1 //===-- X86RegisterInfo.cpp - X86 Register Information --------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains the X86 implementation of the TargetRegisterInfo class.
11 // This file is responsible for the frame pointer elimination optimization
12 // on X86.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "X86RegisterInfo.h"
17 #include "X86FrameLowering.h"
18 #include "X86InstrBuilder.h"
19 #include "X86MachineFunctionInfo.h"
20 #include "X86Subtarget.h"
21 #include "X86TargetMachine.h"
22 #include "llvm/ADT/BitVector.h"
23 #include "llvm/ADT/STLExtras.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineFunctionPass.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineModuleInfo.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/MachineValueType.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/Type.h"
34 #include "llvm/MC/MCAsmInfo.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Target/TargetFrameLowering.h"
38 #include "llvm/Target/TargetInstrInfo.h"
39 #include "llvm/Target/TargetMachine.h"
40 #include "llvm/Target/TargetOptions.h"
41 
42 using namespace llvm;
43 
44 #define GET_REGINFO_TARGET_DESC
45 #include "X86GenRegisterInfo.inc"
46 
47 cl::opt<bool>
48 ForceStackAlign("force-align-stack",
49                  cl::desc("Force align the stack to the minimum alignment"
50                            " needed for the function."),
51                  cl::init(false), cl::Hidden);
52 
53 static cl::opt<bool>
54 EnableBasePointer("x86-use-base-pointer", cl::Hidden, cl::init(true),
55           cl::desc("Enable use of a base pointer for complex stack frames"));
56 
57 X86RegisterInfo::X86RegisterInfo(const Triple &TT)
58     : X86GenRegisterInfo((TT.isArch64Bit() ? X86::RIP : X86::EIP),
59                          X86_MC::getDwarfRegFlavour(TT, false),
60                          X86_MC::getDwarfRegFlavour(TT, true),
61                          (TT.isArch64Bit() ? X86::RIP : X86::EIP)) {
62   X86_MC::InitLLVM2SEHRegisterMapping(this);
63 
64   // Cache some information.
65   Is64Bit = TT.isArch64Bit();
66   IsWin64 = Is64Bit && TT.isOSWindows();
67 
68   // Use a callee-saved register as the base pointer.  These registers must
69   // not conflict with any ABI requirements.  For example, in 32-bit mode PIC
70   // requires GOT in the EBX register before function calls via PLT GOT pointer.
71   if (Is64Bit) {
72     SlotSize = 8;
73     // This matches the simplified 32-bit pointer code in the data layout
74     // computation.
75     // FIXME: Should use the data layout?
76     bool Use64BitReg = TT.getEnvironment() != Triple::GNUX32;
77     StackPtr = Use64BitReg ? X86::RSP : X86::ESP;
78     FramePtr = Use64BitReg ? X86::RBP : X86::EBP;
79     BasePtr = Use64BitReg ? X86::RBX : X86::EBX;
80   } else {
81     SlotSize = 4;
82     StackPtr = X86::ESP;
83     FramePtr = X86::EBP;
84     BasePtr = X86::ESI;
85   }
86 }
87 
88 bool
89 X86RegisterInfo::trackLivenessAfterRegAlloc(const MachineFunction &MF) const {
90   // ExeDepsFixer and PostRAScheduler require liveness.
91   return true;
92 }
93 
94 int
95 X86RegisterInfo::getSEHRegNum(unsigned i) const {
96   return getEncodingValue(i);
97 }
98 
99 const TargetRegisterClass *
100 X86RegisterInfo::getSubClassWithSubReg(const TargetRegisterClass *RC,
101                                        unsigned Idx) const {
102   // The sub_8bit sub-register index is more constrained in 32-bit mode.
103   // It behaves just like the sub_8bit_hi index.
104   if (!Is64Bit && Idx == X86::sub_8bit)
105     Idx = X86::sub_8bit_hi;
106 
107   // Forward to TableGen's default version.
108   return X86GenRegisterInfo::getSubClassWithSubReg(RC, Idx);
109 }
110 
111 const TargetRegisterClass *
112 X86RegisterInfo::getMatchingSuperRegClass(const TargetRegisterClass *A,
113                                           const TargetRegisterClass *B,
114                                           unsigned SubIdx) const {
115   // The sub_8bit sub-register index is more constrained in 32-bit mode.
