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
X86RegisterInfo(const Triple & TT)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
trackLivenessAfterRegAlloc(const MachineFunction & MF) const89 X86RegisterInfo::trackLivenessAfterRegAlloc(const MachineFunction &MF) const {
90 // ExeDepsFixer and PostRAScheduler require liveness.
91 return true;
92 }
93
94 int
getSEHRegNum(unsigned i) const95 X86RegisterInfo::getSEHRegNum(unsigned i) const {
96 return getEncodingValue(i);
97 }
98
99 const TargetRegisterClass *
getSubClassWithSubReg(const TargetRegisterClass * RC,unsigned Idx) const100 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 *
getMatchingSuperRegClass(const TargetRegisterClass * A,const TargetRegisterClass * B,unsigned SubIdx) const112 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 *
getLargestLegalSuperClass(const TargetRegisterClass * RC,const MachineFunction & MF) const125 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 *
getPointerRegClass(const MachineFunction & MF,unsigned Kind) const164 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 *
getCrossCopyRegClass(const TargetRegisterClass * RC) const192 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
getRegPressureLimit(const TargetRegisterClass * RC,MachineFunction & MF) const203 X86RegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC,
204 MachineFunction &MF) const {
205 const X86FrameLowering *TFI = getFrameLowering(MF);
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 *
getCalleeSavedRegs(const MachineFunction * MF) const223 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 *
getCallPreservedMask(const MachineFunction & MF,CallingConv::ID CC) const284 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*
getNoPreservedMask() const340 X86RegisterInfo::getNoPreservedMask() const {
341 return CSR_NoRegs_RegMask;
342 }
343
getReservedRegs(const MachineFunction & MF) const344 BitVector X86RegisterInfo::getReservedRegs(const MachineFunction &MF) const {
345 BitVector Reserved(getNumRegs());
346 const X86FrameLowering *TFI = getFrameLowering(MF);
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
adjustStackMapLiveOutMask(uint32_t * Mask) const422 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
hasBasePointer(const MachineFunction & MF) const442 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->hasOpaqueSPAdjustment();
456 return CantUseFP && CantUseSP;
457 }
458
canRealignStack(const MachineFunction & MF) const459 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
needsStackRealignment(const MachineFunction & MF) const478 bool X86RegisterInfo::needsStackRealignment(const MachineFunction &MF) const {
479 const MachineFrameInfo *MFI = MF.getFrameInfo();
480 const X86FrameLowering *TFI = getFrameLowering(MF);
481 const Function *F = MF.getFunction();
482 unsigned StackAlign = TFI->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
hasReservedSpillSlot(const MachineFunction & MF,unsigned Reg,int & FrameIdx) const493 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
eliminateFrameIndex(MachineBasicBlock::iterator II,int SPAdj,unsigned FIOperandNum,RegScavenger * RS) const501 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 X86FrameLowering *TFI = getFrameLowering(MF);
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 // LOCAL_ESCAPE 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::LOCAL_ESCAPE) {
528 MachineOperand &FI = MI.getOperand(FIOperandNum);
529 bool IsWinEH = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
530 int Offset;
531 if (IsWinEH)
532 Offset = TFI->getFrameIndexOffsetFromSP(MF, FrameIndex);
533 else
534 Offset = TFI->getFrameIndexOffset(MF, FrameIndex);
535 FI.ChangeToImmediate(Offset);
536 return;
537 }
538
539 // For LEA64_32r when BasePtr is 32-bits (X32) we can use full-size 64-bit
540 // register as source operand, semantic is the same and destination is
541 // 32-bits. It saves one byte per lea in code since 0x67 prefix is avoided.
542 if (Opc == X86::LEA64_32r && X86::GR32RegClass.contains(BasePtr))
543 BasePtr = getX86SubSuperRegister(BasePtr, MVT::i64, false);
544
545 // This must be part of a four operand memory reference. Replace the
546 // FrameIndex with base register with EBP. Add an offset to the offset.
547 MI.getOperand(FIOperandNum).ChangeToRegister(BasePtr, false);
548
549 // Now add the frame object offset to the offset from EBP.
550 int FIOffset;
551 if (AfterFPPop) {
552 // Tail call jmp happens after FP is popped.
