1 //===-- X86FrameLowering.cpp - X86 Frame Information ----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the X86 implementation of TargetFrameLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "X86FrameLowering.h"
15 #include "X86InstrBuilder.h"
16 #include "X86InstrInfo.h"
17 #include "X86MachineFunctionInfo.h"
18 #include "X86Subtarget.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/ADT/SmallSet.h"
21 #include "llvm/Analysis/LibCallSemantics.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineModuleInfo.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/WinEHFuncInfo.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/MC/MCAsmInfo.h"
31 #include "llvm/MC/MCSymbol.h"
32 #include "llvm/Target/TargetOptions.h"
33 #include "llvm/Support/Debug.h"
38 X86FrameLowering::X86FrameLowering(const X86Subtarget &STI,
39 unsigned StackAlignOverride)
40 : TargetFrameLowering(StackGrowsDown, StackAlignOverride,
41 STI.is64Bit() ? -8 : -4),
42 STI(STI), TII(*STI.getInstrInfo()), TRI(STI.getRegisterInfo()) {
43 // Cache a bunch of frame-related predicates for this subtarget.
44 SlotSize = TRI->getSlotSize();
45 Is64Bit = STI.is64Bit();
46 IsLP64 = STI.isTarget64BitLP64();
47 // standard x86_64 and NaCl use 64-bit frame/stack pointers, x32 - 32-bit.
48 Uses64BitFramePtr = STI.isTarget64BitLP64() || STI.isTargetNaCl64();
49 StackPtr = TRI->getStackRegister();
52 bool X86FrameLowering::hasReservedCallFrame(const MachineFunction &MF) const {
53 return !MF.getFrameInfo()->hasVarSizedObjects() &&
54 !MF.getInfo<X86MachineFunctionInfo>()->getHasPushSequences();
57 /// canSimplifyCallFramePseudos - If there is a reserved call frame, the
58 /// call frame pseudos can be simplified. Having a FP, as in the default
59 /// implementation, is not sufficient here since we can't always use it.
60 /// Use a more nuanced condition.
62 X86FrameLowering::canSimplifyCallFramePseudos(const MachineFunction &MF) const {
63 return hasReservedCallFrame(MF) ||
64 (hasFP(MF) && !TRI->needsStackRealignment(MF)) ||
65 TRI->hasBasePointer(MF);
68 // needsFrameIndexResolution - Do we need to perform FI resolution for
69 // this function. Normally, this is required only when the function
70 // has any stack objects. However, FI resolution actually has another job,
71 // not apparent from the title - it resolves callframesetup/destroy
72 // that were not simplified earlier.
73 // So, this is required for x86 functions that have push sequences even
74 // when there are no stack objects.
76 X86FrameLowering::needsFrameIndexResolution(const MachineFunction &MF) const {
77 return MF.getFrameInfo()->hasStackObjects() ||
78 MF.getInfo<X86MachineFunctionInfo>()->getHasPushSequences();
81 /// hasFP - Return true if the specified function should have a dedicated frame
82 /// pointer register. This is true if the function has variable sized allocas
83 /// or if frame pointer elimination is disabled.
84 bool X86FrameLowering::hasFP(const MachineFunction &MF) const {
85 const MachineFrameInfo *MFI = MF.getFrameInfo();
86 const MachineModuleInfo &MMI = MF.getMMI();
88 return (MF.getTarget().Options.DisableFramePointerElim(MF) ||
89 TRI->needsStackRealignment(MF) ||
90 MFI->hasVarSizedObjects() ||
91 MFI->isFrameAddressTaken() || MFI->hasOpaqueSPAdjustment() ||
92 MF.getInfo<X86MachineFunctionInfo>()->getForceFramePointer() ||
93 MMI.callsUnwindInit() || MMI.hasEHFunclets() || MMI.callsEHReturn() ||
94 MFI->hasStackMap() || MFI->hasPatchPoint());
97 static unsigned getSUBriOpcode(unsigned IsLP64, int64_t Imm) {
100 return X86::SUB64ri8;
101 return X86::SUB64ri32;
104 return X86::SUB32ri8;
109 static unsigned getADDriOpcode(unsigned IsLP64, int64_t Imm) {
112 return X86::ADD64ri8;
113 return X86::ADD64ri32;
116 return X86::ADD32ri8;
121 static unsigned getSUBrrOpcode(unsigned isLP64) {
122 return isLP64 ? X86::SUB64rr : X86::SUB32rr;
125 static unsigned getADDrrOpcode(unsigned isLP64) {
126 return isLP64 ? X86::ADD64rr : X86::ADD32rr;
129 static unsigned getANDriOpcode(bool IsLP64, int64_t Imm) {
132 return X86::AND64ri8;
133 return X86::AND64ri32;
136 return X86::AND32ri8;
140 static unsigned getLEArOpcode(unsigned IsLP64) {
141 return IsLP64 ? X86::LEA64r : X86::LEA32r;
144 /// findDeadCallerSavedReg - Return a caller-saved register that isn't live
145 /// when it reaches the "return" instruction. We can then pop a stack object
146 /// to this register without worry about clobbering it.
147 static unsigned findDeadCallerSavedReg(MachineBasicBlock &MBB,
148 MachineBasicBlock::iterator &MBBI,
149 const TargetRegisterInfo *TRI,
151 const MachineFunction *MF = MBB.getParent();
152 const Function *F = MF->getFunction();
153 if (!F || MF->getMMI().callsEHReturn())
156 static const uint16_t CallerSavedRegs32Bit[] = {
157 X86::EAX, X86::EDX, X86::ECX, 0
160 static const uint16_t CallerSavedRegs64Bit[] = {
161 X86::RAX, X86::RDX, X86::RCX, X86::RSI, X86::RDI,
162 X86::R8, X86::R9, X86::R10, X86::R11, 0
165 unsigned Opc = MBBI->getOpcode();
172 case X86::TCRETURNdi:
173 case X86::TCRETURNri:
174 case X86::TCRETURNmi:
175 case X86::TCRETURNdi64:
176 case X86::TCRETURNri64:
177 case X86::TCRETURNmi64:
179 case X86::EH_RETURN64: {
180 SmallSet<uint16_t, 8> Uses;
181 for (unsigned i = 0, e = MBBI->getNumOperands(); i != e; ++i) {
182 MachineOperand &MO = MBBI->getOperand(i);
183 if (!MO.isReg() || MO.isDef())
185 unsigned Reg = MO.getReg();
188 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
192 const uint16_t *CS = Is64Bit ? CallerSavedRegs64Bit : CallerSavedRegs32Bit;
194 if (!Uses.count(*CS))
202 static bool isEAXLiveIn(MachineFunction &MF) {
203 for (MachineRegisterInfo::livein_iterator II = MF.getRegInfo().livein_begin(),
204 EE = MF.getRegInfo().livein_end(); II != EE; ++II) {
205 unsigned Reg = II->first;
207 if (Reg == X86::RAX || Reg == X86::EAX || Reg == X86::AX ||
208 Reg == X86::AH || Reg == X86::AL)
215 /// Check whether or not the terminators of \p MBB needs to read EFLAGS.
216 static bool terminatorsNeedFlagsAsInput(const MachineBasicBlock &MBB) {
217 for (const MachineInstr &MI : MBB.terminators()) {
218 bool BreakNext = false;
219 for (const MachineOperand &MO : MI.operands()) {
222 unsigned Reg = MO.getReg();
223 if (Reg != X86::EFLAGS)
226 // This terminator needs an eflag that is not defined
227 // by a previous terminator.
238 /// emitSPUpdate - Emit a series of instructions to increment / decrement the
239 /// stack pointer by a constant value.
240 void X86FrameLowering::emitSPUpdate(MachineBasicBlock &MBB,
241 MachineBasicBlock::iterator &MBBI,
242 int64_t NumBytes, bool InEpilogue) const {
243 bool isSub = NumBytes < 0;
244 uint64_t Offset = isSub ? -NumBytes : NumBytes;
246 uint64_t Chunk = (1LL << 31) - 1;
247 DebugLoc DL = MBB.findDebugLoc(MBBI);
250 if (Offset > Chunk) {
251 // Rather than emit a long series of instructions for large offsets,
252 // load the offset into a register and do one sub/add
255 if (isSub && !isEAXLiveIn(*MBB.getParent()))
256 Reg = (unsigned)(Is64Bit ? X86::RAX : X86::EAX);
258 Reg = findDeadCallerSavedReg(MBB, MBBI, TRI, Is64Bit);
261 unsigned Opc = Is64Bit ? X86::MOV64ri : X86::MOV32ri;
262 BuildMI(MBB, MBBI, DL, TII.get(Opc), Reg)
265 ? getSUBrrOpcode(Is64Bit)
266 : getADDrrOpcode(Is64Bit);
267 MachineInstr *MI = BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr)
270 MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
276 uint64_t ThisVal = std::min(Offset, Chunk);
277 if (ThisVal == (Is64Bit ? 8 : 4)) {
278 // Use push / pop instead.
280 ? (unsigned)(Is64Bit ? X86::RAX : X86::EAX)
281 : findDeadCallerSavedReg(MBB, MBBI, TRI, Is64Bit);
284 ? (Is64Bit ? X86::PUSH64r : X86::PUSH32r)
285 : (Is64Bit ? X86::POP64r : X86::POP32r);
286 MachineInstr *MI = BuildMI(MBB, MBBI, DL, TII.get(Opc))
287 .addReg(Reg, getDefRegState(!isSub) | getUndefRegState(isSub));
289 MI->setFlag(MachineInstr::FrameSetup);
291 MI->setFlag(MachineInstr::FrameDestroy);
297 MachineInstrBuilder MI = BuildStackAdjustment(
298 MBB, MBBI, DL, isSub ? -ThisVal : ThisVal, InEpilogue);
300 MI.setMIFlag(MachineInstr::FrameSetup);
302 MI.setMIFlag(MachineInstr::FrameDestroy);
308 MachineInstrBuilder X86FrameLowering::BuildStackAdjustment(
309 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, DebugLoc DL,
310 int64_t Offset, bool InEpilogue) const {
311 assert(Offset != 0 && "zero offset stack adjustment requested");
313 // On Atom, using LEA to adjust SP is preferred, but using it in the epilogue
317 UseLEA = STI.useLeaForSP();
319 // If we can use LEA for SP but we shouldn't, check that none
320 // of the terminators uses the eflags. Otherwise we will insert
321 // a ADD that will redefine the eflags and break the condition.
322 // Alternatively, we could move the ADD, but this may not be possible
323 // and is an optimization anyway.
324 UseLEA = canUseLEAForSPInEpilogue(*MBB.getParent());
325 if (UseLEA && !STI.useLeaForSP())
326 UseLEA = terminatorsNeedFlagsAsInput(MBB);
327 // If that assert breaks, that means we do not do the right thing
328 // in canUseAsEpilogue.
329 assert((UseLEA || !terminatorsNeedFlagsAsInput(MBB)) &&
330 "We shouldn't have allowed this insertion point");
333 MachineInstrBuilder MI;
335 MI = addRegOffset(BuildMI(MBB, MBBI, DL,
336 TII.get(getLEArOpcode(Uses64BitFramePtr)),
338 StackPtr, false, Offset);
340 bool IsSub = Offset < 0;
341 uint64_t AbsOffset = IsSub ? -Offset : Offset;
342 unsigned Opc = IsSub ? getSUBriOpcode(Uses64BitFramePtr, AbsOffset)
343 : getADDriOpcode(Uses64BitFramePtr, AbsOffset);
344 MI = BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr)
347 MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
352 int X86FrameLowering::mergeSPUpdates(MachineBasicBlock &MBB,
353 MachineBasicBlock::iterator &MBBI,
354 bool doMergeWithPrevious) const {
355 if ((doMergeWithPrevious && MBBI == MBB.begin()) ||
356 (!doMergeWithPrevious && MBBI == MBB.end()))
359 MachineBasicBlock::iterator PI = doMergeWithPrevious ? std::prev(MBBI) : MBBI;
360 MachineBasicBlock::iterator NI = doMergeWithPrevious ? nullptr
362 unsigned Opc = PI->getOpcode();
365 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
366 Opc == X86::ADD32ri || Opc == X86::ADD32ri8 ||
367 Opc == X86::LEA32r || Opc == X86::LEA64_32r) &&
368 PI->getOperand(0).getReg() == StackPtr){
369 Offset += PI->getOperand(2).getImm();
371 if (!doMergeWithPrevious) MBBI = NI;
372 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
373 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
374 PI->getOperand(0).getReg() == StackPtr) {
375 Offset -= PI->getOperand(2).getImm();
377 if (!doMergeWithPrevious) MBBI = NI;
383 void X86FrameLowering::BuildCFI(MachineBasicBlock &MBB,
384 MachineBasicBlock::iterator MBBI, DebugLoc DL,
385 MCCFIInstruction CFIInst) const {
386 MachineFunction &MF = *MBB.getParent();
387 unsigned CFIIndex = MF.getMMI().addFrameInst(CFIInst);
388 BuildMI(MBB, MBBI, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
389 .addCFIIndex(CFIIndex);
393 X86FrameLowering::emitCalleeSavedFrameMoves(MachineBasicBlock &MBB,
394 MachineBasicBlock::iterator MBBI,
396 MachineFunction &MF = *MBB.getParent();
397 MachineFrameInfo *MFI = MF.getFrameInfo();
398 MachineModuleInfo &MMI = MF.getMMI();
399 const MCRegisterInfo *MRI = MMI.getContext().getRegisterInfo();
401 // Add callee saved registers to move list.