116   if (!Is64Bit && SubIdx == X86::sub_8bit) {
117     A = X86GenRegisterInfo::getSubClassWithSubReg(A, X86::sub_8bit_hi);
118     if (!A)
119       return nullptr;
120   }
121   return X86GenRegisterInfo::getMatchingSuperRegClass(A, B, SubIdx);
122 }
123 
124 const TargetRegisterClass *
125 X86RegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC,
126                                            const MachineFunction &MF) const {
127   // Don't allow super-classes of GR8_NOREX.  This class is only used after
128   // extracting sub_8bit_hi sub-registers.  The H sub-registers cannot be copied
129   // to the full GR8 register class in 64-bit mode, so we cannot allow the
130   // reigster class inflation.
131   //
132   // The GR8_NOREX class is always used in a way that won't be constrained to a
133   // sub-class, so sub-classes like GR8_ABCD_L are allowed to expand to the
134   // full GR8 class.
135   if (RC == &X86::GR8_NOREXRegClass)
136     return RC;
137 
138   const TargetRegisterClass *Super = RC;
139   TargetRegisterClass::sc_iterator I = RC->getSuperClasses();
140   do {
141     switch (Super->getID()) {
142     case X86::GR8RegClassID:
143     case X86::GR16RegClassID:
144     case X86::GR32RegClassID:
145     case X86::GR64RegClassID:
146     case X86::FR32RegClassID:
147     case X86::FR64RegClassID:
148     case X86::RFP32RegClassID:
149     case X86::RFP64RegClassID:
150     case X86::RFP80RegClassID:
151     case X86::VR128RegClassID:
152     case X86::VR256RegClassID:
153       // Don't return a super-class that would shrink the spill size.
154       // That can happen with the vector and float classes.
155       if (Super->getSize() == RC->getSize())
156         return Super;
157     }
158     Super = *I++;
159   } while (Super);
160   return RC;
161 }
162 
163 const TargetRegisterClass *
164 X86RegisterInfo::getPointerRegClass(const MachineFunction &MF,
165                                     unsigned Kind) const {
166   const X86Subtarget &Subtarget = MF.getSubtarget<X86Subtarget>();
167   switch (Kind) {
168   default: llvm_unreachable("Unexpected Kind in getPointerRegClass!");
169   case 0: // Normal GPRs.
170     if (Subtarget.isTarget64BitLP64())
171       return &X86::GR64RegClass;
172     return &X86::GR32RegClass;
173   case 1: // Normal GPRs except the stack pointer (for encoding reasons).
174     if (Subtarget.isTarget64BitLP64())
175       return &X86::GR64_NOSPRegClass;
176     return &X86::GR32_NOSPRegClass;
177   case 2: // Available for tailcall (not callee-saved GPRs).
178     const Function *F = MF.getFunction();
179     if (IsWin64 || (F && F->getCallingConv() == CallingConv::X86_64_Win64))
180       return &X86::GR64_TCW64RegClass;
181     else if (Is64Bit)
182       return &X86::GR64_TCRegClass;
183 
184     bool hasHipeCC = (F ? F->getCallingConv() == CallingConv::HiPE : false);
185     if (hasHipeCC)
186       return &X86::GR32RegClass;
187     return &X86::GR32_TCRegClass;
188   }
189 }
190 
191 const TargetRegisterClass *
192 X86RegisterInfo::getCrossCopyRegClass(const TargetRegisterClass *RC) const {
193   if (RC == &X86::CCRRegClass) {
194     if (Is64Bit)
195       return &X86::GR64RegClass;
196     else
197       return &X86::GR32RegClass;
198   }
199   return RC;
200 }
201 
202 unsigned
203 X86RegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC,
204                                      MachineFunction &MF) const {
205   const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
206 
207   unsigned FPDiff = TFI->hasFP(MF) ? 1 : 0;
208   switch (RC->getID()) {
209   default:
210     return 0;
211   case X86::GR32RegClassID:
212     return 4 - FPDiff;
213   case X86::GR64RegClassID:
214     return 12 - FPDiff;
215   case X86::VR128RegClassID:
216     return Is64Bit ? 10 : 4;
