X86RegisterInfo.cpp revision d33fa0f9e0c5ecfa5097c359bdfbb2a0f2895889 
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/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"
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
58  if (Is64Bit) {
59    SlotSize = 8;
60    StackPtr = X86::RSP;
61    FramePtr = X86::RBP;
62  } else {
63    SlotSize = 4;
64    StackPtr = X86::ESP;
65    FramePtr = X86::EBP;
66  }
67}
68
69/// getDwarfRegNum - This function maps LLVM register identifiers to the DWARF
70/// specific numbering, used in debug info and exception tables.
71int X86RegisterInfo::getDwarfRegNum(unsigned RegNo, bool isEH) const {
72  const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
73  unsigned Flavour = DWARFFlavour::X86_64;
74
75  if (!Subtarget->is64Bit()) {
76    if (Subtarget->isTargetDarwin()) {
77      if (isEH)
78        Flavour = DWARFFlavour::X86_32_DarwinEH;
79      else
80        Flavour = DWARFFlavour::X86_32_Generic;
81    } else if (Subtarget->isTargetCygMing()) {
82      // Unsupported by now, just quick fallback
83      Flavour = DWARFFlavour::X86_32_Generic;
84    } else {
85      Flavour = DWARFFlavour::X86_32_Generic;
86    }
87  }
88
89  return X86GenRegisterInfo::getDwarfRegNumFull(RegNo, Flavour);
90}
91
92/// getX86RegNum - This function maps LLVM register identifiers to their X86
93/// specific numbering, which is used in various places encoding instructions.
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    llvm_unreachable("Register allocator hasn't allocated reg correctly yet!");
150    return 0;
151  }
152}
153
154const TargetRegisterClass *
155X86RegisterInfo::getMatchingSuperRegClass(const TargetRegisterClass *A,
156                                          const TargetRegisterClass *B,
157                                          unsigned SubIdx) const {
158  switch (SubIdx) {
159  default: return 0;
160  case 1:
161    // 8-bit
162    if (B == &X86::GR8RegClass) {
163      if (A->getSize() == 2 || A->getSize() == 4 || A->getSize() == 8)
164        return A;
165    } else if (B == &X86::GR8_ABCD_LRegClass || B == &X86::GR8_ABCD_HRegClass) {
166      if (A == &X86::GR64RegClass || A == &X86::GR64_ABCDRegClass ||
167          A == &X86::GR64_NOREXRegClass ||
168          A == &X86::GR64_NOSPRegClass ||
169          A == &X86::GR64_NOREX_NOSPRegClass)
170        return &X86::GR64_ABCDRegClass;
171      else if (A == &X86::GR32RegClass || A == &X86::GR32_ABCDRegClass ||
172               A == &X86::GR32_NOREXRegClass ||
173               A == &X86::GR32_NOSPRegClass)
174        return &X86::GR32_ABCDRegClass;
175      else if (A == &X86::GR16RegClass || A == &X86::GR16_ABCDRegClass ||
176               A == &X86::GR16_NOREXRegClass)
177        return &X86::GR16_ABCDRegClass;
178    } else if (B == &X86::GR8_NOREXRegClass) {
179      if (A == &X86::GR64RegClass || A == &X86::GR64_NOREXRegClass ||
180          A == &X86::GR64_NOSPRegClass || A == &X86::GR64_NOREX_NOSPRegClass)
181        return &X86::GR64_NOREXRegClass;
182      else if (A == &X86::GR64_ABCDRegClass)
183        return &X86::GR64_ABCDRegClass;
184      else if (A == &X86::GR32RegClass || A == &X86::GR32_NOREXRegClass ||
185               A == &X86::GR32_NOSPRegClass)
186        return &X86::GR32_NOREXRegClass;
187      else if (A == &X86::GR32_ABCDRegClass)
188        return &X86::GR32_ABCDRegClass;
189      else if (A == &X86::GR16RegClass || A == &X86::GR16_NOREXRegClass)
190        return &X86::GR16_NOREXRegClass;
191      else if (A == &X86::GR16_ABCDRegClass)
192        return &X86::GR16_ABCDRegClass;
193    } else if (B == &X86::FR32RegClass) {
194      return A;
195    }
196    break;
197  case 2:
198    // 8-bit hi
199    if (B == &X86::GR8_ABCD_HRegClass) {
200      if (A == &X86::GR64RegClass || A == &X86::GR64_ABCDRegClass ||
201          A == &X86::GR64_NOREXRegClass ||
202          A == &X86::GR64_NOSPRegClass ||
203          A == &X86::GR64_NOREX_NOSPRegClass)
204        return &X86::GR64_ABCDRegClass;
205      else if (A == &X86::GR32RegClass || A == &X86::GR32_ABCDRegClass ||
206               A == &X86::GR32_NOREXRegClass || A == &X86::GR32_NOSPRegClass)
207        return &X86::GR32_ABCDRegClass;
208      else if (A == &X86::GR16RegClass || A == &X86::GR16_ABCDRegClass ||
209               A == &X86::GR16_NOREXRegClass)
210        return &X86::GR16_ABCDRegClass;
211    } else if (B == &X86::FR64RegClass) {
212      return A;
213    }
214    break;
215  case 3:
216    // 16-bit
217    if (B == &X86::GR16RegClass) {
218      if (A->getSize() == 4 || A->getSize() == 8)
219        return A;
220    } else if (B == &X86::GR16_ABCDRegClass) {
221      if (A == &X86::GR64RegClass || A == &X86::GR64_ABCDRegClass ||
222          A == &X86::GR64_NOREXRegClass ||
223          A == &X86::GR64_NOSPRegClass ||
224          A == &X86::GR64_NOREX_NOSPRegClass)
225        return &X86::GR64_ABCDRegClass;
226      else if (A == &X86::GR32RegClass || A == &X86::GR32_ABCDRegClass ||
227               A == &X86::GR32_NOREXRegClass || A == &X86::GR32_NOSPRegClass)
228        return &X86::GR32_ABCDRegClass;
229    } else if (B == &X86::GR16_NOREXRegClass) {
230      if (A == &X86::GR64RegClass || A == &X86::GR64_NOREXRegClass ||
231          A == &X86::GR64_NOSPRegClass || A == &X86::GR64_NOREX_NOSPRegClass)
232        return &X86::GR64_NOREXRegClass;
233      else if (A == &X86::GR64_ABCDRegClass)
234        return &X86::GR64_ABCDRegClass;
235      else if (A == &X86::GR32RegClass || A == &X86::GR32_NOREXRegClass ||
236               A == &X86::GR32_NOSPRegClass)
237        return &X86::GR32_NOREXRegClass;
238      else if (A == &X86::GR32_ABCDRegClass)
239        return &X86::GR64_ABCDRegClass;
240    } else if (B == &X86::VR128RegClass) {
241      return A;
242    }
243    break;
244  case 4:
245    // 32-bit
246    if (B == &X86::GR32RegClass || B == &X86::GR32_NOSPRegClass) {
247      if (A->getSize() == 8)
248        return A;
249    } else if (B == &X86::GR32_ABCDRegClass) {
250      if (A == &X86::GR64RegClass || A == &X86::GR64_ABCDRegClass ||
251          A == &X86::GR64_NOREXRegClass ||
252          A == &X86::GR64_NOSPRegClass ||
253          A == &X86::GR64_NOREX_NOSPRegClass)
254        return &X86::GR64_ABCDRegClass;
255    } else if (B == &X86::GR32_NOREXRegClass) {
256      if (A == &X86::GR64RegClass || A == &X86::GR64_NOREXRegClass ||
257          A == &X86::GR64_NOSPRegClass || A == &X86::GR64_NOREX_NOSPRegClass)
258        return &X86::GR64_NOREXRegClass;
259      else if (A == &X86::GR64_ABCDRegClass)
260        return &X86::GR64_ABCDRegClass;
261    }
262    break;
263  }
264  return 0;
265}
266
267const TargetRegisterClass *
268X86RegisterInfo::getPointerRegClass(unsigned Kind) const {
269  switch (Kind) {
270  default: llvm_unreachable("Unexpected Kind in getPointerRegClass!");
271  case 0: // Normal GPRs.
