1 //=======- X86FrameLowering.cpp - X86 Frame Information ------------*- C++ -*-====//
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 "X86TargetMachine.h"
19 #include "llvm/Function.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineModuleInfo.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/Target/TargetOptions.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/ADT/SmallSet.h"
32 // FIXME: completely move here.
33 extern cl::opt<bool> ForceStackAlign;
35 bool X86FrameLowering::hasReservedCallFrame(const MachineFunction &MF) const {
36 return !MF.getFrameInfo()->hasVarSizedObjects();
39 /// hasFP - Return true if the specified function should have a dedicated frame
40 /// pointer register. This is true if the function has variable sized allocas
41 /// or if frame pointer elimination is disabled.
42 bool X86FrameLowering::hasFP(const MachineFunction &MF) const {
43 const MachineFrameInfo *MFI = MF.getFrameInfo();
44 const MachineModuleInfo &MMI = MF.getMMI();
45 const TargetRegisterInfo *RI = TM.getRegisterInfo();
47 return (DisableFramePointerElim(MF) ||
48 RI->needsStackRealignment(MF) ||
49 MFI->hasVarSizedObjects() ||
50 MFI->isFrameAddressTaken() ||
51 MF.getInfo<X86MachineFunctionInfo>()->getForceFramePointer() ||
52 MMI.callsUnwindInit());
55 static unsigned getSUBriOpcode(unsigned is64Bit, int64_t Imm) {
59 return X86::SUB64ri32;
67 static unsigned getADDriOpcode(unsigned is64Bit, int64_t Imm) {
71 return X86::ADD64ri32;
79 /// findDeadCallerSavedReg - Return a caller-saved register that isn't live
80 /// when it reaches the "return" instruction. We can then pop a stack object
81 /// to this register without worry about clobbering it.
82 static unsigned findDeadCallerSavedReg(MachineBasicBlock &MBB,
83 MachineBasicBlock::iterator &MBBI,
84 const TargetRegisterInfo &TRI,
86 const MachineFunction *MF = MBB.getParent();
87 const Function *F = MF->getFunction();
88 if (!F || MF->getMMI().callsEHReturn())
91 static const unsigned CallerSavedRegs32Bit[] = {
92 X86::EAX, X86::EDX, X86::ECX
95 static const unsigned CallerSavedRegs64Bit[] = {
96 X86::RAX, X86::RDX, X86::RCX, X86::RSI, X86::RDI,
97 X86::R8, X86::R9, X86::R10, X86::R11
100 unsigned Opc = MBBI->getOpcode();
105 case X86::TCRETURNdi:
106 case X86::TCRETURNri:
107 case X86::TCRETURNmi:
108 case X86::TCRETURNdi64:
109 case X86::TCRETURNri64:
110 case X86::TCRETURNmi64:
112 case X86::EH_RETURN64: {
113 SmallSet<unsigned, 8> Uses;
114 for (unsigned i = 0, e = MBBI->getNumOperands(); i != e; ++i) {
115 MachineOperand &MO = MBBI->getOperand(i);
116 if (!MO.isReg() || MO.isDef())
118 unsigned Reg = MO.getReg();
121 for (const unsigned *AsI = TRI.getOverlaps(Reg); *AsI; ++AsI)
125 const unsigned *CS = Is64Bit ? CallerSavedRegs64Bit : CallerSavedRegs32Bit;
127 if (!Uses.count(*CS))
136 /// emitSPUpdate - Emit a series of instructions to increment / decrement the
137 /// stack pointer by a constant value.
139 void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
140 unsigned StackPtr, int64_t NumBytes,
141 bool Is64Bit, const TargetInstrInfo &TII,
142 const TargetRegisterInfo &TRI) {
143 bool isSub = NumBytes < 0;
144 uint64_t Offset = isSub ? -NumBytes : NumBytes;
145 unsigned Opc = isSub ?
146 getSUBriOpcode(Is64Bit, Offset) :
147 getADDriOpcode(Is64Bit, Offset);
148 uint64_t Chunk = (1LL << 31) - 1;
149 DebugLoc DL = MBB.findDebugLoc(MBBI);
152 uint64_t ThisVal = (Offset > Chunk) ? Chunk : Offset;
153 if (ThisVal == (Is64Bit ? 8 : 4)) {
154 // Use push / pop instead.
156 ? (unsigned)(Is64Bit ? X86::RAX : X86::EAX)
157 : findDeadCallerSavedReg(MBB, MBBI, TRI, Is64Bit);
160 ? (Is64Bit ? X86::PUSH64r : X86::PUSH32r)
161 : (Is64Bit ? X86::POP64r : X86::POP32r);
162 BuildMI(MBB, MBBI, DL, TII.get(Opc))
163 .addReg(Reg, getDefRegState(!isSub) | getUndefRegState(isSub));
170 BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr)
173 MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
178 /// mergeSPUpdatesUp - Merge two stack-manipulating instructions upper iterator.