402 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
403 if (CSI.empty()) return;
405 // Calculate offsets.
406 for (std::vector<CalleeSavedInfo>::const_iterator
407 I = CSI.begin(), E = CSI.end(); I != E; ++I) {
408 int64_t Offset = MFI->getObjectOffset(I->getFrameIdx());
409 unsigned Reg = I->getReg();
411 unsigned DwarfReg = MRI->getDwarfRegNum(Reg, true);
412 BuildCFI(MBB, MBBI, DL,
413 MCCFIInstruction::createOffset(nullptr, DwarfReg, Offset));
417 /// usesTheStack - This function checks if any of the users of EFLAGS
418 /// copies the EFLAGS. We know that the code that lowers COPY of EFLAGS has
419 /// to use the stack, and if we don't adjust the stack we clobber the first
421 /// See X86InstrInfo::copyPhysReg.
422 static bool usesTheStack(const MachineFunction &MF) {
423 const MachineRegisterInfo &MRI = MF.getRegInfo();
425 for (MachineRegisterInfo::reg_instr_iterator
426 ri = MRI.reg_instr_begin(X86::EFLAGS), re = MRI.reg_instr_end();
434 MachineInstr *X86FrameLowering::emitStackProbe(MachineFunction &MF,
435 MachineBasicBlock &MBB,
436 MachineBasicBlock::iterator MBBI,
438 bool InProlog) const {
439 const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
440 if (STI.isTargetWindowsCoreCLR()) {
442 return emitStackProbeInlineStub(MF, MBB, MBBI, DL, true);
444 return emitStackProbeInline(MF, MBB, MBBI, DL, false);
447 return emitStackProbeCall(MF, MBB, MBBI, DL, InProlog);
451 void X86FrameLowering::inlineStackProbe(MachineFunction &MF,
452 MachineBasicBlock &PrologMBB) const {
453 const StringRef ChkStkStubSymbol = "__chkstk_stub";
454 MachineInstr *ChkStkStub = nullptr;
456 for (MachineInstr &MI : PrologMBB) {
457 if (MI.isCall() && MI.getOperand(0).isSymbol() &&
458 ChkStkStubSymbol == MI.getOperand(0).getSymbolName()) {
464 if (ChkStkStub != nullptr) {
465 MachineBasicBlock::iterator MBBI = std::next(ChkStkStub->getIterator());
466 assert(std::prev(MBBI).operator==(ChkStkStub) &&
467 "MBBI expected after __chkstk_stub.");
468 DebugLoc DL = PrologMBB.findDebugLoc(MBBI);
469 emitStackProbeInline(MF, PrologMBB, MBBI, DL, true);
470 ChkStkStub->eraseFromParent();
474 MachineInstr *X86FrameLowering::emitStackProbeInline(
475 MachineFunction &MF, MachineBasicBlock &MBB,
476 MachineBasicBlock::iterator MBBI, DebugLoc DL, bool InProlog) const {
477 const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
478 assert(STI.is64Bit() && "different expansion needed for 32 bit");
479 assert(STI.isTargetWindowsCoreCLR() && "custom expansion expects CoreCLR");
480 const TargetInstrInfo &TII = *STI.getInstrInfo();
481 const BasicBlock *LLVM_BB = MBB.getBasicBlock();
483 // RAX contains the number of bytes of desired stack adjustment.
484 // The handling here assumes this value has already been updated so as to
485 // maintain stack alignment.
487 // We need to exit with RSP modified by this amount and execute suitable
488 // page touches to notify the OS that we're growing the stack responsibly.
489 // All stack probing must be done without modifying RSP.
495 // Flags, TestReg = CopyReg - SizeReg
496 // FinalReg = !Flags.Ovf ? TestReg : ZeroReg
497 // LimitReg = gs magic thread env access
498 // if FinalReg >= LimitReg goto ContinueMBB
500 // RoundReg = page address of FinalReg
502 // LoopReg = PHI(LimitReg,ProbeReg)
503 // ProbeReg = LoopReg - PageSize
505 // if (ProbeReg > RoundReg) goto LoopMBB
508 // [rest of original MBB]
510 // Set up the new basic blocks
511 MachineBasicBlock *RoundMBB = MF.CreateMachineBasicBlock(LLVM_BB);
512 MachineBasicBlock *LoopMBB = MF.CreateMachineBasicBlock(LLVM_BB);
513 MachineBasicBlock *ContinueMBB = MF.CreateMachineBasicBlock(LLVM_BB);
515 MachineFunction::iterator MBBIter = std::next(MBB.getIterator());
516 MF.insert(MBBIter, RoundMBB);
517 MF.insert(MBBIter, LoopMBB);
518 MF.insert(MBBIter, ContinueMBB);
520 // Split MBB and move the tail portion down to ContinueMBB.
521 MachineBasicBlock::iterator BeforeMBBI = std::prev(MBBI);
522 ContinueMBB->splice(ContinueMBB->begin(), &MBB, MBBI, MBB.end());
523 ContinueMBB->transferSuccessorsAndUpdatePHIs(&MBB);
525 // Some useful constants
526 const int64_t ThreadEnvironmentStackLimit = 0x10;
527 const int64_t PageSize = 0x1000;
528 const int64_t PageMask = ~(PageSize - 1);
530 // Registers we need. For the normal case we use virtual
531 // registers. For the prolog expansion we use RAX, RCX and RDX.
532 MachineRegisterInfo &MRI = MF.getRegInfo();
533 const TargetRegisterClass *RegClass = &X86::GR64RegClass;
534 const unsigned SizeReg = InProlog ? (unsigned)X86::RAX
535 : MRI.createVirtualRegister(RegClass),
536 ZeroReg = InProlog ? (unsigned)X86::RCX
537 : MRI.createVirtualRegister(RegClass),
538 CopyReg = InProlog ? (unsigned)X86::RDX
539 : MRI.createVirtualRegister(RegClass),
540 TestReg = InProlog ? (unsigned)X86::RDX
541 : MRI.createVirtualRegister(RegClass),
542 FinalReg = InProlog ? (unsigned)X86::RDX
543 : MRI.createVirtualRegister(RegClass),
544 RoundedReg = InProlog ? (unsigned)X86::RDX
545 : MRI.createVirtualRegister(RegClass),
546 LimitReg = InProlog ? (unsigned)X86::RCX
547 : MRI.createVirtualRegister(RegClass),
548 JoinReg = InProlog ? (unsigned)X86::RCX
549 : MRI.createVirtualRegister(RegClass),
550 ProbeReg = InProlog ? (unsigned)X86::RCX
551 : MRI.createVirtualRegister(RegClass);
553 // SP-relative offsets where we can save RCX and RDX.
554 int64_t RCXShadowSlot = 0;
555 int64_t RDXShadowSlot = 0;
557 // If inlining in the prolog, save RCX and RDX.
558 // Future optimization: don't save or restore if not live in.
560 // Compute the offsets. We need to account for things already
561 // pushed onto the stack at this point: return address, frame
562 // pointer (if used), and callee saves.
563 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
564 const int64_t CalleeSaveSize = X86FI->getCalleeSavedFrameSize();
565 const bool HasFP = hasFP(MF);
566 RCXShadowSlot = 8 + CalleeSaveSize + (HasFP ? 8 : 0);
567 RDXShadowSlot = RCXShadowSlot + 8;
569 addRegOffset(BuildMI(&MBB, DL, TII.get(X86::MOV64mr)), X86::RSP, false,
572 addRegOffset(BuildMI(&MBB, DL, TII.get(X86::MOV64mr)), X86::RSP, false,
576 // Not in the prolog. Copy RAX to a virtual reg.
577 BuildMI(&MBB, DL, TII.get(X86::MOV64rr), SizeReg).addReg(X86::RAX);
580 // Add code to MBB to check for overflow and set the new target stack pointer
582 BuildMI(&MBB, DL, TII.get(X86::XOR64rr), ZeroReg)
583 .addReg(ZeroReg, RegState::Undef)
584 .addReg(ZeroReg, RegState::Undef);
585 BuildMI(&MBB, DL, TII.get(X86::MOV64rr), CopyReg).addReg(X86::RSP);
586 BuildMI(&MBB, DL, TII.get(X86::SUB64rr), TestReg)
589 BuildMI(&MBB, DL, TII.get(X86::CMOVB64rr), FinalReg)
593 // FinalReg now holds final stack pointer value, or zero if
594 // allocation would overflow. Compare against the current stack
595 // limit from the thread environment block. Note this limit is the
596 // lowest touched page on the stack, not the point at which the OS
597 // will cause an overflow exception, so this is just an optimization
598 // to avoid unnecessarily touching pages that are below the current
599 // SP but already commited to the stack by the OS.
600 BuildMI(&MBB, DL, TII.get(X86::MOV64rm), LimitReg)
604 .addImm(ThreadEnvironmentStackLimit)
606 BuildMI(&MBB, DL, TII.get(X86::CMP64rr)).addReg(FinalReg).addReg(LimitReg);
607 // Jump if the desired stack pointer is at or above the stack limit.
608 BuildMI(&MBB, DL, TII.get(X86::JAE_1)).addMBB(ContinueMBB);
610 // Add code to roundMBB to round the final stack pointer to a page boundary.
611 BuildMI(RoundMBB, DL, TII.get(X86::AND64ri32), RoundedReg)
614 BuildMI(RoundMBB, DL, TII.get(X86::JMP_1)).addMBB(LoopMBB);
616 // LimitReg now holds the current stack limit, RoundedReg page-rounded
617 // final RSP value. Add code to loopMBB to decrement LimitReg page-by-page
618 // and probe until we reach RoundedReg.
620 BuildMI(LoopMBB, DL, TII.get(X86::PHI), JoinReg)
627 addRegOffset(BuildMI(LoopMBB, DL, TII.get(X86::LEA64r), ProbeReg), JoinReg,
630 // Probe by storing a byte onto the stack.
631 BuildMI(LoopMBB, DL, TII.get(X86::MOV8mi))
638 BuildMI(LoopMBB, DL, TII.get(X86::CMP64rr))
641 BuildMI(LoopMBB, DL, TII.get(X86::JNE_1)).addMBB(LoopMBB);
643 MachineBasicBlock::iterator ContinueMBBI = ContinueMBB->getFirstNonPHI();
645 // If in prolog, restore RDX and RCX.
647 addRegOffset(BuildMI(*ContinueMBB, ContinueMBBI, DL, TII.get(X86::MOV64rm),
649 X86::RSP, false, RCXShadowSlot);
650 addRegOffset(BuildMI(*ContinueMBB, ContinueMBBI, DL, TII.get(X86::MOV64rm),
652 X86::RSP, false, RDXShadowSlot);
655 // Now that the probing is done, add code to continueMBB to update
656 // the stack pointer for real.
657 BuildMI(*ContinueMBB, ContinueMBBI, DL, TII.get(X86::SUB64rr), X86::RSP)
661 // Add the control flow edges we need.
662 MBB.addSuccessor(ContinueMBB);
663 MBB.addSuccessor(RoundMBB);
664 RoundMBB->addSuccessor(LoopMBB);
665 LoopMBB->addSuccessor(ContinueMBB);
666 LoopMBB->addSuccessor(LoopMBB);
668 // Mark all the instructions added to the prolog as frame setup.
670 for (++BeforeMBBI; BeforeMBBI != MBB.end(); ++BeforeMBBI) {
671 BeforeMBBI->setFlag(MachineInstr::FrameSetup);
673 for (MachineInstr &MI : *RoundMBB) {
674 MI.setFlag(MachineInstr::FrameSetup);
676 for (MachineInstr &MI : *LoopMBB) {
677 MI.setFlag(MachineInstr::FrameSetup);
679 for (MachineBasicBlock::iterator CMBBI = ContinueMBB->begin();
680 CMBBI != ContinueMBBI; ++CMBBI) {
681 CMBBI->setFlag(MachineInstr::FrameSetup);
685 // Possible TODO: physreg liveness for InProlog case.
690 MachineInstr *X86FrameLowering::emitStackProbeCall(
691 MachineFunction &MF, MachineBasicBlock &MBB,
692 MachineBasicBlock::iterator MBBI, DebugLoc DL, bool InProlog) const {
693 bool IsLargeCodeModel = MF.getTarget().getCodeModel() == CodeModel::Large;
697 CallOp = IsLargeCodeModel ? X86::CALL64r : X86::CALL64pcrel32;
699 CallOp = X86::CALLpcrel32;
703 if (STI.isTargetCygMing()) {
704 Symbol = "___chkstk_ms";
708 } else if (STI.isTargetCygMing())
713 MachineInstrBuilder CI;
714 MachineBasicBlock::iterator ExpansionMBBI = std::prev(MBBI);
716 // All current stack probes take AX and SP as input, clobber flags, and
717 // preserve all registers. x86_64 probes leave RSP unmodified.