272    if (TM.getSubtarget<X86Subtarget>().is64Bit())
273      return &X86::GR64RegClass;
274    return &X86::GR32RegClass;
275  case 1: // Normal GRPs except the stack pointer (for encoding reasons).
276    if (TM.getSubtarget<X86Subtarget>().is64Bit())
277      return &X86::GR64_NOSPRegClass;
278    return &X86::GR32_NOSPRegClass;
279  }
280}
281
282const TargetRegisterClass *
283X86RegisterInfo::getCrossCopyRegClass(const TargetRegisterClass *RC) const {
284  if (RC == &X86::CCRRegClass) {
285    if (Is64Bit)
286      return &X86::GR64RegClass;
287    else
288      return &X86::GR32RegClass;
289  }
290  return NULL;
291}
292
293const unsigned *
294X86RegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
295  bool callsEHReturn = false;
296  bool ghcCall = false;
297
298  if (MF) {
299    callsEHReturn = MF->getMMI().callsEHReturn();
300    const Function *F = MF->getFunction();
301    ghcCall = (F ? F->getCallingConv() == CallingConv::GHC : false);
302  }
303
304  static const unsigned GhcCalleeSavedRegs[] = {
305    0
306  };
307
308  static const unsigned CalleeSavedRegs32Bit[] = {
309    X86::ESI, X86::EDI, X86::EBX, X86::EBP,  0
310  };
311
312  static const unsigned CalleeSavedRegs32EHRet[] = {
313    X86::EAX, X86::EDX, X86::ESI, X86::EDI, X86::EBX, X86::EBP,  0
314  };
315
316  static const unsigned CalleeSavedRegs64Bit[] = {
317    X86::RBX, X86::R12, X86::R13, X86::R14, X86::R15, X86::RBP, 0
318  };
319
320  static const unsigned CalleeSavedRegs64EHRet[] = {
321    X86::RAX, X86::RDX, X86::RBX, X86::R12,
322    X86::R13, X86::R14, X86::R15, X86::RBP, 0
323  };
324
325  static const unsigned CalleeSavedRegsWin64[] = {
326    X86::RBX,   X86::RBP,   X86::RDI,   X86::RSI,
327    X86::R12,   X86::R13,   X86::R14,   X86::R15,
328    X86::XMM6,  X86::XMM7,  X86::XMM8,  X86::XMM9,
329    X86::XMM10, X86::XMM11, X86::XMM12, X86::XMM13,
330    X86::XMM14, X86::XMM15, 0
331  };
332
333  if (ghcCall) {
334    return GhcCalleeSavedRegs;
335  } else if (Is64Bit) {
336    if (IsWin64)
337      return CalleeSavedRegsWin64;
338    else
339      return (callsEHReturn ? CalleeSavedRegs64EHRet : CalleeSavedRegs64Bit);
340  } else {
341    return (callsEHReturn ? CalleeSavedRegs32EHRet : CalleeSavedRegs32Bit);
342  }
343}
344
345const TargetRegisterClass* const*
346X86RegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const {
347  bool callsEHReturn = false;
348  if (MF)
349    callsEHReturn = MF->getMMI().callsEHReturn();
350
351  static const TargetRegisterClass * const CalleeSavedRegClasses32Bit[] = {
352    &X86::GR32RegClass, &X86::GR32RegClass,
353    &X86::GR32RegClass, &X86::GR32RegClass,  0
354  };
355  static const TargetRegisterClass * const CalleeSavedRegClasses32EHRet[] = {
356    &X86::GR32RegClass, &X86::GR32RegClass,
357    &X86::GR32RegClass, &X86::GR32RegClass,
358    &X86::GR32RegClass, &X86::GR32RegClass,  0
359  };
360  static const TargetRegisterClass * const CalleeSavedRegClasses64Bit[] = {
361    &X86::GR64RegClass, &X86::GR64RegClass,
362    &X86::GR64RegClass, &X86::GR64RegClass,
363    &X86::GR64RegClass, &X86::GR64RegClass, 0
364  };
365  static const TargetRegisterClass * const CalleeSavedRegClasses64EHRet[] = {
366    &X86::GR64RegClass, &X86::GR64RegClass,
367    &X86::GR64RegClass, &X86::GR64RegClass,
368    &X86::GR64RegClass, &X86::GR64RegClass,
369    &X86::GR64RegClass, &X86::GR64RegClass, 0
370  };
371  static const TargetRegisterClass * const CalleeSavedRegClassesWin64[] = {
372    &X86::GR64RegClass,  &X86::GR64RegClass,
373    &X86::GR64RegClass,  &X86::GR64RegClass,
374    &X86::GR64RegClass,  &X86::GR64RegClass,
375    &X86::GR64RegClass,  &X86::GR64RegClass,
376    &X86::VR128RegClass, &X86::VR128RegClass,
377    &X86::VR128RegClass, &X86::VR128RegClass,
378    &X86::VR128RegClass, &X86::VR128RegClass,
379    &X86::VR128RegClass, &X86::VR128RegClass,
380    &X86::VR128RegClass, &X86::VR128RegClass, 0
381  };
382
383  if (Is64Bit) {
384    if (IsWin64)
385      return CalleeSavedRegClassesWin64;
386    else
387      return (callsEHReturn ?
388              CalleeSavedRegClasses64EHRet : CalleeSavedRegClasses64Bit);
389  } else {
390    return (callsEHReturn ?
391            CalleeSavedRegClasses32EHRet : CalleeSavedRegClasses32Bit);
392  }
393}
394
395BitVector X86RegisterInfo::getReservedRegs(const MachineFunction &MF) const {
396  BitVector Reserved(getNumRegs());
397  // Set the stack-pointer register and its aliases as reserved.
398  Reserved.set(X86::RSP);
399  Reserved.set(X86::ESP);
400  Reserved.set(X86::SP);
401  Reserved.set(X86::SPL);
402
403  // Set the instruction pointer register and its aliases as reserved.
404  Reserved.set(X86::RIP);
405  Reserved.set(X86::EIP);
406  Reserved.set(X86::IP);
407
408  // Set the frame-pointer register and its aliases as reserved if needed.
409  if (hasFP(MF)) {
410    Reserved.set(X86::RBP);
411    Reserved.set(X86::EBP);
412    Reserved.set(X86::BP);
413    Reserved.set(X86::BPL);
414  }
415
416  // Mark the x87 stack registers as reserved, since they don't behave normally
417  // with respect to liveness. We don't fully model the effects of x87 stack
418  // pushes and pops after stackification.
419  Reserved.set(X86::ST0);
420  Reserved.set(X86::ST1);
421  Reserved.set(X86::ST2);
422  Reserved.set(X86::ST3);
423  Reserved.set(X86::ST4);
424  Reserved.set(X86::ST5);
425  Reserved.set(X86::ST6);
426  Reserved.set(X86::ST7);
427  return Reserved;
428}
429
430//===----------------------------------------------------------------------===//
431// Stack Frame Processing methods
432//===----------------------------------------------------------------------===//
433
434/// hasFP - Return true if the specified function should have a dedicated frame
435/// pointer register.  This is true if the function has variable sized allocas
436/// or if frame pointer elimination is disabled.
