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/EHPersonalities.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 X86RegisterInfo *TRI,
151 const MachineFunction *MF = MBB.getParent();
152 const Function *F = MF->getFunction();
153 if (!F || MF->getMMI().callsEHReturn())
156 const TargetRegisterClass &AvailableRegs = *TRI->getGPRsForTailCall(*MF);
158 unsigned Opc = MBBI->getOpcode();
165 case X86::TCRETURNdi:
166 case X86::TCRETURNri:
167 case X86::TCRETURNmi:
168 case X86::TCRETURNdi64:
169 case X86::TCRETURNri64:
170 case X86::TCRETURNmi64:
172 case X86::EH_RETURN64: {
173 SmallSet<uint16_t, 8> Uses;
174 for (unsigned i = 0, e = MBBI->getNumOperands(); i != e; ++i) {
175 MachineOperand &MO = MBBI->getOperand(i);
176 if (!MO.isReg() || MO.isDef())
178 unsigned Reg = MO.getReg();
181 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
185 for (auto CS : AvailableRegs)
186 if (!Uses.count(CS) && CS != X86::RIP)
194 static bool isEAXLiveIn(MachineFunction &MF) {
195 for (MachineRegisterInfo::livein_iterator II = MF.getRegInfo().livein_begin(),
196 EE = MF.getRegInfo().livein_end(); II != EE; ++II) {
197 unsigned Reg = II->first;
199 if (Reg == X86::RAX || Reg == X86::EAX || Reg == X86::AX ||
200 Reg == X86::AH || Reg == X86::AL)
207 /// Check if the flags need to be preserved before the terminators.
208 /// This would be the case, if the eflags is live-in of the region
209 /// composed by the terminators or live-out of that region, without
210 /// being defined by a terminator.
212 flagsNeedToBePreservedBeforeTheTerminators(const MachineBasicBlock &MBB) {
213 for (const MachineInstr &MI : MBB.terminators()) {
214 bool BreakNext = false;
215 for (const MachineOperand &MO : MI.operands()) {
218 unsigned Reg = MO.getReg();
219 if (Reg != X86::EFLAGS)
222 // This terminator needs an eflags that is not defined
223 // by a previous another terminator:
224 // EFLAGS is live-in of the region composed by the terminators.
227 // This terminator defines the eflags, i.e., we don't need to preserve it.
228 // However, we still need to check this specific terminator does not
229 // read a live-in value.
232 // We found a definition of the eflags, no need to preserve them.
237 // None of the terminators use or define the eflags.
238 // Check if they are live-out, that would imply we need to preserve them.
239 for (const MachineBasicBlock *Succ : MBB.successors())
240 if (Succ->isLiveIn(X86::EFLAGS))
246 /// emitSPUpdate - Emit a series of instructions to increment / decrement the
247 /// stack pointer by a constant value.
248 void X86FrameLowering::emitSPUpdate(MachineBasicBlock &MBB,
249 MachineBasicBlock::iterator &MBBI,
250 int64_t NumBytes, bool InEpilogue) const {
251 bool isSub = NumBytes < 0;
252 uint64_t Offset = isSub ? -NumBytes : NumBytes;
254 uint64_t Chunk = (1LL << 31) - 1;
255 DebugLoc DL = MBB.findDebugLoc(MBBI);
258 if (Offset > Chunk) {
259 // Rather than emit a long series of instructions for large offsets,
260 // load the offset into a register and do one sub/add
263 if (isSub && !isEAXLiveIn(*MBB.getParent()))
264 Reg = (unsigned)(Is64Bit ? X86::RAX : X86::EAX);
266 Reg = findDeadCallerSavedReg(MBB, MBBI, TRI, Is64Bit);
269 unsigned Opc = Is64Bit ? X86::MOV64ri : X86::MOV32ri;
270 BuildMI(MBB, MBBI, DL, TII.get(Opc), Reg)
273 ? getSUBrrOpcode(Is64Bit)
274 : getADDrrOpcode(Is64Bit);
275 MachineInstr *MI = BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr)
278 MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
284 uint64_t ThisVal = std::min(Offset, Chunk);
285 if (ThisVal == (Is64Bit ? 8 : 4)) {
286 // Use push / pop instead.
288 ? (unsigned)(Is64Bit ? X86::RAX : X86::EAX)
289 : findDeadCallerSavedReg(MBB, MBBI, TRI, Is64Bit);
292 ? (Is64Bit ? X86::PUSH64r : X86::PUSH32r)
293 : (Is64Bit ? X86::POP64r : X86::POP32r);
294 MachineInstr *MI = BuildMI(MBB, MBBI, DL, TII.get(Opc))
295 .addReg(Reg, getDefRegState(!isSub) | getUndefRegState(isSub));
297 MI->setFlag(MachineInstr::FrameSetup);
299 MI->setFlag(MachineInstr::FrameDestroy);
305 MachineInstrBuilder MI = BuildStackAdjustment(
306 MBB, MBBI, DL, isSub ? -ThisVal : ThisVal, InEpilogue);
308 MI.setMIFlag(MachineInstr::FrameSetup);
310 MI.setMIFlag(MachineInstr::FrameDestroy);
316 MachineInstrBuilder X86FrameLowering::BuildStackAdjustment(
317 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, DebugLoc DL,
318 int64_t Offset, bool InEpilogue) const {
319 assert(Offset != 0 && "zero offset stack adjustment requested");
321 // On Atom, using LEA to adjust SP is preferred, but using it in the epilogue
325 // Check if inserting the prologue at the beginning
326 // of MBB would require to use LEA operations.
327 // We need to use LEA operations if EFLAGS is live in, because
328 // it means an instruction will read it before it gets defined.
329 UseLEA = STI.useLeaForSP() || MBB.isLiveIn(X86::EFLAGS);
331 // If we can use LEA for SP but we shouldn't, check that none
332 // of the terminators uses the eflags. Otherwise we will insert
333 // a ADD that will redefine the eflags and break the condition.
334 // Alternatively, we could move the ADD, but this may not be possible
335 // and is an optimization anyway.
336 UseLEA = canUseLEAForSPInEpilogue(*MBB.getParent());
337 if (UseLEA && !STI.useLeaForSP())
338 UseLEA = flagsNeedToBePreservedBeforeTheTerminators(MBB);
339 // If that assert breaks, that means we do not do the right thing
340 // in canUseAsEpilogue.
341 assert((UseLEA || !flagsNeedToBePreservedBeforeTheTerminators(MBB)) &&
342 "We shouldn't have allowed this insertion point");
345 MachineInstrBuilder MI;
347 MI = addRegOffset(BuildMI(MBB, MBBI, DL,
348 TII.get(getLEArOpcode(Uses64BitFramePtr)),
350 StackPtr, false, Offset);
352 bool IsSub = Offset < 0;
353 uint64_t AbsOffset = IsSub ? -Offset : Offset;
354 unsigned Opc = IsSub ? getSUBriOpcode(Uses64BitFramePtr, AbsOffset)
355 : getADDriOpcode(Uses64BitFramePtr, AbsOffset);
356 MI = BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr)
359 MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
364 int X86FrameLowering::mergeSPUpdates(MachineBasicBlock &MBB,
365 MachineBasicBlock::iterator &MBBI,
366 bool doMergeWithPrevious) const {
367 if ((doMergeWithPrevious && MBBI == MBB.begin()) ||
368 (!doMergeWithPrevious && MBBI == MBB.end()))
371 MachineBasicBlock::iterator PI = doMergeWithPrevious ? std::prev(MBBI) : MBBI;
372 MachineBasicBlock::iterator NI = doMergeWithPrevious ? nullptr
374 unsigned Opc = PI->getOpcode();
377 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
378 Opc == X86::ADD32ri || Opc == X86::ADD32ri8 ||
379 Opc == X86::LEA32r || Opc == X86::LEA64_32r) &&
380 PI->getOperand(0).getReg() == StackPtr){
381 Offset += PI->getOperand(2).getImm();
383 if (!doMergeWithPrevious) MBBI = NI;
384 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
385 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
386 PI->getOperand(0).getReg() == StackPtr) {
387 Offset -= PI->getOperand(2).getImm();
389 if (!doMergeWithPrevious) MBBI = NI;
395 void X86FrameLowering::BuildCFI(MachineBasicBlock &MBB,
396 MachineBasicBlock::iterator MBBI, DebugLoc DL,
397 MCCFIInstruction CFIInst) const {
398 MachineFunction &MF = *MBB.getParent();
399 unsigned CFIIndex = MF.getMMI().addFrameInst(CFIInst);
400 BuildMI(MBB, MBBI, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
401 .addCFIIndex(CFIIndex);
405 X86FrameLowering::emitCalleeSavedFrameMoves(MachineBasicBlock &MBB,
406 MachineBasicBlock::iterator MBBI,
408 MachineFunction &MF = *MBB.getParent();
409 MachineFrameInfo *MFI = MF.getFrameInfo();
410 MachineModuleInfo &MMI = MF.getMMI();
411 const MCRegisterInfo *MRI = MMI.getContext().getRegisterInfo();
413 // Add callee saved registers to move list.
414 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
415 if (CSI.empty()) return;
417 // Calculate offsets.
418 for (std::vector<CalleeSavedInfo>::const_iterator
419 I = CSI.begin(), E = CSI.end(); I != E; ++I) {
420 int64_t Offset = MFI->getObjectOffset(I->getFrameIdx());
421 unsigned Reg = I->getReg();
423 unsigned DwarfReg = MRI->getDwarfRegNum(Reg, true);
424 BuildCFI(MBB, MBBI, DL,
425 MCCFIInstruction::createOffset(nullptr, DwarfReg, Offset));
429 MachineInstr *X86FrameLowering::emitStackProbe(MachineFunction &MF,
430 MachineBasicBlock &MBB,
431 MachineBasicBlock::iterator MBBI,
433 bool InProlog) const {
434 const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
435 if (STI.isTargetWindowsCoreCLR()) {
437 return emitStackProbeInlineStub(MF, MBB, MBBI, DL, true);
439 return emitStackProbeInline(MF, MBB, MBBI, DL, false);
442 return emitStackProbeCall(MF, MBB, MBBI, DL, InProlog);
446 void X86FrameLowering::inlineStackProbe(MachineFunction &MF,
447 MachineBasicBlock &PrologMBB) const {
448 const StringRef ChkStkStubSymbol = "__chkstk_stub";
449 MachineInstr *ChkStkStub = nullptr;
451 for (MachineInstr &MI : PrologMBB) {
452 if (MI.isCall() && MI.getOperand(0).isSymbol() &&
453 ChkStkStubSymbol == MI.getOperand(0).getSymbolName()) {
459 if (ChkStkStub != nullptr) {
460 MachineBasicBlock::iterator MBBI = std::next(ChkStkStub->getIterator());
461 assert(std::prev(MBBI).operator==(ChkStkStub) &&
462 "MBBI expected after __chkstk_stub.");
463 DebugLoc DL = PrologMBB.findDebugLoc(MBBI);
464 emitStackProbeInline(MF, PrologMBB, MBBI, DL, true);
465 ChkStkStub->eraseFromParent();
469 MachineInstr *X86FrameLowering::emitStackProbeInline(
470 MachineFunction &MF, MachineBasicBlock &MBB,
471 MachineBasicBlock::iterator MBBI, DebugLoc DL, bool InProlog) const {
472 const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
473 assert(STI.is64Bit() && "different expansion needed for 32 bit");
474 assert(STI.isTargetWindowsCoreCLR() && "custom expansion expects CoreCLR");
475 const TargetInstrInfo &TII = *STI.getInstrInfo();
476 const BasicBlock *LLVM_BB = MBB.getBasicBlock();
478 // RAX contains the number of bytes of desired stack adjustment.
479 // The handling here assumes this value has already been updated so as to
480 // maintain stack alignment.
482 // We need to exit with RSP modified by this amount and execute suitable
483 // page touches to notify the OS that we're growing the stack responsibly.
484 // All stack probing must be done without modifying RSP.
490 // Flags, TestReg = CopyReg - SizeReg
491 // FinalReg = !Flags.Ovf ? TestReg : ZeroReg
492 // LimitReg = gs magic thread env access
493 // if FinalReg >= LimitReg goto ContinueMBB
495 // RoundReg = page address of FinalReg
497 // LoopReg = PHI(LimitReg,ProbeReg)
498 // ProbeReg = LoopReg - PageSize
500 // if (ProbeReg > RoundReg) goto LoopMBB
503 // [rest of original MBB]
505 // Set up the new basic blocks
506 MachineBasicBlock *RoundMBB = MF.CreateMachineBasicBlock(LLVM_BB);
507 MachineBasicBlock *LoopMBB = MF.CreateMachineBasicBlock(LLVM_BB);
508 MachineBasicBlock *ContinueMBB = MF.CreateMachineBasicBlock(LLVM_BB);
510 MachineFunction::iterator MBBIter = std::next(MBB.getIterator());
511 MF.insert(MBBIter, RoundMBB);
512 MF.insert(MBBIter, LoopMBB);
513 MF.insert(MBBIter, ContinueMBB);
515 // Split MBB and move the tail portion down to ContinueMBB.