718 if (Is64Bit && MF.getTarget().getCodeModel() == CodeModel::Large) {
719 // For the large code model, we have to call through a register. Use R11,
720 // as it is scratch in all supported calling conventions.
721 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64ri), X86::R11)
722 .addExternalSymbol(Symbol);
723 CI = BuildMI(MBB, MBBI, DL, TII.get(CallOp)).addReg(X86::R11);
725 CI = BuildMI(MBB, MBBI, DL, TII.get(CallOp)).addExternalSymbol(Symbol);
728 unsigned AX = Is64Bit ? X86::RAX : X86::EAX;
729 unsigned SP = Is64Bit ? X86::RSP : X86::ESP;
730 CI.addReg(AX, RegState::Implicit)
731 .addReg(SP, RegState::Implicit)
732 .addReg(AX, RegState::Define | RegState::Implicit)
733 .addReg(SP, RegState::Define | RegState::Implicit)
734 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
737 // MSVC x64's __chkstk and cygwin/mingw's ___chkstk_ms do not adjust %rsp
738 // themselves. It also does not clobber %rax so we can reuse it when
740 BuildMI(MBB, MBBI, DL, TII.get(X86::SUB64rr), X86::RSP)
746 // Apply the frame setup flag to all inserted instrs.
747 for (++ExpansionMBBI; ExpansionMBBI != MBBI; ++ExpansionMBBI)
748 ExpansionMBBI->setFlag(MachineInstr::FrameSetup);
754 MachineInstr *X86FrameLowering::emitStackProbeInlineStub(
755 MachineFunction &MF, MachineBasicBlock &MBB,
756 MachineBasicBlock::iterator MBBI, DebugLoc DL, bool InProlog) const {
758 assert(InProlog && "ChkStkStub called outside prolog!");
760 BuildMI(MBB, MBBI, DL, TII.get(X86::CALLpcrel32))
761 .addExternalSymbol("__chkstk_stub");
766 static unsigned calculateSetFPREG(uint64_t SPAdjust) {
767 // Win64 ABI has a less restrictive limitation of 240; 128 works equally well
768 // and might require smaller successive adjustments.
769 const uint64_t Win64MaxSEHOffset = 128;
770 uint64_t SEHFrameOffset = std::min(SPAdjust, Win64MaxSEHOffset);
771 // Win64 ABI requires 16-byte alignment for the UWOP_SET_FPREG opcode.
772 return SEHFrameOffset & -16;
775 // If we're forcing a stack realignment we can't rely on just the frame
776 // info, we need to know the ABI stack alignment as well in case we
777 // have a call out. Otherwise just make sure we have some alignment - we'll
778 // go with the minimum SlotSize.
779 uint64_t X86FrameLowering::calculateMaxStackAlign(const MachineFunction &MF) const {
780 const MachineFrameInfo *MFI = MF.getFrameInfo();
781 uint64_t MaxAlign = MFI->getMaxAlignment(); // Desired stack alignment.
782 unsigned StackAlign = getStackAlignment();
783 if (MF.getFunction()->hasFnAttribute("stackrealign")) {
785 MaxAlign = (StackAlign > MaxAlign) ? StackAlign : MaxAlign;
786 else if (MaxAlign < SlotSize)
792 void X86FrameLowering::BuildStackAlignAND(MachineBasicBlock &MBB,
793 MachineBasicBlock::iterator MBBI,
794 DebugLoc DL, unsigned Reg,
795 uint64_t MaxAlign) const {
796 uint64_t Val = -MaxAlign;
797 unsigned AndOp = getANDriOpcode(Uses64BitFramePtr, Val);
798 MachineInstr *MI = BuildMI(MBB, MBBI, DL, TII.get(AndOp), Reg)
801 .setMIFlag(MachineInstr::FrameSetup);
803 // The EFLAGS implicit def is dead.
804 MI->getOperand(3).setIsDead();
807 /// emitPrologue - Push callee-saved registers onto the stack, which
808 /// automatically adjust the stack pointer. Adjust the stack pointer to allocate
809 /// space for local variables. Also emit labels used by the exception handler to
810 /// generate the exception handling frames.
813 Here's a gist of what gets emitted:
815 ; Establish frame pointer, if needed
818 .cfi_def_cfa_offset 16
819 .cfi_offset %rbp, -16
822 .cfi_def_cfa_register %rbp
824 ; Spill general-purpose registers
825 [for all callee-saved GPRs]
828 .cfi_def_cfa_offset (offset from RETADDR)
831 ; If the required stack alignment > default stack alignment
832 ; rsp needs to be re-aligned. This creates a "re-alignment gap"
833 ; of unknown size in the stack frame.
834 [if stack needs re-alignment]
837 ; Allocate space for locals
838 [if target is Windows and allocated space > 4096 bytes]
839 ; Windows needs special care for allocations larger
842 call ___chkstk_ms/___chkstk
848 .seh_stackalloc (size of XMM spill slots)
849 .seh_setframe %rbp, SEHFrameOffset ; = size of all spill slots
854 ; Note, that while only Windows 64 ABI specifies XMMs as callee-preserved,
855 ; they may get spilled on any platform, if the current function
856 ; calls @llvm.eh.unwind.init
858 [for all callee-saved XMM registers]
859 movaps %<xmm reg>, -MMM(%rbp)
860 [for all callee-saved XMM registers]
861 .seh_savexmm %<xmm reg>, (-MMM + SEHFrameOffset)
862 ; i.e. the offset relative to (%rbp - SEHFrameOffset)
864 [for all callee-saved XMM registers]
865 movaps %<xmm reg>, KKK(%rsp)
866 [for all callee-saved XMM registers]
867 .seh_savexmm %<xmm reg>, KKK
871 [if needs base pointer]
873 [if needs to restore base pointer]
878 [for all callee-saved registers]
879 .cfi_offset %<reg>, (offset from %rbp)
881 .cfi_def_cfa_offset (offset from RETADDR)
882 [for all callee-saved registers]
883 .cfi_offset %<reg>, (offset from %rsp)
886 - .seh directives are emitted only for Windows 64 ABI
887 - .cfi directives are emitted for all other ABIs
888 - for 32-bit code, substitute %e?? registers for %r??
891 void X86FrameLowering::emitPrologue(MachineFunction &MF,
892 MachineBasicBlock &MBB) const {
893 assert(&STI == &MF.getSubtarget<X86Subtarget>() &&
894 "MF used frame lowering for wrong subtarget");
895 MachineBasicBlock::iterator MBBI = MBB.begin();
896 MachineFrameInfo *MFI = MF.getFrameInfo();
897 const Function *Fn = MF.getFunction();
898 MachineModuleInfo &MMI = MF.getMMI();
899 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
900 uint64_t MaxAlign = calculateMaxStackAlign(MF); // Desired stack alignment.
901 uint64_t StackSize = MFI->getStackSize(); // Number of bytes to allocate.
902 bool IsFunclet = MBB.isEHFuncletEntry();
903 bool FnHasClrFunclet =
904 MMI.hasEHFunclets() &&
905 classifyEHPersonality(Fn->getPersonalityFn()) == EHPersonality::CoreCLR;
906 bool IsClrFunclet = IsFunclet && FnHasClrFunclet;
907 bool HasFP = hasFP(MF);
908 bool IsWin64CC = STI.isCallingConvWin64(Fn->getCallingConv());
909 bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
910 bool NeedsWinCFI = IsWin64Prologue && Fn->needsUnwindTableEntry();
912 !IsWin64Prologue && (MMI.hasDebugInfo() || Fn->needsUnwindTableEntry());
913 unsigned FramePtr = TRI->getFrameRegister(MF);
914 const unsigned MachineFramePtr =
915 STI.isTarget64BitILP32()
916 ? getX86SubSuperRegister(FramePtr, MVT::i64, false)
918 unsigned BasePtr = TRI->getBaseRegister();
920 // Debug location must be unknown since the first debug location is used
921 // to determine the end of the prologue.
924 // Add RETADDR move area to callee saved frame size.
925 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
926 if (TailCallReturnAddrDelta && IsWin64Prologue)
927 report_fatal_error("Can't handle guaranteed tail call under win64 yet");
929 if (TailCallReturnAddrDelta < 0)
930 X86FI->setCalleeSavedFrameSize(
931 X86FI->getCalleeSavedFrameSize() - TailCallReturnAddrDelta);
933 bool UseStackProbe = (STI.isOSWindows() && !STI.isTargetMachO());
935 // The default stack probe size is 4096 if the function has no stackprobesize
937 unsigned StackProbeSize = 4096;
938 if (Fn->hasFnAttribute("stack-probe-size"))
939 Fn->getFnAttribute("stack-probe-size")
941 .getAsInteger(0, StackProbeSize);
943 // If this is x86-64 and the Red Zone is not disabled, if we are a leaf
944 // function, and use up to 128 bytes of stack space, don't have a frame
945 // pointer, calls, or dynamic alloca then we do not need to adjust the
946 // stack pointer (we fit in the Red Zone). We also check that we don't
947 // push and pop from the stack.
948 if (Is64Bit && !Fn->hasFnAttribute(Attribute::NoRedZone) &&
949 !TRI->needsStackRealignment(MF) &&
950 !MFI->hasVarSizedObjects() && // No dynamic alloca.
951 !MFI->adjustsStack() && // No calls.
952 !IsWin64CC && // Win64 has no Red Zone
953 !usesTheStack(MF) && // Don't push and pop.
954 !MF.shouldSplitStack()) { // Regular stack
955 uint64_t MinSize = X86FI->getCalleeSavedFrameSize();
956 if (HasFP) MinSize += SlotSize;
957 StackSize = std::max(MinSize, StackSize > 128 ? StackSize - 128 : 0);
958 MFI->setStackSize(StackSize);
961 // Insert stack pointer adjustment for later moving of return addr. Only
962 // applies to tail call optimized functions where the callee argument stack
963 // size is bigger than the callers.
964 if (TailCallReturnAddrDelta < 0) {
965 BuildStackAdjustment(MBB, MBBI, DL, TailCallReturnAddrDelta,
966 /*InEpilogue=*/false)
967 .setMIFlag(MachineInstr::FrameSetup);
970 // Mapping for machine moves:
972 // DST: VirtualFP AND
973 // SRC: VirtualFP => DW_CFA_def_cfa_offset
974 // ELSE => DW_CFA_def_cfa
976 // SRC: VirtualFP AND
977 // DST: Register => DW_CFA_def_cfa_register
980 // OFFSET < 0 => DW_CFA_offset_extended_sf
981 // REG < 64 => DW_CFA_offset + Reg
982 // ELSE => DW_CFA_offset_extended
984 uint64_t NumBytes = 0;
985 int stackGrowth = -SlotSize;
987 // Find the funclet establisher parameter
988 unsigned Establisher = X86::NoRegister;
990 Establisher = Uses64BitFramePtr ? X86::RCX : X86::ECX;
992 Establisher = Uses64BitFramePtr ? X86::RDX : X86::EDX;
994 if (IsWin64Prologue && IsFunclet & !IsClrFunclet) {
995 // Immediately spill establisher into the home slot.
996 // The runtime cares about this.
997 // MOV64mr %rdx, 16(%rsp)
998 unsigned MOVmr = Uses64BitFramePtr ? X86::MOV64mr : X86::MOV32mr;
999 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(MOVmr)), StackPtr, true, 16)
1000 .addReg(Establisher)
1001 .setMIFlag(MachineInstr::FrameSetup);
1002 MBB.addLiveIn(Establisher);
1006 // Calculate required stack adjustment.
1007 uint64_t FrameSize = StackSize - SlotSize;
1008 // If required, include space for extra hidden slot for stashing base pointer.
1009 if (X86FI->getRestoreBasePointer())
1010 FrameSize += SlotSize;
1012 NumBytes = FrameSize - X86FI->getCalleeSavedFrameSize();
1014 // Callee-saved registers are pushed on stack before the stack is realigned.
1015 if (TRI->needsStackRealignment(MF) && !IsWin64Prologue)
1016 NumBytes = RoundUpToAlignment(NumBytes, MaxAlign);
1018 // Get the offset of the stack slot for the EBP register, which is
1019 // guaranteed to be the last slot by processFunctionBeforeFrameFinalized.
1020 // Update the frame offset adjustment.
1022 MFI->setOffsetAdjustment(-NumBytes);
1024 assert(MFI->getOffsetAdjustment() == -(int)NumBytes &&
1025 "should calculate same local variable offset for funclets");
1027 // Save EBP/RBP into the appropriate stack slot.
1028 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::PUSH64r : X86::PUSH32r))
1029 .addReg(MachineFramePtr, RegState::Kill)
1030 .setMIFlag(MachineInstr::FrameSetup);
1032 if (NeedsDwarfCFI) {
1033 // Mark the place where EBP/RBP was saved.
1034 // Define the current CFA rule to use the provided offset.
1036 BuildCFI(MBB, MBBI, DL,
1037 MCCFIInstruction::createDefCfaOffset(nullptr, 2 * stackGrowth));
1039 // Change the rule for the FramePtr to be an "offset" rule.
