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/CodeGen/MachineFrameInfo.h"
22 #include "llvm/CodeGen/MachineFunction.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineModuleInfo.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/MC/MCAsmInfo.h"
29 #include "llvm/MC/MCSymbol.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Target/TargetOptions.h"
35 // FIXME: completely move here.
36 extern cl::opt<bool> ForceStackAlign;
38 bool X86FrameLowering::hasReservedCallFrame(const MachineFunction &MF) const {
39 return !MF.getFrameInfo()->hasVarSizedObjects();
42 /// hasFP - Return true if the specified function should have a dedicated frame
43 /// pointer register. This is true if the function has variable sized allocas
44 /// or if frame pointer elimination is disabled.
45 bool X86FrameLowering::hasFP(const MachineFunction &MF) const {
46 const MachineFrameInfo *MFI = MF.getFrameInfo();
47 const MachineModuleInfo &MMI = MF.getMMI();
48 const TargetRegisterInfo *RegInfo = TM.getRegisterInfo();
50 return (MF.getTarget().Options.DisableFramePointerElim(MF) ||
51 RegInfo->needsStackRealignment(MF) ||
52 MFI->hasVarSizedObjects() ||
53 MFI->isFrameAddressTaken() || MFI->hasInlineAsmWithSPAdjust() ||
54 MF.getInfo<X86MachineFunctionInfo>()->getForceFramePointer() ||
55 MMI.callsUnwindInit() || MMI.callsEHReturn());
58 static unsigned getSUBriOpcode(unsigned IsLP64, int64_t Imm) {
62 return X86::SUB64ri32;
70 static unsigned getADDriOpcode(unsigned IsLP64, int64_t Imm) {
74 return X86::ADD64ri32;
82 static unsigned getLEArOpcode(unsigned IsLP64) {
83 return IsLP64 ? X86::LEA64r : X86::LEA32r;
86 /// findDeadCallerSavedReg - Return a caller-saved register that isn't live
87 /// when it reaches the "return" instruction. We can then pop a stack object
88 /// to this register without worry about clobbering it.
89 static unsigned findDeadCallerSavedReg(MachineBasicBlock &MBB,
90 MachineBasicBlock::iterator &MBBI,
91 const TargetRegisterInfo &TRI,
93 const MachineFunction *MF = MBB.getParent();
94 const Function *F = MF->getFunction();
95 if (!F || MF->getMMI().callsEHReturn())
98 static const uint16_t CallerSavedRegs32Bit[] = {
99 X86::EAX, X86::EDX, X86::ECX, 0
102 static const uint16_t CallerSavedRegs64Bit[] = {
103 X86::RAX, X86::RDX, X86::RCX, X86::RSI, X86::RDI,
104 X86::R8, X86::R9, X86::R10, X86::R11, 0
107 unsigned Opc = MBBI->getOpcode();
112 case X86::TCRETURNdi:
113 case X86::TCRETURNri:
114 case X86::TCRETURNmi:
115 case X86::TCRETURNdi64:
116 case X86::TCRETURNri64:
117 case X86::TCRETURNmi64:
119 case X86::EH_RETURN64: {
120 SmallSet<uint16_t, 8> Uses;
121 for (unsigned i = 0, e = MBBI->getNumOperands(); i != e; ++i) {
122 MachineOperand &MO = MBBI->getOperand(i);
123 if (!MO.isReg() || MO.isDef())
125 unsigned Reg = MO.getReg();
128 for (MCRegAliasIterator AI(Reg, &TRI, true); AI.isValid(); ++AI)
132 const uint16_t *CS = Is64Bit ? CallerSavedRegs64Bit : CallerSavedRegs32Bit;
134 if (!Uses.count(*CS))
143 /// emitSPUpdate - Emit a series of instructions to increment / decrement the
144 /// stack pointer by a constant value.
146 void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
147 unsigned StackPtr, int64_t NumBytes,
148 bool Is64Bit, bool IsLP64, bool UseLEA,
149 const TargetInstrInfo &TII, const TargetRegisterInfo &TRI) {
150 bool isSub = NumBytes < 0;
151 uint64_t Offset = isSub ? -NumBytes : NumBytes;
154 Opc = getLEArOpcode(IsLP64);
157 ? getSUBriOpcode(IsLP64, Offset)
158 : getADDriOpcode(IsLP64, Offset);
160 uint64_t Chunk = (1LL << 31) - 1;
161 DebugLoc DL = MBB.findDebugLoc(MBBI);
164 uint64_t ThisVal = (Offset > Chunk) ? Chunk : Offset;
165 if (ThisVal == (Is64Bit ? 8 : 4)) {
166 // Use push / pop instead.