180 void mergeSPUpdatesUp(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
181 unsigned StackPtr, uint64_t *NumBytes = NULL) {
182 if (MBBI == MBB.begin()) return;
184 MachineBasicBlock::iterator PI = prior(MBBI);
185 unsigned Opc = PI->getOpcode();
186 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
187 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) &&
188 PI->getOperand(0).getReg() == StackPtr) {
190 *NumBytes += PI->getOperand(2).getImm();
192 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
193 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
194 PI->getOperand(0).getReg() == StackPtr) {
196 *NumBytes -= PI->getOperand(2).getImm();
201 /// mergeSPUpdatesDown - Merge two stack-manipulating instructions lower iterator.
203 void mergeSPUpdatesDown(MachineBasicBlock &MBB,
204 MachineBasicBlock::iterator &MBBI,
205 unsigned StackPtr, uint64_t *NumBytes = NULL) {
206 // FIXME: THIS ISN'T RUN!!!
209 if (MBBI == MBB.end()) return;
211 MachineBasicBlock::iterator NI = llvm::next(MBBI);
212 if (NI == MBB.end()) return;
214 unsigned Opc = NI->getOpcode();
215 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
216 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) &&
217 NI->getOperand(0).getReg() == StackPtr) {
219 *NumBytes -= NI->getOperand(2).getImm();
222 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
223 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
224 NI->getOperand(0).getReg() == StackPtr) {
226 *NumBytes += NI->getOperand(2).getImm();
232 /// mergeSPUpdates - Checks the instruction before/after the passed
233 /// instruction. If it is an ADD/SUB instruction it is deleted argument and the
234 /// stack adjustment is returned as a positive value for ADD and a negative for
236 static int mergeSPUpdates(MachineBasicBlock &MBB,
237 MachineBasicBlock::iterator &MBBI,
239 bool doMergeWithPrevious) {
240 if ((doMergeWithPrevious && MBBI == MBB.begin()) ||
241 (!doMergeWithPrevious && MBBI == MBB.end()))
244 MachineBasicBlock::iterator PI = doMergeWithPrevious ? prior(MBBI) : MBBI;
245 MachineBasicBlock::iterator NI = doMergeWithPrevious ? 0 : llvm::next(MBBI);
246 unsigned Opc = PI->getOpcode();
249 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
250 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) &&
251 PI->getOperand(0).getReg() == StackPtr){
252 Offset += PI->getOperand(2).getImm();
254 if (!doMergeWithPrevious) MBBI = NI;
255 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
256 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
257 PI->getOperand(0).getReg() == StackPtr) {
258 Offset -= PI->getOperand(2).getImm();
260 if (!doMergeWithPrevious) MBBI = NI;
266 static bool isEAXLiveIn(MachineFunction &MF) {
267 for (MachineRegisterInfo::livein_iterator II = MF.getRegInfo().livein_begin(),
268 EE = MF.getRegInfo().livein_end(); II != EE; ++II) {
269 unsigned Reg = II->first;
271 if (Reg == X86::EAX || Reg == X86::AX ||
272 Reg == X86::AH || Reg == X86::AL)
279 void X86FrameLowering::emitCalleeSavedFrameMoves(MachineFunction &MF,
281 unsigned FramePtr) const {
282 MachineFrameInfo *MFI = MF.getFrameInfo();
283 MachineModuleInfo &MMI = MF.getMMI();
285 // Add callee saved registers to move list.
286 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
287 if (CSI.empty()) return;
289 std::vector<MachineMove> &Moves = MMI.getFrameMoves();
290 const TargetData *TD = TM.getTargetData();
291 bool HasFP = hasFP(MF);
293 // Calculate amount of bytes used for return address storing.
295 (TM.getFrameLowering()->getStackGrowthDirection() ==
296 TargetFrameLowering::StackGrowsUp ?
297 TD->getPointerSize() : -TD->getPointerSize());
299 // FIXME: This is dirty hack. The code itself is pretty mess right now.
300 // It should be rewritten from scratch and generalized sometimes.
302 // Determine maximum offset (minumum due to stack growth).
303 int64_t MaxOffset = 0;
304 for (std::vector<CalleeSavedInfo>::const_iterator
305 I = CSI.begin(), E = CSI.end(); I != E; ++I)
306 MaxOffset = std::min(MaxOffset,
307 MFI->getObjectOffset(I->getFrameIdx()));
309 // Calculate offsets.