217   case X86::VR64RegClassID:
218     return 4;
219   }
220 }
221 
222 const MCPhysReg *
223 X86RegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
224   const X86Subtarget &Subtarget = MF->getSubtarget<X86Subtarget>();
225   bool HasAVX = Subtarget.hasAVX();
226   bool HasAVX512 = Subtarget.hasAVX512();
227   bool CallsEHReturn = MF->getMMI().callsEHReturn();
228 
229   assert(MF && "MachineFunction required");
230   switch (MF->getFunction()->getCallingConv()) {
231   case CallingConv::GHC:
232   case CallingConv::HiPE:
233     return CSR_NoRegs_SaveList;
234   case CallingConv::AnyReg:
235     if (HasAVX)
236       return CSR_64_AllRegs_AVX_SaveList;
237     return CSR_64_AllRegs_SaveList;
238   case CallingConv::PreserveMost:
239     return CSR_64_RT_MostRegs_SaveList;
240   case CallingConv::PreserveAll:
241     if (HasAVX)
242       return CSR_64_RT_AllRegs_AVX_SaveList;
243     return CSR_64_RT_AllRegs_SaveList;
244   case CallingConv::Intel_OCL_BI: {
245     if (HasAVX512 && IsWin64)
246       return CSR_Win64_Intel_OCL_BI_AVX512_SaveList;
247     if (HasAVX512 && Is64Bit)
248       return CSR_64_Intel_OCL_BI_AVX512_SaveList;
249     if (HasAVX && IsWin64)
250       return CSR_Win64_Intel_OCL_BI_AVX_SaveList;
251     if (HasAVX && Is64Bit)
252       return CSR_64_Intel_OCL_BI_AVX_SaveList;
253     if (!HasAVX && !IsWin64 && Is64Bit)
254       return CSR_64_Intel_OCL_BI_SaveList;
255     break;
256   }
257   case CallingConv::Cold:
258     if (Is64Bit)
259       return CSR_64_MostRegs_SaveList;
260     break;
261   case CallingConv::X86_64_Win64:
262     return CSR_Win64_SaveList;
263   case CallingConv::X86_64_SysV:
264     if (CallsEHReturn)
265       return CSR_64EHRet_SaveList;
266     return CSR_64_SaveList;
267   default:
268     break;
269   }
270 
271   if (Is64Bit) {
272     if (IsWin64)
273       return CSR_Win64_SaveList;
274     if (CallsEHReturn)
275       return CSR_64EHRet_SaveList;
276     return CSR_64_SaveList;
277   }
278   if (CallsEHReturn)
279     return CSR_32EHRet_SaveList;
280   return CSR_32_SaveList;
281 }
282 
283 const uint32_t *
284 X86RegisterInfo::getCallPreservedMask(const MachineFunction &MF,
285                                       CallingConv::ID CC) const {
286   const X86Subtarget &Subtarget = MF.getSubtarget<X86Subtarget>();
287   bool HasAVX = Subtarget.hasAVX();
288   bool HasAVX512 = Subtarget.hasAVX512();
289 
290   switch (CC) {
291   case CallingConv::GHC:
292   case CallingConv::HiPE:
293     return CSR_NoRegs_RegMask;
294   case CallingConv::AnyReg:
295     if (HasAVX)
296       return CSR_64_AllRegs_AVX_RegMask;
297     return CSR_64_AllRegs_RegMask;
298   case CallingConv::PreserveMost:
299     return CSR_64_RT_MostRegs_RegMask;
300   case CallingConv::PreserveAll:
301     if (HasAVX)
302       return CSR_64_RT_AllRegs_AVX_RegMask;
303     return CSR_64_RT_AllRegs_RegMask;
304   case CallingConv::Intel_OCL_BI: {
305     if (HasAVX512 && IsWin64)
306       return CSR_Win64_Intel_OCL_BI_AVX512_RegMask;
307     if (HasAVX512 && Is64Bit)
308       return CSR_64_Intel_OCL_BI_AVX512_RegMask;
309     if (HasAVX && IsWin64)
310       return CSR_Win64_Intel_OCL_BI_AVX_RegMask;
311     if (HasAVX && Is64Bit)
312       return CSR_64_Intel_OCL_BI_AVX_RegMask;
313     if (!HasAVX && !IsWin64 && Is64Bit)
314       return CSR_64_Intel_OCL_BI_RegMask;
315     break;
316   }
317   case CallingConv::Cold:
318     if (Is64Bit)
319       return CSR_64_MostRegs_RegMask;
320     break;
321   default:
322     break;
323   case CallingConv::X86_64_Win64:
324     return CSR_Win64_RegMask;
325   case CallingConv::X86_64_SysV:
326     return CSR_64_RegMask;
327   }
328 
329   // Unlike getCalleeSavedRegs(), we don't have MMI so we can't check
330   // callsEHReturn().