168 ? (unsigned)(Is64Bit ? X86::RAX : X86::EAX)
169 : findDeadCallerSavedReg(MBB, MBBI, TRI, Is64Bit);
172 ? (Is64Bit ? X86::PUSH64r : X86::PUSH32r)
173 : (Is64Bit ? X86::POP64r : X86::POP32r);
174 MachineInstr *MI = BuildMI(MBB, MBBI, DL, TII.get(Opc))
175 .addReg(Reg, getDefRegState(!isSub) | getUndefRegState(isSub));
177 MI->setFlag(MachineInstr::FrameSetup);
183 MachineInstr *MI = NULL;
186 MI = addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr),
187 StackPtr, false, isSub ? -ThisVal : ThisVal);
189 MI = BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr)
192 MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
196 MI->setFlag(MachineInstr::FrameSetup);
202 /// mergeSPUpdatesUp - Merge two stack-manipulating instructions upper iterator.
204 void mergeSPUpdatesUp(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
205 unsigned StackPtr, uint64_t *NumBytes = NULL) {
206 if (MBBI == MBB.begin()) return;
208 MachineBasicBlock::iterator PI = prior(MBBI);
209 unsigned Opc = PI->getOpcode();
210 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
211 Opc == X86::ADD32ri || Opc == X86::ADD32ri8 ||
212 Opc == X86::LEA32r || Opc == X86::LEA64_32r) &&
213 PI->getOperand(0).getReg() == StackPtr) {
215 *NumBytes += PI->getOperand(2).getImm();
217 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
218 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
219 PI->getOperand(0).getReg() == StackPtr) {
221 *NumBytes -= PI->getOperand(2).getImm();
226 /// mergeSPUpdatesDown - Merge two stack-manipulating instructions lower iterator.
228 void mergeSPUpdatesDown(MachineBasicBlock &MBB,
229 MachineBasicBlock::iterator &MBBI,
230 unsigned StackPtr, uint64_t *NumBytes = NULL) {
231 // FIXME: THIS ISN'T RUN!!!
234 if (MBBI == MBB.end()) return;
236 MachineBasicBlock::iterator NI = llvm::next(MBBI);
237 if (NI == MBB.end()) return;
239 unsigned Opc = NI->getOpcode();
240 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
241 Opc == X86::ADD32ri || Opc == X86::ADD32ri8) &&
242 NI->getOperand(0).getReg() == StackPtr) {
244 *NumBytes -= NI->getOperand(2).getImm();
247 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
248 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
249 NI->getOperand(0).getReg() == StackPtr) {
251 *NumBytes += NI->getOperand(2).getImm();
257 /// mergeSPUpdates - Checks the instruction before/after the passed
258 /// instruction. If it is an ADD/SUB/LEA instruction it is deleted argument and the
259 /// stack adjustment is returned as a positive value for ADD/LEA and a negative for
261 static int mergeSPUpdates(MachineBasicBlock &MBB,
262 MachineBasicBlock::iterator &MBBI,
264 bool doMergeWithPrevious) {
265 if ((doMergeWithPrevious && MBBI == MBB.begin()) ||
266 (!doMergeWithPrevious && MBBI == MBB.end()))
269 MachineBasicBlock::iterator PI = doMergeWithPrevious ? prior(MBBI) : MBBI;
270 MachineBasicBlock::iterator NI = doMergeWithPrevious ? 0 : llvm::next(MBBI);
271 unsigned Opc = PI->getOpcode();
274 if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 ||
275 Opc == X86::ADD32ri || Opc == X86::ADD32ri8 ||
276 Opc == X86::LEA32r || Opc == X86::LEA64_32r) &&
277 PI->getOperand(0).getReg() == StackPtr){
278 Offset += PI->getOperand(2).getImm();
280 if (!doMergeWithPrevious) MBBI = NI;
281 } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 ||
282 Opc == X86::SUB32ri || Opc == X86::SUB32ri8) &&
283 PI->getOperand(0).getReg() == StackPtr) {
284 Offset -= PI->getOperand(2).getImm();
286 if (!doMergeWithPrevious) MBBI = NI;
292 static bool isEAXLiveIn(MachineFunction &MF) {
293 for (MachineRegisterInfo::livein_iterator II = MF.getRegInfo().livein_begin(),
294 EE = MF.getRegInfo().livein_end(); II != EE; ++II) {
295 unsigned Reg = II->first;
297 if (Reg == X86::EAX || Reg == X86::AX ||
298 Reg == X86::AH || Reg == X86::AL)
305 void X86FrameLowering::emitCalleeSavedFrameMoves(MachineFunction &MF,
307 unsigned FramePtr) const {
308 MachineFrameInfo *MFI = MF.getFrameInfo();
309 MachineModuleInfo &MMI = MF.getMMI();
310 const MCRegisterInfo *MRI = MMI.getContext().getRegisterInfo();
312 // Add callee saved registers to move list.
