1 //===- X86RegisterInfo.cpp - X86 Register Information -----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the X86 implementation of the TargetRegisterInfo class.
11 // This file is responsible for the frame pointer elimination optimization
14 //===----------------------------------------------------------------------===//
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/MC/MCAsmInfo.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/ErrorHandling.h"
43 X86RegisterInfo::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),
52 // Cache some information.
53 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
54 Is64Bit = Subtarget->is64Bit();
55 IsWin64 = Subtarget->isTargetWin64();
56 StackAlign = TM.getFrameInfo()->getStackAlignment();
69 /// getDwarfRegNum - This function maps LLVM register identifiers to the DWARF
70 /// specific numbering, used in debug info and exception tables.
71 int X86RegisterInfo::getDwarfRegNum(unsigned RegNo, bool isEH) const {
72 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
73 unsigned Flavour = DWARFFlavour::X86_64;
75 if (!Subtarget->is64Bit()) {
76 if (Subtarget->isTargetDarwin()) {
78 Flavour = DWARFFlavour::X86_32_DarwinEH;
80 Flavour = DWARFFlavour::X86_32_Generic;
81 } else if (Subtarget->isTargetCygMing()) {
82 // Unsupported by now, just quick fallback
83 Flavour = DWARFFlavour::X86_32_Generic;
85 Flavour = DWARFFlavour::X86_32_Generic;
89 return X86GenRegisterInfo::getDwarfRegNumFull(RegNo, Flavour);
92 /// getX86RegNum - This function maps LLVM register identifiers to their X86
93 /// specific numbering, which is used in various places encoding instructions.
94 unsigned X86RegisterInfo::getX86RegNum(unsigned 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:
102 case X86::RBP: case X86::EBP: case X86::BP: case X86::BPL: case X86::CH:
104 case X86::RSI: case X86::ESI: case X86::SI: case X86::SIL: case X86::DH:
106 case X86::RDI: case X86::EDI: case X86::DI: case X86::DIL: case X86::BH:
109 case X86::R8: case X86::R8D: case X86::R8W: case X86::R8B:
111 case X86::R9: case X86::R9D: case X86::R9W: case X86::R9B:
113 case X86::R10: case X86::R10D: case X86::R10W: case X86::R10B:
115 case X86::R11: case X86::R11D: case X86::R11W: case X86::R11B:
117 case X86::R12: case X86::R12D: case X86::R12W: case X86::R12B:
119 case X86::R13: case X86::R13D: case X86::R13W: case X86::R13B:
121 case X86::R14: case X86::R14D: case X86::R14W: case X86::R14B:
123 case X86::R15: case X86::R15D: case X86::R15W: case X86::R15B:
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;
130 case X86::XMM0: case X86::XMM8:
131 case X86::YMM0: case X86::YMM8: case X86::MM0:
133 case X86::XMM1: case X86::XMM9:
134 case X86::YMM1: case X86::YMM9: case X86::MM1:
136 case X86::XMM2: case X86::XMM10:
137 case X86::YMM2: case X86::YMM10: case X86::MM2:
139 case X86::XMM3: case X86::XMM11:
140 case X86::YMM3: case X86::YMM11: case X86::MM3:
142 case X86::XMM4: case X86::XMM12:
143 case X86::YMM4: case X86::YMM12: case X86::MM4:
145 case X86::XMM5: case X86::XMM13:
146 case X86::YMM5: case X86::YMM13: case X86::MM5:
148 case X86::XMM6: case X86::XMM14:
149 case X86::YMM6: case X86::YMM14: case X86::MM6:
151 case X86::XMM7: case X86::XMM15:
152 case X86::YMM7: case X86::YMM15: case X86::MM7:
197 assert(isVirtualRegister(RegNo) && "Unknown physical register!");
198 llvm_unreachable("Register allocator hasn't allocated reg correctly yet!");
203 const TargetRegisterClass *
204 X86RegisterInfo::getMatchingSuperRegClass(const TargetRegisterClass *A,
205 const TargetRegisterClass *B,
206 unsigned SubIdx) const {
210 if (B == &X86::GR8RegClass) {
211 if (A->getSize() == 2 || A->getSize() == 4 || A->getSize() == 8)
213 } else if (B == &X86::GR8_ABCD_LRegClass || B == &X86::GR8_ABCD_HRegClass) {
214 if (A == &X86::GR64RegClass || A == &X86::GR64_ABCDRegClass ||
215 A == &X86::GR64_NOREXRegClass ||
216 A == &X86::GR64_NOSPRegClass ||
217 A == &X86::GR64_NOREX_NOSPRegClass)
218 return &X86::GR64_ABCDRegClass;
219 else if (A == &X86::GR32RegClass || A == &X86::GR32_ABCDRegClass ||
220 A == &X86::GR32_NOREXRegClass ||
221 A == &X86::GR32_NOSPRegClass)
222 return &X86::GR32_ABCDRegClass;
223 else if (A == &X86::GR16RegClass || A == &X86::GR16_ABCDRegClass ||
224 A == &X86::GR16_NOREXRegClass)
225 return &X86::GR16_ABCDRegClass;
226 } else if (B == &X86::GR8_NOREXRegClass) {
227 if (A == &X86::GR64RegClass || A == &X86::GR64_NOREXRegClass ||
228 A == &X86::GR64_NOSPRegClass || A == &X86::GR64_NOREX_NOSPRegClass)
229 return &X86::GR64_NOREXRegClass;
230 else if (A == &X86::GR64_ABCDRegClass)
231 return &X86::GR64_ABCDRegClass;
232 else if (A == &X86::GR32RegClass || A == &X86::GR32_NOREXRegClass ||
233 A == &X86::GR32_NOSPRegClass)
234 return &X86::GR32_NOREXRegClass;
235 else if (A == &X86::GR32_ABCDRegClass)
236 return &X86::GR32_ABCDRegClass;
237 else if (A == &X86::GR16RegClass || A == &X86::GR16_NOREXRegClass)