310 int64_t saveAreaOffset = (HasFP ? 3 : 2) * stackGrowth;
311 for (std::vector<CalleeSavedInfo>::const_iterator
312 I = CSI.begin(), E = CSI.end(); I != E; ++I) {
313 int64_t Offset = MFI->getObjectOffset(I->getFrameIdx());
314 unsigned Reg = I->getReg();
315 Offset = MaxOffset - Offset + saveAreaOffset;
317 // Don't output a new machine move if we're re-saving the frame
318 // pointer. This happens when the PrologEpilogInserter has inserted an extra
319 // "PUSH" of the frame pointer -- the "emitPrologue" method automatically
320 // generates one when frame pointers are used. If we generate a "machine
321 // move" for this extra "PUSH", the linker will lose track of the fact that
322 // the frame pointer should have the value of the first "PUSH" when it's
325 // FIXME: This looks inelegant. It's possibly correct, but it's covering up
326 // another bug. I.e., one where we generate a prolog like this:
334 // The immediate re-push of EBP is unnecessary. At the least, it's an
335 // optimization bug. EBP can be used as a scratch register in certain
336 // cases, but probably not when we have a frame pointer.
337 if (HasFP && FramePtr == Reg)
340 MachineLocation CSDst(MachineLocation::VirtualFP, Offset);
341 MachineLocation CSSrc(Reg);
342 Moves.push_back(MachineMove(Label, CSDst, CSSrc));
346 /// emitPrologue - Push callee-saved registers onto the stack, which
347 /// automatically adjust the stack pointer. Adjust the stack pointer to allocate
348 /// space for local variables. Also emit labels used by the exception handler to
349 /// generate the exception handling frames.
350 void X86FrameLowering::emitPrologue(MachineFunction &MF) const {
351 MachineBasicBlock &MBB = MF.front(); // Prologue goes in entry BB.
352 MachineBasicBlock::iterator MBBI = MBB.begin();
353 MachineFrameInfo *MFI = MF.getFrameInfo();
354 const Function *Fn = MF.getFunction();
355 const X86RegisterInfo *RegInfo = TM.getRegisterInfo();
356 const X86InstrInfo &TII = *TM.getInstrInfo();
357 MachineModuleInfo &MMI = MF.getMMI();
358 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
359 bool needsFrameMoves = MMI.hasDebugInfo() ||
360 !Fn->doesNotThrow() || UnwindTablesMandatory;
361 uint64_t MaxAlign = MFI->getMaxAlignment(); // Desired stack alignment.
362 uint64_t StackSize = MFI->getStackSize(); // Number of bytes to allocate.
363 bool HasFP = hasFP(MF);
364 bool Is64Bit = STI.is64Bit();
365 bool IsWin64 = STI.isTargetWin64();
366 unsigned StackAlign = getStackAlignment();
367 unsigned SlotSize = RegInfo->getSlotSize();
368 unsigned FramePtr = RegInfo->getFrameRegister(MF);
369 unsigned StackPtr = RegInfo->getStackRegister();
373 // If we're forcing a stack realignment we can't rely on just the frame
374 // info, we need to know the ABI stack alignment as well in case we
375 // have a call out. Otherwise just make sure we have some alignment - we'll
376 // go with the minimum SlotSize.
377 if (ForceStackAlign) {
379 MaxAlign = (StackAlign > MaxAlign) ? StackAlign : MaxAlign;
380 else if (MaxAlign < SlotSize)
384 // Add RETADDR move area to callee saved frame size.
385 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
386 if (TailCallReturnAddrDelta < 0)
387 X86FI->setCalleeSavedFrameSize(
388 X86FI->getCalleeSavedFrameSize() - TailCallReturnAddrDelta);
390 // If this is x86-64 and the Red Zone is not disabled, if we are a leaf
391 // function, and use up to 128 bytes of stack space, don't have a frame
392 // pointer, calls, or dynamic alloca then we do not need to adjust the
393 // stack pointer (we fit in the Red Zone).
394 if (Is64Bit && !Fn->hasFnAttr(Attribute::NoRedZone) &&
395 !RegInfo->needsStackRealignment(MF) &&
396 !MFI->hasVarSizedObjects() && // No dynamic alloca.
397 !MFI->adjustsStack() && // No calls.
398 !IsWin64) { // Win64 has no Red Zone
399 uint64_t MinSize = X86FI->getCalleeSavedFrameSize();
400 if (HasFP) MinSize += SlotSize;
401 StackSize = std::max(MinSize, StackSize > 128 ? StackSize - 128 : 0);
402 MFI->setStackSize(StackSize);
403 } else if (IsWin64) {
404 // We need to always allocate 32 bytes as register spill area.
405 // FIXME: We might reuse these 32 bytes for leaf functions.
407 MFI->setStackSize(StackSize);
410 // Insert stack pointer adjustment for later moving of return addr. Only
411 // applies to tail call optimized functions where the callee argument stack
412 // size is bigger than the callers.