437bool X86RegisterInfo::hasFP(const MachineFunction &MF) const {
438  const MachineFrameInfo *MFI = MF.getFrameInfo();
439  const MachineModuleInfo &MMI = MF.getMMI();
440
441  return (DisableFramePointerElim(MF) ||
442          needsStackRealignment(MF) ||
443          MFI->hasVarSizedObjects() ||
444          MFI->isFrameAddressTaken() ||
445          MF.getInfo<X86MachineFunctionInfo>()->getForceFramePointer() ||
446          MMI.callsUnwindInit());
447}
448
449bool X86RegisterInfo::canRealignStack(const MachineFunction &MF) const {
450  const MachineFrameInfo *MFI = MF.getFrameInfo();
451  return (RealignStack &&
452          !MFI->hasVarSizedObjects());
453}
454
455bool X86RegisterInfo::needsStackRealignment(const MachineFunction &MF) const {
456  const MachineFrameInfo *MFI = MF.getFrameInfo();
457  const Function *F = MF.getFunction();
458  bool requiresRealignment =
459    RealignStack && ((MFI->getMaxAlignment() > StackAlign) ||
460                     F->hasFnAttr(Attribute::StackAlignment));
461
462  // FIXME: Currently we don't support stack realignment for functions with
463  //        variable-sized allocas.
464  // FIXME: Temporary disable the error - it seems to be too conservative.
465  if (0 && requiresRealignment && MFI->hasVarSizedObjects())
466    report_fatal_error(
467      "Stack realignment in presense of dynamic allocas is not supported");
468
469  return (requiresRealignment && !MFI->hasVarSizedObjects());
470}
471
472bool X86RegisterInfo::hasReservedCallFrame(MachineFunction &MF) const {
473  return !MF.getFrameInfo()->hasVarSizedObjects();
474}
475
476bool X86RegisterInfo::hasReservedSpillSlot(MachineFunction &MF, unsigned Reg,
477                                           int &FrameIdx) const {
478  if (Reg == FramePtr && hasFP(MF)) {
479    FrameIdx = MF.getFrameInfo()->getObjectIndexBegin();
480    return true;
481  }
482  return false;
483}
484
485int
486X86RegisterInfo::getFrameIndexOffset(const MachineFunction &MF, int FI) const {
487  const TargetFrameInfo &TFI = *MF.getTarget().getFrameInfo();
488  const MachineFrameInfo *MFI = MF.getFrameInfo();
489  int Offset = MFI->getObjectOffset(FI) - TFI.getOffsetOfLocalArea();
490  uint64_t StackSize = MFI->getStackSize();
491
492  if (needsStackRealignment(MF)) {
493    if (FI < 0) {
494      // Skip the saved EBP.
495      Offset += SlotSize;
496    } else {
497      unsigned Align = MFI->getObjectAlignment(FI);
498      assert((-(Offset + StackSize)) % Align == 0);
499      Align = 0;
500      return Offset + StackSize;
501    }
502    // FIXME: Support tail calls
503  } else {
504    if (!hasFP(MF))
505      return Offset + StackSize;
506
507    // Skip the saved EBP.
508    Offset += SlotSize;
509
510    // Skip the RETADDR move area
511    const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
512    int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
513    if (TailCallReturnAddrDelta < 0)
514      Offset -= TailCallReturnAddrDelta;
515  }
516
517  return Offset;
518}
519
520void X86RegisterInfo::
521eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
522                              MachineBasicBlock::iterator I) const {
523  if (!hasReservedCallFrame(MF)) {
524    // If the stack pointer can be changed after prologue, turn the
525    // adjcallstackup instruction into a 'sub ESP, <amt>' and the
526    // adjcallstackdown instruction into 'add ESP, <amt>'
527    // TODO: consider using push / pop instead of sub + store / add
528    MachineInstr *Old = I;
529    uint64_t Amount = Old->getOperand(0).getImm();
530    if (Amount != 0) {
531      // We need to keep the stack aligned properly.  To do this, we round the
532      // amount of space needed for the outgoing arguments up to the next
533      // alignment boundary.
534      Amount = (Amount + StackAlign - 1) / StackAlign * StackAlign;
535
536      MachineInstr *New = 0;
537      if (Old->getOpcode() == getCallFrameSetupOpcode()) {
538        New = BuildMI(MF, Old->getDebugLoc(),
539                      TII.get(Is64Bit ? X86::SUB64ri32 : X86::SUB32ri),
540                      StackPtr)
541          .addReg(StackPtr)
542          .addImm(Amount);
543      } else {
544        assert(Old->getOpcode() == getCallFrameDestroyOpcode());
545
546        // Factor out the amount the callee already popped.
547        uint64_t CalleeAmt = Old->getOperand(1).getImm();
548        Amount -= CalleeAmt;
549
550      if (Amount) {
551          unsigned Opc = (Amount < 128) ?
552            (Is64Bit ? X86::ADD64ri8 : X86::ADD32ri8) :
553            (Is64Bit ? X86::ADD64ri32 : X86::ADD32ri);
554          New = BuildMI(MF, Old->getDebugLoc(), TII.get(Opc), StackPtr)
555            .addReg(StackPtr)
556            .addImm(Amount);
557        }
558      }
559
560      if (New) {
561        // The EFLAGS implicit def is dead.
562        New->getOperand(3).setIsDead();
563
564        // Replace the pseudo instruction with a new instruction.
565        MBB.insert(I, New);
566      }
567    }
568  } else if (I->getOpcode() == getCallFrameDestroyOpcode()) {
569    // If we are performing frame pointer elimination and if the callee pops
570    // something off the stack pointer, add it back.  We do this until we have
571    // more advanced stack pointer tracking ability.
572    if (uint64_t CalleeAmt = I->getOperand(1).getImm()) {
573      unsigned Opc = (CalleeAmt < 128) ?
574        (Is64Bit ? X86::SUB64ri8 : X86::SUB32ri8) :
575        (Is64Bit ? X86::SUB64ri32 : X86::SUB32ri);
576      MachineInstr *Old = I;
577      MachineInstr *New =
578        BuildMI(MF, Old->getDebugLoc(), TII.get(Opc),
579                StackPtr)
580          .addReg(StackPtr)
581          .addImm(CalleeAmt);
582
583      // The EFLAGS implicit def is dead.
584      New->getOperand(3).setIsDead();
585      MBB.insert(I, New);
586    }
587  }
588
589  MBB.erase(I);
590}
591
592unsigned
593X86RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
594                                     int SPAdj, FrameIndexValue *Value,
595                                     RegScavenger *RS) const{
596  assert(SPAdj == 0 && "Unexpected");
597
598  unsigned i = 0;
599  MachineInstr &MI = *II;
600  MachineFunction &MF = *MI.getParent()->getParent();
601
602  while (!MI.getOperand(i).isFI()) {
603    ++i;
604    assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
605  }
606
607  int FrameIndex = MI.getOperand(i).getIndex();
608  unsigned BasePtr;
609
610  unsigned Opc = MI.getOpcode();
611  bool AfterFPPop = Opc == X86::TAILJMPm64 || Opc == X86::TAILJMPm;
612  if (needsStackRealignment(MF))
613    BasePtr = (FrameIndex < 0 ? FramePtr : StackPtr);
614  else if (AfterFPPop)
615    BasePtr = StackPtr;
616  else
617    BasePtr = (hasFP(MF) ? FramePtr : StackPtr);
618
619  // This must be part of a four operand memory reference.  Replace the
620  // FrameIndex with base register with EBP.  Add an offset to the offset.
621  MI.getOperand(i).ChangeToRegister(BasePtr, false);
622
623  // Now add the frame object offset to the offset from EBP.
624  int FIOffset;
625  if (AfterFPPop) {
626    // Tail call jmp happens after FP is popped.