331   if (Is64Bit) {
332     if (IsWin64)
333       return CSR_Win64_RegMask;
334     return CSR_64_RegMask;
335   }
336   return CSR_32_RegMask;
337 }
338 
339 const uint32_t*
340 X86RegisterInfo::getNoPreservedMask() const {
341   return CSR_NoRegs_RegMask;
342 }
343 
344 BitVector X86RegisterInfo::getReservedRegs(const MachineFunction &MF) const {
345   BitVector Reserved(getNumRegs());
346   const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
347 
348   // Set the stack-pointer register and its aliases as reserved.
349   for (MCSubRegIterator I(X86::RSP, this, /*IncludeSelf=*/true); I.isValid();
350        ++I)
351     Reserved.set(*I);
352 
353   // Set the instruction pointer register and its aliases as reserved.
354   for (MCSubRegIterator I(X86::RIP, this, /*IncludeSelf=*/true); I.isValid();
355        ++I)
356     Reserved.set(*I);
357 
358   // Set the frame-pointer register and its aliases as reserved if needed.
359   if (TFI->hasFP(MF)) {
360     for (MCSubRegIterator I(X86::RBP, this, /*IncludeSelf=*/true); I.isValid();
361          ++I)
362       Reserved.set(*I);
363   }
364 
365   // Set the base-pointer register and its aliases as reserved if needed.
366   if (hasBasePointer(MF)) {
367     CallingConv::ID CC = MF.getFunction()->getCallingConv();
368     const uint32_t *RegMask = getCallPreservedMask(MF, CC);
369     if (MachineOperand::clobbersPhysReg(RegMask, getBaseRegister()))
370       report_fatal_error(
371         "Stack realignment in presence of dynamic allocas is not supported with"
372         "this calling convention.");
373 
374     unsigned BasePtr = getX86SubSuperRegister(getBaseRegister(), MVT::i64,
375                                               false);
376     for (MCSubRegIterator I(BasePtr, this, /*IncludeSelf=*/true);
377          I.isValid(); ++I)
378       Reserved.set(*I);
379   }
380 
381   // Mark the segment registers as reserved.
382   Reserved.set(X86::CS);
383   Reserved.set(X86::SS);
384   Reserved.set(X86::DS);
385   Reserved.set(X86::ES);
386   Reserved.set(X86::FS);
387   Reserved.set(X86::GS);
388 
389   // Mark the floating point stack registers as reserved.
390   for (unsigned n = 0; n != 8; ++n)
391     Reserved.set(X86::ST0 + n);
392 
393   // Reserve the registers that only exist in 64-bit mode.
394   if (!Is64Bit) {
395     // These 8-bit registers are part of the x86-64 extension even though their
396     // super-registers are old 32-bits.
397     Reserved.set(X86::SIL);
398     Reserved.set(X86::DIL);
399     Reserved.set(X86::BPL);
400     Reserved.set(X86::SPL);
401 
402     for (unsigned n = 0; n != 8; ++n) {
403       // R8, R9, ...
404       for (MCRegAliasIterator AI(X86::R8 + n, this, true); AI.isValid(); ++AI)
405         Reserved.set(*AI);
406 
407       // XMM8, XMM9, ...
408       for (MCRegAliasIterator AI(X86::XMM8 + n, this, true); AI.isValid(); ++AI)
409         Reserved.set(*AI);
410     }
411   }
412   if (!Is64Bit || !MF.getSubtarget<X86Subtarget>().hasAVX512()) {
413     for (unsigned n = 16; n != 32; ++n) {
414       for (MCRegAliasIterator AI(X86::XMM0 + n, this, true); AI.isValid(); ++AI)
415         Reserved.set(*AI);
416     }
417   }
418 
419   return Reserved;
420 }
421 
422 void X86RegisterInfo::adjustStackMapLiveOutMask(uint32_t *Mask) const {
423   // Check if the EFLAGS register is marked as live-out. This shouldn't happen,
424   // because the calling convention defines the EFLAGS register as NOT
425   // preserved.