313 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
314 if (CSI.empty()) return;
316 const X86RegisterInfo *RegInfo = TM.getRegisterInfo();
317 bool HasFP = hasFP(MF);
319 // Calculate amount of bytes used for return address storing.
320 int stackGrowth = -RegInfo->getSlotSize();
322 // FIXME: This is dirty hack. The code itself is pretty mess right now.
323 // It should be rewritten from scratch and generalized sometimes.
325 // Determine maximum offset (minimum due to stack growth).
326 int64_t MaxOffset = 0;
327 for (std::vector<CalleeSavedInfo>::const_iterator
328 I = CSI.begin(), E = CSI.end(); I != E; ++I)
329 MaxOffset = std::min(MaxOffset,
330 MFI->getObjectOffset(I->getFrameIdx()));
332 // Calculate offsets.
333 int64_t saveAreaOffset = (HasFP ? 3 : 2) * stackGrowth;
334 for (std::vector<CalleeSavedInfo>::const_iterator
335 I = CSI.begin(), E = CSI.end(); I != E; ++I) {
336 int64_t Offset = MFI->getObjectOffset(I->getFrameIdx());
337 unsigned Reg = I->getReg();
338 Offset = MaxOffset - Offset + saveAreaOffset;
340 // Don't output a new machine move if we're re-saving the frame
341 // pointer. This happens when the PrologEpilogInserter has inserted an extra
342 // "PUSH" of the frame pointer -- the "emitPrologue" method automatically
343 // generates one when frame pointers are used. If we generate a "machine
344 // move" for this extra "PUSH", the linker will lose track of the fact that
345 // the frame pointer should have the value of the first "PUSH" when it's
348 // FIXME: This looks inelegant. It's possibly correct, but it's covering up
349 // another bug. I.e., one where we generate a prolog like this:
357 // The immediate re-push of EBP is unnecessary. At the least, it's an
358 // optimization bug. EBP can be used as a scratch register in certain
359 // cases, but probably not when we have a frame pointer.
360 if (HasFP && FramePtr == Reg)
363 unsigned DwarfReg = MRI->getDwarfRegNum(Reg, true);
364 MMI.addFrameInst(MCCFIInstruction::createOffset(Label, DwarfReg, Offset));
368 /// usesTheStack - This function checks if any of the users of EFLAGS
369 /// copies the EFLAGS. We know that the code that lowers COPY of EFLAGS has
370 /// to use the stack, and if we don't adjust the stack we clobber the first
372 /// See X86InstrInfo::copyPhysReg.
373 static bool usesTheStack(const MachineFunction &MF) {
374 const MachineRegisterInfo &MRI = MF.getRegInfo();
376 for (MachineRegisterInfo::reg_iterator ri = MRI.reg_begin(X86::EFLAGS),
377 re = MRI.reg_end(); ri != re; ++ri)
384 /// emitPrologue - Push callee-saved registers onto the stack, which
385 /// automatically adjust the stack pointer. Adjust the stack pointer to allocate
386 /// space for local variables. Also emit labels used by the exception handler to
387 /// generate the exception handling frames.