627    const TargetFrameInfo &TFI = *MF.getTarget().getFrameInfo();
628    const MachineFrameInfo *MFI = MF.getFrameInfo();
629    FIOffset = MFI->getObjectOffset(FrameIndex) - TFI.getOffsetOfLocalArea();
630  } else
631    FIOffset = getFrameIndexOffset(MF, FrameIndex);
632
633  if (MI.getOperand(i+3).isImm()) {
634    // Offset is a 32-bit integer.
635    int Offset = FIOffset + (int)(MI.getOperand(i + 3).getImm());
636    MI.getOperand(i + 3).ChangeToImmediate(Offset);
637  } else {
638    // Offset is symbolic. This is extremely rare.
639    uint64_t Offset = FIOffset + (uint64_t)MI.getOperand(i+3).getOffset();
640    MI.getOperand(i+3).setOffset(Offset);
641  }
642  return 0;
643}
644
645void
646X86RegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
647                                                      RegScavenger *RS) const {
648  MachineFrameInfo *MFI = MF.getFrameInfo();
649
650  X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
651  int32_t TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
652
653  if (TailCallReturnAddrDelta < 0) {
654    // create RETURNADDR area
655    //   arg
656    //   arg
657    //   RETADDR
658    //   { ...
659    //     RETADDR area
660    //     ...
661    //   }
662    //   [EBP]
663    MFI->CreateFixedObject(-TailCallReturnAddrDelta,
664                           (-1U*SlotSize)+TailCallReturnAddrDelta,
665                           true, false);
666  }
667
668  if (hasFP(MF)) {
669    assert((TailCallReturnAddrDelta <= 0) &&
670           "The Delta should always be zero or negative");
671    const TargetFrameInfo &TFI = *MF.getTarget().getFrameInfo();
672
673    // Create a frame entry for the EBP register that must be saved.
674    int FrameIdx = MFI->CreateFixedObject(SlotSize,
675                                          -(int)SlotSize +
676                                          TFI.getOffsetOfLocalArea() +
677                                          TailCallReturnAddrDelta,
678                                          true, false);
679    assert(FrameIdx == MFI->getObjectIndexBegin() &&
680           "Slot for EBP register must be last in order to be found!");
681    FrameIdx = 0;
682  }
683}
684
685/// emitSPUpdate - Emit a series of instructions to increment / decrement the
686/// stack pointer by a constant value.
687static
688void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
689                  unsigned StackPtr, int64_t NumBytes, bool Is64Bit,
690                  const TargetInstrInfo &TII) {
691  bool isSub = NumBytes < 0;
692  uint64_t Offset = isSub ? -NumBytes : NumBytes;
693  unsigned Opc = isSub
694    ? ((Offset < 128) ?
695       (Is64Bit ? X86::SUB64ri8 : X86::SUB32ri8) :
696       (Is64Bit ? X86::SUB64ri32 : X86::SUB32ri))
697    : ((Offset < 128) ?
698       (Is64Bit ? X86::ADD64ri8 : X86::ADD32ri8) :
699       (Is64Bit ? X86::ADD64ri32 : X86::ADD32ri));
700  uint64_t Chunk = (1LL << 31) - 1;
701  DebugLoc DL = MBB.findDebugLoc(MBBI);
702
703  while (Offset) {
704    uint64_t ThisVal = (Offset > Chunk) ? Chunk : Offset;
705    MachineInstr *MI =
706      BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr)
707        .addReg(StackPtr)
708        .addImm(ThisVal);
709    MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
710    Offset -= ThisVal;
711  }
712}
713
714/// mergeSPUpdatesUp - Merge two stack-manipulating instructions upper iterator.
715static
716void mergeSPUpdatesUp(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
717                      unsigned StackPtr, uint64_t *NumBytes = NULL) {
718  if (MBBI == MBB.begin()) return;
719
720  MachineBasicBlock::iterator PI = prior(MBBI);
721  unsigned Opc = PI->getOpcode();
722  if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
723       Opc == X86::ADD32ri || Opc == X86::ADD32ri8) &&
724      PI->getOperand(0).getReg() == StackPtr) {
725    if (NumBytes)
726      *NumBytes += PI->getOperand(2).getImm();
727    MBB.erase(PI);
728  } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
729              Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
730             PI->getOperand(0).getReg() == StackPtr) {
731    if (NumBytes)
732      *NumBytes -= PI->getOperand(2).getImm();
733    MBB.erase(PI);
734  }
735}
736
737/// mergeSPUpdatesUp - Merge two stack-manipulating instructions lower iterator.
738static
739void mergeSPUpdatesDown(MachineBasicBlock &MBB,
740                        MachineBasicBlock::iterator &MBBI,
741                        unsigned StackPtr, uint64_t *NumBytes = NULL) {
742  // FIXME: THIS ISN'T RUN!!!
743  return;
744
745  if (MBBI == MBB.end()) return;
746
747  MachineBasicBlock::iterator NI = llvm::next(MBBI);
748  if (NI == MBB.end()) return;
749
750  unsigned Opc = NI->getOpcode();
751  if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
752       Opc == X86::ADD32ri || Opc == X86::ADD32ri8) &&
753      NI->getOperand(0).getReg() == StackPtr) {
754    if (NumBytes)
755      *NumBytes -= NI->getOperand(2).getImm();
756    MBB.erase(NI);
757    MBBI = NI;
758  } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
759              Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
760             NI->getOperand(0).getReg() == StackPtr) {
761    if (NumBytes)
762      *NumBytes += NI->getOperand(2).getImm();
763    MBB.erase(NI);
764    MBBI = NI;
765  }
766}
767
768/// mergeSPUpdates - Checks the instruction before/after the passed
769/// instruction. If it is an ADD/SUB instruction it is deleted argument and the
770/// stack adjustment is returned as a positive value for ADD and a negative for
771/// SUB.
772static int mergeSPUpdates(MachineBasicBlock &MBB,
773                           MachineBasicBlock::iterator &MBBI,
774                           unsigned StackPtr,
775                           bool doMergeWithPrevious) {
776  if ((doMergeWithPrevious && MBBI == MBB.begin()) ||
777      (!doMergeWithPrevious && MBBI == MBB.end()))
778    return 0;
779
780  MachineBasicBlock::iterator PI = doMergeWithPrevious ? prior(MBBI) : MBBI;
781  MachineBasicBlock::iterator NI = doMergeWithPrevious ? 0 : llvm::next(MBBI);
782  unsigned Opc = PI->getOpcode();
783  int Offset = 0;
784
785  if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
786       Opc == X86::ADD32ri || Opc == X86::ADD32ri8) &&
787      PI->getOperand(0).getReg() == StackPtr){
788    Offset += PI->getOperand(2).getImm();
789    MBB.erase(PI);
790    if (!doMergeWithPrevious) MBBI = NI;
791  } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
792              Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
793             PI->getOperand(0).getReg() == StackPtr) {
794    Offset -= PI->getOperand(2).getImm();
795    MBB.erase(PI);
796    if (!doMergeWithPrevious) MBBI = NI;
797  }
798
799  return Offset;
800}
801
802void X86RegisterInfo::emitCalleeSavedFrameMoves(MachineFunction &MF,
803                                                MCSymbol *Label,
804                                                unsigned FramePtr) const {
805  MachineFrameInfo *MFI = MF.getFrameInfo();
806  MachineModuleInfo &MMI = MF.getMMI();
807
808  // Add callee saved registers to move list.
809  const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
810  if (CSI.empty()) return;
811
812  std::vector<MachineMove> &Moves = MMI.getFrameMoves();
813  const TargetData *TD = MF.getTarget().getTargetData();
814  bool HasFP = hasFP(MF);
815
816  // Calculate amount of bytes used for return address storing.