553 const MachineFrameInfo *MFI = MF.getFrameInfo();
554 FIOffset = MFI->getObjectOffset(FrameIndex) - TFI->getOffsetOfLocalArea();
555 } else
556 FIOffset = TFI->getFrameIndexOffset(MF, FrameIndex);
557
558 if (BasePtr == StackPtr)
559 FIOffset += SPAdj;
560
561 // The frame index format for stackmaps and patchpoints is different from the
562 // X86 format. It only has a FI and an offset.
563 if (Opc == TargetOpcode::STACKMAP || Opc == TargetOpcode::PATCHPOINT) {
564 assert(BasePtr == FramePtr && "Expected the FP as base register");
565 int64_t Offset = MI.getOperand(FIOperandNum + 1).getImm() + FIOffset;
566 MI.getOperand(FIOperandNum + 1).ChangeToImmediate(Offset);
567 return;
568 }
569
570 if (MI.getOperand(FIOperandNum+3).isImm()) {
571 // Offset is a 32-bit integer.
572 int Imm = (int)(MI.getOperand(FIOperandNum + 3).getImm());
573 int Offset = FIOffset + Imm;
574 assert((!Is64Bit || isInt<32>((long long)FIOffset + Imm)) &&
575 "Requesting 64-bit offset in 32-bit immediate!");
576 MI.getOperand(FIOperandNum + 3).ChangeToImmediate(Offset);
577 } else {
578 // Offset is symbolic. This is extremely rare.
579 uint64_t Offset = FIOffset +
580 (uint64_t)MI.getOperand(FIOperandNum+3).getOffset();
581 MI.getOperand(FIOperandNum + 3).setOffset(Offset);
582 }
583 }
584
getFrameRegister(const MachineFunction & MF) const585 unsigned X86RegisterInfo::getFrameRegister(const MachineFunction &MF) const {
586 const X86FrameLowering *TFI = getFrameLowering(MF);
587 return TFI->hasFP(MF) ? FramePtr : StackPtr;
588 }
589
590 unsigned
getPtrSizedFrameRegister(const MachineFunction & MF) const591 X86RegisterInfo::getPtrSizedFrameRegister(const MachineFunction &MF) const {
592 const X86Subtarget &Subtarget = MF.getSubtarget<X86Subtarget>();
593 unsigned FrameReg = getFrameRegister(MF);
594 if (Subtarget.isTarget64BitILP32())
595 FrameReg = getX86SubSuperRegister(FrameReg, MVT::i32, false);
596 return FrameReg;
597 }
598
599 namespace llvm {
getX86SubSuperRegisterOrZero(unsigned Reg,MVT::SimpleValueType VT,bool High)600 unsigned getX86SubSuperRegisterOrZero(unsigned Reg, MVT::SimpleValueType VT,
601 bool High) {
602 switch (VT) {
603 default: return 0;
604 case MVT::i8:
605 if (High) {
606 switch (Reg) {
607 default: return getX86SubSuperRegister(Reg, MVT::i64);
608 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
609 return X86::SI;
610 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
611 return X86::DI;
612 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
613 return X86::BP;
614 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
615 return X86::SP;
616 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
617 return X86::AH;
618 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
619 return X86::DH;
620 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
621 return X86::CH;
622 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
623 return X86::BH;
624 }
625 } else {
626 switch (Reg) {
627 default: return 0;
628 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
629 return X86::AL;
630 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
631 return X86::DL;
632 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
633 return X86::CL;
634 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
635 return X86::BL;
636 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
637 return X86::SIL;
638 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
639 return X86::DIL;
640 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
641 return X86::BPL;
642 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
643 return X86::SPL;
644 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
645 return X86::R8B;
646 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
647 return X86::R9B;
648 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
649 return X86::R10B;
650 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
651 return X86::R11B;
652 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
653 return X86::R12B;
654 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
655 return X86::R13B;
656 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
657 return X86::R14B;
658 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
659 return X86::R15B;
660 }
661 }
662 case MVT::i16:
663 switch (Reg) {
664 default: return 0;
665 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
666 return X86::AX;
667 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
668 return X86::DX;
669 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
670 return X86::CX;