388 void X86FrameLowering::emitPrologue(MachineFunction &MF) const {
389 MachineBasicBlock &MBB = MF.front(); // Prologue goes in entry BB.
390 MachineBasicBlock::iterator MBBI = MBB.begin();
391 MachineFrameInfo *MFI = MF.getFrameInfo();
392 const Function *Fn = MF.getFunction();
393 const X86RegisterInfo *RegInfo = TM.getRegisterInfo();
394 const X86InstrInfo &TII = *TM.getInstrInfo();
395 MachineModuleInfo &MMI = MF.getMMI();
396 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
397 bool needsFrameMoves = MMI.hasDebugInfo() ||
398 Fn->needsUnwindTableEntry();
399 uint64_t MaxAlign = MFI->getMaxAlignment(); // Desired stack alignment.
400 uint64_t StackSize = MFI->getStackSize(); // Number of bytes to allocate.
401 bool HasFP = hasFP(MF);
402 bool Is64Bit = STI.is64Bit();
403 bool IsLP64 = STI.isTarget64BitLP64();
404 bool IsWin64 = STI.isTargetWin64();
405 bool UseLEA = STI.useLeaForSP();
406 unsigned StackAlign = getStackAlignment();
407 unsigned SlotSize = RegInfo->getSlotSize();
408 unsigned FramePtr = RegInfo->getFrameRegister(MF);
409 unsigned StackPtr = RegInfo->getStackRegister();
410 unsigned BasePtr = RegInfo->getBaseRegister();
413 // If we're forcing a stack realignment we can't rely on just the frame
414 // info, we need to know the ABI stack alignment as well in case we
415 // have a call out. Otherwise just make sure we have some alignment - we'll
416 // go with the minimum SlotSize.
417 if (ForceStackAlign) {
419 MaxAlign = (StackAlign > MaxAlign) ? StackAlign : MaxAlign;
420 else if (MaxAlign < SlotSize)
424 // Add RETADDR move area to callee saved frame size.
425 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
426 if (TailCallReturnAddrDelta < 0)
427 X86FI->setCalleeSavedFrameSize(
428 X86FI->getCalleeSavedFrameSize() - TailCallReturnAddrDelta);
430 // If this is x86-64 and the Red Zone is not disabled, if we are a leaf
431 // function, and use up to 128 bytes of stack space, don't have a frame
432 // pointer, calls, or dynamic alloca then we do not need to adjust the
433 // stack pointer (we fit in the Red Zone). We also check that we don't
434 // push and pop from the stack.
435 if (Is64Bit && !Fn->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
436 Attribute::NoRedZone) &&
437 !RegInfo->needsStackRealignment(MF) &&
438 !MFI->hasVarSizedObjects() && // No dynamic alloca.
439 !MFI->adjustsStack() && // No calls.
440 !IsWin64 && // Win64 has no Red Zone
441 !usesTheStack(MF) && // Don't push and pop.
442 !MF.getTarget().Options.EnableSegmentedStacks) { // Regular stack
443 uint64_t MinSize = X86FI->getCalleeSavedFrameSize();
444 if (HasFP) MinSize += SlotSize;
445 StackSize = std::max(MinSize, StackSize > 128 ? StackSize - 128 : 0);
446 MFI->setStackSize(StackSize);
449 // Insert stack pointer adjustment for later moving of return addr. Only
450 // applies to tail call optimized functions where the callee argument stack
451 // size is bigger than the callers.
452 if (TailCallReturnAddrDelta < 0) {
454 BuildMI(MBB, MBBI, DL,
455 TII.get(getSUBriOpcode(IsLP64, -TailCallReturnAddrDelta)),
458 .addImm(-TailCallReturnAddrDelta)
459 .setMIFlag(MachineInstr::FrameSetup);
460 MI->getOperand(3).setIsDead(); // The EFLAGS implicit def is dead.
463 // Mapping for machine moves:
465 // DST: VirtualFP AND
466 // SRC: VirtualFP => DW_CFA_def_cfa_offset
467 // ELSE => DW_CFA_def_cfa
469 // SRC: VirtualFP AND
470 // DST: Register => DW_CFA_def_cfa_register
473 // OFFSET < 0 => DW_CFA_offset_extended_sf
474 // REG < 64 => DW_CFA_offset + Reg
475 // ELSE => DW_CFA_offset_extended
477 uint64_t NumBytes = 0;
478 int stackGrowth = -SlotSize;
481 // Calculate required stack adjustment.