426   //
427   // Unfortunatelly the EFLAGS show up as live-out after branch folding. Adding
428   // an assert to track this and clear the register afterwards to avoid
429   // unnecessary crashes during release builds.
430   assert(!(Mask[X86::EFLAGS / 32] & (1U << (X86::EFLAGS % 32))) &&
431          "EFLAGS are not live-out from a patchpoint.");
432 
433   // Also clean other registers that don't need preserving (IP).
434   for (auto Reg : {X86::EFLAGS, X86::RIP, X86::EIP, X86::IP})
435     Mask[Reg / 32] &= ~(1U << (Reg % 32));
436 }
437 
438 //===----------------------------------------------------------------------===//
439 // Stack Frame Processing methods
440 //===----------------------------------------------------------------------===//
441 
442 bool X86RegisterInfo::hasBasePointer(const MachineFunction &MF) const {
443    const MachineFrameInfo *MFI = MF.getFrameInfo();
444 
445    if (!EnableBasePointer)
446      return false;
447 
448    // When we need stack realignment, we can't address the stack from the frame
449    // pointer.  When we have dynamic allocas or stack-adjusting inline asm, we
450    // can't address variables from the stack pointer.  MS inline asm can
451    // reference locals while also adjusting the stack pointer.  When we can't
452    // use both the SP and the FP, we need a separate base pointer register.
453    bool CantUseFP = needsStackRealignment(MF);
454    bool CantUseSP =
455        MFI->hasVarSizedObjects() || MFI->hasInlineAsmWithSPAdjust();
456    return CantUseFP && CantUseSP;
457 }
458 
459 bool X86RegisterInfo::canRealignStack(const MachineFunction &MF) const {
460   if (MF.getFunction()->hasFnAttribute("no-realign-stack"))
461     return false;
462 
463   const MachineFrameInfo *MFI = MF.getFrameInfo();
464   const MachineRegisterInfo *MRI = &MF.getRegInfo();
465 
466   // Stack realignment requires a frame pointer.  If we already started
467   // register allocation with frame pointer elimination, it is too late now.
468   if (!MRI->canReserveReg(FramePtr))
469     return false;
470 
471   // If a base pointer is necessary.  Check that it isn't too late to reserve
472   // it.
473   if (MFI->hasVarSizedObjects())
474     return MRI->canReserveReg(BasePtr);
475   return true;
476 }
477 
478 bool X86RegisterInfo::needsStackRealignment(const MachineFunction &MF) const {
479   const MachineFrameInfo *MFI = MF.getFrameInfo();
480   const Function *F = MF.getFunction();
481   unsigned StackAlign =
482     MF.getSubtarget().getFrameLowering()->getStackAlignment();
483   bool requiresRealignment = ((MFI->getMaxAlignment() > StackAlign) ||
484                               F->hasFnAttribute(Attribute::StackAlignment));
485 
486   // If we've requested that we force align the stack do so now.
487   if (ForceStackAlign)
488     return canRealignStack(MF);
489 
490   return requiresRealignment && canRealignStack(MF);
491 }
492 
493 bool X86RegisterInfo::hasReservedSpillSlot(const MachineFunction &MF,
494                                            unsigned Reg, int &FrameIdx) const {
495   // Since X86 defines assignCalleeSavedSpillSlots which always return true
496   // this function neither used nor tested.
497   llvm_unreachable("Unused function on X86. Otherwise need a test case.");
498 }
499 
500 void
501 X86RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
502                                      int SPAdj, unsigned FIOperandNum,
503                                      RegScavenger *RS) const {
504   MachineInstr &MI = *II;
505   MachineFunction &MF = *MI.getParent()->getParent();
506   const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
507   int FrameIndex = MI.getOperand(FIOperandNum).getIndex();
508   unsigned BasePtr;
509 
510   unsigned Opc = MI.getOpcode();
511   bool AfterFPPop = Opc == X86::TAILJMPm64 || Opc == X86::TAILJMPm ||
512                     Opc == X86::TCRETURNmi || Opc == X86::TCRETURNmi64;
513   if (hasBasePointer(MF))
514     BasePtr = (FrameIndex < 0 ? FramePtr : getBaseRegister());
515   else if (needsStackRealignment(MF))
516     BasePtr = (FrameIndex < 0 ? FramePtr : StackPtr);
517   else if (AfterFPPop)
518     BasePtr = StackPtr;
519   else
520     BasePtr = (TFI->hasFP(MF) ? FramePtr : StackPtr);
521 
522   // FRAME_ALLOC uses a single offset, with no register. It only works in the
523   // simple FP case, and doesn't work with stack realignment. On 32-bit, the
524   // offset is from the traditional base pointer location.  On 64-bit, the
525   // offset is from the SP at the end of the prologue, not the FP location. This
526   // matches the behavior of llvm.frameaddress.