413 if (TailCallReturnAddrDelta < 0) {
415 BuildMI(MBB, MBBI, DL,
416 TII.get(getSUBriOpcode(Is64Bit, -TailCallReturnAddrDelta)),
419 .addImm(-TailCallReturnAddrDelta);
420 MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
423 // Mapping for machine moves:
425 // DST: VirtualFP AND
426 // SRC: VirtualFP => DW_CFA_def_cfa_offset
427 // ELSE => DW_CFA_def_cfa
429 // SRC: VirtualFP AND
430 // DST: Register => DW_CFA_def_cfa_register
433 // OFFSET < 0 => DW_CFA_offset_extended_sf
434 // REG < 64 => DW_CFA_offset + Reg
435 // ELSE => DW_CFA_offset_extended
437 std::vector<MachineMove> &Moves = MMI.getFrameMoves();
438 const TargetData *TD = MF.getTarget().getTargetData();
439 uint64_t NumBytes = 0;
440 int stackGrowth = -TD->getPointerSize();
443 // Calculate required stack adjustment.
444 uint64_t FrameSize = StackSize - SlotSize;
445 if (RegInfo->needsStackRealignment(MF))
446 FrameSize = (FrameSize + MaxAlign - 1) / MaxAlign * MaxAlign;
448 NumBytes = FrameSize - X86FI->getCalleeSavedFrameSize();
450 // Get the offset of the stack slot for the EBP register, which is
451 // guaranteed to be the last slot by processFunctionBeforeFrameFinalized.
452 // Update the frame offset adjustment.
453 MFI->setOffsetAdjustment(-NumBytes);
455 // Save EBP/RBP into the appropriate stack slot.
456 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::PUSH64r : X86::PUSH32r))
457 .addReg(FramePtr, RegState::Kill);
459 if (needsFrameMoves) {
460 // Mark the place where EBP/RBP was saved.
461 MCSymbol *FrameLabel = MMI.getContext().CreateTempSymbol();
462 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL)).addSym(FrameLabel);
464 // Define the current CFA rule to use the provided offset.
466 MachineLocation SPDst(MachineLocation::VirtualFP);
467 MachineLocation SPSrc(MachineLocation::VirtualFP, 2 * stackGrowth);
468 Moves.push_back(MachineMove(FrameLabel, SPDst, SPSrc));
470 // FIXME: Verify & implement for FP
471 MachineLocation SPDst(StackPtr);
472 MachineLocation SPSrc(StackPtr, stackGrowth);
473 Moves.push_back(MachineMove(FrameLabel, SPDst, SPSrc));
476 // Change the rule for the FramePtr to be an "offset" rule.
477 MachineLocation FPDst(MachineLocation::VirtualFP, 2 * stackGrowth);
478 MachineLocation FPSrc(FramePtr);
479 Moves.push_back(MachineMove(FrameLabel, FPDst, FPSrc));
482 // Update EBP with the new base value...
483 BuildMI(MBB, MBBI, DL,
484 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), FramePtr)
487 if (needsFrameMoves) {
488 // Mark effective beginning of when frame pointer becomes valid.
489 MCSymbol *FrameLabel = MMI.getContext().CreateTempSymbol();
490 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL)).addSym(FrameLabel);
492 // Define the current CFA to use the EBP/RBP register.
493 MachineLocation FPDst(FramePtr);
494 MachineLocation FPSrc(MachineLocation::VirtualFP);
495 Moves.push_back(MachineMove(FrameLabel, FPDst, FPSrc));
498 // Mark the FramePtr as live-in in every block except the entry.
499 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
501 I->addLiveIn(FramePtr);
504 if (RegInfo->needsStackRealignment(MF)) {
506 BuildMI(MBB, MBBI, DL,
507 TII.get(Is64Bit ? X86::AND64ri32 : X86::AND32ri),
508 StackPtr).addReg(StackPtr).addImm(-MaxAlign);
510 // The EFLAGS implicit def is dead.
511 MI->getOperand(3).setIsDead();
514 NumBytes = StackSize - X86FI->getCalleeSavedFrameSize();
517 // Skip the callee-saved push instructions.
518 bool PushedRegs = false;
519 int StackOffset = 2 * stackGrowth;
521 while (MBBI != MBB.end() &&
522 (MBBI->getOpcode() == X86::PUSH32r ||
523 MBBI->getOpcode() == X86::PUSH64r)) {
527 if (!HasFP && needsFrameMoves) {
528 // Mark callee-saved push instruction.
529 MCSymbol *Label = MMI.getContext().CreateTempSymbol();
530 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL)).addSym(Label);
532 // Define the current CFA rule to use the provided offset.
533 unsigned Ptr = StackSize ?