817  int stackGrowth =
818    (MF.getTarget().getFrameInfo()->getStackGrowthDirection() ==
819     TargetFrameInfo::StackGrowsUp ?
820     TD->getPointerSize() : -TD->getPointerSize());
821
822  // FIXME: This is dirty hack. The code itself is pretty mess right now.
823  // It should be rewritten from scratch and generalized sometimes.
824
825  // Determine maximum offset (minumum due to stack growth).
826  int64_t MaxOffset = 0;
827  for (std::vector<CalleeSavedInfo>::const_iterator
828         I = CSI.begin(), E = CSI.end(); I != E; ++I)
829    MaxOffset = std::min(MaxOffset,
830                         MFI->getObjectOffset(I->getFrameIdx()));
831
832  // Calculate offsets.
833  int64_t saveAreaOffset = (HasFP ? 3 : 2) * stackGrowth;
834  for (std::vector<CalleeSavedInfo>::const_iterator
835         I = CSI.begin(), E = CSI.end(); I != E; ++I) {
836    int64_t Offset = MFI->getObjectOffset(I->getFrameIdx());
837    unsigned Reg = I->getReg();
838    Offset = MaxOffset - Offset + saveAreaOffset;
839
840    // Don't output a new machine move if we're re-saving the frame
841    // pointer. This happens when the PrologEpilogInserter has inserted an extra
842    // "PUSH" of the frame pointer -- the "emitPrologue" method automatically
843    // generates one when frame pointers are used. If we generate a "machine
844    // move" for this extra "PUSH", the linker will lose track of the fact that
845    // the frame pointer should have the value of the first "PUSH" when it's
846    // trying to unwind.
847    //
848    // FIXME: This looks inelegant. It's possibly correct, but it's covering up
849    //        another bug. I.e., one where we generate a prolog like this:
850    //
851    //          pushl  %ebp
852    //          movl   %esp, %ebp
853    //          pushl  %ebp
854    //          pushl  %esi
855    //           ...
856    //
857    //        The immediate re-push of EBP is unnecessary. At the least, it's an
858    //        optimization bug. EBP can be used as a scratch register in certain
859    //        cases, but probably not when we have a frame pointer.
860    if (HasFP && FramePtr == Reg)
861      continue;
862
863    MachineLocation CSDst(MachineLocation::VirtualFP, Offset);
864    MachineLocation CSSrc(Reg);
865    Moves.push_back(MachineMove(Label, CSDst, CSSrc));
866  }
867}
868
869/// emitPrologue - Push callee-saved registers onto the stack, which
870/// automatically adjust the stack pointer. Adjust the stack pointer to allocate
871/// space for local variables. Also emit labels used by the exception handler to
872/// generate the exception handling frames.
873void X86RegisterInfo::emitPrologue(MachineFunction &MF) const {
874  MachineBasicBlock &MBB = MF.front(); // Prologue goes in entry BB.
875  MachineBasicBlock::iterator MBBI = MBB.begin();
876  MachineFrameInfo *MFI = MF.getFrameInfo();
877  const Function *Fn = MF.getFunction();
878  const X86Subtarget *Subtarget = &MF.getTarget().getSubtarget<X86Subtarget>();
879  MachineModuleInfo &MMI = MF.getMMI();
880  X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
881  bool needsFrameMoves = MMI.hasDebugInfo() ||
882                          !Fn->doesNotThrow() || UnwindTablesMandatory;
883  uint64_t MaxAlign  = MFI->getMaxAlignment(); // Desired stack alignment.
884  uint64_t StackSize = MFI->getStackSize();    // Number of bytes to allocate.
885  bool HasFP = hasFP(MF);
886  DebugLoc DL;
887
888  // Add RETADDR move area to callee saved frame size.
889  int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
890  if (TailCallReturnAddrDelta < 0)
891    X86FI->setCalleeSavedFrameSize(
892      X86FI->getCalleeSavedFrameSize() - TailCallReturnAddrDelta);
893
894  // If this is x86-64 and the Red Zone is not disabled, if we are a leaf
895  // function, and use up to 128 bytes of stack space, don't have a frame
896  // pointer, calls, or dynamic alloca then we do not need to adjust the
897  // stack pointer (we fit in the Red Zone).
898  if (Is64Bit && !Fn->hasFnAttr(Attribute::NoRedZone) &&
899      !needsStackRealignment(MF) &&
900      !MFI->hasVarSizedObjects() &&                // No dynamic alloca.
901      !MFI->adjustsStack() &&                      // Doesn't adjust stack.
902      !Subtarget->isTargetWin64()) {               // Win64 has no Red Zone
903    uint64_t MinSize = X86FI->getCalleeSavedFrameSize();
904    if (HasFP) MinSize += SlotSize;
905    StackSize = std::max(MinSize, StackSize > 128 ? StackSize - 128 : 0);
906    MFI->setStackSize(StackSize);
907  } else if (Subtarget->isTargetWin64()) {
908    // We need to always allocate 32 bytes as register spill area.
909    // FIXME: We might reuse these 32 bytes for leaf functions.
910    StackSize += 32;
911    MFI->setStackSize(StackSize);
912  }
913
914  // Insert stack pointer adjustment for later moving of return addr.  Only
915  // applies to tail call optimized functions where the callee argument stack
916  // size is bigger than the callers.
917  if (TailCallReturnAddrDelta < 0) {
918    MachineInstr *MI =
919      BuildMI(MBB, MBBI, DL, TII.get(Is64Bit? X86::SUB64ri32 : X86::SUB32ri),
920              StackPtr)
921        .addReg(StackPtr)
922        .addImm(-TailCallReturnAddrDelta);
923    MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
924  }
925
926  // Mapping for machine moves:
927  //
928  //   DST: VirtualFP AND
929  //        SRC: VirtualFP              => DW_CFA_def_cfa_offset
930  //        ELSE                        => DW_CFA_def_cfa
931  //
932  //   SRC: VirtualFP AND
933  //        DST: Register               => DW_CFA_def_cfa_register
934  //
935  //   ELSE
936  //        OFFSET < 0                  => DW_CFA_offset_extended_sf
937  //        REG < 64                    => DW_CFA_offset + Reg
938  //        ELSE                        => DW_CFA_offset_extended
939
940  std::vector<MachineMove> &Moves = MMI.getFrameMoves();
941  const TargetData *TD = MF.getTarget().getTargetData();
942  uint64_t NumBytes = 0;
943  int stackGrowth = -TD->getPointerSize();
944
945  if (HasFP) {
946    // Calculate required stack adjustment.
947    uint64_t FrameSize = StackSize - SlotSize;
948    if (needsStackRealignment(MF))
949      FrameSize = (FrameSize + MaxAlign - 1) / MaxAlign * MaxAlign;
950
951    NumBytes = FrameSize - X86FI->getCalleeSavedFrameSize();
952
953    // Get the offset of the stack slot for the EBP register, which is
954    // guaranteed to be the last slot by processFunctionBeforeFrameFinalized.
955    // Update the frame offset adjustment.
956    MFI->setOffsetAdjustment(-NumBytes);
957
958    // Save EBP/RBP into the appropriate stack slot.
959    BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::PUSH64r : X86::PUSH32r))
960      .addReg(FramePtr, RegState::Kill);
961
962    if (needsFrameMoves) {
963      // Mark the place where EBP/RBP was saved.
964      MCSymbol *FrameLabel = MMI.getContext().CreateTempSymbol();
965      BuildMI(MBB, MBBI, DL, TII.get(X86::DBG_LABEL)).addSym(FrameLabel);
966
967      // Define the current CFA rule to use the provided offset.