527   if (Opc == TargetOpcode::FRAME_ALLOC) {
528     MachineOperand &FI = MI.getOperand(FIOperandNum);
529     bool IsWinEH = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
530     int Offset;
531     if (IsWinEH)
532       Offset = static_cast<const X86FrameLowering *>(TFI)
533                      ->getFrameIndexOffsetFromSP(MF, FrameIndex);
534     else
535       Offset = TFI->getFrameIndexOffset(MF, FrameIndex);
536     FI.ChangeToImmediate(Offset);
537     return;
538   }
539 
540   // For LEA64_32r when BasePtr is 32-bits (X32) we can use full-size 64-bit
541   // register as source operand, semantic is the same and destination is
542   // 32-bits. It saves one byte per lea in code since 0x67 prefix is avoided.
543   if (Opc == X86::LEA64_32r && X86::GR32RegClass.contains(BasePtr))
544     BasePtr = getX86SubSuperRegister(BasePtr, MVT::i64, false);
545 
546   // This must be part of a four operand memory reference.  Replace the
547   // FrameIndex with base register with EBP.  Add an offset to the offset.
548   MI.getOperand(FIOperandNum).ChangeToRegister(BasePtr, false);
549 
550   // Now add the frame object offset to the offset from EBP.
551   int FIOffset;
552   if (AfterFPPop) {
553     // Tail call jmp happens after FP is popped.
554     const MachineFrameInfo *MFI = MF.getFrameInfo();
555     FIOffset = MFI->getObjectOffset(FrameIndex) - TFI->getOffsetOfLocalArea();
556   } else
557     FIOffset = TFI->getFrameIndexOffset(MF, FrameIndex);
558 
559   if (BasePtr == StackPtr)
560     FIOffset += SPAdj;
561 
562   // The frame index format for stackmaps and patchpoints is different from the
563   // X86 format. It only has a FI and an offset.
564   if (Opc == TargetOpcode::STACKMAP || Opc == TargetOpcode::PATCHPOINT) {
565     assert(BasePtr == FramePtr && "Expected the FP as base register");
566     int64_t Offset = MI.getOperand(FIOperandNum + 1).getImm() + FIOffset;
567     MI.getOperand(FIOperandNum + 1).ChangeToImmediate(Offset);
568     return;
569   }
570 
571   if (MI.getOperand(FIOperandNum+3).isImm()) {
572     // Offset is a 32-bit integer.
573     int Imm = (int)(MI.getOperand(FIOperandNum + 3).getImm());
574     int Offset = FIOffset + Imm;
575     assert((!Is64Bit || isInt<32>((long long)FIOffset + Imm)) &&
576            "Requesting 64-bit offset in 32-bit immediate!");
577     MI.getOperand(FIOperandNum + 3).ChangeToImmediate(Offset);
578   } else {
579     // Offset is symbolic. This is extremely rare.