516 MachineBasicBlock::iterator BeforeMBBI = std::prev(MBBI);
517 ContinueMBB->splice(ContinueMBB->begin(), &MBB, MBBI, MBB.end());
518 ContinueMBB->transferSuccessorsAndUpdatePHIs(&MBB);
520 // Some useful constants
521 const int64_t ThreadEnvironmentStackLimit = 0x10;
522 const int64_t PageSize = 0x1000;
523 const int64_t PageMask = ~(PageSize - 1);
525 // Registers we need. For the normal case we use virtual
526 // registers. For the prolog expansion we use RAX, RCX and RDX.
527 MachineRegisterInfo &MRI = MF.getRegInfo();
528 const TargetRegisterClass *RegClass = &X86::GR64RegClass;
529 const unsigned SizeReg = InProlog ? (unsigned)X86::RAX
530 : MRI.createVirtualRegister(RegClass),
531 ZeroReg = InProlog ? (unsigned)X86::RCX
532 : MRI.createVirtualRegister(RegClass),
533 CopyReg = InProlog ? (unsigned)X86::RDX
534 : MRI.createVirtualRegister(RegClass),
535 TestReg = InProlog ? (unsigned)X86::RDX
536 : MRI.createVirtualRegister(RegClass),
537 FinalReg = InProlog ? (unsigned)X86::RDX
538 : MRI.createVirtualRegister(RegClass),
539 RoundedReg = InProlog ? (unsigned)X86::RDX
540 : MRI.createVirtualRegister(RegClass),
541 LimitReg = InProlog ? (unsigned)X86::RCX
542 : MRI.createVirtualRegister(RegClass),
543 JoinReg = InProlog ? (unsigned)X86::RCX
544 : MRI.createVirtualRegister(RegClass),
545 ProbeReg = InProlog ? (unsigned)X86::RCX
546 : MRI.createVirtualRegister(RegClass);
548 // SP-relative offsets where we can save RCX and RDX.
549 int64_t RCXShadowSlot = 0;
550 int64_t RDXShadowSlot = 0;
552 // If inlining in the prolog, save RCX and RDX.
553 // Future optimization: don't save or restore if not live in.
555 // Compute the offsets. We need to account for things already
556 // pushed onto the stack at this point: return address, frame
557 // pointer (if used), and callee saves.
558 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
559 const int64_t CalleeSaveSize = X86FI->getCalleeSavedFrameSize();
560 const bool HasFP = hasFP(MF);
561 RCXShadowSlot = 8 + CalleeSaveSize + (HasFP ? 8 : 0);
562 RDXShadowSlot = RCXShadowSlot + 8;
564 addRegOffset(BuildMI(&MBB, DL, TII.get(X86::MOV64mr)), X86::RSP, false,
567 addRegOffset(BuildMI(&MBB, DL, TII.get(X86::MOV64mr)), X86::RSP, false,
571 // Not in the prolog. Copy RAX to a virtual reg.
572 BuildMI(&MBB, DL, TII.get(X86::MOV64rr), SizeReg).addReg(X86::RAX);
575 // Add code to MBB to check for overflow and set the new target stack pointer
577 BuildMI(&MBB, DL, TII.get(X86::XOR64rr), ZeroReg)
578 .addReg(ZeroReg, RegState::Undef)
579 .addReg(ZeroReg, RegState::Undef);
580 BuildMI(&MBB, DL, TII.get(X86::MOV64rr), CopyReg).addReg(X86::RSP);
581 BuildMI(&MBB, DL, TII.get(X86::SUB64rr), TestReg)
584 BuildMI(&MBB, DL, TII.get(X86::CMOVB64rr), FinalReg)
588 // FinalReg now holds final stack pointer value, or zero if
589 // allocation would overflow. Compare against the current stack
590 // limit from the thread environment block. Note this limit is the
591 // lowest touched page on the stack, not the point at which the OS
592 // will cause an overflow exception, so this is just an optimization
593 // to avoid unnecessarily touching pages that are below the current
594 // SP but already commited to the stack by the OS.
595 BuildMI(&MBB, DL, TII.get(X86::MOV64rm), LimitReg)
599 .addImm(ThreadEnvironmentStackLimit)
601 BuildMI(&MBB, DL, TII.get(X86::CMP64rr)).addReg(FinalReg).addReg(LimitReg);
602 // Jump if the desired stack pointer is at or above the stack limit.
603 BuildMI(&MBB, DL, TII.get(X86::JAE_1)).addMBB(ContinueMBB);
605 // Add code to roundMBB to round the final stack pointer to a page boundary.
606 BuildMI(RoundMBB, DL, TII.get(X86::AND64ri32), RoundedReg)
609 BuildMI(RoundMBB, DL, TII.get(X86::JMP_1)).addMBB(LoopMBB);
611 // LimitReg now holds the current stack limit, RoundedReg page-rounded
612 // final RSP value. Add code to loopMBB to decrement LimitReg page-by-page
613 // and probe until we reach RoundedReg.
615 BuildMI(LoopMBB, DL, TII.get(X86::PHI), JoinReg)
622 addRegOffset(BuildMI(LoopMBB, DL, TII.get(X86::LEA64r), ProbeReg), JoinReg,
625 // Probe by storing a byte onto the stack.
626 BuildMI(LoopMBB, DL, TII.get(X86::MOV8mi))
633 BuildMI(LoopMBB, DL, TII.get(X86::CMP64rr))
636 BuildMI(LoopMBB, DL, TII.get(X86::JNE_1)).addMBB(LoopMBB);
638 MachineBasicBlock::iterator ContinueMBBI = ContinueMBB->getFirstNonPHI();
640 // If in prolog, restore RDX and RCX.
642 addRegOffset(BuildMI(*ContinueMBB, ContinueMBBI, DL, TII.get(X86::MOV64rm),
644 X86::RSP, false, RCXShadowSlot);
645 addRegOffset(BuildMI(*ContinueMBB, ContinueMBBI, DL, TII.get(X86::MOV64rm),
647 X86::RSP, false, RDXShadowSlot);
650 // Now that the probing is done, add code to continueMBB to update
651 // the stack pointer for real.
652 BuildMI(*ContinueMBB, ContinueMBBI, DL, TII.get(X86::SUB64rr), X86::RSP)
656 // Add the control flow edges we need.
657 MBB.addSuccessor(ContinueMBB);
658 MBB.addSuccessor(RoundMBB);
659 RoundMBB->addSuccessor(LoopMBB);
660 LoopMBB->addSuccessor(ContinueMBB);
661 LoopMBB->addSuccessor(LoopMBB);
663 // Mark all the instructions added to the prolog as frame setup.
665 for (++BeforeMBBI; BeforeMBBI != MBB.end(); ++BeforeMBBI) {
666 BeforeMBBI->setFlag(MachineInstr::FrameSetup);
668 for (MachineInstr &MI : *RoundMBB) {
669 MI.setFlag(MachineInstr::FrameSetup);
671 for (MachineInstr &MI : *LoopMBB) {
672 MI.setFlag(MachineInstr::FrameSetup);
674 for (MachineBasicBlock::iterator CMBBI = ContinueMBB->begin();
675 CMBBI != ContinueMBBI; ++CMBBI) {
676 CMBBI->setFlag(MachineInstr::FrameSetup);
680 // Possible TODO: physreg liveness for InProlog case.
685 MachineInstr *X86FrameLowering::emitStackProbeCall(
686 MachineFunction &MF, MachineBasicBlock &MBB,
687 MachineBasicBlock::iterator MBBI, DebugLoc DL, bool InProlog) const {
688 bool IsLargeCodeModel = MF.getTarget().getCodeModel() == CodeModel::Large;
692 CallOp = IsLargeCodeModel ? X86::CALL64r : X86::CALL64pcrel32;
694 CallOp = X86::CALLpcrel32;
698 if (STI.isTargetCygMing()) {
699 Symbol = "___chkstk_ms";
703 } else if (STI.isTargetCygMing())
708 MachineInstrBuilder CI;
709 MachineBasicBlock::iterator ExpansionMBBI = std::prev(MBBI);
711 // All current stack probes take AX and SP as input, clobber flags, and
712 // preserve all registers. x86_64 probes leave RSP unmodified.
713 if (Is64Bit && MF.getTarget().getCodeModel() == CodeModel::Large) {
714 // For the large code model, we have to call through a register. Use R11,
715 // as it is scratch in all supported calling conventions.
716 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64ri), X86::R11)
717 .addExternalSymbol(Symbol);
718 CI = BuildMI(MBB, MBBI, DL, TII.get(CallOp)).addReg(X86::R11);
720 CI = BuildMI(MBB, MBBI, DL, TII.get(CallOp)).addExternalSymbol(Symbol);
723 unsigned AX = Is64Bit ? X86::RAX : X86::EAX;
724 unsigned SP = Is64Bit ? X86::RSP : X86::ESP;
725 CI.addReg(AX, RegState::Implicit)
726 .addReg(SP, RegState::Implicit)
727 .addReg(AX, RegState::Define | RegState::Implicit)
728 .addReg(SP, RegState::Define | RegState::Implicit)
729 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
732 // MSVC x64's __chkstk and cygwin/mingw's ___chkstk_ms do not adjust %rsp
733 // themselves. It also does not clobber %rax so we can reuse it when
735 BuildMI(MBB, MBBI, DL, TII.get(X86::SUB64rr), X86::RSP)
741 // Apply the frame setup flag to all inserted instrs.
742 for (++ExpansionMBBI; ExpansionMBBI != MBBI; ++ExpansionMBBI)
743 ExpansionMBBI->setFlag(MachineInstr::FrameSetup);
749 MachineInstr *X86FrameLowering::emitStackProbeInlineStub(
750 MachineFunction &MF, MachineBasicBlock &MBB,
751 MachineBasicBlock::iterator MBBI, DebugLoc DL, bool InProlog) const {
753 assert(InProlog && "ChkStkStub called outside prolog!");
755 BuildMI(MBB, MBBI, DL, TII.get(X86::CALLpcrel32))
756 .addExternalSymbol("__chkstk_stub");
761 static unsigned calculateSetFPREG(uint64_t SPAdjust) {
762 // Win64 ABI has a less restrictive limitation of 240; 128 works equally well
763 // and might require smaller successive adjustments.
764 const uint64_t Win64MaxSEHOffset = 128;
765 uint64_t SEHFrameOffset = std::min(SPAdjust, Win64MaxSEHOffset);
766 // Win64 ABI requires 16-byte alignment for the UWOP_SET_FPREG opcode.
767 return SEHFrameOffset & -16;
770 // If we're forcing a stack realignment we can't rely on just the frame
771 // info, we need to know the ABI stack alignment as well in case we
772 // have a call out. Otherwise just make sure we have some alignment - we'll
773 // go with the minimum SlotSize.
774 uint64_t X86FrameLowering::calculateMaxStackAlign(const MachineFunction &MF) const {
775 const MachineFrameInfo *MFI = MF.getFrameInfo();
776 uint64_t MaxAlign = MFI->getMaxAlignment(); // Desired stack alignment.
777 unsigned StackAlign = getStackAlignment();
778 if (MF.getFunction()->hasFnAttribute("stackrealign")) {
780 MaxAlign = (StackAlign > MaxAlign) ? StackAlign : MaxAlign;
781 else if (MaxAlign < SlotSize)
787 void X86FrameLowering::BuildStackAlignAND(MachineBasicBlock &MBB,
788 MachineBasicBlock::iterator MBBI,
789 DebugLoc DL, unsigned Reg,
790 uint64_t MaxAlign) const {
791 uint64_t Val = -MaxAlign;
792 unsigned AndOp = getANDriOpcode(Uses64BitFramePtr, Val);
793 MachineInstr *MI = BuildMI(MBB, MBBI, DL, TII.get(AndOp), Reg)
796 .setMIFlag(MachineInstr::FrameSetup);
798 // The EFLAGS implicit def is dead.
799 MI->getOperand(3).setIsDead();
802 /// emitPrologue - Push callee-saved registers onto the stack, which
803 /// automatically adjust the stack pointer. Adjust the stack pointer to allocate
804 /// space for local variables. Also emit labels used by the exception handler to
805 /// generate the exception handling frames.
808 Here's a gist of what gets emitted:
810 ; Establish frame pointer, if needed
813 .cfi_def_cfa_offset 16
814 .cfi_offset %rbp, -16
817 .cfi_def_cfa_register %rbp
819 ; Spill general-purpose registers
820 [for all callee-saved GPRs]
823 .cfi_def_cfa_offset (offset from RETADDR)
826 ; If the required stack alignment > default stack alignment
827 ; rsp needs to be re-aligned. This creates a "re-alignment gap"
828 ; of unknown size in the stack frame.