238 return &X86::GR16_NOREXRegClass;
239 else if (A == &X86::GR16_ABCDRegClass)
240 return &X86::GR16_ABCDRegClass;
243 case X86::sub_8bit_hi:
244 if (B == &X86::GR8_ABCD_HRegClass) {
245 if (A == &X86::GR64RegClass || A == &X86::GR64_ABCDRegClass ||
246 A == &X86::GR64_NOREXRegClass ||
247 A == &X86::GR64_NOSPRegClass ||
248 A == &X86::GR64_NOREX_NOSPRegClass)
249 return &X86::GR64_ABCDRegClass;
250 else if (A == &X86::GR32RegClass || A == &X86::GR32_ABCDRegClass ||
251 A == &X86::GR32_NOREXRegClass || A == &X86::GR32_NOSPRegClass)
252 return &X86::GR32_ABCDRegClass;
253 else if (A == &X86::GR16RegClass || A == &X86::GR16_ABCDRegClass ||
254 A == &X86::GR16_NOREXRegClass)
255 return &X86::GR16_ABCDRegClass;
259 if (B == &X86::GR16RegClass) {
260 if (A->getSize() == 4 || A->getSize() == 8)
262 } else if (B == &X86::GR16_ABCDRegClass) {
263 if (A == &X86::GR64RegClass || A == &X86::GR64_ABCDRegClass ||
264 A == &X86::GR64_NOREXRegClass ||
265 A == &X86::GR64_NOSPRegClass ||
266 A == &X86::GR64_NOREX_NOSPRegClass)
267 return &X86::GR64_ABCDRegClass;
268 else if (A == &X86::GR32RegClass || A == &X86::GR32_ABCDRegClass ||
269 A == &X86::GR32_NOREXRegClass || A == &X86::GR32_NOSPRegClass)
270 return &X86::GR32_ABCDRegClass;
271 } else if (B == &X86::GR16_NOREXRegClass) {
272 if (A == &X86::GR64RegClass || A == &X86::GR64_NOREXRegClass ||
273 A == &X86::GR64_NOSPRegClass || A == &X86::GR64_NOREX_NOSPRegClass)
274 return &X86::GR64_NOREXRegClass;
275 else if (A == &X86::GR64_ABCDRegClass)
276 return &X86::GR64_ABCDRegClass;
277 else if (A == &X86::GR32RegClass || A == &X86::GR32_NOREXRegClass ||
278 A == &X86::GR32_NOSPRegClass)
279 return &X86::GR32_NOREXRegClass;
280 else if (A == &X86::GR32_ABCDRegClass)
281 return &X86::GR64_ABCDRegClass;
285 if (B == &X86::GR32RegClass || B == &X86::GR32_NOSPRegClass) {
286 if (A->getSize() == 8)
288 } else if (B == &X86::GR32_ABCDRegClass) {
289 if (A == &X86::GR64RegClass || A == &X86::GR64_ABCDRegClass ||
290 A == &X86::GR64_NOREXRegClass ||
291 A == &X86::GR64_NOSPRegClass ||
292 A == &X86::GR64_NOREX_NOSPRegClass)
293 return &X86::GR64_ABCDRegClass;
294 } else if (B == &X86::GR32_NOREXRegClass) {
295 if (A == &X86::GR64RegClass || A == &X86::GR64_NOREXRegClass ||
296 A == &X86::GR64_NOSPRegClass || A == &X86::GR64_NOREX_NOSPRegClass)
297 return &X86::GR64_NOREXRegClass;
298 else if (A == &X86::GR64_ABCDRegClass)
299 return &X86::GR64_ABCDRegClass;
303 if (B == &X86::FR32RegClass)
307 if (B == &X86::FR64RegClass)
311 if (B == &X86::VR128RegClass)
318 const TargetRegisterClass *
319 X86RegisterInfo::getPointerRegClass(unsigned Kind) const {
321 default: llvm_unreachable("Unexpected Kind in getPointerRegClass!");
322 case 0: // Normal GPRs.
323 if (TM.getSubtarget<X86Subtarget>().is64Bit())
324 return &X86::GR64RegClass;
325 return &X86::GR32RegClass;
326 case 1: // Normal GRPs except the stack pointer (for encoding reasons).
327 if (TM.getSubtarget<X86Subtarget>().is64Bit())
328 return &X86::GR64_NOSPRegClass;
329 return &X86::GR32_NOSPRegClass;
333 const TargetRegisterClass *
334 X86RegisterInfo::getCrossCopyRegClass(const TargetRegisterClass *RC) const {
335 if (RC == &X86::CCRRegClass) {
337 return &X86::GR64RegClass;
339 return &X86::GR32RegClass;
345 X86RegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
346 bool callsEHReturn = false;
347 bool ghcCall = false;
350 callsEHReturn = MF->getMMI().callsEHReturn();
351 const Function *F = MF->getFunction();
352 ghcCall = (F ? F->getCallingConv() == CallingConv::GHC : false);
355 static const unsigned GhcCalleeSavedRegs[] = {
359 static const unsigned CalleeSavedRegs32Bit[] = {
360 X86::ESI, X86::EDI, X86::EBX, X86::EBP, 0
363 static const unsigned CalleeSavedRegs32EHRet[] = {
364 X86::EAX, X86::EDX, X86::ESI, X86::EDI, X86::EBX, X86::EBP, 0
367 static const unsigned CalleeSavedRegs64Bit[] = {
368 X86::RBX, X86::R12, X86::R13, X86::R14, X86::R15, X86::RBP, 0
371 static const unsigned CalleeSavedRegs64EHRet[] = {
372 X86::RAX, X86::RDX, X86::RBX, X86::R12,
373 X86::R13, X86::R14, X86::R15, X86::RBP, 0
376 static const unsigned CalleeSavedRegsWin64[] = {
377 X86::RBX, X86::RBP, X86::RDI, X86::RSI,
378 X86::R12, X86::R13, X86::R14, X86::R15,
379 X86::XMM6, X86::XMM7, X86::XMM8, X86::XMM9,
380 X86::XMM10, X86::XMM11, X86::XMM12, X86::XMM13,
381 X86::XMM14, X86::XMM15, 0
385 return GhcCalleeSavedRegs;
386 } else if (Is64Bit) {
388 return CalleeSavedRegsWin64;
390 return (callsEHReturn ? CalleeSavedRegs64EHRet : CalleeSavedRegs64Bit);
392 return (callsEHReturn ? CalleeSavedRegs32EHRet : CalleeSavedRegs32Bit);
396 BitVector X86RegisterInfo::getReservedRegs(const MachineFunction &MF) const {
397 BitVector Reserved(getNumRegs());
398 // Set the stack-pointer register and its aliases as reserved.
399 Reserved.set(X86::RSP);
400 Reserved.set(X86::ESP);
401 Reserved.set(X86::SP);
402 Reserved.set(X86::SPL);
404 // Set the instruction pointer register and its aliases as reserved.