482 uint64_t FrameSize = StackSize - SlotSize;
483 if (RegInfo->needsStackRealignment(MF)) {
484 // Callee-saved registers are pushed on stack before the stack
486 FrameSize -= X86FI->getCalleeSavedFrameSize();
487 NumBytes = (FrameSize + MaxAlign - 1) / MaxAlign * MaxAlign;
489 NumBytes = FrameSize - X86FI->getCalleeSavedFrameSize();
492 // Get the offset of the stack slot for the EBP register, which is
493 // guaranteed to be the last slot by processFunctionBeforeFrameFinalized.
494 // Update the frame offset adjustment.
495 MFI->setOffsetAdjustment(-NumBytes);
497 // Save EBP/RBP into the appropriate stack slot.
498 BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::PUSH64r : X86::PUSH32r))
499 .addReg(FramePtr, RegState::Kill)
500 .setMIFlag(MachineInstr::FrameSetup);
502 if (needsFrameMoves) {
503 // Mark the place where EBP/RBP was saved.
504 MCSymbol *FrameLabel = MMI.getContext().CreateTempSymbol();
505 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL))
508 // Define the current CFA rule to use the provided offset.
511 MCCFIInstruction::createDefCfaOffset(FrameLabel, 2 * stackGrowth));
513 // Change the rule for the FramePtr to be an "offset" rule.
514 unsigned DwarfFramePtr = RegInfo->getDwarfRegNum(FramePtr, true);
515 MMI.addFrameInst(MCCFIInstruction::createOffset(FrameLabel, DwarfFramePtr,
519 // Update EBP with the new base value.
520 BuildMI(MBB, MBBI, DL,
521 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr), FramePtr)
523 .setMIFlag(MachineInstr::FrameSetup);
525 if (needsFrameMoves) {
526 // Mark effective beginning of when frame pointer becomes valid.
527 MCSymbol *FrameLabel = MMI.getContext().CreateTempSymbol();
528 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL))
531 // Define the current CFA to use the EBP/RBP register.
532 unsigned DwarfFramePtr = RegInfo->getDwarfRegNum(FramePtr, true);
534 MCCFIInstruction::createDefCfaRegister(FrameLabel, DwarfFramePtr));
537 // Mark the FramePtr as live-in in every block except the entry.
538 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
540 I->addLiveIn(FramePtr);
542 NumBytes = StackSize - X86FI->getCalleeSavedFrameSize();
545 // Skip the callee-saved push instructions.
546 bool PushedRegs = false;
547 int StackOffset = 2 * stackGrowth;
549 while (MBBI != MBB.end() &&
550 (MBBI->getOpcode() == X86::PUSH32r ||
551 MBBI->getOpcode() == X86::PUSH64r)) {
553 MBBI->setFlag(MachineInstr::FrameSetup);
556 if (!HasFP && needsFrameMoves) {
557 // Mark callee-saved push instruction.
558 MCSymbol *Label = MMI.getContext().CreateTempSymbol();
559 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL)).addSym(Label);
561 // Define the current CFA rule to use the provided offset.
564 MCCFIInstruction::createDefCfaOffset(Label, StackOffset));
565 StackOffset += stackGrowth;
569 // Realign stack after we pushed callee-saved registers (so that we'll be
570 // able to calculate their offsets from the frame pointer).
572 // NOTE: We push the registers before realigning the stack, so
573 // vector callee-saved (xmm) registers may be saved w/o proper
574 // alignment in this way. However, currently these regs are saved in
575 // stack slots (see X86FrameLowering::spillCalleeSavedRegisters()), so
576 // this shouldn't be a problem.
577 if (RegInfo->needsStackRealignment(MF)) {
578 assert(HasFP && "There should be a frame pointer if stack is realigned.");
580 BuildMI(MBB, MBBI, DL,
581 TII.get(Is64Bit ? X86::AND64ri32 : X86::AND32ri), StackPtr)
584 .setMIFlag(MachineInstr::FrameSetup);
586 // The EFLAGS implicit def is dead.
587 MI->getOperand(3).setIsDead();
590 // If there is an SUB32ri of ESP immediately before this instruction, merge
591 // the two. This can be the case when tail call elimination is enabled and
592 // the callee has more arguments then the caller.
593 NumBytes -= mergeSPUpdates(MBB, MBBI, StackPtr, true);
595 // If there is an ADD32ri or SUB32ri of ESP immediately after this
596 // instruction, merge the two instructions.