829 [if stack needs re-alignment]
832 ; Allocate space for locals
833 [if target is Windows and allocated space > 4096 bytes]
834 ; Windows needs special care for allocations larger
837 call ___chkstk_ms/___chkstk
843 .seh_stackalloc (size of XMM spill slots)
844 .seh_setframe %rbp, SEHFrameOffset ; = size of all spill slots
849 ; Note, that while only Windows 64 ABI specifies XMMs as callee-preserved,
850 ; they may get spilled on any platform, if the current function
851 ; calls @llvm.eh.unwind.init
853 [for all callee-saved XMM registers]
854 movaps %<xmm reg>, -MMM(%rbp)
855 [for all callee-saved XMM registers]
856 .seh_savexmm %<xmm reg>, (-MMM + SEHFrameOffset)
857 ; i.e. the offset relative to (%rbp - SEHFrameOffset)
859 [for all callee-saved XMM registers]
860 movaps %<xmm reg>, KKK(%rsp)
861 [for all callee-saved XMM registers]
862 .seh_savexmm %<xmm reg>, KKK
866 [if needs base pointer]
868 [if needs to restore base pointer]
873 [for all callee-saved registers]
874 .cfi_offset %<reg>, (offset from %rbp)
876 .cfi_def_cfa_offset (offset from RETADDR)
877 [for all callee-saved registers]
878 .cfi_offset %<reg>, (offset from %rsp)
881 - .seh directives are emitted only for Windows 64 ABI
882 - .cfi directives are emitted for all other ABIs
883 - for 32-bit code, substitute %e?? registers for %r??
886 void X86FrameLowering::emitPrologue(MachineFunction &MF,
887 MachineBasicBlock &MBB) const {
888 assert(&STI == &MF.getSubtarget<X86Subtarget>() &&
889 "MF used frame lowering for wrong subtarget");
890 MachineBasicBlock::iterator MBBI = MBB.begin();
891 MachineFrameInfo *MFI = MF.getFrameInfo();
892 const Function *Fn = MF.getFunction();
893 MachineModuleInfo &MMI = MF.getMMI();
894 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
895 uint64_t MaxAlign = calculateMaxStackAlign(MF); // Desired stack alignment.
896 uint64_t StackSize = MFI->getStackSize(); // Number of bytes to allocate.
897 bool IsFunclet = MBB.isEHFuncletEntry();
898 EHPersonality Personality = EHPersonality::Unknown;
899 if (Fn->hasPersonalityFn())
900 Personality = classifyEHPersonality(Fn->getPersonalityFn());
901 bool FnHasClrFunclet =
902 MMI.hasEHFunclets() && Personality == EHPersonality::CoreCLR;
903 bool IsClrFunclet = IsFunclet && FnHasClrFunclet;
904 bool HasFP = hasFP(MF);
905 bool IsWin64CC = STI.isCallingConvWin64(Fn->getCallingConv());
906 bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
907 bool NeedsWinCFI = IsWin64Prologue && Fn->needsUnwindTableEntry();
909 !IsWin64Prologue && (MMI.hasDebugInfo() || Fn->needsUnwindTableEntry());
910 unsigned FramePtr = TRI->getFrameRegister(MF);
911 const unsigned MachineFramePtr =
912 STI.isTarget64BitILP32()
913 ? getX86SubSuperRegister(FramePtr, 64) : FramePtr;
914 unsigned BasePtr = TRI->getBaseRegister();
916 // Debug location must be unknown since the first debug location is used
917 // to determine the end of the prologue.
920 // Add RETADDR move area to callee saved frame size.
921 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
922 if (TailCallReturnAddrDelta && IsWin64Prologue)
923 report_fatal_error("Can't handle guaranteed tail call under win64 yet");
925 if (TailCallReturnAddrDelta < 0)
926 X86FI->setCalleeSavedFrameSize(
927 X86FI->getCalleeSavedFrameSize() - TailCallReturnAddrDelta);
929 bool UseStackProbe = (STI.isOSWindows() && !STI.isTargetMachO());
931 // The default stack probe size is 4096 if the function has no stackprobesize
933 unsigned StackProbeSize = 4096;
934 if (Fn->hasFnAttribute("stack-probe-size"))
935 Fn->getFnAttribute("stack-probe-size")
937 .getAsInteger(0, StackProbeSize);
939 // If this is x86-64 and the Red Zone is not disabled, if we are a leaf
940 // function, and use up to 128 bytes of stack space, don't have a frame
941 // pointer, calls, or dynamic alloca then we do not need to adjust the
942 // stack pointer (we fit in the Red Zone). We also check that we don't
943 // push and pop from the stack.
944 if (Is64Bit && !Fn->hasFnAttribute(Attribute::NoRedZone) &&
945 !TRI->needsStackRealignment(MF) &&
946 !MFI->hasVarSizedObjects() && // No dynamic alloca.
947 !MFI->adjustsStack() && // No calls.
948 !IsWin64CC && // Win64 has no Red Zone
949 !MFI->hasOpaqueSPAdjustment() && // Don't push and pop.
950 !MF.shouldSplitStack()) { // Regular stack
951 uint64_t MinSize = X86FI->getCalleeSavedFrameSize();
952 if (HasFP) MinSize += SlotSize;
953 StackSize = std::max(MinSize, StackSize > 128 ? StackSize - 128 : 0);
954 MFI->setStackSize(StackSize);
957 // Insert stack pointer adjustment for later moving of return addr. Only
958 // applies to tail call optimized functions where the callee argument stack
959 // size is bigger than the callers.
960 if (TailCallReturnAddrDelta < 0) {
961 BuildStackAdjustment(MBB, MBBI, DL, TailCallReturnAddrDelta,
962 /*InEpilogue=*/false)
963 .setMIFlag(MachineInstr::FrameSetup);
966 // Mapping for machine moves:
968 // DST: VirtualFP AND
969 // SRC: VirtualFP => DW_CFA_def_cfa_offset
970 // ELSE => DW_CFA_def_cfa
972 // SRC: VirtualFP AND
973 // DST: Register => DW_CFA_def_cfa_register
976 // OFFSET < 0 => DW_CFA_offset_extended_sf
977 // REG < 64 => DW_CFA_offset + Reg
978 // ELSE => DW_CFA_offset_extended
980 uint64_t NumBytes = 0;
981 int stackGrowth = -SlotSize;
983 // Find the funclet establisher parameter
984 unsigned Establisher = X86::NoRegister;
986 Establisher = Uses64BitFramePtr ? X86::RCX : X86::ECX;
988 Establisher = Uses64BitFramePtr ? X86::RDX : X86::EDX;
990 if (IsWin64Prologue && IsFunclet && !IsClrFunclet) {
991 // Immediately spill establisher into the home slot.
992 // The runtime cares about this.
993 // MOV64mr %rdx, 16(%rsp)
994 unsigned MOVmr = Uses64BitFramePtr ? X86::MOV64mr : X86::MOV32mr;
995 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(MOVmr)), StackPtr, true, 16)
997 .setMIFlag(MachineInstr::FrameSetup);
998 MBB.addLiveIn(Establisher);
1002 // Calculate required stack adjustment.
1003 uint64_t FrameSize = StackSize - SlotSize;
1004 // If required, include space for extra hidden slot for stashing base pointer.
1005 if (X86FI->getRestoreBasePointer())
1006 FrameSize += SlotSize;
1008 NumBytes = FrameSize - X86FI->getCalleeSavedFrameSize();
1010 // Callee-saved registers are pushed on stack before the stack is realigned.
1011 if (TRI->needsStackRealignment(MF) && !IsWin64Prologue)
1012 NumBytes = RoundUpToAlignment(NumBytes, MaxAlign);
1014 // Get the offset of the stack slot for the EBP register, which is
1015 // guaranteed to be the last slot by processFunctionBeforeFrameFinalized.
1016 // Update the frame offset adjustment.
1018 MFI->setOffsetAdjustment(-NumBytes);
1020 assert(MFI->getOffsetAdjustment() == -(int)NumBytes &&
1021 "should calculate same local variable offset for funclets");
1023 // Save EBP/RBP into the appropriate stack slot.
1024 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::PUSH64r : X86::PUSH32r))
1025 .addReg(MachineFramePtr, RegState::Kill)
1026 .setMIFlag(MachineInstr::FrameSetup);
1028 if (NeedsDwarfCFI) {
1029 // Mark the place where EBP/RBP was saved.
1030 // Define the current CFA rule to use the provided offset.
1032 BuildCFI(MBB, MBBI, DL,
1033 MCCFIInstruction::createDefCfaOffset(nullptr, 2 * stackGrowth));
1035 // Change the rule for the FramePtr to be an "offset" rule.
1036 unsigned DwarfFramePtr = TRI->getDwarfRegNum(MachineFramePtr, true);
1037 BuildCFI(MBB, MBBI, DL, MCCFIInstruction::createOffset(
1038 nullptr, DwarfFramePtr, 2 * stackGrowth));
1042 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_PushReg))
1044 .setMIFlag(MachineInstr::FrameSetup);
1047 if (!IsWin64Prologue && !IsFunclet) {
1048 // Update EBP with the new base value.
1049 BuildMI(MBB, MBBI, DL,
1050 TII.get(Uses64BitFramePtr ? X86::MOV64rr : X86::MOV32rr),
1053 .setMIFlag(MachineInstr::FrameSetup);
1055 if (NeedsDwarfCFI) {
1056 // Mark effective beginning of when frame pointer becomes valid.
1057 // Define the current CFA to use the EBP/RBP register.
1058 unsigned DwarfFramePtr = TRI->getDwarfRegNum(MachineFramePtr, true);
1059 BuildCFI(MBB, MBBI, DL, MCCFIInstruction::createDefCfaRegister(
1060 nullptr, DwarfFramePtr));
1064 // Mark the FramePtr as live-in in every block. Don't do this again for
1065 // funclet prologues.
1067 for (MachineBasicBlock &EveryMBB : MF)
1068 EveryMBB.addLiveIn(MachineFramePtr);
1071 assert(!IsFunclet && "funclets without FPs not yet implemented");
1072 NumBytes = StackSize - X86FI->getCalleeSavedFrameSize();
1075 // For EH funclets, only allocate enough space for outgoing calls. Save the
1076 // NumBytes value that we would've used for the parent frame.
1077 unsigned ParentFrameNumBytes = NumBytes;
1079 NumBytes = getWinEHFuncletFrameSize(MF);
1081 // Skip the callee-saved push instructions.
1082 bool PushedRegs = false;
1083 int StackOffset = 2 * stackGrowth;
1085 while (MBBI != MBB.end() &&
1086 MBBI->getFlag(MachineInstr::FrameSetup) &&
1087 (MBBI->getOpcode() == X86::PUSH32r ||
1088 MBBI->getOpcode() == X86::PUSH64r)) {
1090 unsigned Reg = MBBI->getOperand(0).getReg();
1093 if (!HasFP && NeedsDwarfCFI) {
1094 // Mark callee-saved push instruction.
1095 // Define the current CFA rule to use the provided offset.
1097 BuildCFI(MBB, MBBI, DL,
1098 MCCFIInstruction::createDefCfaOffset(nullptr, StackOffset));
1099 StackOffset += stackGrowth;
1103 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_PushReg)).addImm(Reg).setMIFlag(
1104 MachineInstr::FrameSetup);
1108 // Realign stack after we pushed callee-saved registers (so that we'll be
1109 // able to calculate their offsets from the frame pointer).
1110 // Don't do this for Win64, it needs to realign the stack after the prologue.
1111 if (!IsWin64Prologue && !IsFunclet && TRI->needsStackRealignment(MF)) {
1112 assert(HasFP && "There should be a frame pointer if stack is realigned.");
1113 BuildStackAlignAND(MBB, MBBI, DL, StackPtr, MaxAlign);
1116 // If there is an SUB32ri of ESP immediately before this instruction, merge
1117 // the two. This can be the case when tail call elimination is enabled and
1118 // the callee has more arguments then the caller.
1119 NumBytes -= mergeSPUpdates(MBB, MBBI, true);
1121 // Adjust stack pointer: ESP -= numbytes.
1123 // Windows and cygwin/mingw require a prologue helper routine when allocating
1124 // more than 4K bytes on the stack. Windows uses __chkstk and cygwin/mingw
1125 // uses __alloca. __alloca and the 32-bit version of __chkstk will probe the
1126 // stack and adjust the stack pointer in one go. The 64-bit version of
1127 // __chkstk is only responsible for probing the stack. The 64-bit prologue is
1128 // responsible for adjusting the stack pointer. Touching the stack at 4K
1129 // increments is necessary to ensure that the guard pages used by the OS
1130 // virtual memory manager are allocated in correct sequence.