405 Reserved.set(X86::RIP);
406 Reserved.set(X86::EIP);
407 Reserved.set(X86::IP);
409 // Set the frame-pointer register and its aliases as reserved if needed.
411 Reserved.set(X86::RBP);
412 Reserved.set(X86::EBP);
413 Reserved.set(X86::BP);
414 Reserved.set(X86::BPL);
417 // Mark the x87 stack registers as reserved, since they don't behave normally
418 // with respect to liveness. We don't fully model the effects of x87 stack
419 // pushes and pops after stackification.
420 Reserved.set(X86::ST0);
421 Reserved.set(X86::ST1);
422 Reserved.set(X86::ST2);
423 Reserved.set(X86::ST3);
424 Reserved.set(X86::ST4);
425 Reserved.set(X86::ST5);
426 Reserved.set(X86::ST6);
427 Reserved.set(X86::ST7);
431 //===----------------------------------------------------------------------===//
432 // Stack Frame Processing methods
433 //===----------------------------------------------------------------------===//
435 /// hasFP - Return true if the specified function should have a dedicated frame
436 /// pointer register. This is true if the function has variable sized allocas
437 /// or if frame pointer elimination is disabled.
438 bool X86RegisterInfo::hasFP(const MachineFunction &MF) const {
439 const MachineFrameInfo *MFI = MF.getFrameInfo();
440 const MachineModuleInfo &MMI = MF.getMMI();
442 return (DisableFramePointerElim(MF) ||
443 needsStackRealignment(MF) ||
444 MFI->hasVarSizedObjects() ||
445 MFI->isFrameAddressTaken() ||
446 MF.getInfo<X86MachineFunctionInfo>()->getForceFramePointer() ||
447 MMI.callsUnwindInit());
450 bool X86RegisterInfo::canRealignStack(const MachineFunction &MF) const {
451 const MachineFrameInfo *MFI = MF.getFrameInfo();
452 return (RealignStack &&
453 !MFI->hasVarSizedObjects());
456 bool X86RegisterInfo::needsStackRealignment(const MachineFunction &MF) const {
457 const MachineFrameInfo *MFI = MF.getFrameInfo();
458 const Function *F = MF.getFunction();
459 bool requiresRealignment =
460 RealignStack && ((MFI->getMaxAlignment() > StackAlign) ||
461 F->hasFnAttr(Attribute::StackAlignment));
463 // FIXME: Currently we don't support stack realignment for functions with
464 // variable-sized allocas.
465 // FIXME: Temporary disable the error - it seems to be too conservative.
466 if (0 && requiresRealignment && MFI->hasVarSizedObjects())
468 "Stack realignment in presense of dynamic allocas is not supported");
470 return (requiresRealignment && !MFI->hasVarSizedObjects());
473 bool X86RegisterInfo::hasReservedCallFrame(MachineFunction &MF) const {
474 return !MF.getFrameInfo()->hasVarSizedObjects();
477 bool X86RegisterInfo::hasReservedSpillSlot(MachineFunction &MF, unsigned Reg,
478 int &FrameIdx) const {
479 if (Reg == FramePtr && hasFP(MF)) {
480 FrameIdx = MF.getFrameInfo()->getObjectIndexBegin();
487 X86RegisterInfo::getFrameIndexOffset(const MachineFunction &MF, int FI) const {
488 const TargetFrameInfo &TFI = *MF.getTarget().getFrameInfo();
489 const MachineFrameInfo *MFI = MF.getFrameInfo();
490 int Offset = MFI->getObjectOffset(FI) - TFI.getOffsetOfLocalArea();
491 uint64_t StackSize = MFI->getStackSize();
493 if (needsStackRealignment(MF)) {
495 // Skip the saved EBP.
498 unsigned Align = MFI->getObjectAlignment(FI);
499 assert((-(Offset + StackSize)) % Align == 0);
501 return Offset + StackSize;
503 // FIXME: Support tail calls
506 return Offset + StackSize;
508 // Skip the saved EBP.
511 // Skip the RETADDR move area
512 const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
513 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
514 if (TailCallReturnAddrDelta < 0)
515 Offset -= TailCallReturnAddrDelta;
521 static unsigned getSUBriOpcode(unsigned is64Bit, int64_t Imm) {
524 return X86::SUB64ri8;
525 return X86::SUB64ri32;
528 return X86::SUB32ri8;
533 static unsigned getADDriOpcode(unsigned is64Bit, int64_t Imm) {
536 return X86::ADD64ri8;
537 return X86::ADD64ri32;
540 return X86::ADD32ri8;
545 void X86RegisterInfo::
546 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
547 MachineBasicBlock::iterator I) const {
548 if (!hasReservedCallFrame(MF)) {
549 // If the stack pointer can be changed after prologue, turn the
550 // adjcallstackup instruction into a 'sub ESP, <amt>' and the
551 // adjcallstackdown instruction into 'add ESP, <amt>'
552 // TODO: consider using push / pop instead of sub + store / add
553 MachineInstr *Old = I;
554 uint64_t Amount = Old->getOperand(0).getImm();
556 // We need to keep the stack aligned properly. To do this, we round the
557 // amount of space needed for the outgoing arguments up to the next
558 // alignment boundary.
559 Amount = (Amount + StackAlign - 1) / StackAlign * StackAlign;
561 MachineInstr *New = 0;
562 if (Old->getOpcode() == getCallFrameSetupOpcode()) {
563 New = BuildMI(MF, Old->getDebugLoc(),
564 TII.get(getSUBriOpcode(Is64Bit, Amount)),
569 assert(Old->getOpcode() == getCallFrameDestroyOpcode());
571 // Factor out the amount the callee already popped.
572 uint64_t CalleeAmt = Old->getOperand(1).getImm();
576 unsigned Opc = getADDriOpcode(Is64Bit, Amount);
577 New = BuildMI(MF, Old->getDebugLoc(), TII.get(Opc), StackPtr)
584 // The EFLAGS implicit def is dead.
585 New->getOperand(3).setIsDead();
587 // Replace the pseudo instruction with a new instruction.
591 } else if (I->getOpcode() == getCallFrameDestroyOpcode()) {
592 // If we are performing frame pointer elimination and if the callee pops
593 // something off the stack pointer, add it back. We do this until we have
594 // more advanced stack pointer tracking ability.
595 if (uint64_t CalleeAmt = I->getOperand(1).getImm()) {
596 unsigned Opc = getSUBriOpcode(Is64Bit, CalleeAmt);
597 MachineInstr *Old = I;
599 BuildMI(MF, Old->getDebugLoc(), TII.get(Opc),
604 // The EFLAGS implicit def is dead.