1040 unsigned DwarfFramePtr = TRI->getDwarfRegNum(MachineFramePtr, true);
1041 BuildCFI(MBB, MBBI, DL, MCCFIInstruction::createOffset(
1042 nullptr, DwarfFramePtr, 2 * stackGrowth));
1046 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_PushReg))
1048 .setMIFlag(MachineInstr::FrameSetup);
1051 if (!IsWin64Prologue && !IsFunclet) {
1052 // Update EBP with the new base value.
1053 BuildMI(MBB, MBBI, DL,
1054 TII.get(Uses64BitFramePtr ? X86::MOV64rr : X86::MOV32rr),
1057 .setMIFlag(MachineInstr::FrameSetup);
1059 if (NeedsDwarfCFI) {
1060 // Mark effective beginning of when frame pointer becomes valid.
1061 // Define the current CFA to use the EBP/RBP register.
1062 unsigned DwarfFramePtr = TRI->getDwarfRegNum(MachineFramePtr, true);
1063 BuildCFI(MBB, MBBI, DL, MCCFIInstruction::createDefCfaRegister(
1064 nullptr, DwarfFramePtr));
1068 // Mark the FramePtr as live-in in every block. Don't do this again for
1069 // funclet prologues.
1071 for (MachineBasicBlock &EveryMBB : MF)
1072 EveryMBB.addLiveIn(MachineFramePtr);
1075 assert(!IsFunclet && "funclets without FPs not yet implemented");
1076 NumBytes = StackSize - X86FI->getCalleeSavedFrameSize();
1079 // For EH funclets, only allocate enough space for outgoing calls. Save the
1080 // NumBytes value that we would've used for the parent frame.
1081 unsigned ParentFrameNumBytes = NumBytes;
1083 NumBytes = getWinEHFuncletFrameSize(MF);
1085 // Skip the callee-saved push instructions.
1086 bool PushedRegs = false;
1087 int StackOffset = 2 * stackGrowth;
1089 while (MBBI != MBB.end() &&
1090 MBBI->getFlag(MachineInstr::FrameSetup) &&
1091 (MBBI->getOpcode() == X86::PUSH32r ||
1092 MBBI->getOpcode() == X86::PUSH64r)) {
1094 unsigned Reg = MBBI->getOperand(0).getReg();
1097 if (!HasFP && NeedsDwarfCFI) {
1098 // Mark callee-saved push instruction.
1099 // Define the current CFA rule to use the provided offset.
1101 BuildCFI(MBB, MBBI, DL,
1102 MCCFIInstruction::createDefCfaOffset(nullptr, StackOffset));
1103 StackOffset += stackGrowth;
1107 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_PushReg)).addImm(Reg).setMIFlag(
1108 MachineInstr::FrameSetup);
1112 // Realign stack after we pushed callee-saved registers (so that we'll be
1113 // able to calculate their offsets from the frame pointer).
1114 // Don't do this for Win64, it needs to realign the stack after the prologue.
1115 if (!IsWin64Prologue && !IsFunclet && TRI->needsStackRealignment(MF)) {
1116 assert(HasFP && "There should be a frame pointer if stack is realigned.");
1117 BuildStackAlignAND(MBB, MBBI, DL, StackPtr, MaxAlign);
1120 // If there is an SUB32ri of ESP immediately before this instruction, merge
1121 // the two. This can be the case when tail call elimination is enabled and
1122 // the callee has more arguments then the caller.
1123 NumBytes -= mergeSPUpdates(MBB, MBBI, true);
1125 // Adjust stack pointer: ESP -= numbytes.
1127 // Windows and cygwin/mingw require a prologue helper routine when allocating
1128 // more than 4K bytes on the stack. Windows uses __chkstk and cygwin/mingw
1129 // uses __alloca. __alloca and the 32-bit version of __chkstk will probe the
1130 // stack and adjust the stack pointer in one go. The 64-bit version of
1131 // __chkstk is only responsible for probing the stack. The 64-bit prologue is
1132 // responsible for adjusting the stack pointer. Touching the stack at 4K
1133 // increments is necessary to ensure that the guard pages used by the OS
1134 // virtual memory manager are allocated in correct sequence.
1135 uint64_t AlignedNumBytes = NumBytes;
1136 if (IsWin64Prologue && !IsFunclet && TRI->needsStackRealignment(MF))
1137 AlignedNumBytes = RoundUpToAlignment(AlignedNumBytes, MaxAlign);
1138 if (AlignedNumBytes >= StackProbeSize && UseStackProbe) {
1139 // Check whether EAX is livein for this function.
1140 bool isEAXAlive = isEAXLiveIn(MF);
1143 // Sanity check that EAX is not livein for this function.
1144 // It should not be, so throw an assert.
1145 assert(!Is64Bit && "EAX is livein in x64 case!");
1148 BuildMI(MBB, MBBI, DL, TII.get(X86::PUSH32r))
1149 .addReg(X86::EAX, RegState::Kill)
1150 .setMIFlag(MachineInstr::FrameSetup);
1154 // Handle the 64-bit Windows ABI case where we need to call __chkstk.
1155 // Function prologue is responsible for adjusting the stack pointer.
1156 if (isUInt<32>(NumBytes)) {
1157 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
1159 .setMIFlag(MachineInstr::FrameSetup);
1160 } else if (isInt<32>(NumBytes)) {
1161 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64ri32), X86::RAX)
1163 .setMIFlag(MachineInstr::FrameSetup);
1165 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64ri), X86::RAX)
1167 .setMIFlag(MachineInstr::FrameSetup);
1170 // Allocate NumBytes-4 bytes on stack in case of isEAXAlive.
1171 // We'll also use 4 already allocated bytes for EAX.
1172 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
1173 .addImm(isEAXAlive ? NumBytes - 4 : NumBytes)
1174 .setMIFlag(MachineInstr::FrameSetup);
1177 // Call __chkstk, __chkstk_ms, or __alloca.
1178 emitStackProbe(MF, MBB, MBBI, DL, true);
1183 addRegOffset(BuildMI(MF, DL, TII.get(X86::MOV32rm), X86::EAX),
1184 StackPtr, false, NumBytes - 4);
1185 MI->setFlag(MachineInstr::FrameSetup);
1186 MBB.insert(MBBI, MI);
1188 } else if (NumBytes) {
1189 emitSPUpdate(MBB, MBBI, -(int64_t)NumBytes, /*InEpilogue=*/false);
1192 if (NeedsWinCFI && NumBytes)
1193 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_StackAlloc))
1195 .setMIFlag(MachineInstr::FrameSetup);
1197 int SEHFrameOffset = 0;
1198 unsigned SPOrEstablisher;
1201 // The establisher parameter passed to a CLR funclet is actually a pointer
1202 // to the (mostly empty) frame of its nearest enclosing funclet; we have
1203 // to find the root function establisher frame by loading the PSPSym from
1204 // the intermediate frame.
1205 unsigned PSPSlotOffset = getPSPSlotOffsetFromSP(MF);
1206 MachinePointerInfo NoInfo;
1207 MBB.addLiveIn(Establisher);
1208 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64rm), Establisher),
1209 Establisher, false, PSPSlotOffset)
1210 .addMemOperand(MF.getMachineMemOperand(
1211 NoInfo, MachineMemOperand::MOLoad, SlotSize, SlotSize));
1213 // Save the root establisher back into the current funclet's (mostly
1214 // empty) frame, in case a sub-funclet or the GC needs it.
1215 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64mr)), StackPtr,
1216 false, PSPSlotOffset)
1217 .addReg(Establisher)
1219 MF.getMachineMemOperand(NoInfo, MachineMemOperand::MOStore |
1220 MachineMemOperand::MOVolatile,
1221 SlotSize, SlotSize));
1223 SPOrEstablisher = Establisher;
1225 SPOrEstablisher = StackPtr;
1228 if (IsWin64Prologue && HasFP) {
1229 // Set RBP to a small fixed offset from RSP. In the funclet case, we base
1230 // this calculation on the incoming establisher, which holds the value of
1231 // RSP from the parent frame at the end of the prologue.
1232 SEHFrameOffset = calculateSetFPREG(ParentFrameNumBytes);
1234 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::LEA64r), FramePtr),
1235 SPOrEstablisher, false, SEHFrameOffset);
1237 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64rr), FramePtr)
1238 .addReg(SPOrEstablisher);
1240 // If this is not a funclet, emit the CFI describing our frame pointer.
1241 if (NeedsWinCFI && !IsFunclet)
1242 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_SetFrame))
1244 .addImm(SEHFrameOffset)
1245 .setMIFlag(MachineInstr::FrameSetup);
1246 } else if (IsFunclet && STI.is32Bit()) {
1247 // Reset EBP / ESI to something good for funclets.
1248 MBBI = restoreWin32EHStackPointers(MBB, MBBI, DL);
1251 while (MBBI != MBB.end() && MBBI->getFlag(MachineInstr::FrameSetup)) {
1252 const MachineInstr *FrameInstr = &*MBBI;
1257 if (unsigned Reg = TII.isStoreToStackSlot(FrameInstr, FI)) {
1258 if (X86::FR64RegClass.contains(Reg)) {
1259 unsigned IgnoredFrameReg;
1260 int Offset = getFrameIndexReference(MF, FI, IgnoredFrameReg);
1261 Offset += SEHFrameOffset;
1263 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_SaveXMM))
1266 .setMIFlag(MachineInstr::FrameSetup);
1273 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_EndPrologue))
1274 .setMIFlag(MachineInstr::FrameSetup);
1276 if (FnHasClrFunclet && !IsFunclet) {
1277 // Save the so-called Initial-SP (i.e. the value of the stack pointer
1278 // immediately after the prolog) into the PSPSlot so that funclets
1279 // and the GC can recover it.
1280 unsigned PSPSlotOffset = getPSPSlotOffsetFromSP(MF);
1281 auto PSPInfo = MachinePointerInfo::getFixedStack(
1282 MF, MF.getMMI().getWinEHFuncInfo(Fn).PSPSymFrameIdx);
1283 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64mr)), StackPtr, false,
1286 .addMemOperand(MF.getMachineMemOperand(
1287 PSPInfo, MachineMemOperand::MOStore | MachineMemOperand::MOVolatile,
1288 SlotSize, SlotSize));
1291 // Realign stack after we spilled callee-saved registers (so that we'll be
1292 // able to calculate their offsets from the frame pointer).
1293 // Win64 requires aligning the stack after the prologue.
1294 if (IsWin64Prologue && TRI->needsStackRealignment(MF)) {
1295 assert(HasFP && "There should be a frame pointer if stack is realigned.");
1296 BuildStackAlignAND(MBB, MBBI, DL, SPOrEstablisher, MaxAlign);
1299 // We already dealt with stack realignment and funclets above.
1300 if (IsFunclet && STI.is32Bit())
1303 // If we need a base pointer, set it up here. It's whatever the value
1304 // of the stack pointer is at this point. Any variable size objects
1305 // will be allocated after this, so we can still use the base pointer
1306 // to reference locals.
1307 if (TRI->hasBasePointer(MF)) {
1308 // Update the base pointer with the current stack pointer.
1309 unsigned Opc = Uses64BitFramePtr ? X86::MOV64rr : X86::MOV32rr;
1310 BuildMI(MBB, MBBI, DL, TII.get(Opc), BasePtr)
1311 .addReg(SPOrEstablisher)
1312 .setMIFlag(MachineInstr::FrameSetup);
1313 if (X86FI->getRestoreBasePointer()) {
1314 // Stash value of base pointer. Saving RSP instead of EBP shortens
1315 // dependence chain. Used by SjLj EH.
1316 unsigned Opm = Uses64BitFramePtr ? X86::MOV64mr : X86::MOV32mr;
1317 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(Opm)),
1318 FramePtr, true, X86FI->getRestoreBasePointerOffset())
1319 .addReg(SPOrEstablisher)
1320 .setMIFlag(MachineInstr::FrameSetup);
1323 if (X86FI->getHasSEHFramePtrSave() && !IsFunclet) {
1324 // Stash the value of the frame pointer relative to the base pointer for
1325 // Win32 EH. This supports Win32 EH, which does the inverse of the above:
1326 // it recovers the frame pointer from the base pointer rather than the
1327 // other way around.
1328 unsigned Opm = Uses64BitFramePtr ? X86::MOV64mr : X86::MOV32mr;
1331 getFrameIndexReference(MF, X86FI->getSEHFramePtrSaveIndex(), UsedReg);
1332 assert(UsedReg == BasePtr);
1333 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(Opm)), UsedReg, true, Offset)
1335 .setMIFlag(MachineInstr::FrameSetup);
1339 if (((!HasFP && NumBytes) || PushedRegs) && NeedsDwarfCFI) {
1340 // Mark end of stack pointer adjustment.
1341 if (!HasFP && NumBytes) {
1342 // Define the current CFA rule to use the provided offset.
1344 BuildCFI(MBB, MBBI, DL, MCCFIInstruction::createDefCfaOffset(
1345 nullptr, -StackSize + stackGrowth));
1348 // Emit DWARF info specifying the offsets of the callee-saved registers.