1131 uint64_t AlignedNumBytes = NumBytes;
1132 if (IsWin64Prologue && !IsFunclet && TRI->needsStackRealignment(MF))
1133 AlignedNumBytes = RoundUpToAlignment(AlignedNumBytes, MaxAlign);
1134 if (AlignedNumBytes >= StackProbeSize && UseStackProbe) {
1135 // Check whether EAX is livein for this function.
1136 bool isEAXAlive = isEAXLiveIn(MF);
1139 // Sanity check that EAX is not livein for this function.
1140 // It should not be, so throw an assert.
1141 assert(!Is64Bit && "EAX is livein in x64 case!");
1144 BuildMI(MBB, MBBI, DL, TII.get(X86::PUSH32r))
1145 .addReg(X86::EAX, RegState::Kill)
1146 .setMIFlag(MachineInstr::FrameSetup);
1150 // Handle the 64-bit Windows ABI case where we need to call __chkstk.
1151 // Function prologue is responsible for adjusting the stack pointer.
1152 if (isUInt<32>(NumBytes)) {
1153 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
1155 .setMIFlag(MachineInstr::FrameSetup);
1156 } else if (isInt<32>(NumBytes)) {
1157 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64ri32), X86::RAX)
1159 .setMIFlag(MachineInstr::FrameSetup);
1161 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64ri), X86::RAX)
1163 .setMIFlag(MachineInstr::FrameSetup);
1166 // Allocate NumBytes-4 bytes on stack in case of isEAXAlive.
1167 // We'll also use 4 already allocated bytes for EAX.
1168 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
1169 .addImm(isEAXAlive ? NumBytes - 4 : NumBytes)
1170 .setMIFlag(MachineInstr::FrameSetup);
1173 // Call __chkstk, __chkstk_ms, or __alloca.
1174 emitStackProbe(MF, MBB, MBBI, DL, true);
1179 addRegOffset(BuildMI(MF, DL, TII.get(X86::MOV32rm), X86::EAX),
1180 StackPtr, false, NumBytes - 4);
1181 MI->setFlag(MachineInstr::FrameSetup);
1182 MBB.insert(MBBI, MI);
1184 } else if (NumBytes) {
1185 emitSPUpdate(MBB, MBBI, -(int64_t)NumBytes, /*InEpilogue=*/false);
1188 if (NeedsWinCFI && NumBytes)
1189 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_StackAlloc))
1191 .setMIFlag(MachineInstr::FrameSetup);
1193 int SEHFrameOffset = 0;
1194 unsigned SPOrEstablisher;
1197 // The establisher parameter passed to a CLR funclet is actually a pointer
1198 // to the (mostly empty) frame of its nearest enclosing funclet; we have
1199 // to find the root function establisher frame by loading the PSPSym from
1200 // the intermediate frame.
1201 unsigned PSPSlotOffset = getPSPSlotOffsetFromSP(MF);
1202 MachinePointerInfo NoInfo;
1203 MBB.addLiveIn(Establisher);
1204 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64rm), Establisher),
1205 Establisher, false, PSPSlotOffset)
1206 .addMemOperand(MF.getMachineMemOperand(
1207 NoInfo, MachineMemOperand::MOLoad, SlotSize, SlotSize));
1209 // Save the root establisher back into the current funclet's (mostly
1210 // empty) frame, in case a sub-funclet or the GC needs it.
1211 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64mr)), StackPtr,
1212 false, PSPSlotOffset)
1213 .addReg(Establisher)
1215 MF.getMachineMemOperand(NoInfo, MachineMemOperand::MOStore |
1216 MachineMemOperand::MOVolatile,
1217 SlotSize, SlotSize));
1219 SPOrEstablisher = Establisher;
1221 SPOrEstablisher = StackPtr;
1224 if (IsWin64Prologue && HasFP) {
1225 // Set RBP to a small fixed offset from RSP. In the funclet case, we base
1226 // this calculation on the incoming establisher, which holds the value of
1227 // RSP from the parent frame at the end of the prologue.
1228 SEHFrameOffset = calculateSetFPREG(ParentFrameNumBytes);
1230 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::LEA64r), FramePtr),
1231 SPOrEstablisher, false, SEHFrameOffset);
1233 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64rr), FramePtr)
1234 .addReg(SPOrEstablisher);
1236 // If this is not a funclet, emit the CFI describing our frame pointer.
1237 if (NeedsWinCFI && !IsFunclet) {
1238 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_SetFrame))
1240 .addImm(SEHFrameOffset)
1241 .setMIFlag(MachineInstr::FrameSetup);
1242 if (isAsynchronousEHPersonality(Personality))
1243 MF.getWinEHFuncInfo()->SEHSetFrameOffset = SEHFrameOffset;
1245 } else if (IsFunclet && STI.is32Bit()) {
1246 // Reset EBP / ESI to something good for funclets.
1247 MBBI = restoreWin32EHStackPointers(MBB, MBBI, DL);
1248 // If we're a catch funclet, we can be returned to via catchret. Save ESP
1249 // into the registration node so that the runtime will restore it for us.
1250 if (!MBB.isCleanupFuncletEntry()) {
1251 assert(Personality == EHPersonality::MSVC_CXX);
1253 int FI = MF.getWinEHFuncInfo()->EHRegNodeFrameIndex;
1254 int64_t EHRegOffset = getFrameIndexReference(MF, FI, FrameReg);
1255 // ESP is the first field, so no extra displacement is needed.
1256 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32mr)), FrameReg,
1262 while (MBBI != MBB.end() && MBBI->getFlag(MachineInstr::FrameSetup)) {
1263 const MachineInstr *FrameInstr = &*MBBI;
1268 if (unsigned Reg = TII.isStoreToStackSlot(FrameInstr, FI)) {
1269 if (X86::FR64RegClass.contains(Reg)) {
1270 unsigned IgnoredFrameReg;
1271 int Offset = getFrameIndexReference(MF, FI, IgnoredFrameReg);
1272 Offset += SEHFrameOffset;
1274 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_SaveXMM))
1277 .setMIFlag(MachineInstr::FrameSetup);
1284 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_EndPrologue))
1285 .setMIFlag(MachineInstr::FrameSetup);
1287 if (FnHasClrFunclet && !IsFunclet) {
1288 // Save the so-called Initial-SP (i.e. the value of the stack pointer
1289 // immediately after the prolog) into the PSPSlot so that funclets
1290 // and the GC can recover it.
1291 unsigned PSPSlotOffset = getPSPSlotOffsetFromSP(MF);
1292 auto PSPInfo = MachinePointerInfo::getFixedStack(
1293 MF, MF.getWinEHFuncInfo()->PSPSymFrameIdx);
1294 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64mr)), StackPtr, false,
1297 .addMemOperand(MF.getMachineMemOperand(
1298 PSPInfo, MachineMemOperand::MOStore | MachineMemOperand::MOVolatile,
1299 SlotSize, SlotSize));
1302 // Realign stack after we spilled callee-saved registers (so that we'll be
1303 // able to calculate their offsets from the frame pointer).
1304 // Win64 requires aligning the stack after the prologue.
1305 if (IsWin64Prologue && TRI->needsStackRealignment(MF)) {
1306 assert(HasFP && "There should be a frame pointer if stack is realigned.");
1307 BuildStackAlignAND(MBB, MBBI, DL, SPOrEstablisher, MaxAlign);
1310 // We already dealt with stack realignment and funclets above.
1311 if (IsFunclet && STI.is32Bit())
1314 // If we need a base pointer, set it up here. It's whatever the value
1315 // of the stack pointer is at this point. Any variable size objects
1316 // will be allocated after this, so we can still use the base pointer
1317 // to reference locals.
1318 if (TRI->hasBasePointer(MF)) {
1319 // Update the base pointer with the current stack pointer.
1320 unsigned Opc = Uses64BitFramePtr ? X86::MOV64rr : X86::MOV32rr;
1321 BuildMI(MBB, MBBI, DL, TII.get(Opc), BasePtr)
1322 .addReg(SPOrEstablisher)
1323 .setMIFlag(MachineInstr::FrameSetup);
1324 if (X86FI->getRestoreBasePointer()) {
1325 // Stash value of base pointer. Saving RSP instead of EBP shortens
1326 // dependence chain. Used by SjLj EH.
1327 unsigned Opm = Uses64BitFramePtr ? X86::MOV64mr : X86::MOV32mr;
1328 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(Opm)),
1329 FramePtr, true, X86FI->getRestoreBasePointerOffset())
1330 .addReg(SPOrEstablisher)
1331 .setMIFlag(MachineInstr::FrameSetup);
1334 if (X86FI->getHasSEHFramePtrSave() && !IsFunclet) {
1335 // Stash the value of the frame pointer relative to the base pointer for
1336 // Win32 EH. This supports Win32 EH, which does the inverse of the above:
1337 // it recovers the frame pointer from the base pointer rather than the
1338 // other way around.
1339 unsigned Opm = Uses64BitFramePtr ? X86::MOV64mr : X86::MOV32mr;
1342 getFrameIndexReference(MF, X86FI->getSEHFramePtrSaveIndex(), UsedReg);
1343 assert(UsedReg == BasePtr);
1344 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(Opm)), UsedReg, true, Offset)
1346 .setMIFlag(MachineInstr::FrameSetup);
1350 if (((!HasFP && NumBytes) || PushedRegs) && NeedsDwarfCFI) {
1351 // Mark end of stack pointer adjustment.
1352 if (!HasFP && NumBytes) {
1353 // Define the current CFA rule to use the provided offset.
1355 BuildCFI(MBB, MBBI, DL, MCCFIInstruction::createDefCfaOffset(
1356 nullptr, -StackSize + stackGrowth));
1359 // Emit DWARF info specifying the offsets of the callee-saved registers.
1361 emitCalleeSavedFrameMoves(MBB, MBBI, DL);
1365 bool X86FrameLowering::canUseLEAForSPInEpilogue(
1366 const MachineFunction &MF) const {
1367 // We can't use LEA instructions for adjusting the stack pointer if this is a
1368 // leaf function in the Win64 ABI. Only ADD instructions may be used to
1369 // deallocate the stack.
1370 // This means that we can use LEA for SP in two situations:
1371 // 1. We *aren't* using the Win64 ABI which means we are free to use LEA.
1372 // 2. We *have* a frame pointer which means we are permitted to use LEA.
1373 return !MF.getTarget().getMCAsmInfo()->usesWindowsCFI() || hasFP(MF);
1376 static bool isFuncletReturnInstr(MachineInstr *MI) {
1377 switch (MI->getOpcode()) {
1379 case X86::CLEANUPRET:
1384 llvm_unreachable("impossible");
1387 // CLR funclets use a special "Previous Stack Pointer Symbol" slot on the
1388 // stack. It holds a pointer to the bottom of the root function frame. The
1389 // establisher frame pointer passed to a nested funclet may point to the
1390 // (mostly empty) frame of its parent funclet, but it will need to find
1391 // the frame of the root function to access locals. To facilitate this,
1392 // every funclet copies the pointer to the bottom of the root function
1393 // frame into a PSPSym slot in its own (mostly empty) stack frame. Using the
1394 // same offset for the PSPSym in the root function frame that's used in the
1395 // funclets' frames allows each funclet to dynamically accept any ancestor
1396 // frame as its establisher argument (the runtime doesn't guarantee the
1397 // immediate parent for some reason lost to history), and also allows the GC,
1398 // which uses the PSPSym for some bookkeeping, to find it in any funclet's
1399 // frame with only a single offset reported for the entire method.
1401 X86FrameLowering::getPSPSlotOffsetFromSP(const MachineFunction &MF) const {
1402 const WinEHFuncInfo &Info = *MF.getWinEHFuncInfo();
1403 // getFrameIndexReferenceFromSP has an out ref parameter for the stack
1404 // pointer register; pass a dummy that we ignore
1406 int Offset = getFrameIndexReferenceFromSP(MF, Info.PSPSymFrameIdx, SPReg);
1407 assert(Offset >= 0);
1408 return static_cast<unsigned>(Offset);
1412 X86FrameLowering::getWinEHFuncletFrameSize(const MachineFunction &MF) const {
1413 // This is the size of the pushed CSRs.
1415 MF.getInfo<X86MachineFunctionInfo>()->getCalleeSavedFrameSize();
1416 // This is the amount of stack a funclet needs to allocate.
1418 EHPersonality Personality =
1419 classifyEHPersonality(MF.getFunction()->getPersonalityFn());
1420 if (Personality == EHPersonality::CoreCLR) {
1421 // CLR funclets need to hold enough space to include the PSPSym, at the
1422 // same offset from the stack pointer (immediately after the prolog) as it
1423 // resides at in the main function.
1424 UsedSize = getPSPSlotOffsetFromSP(MF) + SlotSize;
1426 // Other funclets just need enough stack for outgoing call arguments.