605 New->getOperand(3).setIsDead();
614 X86RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
615 int SPAdj, FrameIndexValue *Value,
616 RegScavenger *RS) const{
617 assert(SPAdj == 0 && "Unexpected");
620 MachineInstr &MI = *II;
621 MachineFunction &MF = *MI.getParent()->getParent();
623 while (!MI.getOperand(i).isFI()) {
625 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
628 int FrameIndex = MI.getOperand(i).getIndex();
631 unsigned Opc = MI.getOpcode();
632 bool AfterFPPop = Opc == X86::TAILJMPm64 || Opc == X86::TAILJMPm;
633 if (needsStackRealignment(MF))
634 BasePtr = (FrameIndex < 0 ? FramePtr : StackPtr);
638 BasePtr = (hasFP(MF) ? FramePtr : StackPtr);
640 // This must be part of a four operand memory reference. Replace the
641 // FrameIndex with base register with EBP. Add an offset to the offset.
642 MI.getOperand(i).ChangeToRegister(BasePtr, false);
644 // Now add the frame object offset to the offset from EBP.
647 // Tail call jmp happens after FP is popped.
648 const TargetFrameInfo &TFI = *MF.getTarget().getFrameInfo();
649 const MachineFrameInfo *MFI = MF.getFrameInfo();
650 FIOffset = MFI->getObjectOffset(FrameIndex) - TFI.getOffsetOfLocalArea();
652 FIOffset = getFrameIndexOffset(MF, FrameIndex);
654 if (MI.getOperand(i+3).isImm()) {
655 // Offset is a 32-bit integer.
656 int Offset = FIOffset + (int)(MI.getOperand(i + 3).getImm());
657 MI.getOperand(i + 3).ChangeToImmediate(Offset);
659 // Offset is symbolic. This is extremely rare.
660 uint64_t Offset = FIOffset + (uint64_t)MI.getOperand(i+3).getOffset();
661 MI.getOperand(i+3).setOffset(Offset);
667 X86RegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
668 RegScavenger *RS) const {
669 MachineFrameInfo *MFI = MF.getFrameInfo();
671 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
672 int32_t TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
674 if (TailCallReturnAddrDelta < 0) {
675 // create RETURNADDR area
684 MFI->CreateFixedObject(-TailCallReturnAddrDelta,
685 (-1U*SlotSize)+TailCallReturnAddrDelta, true);
689 assert((TailCallReturnAddrDelta <= 0) &&
690 "The Delta should always be zero or negative");
691 const TargetFrameInfo &TFI = *MF.getTarget().getFrameInfo();
693 // Create a frame entry for the EBP register that must be saved.
694 int FrameIdx = MFI->CreateFixedObject(SlotSize,
696 TFI.getOffsetOfLocalArea() +
697 TailCallReturnAddrDelta,
699 assert(FrameIdx == MFI->getObjectIndexBegin() &&
700 "Slot for EBP register must be last in order to be found!");
705 /// emitSPUpdate - Emit a series of instructions to increment / decrement the
706 /// stack pointer by a constant value.
708 void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
709 unsigned StackPtr, int64_t NumBytes, bool Is64Bit,
710 const TargetInstrInfo &TII) {
711 bool isSub = NumBytes < 0;
712 uint64_t Offset = isSub ? -NumBytes : NumBytes;
713 unsigned Opc = isSub ?
714 getSUBriOpcode(Is64Bit, Offset) :
715 getADDriOpcode(Is64Bit, Offset);
716 uint64_t Chunk = (1LL << 31) - 1;
717 DebugLoc DL = MBB.findDebugLoc(MBBI);
720 uint64_t ThisVal = (Offset > Chunk) ? Chunk : Offset;
722 BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr)
725 MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
730 /// mergeSPUpdatesUp - Merge two stack-manipulating instructions upper iterator.
732 void mergeSPUpdatesUp(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
733 unsigned StackPtr, uint64_t *NumBytes = NULL) {
734 if (MBBI == MBB.begin()) return;
736 MachineBasicBlock::iterator PI = prior(MBBI);
737 unsigned Opc = PI->getOpcode();
738 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
739 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) &&
740 PI->getOperand(0).getReg() == StackPtr) {
742 *NumBytes += PI->getOperand(2).getImm();
744 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
745 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
746 PI->getOperand(0).getReg() == StackPtr) {
748 *NumBytes -= PI->getOperand(2).getImm();
753 /// mergeSPUpdatesUp - Merge two stack-manipulating instructions lower iterator.
755 void mergeSPUpdatesDown(MachineBasicBlock &MBB,
756 MachineBasicBlock::iterator &MBBI,
757 unsigned StackPtr, uint64_t *NumBytes = NULL) {
758 // FIXME: THIS ISN'T RUN!!!
761 if (MBBI == MBB.end()) return;
763 MachineBasicBlock::iterator NI = llvm::next(MBBI);
764 if (NI == MBB.end()) return;
766 unsigned Opc = NI->getOpcode();
767 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
768 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) &&
769 NI->getOperand(0).getReg() == StackPtr) {
771 *NumBytes -= NI->getOperand(2).getImm();
774 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
775 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
776 NI->getOperand(0).getReg() == StackPtr) {
778 *NumBytes += NI->getOperand(2).getImm();
784 /// mergeSPUpdates - Checks the instruction before/after the passed
785 /// instruction. If it is an ADD/SUB instruction it is deleted argument and the
786 /// stack adjustment is returned as a positive value for ADD and a negative for
788 static int mergeSPUpdates(MachineBasicBlock &MBB,
789 MachineBasicBlock::iterator &MBBI,
791 bool doMergeWithPrevious) {
792 if ((doMergeWithPrevious && MBBI == MBB.begin()) ||
793 (!doMergeWithPrevious && MBBI == MBB.end()))
796 MachineBasicBlock::iterator PI = doMergeWithPrevious ? prior(MBBI) : MBBI;
797 MachineBasicBlock::iterator NI = doMergeWithPrevious ? 0 : llvm::next(MBBI);
798 unsigned Opc = PI->getOpcode();
801 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
802 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) &&
803 PI->getOperand(0).getReg() == StackPtr){
804 Offset += PI->getOperand(2).getImm();
806 if (!doMergeWithPrevious) MBBI = NI;
807 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
808 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
809 PI->getOperand(0).getReg() == StackPtr) {
810 Offset -= PI->getOperand(2).getImm();
812 if (!doMergeWithPrevious) MBBI = NI;
818 void X86RegisterInfo::emitCalleeSavedFrameMoves(MachineFunction &MF,
820 unsigned FramePtr) const {
821 MachineFrameInfo *MFI = MF.getFrameInfo();
822 MachineModuleInfo &MMI = MF.getMMI();
824 // Add callee saved registers to move list.