671 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
672 return X86::BX;
673 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
674 return X86::SI;
675 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
676 return X86::DI;
677 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
678 return X86::BP;
679 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
680 return X86::SP;
681 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
682 return X86::R8W;
683 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
684 return X86::R9W;
685 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
686 return X86::R10W;
687 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
688 return X86::R11W;
689 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
690 return X86::R12W;
691 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
692 return X86::R13W;
693 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
694 return X86::R14W;
695 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
696 return X86::R15W;
697 }
698 case MVT::i32:
699 switch (Reg) {
700 default: return 0;
701 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
702 return X86::EAX;
703 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
704 return X86::EDX;
705 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
706 return X86::ECX;
707 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
708 return X86::EBX;
709 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
710 return X86::ESI;
711 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
712 return X86::EDI;
713 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
714 return X86::EBP;
715 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
716 return X86::ESP;
717 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
718 return X86::R8D;
719 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
720 return X86::R9D;
721 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
722 return X86::R10D;
723 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
724 return X86::R11D;
725 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
726 return X86::R12D;
727 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
728 return X86::R13D;
729 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
730 return X86::R14D;
731 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
732 return X86::R15D;
733 }
734 case MVT::i64:
735 switch (Reg) {
736 default: return 0;
737 case X86::AH: case X86::AL: case X86::AX: case X86::EAX: case X86::RAX:
738 return X86::RAX;
739 case X86::DH: case X86::DL: case X86::DX: case X86::EDX: case X86::RDX:
740 return X86::RDX;
741 case X86::CH: case X86::CL: case X86::CX: case X86::ECX: case X86::RCX:
742 return X86::RCX;
743 case X86::BH: case X86::BL: case X86::BX: case X86::EBX: case X86::RBX:
744 return X86::RBX;
745 case X86::SIL: case X86::SI: case X86::ESI: case X86::RSI:
746 return X86::RSI;
747 case X86::DIL: case X86::DI: case X86::EDI: case X86::RDI:
748 return X86::RDI;
749 case X86::BPL: case X86::BP: case X86::EBP: case X86::RBP:
750 return X86::RBP;
751 case X86::SPL: case X86::SP: case X86::ESP: case X86::RSP:
752 return X86::RSP;
753 case X86::R8B: case X86::R8W: case X86::R8D: case X86::R8:
754 return X86::R8;
755 case X86::R9B: case X86::R9W: case X86::R9D: case X86::R9:
756 return X86::R9;
757 case X86::R10B: case X86::R10W: case X86::R10D: case X86::R10:
758 return X86::R10;
759 case X86::R11B: case X86::R11W: case X86::R11D: case X86::R11:
760 return X86::R11;
761 case X86::R12B: case X86::R12W: case X86::R12D: case X86::R12:
762 return X86::R12;
763 case X86::R13B: case X86::R13W: case X86::R13D: case X86::R13:
764 return X86::R13;
765 case X86::R14B: case X86::R14W: case X86::R14D: case X86::R14:
766 return X86::R14;
767 case X86::R15B: case X86::R15W: case X86::R15D: case X86::R15:
768 return X86::R15;
769 }
770 }
771 }
772
getX86SubSuperRegister(unsigned Reg,MVT::SimpleValueType VT,bool High)773 unsigned getX86SubSuperRegister(unsigned Reg, MVT::SimpleValueType VT,
774 bool High) {
775 unsigned Res = getX86SubSuperRegisterOrZero(Reg, VT, High);
776 if (Res == 0)
777 llvm_unreachable("Unexpected register or VT");
778 return Res;
779 }
780
get512BitSuperRegister(unsigned Reg)781 unsigned get512BitSuperRegister(unsigned Reg) {
782 if (Reg >= X86::XMM0 && Reg <= X86::XMM31)
783 return X86::ZMM0 + (Reg - X86::XMM0);
784 if (Reg >= X86::YMM0 && Reg <= X86::YMM31)
785 return X86::ZMM0 + (Reg - X86::YMM0);
786 if (Reg >= X86::ZMM0 && Reg <= X86::ZMM31)
787 return Reg;
788 llvm_unreachable("Unexpected SIMD register");
789 }
790
791 }
792