1350 emitCalleeSavedFrameMoves(MBB, MBBI, DL);
1354 bool X86FrameLowering::canUseLEAForSPInEpilogue(
1355 const MachineFunction &MF) const {
1356 // We can't use LEA instructions for adjusting the stack pointer if this is a
1357 // leaf function in the Win64 ABI. Only ADD instructions may be used to
1358 // deallocate the stack.
1359 // This means that we can use LEA for SP in two situations:
1360 // 1. We *aren't* using the Win64 ABI which means we are free to use LEA.
1361 // 2. We *have* a frame pointer which means we are permitted to use LEA.
1362 return !MF.getTarget().getMCAsmInfo()->usesWindowsCFI() || hasFP(MF);
1365 static bool isFuncletReturnInstr(MachineInstr *MI) {
1366 switch (MI->getOpcode()) {
1368 case X86::CLEANUPRET:
1373 llvm_unreachable("impossible");
1376 // CLR funclets use a special "Previous Stack Pointer Symbol" slot on the
1377 // stack. It holds a pointer to the bottom of the root function frame. The
1378 // establisher frame pointer passed to a nested funclet may point to the
1379 // (mostly empty) frame of its parent funclet, but it will need to find
1380 // the frame of the root function to access locals. To facilitate this,
1381 // every funclet copies the pointer to the bottom of the root function
1382 // frame into a PSPSym slot in its own (mostly empty) stack frame. Using the
1383 // same offset for the PSPSym in the root function frame that's used in the
1384 // funclets' frames allows each funclet to dynamically accept any ancestor
1385 // frame as its establisher argument (the runtime doesn't guarantee the
1386 // immediate parent for some reason lost to history), and also allows the GC,
1387 // which uses the PSPSym for some bookkeeping, to find it in any funclet's
1388 // frame with only a single offset reported for the entire method.
1390 X86FrameLowering::getPSPSlotOffsetFromSP(const MachineFunction &MF) const {
1391 MachineModuleInfo &MMI = MF.getMMI();
1392 WinEHFuncInfo &Info = MMI.getWinEHFuncInfo(MF.getFunction());
1393 // getFrameIndexReferenceFromSP has an out ref parameter for the stack
1394 // pointer register; pass a dummy that we ignore
1396 int Offset = getFrameIndexReferenceFromSP(MF, Info.PSPSymFrameIdx, SPReg);
1397 assert(Offset >= 0);
1398 return static_cast<unsigned>(Offset);
1402 X86FrameLowering::getWinEHFuncletFrameSize(const MachineFunction &MF) const {
1403 // This is the size of the pushed CSRs.
1405 MF.getInfo<X86MachineFunctionInfo>()->getCalleeSavedFrameSize();
1406 // This is the amount of stack a funclet needs to allocate.
1408 EHPersonality Personality =
1409 classifyEHPersonality(MF.getFunction()->getPersonalityFn());
1410 if (Personality == EHPersonality::CoreCLR) {
1411 // CLR funclets need to hold enough space to include the PSPSym, at the
1412 // same offset from the stack pointer (immediately after the prolog) as it
1413 // resides at in the main function.
1414 UsedSize = getPSPSlotOffsetFromSP(MF) + SlotSize;
1416 // Other funclets just need enough stack for outgoing call arguments.
1417 UsedSize = MF.getFrameInfo()->getMaxCallFrameSize();
1419 // RBP is not included in the callee saved register block. After pushing RBP,
1420 // everything is 16 byte aligned. Everything we allocate before an outgoing
1421 // call must also be 16 byte aligned.
1422 unsigned FrameSizeMinusRBP =
1423 RoundUpToAlignment(CSSize + UsedSize, getStackAlignment());
1424 // Subtract out the size of the callee saved registers. This is how much stack
1425 // each funclet will allocate.
1426 return FrameSizeMinusRBP - CSSize;
1429 void X86FrameLowering::emitEpilogue(MachineFunction &MF,
1430 MachineBasicBlock &MBB) const {
1431 const MachineFrameInfo *MFI = MF.getFrameInfo();
1432 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1433 MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator();
1435 if (MBBI != MBB.end())
1436 DL = MBBI->getDebugLoc();
1437 // standard x86_64 and NaCl use 64-bit frame/stack pointers, x32 - 32-bit.
1438 const bool Is64BitILP32 = STI.isTarget64BitILP32();
1439 unsigned FramePtr = TRI->getFrameRegister(MF);
1440 unsigned MachineFramePtr =
1441 Is64BitILP32 ? getX86SubSuperRegister(FramePtr, MVT::i64, false)
1444 bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
1446 IsWin64Prologue && MF.getFunction()->needsUnwindTableEntry();
1447 bool IsFunclet = isFuncletReturnInstr(MBBI);
1448 MachineBasicBlock *TargetMBB = nullptr;
1450 // Get the number of bytes to allocate from the FrameInfo.
1451 uint64_t StackSize = MFI->getStackSize();
1452 uint64_t MaxAlign = calculateMaxStackAlign(MF);
1453 unsigned CSSize = X86FI->getCalleeSavedFrameSize();
1454 uint64_t NumBytes = 0;
1456 if (MBBI->getOpcode() == X86::CATCHRET) {
1457 // SEH shouldn't use catchret.
1458 assert(!isAsynchronousEHPersonality(
1459 classifyEHPersonality(MF.getFunction()->getPersonalityFn())) &&
1460 "SEH should not use CATCHRET");
1462 NumBytes = getWinEHFuncletFrameSize(MF);
1463 assert(hasFP(MF) && "EH funclets without FP not yet implemented");
1464 TargetMBB = MBBI->getOperand(0).getMBB();
1467 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::POP64r : X86::POP32r),
1469 .setMIFlag(MachineInstr::FrameDestroy);
1470 } else if (MBBI->getOpcode() == X86::CLEANUPRET) {
1471 NumBytes = getWinEHFuncletFrameSize(MF);
1472 assert(hasFP(MF) && "EH funclets without FP not yet implemented");
1473 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::POP64r : X86::POP32r),
1475 .setMIFlag(MachineInstr::FrameDestroy);
1476 } else if (hasFP(MF)) {
1477 // Calculate required stack adjustment.
1478 uint64_t FrameSize = StackSize - SlotSize;
1479 NumBytes = FrameSize - CSSize;
1481 // Callee-saved registers were pushed on stack before the stack was
1483 if (TRI->needsStackRealignment(MF) && !IsWin64Prologue)
1484 NumBytes = RoundUpToAlignment(FrameSize, MaxAlign);
1487 BuildMI(MBB, MBBI, DL,
1488 TII.get(Is64Bit ? X86::POP64r : X86::POP32r), MachineFramePtr)
1489 .setMIFlag(MachineInstr::FrameDestroy);
1491 NumBytes = StackSize - CSSize;
1493 uint64_t SEHStackAllocAmt = NumBytes;
1495 // Skip the callee-saved pop instructions.
1496 while (MBBI != MBB.begin()) {
1497 MachineBasicBlock::iterator PI = std::prev(MBBI);
1498 unsigned Opc = PI->getOpcode();
1500 if ((Opc != X86::POP32r || !PI->getFlag(MachineInstr::FrameDestroy)) &&
1501 (Opc != X86::POP64r || !PI->getFlag(MachineInstr::FrameDestroy)) &&
1502 Opc != X86::DBG_VALUE && !PI->isTerminator())
1507 MachineBasicBlock::iterator FirstCSPop = MBBI;
1510 // Fill EAX/RAX with the address of the target block.
1511 unsigned ReturnReg = STI.is64Bit() ? X86::RAX : X86::EAX;
1512 if (STI.is64Bit()) {
1513 // LEA64r TargetMBB(%rip), %rax
1514 BuildMI(MBB, FirstCSPop, DL, TII.get(X86::LEA64r), ReturnReg)
1521 // MOV32ri $TargetMBB, %eax
1522 BuildMI(MBB, FirstCSPop, DL, TII.get(X86::MOV32ri), ReturnReg)
1525 // Record that we've taken the address of TargetMBB and no longer just
1526 // reference it in a terminator.
1527 TargetMBB->setHasAddressTaken();
1530 if (MBBI != MBB.end())
1531 DL = MBBI->getDebugLoc();
1533 // If there is an ADD32ri or SUB32ri of ESP immediately before this
1534 // instruction, merge the two instructions.
1535 if (NumBytes || MFI->hasVarSizedObjects())
1536 NumBytes += mergeSPUpdates(MBB, MBBI, true);
1538 // If dynamic alloca is used, then reset esp to point to the last callee-saved
1539 // slot before popping them off! Same applies for the case, when stack was
1540 // realigned. Don't do this if this was a funclet epilogue, since the funclets
1541 // will not do realignment or dynamic stack allocation.
1542 if ((TRI->needsStackRealignment(MF) || MFI->hasVarSizedObjects()) &&
1544 if (TRI->needsStackRealignment(MF))
1546 unsigned SEHFrameOffset = calculateSetFPREG(SEHStackAllocAmt);
1547 uint64_t LEAAmount =
1548 IsWin64Prologue ? SEHStackAllocAmt - SEHFrameOffset : -CSSize;
1550 // There are only two legal forms of epilogue:
1551 // - add SEHAllocationSize, %rsp
1552 // - lea SEHAllocationSize(%FramePtr), %rsp
1554 // 'mov %FramePtr, %rsp' will not be recognized as an epilogue sequence.
1555 // However, we may use this sequence if we have a frame pointer because the
1556 // effects of the prologue can safely be undone.
1557 if (LEAAmount != 0) {
1558 unsigned Opc = getLEArOpcode(Uses64BitFramePtr);
1559 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr),
1560 FramePtr, false, LEAAmount);
1563 unsigned Opc = (Uses64BitFramePtr ? X86::MOV64rr : X86::MOV32rr);
1564 BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr)
1568 } else if (NumBytes) {
1569 // Adjust stack pointer back: ESP += numbytes.
1570 emitSPUpdate(MBB, MBBI, NumBytes, /*InEpilogue=*/true);
1574 // Windows unwinder will not invoke function's exception handler if IP is
1575 // either in prologue or in epilogue. This behavior causes a problem when a
1576 // call immediately precedes an epilogue, because the return address points
1577 // into the epilogue. To cope with that, we insert an epilogue marker here,
1578 // then replace it with a 'nop' if it ends up immediately after a CALL in the
1579 // final emitted code.
1581 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_Epilogue));
1583 // Add the return addr area delta back since we are not tail calling.
1584 int Offset = -1 * X86FI->getTCReturnAddrDelta();
1585 assert(Offset >= 0 && "TCDelta should never be positive");
1587 MBBI = MBB.getFirstTerminator();
1589 // Check for possible merge with preceding ADD instruction.
1590 Offset += mergeSPUpdates(MBB, MBBI, true);
1591 emitSPUpdate(MBB, MBBI, Offset, /*InEpilogue=*/true);
1595 // NOTE: this only has a subset of the full frame index logic. In
1596 // particular, the FI < 0 and AfterFPPop logic is handled in
1597 // X86RegisterInfo::eliminateFrameIndex, but not here. Possibly
1598 // (probably?) it should be moved into here.
1599 int X86FrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI,
1600 unsigned &FrameReg) const {
1601 const MachineFrameInfo *MFI = MF.getFrameInfo();
1603 // We can't calculate offset from frame pointer if the stack is realigned,
1604 // so enforce usage of stack/base pointer. The base pointer is used when we
1605 // have dynamic allocas in addition to dynamic realignment.
1606 if (TRI->hasBasePointer(MF))
1607 FrameReg = TRI->getBaseRegister();
1608 else if (TRI->needsStackRealignment(MF))
1609 FrameReg = TRI->getStackRegister();
1611 FrameReg = TRI->getFrameRegister(MF);
1613 // Offset will hold the offset from the stack pointer at function entry to the
1615 // We need to factor in additional offsets applied during the prologue to the
1616 // frame, base, and stack pointer depending on which is used.
1617 int Offset = MFI->getObjectOffset(FI) - getOffsetOfLocalArea();
1618 const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1619 unsigned CSSize = X86FI->getCalleeSavedFrameSize();
1620 uint64_t StackSize = MFI->getStackSize();
1621 bool HasFP = hasFP(MF);
1622 bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
1623 int64_t FPDelta = 0;
1625 if (IsWin64Prologue) {
1626 assert(!MFI->hasCalls() || (StackSize % 16) == 8);
1628 // Calculate required stack adjustment.
1629 uint64_t FrameSize = StackSize - SlotSize;
1630 // If required, include space for extra hidden slot for stashing base pointer.
1631 if (X86FI->getRestoreBasePointer())
1632 FrameSize += SlotSize;
1633 uint64_t NumBytes = FrameSize - CSSize;
1635 uint64_t SEHFrameOffset = calculateSetFPREG(NumBytes);
1636 if (FI && FI == X86FI->getFAIndex())
1637 return -SEHFrameOffset;
1639 // FPDelta is the offset from the "traditional" FP location of the old base
1640 // pointer followed by return address and the location required by the
1641 // restricted Win64 prologue.
1642 // Add FPDelta to all offsets below that go through the frame pointer.