825 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
826 if (CSI.empty()) return;
828 std::vector<MachineMove> &Moves = MMI.getFrameMoves();
829 const TargetData *TD = MF.getTarget().getTargetData();
830 bool HasFP = hasFP(MF);
832 // Calculate amount of bytes used for return address storing.
834 (MF.getTarget().getFrameInfo()->getStackGrowthDirection() ==
835 TargetFrameInfo::StackGrowsUp ?
836 TD->getPointerSize() : -TD->getPointerSize());
838 // FIXME: This is dirty hack. The code itself is pretty mess right now.
839 // It should be rewritten from scratch and generalized sometimes.
841 // Determine maximum offset (minumum due to stack growth).
842 int64_t MaxOffset = 0;
843 for (std::vector<CalleeSavedInfo>::const_iterator
844 I = CSI.begin(), E = CSI.end(); I != E; ++I)
845 MaxOffset = std::min(MaxOffset,
846 MFI->getObjectOffset(I->getFrameIdx()));
848 // Calculate offsets.
849 int64_t saveAreaOffset = (HasFP ? 3 : 2) * stackGrowth;
850 for (std::vector<CalleeSavedInfo>::const_iterator
851 I = CSI.begin(), E = CSI.end(); I != E; ++I) {
852 int64_t Offset = MFI->getObjectOffset(I->getFrameIdx());
853 unsigned Reg = I->getReg();
854 Offset = MaxOffset - Offset + saveAreaOffset;
856 // Don't output a new machine move if we're re-saving the frame
857 // pointer. This happens when the PrologEpilogInserter has inserted an extra
858 // "PUSH" of the frame pointer -- the "emitPrologue" method automatically
859 // generates one when frame pointers are used. If we generate a "machine
860 // move" for this extra "PUSH", the linker will lose track of the fact that
861 // the frame pointer should have the value of the first "PUSH" when it's
864 // FIXME: This looks inelegant. It's possibly correct, but it's covering up
865 // another bug. I.e., one where we generate a prolog like this:
873 // The immediate re-push of EBP is unnecessary. At the least, it's an
874 // optimization bug. EBP can be used as a scratch register in certain
875 // cases, but probably not when we have a frame pointer.
876 if (HasFP && FramePtr == Reg)
879 MachineLocation CSDst(MachineLocation::VirtualFP, Offset);
880 MachineLocation CSSrc(Reg);
881 Moves.push_back(MachineMove(Label, CSDst, CSSrc));
885 /// emitPrologue - Push callee-saved registers onto the stack, which
886 /// automatically adjust the stack pointer. Adjust the stack pointer to allocate
887 /// space for local variables. Also emit labels used by the exception handler to
888 /// generate the exception handling frames.
889 void X86RegisterInfo::emitPrologue(MachineFunction &MF) const {
890 MachineBasicBlock &MBB = MF.front(); // Prologue goes in entry BB.
891 MachineBasicBlock::iterator MBBI = MBB.begin();
892 MachineFrameInfo *MFI = MF.getFrameInfo();
893 const Function *Fn = MF.getFunction();
894 const X86Subtarget *Subtarget = &MF.getTarget().getSubtarget<X86Subtarget>();
895 MachineModuleInfo &MMI = MF.getMMI();
896 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
897 bool needsFrameMoves = MMI.hasDebugInfo() ||
898 !Fn->doesNotThrow() || UnwindTablesMandatory;
899 uint64_t MaxAlign = MFI->getMaxAlignment(); // Desired stack alignment.
900 uint64_t StackSize = MFI->getStackSize(); // Number of bytes to allocate.
901 bool HasFP = hasFP(MF);
904 // Add RETADDR move area to callee saved frame size.
905 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
906 if (TailCallReturnAddrDelta < 0)
907 X86FI->setCalleeSavedFrameSize(
908 X86FI->getCalleeSavedFrameSize() - TailCallReturnAddrDelta);
910 // If this is x86-64 and the Red Zone is not disabled, if we are a leaf
911 // function, and use up to 128 bytes of stack space, don't have a frame
912 // pointer, calls, or dynamic alloca then we do not need to adjust the
913 // stack pointer (we fit in the Red Zone).
914 if (Is64Bit && !Fn->hasFnAttr(Attribute::NoRedZone) &&
915 !needsStackRealignment(MF) &&
916 !MFI->hasVarSizedObjects() && // No dynamic alloca.
917 !MFI->adjustsStack() && // No calls.
918 !Subtarget->isTargetWin64()) { // Win64 has no Red Zone
919 uint64_t MinSize = X86FI->getCalleeSavedFrameSize();
920 if (HasFP) MinSize += SlotSize;
921 StackSize = std::max(MinSize, StackSize > 128 ? StackSize - 128 : 0);
922 MFI->setStackSize(StackSize);
923 } else if (Subtarget->isTargetWin64()) {
924 // We need to always allocate 32 bytes as register spill area.
925 // FIXME: We might reuse these 32 bytes for leaf functions.
927 MFI->setStackSize(StackSize);
930 // Insert stack pointer adjustment for later moving of return addr. Only
931 // applies to tail call optimized functions where the callee argument stack
932 // size is bigger than the callers.
933 if (TailCallReturnAddrDelta < 0) {
935 BuildMI(MBB, MBBI, DL,
936 TII.get(getSUBriOpcode(Is64Bit, -TailCallReturnAddrDelta)),
939 .addImm(-TailCallReturnAddrDelta);
940 MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
943 // Mapping for machine moves:
945 // DST: VirtualFP AND
946 // SRC: VirtualFP => DW_CFA_def_cfa_offset
947 // ELSE => DW_CFA_def_cfa
949 // SRC: VirtualFP AND
950 // DST: Register => DW_CFA_def_cfa_register
953 // OFFSET < 0 => DW_CFA_offset_extended_sf
954 // REG < 64 => DW_CFA_offset + Reg
955 // ELSE => DW_CFA_offset_extended
957 std::vector<MachineMove> &Moves = MMI.getFrameMoves();
958 const TargetData *TD = MF.getTarget().getTargetData();
959 uint64_t NumBytes = 0;
960 int stackGrowth = -TD->getPointerSize();
963 // Calculate required stack adjustment.