534 MachineLocation::VirtualFP : StackPtr;
535 MachineLocation SPDst(Ptr);
536 MachineLocation SPSrc(Ptr, StackOffset);
537 Moves.push_back(MachineMove(Label, SPDst, SPSrc));
538 StackOffset += stackGrowth;
542 DL = MBB.findDebugLoc(MBBI);
544 // If there is an SUB32ri of ESP immediately before this instruction, merge
545 // the two. This can be the case when tail call elimination is enabled and
546 // the callee has more arguments then the caller.
547 NumBytes -= mergeSPUpdates(MBB, MBBI, StackPtr, true);
549 // If there is an ADD32ri or SUB32ri of ESP immediately after this
550 // instruction, merge the two instructions.
551 mergeSPUpdatesDown(MBB, MBBI, StackPtr, &NumBytes);
553 // Adjust stack pointer: ESP -= numbytes.
555 // Windows and cygwin/mingw require a prologue helper routine when allocating
556 // more than 4K bytes on the stack. Windows uses __chkstk and cygwin/mingw
557 // uses __alloca. __alloca and the 32-bit version of __chkstk will probe the
558 // stack and adjust the stack pointer in one go. The 64-bit version of
559 // __chkstk is only responsible for probing the stack. The 64-bit prologue is
560 // responsible for adjusting the stack pointer. Touching the stack at 4K
561 // increments is necessary to ensure that the guard pages used by the OS
562 // virtual memory manager are allocated in correct sequence.
563 if (NumBytes >= 4096 && (STI.isTargetCygMing() || STI.isTargetWin32())) {
564 // Check whether EAX is livein for this function.
565 bool isEAXAlive = isEAXLiveIn(MF);
567 const char *StackProbeSymbol =
568 STI.isTargetWindows() ? "_chkstk" : "_alloca";
569 if (Is64Bit && STI.isTargetCygMing())
570 StackProbeSymbol = "__chkstk";
571 unsigned CallOp = Is64Bit ? X86::CALL64pcrel32 : X86::CALLpcrel32;
573 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
575 BuildMI(MBB, MBBI, DL, TII.get(CallOp))
576 .addExternalSymbol(StackProbeSymbol)
577 .addReg(StackPtr, RegState::Define | RegState::Implicit)
578 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
581 BuildMI(MBB, MBBI, DL, TII.get(X86::PUSH32r))
582 .addReg(X86::EAX, RegState::Kill);
584 // Allocate NumBytes-4 bytes on stack. We'll also use 4 already
585 // allocated bytes for EAX.
586 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
587 .addImm(NumBytes - 4);
588 BuildMI(MBB, MBBI, DL, TII.get(CallOp))
589 .addExternalSymbol(StackProbeSymbol)
590 .addReg(StackPtr, RegState::Define | RegState::Implicit)
591 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
594 MachineInstr *MI = addRegOffset(BuildMI(MF, DL, TII.get(X86::MOV32rm),
596 StackPtr, false, NumBytes - 4);
597 MBB.insert(MBBI, MI);
599 } else if (NumBytes >= 4096 && STI.isTargetWin64()) {
600 // Sanity check that EAX is not livein for this function. It should
601 // should not be, so throw an assert.
602 assert(!isEAXLiveIn(MF) && "EAX is livein in the Win64 case!");
604 // Handle the 64-bit Windows ABI case where we need to call __chkstk.
605 // Function prologue is responsible for adjusting the stack pointer.
606 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
608 BuildMI(MBB, MBBI, DL, TII.get(X86::WINCALL64pcrel32))
609 .addExternalSymbol("__chkstk")
610 .addReg(StackPtr, RegState::Define | RegState::Implicit);
611 emitSPUpdate(MBB, MBBI, StackPtr, -(int64_t)NumBytes, Is64Bit,
614 emitSPUpdate(MBB, MBBI, StackPtr, -(int64_t)NumBytes, Is64Bit,
617 if ((NumBytes || PushedRegs) && needsFrameMoves) {
618 // Mark end of stack pointer adjustment.
619 MCSymbol *Label = MMI.getContext().CreateTempSymbol();
620 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL)).addSym(Label);
622 if (!HasFP && NumBytes) {
623 // Define the current CFA rule to use the provided offset.
625 MachineLocation SPDst(MachineLocation::VirtualFP);
626 MachineLocation SPSrc(MachineLocation::VirtualFP,
627 -StackSize + stackGrowth);
628 Moves.push_back(MachineMove(Label, SPDst, SPSrc));
630 // FIXME: Verify & implement for FP
631 MachineLocation SPDst(StackPtr);
632 MachineLocation SPSrc(StackPtr, stackGrowth);
633 Moves.push_back(MachineMove(Label, SPDst, SPSrc));
637 // Emit DWARF info specifying the offsets of the callee-saved registers.