580     uint64_t Offset = FIOffset +
581       (uint64_t)MI.getOperand(FIOperandNum+3).getOffset();
582     MI.getOperand(FIOperandNum + 3).setOffset(Offset);
583   }
584 }
585 
586 unsigned X86RegisterInfo::getFrameRegister(const MachineFunction &MF) const {
587   const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
588   return TFI->hasFP(MF) ? FramePtr : StackPtr;
589 }
590 
591 unsigned
592 X86RegisterInfo::getPtrSizedFrameRegister(const MachineFunction &MF) const {
593   const X86Subtarget &Subtarget = MF.getSubtarget<X86Subtarget>();
594   unsigned FrameReg = getFrameRegister(MF);
595   if (Subtarget.isTarget64BitILP32())
596     FrameReg = getX86SubSuperRegister(FrameReg, MVT::i32, false);
597   return FrameReg;
598 }
599 
600 namespace llvm {
601 unsigned getX86SubSuperRegister(unsigned Reg, MVT::SimpleValueType VT,
602                                 bool High) {
603   switch (VT) {
604   default: llvm_unreachable("Unexpected VT");
605   case MVT::i8:
606     if (High) {
607       switch (Reg) {
608       default: return getX86SubSuperRegister(Reg, MVT::i64);
609       case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
610         return X86::SI;
611       case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
612         return X86::DI;
613       case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
614         return X86::BP;
615       case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
616         return X86::SP;
617       case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
618         return X86::AH;
619       case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
620         return X86::DH;
621       case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
622         return X86::CH;
623       case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
624         return X86::BH;
625       }
626     } else {
627       switch (Reg) {
628       default: llvm_unreachable("Unexpected register");
629       case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
630         return X86::AL;
631       case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
632         return X86::DL;
633       case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
634         return X86::CL;
635       case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
636         return X86::BL;
637       case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
638         return X86::SIL;
639       case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
640         return X86::DIL;
641       case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
642         return X86::BPL;
643       case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
644         return X86::SPL;
645       case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
646         return X86::R8B;
647       case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
648         return X86::R9B;
649       case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
650         return X86::R10B;
651       case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
652         return X86::R11B;
653       case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
654         return X86::R12B;
655       case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
656         return X86::R13B;
657       case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
658         return X86::R14B;
659       case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
660         return X86::R15B;
661       }
662     }
663   case MVT::i16:
664     switch (Reg) {
665     default: llvm_unreachable("Unexpected register");
666     case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
667       return X86::AX;
668     case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
669       return X86::DX;
670     case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
671       return X86::CX;
672     case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
673       return X86::BX;
674     case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
675       return X86::SI;
676     case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
677       return X86::DI;
678     case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
679       return X86::BP;
680     case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
681       return X86::SP;
682     case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
683       return X86::R8W;
684     case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
685       return X86::R9W;
686     case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
687       return X86::R10W;
688     case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
689       return X86::R11W;
690     case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
691       return X86::R12W;
692     case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
693       return X86::R13W;
694     case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
695       return X86::R14W;
696     case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
697       return X86::R15W;
698     }
699   case MVT::i32:
700     switch (Reg) {
701     default: llvm_unreachable("Unexpected register");
702     case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
703       return X86::EAX;
704     case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
705       return X86::EDX;
706     case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
707       return X86::ECX;
708     case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
709       return X86::EBX;
710     case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
711       return X86::ESI;
712     case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
713       return X86::EDI;
714     case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
715       return X86::EBP;
716     case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
717       return X86::ESP;
718     case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
719       return X86::R8D;
720     case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
721       return X86::R9D;
722     case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
723       return X86::R10D;
724     case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
725       return X86::R11D;
726     case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
727       return X86::R12D;
728     case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
729       return X86::R13D;
730     case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
731       return X86::R14D;
732     case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
733       return X86::R15D;
734     }
735   case MVT::i64:
736     switch (Reg) {
737     default: llvm_unreachable("Unexpected register");
738     case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
739       return X86::RAX;
740     case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
741       return X86::RDX;
742     case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
743       return X86::RCX;
744     case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
745       return X86::RBX;
746     case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
747       return X86::RSI;
748     case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
749       return X86::RDI;
750     case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
751       return X86::RBP;
752     case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
753       return X86::RSP;
754     case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
755       return X86::R8;
756     case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
757       return X86::R9;
758     case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
759       return X86::R10;
760     case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
761       return X86::R11;
762     case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
763       return X86::R12;
764     case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
765       return X86::R13;
766     case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
767       return X86::R14;
768     case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
769       return X86::R15;
770     }
771   }
772 }
773 
774 unsigned get512BitSuperRegister(unsigned Reg) {
775   if (Reg >= X86::XMM0 && Reg <= X86::XMM31)
776     return X86::ZMM0 + (Reg - X86::XMM0);
777   if (Reg >= X86::YMM0 && Reg <= X86::YMM31)
778     return X86::ZMM0 + (Reg - X86::YMM0);
779   if (Reg >= X86::ZMM0 && Reg <= X86::ZMM31)
780     return Reg;
781   llvm_unreachable("Unexpected SIMD register");
782 }
783 
784 }
785