964 uint64_t FrameSize = StackSize - SlotSize;
965 if (needsStackRealignment(MF))
966 FrameSize = (FrameSize + MaxAlign - 1) / MaxAlign * MaxAlign;
968 NumBytes = FrameSize - X86FI->getCalleeSavedFrameSize();
970 // Get the offset of the stack slot for the EBP register, which is
971 // guaranteed to be the last slot by processFunctionBeforeFrameFinalized.
972 // Update the frame offset adjustment.
973 MFI->setOffsetAdjustment(-NumBytes);
975 // Save EBP/RBP into the appropriate stack slot.
976 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::PUSH64r : X86::PUSH32r))
977 .addReg(FramePtr, RegState::Kill);
979 if (needsFrameMoves) {
980 // Mark the place where EBP/RBP was saved.
981 MCSymbol *FrameLabel = MMI.getContext().CreateTempSymbol();
982 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL)).addSym(FrameLabel);
984 // Define the current CFA rule to use the provided offset.
986 MachineLocation SPDst(MachineLocation::VirtualFP);
987 MachineLocation SPSrc(MachineLocation::VirtualFP, 2 * stackGrowth);
988 Moves.push_back(MachineMove(FrameLabel, SPDst, SPSrc));
990 // FIXME: Verify & implement for FP
991 MachineLocation SPDst(StackPtr);
992 MachineLocation SPSrc(StackPtr, stackGrowth);
993 Moves.push_back(MachineMove(FrameLabel, SPDst, SPSrc));
996 // Change the rule for the FramePtr to be an "offset" rule.
997 MachineLocation FPDst(MachineLocation::VirtualFP, 2 * stackGrowth);
998 MachineLocation FPSrc(FramePtr);
999 Moves.push_back(MachineMove(FrameLabel, FPDst, FPSrc));
1002 // Update EBP with the new base value...
1003 BuildMI(MBB, MBBI, DL,
1004 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), FramePtr)
1007 if (needsFrameMoves) {
1008 // Mark effective beginning of when frame pointer becomes valid.
1009 MCSymbol *FrameLabel = MMI.getContext().CreateTempSymbol();
1010 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL)).addSym(FrameLabel);
1012 // Define the current CFA to use the EBP/RBP register.
1013 MachineLocation FPDst(FramePtr);
1014 MachineLocation FPSrc(MachineLocation::VirtualFP);
1015 Moves.push_back(MachineMove(FrameLabel, FPDst, FPSrc));
1018 // Mark the FramePtr as live-in in every block except the entry.
1019 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
1021 I->addLiveIn(FramePtr);
1024 if (needsStackRealignment(MF)) {
1026 BuildMI(MBB, MBBI, DL,
1027 TII.get(Is64Bit ? X86::AND64ri32 : X86::AND32ri),
1028 StackPtr).addReg(StackPtr).addImm(-MaxAlign);
1030 // The EFLAGS implicit def is dead.
1031 MI->getOperand(3).setIsDead();
1034 NumBytes = StackSize - X86FI->getCalleeSavedFrameSize();
1037 // Skip the callee-saved push instructions.
1038 bool PushedRegs = false;
1039 int StackOffset = 2 * stackGrowth;
1041 while (MBBI != MBB.end() &&
1042 (MBBI->getOpcode() == X86::PUSH32r ||
1043 MBBI->getOpcode() == X86::PUSH64r)) {
1047 if (!HasFP && needsFrameMoves) {
1048 // Mark callee-saved push instruction.
1049 MCSymbol *Label = MMI.getContext().CreateTempSymbol();
1050 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL)).addSym(Label);
1052 // Define the current CFA rule to use the provided offset.
1053 unsigned Ptr = StackSize ?
1054 MachineLocation::VirtualFP : StackPtr;
1055 MachineLocation SPDst(Ptr);
1056 MachineLocation SPSrc(Ptr, StackOffset);
1057 Moves.push_back(MachineMove(Label, SPDst, SPSrc));
1058 StackOffset += stackGrowth;
1062 DL = MBB.findDebugLoc(MBBI);
1064 // Adjust stack pointer: ESP -= numbytes.
1065 if (NumBytes >= 4096 && Subtarget->isTargetCygMing()) {
1066 // Check, whether EAX is livein for this function.
1067 bool isEAXAlive = false;
1068 for (MachineRegisterInfo::livein_iterator
1069 II = MF.getRegInfo().livein_begin(),
1070 EE = MF.getRegInfo().livein_end(); (II != EE) && !isEAXAlive; ++II) {
1071 unsigned Reg = II->first;
1072 isEAXAlive = (Reg == X86::EAX || Reg == X86::AX ||
1073 Reg == X86::AH || Reg == X86::AL);
1076 // Function prologue calls _alloca to probe the stack when allocating more
1077 // than 4k bytes in one go. Touching the stack at 4K increments is necessary
1078 // to ensure that the guard pages used by the OS virtual memory manager are
1079 // allocated in correct sequence.
1081 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
1083 BuildMI(MBB, MBBI, DL, TII.get(X86::CALLpcrel32))
1084 .addExternalSymbol("_alloca")
1085 .addReg(StackPtr, RegState::Define | RegState::Implicit);
1088 BuildMI(MBB, MBBI, DL, TII.get(X86::PUSH32r))
1089 .addReg(X86::EAX, RegState::Kill);
1091 // Allocate NumBytes-4 bytes on stack. We'll also use 4 already
1092 // allocated bytes for EAX.
1093 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
1094 .addImm(NumBytes - 4);
1095 BuildMI(MBB, MBBI, DL, TII.get(X86::CALLpcrel32))
1096 .addExternalSymbol("_alloca")
1097 .addReg(StackPtr, RegState::Define | RegState::Implicit);
1100 MachineInstr *MI = addRegOffset(BuildMI(MF, DL, TII.get(X86::MOV32rm),
1102 StackPtr, false, NumBytes - 4);
1103 MBB.insert(MBBI, MI);
1105 } else if (NumBytes) {
1106 // If there is an SUB32ri of ESP immediately before this instruction, merge
1107 // the two. This can be the case when tail call elimination is enabled and
1108 // the callee has more arguments then the caller.
1109 NumBytes -= mergeSPUpdates(MBB, MBBI, StackPtr, true);
1111 // If there is an ADD32ri or SUB32ri of ESP immediately after this
1112 // instruction, merge the two instructions.