639 emitCalleeSavedFrameMoves(MF, Label, HasFP ? FramePtr : StackPtr);
643 void X86FrameLowering::emitEpilogue(MachineFunction &MF,
644 MachineBasicBlock &MBB) const {
645 const MachineFrameInfo *MFI = MF.getFrameInfo();
646 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
647 const X86RegisterInfo *RegInfo = TM.getRegisterInfo();
648 const X86InstrInfo &TII = *TM.getInstrInfo();
649 MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr();
650 assert(MBBI != MBB.end() && "Returning block has no instructions");
651 unsigned RetOpcode = MBBI->getOpcode();
652 DebugLoc DL = MBBI->getDebugLoc();
653 bool Is64Bit = STI.is64Bit();
654 unsigned StackAlign = getStackAlignment();
655 unsigned SlotSize = RegInfo->getSlotSize();
656 unsigned FramePtr = RegInfo->getFrameRegister(MF);
657 unsigned StackPtr = RegInfo->getStackRegister();
661 llvm_unreachable("Can only insert epilog into returning blocks");
664 case X86::TCRETURNdi:
665 case X86::TCRETURNri:
666 case X86::TCRETURNmi:
667 case X86::TCRETURNdi64:
668 case X86::TCRETURNri64:
669 case X86::TCRETURNmi64:
671 case X86::EH_RETURN64:
672 break; // These are ok
675 // Get the number of bytes to allocate from the FrameInfo.
676 uint64_t StackSize = MFI->getStackSize();
677 uint64_t MaxAlign = MFI->getMaxAlignment();
678 unsigned CSSize = X86FI->getCalleeSavedFrameSize();
679 uint64_t NumBytes = 0;
681 // If we're forcing a stack realignment we can't rely on just the frame
682 // info, we need to know the ABI stack alignment as well in case we
683 // have a call out. Otherwise just make sure we have some alignment - we'll
684 // go with the minimum.
685 if (ForceStackAlign) {
687 MaxAlign = (StackAlign > MaxAlign) ? StackAlign : MaxAlign;
689 MaxAlign = MaxAlign ? MaxAlign : 4;
693 // Calculate required stack adjustment.
694 uint64_t FrameSize = StackSize - SlotSize;
695 if (RegInfo->needsStackRealignment(MF))
696 FrameSize = (FrameSize + MaxAlign - 1)/MaxAlign*MaxAlign;
698 NumBytes = FrameSize - CSSize;
701 BuildMI(MBB, MBBI, DL,
702 TII.get(Is64Bit ? X86::POP64r : X86::POP32r), FramePtr);
704 NumBytes = StackSize - CSSize;
707 // Skip the callee-saved pop instructions.
708 MachineBasicBlock::iterator LastCSPop = MBBI;
709 while (MBBI != MBB.begin()) {
710 MachineBasicBlock::iterator PI = prior(MBBI);
711 unsigned Opc = PI->getOpcode();
713 if (Opc != X86::POP32r && Opc != X86::POP64r && Opc != X86::DBG_VALUE &&
714 !PI->getDesc().isTerminator())
720 DL = MBBI->getDebugLoc();
722 // If there is an ADD32ri or SUB32ri of ESP immediately before this
723 // instruction, merge the two instructions.
724 if (NumBytes || MFI->hasVarSizedObjects())
725 mergeSPUpdatesUp(MBB, MBBI, StackPtr, &NumBytes);
727 // If dynamic alloca is used, then reset esp to point to the last callee-saved
728 // slot before popping them off! Same applies for the case, when stack was
730 if (RegInfo->needsStackRealignment(MF)) {
731 // We cannot use LEA here, because stack pointer was realigned. We need to
732 // deallocate local frame back.
734 emitSPUpdate(MBB, MBBI, StackPtr, NumBytes, Is64Bit, TII, *RegInfo);
735 MBBI = prior(LastCSPop);
738 BuildMI(MBB, MBBI, DL,
739 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr),
740 StackPtr).addReg(FramePtr);
741 } else if (MFI->hasVarSizedObjects()) {
743 unsigned Opc = Is64Bit ? X86::LEA64r : X86::LEA32r;
745 addRegOffset(BuildMI(MF, DL, TII.get(Opc), StackPtr),
746 FramePtr, false, -CSSize);
747 MBB.insert(MBBI, MI);
749 BuildMI(MBB, MBBI, DL,
750 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), StackPtr)
753 } else if (NumBytes) {
754 // Adjust stack pointer back: ESP += numbytes.
755 emitSPUpdate(MBB, MBBI, StackPtr, NumBytes, Is64Bit, TII, *RegInfo);
758 // We're returning from function via eh_return.