1643 FPDelta = FrameSize - SEHFrameOffset;
1644 assert((!MFI->hasCalls() || (FPDelta % 16) == 0) &&
1645 "FPDelta isn't aligned per the Win64 ABI!");
1649 if (TRI->hasBasePointer(MF)) {
1650 assert(HasFP && "VLAs and dynamic stack realign, but no FP?!");
1652 // Skip the saved EBP.
1653 return Offset + SlotSize + FPDelta;
1655 assert((-(Offset + StackSize)) % MFI->getObjectAlignment(FI) == 0);
1656 return Offset + StackSize;
1658 } else if (TRI->needsStackRealignment(MF)) {
1660 // Skip the saved EBP.
1661 return Offset + SlotSize + FPDelta;
1663 assert((-(Offset + StackSize)) % MFI->getObjectAlignment(FI) == 0);
1664 return Offset + StackSize;
1666 // FIXME: Support tail calls
1669 return Offset + StackSize;
1671 // Skip the saved EBP.
1674 // Skip the RETADDR move area
1675 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
1676 if (TailCallReturnAddrDelta < 0)
1677 Offset -= TailCallReturnAddrDelta;
1680 return Offset + FPDelta;
1683 // Simplified from getFrameIndexReference keeping only StackPointer cases
1684 int X86FrameLowering::getFrameIndexReferenceFromSP(const MachineFunction &MF,
1686 unsigned &FrameReg) const {
1687 const MachineFrameInfo *MFI = MF.getFrameInfo();
1688 // Does not include any dynamic realign.
1689 const uint64_t StackSize = MFI->getStackSize();
1692 // LLVM arranges the stack as follows:
1697 // PUSH RBP <-- RBP points here
1699 // ~~~~~~~ <-- optional stack realignment dynamic adjustment
1702 // ... <-- RSP after prologue points here
1704 // if (hasVarSizedObjects()):
1705 // ... <-- "base pointer" (ESI/RBX) points here
1707 // ... <-- RSP points here
1709 // Case 1: In the simple case of no stack realignment and no dynamic
1710 // allocas, both "fixed" stack objects (arguments and CSRs) are addressable
1711 // with fixed offsets from RSP.
1713 // Case 2: In the case of stack realignment with no dynamic allocas, fixed
1714 // stack objects are addressed with RBP and regular stack objects with RSP.
1716 // Case 3: In the case of dynamic allocas and stack realignment, RSP is used
1717 // to address stack arguments for outgoing calls and nothing else. The "base
1718 // pointer" points to local variables, and RBP points to fixed objects.
1720 // In cases 2 and 3, we can only answer for non-fixed stack objects, and the
1721 // answer we give is relative to the SP after the prologue, and not the
1722 // SP in the middle of the function.
1724 assert((!TRI->needsStackRealignment(MF) || !MFI->isFixedObjectIndex(FI)) &&
1725 "offset from fixed object to SP is not static");
1727 // We don't handle tail calls, and shouldn't be seeing them either.
1728 int TailCallReturnAddrDelta =
1729 MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta();
1730 assert(!(TailCallReturnAddrDelta < 0) && "we don't handle this case!");
1734 // Fill in FrameReg output argument.
1735 FrameReg = TRI->getStackRegister();
1737 // This is how the math works out:
1739 // %rsp grows (i.e. gets lower) left to right. Each box below is
1740 // one word (eight bytes). Obj0 is the stack slot we're trying to
1743 // ----------------------------------
1744 // | BP | Obj0 | Obj1 | ... | ObjN |
1745 // ----------------------------------
1749 // A is the incoming stack pointer.
1750 // (B - A) is the local area offset (-8 for x86-64) [1]
1751 // (C - A) is the Offset returned by MFI->getObjectOffset for Obj0 [2]
1753 // |(E - B)| is the StackSize (absolute value, positive). For a
1754 // stack that grown down, this works out to be (B - E). [3]
1756 // E is also the value of %rsp after stack has been set up, and we
1757 // want (C - E) -- the value we can add to %rsp to get to Obj0. Now
1758 // (C - E) == (C - A) - (B - A) + (B - E)
1759 // { Using [1], [2] and [3] above }
1760 // == getObjectOffset - LocalAreaOffset + StackSize
1763 // Get the Offset from the StackPointer
1764 int Offset = MFI->getObjectOffset(FI) - getOffsetOfLocalArea();
1766 return Offset + StackSize;
1769 bool X86FrameLowering::assignCalleeSavedSpillSlots(
1770 MachineFunction &MF, const TargetRegisterInfo *TRI,
1771 std::vector<CalleeSavedInfo> &CSI) const {
1772 MachineFrameInfo *MFI = MF.getFrameInfo();
1773 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1775 unsigned CalleeSavedFrameSize = 0;
1776 int SpillSlotOffset = getOffsetOfLocalArea() + X86FI->getTCReturnAddrDelta();
1779 // emitPrologue always spills frame register the first thing.
1780 SpillSlotOffset -= SlotSize;
1781 MFI->CreateFixedSpillStackObject(SlotSize, SpillSlotOffset);
1783 // Since emitPrologue and emitEpilogue will handle spilling and restoring of
1784 // the frame register, we can delete it from CSI list and not have to worry
1785 // about avoiding it later.
1786 unsigned FPReg = TRI->getFrameRegister(MF);
1787 for (unsigned i = 0; i < CSI.size(); ++i) {
1788 if (TRI->regsOverlap(CSI[i].getReg(),FPReg)) {
1789 CSI.erase(CSI.begin() + i);
1795 // Assign slots for GPRs. It increases frame size.
1796 for (unsigned i = CSI.size(); i != 0; --i) {
1797 unsigned Reg = CSI[i - 1].getReg();
1799 if (!X86::GR64RegClass.contains(Reg) && !X86::GR32RegClass.contains(Reg))
1802 SpillSlotOffset -= SlotSize;
1803 CalleeSavedFrameSize += SlotSize;
1805 int SlotIndex = MFI->CreateFixedSpillStackObject(SlotSize, SpillSlotOffset);
1806 CSI[i - 1].setFrameIdx(SlotIndex);
1809 X86FI->setCalleeSavedFrameSize(CalleeSavedFrameSize);
1811 // Assign slots for XMMs.
1812 for (unsigned i = CSI.size(); i != 0; --i) {
1813 unsigned Reg = CSI[i - 1].getReg();
1814 if (X86::GR64RegClass.contains(Reg) || X86::GR32RegClass.contains(Reg))
1817 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
1819 SpillSlotOffset -= std::abs(SpillSlotOffset) % RC->getAlignment();
1821 SpillSlotOffset -= RC->getSize();
1823 MFI->CreateFixedSpillStackObject(RC->getSize(), SpillSlotOffset);
1824 CSI[i - 1].setFrameIdx(SlotIndex);
1825 MFI->ensureMaxAlignment(RC->getAlignment());
1831 bool X86FrameLowering::spillCalleeSavedRegisters(
1832 MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
1833 const std::vector<CalleeSavedInfo> &CSI,
1834 const TargetRegisterInfo *TRI) const {
1835 DebugLoc DL = MBB.findDebugLoc(MI);
1837 // Don't save CSRs in 32-bit EH funclets. The caller saves EBX, EBP, ESI, EDI
1838 // for us, and there are no XMM CSRs on Win32.
1839 if (MBB.isEHFuncletEntry() && STI.is32Bit() && STI.isOSWindows())
1842 // Push GPRs. It increases frame size.
1843 unsigned Opc = STI.is64Bit() ? X86::PUSH64r : X86::PUSH32r;
1844 for (unsigned i = CSI.size(); i != 0; --i) {
1845 unsigned Reg = CSI[i - 1].getReg();
1847 if (!X86::GR64RegClass.contains(Reg) && !X86::GR32RegClass.contains(Reg))
1849 // Add the callee-saved register as live-in. It's killed at the spill.
1852 BuildMI(MBB, MI, DL, TII.get(Opc)).addReg(Reg, RegState::Kill)
1853 .setMIFlag(MachineInstr::FrameSetup);
1856 // Make XMM regs spilled. X86 does not have ability of push/pop XMM.
1857 // It can be done by spilling XMMs to stack frame.
1858 for (unsigned i = CSI.size(); i != 0; --i) {
1859 unsigned Reg = CSI[i-1].getReg();
1860 if (X86::GR64RegClass.contains(Reg) || X86::GR32RegClass.contains(Reg))
1862 // Add the callee-saved register as live-in. It's killed at the spill.
1864 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
1866 TII.storeRegToStackSlot(MBB, MI, Reg, true, CSI[i - 1].getFrameIdx(), RC,
1869 MI->setFlag(MachineInstr::FrameSetup);
1876 bool X86FrameLowering::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1877 MachineBasicBlock::iterator MI,
1878 const std::vector<CalleeSavedInfo> &CSI,
1879 const TargetRegisterInfo *TRI) const {
1883 if (isFuncletReturnInstr(MI) && STI.isOSWindows()) {
1884 // Don't restore CSRs in 32-bit EH funclets. Matches
1885 // spillCalleeSavedRegisters.
1888 // Don't restore CSRs before an SEH catchret. SEH except blocks do not form
1889 // funclets. emitEpilogue transforms these to normal jumps.
1890 if (MI->getOpcode() == X86::CATCHRET) {
1891 const Function *Func = MBB.getParent()->getFunction();
1892 bool IsSEH = isAsynchronousEHPersonality(
1893 classifyEHPersonality(Func->getPersonalityFn()));
1899 DebugLoc DL = MBB.findDebugLoc(MI);
1901 // Reload XMMs from stack frame.
1902 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1903 unsigned Reg = CSI[i].getReg();
1904 if (X86::GR64RegClass.contains(Reg) ||
1905 X86::GR32RegClass.contains(Reg))
1908 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
1909 TII.loadRegFromStackSlot(MBB, MI, Reg, CSI[i].getFrameIdx(), RC, TRI);
1913 unsigned Opc = STI.is64Bit() ? X86::POP64r : X86::POP32r;
1914 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1915 unsigned Reg = CSI[i].getReg();
1916 if (!X86::GR64RegClass.contains(Reg) &&
1917 !X86::GR32RegClass.contains(Reg))
1920 BuildMI(MBB, MI, DL, TII.get(Opc), Reg)
1921 .setMIFlag(MachineInstr::FrameDestroy);
1926 void X86FrameLowering::determineCalleeSaves(MachineFunction &MF,
1927 BitVector &SavedRegs,
1928 RegScavenger *RS) const {
1929 TargetFrameLowering::determineCalleeSaves(MF, SavedRegs, RS);
1931 MachineFrameInfo *MFI = MF.getFrameInfo();
1933 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1934 int64_t TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
1936 if (TailCallReturnAddrDelta < 0) {
1937 // create RETURNADDR area
1946 MFI->CreateFixedObject(-TailCallReturnAddrDelta,
1947 TailCallReturnAddrDelta - SlotSize, true);
1950 // Spill the BasePtr if it's used.
1951 if (TRI->hasBasePointer(MF)) {
1952 SavedRegs.set(TRI->getBaseRegister());
1954 // Allocate a spill slot for EBP if we have a base pointer and EH funclets.
1955 if (MF.getMMI().hasEHFunclets()) {
1956 int FI = MFI->CreateSpillStackObject(SlotSize, SlotSize);
1957 X86FI->setHasSEHFramePtrSave(true);
1958 X86FI->setSEHFramePtrSaveIndex(FI);
1964 HasNestArgument(const MachineFunction *MF) {
1965 const Function *F = MF->getFunction();
1966 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
1968 if (I->hasNestAttr())
1974 /// GetScratchRegister - Get a temp register for performing work in the
1975 /// segmented stack and the Erlang/HiPE stack prologue. Depending on platform
1976 /// and the properties of the function either one or two registers will be
1977 /// needed. Set primary to true for the first register, false for the second.
1979 GetScratchRegister(bool Is64Bit, bool IsLP64, const MachineFunction &MF, bool Primary) {
1980 CallingConv::ID CallingConvention = MF.getFunction()->getCallingConv();
1983 if (CallingConvention == CallingConv::HiPE) {
1985 return Primary ? X86::R14 : X86::R13;
1987 return Primary ? X86::EBX : X86::EDI;
1992 return Primary ? X86::R11 : X86::R12;
1994 return Primary ? X86::R11D : X86::R12D;
1997 bool IsNested = HasNestArgument(&MF);
1999 if (CallingConvention == CallingConv::X86_FastCall ||
2000 CallingConvention == CallingConv::Fast) {
2002 report_fatal_error("Segmented stacks does not support fastcall with "
2003 "nested function.");
2004 return Primary ? X86::EAX : X86::ECX;
2007 return Primary ? X86::EDX : X86::EAX;
2008 return Primary ? X86::ECX : X86::EAX;
2011 // The stack limit in the TCB is set to this many bytes above the actual stack
2013 static const uint64_t kSplitStackAvailable = 256;
2015 void X86FrameLowering::adjustForSegmentedStacks(
2016 MachineFunction &MF, MachineBasicBlock &PrologueMBB) const {
2017 MachineFrameInfo *MFI = MF.getFrameInfo();
2019 unsigned TlsReg, TlsOffset;
2022 unsigned ScratchReg = GetScratchRegister(Is64Bit, IsLP64, MF, true);
2023 assert(!MF.getRegInfo().isLiveIn(ScratchReg) &&
2024 "Scratch register is live-in");
2026 if (MF.getFunction()->isVarArg())
2027 report_fatal_error("Segmented stacks do not support vararg functions.");
2028 if (!STI.isTargetLinux() && !STI.isTargetDarwin() && !STI.isTargetWin32() &&
2029 !STI.isTargetWin64() && !STI.isTargetFreeBSD() &&
2030 !STI.isTargetDragonFly())
2031 report_fatal_error("Segmented stacks not supported on this platform.");
2033 // Eventually StackSize will be calculated by a link-time pass; which will
2034 // also decide whether checking code needs to be injected into this particular
2036 StackSize = MFI->getStackSize();
2038 // Do not generate a prologue for functions with a stack of size zero
2042 MachineBasicBlock *allocMBB = MF.CreateMachineBasicBlock();
2043 MachineBasicBlock *checkMBB = MF.CreateMachineBasicBlock();
2044 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
2045 bool IsNested = false;
2047 // We need to know if the function has a nest argument only in 64 bit mode.