1113 mergeSPUpdatesDown(MBB, MBBI, StackPtr, &NumBytes);
1116 emitSPUpdate(MBB, MBBI, StackPtr, -(int64_t)NumBytes, Is64Bit, TII);
1119 if ((NumBytes || PushedRegs) && needsFrameMoves) {
1120 // Mark end of stack pointer adjustment.
1121 MCSymbol *Label = MMI.getContext().CreateTempSymbol();
1122 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL)).addSym(Label);
1124 if (!HasFP && NumBytes) {
1125 // Define the current CFA rule to use the provided offset.
1127 MachineLocation SPDst(MachineLocation::VirtualFP);
1128 MachineLocation SPSrc(MachineLocation::VirtualFP,
1129 -StackSize + stackGrowth);
1130 Moves.push_back(MachineMove(Label, SPDst, SPSrc));
1132 // FIXME: Verify & implement for FP
1133 MachineLocation SPDst(StackPtr);
1134 MachineLocation SPSrc(StackPtr, stackGrowth);
1135 Moves.push_back(MachineMove(Label, SPDst, SPSrc));
1139 // Emit DWARF info specifying the offsets of the callee-saved registers.
1141 emitCalleeSavedFrameMoves(MF, Label, HasFP ? FramePtr : StackPtr);
1145 void X86RegisterInfo::emitEpilogue(MachineFunction &MF,
1146 MachineBasicBlock &MBB) const {
1147 const MachineFrameInfo *MFI = MF.getFrameInfo();
1148 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1149 MachineBasicBlock::iterator MBBI = prior(MBB.end());
1150 unsigned RetOpcode = MBBI->getOpcode();
1151 DebugLoc DL = MBBI->getDebugLoc();
1153 switch (RetOpcode) {
1155 llvm_unreachable("Can only insert epilog into returning blocks");
1158 case X86::TCRETURNdi:
1159 case X86::TCRETURNri:
1160 case X86::TCRETURNmi:
1161 case X86::TCRETURNdi64:
1162 case X86::TCRETURNri64:
1163 case X86::TCRETURNmi64:
1164 case X86::EH_RETURN:
1165 case X86::EH_RETURN64:
1166 break; // These are ok
1169 // Get the number of bytes to allocate from the FrameInfo.
1170 uint64_t StackSize = MFI->getStackSize();
1171 uint64_t MaxAlign = MFI->getMaxAlignment();
1172 unsigned CSSize = X86FI->getCalleeSavedFrameSize();
1173 uint64_t NumBytes = 0;
1176 // Calculate required stack adjustment.
1177 uint64_t FrameSize = StackSize - SlotSize;
1178 if (needsStackRealignment(MF))
1179 FrameSize = (FrameSize + MaxAlign - 1)/MaxAlign*MaxAlign;
1181 NumBytes = FrameSize - CSSize;
1184 BuildMI(MBB, MBBI, DL,
1185 TII.get(Is64Bit ? X86::POP64r : X86::POP32r), FramePtr);
1187 NumBytes = StackSize - CSSize;
1190 // Skip the callee-saved pop instructions.
1191 MachineBasicBlock::iterator LastCSPop = MBBI;
1192 while (MBBI != MBB.begin()) {
1193 MachineBasicBlock::iterator PI = prior(MBBI);
1194 unsigned Opc = PI->getOpcode();
1196 if (Opc != X86::POP32r && Opc != X86::POP64r &&
1197 !PI->getDesc().isTerminator())
1203 DL = MBBI->getDebugLoc();
1205 // If there is an ADD32ri or SUB32ri of ESP immediately before this
1206 // instruction, merge the two instructions.
1207 if (NumBytes || MFI->hasVarSizedObjects())
1208 mergeSPUpdatesUp(MBB, MBBI, StackPtr, &NumBytes);
1210 // If dynamic alloca is used, then reset esp to point to the last callee-saved
1211 // slot before popping them off! Same applies for the case, when stack was
1213 if (needsStackRealignment(MF)) {
1214 // We cannot use LEA here, because stack pointer was realigned. We need to
1215 // deallocate local frame back.
1217 emitSPUpdate(MBB, MBBI, StackPtr, NumBytes, Is64Bit, TII);
1218 MBBI = prior(LastCSPop);
1221 BuildMI(MBB, MBBI, DL,
1222 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr),
1223 StackPtr).addReg(FramePtr);
1224 } else if (MFI->hasVarSizedObjects()) {
1226 unsigned Opc = Is64Bit ? X86::LEA64r : X86::LEA32r;
1228 addRegOffset(BuildMI(MF, DL, TII.get(Opc), StackPtr),
1229 FramePtr, false, -CSSize);
1230 MBB.insert(MBBI, MI);
1232 BuildMI(MBB, MBBI, DL,
1233 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), StackPtr)
1236 } else if (NumBytes) {
1237 // Adjust stack pointer back: ESP += numbytes.
1238 emitSPUpdate(MBB, MBBI, StackPtr, NumBytes, Is64Bit, TII);
1241 // We're returning from function via eh_return.
1242 if (RetOpcode == X86::EH_RETURN || RetOpcode == X86::EH_RETURN64) {
1243 MBBI = prior(MBB.end());
1244 MachineOperand &DestAddr = MBBI->getOperand(0);
1245 assert(DestAddr.isReg() && "Offset should be in register!");
1246 BuildMI(MBB, MBBI, DL,
1247 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr),
1248 StackPtr).addReg(DestAddr.getReg());
1249 } else if (RetOpcode == X86::TCRETURNri || RetOpcode == X86::TCRETURNdi ||
1250 RetOpcode == X86::TCRETURNmi ||
1251 RetOpcode == X86::TCRETURNri64 || RetOpcode == X86::TCRETURNdi64 ||
1252 RetOpcode == X86::TCRETURNmi64) {
1253 bool isMem = RetOpcode == X86::TCRETURNmi || RetOpcode == X86::TCRETURNmi64;
1254 // Tail call return: adjust the stack pointer and jump to callee.
1255 MBBI = prior(MBB.end());
1256 MachineOperand &JumpTarget = MBBI->getOperand(0);
1257 MachineOperand &StackAdjust = MBBI->getOperand(isMem ? 5 : 1);
1258 assert(StackAdjust.isImm() && "Expecting immediate value.");
1260 // Adjust stack pointer.
1261 int StackAdj = StackAdjust.getImm();
1262 int MaxTCDelta = X86FI->getTCReturnAddrDelta();
1264 assert(MaxTCDelta <= 0 && "MaxTCDelta should never be positive");
1266 // Incoporate the retaddr area.