1427 UsedSize = MF.getFrameInfo()->getMaxCallFrameSize();
1429 // RBP is not included in the callee saved register block. After pushing RBP,
1430 // everything is 16 byte aligned. Everything we allocate before an outgoing
1431 // call must also be 16 byte aligned.
1432 unsigned FrameSizeMinusRBP =
1433 RoundUpToAlignment(CSSize + UsedSize, getStackAlignment());
1434 // Subtract out the size of the callee saved registers. This is how much stack
1435 // each funclet will allocate.
1436 return FrameSizeMinusRBP - CSSize;
1439 void X86FrameLowering::emitEpilogue(MachineFunction &MF,
1440 MachineBasicBlock &MBB) const {
1441 const MachineFrameInfo *MFI = MF.getFrameInfo();
1442 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1443 MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator();
1445 if (MBBI != MBB.end())
1446 DL = MBBI->getDebugLoc();
1447 // standard x86_64 and NaCl use 64-bit frame/stack pointers, x32 - 32-bit.
1448 const bool Is64BitILP32 = STI.isTarget64BitILP32();
1449 unsigned FramePtr = TRI->getFrameRegister(MF);
1450 unsigned MachineFramePtr =
1451 Is64BitILP32 ? getX86SubSuperRegister(FramePtr, 64) : FramePtr;
1453 bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
1455 IsWin64Prologue && MF.getFunction()->needsUnwindTableEntry();
1456 bool IsFunclet = isFuncletReturnInstr(MBBI);
1457 MachineBasicBlock *TargetMBB = nullptr;
1459 // Get the number of bytes to allocate from the FrameInfo.
1460 uint64_t StackSize = MFI->getStackSize();
1461 uint64_t MaxAlign = calculateMaxStackAlign(MF);
1462 unsigned CSSize = X86FI->getCalleeSavedFrameSize();
1463 uint64_t NumBytes = 0;
1465 if (MBBI->getOpcode() == X86::CATCHRET) {
1466 // SEH shouldn't use catchret.
1467 assert(!isAsynchronousEHPersonality(
1468 classifyEHPersonality(MF.getFunction()->getPersonalityFn())) &&
1469 "SEH should not use CATCHRET");
1471 NumBytes = getWinEHFuncletFrameSize(MF);
1472 assert(hasFP(MF) && "EH funclets without FP not yet implemented");
1473 TargetMBB = MBBI->getOperand(0).getMBB();
1476 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::POP64r : X86::POP32r),
1478 .setMIFlag(MachineInstr::FrameDestroy);
1479 } else if (MBBI->getOpcode() == X86::CLEANUPRET) {
1480 NumBytes = getWinEHFuncletFrameSize(MF);
1481 assert(hasFP(MF) && "EH funclets without FP not yet implemented");
1482 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::POP64r : X86::POP32r),
1484 .setMIFlag(MachineInstr::FrameDestroy);
1485 } else if (hasFP(MF)) {
1486 // Calculate required stack adjustment.
1487 uint64_t FrameSize = StackSize - SlotSize;
1488 NumBytes = FrameSize - CSSize;
1490 // Callee-saved registers were pushed on stack before the stack was
1492 if (TRI->needsStackRealignment(MF) && !IsWin64Prologue)
1493 NumBytes = RoundUpToAlignment(FrameSize, MaxAlign);
1496 BuildMI(MBB, MBBI, DL,
1497 TII.get(Is64Bit ? X86::POP64r : X86::POP32r), MachineFramePtr)
1498 .setMIFlag(MachineInstr::FrameDestroy);
1500 NumBytes = StackSize - CSSize;
1502 uint64_t SEHStackAllocAmt = NumBytes;
1504 // Skip the callee-saved pop instructions.
1505 while (MBBI != MBB.begin()) {
1506 MachineBasicBlock::iterator PI = std::prev(MBBI);
1507 unsigned Opc = PI->getOpcode();
1509 if ((Opc != X86::POP32r || !PI->getFlag(MachineInstr::FrameDestroy)) &&
1510 (Opc != X86::POP64r || !PI->getFlag(MachineInstr::FrameDestroy)) &&
1511 Opc != X86::DBG_VALUE && !PI->isTerminator())
1516 MachineBasicBlock::iterator FirstCSPop = MBBI;
1519 // Fill EAX/RAX with the address of the target block.
1520 unsigned ReturnReg = STI.is64Bit() ? X86::RAX : X86::EAX;
1521 if (STI.is64Bit()) {
1522 // LEA64r TargetMBB(%rip), %rax
1523 BuildMI(MBB, FirstCSPop, DL, TII.get(X86::LEA64r), ReturnReg)
1530 // MOV32ri $TargetMBB, %eax
1531 BuildMI(MBB, FirstCSPop, DL, TII.get(X86::MOV32ri), ReturnReg)
1534 // Record that we've taken the address of TargetMBB and no longer just
1535 // reference it in a terminator.
1536 TargetMBB->setHasAddressTaken();
1539 if (MBBI != MBB.end())
1540 DL = MBBI->getDebugLoc();
1542 // If there is an ADD32ri or SUB32ri of ESP immediately before this
1543 // instruction, merge the two instructions.
1544 if (NumBytes || MFI->hasVarSizedObjects())
1545 NumBytes += mergeSPUpdates(MBB, MBBI, true);
1547 // If dynamic alloca is used, then reset esp to point to the last callee-saved
1548 // slot before popping them off! Same applies for the case, when stack was
1549 // realigned. Don't do this if this was a funclet epilogue, since the funclets
1550 // will not do realignment or dynamic stack allocation.
1551 if ((TRI->needsStackRealignment(MF) || MFI->hasVarSizedObjects()) &&
1553 if (TRI->needsStackRealignment(MF))
1555 unsigned SEHFrameOffset = calculateSetFPREG(SEHStackAllocAmt);
1556 uint64_t LEAAmount =
1557 IsWin64Prologue ? SEHStackAllocAmt - SEHFrameOffset : -CSSize;
1559 // There are only two legal forms of epilogue:
1560 // - add SEHAllocationSize, %rsp
1561 // - lea SEHAllocationSize(%FramePtr), %rsp
1563 // 'mov %FramePtr, %rsp' will not be recognized as an epilogue sequence.
1564 // However, we may use this sequence if we have a frame pointer because the
1565 // effects of the prologue can safely be undone.
1566 if (LEAAmount != 0) {
1567 unsigned Opc = getLEArOpcode(Uses64BitFramePtr);
1568 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr),
1569 FramePtr, false, LEAAmount);
1572 unsigned Opc = (Uses64BitFramePtr ? X86::MOV64rr : X86::MOV32rr);
1573 BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr)
1577 } else if (NumBytes) {
1578 // Adjust stack pointer back: ESP += numbytes.
1579 emitSPUpdate(MBB, MBBI, NumBytes, /*InEpilogue=*/true);
1583 // Windows unwinder will not invoke function's exception handler if IP is
1584 // either in prologue or in epilogue. This behavior causes a problem when a
1585 // call immediately precedes an epilogue, because the return address points
1586 // into the epilogue. To cope with that, we insert an epilogue marker here,
1587 // then replace it with a 'nop' if it ends up immediately after a CALL in the
1588 // final emitted code.
1590 BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_Epilogue));
1592 // Add the return addr area delta back since we are not tail calling.
1593 int Offset = -1 * X86FI->getTCReturnAddrDelta();
1594 assert(Offset >= 0 && "TCDelta should never be positive");
1596 MBBI = MBB.getFirstTerminator();
1598 // Check for possible merge with preceding ADD instruction.
1599 Offset += mergeSPUpdates(MBB, MBBI, true);
1600 emitSPUpdate(MBB, MBBI, Offset, /*InEpilogue=*/true);
1604 // NOTE: this only has a subset of the full frame index logic. In
1605 // particular, the FI < 0 and AfterFPPop logic is handled in
1606 // X86RegisterInfo::eliminateFrameIndex, but not here. Possibly
1607 // (probably?) it should be moved into here.
1608 int X86FrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI,
1609 unsigned &FrameReg) const {
1610 const MachineFrameInfo *MFI = MF.getFrameInfo();
1612 // We can't calculate offset from frame pointer if the stack is realigned,
1613 // so enforce usage of stack/base pointer. The base pointer is used when we
1614 // have dynamic allocas in addition to dynamic realignment.
1615 if (TRI->hasBasePointer(MF))
1616 FrameReg = TRI->getBaseRegister();
1617 else if (TRI->needsStackRealignment(MF))
1618 FrameReg = TRI->getStackRegister();
1620 FrameReg = TRI->getFrameRegister(MF);
1622 // Offset will hold the offset from the stack pointer at function entry to the
1624 // We need to factor in additional offsets applied during the prologue to the
1625 // frame, base, and stack pointer depending on which is used.
1626 int Offset = MFI->getObjectOffset(FI) - getOffsetOfLocalArea();
1627 const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1628 unsigned CSSize = X86FI->getCalleeSavedFrameSize();
1629 uint64_t StackSize = MFI->getStackSize();
1630 bool HasFP = hasFP(MF);
1631 bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
1632 int64_t FPDelta = 0;
1634 if (IsWin64Prologue) {
1635 assert(!MFI->hasCalls() || (StackSize % 16) == 8);
1637 // Calculate required stack adjustment.
1638 uint64_t FrameSize = StackSize - SlotSize;
1639 // If required, include space for extra hidden slot for stashing base pointer.
1640 if (X86FI->getRestoreBasePointer())
1641 FrameSize += SlotSize;
1642 uint64_t NumBytes = FrameSize - CSSize;
1644 uint64_t SEHFrameOffset = calculateSetFPREG(NumBytes);
1645 if (FI && FI == X86FI->getFAIndex())
1646 return -SEHFrameOffset;
1648 // FPDelta is the offset from the "traditional" FP location of the old base
1649 // pointer followed by return address and the location required by the
1650 // restricted Win64 prologue.
1651 // Add FPDelta to all offsets below that go through the frame pointer.
1652 FPDelta = FrameSize - SEHFrameOffset;
1653 assert((!MFI->hasCalls() || (FPDelta % 16) == 0) &&
1654 "FPDelta isn't aligned per the Win64 ABI!");
1658 if (TRI->hasBasePointer(MF)) {
1659 assert(HasFP && "VLAs and dynamic stack realign, but no FP?!");
1661 // Skip the saved EBP.
1662 return Offset + SlotSize + FPDelta;
1664 assert((-(Offset + StackSize)) % MFI->getObjectAlignment(FI) == 0);
1665 return Offset + StackSize;
1667 } else if (TRI->needsStackRealignment(MF)) {
1669 // Skip the saved EBP.
1670 return Offset + SlotSize + FPDelta;
1672 assert((-(Offset + StackSize)) % MFI->getObjectAlignment(FI) == 0);
1673 return Offset + StackSize;
1675 // FIXME: Support tail calls
1678 return Offset + StackSize;
1680 // Skip the saved EBP.
1683 // Skip the RETADDR move area
1684 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
1685 if (TailCallReturnAddrDelta < 0)
1686 Offset -= TailCallReturnAddrDelta;
1689 return Offset + FPDelta;
1692 // Simplified from getFrameIndexReference keeping only StackPointer cases
1693 int X86FrameLowering::getFrameIndexReferenceFromSP(const MachineFunction &MF,
1695 unsigned &FrameReg) const {
1696 const MachineFrameInfo *MFI = MF.getFrameInfo();
1697 // Does not include any dynamic realign.
1698 const uint64_t StackSize = MFI->getStackSize();
1701 // LLVM arranges the stack as follows:
1706 // PUSH RBP <-- RBP points here
1708 // ~~~~~~~ <-- possible stack realignment (non-win64)
1711 // ... <-- RSP after prologue points here
1712 // ~~~~~~~ <-- possible stack realignment (win64)
1714 // if (hasVarSizedObjects()):
1715 // ... <-- "base pointer" (ESI/RBX) points here
1717 // ... <-- RSP points here
1719 // Case 1: In the simple case of no stack realignment and no dynamic
1720 // allocas, both "fixed" stack objects (arguments and CSRs) are addressable
1721 // with fixed offsets from RSP.
1723 // Case 2: In the case of stack realignment with no dynamic allocas, fixed
1724 // stack objects are addressed with RBP and regular stack objects with RSP.
1726 // Case 3: In the case of dynamic allocas and stack realignment, RSP is used
1727 // to address stack arguments for outgoing calls and nothing else. The "base
1728 // pointer" points to local variables, and RBP points to fixed objects.
1730 // In cases 2 and 3, we can only answer for non-fixed stack objects, and the
1731 // answer we give is relative to the SP after the prologue, and not the
1732 // SP in the middle of the function.
1734 assert((!MFI->isFixedObjectIndex(FI) || !TRI->needsStackRealignment(MF) ||
1735 STI.isTargetWin64()) &&
1736 "offset from fixed object to SP is not static");
1738 // We don't handle tail calls, and shouldn't be seeing them either.