968      if (StackSize) {
969        MachineLocation SPDst(MachineLocation::VirtualFP);
970        MachineLocation SPSrc(MachineLocation::VirtualFP, 2 * stackGrowth);
971        Moves.push_back(MachineMove(FrameLabel, SPDst, SPSrc));
972      } else {
973        // FIXME: Verify & implement for FP
974        MachineLocation SPDst(StackPtr);
975        MachineLocation SPSrc(StackPtr, stackGrowth);
976        Moves.push_back(MachineMove(FrameLabel, SPDst, SPSrc));
977      }
978
979      // Change the rule for the FramePtr to be an "offset" rule.
980      MachineLocation FPDst(MachineLocation::VirtualFP, 2 * stackGrowth);
981      MachineLocation FPSrc(FramePtr);
982      Moves.push_back(MachineMove(FrameLabel, FPDst, FPSrc));
983    }
984
985    // Update EBP with the new base value...
986    BuildMI(MBB, MBBI, DL,
987            TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), FramePtr)
988        .addReg(StackPtr);
989
990    if (needsFrameMoves) {
991      // Mark effective beginning of when frame pointer becomes valid.
992      MCSymbol *FrameLabel = MMI.getContext().CreateTempSymbol();
993      BuildMI(MBB, MBBI, DL, TII.get(X86::DBG_LABEL)).addSym(FrameLabel);
994
995      // Define the current CFA to use the EBP/RBP register.
996      MachineLocation FPDst(FramePtr);
997      MachineLocation FPSrc(MachineLocation::VirtualFP);
998      Moves.push_back(MachineMove(FrameLabel, FPDst, FPSrc));
999    }
1000
1001    // Mark the FramePtr as live-in in every block except the entry.
1002    for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
1003         I != E; ++I)
1004      I->addLiveIn(FramePtr);
1005
1006    // Realign stack
1007    if (needsStackRealignment(MF)) {
1008      MachineInstr *MI =
1009        BuildMI(MBB, MBBI, DL,
1010                TII.get(Is64Bit ? X86::AND64ri32 : X86::AND32ri),
1011                StackPtr).addReg(StackPtr).addImm(-MaxAlign);
1012
1013      // The EFLAGS implicit def is dead.
1014      MI->getOperand(3).setIsDead();
1015    }
1016  } else {
1017    NumBytes = StackSize - X86FI->getCalleeSavedFrameSize();
1018  }
1019
1020  // Skip the callee-saved push instructions.
1021  bool PushedRegs = false;
1022  int StackOffset = 2 * stackGrowth;
1023
1024  while (MBBI != MBB.end() &&
1025         (MBBI->getOpcode() == X86::PUSH32r ||
1026          MBBI->getOpcode() == X86::PUSH64r)) {
1027    PushedRegs = true;
1028    ++MBBI;
1029
1030    if (!HasFP && needsFrameMoves) {
1031      // Mark callee-saved push instruction.
1032      MCSymbol *Label = MMI.getContext().CreateTempSymbol();
1033      BuildMI(MBB, MBBI, DL, TII.get(X86::DBG_LABEL)).addSym(Label);
1034
1035      // Define the current CFA rule to use the provided offset.
1036      unsigned Ptr = StackSize ?
1037        MachineLocation::VirtualFP : StackPtr;
1038      MachineLocation SPDst(Ptr);
1039      MachineLocation SPSrc(Ptr, StackOffset);
1040      Moves.push_back(MachineMove(Label, SPDst, SPSrc));
1041      StackOffset += stackGrowth;
1042    }
1043  }
1044
1045  DL = MBB.findDebugLoc(MBBI);
1046
1047  // Adjust stack pointer: ESP -= numbytes.
1048  if (NumBytes >= 4096 && Subtarget->isTargetCygMing()) {
1049    // Check, whether EAX is livein for this function.
1050    bool isEAXAlive = false;
1051    for (MachineRegisterInfo::livein_iterator
1052           II = MF.getRegInfo().livein_begin(),
1053           EE = MF.getRegInfo().livein_end(); (II != EE) && !isEAXAlive; ++II) {
1054      unsigned Reg = II->first;
1055      isEAXAlive = (Reg == X86::EAX || Reg == X86::AX ||
1056                    Reg == X86::AH || Reg == X86::AL);
1057    }
1058
1059    // Function prologue calls _alloca to probe the stack when allocating more
1060    // than 4k bytes in one go. Touching the stack at 4K increments is necessary
1061    // to ensure that the guard pages used by the OS virtual memory manager are
1062    // allocated in correct sequence.
1063    if (!isEAXAlive) {
1064      BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
1065        .addImm(NumBytes);
1066      BuildMI(MBB, MBBI, DL, TII.get(X86::CALLpcrel32))
1067        .addExternalSymbol("_alloca")
1068        .addReg(StackPtr, RegState::Define | RegState::Implicit);
1069    } else {
1070      // Save EAX
1071      BuildMI(MBB, MBBI, DL, TII.get(X86::PUSH32r))
1072        .addReg(X86::EAX, RegState::Kill);
1073
1074      // Allocate NumBytes-4 bytes on stack. We'll also use 4 already
1075      // allocated bytes for EAX.
1076      BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
1077        .addImm(NumBytes - 4);
1078      BuildMI(MBB, MBBI, DL, TII.get(X86::CALLpcrel32))
1079        .addExternalSymbol("_alloca")
1080        .addReg(StackPtr, RegState::Define | RegState::Implicit);
1081
1082      // Restore EAX
1083      MachineInstr *MI = addRegOffset(BuildMI(MF, DL, TII.get(X86::MOV32rm),
1084                                              X86::EAX),
1085                                      StackPtr, false, NumBytes - 4);
1086      MBB.insert(MBBI, MI);
1087    }
1088  } else if (NumBytes) {
1089    // If there is an SUB32ri of ESP immediately before this instruction, merge
1090    // the two. This can be the case when tail call elimination is enabled and
1091    // the callee has more arguments then the caller.
1092    NumBytes -= mergeSPUpdates(MBB, MBBI, StackPtr, true);
1093
1094    // If there is an ADD32ri or SUB32ri of ESP immediately after this
1095    // instruction, merge the two instructions.
1096    mergeSPUpdatesDown(MBB, MBBI, StackPtr, &NumBytes);
1097
1098    if (NumBytes)
1099      emitSPUpdate(MBB, MBBI, StackPtr, -(int64_t)NumBytes, Is64Bit, TII);
1100  }
1101
1102  if ((NumBytes || PushedRegs) && needsFrameMoves) {
1103    // Mark end of stack pointer adjustment.
1104    MCSymbol *Label = MMI.getContext().CreateTempSymbol();
1105    BuildMI(MBB, MBBI, DL, TII.get(X86::DBG_LABEL)).addSym(Label);
1106
1107    if (!HasFP && NumBytes) {
1108      // Define the current CFA rule to use the provided offset.
1109      if (StackSize) {
1110        MachineLocation SPDst(MachineLocation::VirtualFP);
1111        MachineLocation SPSrc(MachineLocation::VirtualFP,
1112                              -StackSize + stackGrowth);
1113        Moves.push_back(MachineMove(Label, SPDst, SPSrc));
1114      } else {
1115        // FIXME: Verify & implement for FP
1116        MachineLocation SPDst(StackPtr);
1117        MachineLocation SPSrc(StackPtr, stackGrowth);
1118        Moves.push_back(MachineMove(Label, SPDst, SPSrc));
1119      }
1120    }
1121
1122    // Emit DWARF info specifying the offsets of the callee-saved registers.