2049 IsNested = HasNestArgument(&MF);
2051 // The MOV R10, RAX needs to be in a different block, since the RET we emit in
2052 // allocMBB needs to be last (terminating) instruction.
2054 for (const auto &LI : PrologueMBB.liveins()) {
2055 allocMBB->addLiveIn(LI);
2056 checkMBB->addLiveIn(LI);
2060 allocMBB->addLiveIn(IsLP64 ? X86::R10 : X86::R10D);
2062 MF.push_front(allocMBB);
2063 MF.push_front(checkMBB);
2065 // When the frame size is less than 256 we just compare the stack
2066 // boundary directly to the value of the stack pointer, per gcc.
2067 bool CompareStackPointer = StackSize < kSplitStackAvailable;
2069 // Read the limit off the current stacklet off the stack_guard location.
2071 if (STI.isTargetLinux()) {
2073 TlsOffset = IsLP64 ? 0x70 : 0x40;
2074 } else if (STI.isTargetDarwin()) {
2076 TlsOffset = 0x60 + 90*8; // See pthread_machdep.h. Steal TLS slot 90.
2077 } else if (STI.isTargetWin64()) {
2079 TlsOffset = 0x28; // pvArbitrary, reserved for application use
2080 } else if (STI.isTargetFreeBSD()) {
2083 } else if (STI.isTargetDragonFly()) {
2085 TlsOffset = 0x20; // use tls_tcb.tcb_segstack
2087 report_fatal_error("Segmented stacks not supported on this platform.");
2090 if (CompareStackPointer)
2091 ScratchReg = IsLP64 ? X86::RSP : X86::ESP;
2093 BuildMI(checkMBB, DL, TII.get(IsLP64 ? X86::LEA64r : X86::LEA64_32r), ScratchReg).addReg(X86::RSP)
2094 .addImm(1).addReg(0).addImm(-StackSize).addReg(0);
2096 BuildMI(checkMBB, DL, TII.get(IsLP64 ? X86::CMP64rm : X86::CMP32rm)).addReg(ScratchReg)
2097 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg);
2099 if (STI.isTargetLinux()) {
2102 } else if (STI.isTargetDarwin()) {
2104 TlsOffset = 0x48 + 90*4;
2105 } else if (STI.isTargetWin32()) {
2107 TlsOffset = 0x14; // pvArbitrary, reserved for application use
2108 } else if (STI.isTargetDragonFly()) {
2110 TlsOffset = 0x10; // use tls_tcb.tcb_segstack
2111 } else if (STI.isTargetFreeBSD()) {
2112 report_fatal_error("Segmented stacks not supported on FreeBSD i386.");
2114 report_fatal_error("Segmented stacks not supported on this platform.");
2117 if (CompareStackPointer)
2118 ScratchReg = X86::ESP;
2120 BuildMI(checkMBB, DL, TII.get(X86::LEA32r), ScratchReg).addReg(X86::ESP)
2121 .addImm(1).addReg(0).addImm(-StackSize).addReg(0);
2123 if (STI.isTargetLinux() || STI.isTargetWin32() || STI.isTargetWin64() ||
2124 STI.isTargetDragonFly()) {
2125 BuildMI(checkMBB, DL, TII.get(X86::CMP32rm)).addReg(ScratchReg)
2126 .addReg(0).addImm(0).addReg(0).addImm(TlsOffset).addReg(TlsReg);
2127 } else if (STI.isTargetDarwin()) {
2129 // TlsOffset doesn't fit into a mod r/m byte so we need an extra register.
2130 unsigned ScratchReg2;
2132 if (CompareStackPointer) {
2133 // The primary scratch register is available for holding the TLS offset.
2134 ScratchReg2 = GetScratchRegister(Is64Bit, IsLP64, MF, true);
2135 SaveScratch2 = false;
2137 // Need to use a second register to hold the TLS offset
2138 ScratchReg2 = GetScratchRegister(Is64Bit, IsLP64, MF, false);
2140 // Unfortunately, with fastcc the second scratch register may hold an
2142 SaveScratch2 = MF.getRegInfo().isLiveIn(ScratchReg2);
2145 // If Scratch2 is live-in then it needs to be saved.
2146 assert((!MF.getRegInfo().isLiveIn(ScratchReg2) || SaveScratch2) &&
2147 "Scratch register is live-in and not saved");
2150 BuildMI(checkMBB, DL, TII.get(X86::PUSH32r))
2151 .addReg(ScratchReg2, RegState::Kill);
2153 BuildMI(checkMBB, DL, TII.get(X86::MOV32ri), ScratchReg2)
2155 BuildMI(checkMBB, DL, TII.get(X86::CMP32rm))
2157 .addReg(ScratchReg2).addImm(1).addReg(0)
2162 BuildMI(checkMBB, DL, TII.get(X86::POP32r), ScratchReg2);
2166 // This jump is taken if SP >= (Stacklet Limit + Stack Space required).
2167 // It jumps to normal execution of the function body.
2168 BuildMI(checkMBB, DL, TII.get(X86::JA_1)).addMBB(&PrologueMBB);
2170 // On 32 bit we first push the arguments size and then the frame size. On 64
2171 // bit, we pass the stack frame size in r10 and the argument size in r11.
2173 // Functions with nested arguments use R10, so it needs to be saved across
2174 // the call to _morestack
2176 const unsigned RegAX = IsLP64 ? X86::RAX : X86::EAX;
2177 const unsigned Reg10 = IsLP64 ? X86::R10 : X86::R10D;
2178 const unsigned Reg11 = IsLP64 ? X86::R11 : X86::R11D;
2179 const unsigned MOVrr = IsLP64 ? X86::MOV64rr : X86::MOV32rr;
2180 const unsigned MOVri = IsLP64 ? X86::MOV64ri : X86::MOV32ri;
2183 BuildMI(allocMBB, DL, TII.get(MOVrr), RegAX).addReg(Reg10);
2185 BuildMI(allocMBB, DL, TII.get(MOVri), Reg10)
2187 BuildMI(allocMBB, DL, TII.get(MOVri), Reg11)
2188 .addImm(X86FI->getArgumentStackSize());
2190 BuildMI(allocMBB, DL, TII.get(X86::PUSHi32))
2191 .addImm(X86FI->getArgumentStackSize());
2192 BuildMI(allocMBB, DL, TII.get(X86::PUSHi32))
2196 // __morestack is in libgcc
2197 if (Is64Bit && MF.getTarget().getCodeModel() == CodeModel::Large) {
2198 // Under the large code model, we cannot assume that __morestack lives
2199 // within 2^31 bytes of the call site, so we cannot use pc-relative
2200 // addressing. We cannot perform the call via a temporary register,
2201 // as the rax register may be used to store the static chain, and all
2202 // other suitable registers may be either callee-save or used for
2203 // parameter passing. We cannot use the stack at this point either
2204 // because __morestack manipulates the stack directly.
2206 // To avoid these issues, perform an indirect call via a read-only memory
2207 // location containing the address.
2209 // This solution is not perfect, as it assumes that the .rodata section
2210 // is laid out within 2^31 bytes of each function body, but this seems
2211 // to be sufficient for JIT.
2212 BuildMI(allocMBB, DL, TII.get(X86::CALL64m))
2216 .addExternalSymbol("__morestack_addr")
2218 MF.getMMI().setUsesMorestackAddr(true);
2221 BuildMI(allocMBB, DL, TII.get(X86::CALL64pcrel32))
2222 .addExternalSymbol("__morestack");
2224 BuildMI(allocMBB, DL, TII.get(X86::CALLpcrel32))
2225 .addExternalSymbol("__morestack");
2229 BuildMI(allocMBB, DL, TII.get(X86::MORESTACK_RET_RESTORE_R10));
2231 BuildMI(allocMBB, DL, TII.get(X86::MORESTACK_RET));
2233 allocMBB->addSuccessor(&PrologueMBB);
2235 checkMBB->addSuccessor(allocMBB);
2236 checkMBB->addSuccessor(&PrologueMBB);
2243 /// Erlang programs may need a special prologue to handle the stack size they
2244 /// might need at runtime. That is because Erlang/OTP does not implement a C
2245 /// stack but uses a custom implementation of hybrid stack/heap architecture.
2246 /// (for more information see Eric Stenman's Ph.D. thesis:
2247 /// http://publications.uu.se/uu/fulltext/nbn_se_uu_diva-2688.pdf)
2250 /// temp0 = sp - MaxStack
2251 /// if( temp0 < SP_LIMIT(P) ) goto IncStack else goto OldStart
2255 /// call inc_stack # doubles the stack space
2256 /// temp0 = sp - MaxStack
2257 /// if( temp0 < SP_LIMIT(P) ) goto IncStack else goto OldStart
2258 void X86FrameLowering::adjustForHiPEPrologue(
2259 MachineFunction &MF, MachineBasicBlock &PrologueMBB) const {
2260 MachineFrameInfo *MFI = MF.getFrameInfo();
2262 // HiPE-specific values
2263 const unsigned HipeLeafWords = 24;
2264 const unsigned CCRegisteredArgs = Is64Bit ? 6 : 5;
2265 const unsigned Guaranteed = HipeLeafWords * SlotSize;
2266 unsigned CallerStkArity = MF.getFunction()->arg_size() > CCRegisteredArgs ?
2267 MF.getFunction()->arg_size() - CCRegisteredArgs : 0;
2268 unsigned MaxStack = MFI->getStackSize() + CallerStkArity*SlotSize + SlotSize;
2270 assert(STI.isTargetLinux() &&
2271 "HiPE prologue is only supported on Linux operating systems.");
2273 // Compute the largest caller's frame that is needed to fit the callees'
2274 // frames. This 'MaxStack' is computed from:
2276 // a) the fixed frame size, which is the space needed for all spilled temps,
2277 // b) outgoing on-stack parameter areas, and
2278 // c) the minimum stack space this function needs to make available for the
2279 // functions it calls (a tunable ABI property).
2280 if (MFI->hasCalls()) {
2281 unsigned MoreStackForCalls = 0;
2283 for (MachineFunction::iterator MBBI = MF.begin(), MBBE = MF.end();
2284 MBBI != MBBE; ++MBBI)
2285 for (MachineBasicBlock::iterator MI = MBBI->begin(), ME = MBBI->end();
2290 // Get callee operand.
2291 const MachineOperand &MO = MI->getOperand(0);
2293 // Only take account of global function calls (no closures etc.).
2297 const Function *F = dyn_cast<Function>(MO.getGlobal());
2301 // Do not update 'MaxStack' for primitive and built-in functions
2302 // (encoded with names either starting with "erlang."/"bif_" or not
2303 // having a ".", such as a simple <Module>.<Function>.<Arity>, or an
2304 // "_", such as the BIF "suspend_0") as they are executed on another
2306 if (F->getName().find("erlang.") != StringRef::npos ||
2307 F->getName().find("bif_") != StringRef::npos ||
2308 F->getName().find_first_of("._") == StringRef::npos)
2311 unsigned CalleeStkArity =
2312 F->arg_size() > CCRegisteredArgs ? F->arg_size()-CCRegisteredArgs : 0;
2313 if (HipeLeafWords - 1 > CalleeStkArity)
2314 MoreStackForCalls = std::max(MoreStackForCalls,
2315 (HipeLeafWords - 1 - CalleeStkArity) * SlotSize);
2317 MaxStack += MoreStackForCalls;
2320 // If the stack frame needed is larger than the guaranteed then runtime checks
2321 // and calls to "inc_stack_0" BIF should be inserted in the assembly prologue.