1739 int TailCallReturnAddrDelta =
1740 MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta();
1741 assert(!(TailCallReturnAddrDelta < 0) && "we don't handle this case!");
1745 // Fill in FrameReg output argument.
1746 FrameReg = TRI->getStackRegister();
1748 // This is how the math works out:
1750 // %rsp grows (i.e. gets lower) left to right. Each box below is
1751 // one word (eight bytes). Obj0 is the stack slot we're trying to
1754 // ----------------------------------
1755 // | BP | Obj0 | Obj1 | ... | ObjN |
1756 // ----------------------------------
1760 // A is the incoming stack pointer.
1761 // (B - A) is the local area offset (-8 for x86-64) [1]
1762 // (C - A) is the Offset returned by MFI->getObjectOffset for Obj0 [2]
1764 // |(E - B)| is the StackSize (absolute value, positive). For a
1765 // stack that grown down, this works out to be (B - E). [3]
1767 // E is also the value of %rsp after stack has been set up, and we
1768 // want (C - E) -- the value we can add to %rsp to get to Obj0. Now
1769 // (C - E) == (C - A) - (B - A) + (B - E)
1770 // { Using [1], [2] and [3] above }
1771 // == getObjectOffset - LocalAreaOffset + StackSize
1774 // Get the Offset from the StackPointer
1775 int Offset = MFI->getObjectOffset(FI) - getOffsetOfLocalArea();
1777 return Offset + StackSize;
1780 bool X86FrameLowering::assignCalleeSavedSpillSlots(
1781 MachineFunction &MF, const TargetRegisterInfo *TRI,
1782 std::vector<CalleeSavedInfo> &CSI) const {
1783 MachineFrameInfo *MFI = MF.getFrameInfo();
1784 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1786 unsigned CalleeSavedFrameSize = 0;
1787 int SpillSlotOffset = getOffsetOfLocalArea() + X86FI->getTCReturnAddrDelta();
1790 // emitPrologue always spills frame register the first thing.
1791 SpillSlotOffset -= SlotSize;
1792 MFI->CreateFixedSpillStackObject(SlotSize, SpillSlotOffset);
1794 // Since emitPrologue and emitEpilogue will handle spilling and restoring of
1795 // the frame register, we can delete it from CSI list and not have to worry
1796 // about avoiding it later.
1797 unsigned FPReg = TRI->getFrameRegister(MF);
1798 for (unsigned i = 0; i < CSI.size(); ++i) {
1799 if (TRI->regsOverlap(CSI[i].getReg(),FPReg)) {
1800 CSI.erase(CSI.begin() + i);
1806 // Assign slots for GPRs. It increases frame size.
1807 for (unsigned i = CSI.size(); i != 0; --i) {
1808 unsigned Reg = CSI[i - 1].getReg();
1810 if (!X86::GR64RegClass.contains(Reg) && !X86::GR32RegClass.contains(Reg))
1813 SpillSlotOffset -= SlotSize;
1814 CalleeSavedFrameSize += SlotSize;
1816 int SlotIndex = MFI->CreateFixedSpillStackObject(SlotSize, SpillSlotOffset);
1817 CSI[i - 1].setFrameIdx(SlotIndex);
1820 X86FI->setCalleeSavedFrameSize(CalleeSavedFrameSize);
1822 // Assign slots for XMMs.
1823 for (unsigned i = CSI.size(); i != 0; --i) {
1824 unsigned Reg = CSI[i - 1].getReg();
1825 if (X86::GR64RegClass.contains(Reg) || X86::GR32RegClass.contains(Reg))
1828 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
1830 SpillSlotOffset -= std::abs(SpillSlotOffset) % RC->getAlignment();
1832 SpillSlotOffset -= RC->getSize();
1834 MFI->CreateFixedSpillStackObject(RC->getSize(), SpillSlotOffset);
1835 CSI[i - 1].setFrameIdx(SlotIndex);
1836 MFI->ensureMaxAlignment(RC->getAlignment());
1842 bool X86FrameLowering::spillCalleeSavedRegisters(
1843 MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
1844 const std::vector<CalleeSavedInfo> &CSI,
1845 const TargetRegisterInfo *TRI) const {
1846 DebugLoc DL = MBB.findDebugLoc(MI);
1848 // Don't save CSRs in 32-bit EH funclets. The caller saves EBX, EBP, ESI, EDI
1849 // for us, and there are no XMM CSRs on Win32.
1850 if (MBB.isEHFuncletEntry() && STI.is32Bit() && STI.isOSWindows())
1853 // Push GPRs. It increases frame size.
1854 unsigned Opc = STI.is64Bit() ? X86::PUSH64r : X86::PUSH32r;
1855 for (unsigned i = CSI.size(); i != 0; --i) {
1856 unsigned Reg = CSI[i - 1].getReg();
1858 if (!X86::GR64RegClass.contains(Reg) && !X86::GR32RegClass.contains(Reg))
1860 // Add the callee-saved register as live-in. It's killed at the spill.
1863 BuildMI(MBB, MI, DL, TII.get(Opc)).addReg(Reg, RegState::Kill)
1864 .setMIFlag(MachineInstr::FrameSetup);
1867 // Make XMM regs spilled. X86 does not have ability of push/pop XMM.
1868 // It can be done by spilling XMMs to stack frame.
1869 for (unsigned i = CSI.size(); i != 0; --i) {
1870 unsigned Reg = CSI[i-1].getReg();
1871 if (X86::GR64RegClass.contains(Reg) || X86::GR32RegClass.contains(Reg))
1873 // Add the callee-saved register as live-in. It's killed at the spill.
1875 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
1877 TII.storeRegToStackSlot(MBB, MI, Reg, true, CSI[i - 1].getFrameIdx(), RC,
1880 MI->setFlag(MachineInstr::FrameSetup);
1887 bool X86FrameLowering::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1888 MachineBasicBlock::iterator MI,
1889 const std::vector<CalleeSavedInfo> &CSI,
1890 const TargetRegisterInfo *TRI) const {
1894 if (isFuncletReturnInstr(MI) && STI.isOSWindows()) {
1895 // Don't restore CSRs in 32-bit EH funclets. Matches
1896 // spillCalleeSavedRegisters.
1899 // Don't restore CSRs before an SEH catchret. SEH except blocks do not form
1900 // funclets. emitEpilogue transforms these to normal jumps.
1901 if (MI->getOpcode() == X86::CATCHRET) {
1902 const Function *Func = MBB.getParent()->getFunction();
1903 bool IsSEH = isAsynchronousEHPersonality(
1904 classifyEHPersonality(Func->getPersonalityFn()));
1910 DebugLoc DL = MBB.findDebugLoc(MI);
1912 // Reload XMMs from stack frame.
1913 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1914 unsigned Reg = CSI[i].getReg();
1915 if (X86::GR64RegClass.contains(Reg) ||
1916 X86::GR32RegClass.contains(Reg))
1919 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
1920 TII.loadRegFromStackSlot(MBB, MI, Reg, CSI[i].getFrameIdx(), RC, TRI);
1924 unsigned Opc = STI.is64Bit() ? X86::POP64r : X86::POP32r;
1925 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1926 unsigned Reg = CSI[i].getReg();
1927 if (!X86::GR64RegClass.contains(Reg) &&
1928 !X86::GR32RegClass.contains(Reg))
1931 BuildMI(MBB, MI, DL, TII.get(Opc), Reg)
1932 .setMIFlag(MachineInstr::FrameDestroy);
1937 void X86FrameLowering::determineCalleeSaves(MachineFunction &MF,
1938 BitVector &SavedRegs,
1939 RegScavenger *RS) const {
1940 TargetFrameLowering::determineCalleeSaves(MF, SavedRegs, RS);
1942 MachineFrameInfo *MFI = MF.getFrameInfo();
1944 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1945 int64_t TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
1947 if (TailCallReturnAddrDelta < 0) {
1948 // create RETURNADDR area
1957 MFI->CreateFixedObject(-TailCallReturnAddrDelta,
1958 TailCallReturnAddrDelta - SlotSize, true);
1961 // Spill the BasePtr if it's used.
1962 if (TRI->hasBasePointer(MF)) {
1963 SavedRegs.set(TRI->getBaseRegister());
1965 // Allocate a spill slot for EBP if we have a base pointer and EH funclets.
1966 if (MF.getMMI().hasEHFunclets()) {
1967 int FI = MFI->CreateSpillStackObject(SlotSize, SlotSize);
1968 X86FI->setHasSEHFramePtrSave(true);
1969 X86FI->setSEHFramePtrSaveIndex(FI);
1975 HasNestArgument(const MachineFunction *MF) {
1976 const Function *F = MF->getFunction();
1977 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
1979 if (I->hasNestAttr())
1985 /// GetScratchRegister - Get a temp register for performing work in the
1986 /// segmented stack and the Erlang/HiPE stack prologue. Depending on platform
1987 /// and the properties of the function either one or two registers will be
1988 /// needed. Set primary to true for the first register, false for the second.
1990 GetScratchRegister(bool Is64Bit, bool IsLP64, const MachineFunction &MF, bool Primary) {
1991 CallingConv::ID CallingConvention = MF.getFunction()->getCallingConv();
1994 if (CallingConvention == CallingConv::HiPE) {
1996 return Primary ? X86::R14 : X86::R13;
1998 return Primary ? X86::EBX : X86::EDI;
2003 return Primary ? X86::R11 : X86::R12;
2005 return Primary ? X86::R11D : X86::R12D;
2008 bool IsNested = HasNestArgument(&MF);
2010 if (CallingConvention == CallingConv::X86_FastCall ||
2011 CallingConvention == CallingConv::Fast) {
2013 report_fatal_error("Segmented stacks does not support fastcall with "
2014 "nested function.");
2015 return Primary ? X86::EAX : X86::ECX;
2018 return Primary ? X86::EDX : X86::EAX;
2019 return Primary ? X86::ECX : X86::EAX;
2022 // The stack limit in the TCB is set to this many bytes above the actual stack
2024 static const uint64_t kSplitStackAvailable = 256;
2026 void X86FrameLowering::adjustForSegmentedStacks(
2027 MachineFunction &MF, MachineBasicBlock &PrologueMBB) const {
2028 MachineFrameInfo *MFI = MF.getFrameInfo();
2030 unsigned TlsReg, TlsOffset;
2033 unsigned ScratchReg = GetScratchRegister(Is64Bit, IsLP64, MF, true);
2034 assert(!MF.getRegInfo().isLiveIn(ScratchReg) &&
2035 "Scratch register is live-in");
2037 if (MF.getFunction()->isVarArg())
2038 report_fatal_error("Segmented stacks do not support vararg functions.");
2039 if (!STI.isTargetLinux() && !STI.isTargetDarwin() && !STI.isTargetWin32() &&
2040 !STI.isTargetWin64() && !STI.isTargetFreeBSD() &&
2041 !STI.isTargetDragonFly())
2042 report_fatal_error("Segmented stacks not supported on this platform.");
2044 // Eventually StackSize will be calculated by a link-time pass; which will
2045 // also decide whether checking code needs to be injected into this particular
2047 StackSize = MFI->getStackSize();
2049 // Do not generate a prologue for functions with a stack of size zero
2053 MachineBasicBlock *allocMBB = MF.CreateMachineBasicBlock();
2054 MachineBasicBlock *checkMBB = MF.CreateMachineBasicBlock();
2055 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
2056 bool IsNested = false;
2058 // We need to know if the function has a nest argument only in 64 bit mode.
2060 IsNested = HasNestArgument(&MF);
2062 // The MOV R10, RAX needs to be in a different block, since the RET we emit in
2063 // allocMBB needs to be last (terminating) instruction.
2065 for (const auto &LI : PrologueMBB.liveins()) {
2066 allocMBB->addLiveIn(LI);
2067 checkMBB->addLiveIn(LI);
2071 allocMBB->addLiveIn(IsLP64 ? X86::R10 : X86::R10D);
2073 MF.push_front(allocMBB);
2074 MF.push_front(checkMBB);
2076 // When the frame size is less than 256 we just compare the stack
2077 // boundary directly to the value of the stack pointer, per gcc.
2078 bool CompareStackPointer = StackSize < kSplitStackAvailable;
2080 // Read the limit off the current stacklet off the stack_guard location.