1267 Offset = StackAdj-MaxTCDelta;
1268 assert(Offset >= 0 && "Offset should never be negative");
1271 // Check for possible merge with preceeding ADD instruction.
1272 Offset += mergeSPUpdates(MBB, MBBI, StackPtr, true);
1273 emitSPUpdate(MBB, MBBI, StackPtr, Offset, Is64Bit, TII);
1276 // Jump to label or value in register.
1277 if (RetOpcode == X86::TCRETURNdi || RetOpcode == X86::TCRETURNdi64) {
1278 BuildMI(MBB, MBBI, DL, TII.get((RetOpcode == X86::TCRETURNdi)
1279 ? X86::TAILJMPd : X86::TAILJMPd64)).
1280 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(),
1281 JumpTarget.getTargetFlags());
1282 } else if (RetOpcode == X86::TCRETURNmi || RetOpcode == X86::TCRETURNmi64) {
1283 MachineInstrBuilder MIB =
1284 BuildMI(MBB, MBBI, DL, TII.get((RetOpcode == X86::TCRETURNmi)
1285 ? X86::TAILJMPm : X86::TAILJMPm64));
1286 for (unsigned i = 0; i != 5; ++i)
1287 MIB.addOperand(MBBI->getOperand(i));
1288 } else if (RetOpcode == X86::TCRETURNri64) {
1289 BuildMI(MBB, MBBI, DL, TII.get(X86::TAILJMPr64)).
1290 addReg(JumpTarget.getReg(), RegState::Kill);
1292 BuildMI(MBB, MBBI, DL, TII.get(X86::TAILJMPr)).
1293 addReg(JumpTarget.getReg(), RegState::Kill);
1296 MachineInstr *NewMI = prior(MBBI);
1297 for (unsigned i = 2, e = MBBI->getNumOperands(); i != e; ++i)
1298 NewMI->addOperand(MBBI->getOperand(i));
1300 // Delete the pseudo instruction TCRETURN.
1302 } else if ((RetOpcode == X86::RET || RetOpcode == X86::RETI) &&
1303 (X86FI->getTCReturnAddrDelta() < 0)) {
1304 // Add the return addr area delta back since we are not tail calling.
1305 int delta = -1*X86FI->getTCReturnAddrDelta();
1306 MBBI = prior(MBB.end());
1308 // Check for possible merge with preceeding ADD instruction.
1309 delta += mergeSPUpdates(MBB, MBBI, StackPtr, true);
1310 emitSPUpdate(MBB, MBBI, StackPtr, delta, Is64Bit, TII);
1314 unsigned X86RegisterInfo::getRARegister() const {
1315 return Is64Bit ? X86::RIP // Should have dwarf #16.
1316 : X86::EIP; // Should have dwarf #8.
1319 unsigned X86RegisterInfo::getFrameRegister(const MachineFunction &MF) const {
1320 return hasFP(MF) ? FramePtr : StackPtr;
1324 X86RegisterInfo::getInitialFrameState(std::vector<MachineMove> &Moves) const {
1325 // Calculate amount of bytes used for return address storing
1326 int stackGrowth = (Is64Bit ? -8 : -4);
1328 // Initial state of the frame pointer is esp+stackGrowth.
1329 MachineLocation Dst(MachineLocation::VirtualFP);
1330 MachineLocation Src(StackPtr, stackGrowth);
1331 Moves.push_back(MachineMove(0, Dst, Src));
1333 // Add return address to move list
1334 MachineLocation CSDst(StackPtr, stackGrowth);
1335 MachineLocation CSSrc(getRARegister());
1336 Moves.push_back(MachineMove(0, CSDst, CSSrc));
1339 unsigned X86RegisterInfo::getEHExceptionRegister() const {
1340 llvm_unreachable("What is the exception register");
1344 unsigned X86RegisterInfo::getEHHandlerRegister() const {
1345 llvm_unreachable("What is the exception handler register");
1350 unsigned getX86SubSuperRegister(unsigned Reg, EVT VT, bool High) {
1351 switch (VT.getSimpleVT().SimpleTy) {
1352 default: return Reg;
1357 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
1359 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
1361 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
1363 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
1369 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
1371 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
1373 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
1375 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
1377 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
1379 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
1381 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
1383 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
1385 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
1387 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
1389 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
1391 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
1393 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
1395 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
1397 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
1399 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
1405 default: return Reg;
1406 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
1408 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
1410 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
1412 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
1414 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
1416 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
1418 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
1420 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
1422 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
1424 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
1426 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
1428 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
1430 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
1432 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
1434 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
1436 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
1441 default: return Reg;
1442 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
1444 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
1446 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
1448 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
1450 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
1452 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
1454 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
1456 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
1458 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
1460 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
1462 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
1464 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
1466 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
1468 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
1470 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
1472 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
1477 default: return Reg;
1478 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
1480 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
1482 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
1484 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
1486 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
1488 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
1490 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
1492 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
1494 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
1496 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
1498 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
1500 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
1502 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
1504 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
1506 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
1508 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
1517 #include "X86GenRegisterInfo.inc"
1520 struct MSAH : public MachineFunctionPass {
1522 MSAH() : MachineFunctionPass(&ID) {}
1524 virtual bool runOnMachineFunction(MachineFunction &MF) {
1525 const X86TargetMachine *TM =
1526 static_cast<const X86TargetMachine *>(&MF.getTarget());
1527 const X86RegisterInfo *X86RI = TM->getRegisterInfo();
1528 MachineRegisterInfo &RI = MF.getRegInfo();
1529 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1530 unsigned StackAlignment = X86RI->getStackAlignment();
1532 // Be over-conservative: scan over all vreg defs and find whether vector
1533 // registers are used. If yes, there is a possibility that vector register
1534 // will be spilled and thus require dynamic stack realignment.
1535 for (unsigned RegNum = TargetRegisterInfo::FirstVirtualRegister;
1536 RegNum < RI.getLastVirtReg(); ++RegNum)
1537 if (RI.getRegClass(RegNum)->getAlignment() > StackAlignment) {
1538 FuncInfo->setReserveFP(true);
1546 virtual const char *getPassName() const {
1547 return "X86 Maximal Stack Alignment Check";
1550 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1551 AU.setPreservesCFG();
1552 MachineFunctionPass::getAnalysisUsage(AU);
1560 llvm::createX86MaxStackAlignmentHeuristicPass() { return new MSAH(); }