597 mergeSPUpdatesDown(MBB, MBBI, StackPtr, &NumBytes);
599 // Adjust stack pointer: ESP -= numbytes.
601 // Windows and cygwin/mingw require a prologue helper routine when allocating
602 // more than 4K bytes on the stack. Windows uses __chkstk and cygwin/mingw
603 // uses __alloca. __alloca and the 32-bit version of __chkstk will probe the
604 // stack and adjust the stack pointer in one go. The 64-bit version of
605 // __chkstk is only responsible for probing the stack. The 64-bit prologue is
606 // responsible for adjusting the stack pointer. Touching the stack at 4K
607 // increments is necessary to ensure that the guard pages used by the OS
608 // virtual memory manager are allocated in correct sequence.
609 if (NumBytes >= 4096 && STI.isOSWindows() && !STI.isTargetMacho()) {
610 const char *StackProbeSymbol;
613 if (STI.isTargetCygMing()) {
614 StackProbeSymbol = "___chkstk_ms";
616 StackProbeSymbol = "__chkstk";
618 } else if (STI.isTargetCygMing())
619 StackProbeSymbol = "_alloca";
621 StackProbeSymbol = "_chkstk";
623 // Check whether EAX is livein for this function.
624 bool isEAXAlive = isEAXLiveIn(MF);
627 // Sanity check that EAX is not livein for this function.
628 // It should not be, so throw an assert.
629 assert(!Is64Bit && "EAX is livein in x64 case!");
632 BuildMI(MBB, MBBI, DL, TII.get(X86::PUSH32r))
633 .addReg(X86::EAX, RegState::Kill)
634 .setMIFlag(MachineInstr::FrameSetup);
638 // Handle the 64-bit Windows ABI case where we need to call __chkstk.
639 // Function prologue is responsible for adjusting the stack pointer.
640 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64ri), X86::RAX)
642 .setMIFlag(MachineInstr::FrameSetup);
644 // Allocate NumBytes-4 bytes on stack in case of isEAXAlive.
645 // We'll also use 4 already allocated bytes for EAX.
646 BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), X86::EAX)
647 .addImm(isEAXAlive ? NumBytes - 4 : NumBytes)
648 .setMIFlag(MachineInstr::FrameSetup);
651 BuildMI(MBB, MBBI, DL,
652 TII.get(Is64Bit ? X86::W64ALLOCA : X86::CALLpcrel32))
653 .addExternalSymbol(StackProbeSymbol)
654 .addReg(StackPtr, RegState::Define | RegState::Implicit)
655 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit)
656 .setMIFlag(MachineInstr::FrameSetup);
659 // MSVC x64's __chkstk and cygwin/mingw's ___chkstk_ms do not adjust %rsp
660 // themself. It also does not clobber %rax so we can reuse it when
662 BuildMI(MBB, MBBI, DL, TII.get(X86::SUB64rr), StackPtr)
665 .setMIFlag(MachineInstr::FrameSetup);
669 MachineInstr *MI = addRegOffset(BuildMI(MF, DL, TII.get(X86::MOV32rm),
671 StackPtr, false, NumBytes - 4);
672 MI->setFlag(MachineInstr::FrameSetup);
673 MBB.insert(MBBI, MI);
676 emitSPUpdate(MBB, MBBI, StackPtr, -(int64_t)NumBytes, Is64Bit, IsLP64,
677 UseLEA, TII, *RegInfo);
679 // If we need a base pointer, set it up here. It's whatever the value
680 // of the stack pointer is at this point. Any variable size objects
681 // will be allocated after this, so we can still use the base pointer
682 // to reference locals.
683 if (RegInfo->hasBasePointer(MF)) {
684 // Update the frame pointer with the current stack pointer.
685 unsigned Opc = Is64Bit ? X86::MOV64rr : X86::MOV32rr;
686 BuildMI(MBB, MBBI, DL, TII.get(Opc), BasePtr)
688 .setMIFlag(MachineInstr::FrameSetup);
691 if (( (!HasFP && NumBytes) || PushedRegs) && needsFrameMoves) {
692 // Mark end of stack pointer adjustment.