759 if (RetOpcode == X86::EH_RETURN || RetOpcode == X86::EH_RETURN64) {
760 MBBI = MBB.getLastNonDebugInstr();
761 MachineOperand &DestAddr = MBBI->getOperand(0);
762 assert(DestAddr.isReg() && "Offset should be in register!");
763 BuildMI(MBB, MBBI, DL,
764 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr),
765 StackPtr).addReg(DestAddr.getReg());
766 } else if (RetOpcode == X86::TCRETURNri || RetOpcode == X86::TCRETURNdi ||
767 RetOpcode == X86::TCRETURNmi ||
768 RetOpcode == X86::TCRETURNri64 || RetOpcode == X86::TCRETURNdi64 ||
769 RetOpcode == X86::TCRETURNmi64) {
770 bool isMem = RetOpcode == X86::TCRETURNmi || RetOpcode == X86::TCRETURNmi64;
771 // Tail call return: adjust the stack pointer and jump to callee.
772 MBBI = MBB.getLastNonDebugInstr();
773 MachineOperand &JumpTarget = MBBI->getOperand(0);
774 MachineOperand &StackAdjust = MBBI->getOperand(isMem ? 5 : 1);
775 assert(StackAdjust.isImm() && "Expecting immediate value.");
777 // Adjust stack pointer.
778 int StackAdj = StackAdjust.getImm();
779 int MaxTCDelta = X86FI->getTCReturnAddrDelta();
781 assert(MaxTCDelta <= 0 && "MaxTCDelta should never be positive");
783 // Incoporate the retaddr area.
784 Offset = StackAdj-MaxTCDelta;
785 assert(Offset >= 0 && "Offset should never be negative");
788 // Check for possible merge with preceeding ADD instruction.
789 Offset += mergeSPUpdates(MBB, MBBI, StackPtr, true);
790 emitSPUpdate(MBB, MBBI, StackPtr, Offset, Is64Bit, TII, *RegInfo);
793 // Jump to label or value in register.
794 if (RetOpcode == X86::TCRETURNdi || RetOpcode == X86::TCRETURNdi64) {
795 MachineInstrBuilder MIB =
796 BuildMI(MBB, MBBI, DL, TII.get((RetOpcode == X86::TCRETURNdi)
797 ? X86::TAILJMPd : X86::TAILJMPd64));
798 if (JumpTarget.isGlobal())
799 MIB.addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(),
800 JumpTarget.getTargetFlags());
802 assert(JumpTarget.isSymbol());
803 MIB.addExternalSymbol(JumpTarget.getSymbolName(),
804 JumpTarget.getTargetFlags());
806 } else if (RetOpcode == X86::TCRETURNmi || RetOpcode == X86::TCRETURNmi64) {
807 MachineInstrBuilder MIB =
808 BuildMI(MBB, MBBI, DL, TII.get((RetOpcode == X86::TCRETURNmi)
809 ? X86::TAILJMPm : X86::TAILJMPm64));
810 for (unsigned i = 0; i != 5; ++i)
811 MIB.addOperand(MBBI->getOperand(i));
812 } else if (RetOpcode == X86::TCRETURNri64) {
813 BuildMI(MBB, MBBI, DL, TII.get(X86::TAILJMPr64)).
814 addReg(JumpTarget.getReg(), RegState::Kill);
816 BuildMI(MBB, MBBI, DL, TII.get(X86::TAILJMPr)).
817 addReg(JumpTarget.getReg(), RegState::Kill);
820 MachineInstr *NewMI = prior(MBBI);
821 for (unsigned i = 2, e = MBBI->getNumOperands(); i != e; ++i)
822 NewMI->addOperand(MBBI->getOperand(i));
824 // Delete the pseudo instruction TCRETURN.
826 } else if ((RetOpcode == X86::RET || RetOpcode == X86::RETI) &&
827 (X86FI->getTCReturnAddrDelta() < 0)) {
828 // Add the return addr area delta back since we are not tail calling.
829 int delta = -1*X86FI->getTCReturnAddrDelta();
830 MBBI = MBB.getLastNonDebugInstr();
832 // Check for possible merge with preceeding ADD instruction.
833 delta += mergeSPUpdates(MBB, MBBI, StackPtr, true);
834 emitSPUpdate(MBB, MBBI, StackPtr, delta, Is64Bit, TII, *RegInfo);
839 X86FrameLowering::getInitialFrameState(std::vector<MachineMove> &Moves) const {
840 // Calculate amount of bytes used for return address storing
841 int stackGrowth = (STI.is64Bit() ? -8 : -4);
842 const X86RegisterInfo *RI = TM.getRegisterInfo();
844 // Initial state of the frame pointer is esp+stackGrowth.