2322 if (MaxStack > Guaranteed) {
2323 MachineBasicBlock *stackCheckMBB = MF.CreateMachineBasicBlock();
2324 MachineBasicBlock *incStackMBB = MF.CreateMachineBasicBlock();
2326 for (const auto &LI : PrologueMBB.liveins()) {
2327 stackCheckMBB->addLiveIn(LI);
2328 incStackMBB->addLiveIn(LI);
2331 MF.push_front(incStackMBB);
2332 MF.push_front(stackCheckMBB);
2334 unsigned ScratchReg, SPReg, PReg, SPLimitOffset;
2335 unsigned LEAop, CMPop, CALLop;
2339 LEAop = X86::LEA64r;
2340 CMPop = X86::CMP64rm;
2341 CALLop = X86::CALL64pcrel32;
2342 SPLimitOffset = 0x90;
2346 LEAop = X86::LEA32r;
2347 CMPop = X86::CMP32rm;
2348 CALLop = X86::CALLpcrel32;
2349 SPLimitOffset = 0x4c;
2352 ScratchReg = GetScratchRegister(Is64Bit, IsLP64, MF, true);
2353 assert(!MF.getRegInfo().isLiveIn(ScratchReg) &&
2354 "HiPE prologue scratch register is live-in");
2356 // Create new MBB for StackCheck:
2357 addRegOffset(BuildMI(stackCheckMBB, DL, TII.get(LEAop), ScratchReg),
2358 SPReg, false, -MaxStack);
2359 // SPLimitOffset is in a fixed heap location (pointed by BP).
2360 addRegOffset(BuildMI(stackCheckMBB, DL, TII.get(CMPop))
2361 .addReg(ScratchReg), PReg, false, SPLimitOffset);
2362 BuildMI(stackCheckMBB, DL, TII.get(X86::JAE_1)).addMBB(&PrologueMBB);
2364 // Create new MBB for IncStack:
2365 BuildMI(incStackMBB, DL, TII.get(CALLop)).
2366 addExternalSymbol("inc_stack_0");
2367 addRegOffset(BuildMI(incStackMBB, DL, TII.get(LEAop), ScratchReg),
2368 SPReg, false, -MaxStack);
2369 addRegOffset(BuildMI(incStackMBB, DL, TII.get(CMPop))
2370 .addReg(ScratchReg), PReg, false, SPLimitOffset);
2371 BuildMI(incStackMBB, DL, TII.get(X86::JLE_1)).addMBB(incStackMBB);
2373 stackCheckMBB->addSuccessor(&PrologueMBB, 99);
2374 stackCheckMBB->addSuccessor(incStackMBB, 1);
2375 incStackMBB->addSuccessor(&PrologueMBB, 99);
2376 incStackMBB->addSuccessor(incStackMBB, 1);
2383 bool X86FrameLowering::adjustStackWithPops(MachineBasicBlock &MBB,
2384 MachineBasicBlock::iterator MBBI, DebugLoc DL, int Offset) const {
2389 if (Offset % SlotSize)
2392 int NumPops = Offset / SlotSize;
2393 // This is only worth it if we have at most 2 pops.
2394 if (NumPops != 1 && NumPops != 2)
2397 // Handle only the trivial case where the adjustment directly follows
2398 // a call. This is the most common one, anyway.
2399 if (MBBI == MBB.begin())
2401 MachineBasicBlock::iterator Prev = std::prev(MBBI);
2402 if (!Prev->isCall() || !Prev->getOperand(1).isRegMask())
2406 unsigned FoundRegs = 0;
2408 auto RegMask = Prev->getOperand(1);
2411 Is64Bit ? X86::GR64_NOREX_NOSPRegClass : X86::GR32_NOREX_NOSPRegClass;
2412 // Try to find up to NumPops free registers.
2413 for (auto Candidate : RegClass) {
2415 // Poor man's liveness:
2416 // Since we're immediately after a call, any register that is clobbered
2417 // by the call and not defined by it can be considered dead.
2418 if (!RegMask.clobbersPhysReg(Candidate))
2422 for (const MachineOperand &MO : Prev->implicit_operands()) {
2423 if (MO.isReg() && MO.isDef() && MO.getReg() == Candidate) {
2432 Regs[FoundRegs++] = Candidate;
2433 if (FoundRegs == (unsigned)NumPops)
2440 // If we found only one free register, but need two, reuse the same one twice.
2441 while (FoundRegs < (unsigned)NumPops)
2442 Regs[FoundRegs++] = Regs[0];
2444 for (int i = 0; i < NumPops; ++i)
2445 BuildMI(MBB, MBBI, DL,
2446 TII.get(STI.is64Bit() ? X86::POP64r : X86::POP32r), Regs[i]);
2451 void X86FrameLowering::
2452 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
2453 MachineBasicBlock::iterator I) const {
2454 bool reserveCallFrame = hasReservedCallFrame(MF);
2455 unsigned Opcode = I->getOpcode();
2456 bool isDestroy = Opcode == TII.getCallFrameDestroyOpcode();
2457 DebugLoc DL = I->getDebugLoc();
2458 uint64_t Amount = !reserveCallFrame ? I->getOperand(0).getImm() : 0;
2459 uint64_t InternalAmt = (isDestroy || Amount) ? I->getOperand(1).getImm() : 0;
2462 if (!reserveCallFrame) {
2463 // If the stack pointer can be changed after prologue, turn the
2464 // adjcallstackup instruction into a 'sub ESP, <amt>' and the
2465 // adjcallstackdown instruction into 'add ESP, <amt>'
2467 // We need to keep the stack aligned properly. To do this, we round the
2468 // amount of space needed for the outgoing arguments up to the next
2469 // alignment boundary.
2470 unsigned StackAlign = getStackAlignment();
2471 Amount = RoundUpToAlignment(Amount, StackAlign);
2473 MachineModuleInfo &MMI = MF.getMMI();
2474 const Function *Fn = MF.getFunction();
2475 bool WindowsCFI = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
2476 bool DwarfCFI = !WindowsCFI &&
2477 (MMI.hasDebugInfo() || Fn->needsUnwindTableEntry());
2479 // If we have any exception handlers in this function, and we adjust
2480 // the SP before calls, we may need to indicate this to the unwinder
2481 // using GNU_ARGS_SIZE. Note that this may be necessary even when
2482 // Amount == 0, because the preceding function may have set a non-0
2484 // TODO: We don't need to reset this between subsequent functions,
2485 // if it didn't change.
2486 bool HasDwarfEHHandlers = !WindowsCFI &&
2487 !MF.getMMI().getLandingPads().empty();
2489 if (HasDwarfEHHandlers && !isDestroy &&
2490 MF.getInfo<X86MachineFunctionInfo>()->getHasPushSequences())
2491 BuildCFI(MBB, I, DL,
2492 MCCFIInstruction::createGnuArgsSize(nullptr, Amount));
2497 // Factor out the amount that gets handled inside the sequence
2498 // (Pushes of argument for frame setup, callee pops for frame destroy)
2499 Amount -= InternalAmt;
2501 // If this is a callee-pop calling convention, and we're emitting precise
2502 // SP-based CFI, emit a CFA adjust for the amount the callee popped.
2503 if (isDestroy && InternalAmt && DwarfCFI && !hasFP(MF) &&
2504 MMI.usePreciseUnwindInfo())
2505 BuildCFI(MBB, I, DL,
2506 MCCFIInstruction::createAdjustCfaOffset(nullptr, -InternalAmt));
2509 // Add Amount to SP to destroy a frame, and subtract to setup.
2510 int Offset = isDestroy ? Amount : -Amount;
2512 if (!(Fn->optForMinSize() &&
2513 adjustStackWithPops(MBB, I, DL, Offset)))
2514 BuildStackAdjustment(MBB, I, DL, Offset, /*InEpilogue=*/false);
2517 if (DwarfCFI && !hasFP(MF)) {
2518 // If we don't have FP, but need to generate unwind information,
2519 // we need to set the correct CFA offset after the stack adjustment.
2520 // How much we adjust the CFA offset depends on whether we're emitting
2521 // CFI only for EH purposes or for debugging. EH only requires the CFA
2522 // offset to be correct at each call site, while for debugging we want
2523 // it to be more precise.
2524 int CFAOffset = Amount;
2525 if (!MMI.usePreciseUnwindInfo())
2526 CFAOffset += InternalAmt;
2527 CFAOffset = isDestroy ? -CFAOffset : CFAOffset;
2528 BuildCFI(MBB, I, DL,
2529 MCCFIInstruction::createAdjustCfaOffset(nullptr, CFAOffset));
2535 if (isDestroy && InternalAmt) {
2536 // If we are performing frame pointer elimination and if the callee pops
2537 // something off the stack pointer, add it back. We do this until we have
2538 // more advanced stack pointer tracking ability.
2539 // We are not tracking the stack pointer adjustment by the callee, so make
2540 // sure we restore the stack pointer immediately after the call, there may
2541 // be spill code inserted between the CALL and ADJCALLSTACKUP instructions.
2542 MachineBasicBlock::iterator B = MBB.begin();
2543 while (I != B && !std::prev(I)->isCall())
2545 BuildStackAdjustment(MBB, I, DL, -InternalAmt, /*InEpilogue=*/false);
2549 bool X86FrameLowering::canUseAsEpilogue(const MachineBasicBlock &MBB) const {
2550 assert(MBB.getParent() && "Block is not attached to a function!");
2552 // Win64 has strict requirements in terms of epilogue and we are
2553 // not taking a chance at messing with them.
2554 // I.e., unless this block is already an exit block, we can't use
2555 // it as an epilogue.
2556 if (MBB.getParent()->getSubtarget<X86Subtarget>().isTargetWin64() &&
2557 !MBB.succ_empty() && !MBB.isReturnBlock())
2560 if (canUseLEAForSPInEpilogue(*MBB.getParent()))
2563 // If we cannot use LEA to adjust SP, we may need to use ADD, which
2564 // clobbers the EFLAGS. Check that none of the terminators reads the
2565 // EFLAGS, and if one uses it, conservatively assume this is not
2566 // safe to insert the epilogue here.
2567 return !terminatorsNeedFlagsAsInput(MBB);
2570 MachineBasicBlock::iterator X86FrameLowering::restoreWin32EHStackPointers(
2571 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
2572 DebugLoc DL, bool RestoreSP) const {
2573 assert(STI.isTargetWindowsMSVC() && "funclets only supported in MSVC env");
2574 assert(STI.isTargetWin32() && "EBP/ESI restoration only required on win32");
2575 assert(STI.is32Bit() && !Uses64BitFramePtr &&
2576 "restoring EBP/ESI on non-32-bit target");
2578 MachineFunction &MF = *MBB.getParent();
2579 unsigned FramePtr = TRI->getFrameRegister(MF);
2580 unsigned BasePtr = TRI->getBaseRegister();
2581 MachineModuleInfo &MMI = MF.getMMI();
2582 const Function *Fn = MF.getFunction();
2583 WinEHFuncInfo &FuncInfo = MMI.getWinEHFuncInfo(Fn);
2584 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
2585 MachineFrameInfo *MFI = MF.getFrameInfo();
2587 // FIXME: Don't set FrameSetup flag in catchret case.
2589 int FI = FuncInfo.EHRegNodeFrameIndex;
2590 int EHRegSize = MFI->getObjectSize(FI);
2593 // MOV32rm -EHRegSize(%ebp), %esp
2594 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32rm), X86::ESP),
2595 X86::EBP, true, -EHRegSize)
2596 .setMIFlag(MachineInstr::FrameSetup);
2600 int EHRegOffset = getFrameIndexReference(MF, FI, UsedReg);
2601 int EndOffset = -EHRegOffset - EHRegSize;
2602 FuncInfo.EHRegNodeEndOffset = EndOffset;
2604 if (UsedReg == FramePtr) {
2605 // ADD $offset, %ebp
2606 unsigned ADDri = getADDriOpcode(false, EndOffset);
2607 BuildMI(MBB, MBBI, DL, TII.get(ADDri), FramePtr)
2610 .setMIFlag(MachineInstr::FrameSetup)
2613 assert(EndOffset >= 0 &&
2614 "end of registration object above normal EBP position!");
2615 } else if (UsedReg == BasePtr) {
2616 // LEA offset(%ebp), %esi
2617 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::LEA32r), BasePtr),
2618 FramePtr, false, EndOffset)
2619 .setMIFlag(MachineInstr::FrameSetup);
2620 // MOV32rm SavedEBPOffset(%esi), %ebp
2621 assert(X86FI->getHasSEHFramePtrSave());
2623 getFrameIndexReference(MF, X86FI->getSEHFramePtrSaveIndex(), UsedReg);
2624 assert(UsedReg == BasePtr);
2625 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32rm), FramePtr),
2626 UsedReg, true, Offset)
2627 .setMIFlag(MachineInstr::FrameSetup);
2629 llvm_unreachable("32-bit frames with WinEH must use FramePtr or BasePtr");
2634 unsigned X86FrameLowering::getWinEHParentFrameOffset(const MachineFunction &MF) const {
2635 // RDX, the parent frame pointer, is homed into 16(%rsp) in the prologue.
2636 unsigned Offset = 16;
2637 // RBP is immediately pushed.
2639 // All callee-saved registers are then pushed.
2640 Offset += MF.getInfo<X86MachineFunctionInfo>()->getCalleeSavedFrameSize();
2641 // Every funclet allocates enough stack space for the largest outgoing call.
2642 Offset += getWinEHFuncletFrameSize(MF);