2082 if (STI.isTargetLinux()) {
2084 TlsOffset = IsLP64 ? 0x70 : 0x40;
2085 } else if (STI.isTargetDarwin()) {
2087 TlsOffset = 0x60 + 90*8; // See pthread_machdep.h. Steal TLS slot 90.
2088 } else if (STI.isTargetWin64()) {
2090 TlsOffset = 0x28; // pvArbitrary, reserved for application use
2091 } else if (STI.isTargetFreeBSD()) {
2094 } else if (STI.isTargetDragonFly()) {
2096 TlsOffset = 0x20; // use tls_tcb.tcb_segstack
2098 report_fatal_error("Segmented stacks not supported on this platform.");
2101 if (CompareStackPointer)
2102 ScratchReg = IsLP64 ? X86::RSP : X86::ESP;
2104 BuildMI(checkMBB, DL, TII.get(IsLP64 ? X86::LEA64r : X86::LEA64_32r), ScratchReg).addReg(X86::RSP)
2105 .addImm(1).addReg(0).addImm(-StackSize).addReg(0);
2107 BuildMI(checkMBB, DL, TII.get(IsLP64 ? X86::CMP64rm : X86::CMP32rm)).addReg(ScratchReg)
2108 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg);
2110 if (STI.isTargetLinux()) {
2113 } else if (STI.isTargetDarwin()) {
2115 TlsOffset = 0x48 + 90*4;
2116 } else if (STI.isTargetWin32()) {
2118 TlsOffset = 0x14; // pvArbitrary, reserved for application use
2119 } else if (STI.isTargetDragonFly()) {
2121 TlsOffset = 0x10; // use tls_tcb.tcb_segstack
2122 } else if (STI.isTargetFreeBSD()) {
2123 report_fatal_error("Segmented stacks not supported on FreeBSD i386.");
2125 report_fatal_error("Segmented stacks not supported on this platform.");
2128 if (CompareStackPointer)
2129 ScratchReg = X86::ESP;
2131 BuildMI(checkMBB, DL, TII.get(X86::LEA32r), ScratchReg).addReg(X86::ESP)
2132 .addImm(1).addReg(0).addImm(-StackSize).addReg(0);
2134 if (STI.isTargetLinux() || STI.isTargetWin32() || STI.isTargetWin64() ||
2135 STI.isTargetDragonFly()) {
2136 BuildMI(checkMBB, DL, TII.get(X86::CMP32rm)).addReg(ScratchReg)
2137 .addReg(0).addImm(0).addReg(0).addImm(TlsOffset).addReg(TlsReg);
2138 } else if (STI.isTargetDarwin()) {
2140 // TlsOffset doesn't fit into a mod r/m byte so we need an extra register.
2141 unsigned ScratchReg2;
2143 if (CompareStackPointer) {
2144 // The primary scratch register is available for holding the TLS offset.
2145 ScratchReg2 = GetScratchRegister(Is64Bit, IsLP64, MF, true);
2146 SaveScratch2 = false;
2148 // Need to use a second register to hold the TLS offset
2149 ScratchReg2 = GetScratchRegister(Is64Bit, IsLP64, MF, false);
2151 // Unfortunately, with fastcc the second scratch register may hold an
2153 SaveScratch2 = MF.getRegInfo().isLiveIn(ScratchReg2);
2156 // If Scratch2 is live-in then it needs to be saved.
2157 assert((!MF.getRegInfo().isLiveIn(ScratchReg2) || SaveScratch2) &&
2158 "Scratch register is live-in and not saved");
2161 BuildMI(checkMBB, DL, TII.get(X86::PUSH32r))
2162 .addReg(ScratchReg2, RegState::Kill);
2164 BuildMI(checkMBB, DL, TII.get(X86::MOV32ri), ScratchReg2)
2166 BuildMI(checkMBB, DL, TII.get(X86::CMP32rm))
2168 .addReg(ScratchReg2).addImm(1).addReg(0)
2173 BuildMI(checkMBB, DL, TII.get(X86::POP32r), ScratchReg2);
2177 // This jump is taken if SP >= (Stacklet Limit + Stack Space required).
2178 // It jumps to normal execution of the function body.
2179 BuildMI(checkMBB, DL, TII.get(X86::JA_1)).addMBB(&PrologueMBB);
2181 // On 32 bit we first push the arguments size and then the frame size. On 64
2182 // bit, we pass the stack frame size in r10 and the argument size in r11.
2184 // Functions with nested arguments use R10, so it needs to be saved across
2185 // the call to _morestack
2187 const unsigned RegAX = IsLP64 ? X86::RAX : X86::EAX;
2188 const unsigned Reg10 = IsLP64 ? X86::R10 : X86::R10D;
2189 const unsigned Reg11 = IsLP64 ? X86::R11 : X86::R11D;
2190 const unsigned MOVrr = IsLP64 ? X86::MOV64rr : X86::MOV32rr;
2191 const unsigned MOVri = IsLP64 ? X86::MOV64ri : X86::MOV32ri;
2194 BuildMI(allocMBB, DL, TII.get(MOVrr), RegAX).addReg(Reg10);
2196 BuildMI(allocMBB, DL, TII.get(MOVri), Reg10)
2198 BuildMI(allocMBB, DL, TII.get(MOVri), Reg11)
2199 .addImm(X86FI->getArgumentStackSize());
2201 BuildMI(allocMBB, DL, TII.get(X86::PUSHi32))
2202 .addImm(X86FI->getArgumentStackSize());
2203 BuildMI(allocMBB, DL, TII.get(X86::PUSHi32))
2207 // __morestack is in libgcc
2208 if (Is64Bit && MF.getTarget().getCodeModel() == CodeModel::Large) {
2209 // Under the large code model, we cannot assume that __morestack lives
2210 // within 2^31 bytes of the call site, so we cannot use pc-relative
2211 // addressing. We cannot perform the call via a temporary register,
2212 // as the rax register may be used to store the static chain, and all
2213 // other suitable registers may be either callee-save or used for
2214 // parameter passing. We cannot use the stack at this point either
2215 // because __morestack manipulates the stack directly.
2217 // To avoid these issues, perform an indirect call via a read-only memory
2218 // location containing the address.
2220 // This solution is not perfect, as it assumes that the .rodata section
2221 // is laid out within 2^31 bytes of each function body, but this seems
2222 // to be sufficient for JIT.
2223 BuildMI(allocMBB, DL, TII.get(X86::CALL64m))
2227 .addExternalSymbol("__morestack_addr")
2229 MF.getMMI().setUsesMorestackAddr(true);
2232 BuildMI(allocMBB, DL, TII.get(X86::CALL64pcrel32))
2233 .addExternalSymbol("__morestack");
2235 BuildMI(allocMBB, DL, TII.get(X86::CALLpcrel32))
2236 .addExternalSymbol("__morestack");
2240 BuildMI(allocMBB, DL, TII.get(X86::MORESTACK_RET_RESTORE_R10));
2242 BuildMI(allocMBB, DL, TII.get(X86::MORESTACK_RET));
2244 allocMBB->addSuccessor(&PrologueMBB);
2246 checkMBB->addSuccessor(allocMBB);
2247 checkMBB->addSuccessor(&PrologueMBB);
2254 /// Erlang programs may need a special prologue to handle the stack size they
2255 /// might need at runtime. That is because Erlang/OTP does not implement a C
2256 /// stack but uses a custom implementation of hybrid stack/heap architecture.
2257 /// (for more information see Eric Stenman's Ph.D. thesis:
2258 /// http://publications.uu.se/uu/fulltext/nbn_se_uu_diva-2688.pdf)
2261 /// temp0 = sp - MaxStack
2262 /// if( temp0 < SP_LIMIT(P) ) goto IncStack else goto OldStart
2266 /// call inc_stack # doubles the stack space
2267 /// temp0 = sp - MaxStack
2268 /// if( temp0 < SP_LIMIT(P) ) goto IncStack else goto OldStart
2269 void X86FrameLowering::adjustForHiPEPrologue(
2270 MachineFunction &MF, MachineBasicBlock &PrologueMBB) const {
2271 MachineFrameInfo *MFI = MF.getFrameInfo();
2273 // HiPE-specific values
2274 const unsigned HipeLeafWords = 24;
2275 const unsigned CCRegisteredArgs = Is64Bit ? 6 : 5;
2276 const unsigned Guaranteed = HipeLeafWords * SlotSize;
2277 unsigned CallerStkArity = MF.getFunction()->arg_size() > CCRegisteredArgs ?
2278 MF.getFunction()->arg_size() - CCRegisteredArgs : 0;
2279 unsigned MaxStack = MFI->getStackSize() + CallerStkArity*SlotSize + SlotSize;
2281 assert(STI.isTargetLinux() &&
2282 "HiPE prologue is only supported on Linux operating systems.");
2284 // Compute the largest caller's frame that is needed to fit the callees'
2285 // frames. This 'MaxStack' is computed from:
2287 // a) the fixed frame size, which is the space needed for all spilled temps,
2288 // b) outgoing on-stack parameter areas, and
2289 // c) the minimum stack space this function needs to make available for the
2290 // functions it calls (a tunable ABI property).
2291 if (MFI->hasCalls()) {
2292 unsigned MoreStackForCalls = 0;
2294 for (MachineFunction::iterator MBBI = MF.begin(), MBBE = MF.end();
2295 MBBI != MBBE; ++MBBI)
2296 for (MachineBasicBlock::iterator MI = MBBI->begin(), ME = MBBI->end();
2301 // Get callee operand.
2302 const MachineOperand &MO = MI->getOperand(0);
2304 // Only take account of global function calls (no closures etc.).
2308 const Function *F = dyn_cast<Function>(MO.getGlobal());
2312 // Do not update 'MaxStack' for primitive and built-in functions
2313 // (encoded with names either starting with "erlang."/"bif_" or not
2314 // having a ".", such as a simple <Module>.<Function>.<Arity>, or an
2315 // "_", such as the BIF "suspend_0") as they are executed on another
2317 if (F->getName().find("erlang.") != StringRef::npos ||
2318 F->getName().find("bif_") != StringRef::npos ||
2319 F->getName().find_first_of("._") == StringRef::npos)
2322 unsigned CalleeStkArity =
2323 F->arg_size() > CCRegisteredArgs ? F->arg_size()-CCRegisteredArgs : 0;
2324 if (HipeLeafWords - 1 > CalleeStkArity)
2325 MoreStackForCalls = std::max(MoreStackForCalls,
2326 (HipeLeafWords - 1 - CalleeStkArity) * SlotSize);
2328 MaxStack += MoreStackForCalls;
2331 // If the stack frame needed is larger than the guaranteed then runtime checks
2332 // and calls to "inc_stack_0" BIF should be inserted in the assembly prologue.
2333 if (MaxStack > Guaranteed) {
2334 MachineBasicBlock *stackCheckMBB = MF.CreateMachineBasicBlock();
2335 MachineBasicBlock *incStackMBB = MF.CreateMachineBasicBlock();
2337 for (const auto &LI : PrologueMBB.liveins()) {
2338 stackCheckMBB->addLiveIn(LI);
2339 incStackMBB->addLiveIn(LI);
2342 MF.push_front(incStackMBB);
2343 MF.push_front(stackCheckMBB);
2345 unsigned ScratchReg, SPReg, PReg, SPLimitOffset;
2346 unsigned LEAop, CMPop, CALLop;
2350 LEAop = X86::LEA64r;
2351 CMPop = X86::CMP64rm;
2352 CALLop = X86::CALL64pcrel32;
2353 SPLimitOffset = 0x90;
2357 LEAop = X86::LEA32r;
2358 CMPop = X86::CMP32rm;
2359 CALLop = X86::CALLpcrel32;
2360 SPLimitOffset = 0x4c;
2363 ScratchReg = GetScratchRegister(Is64Bit, IsLP64, MF, true);
2364 assert(!MF.getRegInfo().isLiveIn(ScratchReg) &&
2365 "HiPE prologue scratch register is live-in");
2367 // Create new MBB for StackCheck:
2368 addRegOffset(BuildMI(stackCheckMBB, DL, TII.get(LEAop), ScratchReg),
2369 SPReg, false, -MaxStack);
2370 // SPLimitOffset is in a fixed heap location (pointed by BP).
2371 addRegOffset(BuildMI(stackCheckMBB, DL, TII.get(CMPop))
2372 .addReg(ScratchReg), PReg, false, SPLimitOffset);
2373 BuildMI(stackCheckMBB, DL, TII.get(X86::JAE_1)).addMBB(&PrologueMBB);
2375 // Create new MBB for IncStack:
2376 BuildMI(incStackMBB, DL, TII.get(CALLop)).
2377 addExternalSymbol("inc_stack_0");
2378 addRegOffset(BuildMI(incStackMBB, DL, TII.get(LEAop), ScratchReg),
2379 SPReg, false, -MaxStack);
2380 addRegOffset(BuildMI(incStackMBB, DL, TII.get(CMPop))
2381 .addReg(ScratchReg), PReg, false, SPLimitOffset);
2382 BuildMI(incStackMBB, DL, TII.get(X86::JLE_1)).addMBB(incStackMBB);
2384 stackCheckMBB->addSuccessor(&PrologueMBB, {99, 100});
2385 stackCheckMBB->addSuccessor(incStackMBB, {1, 100});
2386 incStackMBB->addSuccessor(&PrologueMBB, {99, 100});
2387 incStackMBB->addSuccessor(incStackMBB, {1, 100});
2394 bool X86FrameLowering::adjustStackWithPops(MachineBasicBlock &MBB,
2395 MachineBasicBlock::iterator MBBI, DebugLoc DL, int Offset) const {
2400 if (Offset % SlotSize)
2403 int NumPops = Offset / SlotSize;
2404 // This is only worth it if we have at most 2 pops.