845 MachineLocation Dst(MachineLocation::VirtualFP);
846 MachineLocation Src(RI->getStackRegister(), stackGrowth);
847 Moves.push_back(MachineMove(0, Dst, Src));
849 // Add return address to move list
850 MachineLocation CSDst(RI->getStackRegister(), stackGrowth);
851 MachineLocation CSSrc(RI->getRARegister());
852 Moves.push_back(MachineMove(0, CSDst, CSSrc));
855 int X86FrameLowering::getFrameIndexOffset(const MachineFunction &MF, int FI) const {
856 const X86RegisterInfo *RI =
857 static_cast<const X86RegisterInfo*>(MF.getTarget().getRegisterInfo());
858 const MachineFrameInfo *MFI = MF.getFrameInfo();
859 int Offset = MFI->getObjectOffset(FI) - getOffsetOfLocalArea();
860 uint64_t StackSize = MFI->getStackSize();
862 if (RI->needsStackRealignment(MF)) {
864 // Skip the saved EBP.
865 Offset += RI->getSlotSize();
867 unsigned Align = MFI->getObjectAlignment(FI);
868 assert((-(Offset + StackSize)) % Align == 0);
870 return Offset + StackSize;
872 // FIXME: Support tail calls
875 return Offset + StackSize;
877 // Skip the saved EBP.
878 Offset += RI->getSlotSize();
880 // Skip the RETADDR move area
881 const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
882 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
883 if (TailCallReturnAddrDelta < 0)
884 Offset -= TailCallReturnAddrDelta;
890 bool X86FrameLowering::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
891 MachineBasicBlock::iterator MI,
892 const std::vector<CalleeSavedInfo> &CSI,
893 const TargetRegisterInfo *TRI) const {
897 DebugLoc DL = MBB.findDebugLoc(MI);
899 MachineFunction &MF = *MBB.getParent();
901 bool isWin64 = STI.isTargetWin64();
902 unsigned SlotSize = STI.is64Bit() ? 8 : 4;
903 unsigned FPReg = TRI->getFrameRegister(MF);
904 unsigned CalleeFrameSize = 0;
906 const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
907 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
909 unsigned Opc = STI.is64Bit() ? X86::PUSH64r : X86::PUSH32r;
910 for (unsigned i = CSI.size(); i != 0; --i) {
911 unsigned Reg = CSI[i-1].getReg();
912 // Add the callee-saved register as live-in. It's killed at the spill.
915 // X86RegisterInfo::emitPrologue will handle spilling of frame register.
917 if (!X86::VR128RegClass.contains(Reg) && !isWin64) {
918 CalleeFrameSize += SlotSize;
919 BuildMI(MBB, MI, DL, TII.get(Opc)).addReg(Reg, RegState::Kill);
921 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
922 TII.storeRegToStackSlot(MBB, MI, Reg, true, CSI[i-1].getFrameIdx(),
927 X86FI->setCalleeSavedFrameSize(CalleeFrameSize);
931 bool X86FrameLowering::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
932 MachineBasicBlock::iterator MI,
933 const std::vector<CalleeSavedInfo> &CSI,
934 const TargetRegisterInfo *TRI) const {
938 DebugLoc DL = MBB.findDebugLoc(MI);
940 MachineFunction &MF = *MBB.getParent();
941 const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
942 unsigned FPReg = TRI->getFrameRegister(MF);
943 bool isWin64 = STI.isTargetWin64();
944 unsigned Opc = STI.is64Bit() ? X86::POP64r : X86::POP32r;
945 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
946 unsigned Reg = CSI[i].getReg();
948 // X86RegisterInfo::emitEpilogue will handle restoring of frame register.
950 if (!X86::VR128RegClass.contains(Reg) && !isWin64) {
951 BuildMI(MBB, MI, DL, TII.get(Opc), Reg);
953 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
954 TII.loadRegFromStackSlot(MBB, MI, Reg, CSI[i].getFrameIdx(),
962 X86FrameLowering::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
963 RegScavenger *RS) const {
964 MachineFrameInfo *MFI = MF.getFrameInfo();
965 const X86RegisterInfo *RegInfo = TM.getRegisterInfo();
966 unsigned SlotSize = RegInfo->getSlotSize();
968 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
969 int32_t TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
971 if (TailCallReturnAddrDelta < 0) {
972 // create RETURNADDR area
981 MFI->CreateFixedObject(-TailCallReturnAddrDelta,
982 (-1U*SlotSize)+TailCallReturnAddrDelta, true);
986 assert((TailCallReturnAddrDelta <= 0) &&
987 "The Delta should always be zero or negative");
988 const TargetFrameLowering &TFI = *MF.getTarget().getFrameLowering();
990 // Create a frame entry for the EBP register that must be saved.
991 int FrameIdx = MFI->CreateFixedObject(SlotSize,
993 TFI.getOffsetOfLocalArea() +
994 TailCallReturnAddrDelta,
996 assert(FrameIdx == MFI->getObjectIndexBegin() &&
997 "Slot for EBP register must be last in order to be found!");