 0; 488 MachineInstr &MI = *II; 489 MachineFunction &MF = *MI.getParent()->getParent(); 490 const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering(); 491 492 while (!MI.getOperand(i).isFI()) { 493 ++i; 494 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!"); 495 } 496 497 int FrameIndex = MI.getOperand(i).getIndex(); 498 unsigned BasePtr; 499 500 unsigned Opc = MI.getOpcode(); 501 bool AfterFPPop = Opc == X86::TAILJMPm64 || Opc == X86::TAILJMPm; 502 if (needsStackRealignment(MF)) 503 BasePtr = (FrameIndex < 0 ? FramePtr : StackPtr); 504 else if (AfterFPPop) 505 BasePtr = StackPtr; 506 else 507 BasePtr = (TFI->hasFP(MF) ? FramePtr : StackPtr); 508 509 // This must be part of a four operand memory reference. Replace the 510 // FrameIndex with base register with EBP. Add an offset to the offset. 511 MI.getOperand(i).ChangeToRegister(BasePtr, false); 512 513 // Now add the frame object offset to the offset from EBP. 514 int FIOffset; 515 if (AfterFPPop) { 516 // Tail call jmp happens after FP is popped. 517 const MachineFrameInfo *MFI = MF.getFrameInfo(); 518 FIOffset = MFI->getObjectOffset(FrameIndex) - TFI->getOffsetOfLocalArea(); 519 } else 520 FIOffset = TFI->getFrameIndexOffset(MF, FrameIndex); 521 522 if (MI.getOperand(i+3).isImm()) { 523 // Offset is a 32-bit integer. 524 int Imm = (int)(MI.getOperand(i + 3).getImm()); 525 int Offset = FIOffset + Imm; 526 assert((!Is64Bit || isInt<32>((long long)FIOffset + Imm)) && 527 "Requesting 64-bit offset in 32-bit immediate!"); 528 MI.getOperand(i + 3).ChangeToImmediate(Offset); 529 } else { 530 // Offset is symbolic. This is extremely rare. 531 uint64_t Offset = FIOffset + (uint64_t)MI.getOperand(i+3).getOffset(); 532 MI.getOperand(i+3).setOffset(Offset); 533 } 534} 535 536unsigned X86RegisterInfo::getFrameRegister(const MachineFunction &MF) const { 537 const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering(); 538 return TFI->hasFP(MF) ? FramePtr : StackPtr; 539} 540 541unsigned X86RegisterInfo::getEHExceptionRegister() const { 542 llvm_unreachable("What is the exception register"); 543} 544 545unsigned X86RegisterInfo::getEHHandlerRegister() const { 546 llvm_unreachable("What is the exception handler register"); 547} 548 549namespace llvm { 550unsigned getX86SubSuperRegister(unsigned Reg, EVT VT, bool High) { 551 switch (VT.getSimpleVT().SimpleTy) { 552 default: return Reg; 553 case MVT::i8: 554 if (High) { 555 switch (Reg) { 556 default: return getX86SubSuperRegister(Reg, MVT::i64, High); 557 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 558 return X86::AH; 559 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 560 return X86::DH; 561 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 562 return X86::CH; 563 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 564 return X86::BH; 565 } 566 } else { 567 switch (Reg) { 568 default: return 0; 569 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 570 return X86::AL; 571 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 572 return X86::DL; 573 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 574 return X86::CL; 575 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 576 return X86::BL; 577 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 578 return X86::SIL; 579 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 580 return X86::DIL; 581 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 582 return X86::BPL; 583 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 584 return X86::SPL; 585 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 586 return X86::R8B; 587 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 588 return X86::R9B; 589 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 590 return X86::R10B; 591 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 592 return X86::R11B; 593 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 594 return X86::R12B; 595 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 596 return X86::R13B; 597 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 598 return X86::R14B; 599 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 600 return X86::R15B; 601 } 602 } 603 case MVT::i16: 604 switch (Reg) { 605 default: return Reg; 606 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 607 return X86::AX; 608 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 609 return X86::DX; 610 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 611 return X86::CX; 612 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 613 return X86::BX; 614 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 615 return X86::SI; 616 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 617 return X86::DI; 618 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 619 return X86::BP; 620 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 621 return X86::SP; 622 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 623 return X86::R8W; 624 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 625 return X86::R9W; 626 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 627 return X86::R10W; 628 