2405 if (NumPops != 1 && NumPops != 2)
2408 // Handle only the trivial case where the adjustment directly follows
2409 // a call. This is the most common one, anyway.
2410 if (MBBI == MBB.begin())
2412 MachineBasicBlock::iterator Prev = std::prev(MBBI);
2413 if (!Prev->isCall() || !Prev->getOperand(1).isRegMask())
2417 unsigned FoundRegs = 0;
2419 auto RegMask = Prev->getOperand(1);
2422 Is64Bit ? X86::GR64_NOREX_NOSPRegClass : X86::GR32_NOREX_NOSPRegClass;
2423 // Try to find up to NumPops free registers.
2424 for (auto Candidate : RegClass) {
2426 // Poor man's liveness:
2427 // Since we're immediately after a call, any register that is clobbered
2428 // by the call and not defined by it can be considered dead.
2429 if (!RegMask.clobbersPhysReg(Candidate))
2433 for (const MachineOperand &MO : Prev->implicit_operands()) {
2434 if (MO.isReg() && MO.isDef() && MO.getReg() == Candidate) {
2443 Regs[FoundRegs++] = Candidate;
2444 if (FoundRegs == (unsigned)NumPops)
2451 // If we found only one free register, but need two, reuse the same one twice.
2452 while (FoundRegs < (unsigned)NumPops)
2453 Regs[FoundRegs++] = Regs[0];
2455 for (int i = 0; i < NumPops; ++i)
2456 BuildMI(MBB, MBBI, DL,
2457 TII.get(STI.is64Bit() ? X86::POP64r : X86::POP32r), Regs[i]);
2462 void X86FrameLowering::
2463 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
2464 MachineBasicBlock::iterator I) const {
2465 bool reserveCallFrame = hasReservedCallFrame(MF);
2466 unsigned Opcode = I->getOpcode();
2467 bool isDestroy = Opcode == TII.getCallFrameDestroyOpcode();
2468 DebugLoc DL = I->getDebugLoc();
2469 uint64_t Amount = !reserveCallFrame ? I->getOperand(0).getImm() : 0;
2470 uint64_t InternalAmt = (isDestroy || Amount) ? I->getOperand(1).getImm() : 0;
2473 if (!reserveCallFrame) {
2474 // If the stack pointer can be changed after prologue, turn the
2475 // adjcallstackup instruction into a 'sub ESP, <amt>' and the
2476 // adjcallstackdown instruction into 'add ESP, <amt>'
2478 // We need to keep the stack aligned properly. To do this, we round the
2479 // amount of space needed for the outgoing arguments up to the next
2480 // alignment boundary.
2481 unsigned StackAlign = getStackAlignment();
2482 Amount = RoundUpToAlignment(Amount, StackAlign);
2484 MachineModuleInfo &MMI = MF.getMMI();
2485 const Function *Fn = MF.getFunction();
2486 bool WindowsCFI = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
2487 bool DwarfCFI = !WindowsCFI &&
2488 (MMI.hasDebugInfo() || Fn->needsUnwindTableEntry());
2490 // If we have any exception handlers in this function, and we adjust
2491 // the SP before calls, we may need to indicate this to the unwinder
2492 // using GNU_ARGS_SIZE. Note that this may be necessary even when
2493 // Amount == 0, because the preceding function may have set a non-0
2495 // TODO: We don't need to reset this between subsequent functions,
2496 // if it didn't change.
2497 bool HasDwarfEHHandlers = !WindowsCFI &&
2498 !MF.getMMI().getLandingPads().empty();
2500 if (HasDwarfEHHandlers && !isDestroy &&
2501 MF.getInfo<X86MachineFunctionInfo>()->getHasPushSequences())
2502 BuildCFI(MBB, I, DL,
2503 MCCFIInstruction::createGnuArgsSize(nullptr, Amount));
2508 // Factor out the amount that gets handled inside the sequence
2509 // (Pushes of argument for frame setup, callee pops for frame destroy)
2510 Amount -= InternalAmt;
2512 // TODO: This is needed only if we require precise CFA.
2513 // If this is a callee-pop calling convention, emit a CFA adjust for
2514 // the amount the callee popped.
2515 if (isDestroy && InternalAmt && DwarfCFI && !hasFP(MF))
2516 BuildCFI(MBB, I, DL,
2517 MCCFIInstruction::createAdjustCfaOffset(nullptr, -InternalAmt));
2520 // Add Amount to SP to destroy a frame, and subtract to setup.
2521 int Offset = isDestroy ? Amount : -Amount;
2523 if (!(Fn->optForMinSize() &&
2524 adjustStackWithPops(MBB, I, DL, Offset)))
2525 BuildStackAdjustment(MBB, I, DL, Offset, /*InEpilogue=*/false);
2528 if (DwarfCFI && !hasFP(MF)) {
2529 // If we don't have FP, but need to generate unwind information,
2530 // we need to set the correct CFA offset after the stack adjustment.
2531 // How much we adjust the CFA offset depends on whether we're emitting
2532 // CFI only for EH purposes or for debugging. EH only requires the CFA
2533 // offset to be correct at each call site, while for debugging we want
2534 // it to be more precise.
2535 int CFAOffset = Amount;
2536 // TODO: When not using precise CFA, we also need to adjust for the
2537 // InternalAmt here.
2540 CFAOffset = isDestroy ? -CFAOffset : CFAOffset;
2541 BuildCFI(MBB, I, DL,
2542 MCCFIInstruction::createAdjustCfaOffset(nullptr, CFAOffset));
2549 if (isDestroy && InternalAmt) {
2550 // If we are performing frame pointer elimination and if the callee pops
2551 // something off the stack pointer, add it back. We do this until we have
2552 // more advanced stack pointer tracking ability.
2553 // We are not tracking the stack pointer adjustment by the callee, so make
2554 // sure we restore the stack pointer immediately after the call, there may
2555 // be spill code inserted between the CALL and ADJCALLSTACKUP instructions.
2556 MachineBasicBlock::iterator B = MBB.begin();
2557 while (I != B && !std::prev(I)->isCall())
2559 BuildStackAdjustment(MBB, I, DL, -InternalAmt, /*InEpilogue=*/false);
2563 bool X86FrameLowering::canUseAsEpilogue(const MachineBasicBlock &MBB) const {
2564 assert(MBB.getParent() && "Block is not attached to a function!");
2566 // Win64 has strict requirements in terms of epilogue and we are
2567 // not taking a chance at messing with them.
2568 // I.e., unless this block is already an exit block, we can't use
2569 // it as an epilogue.
2570 if (STI.isTargetWin64() && !MBB.succ_empty() && !MBB.isReturnBlock())
2573 if (canUseLEAForSPInEpilogue(*MBB.getParent()))
2576 // If we cannot use LEA to adjust SP, we may need to use ADD, which
2577 // clobbers the EFLAGS. Check that we do not need to preserve it,
2578 // otherwise, conservatively assume this is not
2579 // safe to insert the epilogue here.
2580 return !flagsNeedToBePreservedBeforeTheTerminators(MBB);
2583 bool X86FrameLowering::enableShrinkWrapping(const MachineFunction &MF) const {
2584 // If we may need to emit frameless compact unwind information, give
2585 // up as this is currently broken: PR25614.
2586 return MF.getFunction()->hasFnAttribute(Attribute::NoUnwind) || hasFP(MF);
2589 MachineBasicBlock::iterator X86FrameLowering::restoreWin32EHStackPointers(
2590 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
2591 DebugLoc DL, bool RestoreSP) const {
2592 assert(STI.isTargetWindowsMSVC() && "funclets only supported in MSVC env");
2593 assert(STI.isTargetWin32() && "EBP/ESI restoration only required on win32");
2594 assert(STI.is32Bit() && !Uses64BitFramePtr &&
2595 "restoring EBP/ESI on non-32-bit target");
2597 MachineFunction &MF = *MBB.getParent();
2598 unsigned FramePtr = TRI->getFrameRegister(MF);
2599 unsigned BasePtr = TRI->getBaseRegister();
2600 WinEHFuncInfo &FuncInfo = *MF.getWinEHFuncInfo();
2601 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
2602 MachineFrameInfo *MFI = MF.getFrameInfo();
2604 // FIXME: Don't set FrameSetup flag in catchret case.
2606 int FI = FuncInfo.EHRegNodeFrameIndex;
2607 int EHRegSize = MFI->getObjectSize(FI);
2610 // MOV32rm -EHRegSize(%ebp), %esp
2611 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32rm), X86::ESP),
2612 X86::EBP, true, -EHRegSize)
2613 .setMIFlag(MachineInstr::FrameSetup);
2617 int EHRegOffset = getFrameIndexReference(MF, FI, UsedReg);
2618 int EndOffset = -EHRegOffset - EHRegSize;
2619 FuncInfo.EHRegNodeEndOffset = EndOffset;
2621 if (UsedReg == FramePtr) {
2622 // ADD $offset, %ebp
2623 unsigned ADDri = getADDriOpcode(false, EndOffset);
2624 BuildMI(MBB, MBBI, DL, TII.get(ADDri), FramePtr)
2627 .setMIFlag(MachineInstr::FrameSetup)
2630 assert(EndOffset >= 0 &&
2631 "end of registration object above normal EBP position!");
2632 } else if (UsedReg == BasePtr) {
2633 // LEA offset(%ebp), %esi
2634 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::LEA32r), BasePtr),
2635 FramePtr, false, EndOffset)
2636 .setMIFlag(MachineInstr::FrameSetup);
2637 // MOV32rm SavedEBPOffset(%esi), %ebp
2638 assert(X86FI->getHasSEHFramePtrSave());
2640 getFrameIndexReference(MF, X86FI->getSEHFramePtrSaveIndex(), UsedReg);
2641 assert(UsedReg == BasePtr);
2642 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32rm), FramePtr),
2643 UsedReg, true, Offset)
2644 .setMIFlag(MachineInstr::FrameSetup);
2646 llvm_unreachable("32-bit frames with WinEH must use FramePtr or BasePtr");
2651 unsigned X86FrameLowering::getWinEHParentFrameOffset(const MachineFunction &MF) const {
2652 // RDX, the parent frame pointer, is homed into 16(%rsp) in the prologue.
2653 unsigned Offset = 16;
2654 // RBP is immediately pushed.
2656 // All callee-saved registers are then pushed.
2657 Offset += MF.getInfo<X86MachineFunctionInfo>()->getCalleeSavedFrameSize();
2658 // Every funclet allocates enough stack space for the largest outgoing call.
2659 Offset += getWinEHFuncletFrameSize(MF);
2663 void X86FrameLowering::processFunctionBeforeFrameFinalized(
2664 MachineFunction &MF, RegScavenger *RS) const {
2665 // If this function isn't doing Win64-style C++ EH, we don't need to do
2667 const Function *Fn = MF.getFunction();
2668 if (!STI.is64Bit() || !MF.getMMI().hasEHFunclets() ||
2669 classifyEHPersonality(Fn->getPersonalityFn()) != EHPersonality::MSVC_CXX)
2672 // Win64 C++ EH needs to allocate the UnwindHelp object at some fixed offset
2673 // relative to RSP after the prologue. Find the offset of the last fixed
2674 // object, so that we can allocate a slot immediately following it. If there
2675 // were no fixed objects, use offset -SlotSize, which is immediately after the
2676 // return address. Fixed objects have negative frame indices.
2677 MachineFrameInfo *MFI = MF.getFrameInfo();
2678 int64_t MinFixedObjOffset = -SlotSize;
2679 for (int I = MFI->getObjectIndexBegin(); I < 0; ++I)
2680 MinFixedObjOffset = std::min(MinFixedObjOffset, MFI->getObjectOffset(I));
2682 int64_t UnwindHelpOffset = MinFixedObjOffset - SlotSize;
2684 MFI->CreateFixedObject(SlotSize, UnwindHelpOffset, /*Immutable=*/false);
2685 MF.getWinEHFuncInfo()->UnwindHelpFrameIdx = UnwindHelpFI;
2687 // Store -2 into UnwindHelp on function entry. We have to scan forwards past
2688 // other frame setup instructions.
2689 MachineBasicBlock &MBB = MF.front();
2690 auto MBBI = MBB.begin();
2691 while (MBBI != MBB.end() && MBBI->getFlag(MachineInstr::FrameSetup))
2694 DebugLoc DL = MBB.findDebugLoc(MBBI);
2695 addFrameReference(BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64mi32)),