1123    if (PushedRegs)
1124      emitCalleeSavedFrameMoves(MF, Label, HasFP ? FramePtr : StackPtr);
1125  }
1126}
1127
1128void X86RegisterInfo::emitEpilogue(MachineFunction &MF,
1129                                   MachineBasicBlock &MBB) const {
1130  const MachineFrameInfo *MFI = MF.getFrameInfo();
1131  X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1132  MachineBasicBlock::iterator MBBI = prior(MBB.end());
1133  unsigned RetOpcode = MBBI->getOpcode();
1134  DebugLoc DL = MBBI->getDebugLoc();
1135
1136  switch (RetOpcode) {
1137  default:
1138    llvm_unreachable("Can only insert epilog into returning blocks");
1139  case X86::RET:
1140  case X86::RETI:
1141  case X86::TCRETURNdi:
1142  case X86::TCRETURNri:
1143  case X86::TCRETURNmi:
1144  case X86::TCRETURNdi64:
1145  case X86::TCRETURNri64:
1146  case X86::TCRETURNmi64:
1147  case X86::EH_RETURN:
1148  case X86::EH_RETURN64:
1149    break;  // These are ok
1150  }
1151
1152  // Get the number of bytes to allocate from the FrameInfo.
1153  uint64_t StackSize = MFI->getStackSize();
1154  uint64_t MaxAlign  = MFI->getMaxAlignment();
1155  unsigned CSSize = X86FI->getCalleeSavedFrameSize();
1156  uint64_t NumBytes = 0;
1157
1158  if (hasFP(MF)) {
1159    // Calculate required stack adjustment.
1160    uint64_t FrameSize = StackSize - SlotSize;
1161    if (needsStackRealignment(MF))
1162      FrameSize = (FrameSize + MaxAlign - 1)/MaxAlign*MaxAlign;
1163
1164    NumBytes = FrameSize - CSSize;
1165
1166    // Pop EBP.
1167    BuildMI(MBB, MBBI, DL,
1168            TII.get(Is64Bit ? X86::POP64r : X86::POP32r), FramePtr);
1169  } else {
1170    NumBytes = StackSize - CSSize;
1171  }
1172
1173  // Skip the callee-saved pop instructions.
1174  MachineBasicBlock::iterator LastCSPop = MBBI;
1175  while (MBBI != MBB.begin()) {
1176    MachineBasicBlock::iterator PI = prior(MBBI);
1177    unsigned Opc = PI->getOpcode();
1178
1179    if (Opc != X86::POP32r && Opc != X86::POP64r &&
1180        !PI->getDesc().isTerminator())
1181      break;
1182
1183    --MBBI;
1184  }
1185
1186  DL = MBBI->getDebugLoc();
1187
1188  // If there is an ADD32ri or SUB32ri of ESP immediately before this
1189  // instruction, merge the two instructions.
1190  if (NumBytes || MFI->hasVarSizedObjects())
1191    mergeSPUpdatesUp(MBB, MBBI, StackPtr, &NumBytes);
1192
1193  // If dynamic alloca is used, then reset esp to point to the last callee-saved
1194  // slot before popping them off! Same applies for the case, when stack was
1195  // realigned.
1196  if (needsStackRealignment(MF)) {
1197    // We cannot use LEA here, because stack pointer was realigned. We need to
1198    // deallocate local frame back.
1199    if (CSSize) {
1200      emitSPUpdate(MBB, MBBI, StackPtr, NumBytes, Is64Bit, TII);
1201      MBBI = prior(LastCSPop);
1202    }
1203
1204    BuildMI(MBB, MBBI, DL,
1205            TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr),
1206            StackPtr).addReg(FramePtr);
1207  } else if (MFI->hasVarSizedObjects()) {
1208    if (CSSize) {
1209      unsigned Opc = Is64Bit ? X86::LEA64r : X86::LEA32r;
1210      MachineInstr *MI =
1211        addLeaRegOffset(BuildMI(MF, DL, TII.get(Opc), StackPtr),
1212                        FramePtr, false, -CSSize);
1213      MBB.insert(MBBI, MI);
1214    } else {
1215      BuildMI(MBB, MBBI, DL,
1216              TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), StackPtr)
1217        .addReg(FramePtr);
1218    }
1219  } else if (NumBytes) {
1220    // Adjust stack pointer back: ESP += numbytes.
1221    emitSPUpdate(MBB, MBBI, StackPtr, NumBytes, Is64Bit, TII);
1222  }
1223
1224  // We're returning from function via eh_return.
1225  if (RetOpcode == X86::EH_RETURN || RetOpcode == X86::EH_RETURN64) {
1226    MBBI = prior(MBB.end());
1227    MachineOperand &DestAddr  = MBBI->getOperand(0);
1228    assert(DestAddr.isReg() && "Offset should be in register!");
1229    BuildMI(MBB, MBBI, DL,
1230            TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr),
1231            StackPtr).addReg(DestAddr.getReg());
1232  } else if (RetOpcode == X86::TCRETURNri || RetOpcode == X86::TCRETURNdi ||
1233             RetOpcode == X86::TCRETURNmi ||
1234             RetOpcode == X86::TCRETURNri64 || RetOpcode == X86::TCRETURNdi64 ||
1235             RetOpcode == X86::TCRETURNmi64) {
1236    bool isMem = RetOpcode == X86::TCRETURNmi || RetOpcode == X86::TCRETURNmi64;
1237    // Tail call return: adjust the stack pointer and jump to callee.
1238    MBBI = prior(MBB.end());
1239    MachineOperand &JumpTarget = MBBI->getOperand(0);
1240    MachineOperand &StackAdjust = MBBI->getOperand(isMem ? 5 : 1);
1241    assert(StackAdjust.isImm() && "Expecting immediate value.");
1242
1243    // Adjust stack pointer.
1244    int StackAdj = StackAdjust.getImm();
1245    int MaxTCDelta = X86FI->getTCReturnAddrDelta();
1246    int Offset = 0;
1247    assert(MaxTCDelta <= 0 && "MaxTCDelta should never be positive");
1248
1249    // Incoporate the retaddr area.
1250    Offset = StackAdj-MaxTCDelta;
1251    assert(Offset >= 0 && "Offset should never be negative");
1252
1253    if (Offset) {
1254      // Check for possible merge with preceeding ADD instruction.
1255      Offset += mergeSPUpdates(MBB, MBBI, StackPtr, true);
1256      emitSPUpdate(MBB, MBBI, StackPtr, Offset, Is64Bit, TII);
1257    }
1258
1259    // Jump to label or value in register.
1260    if (RetOpcode == X86::TCRETURNdi || RetOpcode == X86::TCRETURNdi64) {
1261      BuildMI(MBB, MBBI, DL, TII.get((RetOpcode == X86::TCRETURNdi)
1262                                     ? X86::TAILJMPd : X86::TAILJMPd64)).
1263        addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(),
1264                         JumpTarget.getTargetFlags());
1265    } else if (RetOpcode == X86::TCRETURNmi || RetOpcode == X86::TCRETURNmi64) {
1266      MachineInstrBuilder MIB =
1267        BuildMI(MBB, MBBI, DL, TII.get((RetOpcode == X86::TCRETURNmi)
1268                                       ? X86::TAILJMPm : X86::TAILJMPm64));
1269      for (unsigned i = 0; i != 5; ++i)
1270        MIB.addOperand(MBBI->getOperand(i));
1271    } else if (RetOpcode == X86::TCRETURNri64) {
1272      BuildMI(MBB, MBBI, DL, TII.get(X86::TAILJMPr64), JumpTarget.getReg());
1273    } else {
1274      BuildMI(MBB, MBBI, DL, TII.get(X86::TAILJMPr), JumpTarget.getReg());
1275    }
1276
1277    MachineInstr *NewMI = prior(MBBI);
1278    for (unsigned i = 2, e = MBBI->getNumOperands(); i != e; ++i)
1279      NewMI->addOperand(MBBI->getOperand(i));
1280
1281    // Delete the pseudo instruction TCRETURN.