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 629 return X86::R11W; 630 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 631 return X86::R12W; 632 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 633 return X86::R13W; 634 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 635 return X86::R14W; 636 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 637 return X86::R15W; 638 } 639 case MVT::i32: 640 switch (Reg) { 641 default: return Reg; 642 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 643 return X86::EAX; 644 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 645 return X86::EDX; 646 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 647 return X86::ECX; 648 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 649 return X86::EBX; 650 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 651 return X86::ESI; 652 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 653 return X86::EDI; 654 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 655 return X86::EBP; 656 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 657 return X86::ESP; 658 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 659 return X86::R8D; 660 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 661 return X86::R9D; 662 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 663 return X86::R10D; 664 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 665 return X86::R11D; 666 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 667 return X86::R12D; 668 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 669 return X86::R13D; 670 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 671 return X86::R14D; 672 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 673 return X86::R15D; 674 } 675 case MVT::i64: 676 // For 64-bit mode if we've requested a "high" register and the 677 // Q or r constraints we want one of these high registers or 678 // just the register name otherwise. 679 if (High) { 680 switch (Reg) { 681 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 682 return X86::SI; 683 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 684 return X86::DI; 685 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 686 return X86::BP; 687 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 688 return X86::SP; 689 // Fallthrough. 690 } 691 } 692 switch (Reg) { 693 default: return Reg; 694 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX: 695 return X86::RAX; 696 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX: 697 return X86::RDX; 698 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX: 699 return X86::RCX; 700 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX: 701 return X86::RBX; 702 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI: 703 return X86::RSI; 704 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI: 705 return X86::RDI; 706 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP: 707 return X86::RBP; 708 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP: 709 return X86::RSP; 710 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8: 711 return X86::R8; 712 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9: 713 return X86::R9; 714 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10: 715 return X86::R10; 716 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11: 717 return X86::R11; 718 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12: 719 return X86::R12; 720 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13: 721 return X86::R13; 722 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14: 723 return X86::R14; 724 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15: 725 return X86::R15; 726 } 727 } 728} 729} 730 731namespace { 732 struct MSAH : public MachineFunctionPass { 733 static char ID; 734 MSAH() : MachineFunctionPass(ID) {} 735 736 virtual bool runOnMachineFunction(MachineFunction &MF) { 737 const X86TargetMachine *TM = 738 static_cast<const X86TargetMachine *>(&MF.getTarget()); 739 const TargetFrameLowering *TFI = TM->getFrameLowering(); 740 MachineRegisterInfo &RI = MF.getRegInfo(); 741 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); 742 unsigned StackAlignment = TFI->getStackAlignment(); 743 744 // Be over-conservative: scan over all vreg defs and find whether vector 745 // registers are used. If yes, there is a possibility that vector register 746 // will be spilled and thus require dynamic stack realignment. 747 for (unsigned i = 0, e = RI.getNumVirtRegs(); i != e; ++i) { 748 unsigned Reg = TargetRegisterInfo::index2VirtReg(i); 749 if (RI.getRegClass(Reg)->getAlignment() > StackAlignment) { 750 FuncInfo->setForceFramePointer(true); 751 return true; 752 } 753 } 754 // Nothing to do 755 return false; 756 } 757 758 virtual const char *getPassName() const { 759 return "X86 Maximal Stack Alignment Check"; 760 } 761 762 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 763 AU.setPreservesCFG(); 764 MachineFunctionPass::getAnalysisUsage(AU); 765 } 766 }; 767 768 char MSAH::ID = 0; 769} 770 771FunctionPass* 772llvm::createX86MaxStackAlignmentHeuristicPass() { return new MSAH(); } 773