1282    MBB.erase(MBBI);
1283  } else if ((RetOpcode == X86::RET || RetOpcode == X86::RETI) &&
1284             (X86FI->getTCReturnAddrDelta() < 0)) {
1285    // Add the return addr area delta back since we are not tail calling.
1286    int delta = -1*X86FI->getTCReturnAddrDelta();
1287    MBBI = prior(MBB.end());
1288
1289    // Check for possible merge with preceeding ADD instruction.
1290    delta += mergeSPUpdates(MBB, MBBI, StackPtr, true);
1291    emitSPUpdate(MBB, MBBI, StackPtr, delta, Is64Bit, TII);
1292  }
1293}
1294
1295unsigned X86RegisterInfo::getRARegister() const {
1296  return Is64Bit ? X86::RIP     // Should have dwarf #16.
1297                 : X86::EIP;    // Should have dwarf #8.
1298}
1299
1300unsigned X86RegisterInfo::getFrameRegister(const MachineFunction &MF) const {
1301  return hasFP(MF) ? FramePtr : StackPtr;
1302}
1303
1304void
1305X86RegisterInfo::getInitialFrameState(std::vector<MachineMove> &Moves) const {
1306  // Calculate amount of bytes used for return address storing
1307  int stackGrowth = (Is64Bit ? -8 : -4);
1308
1309  // Initial state of the frame pointer is esp+stackGrowth.
1310  MachineLocation Dst(MachineLocation::VirtualFP);
1311  MachineLocation Src(StackPtr, stackGrowth);
1312  Moves.push_back(MachineMove(0, Dst, Src));
1313
1314  // Add return address to move list
1315  MachineLocation CSDst(StackPtr, stackGrowth);
1316  MachineLocation CSSrc(getRARegister());
1317  Moves.push_back(MachineMove(0, CSDst, CSSrc));
1318}
1319
1320unsigned X86RegisterInfo::getEHExceptionRegister() const {
1321  llvm_unreachable("What is the exception register");
1322  return 0;
1323}
1324
1325unsigned X86RegisterInfo::getEHHandlerRegister() const {
1326  llvm_unreachable("What is the exception handler register");
1327  return 0;
1328}
1329
1330namespace llvm {
1331unsigned getX86SubSuperRegister(unsigned Reg, EVT VT, bool High) {
1332  switch (VT.getSimpleVT().SimpleTy) {
1333  default: return Reg;
1334  case MVT::i8:
1335    if (High) {
1336      switch (Reg) {
1337      default: return 0;
1338      case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
1339        return X86::AH;
1340      case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
1341        return X86::DH;
1342      case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
1343        return X86::CH;
1344      case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
1345        return X86::BH;
1346      }
1347    } else {
1348      switch (Reg) {
1349      default: return 0;
1350      case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
1351        return X86::AL;
1352      case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
1353        return X86::DL;
1354      case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
1355        return X86::CL;
1356      case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
1357        return X86::BL;
1358      case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
1359        return X86::SIL;
1360      case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
1361        return X86::DIL;
1362      case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
1363        return X86::BPL;
1364      case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
1365        return X86::SPL;
1366      case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
1367        return X86::R8B;
1368      case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
1369        return X86::R9B;
1370      case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
1371        return X86::R10B;
1372      case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
1373        return X86::R11B;
1374      case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
1375        return X86::R12B;
1376      case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
1377        return X86::R13B;
1378      case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
1379        return X86::R14B;
1380      case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
1381        return X86::R15B;
1382      }
1383    }
1384  case MVT::i16:
1385    switch (Reg) {
1386    default: return Reg;
1387    case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
1388      return X86::AX;
1389    case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
1390      return X86::DX;
1391    case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
1392      return X86::CX;
1393    case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
1394      return X86::BX;
1395    case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
1396      return X86::SI;
1397    case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
1398      return X86::DI;
1399    case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
1400      return X86::BP;
1401    case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
1402      return X86::SP;
1403    case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
1404      return X86::R8W;
1405    case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
1406      return X86::R9W;
1407    case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
1408      return X86::R10W;
1409    case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
1410      return X86::R11W;
1411    case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
1412      return X86::R12W;
1413    case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
1414      return X86::R13W;
1415    case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
1416      return X86::R14W;
1417    case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
1418      return X86::R15W;
1419    }
1420  case MVT::i32:
1421    switch (Reg) {
1422    default: return Reg;
1423    case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
1424      return X86::EAX;
1425    case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
1426      return X86::EDX;
1427    case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
1428      return X86::ECX;
1429    case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
1430      return X86::EBX;
1431    case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
1432      return X86::ESI;
1433    case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
1434      return X86::EDI;
1435    case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
1436      return X86::EBP;
1437    case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
1438      return X86::ESP;
1439    case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
1440      return X86::R8D;
1441    case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
1442      return X86::R9D;
1443    case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
1444      return X86::R10D;
1445    case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
1446      return X86::R11D;
1447    case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
1448      return X86::R12D;
1449    case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
1450      return X86::R13D;
1451    case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
1452      return X86::R14D;
1453    case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
1454      return X86::R15D;
1455    }
1456  case MVT::i64:
1457    switch (Reg) {
1458    default: return Reg;
1459    case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
1460      return X86::RAX;
1461    case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
1462      return X86::RDX;
1463    case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
1464      return X86::RCX;
1465    case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
1466      return X86::RBX;
1467    case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
1468      return X86::RSI;
1469    case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
1470      return X86::RDI;
1471    case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
1472      return X86::RBP;
1473    case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
1474      return X86::RSP;
1475    case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
1476      return X86::R8;
1477    case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
1478      return X86::R9;
1479    case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
1480      return X86::R10;
1481    case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
1482      return X86::R11;
1483    case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
1484      return X86::R12;
1485    case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
1486      return X86::R13;
1487    case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
1488      return X86::R14;
1489    case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
1490      return X86::R15;
1491    }
1492  }
1493
1494  return Reg;
1495}
1496}
1497
1498#include "X86GenRegisterInfo.inc"
1499
1500namespace {
1501  struct MSAH : public MachineFunctionPass {
1502    static char ID;
1503    MSAH() : MachineFunctionPass(&ID) {}
1504
1505    virtual bool runOnMachineFunction(MachineFunction &MF) {
1506      const X86TargetMachine *TM =
1507        static_cast<const X86TargetMachine *>(&MF.getTarget());
1508      const X86RegisterInfo *X86RI = TM->getRegisterInfo();
1509      MachineRegisterInfo &RI = MF.getRegInfo();
1510      X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1511      unsigned StackAlignment = X86RI->getStackAlignment();
1512
1513      // Be over-conservative: scan over all vreg defs and find whether vector
1514      // registers are used. If yes, there is a possibility that vector register
1515      // will be spilled and thus require dynamic stack realignment.
1516      for (unsigned RegNum = TargetRegisterInfo::FirstVirtualRegister;
1517           RegNum < RI.getLastVirtReg(); ++RegNum)
1518        if (RI.getRegClass(RegNum)->getAlignment() > StackAlignment) {
1519          FuncInfo->setReserveFP(true);
1520          return true;
1521        }
1522
1523      // Nothing to do
1524      return false;
1525    }
1526
1527    virtual const char *getPassName() const {
1528      return "X86 Maximal Stack Alignment Check";
1529    }
1530
1531    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1532      AU.setPreservesCFG();
1533      MachineFunctionPass::getAnalysisUsage(AU);
1534    }
1535  };
1536
1537  char MSAH::ID = 0;
1538}
1539
1540FunctionPass*
1541llvm::createX86MaxStackAlignmentHeuristicPass() { return new MSAH(); }
1542