693 MCSymbol *Label = MMI.getContext().CreateTempSymbol();
694 BuildMI(MBB, MBBI, DL, TII.get(X86::PROLOG_LABEL))
697 if (!HasFP && NumBytes) {
698 // Define the current CFA rule to use the provided offset.
700 MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(
701 Label, -StackSize + stackGrowth));
704 // Emit DWARF info specifying the offsets of the callee-saved registers.
706 emitCalleeSavedFrameMoves(MF, Label, HasFP ? FramePtr : StackPtr);
710 void X86FrameLowering::emitEpilogue(MachineFunction &MF,
711 MachineBasicBlock &MBB) const {
712 const MachineFrameInfo *MFI = MF.getFrameInfo();
713 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
714 const X86RegisterInfo *RegInfo = TM.getRegisterInfo();
715 const X86InstrInfo &TII = *TM.getInstrInfo();
716 MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr();
717 assert(MBBI != MBB.end() && "Returning block has no instructions");
718 unsigned RetOpcode = MBBI->getOpcode();
719 DebugLoc DL = MBBI->getDebugLoc();
720 bool Is64Bit = STI.is64Bit();
721 bool IsLP64 = STI.isTarget64BitLP64();
722 bool UseLEA = STI.useLeaForSP();
723 unsigned StackAlign = getStackAlignment();
724 unsigned SlotSize = RegInfo->getSlotSize();
725 unsigned FramePtr = RegInfo->getFrameRegister(MF);
726 unsigned StackPtr = RegInfo->getStackRegister();
730 llvm_unreachable("Can only insert epilog into returning blocks");
733 case X86::TCRETURNdi:
734 case X86::TCRETURNri:
735 case X86::TCRETURNmi:
736 case X86::TCRETURNdi64:
737 case X86::TCRETURNri64:
738 case X86::TCRETURNmi64:
740 case X86::EH_RETURN64:
741 break; // These are ok
744 // Get the number of bytes to allocate from the FrameInfo.
745 uint64_t StackSize = MFI->getStackSize();
746 uint64_t MaxAlign = MFI->getMaxAlignment();
747 unsigned CSSize = X86FI->getCalleeSavedFrameSize();
748 uint64_t NumBytes = 0;
750 // If we're forcing a stack realignment we can't rely on just the frame
751 // info, we need to know the ABI stack alignment as well in case we
752 // have a call out. Otherwise just make sure we have some alignment - we'll
753 // go with the minimum.
754 if (ForceStackAlign) {
756 MaxAlign = (StackAlign > MaxAlign) ? StackAlign : MaxAlign;
758 MaxAlign = MaxAlign ? MaxAlign : 4;
762 // Calculate required stack adjustment.
763 uint64_t FrameSize = StackSize - SlotSize;
764 if (RegInfo->needsStackRealignment(MF)) {
765 // Callee-saved registers were pushed on stack before the stack
768 NumBytes = (FrameSize + MaxAlign - 1) / MaxAlign * MaxAlign;
770 NumBytes = FrameSize - CSSize;
774 BuildMI(MBB, MBBI, DL,
775 TII.get(Is64Bit ? X86::POP64r : X86::POP32r), FramePtr);
777 NumBytes = StackSize - CSSize;
780 // Skip the callee-saved pop instructions.
781 while (MBBI != MBB.begin()) {
782 MachineBasicBlock::iterator PI = prior(MBBI);
783 unsigned Opc = PI->getOpcode();
785 if (Opc != X86::POP32r && Opc != X86::POP64r && Opc != X86::DBG_VALUE &&
791 MachineBasicBlock::iterator FirstCSPop = MBBI;
793 DL = MBBI->getDebugLoc();
795 // If there is an ADD32ri or SUB32ri of ESP immediately before this
796 // instruction, merge the two instructions.
797 if (NumBytes || MFI->hasVarSizedObjects())
798 mergeSPUpdatesUp(MBB, MBBI, StackPtr, &NumBytes);
800 // If dynamic alloca is used, then reset esp to point to the last callee-saved
801 // slot before popping them off! Same applies for the case, when stack was
803 if (RegInfo->needsStackRealignment(MF) || MFI->hasVarSizedObjects()) {
804 if (RegInfo->needsStackRealignment(MF))
807 unsigned Opc = getLEArOpcode(IsLP64);
808 addRegOffset(BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr),
809 FramePtr, false, -CSSize);
811 unsigned Opc = (Is64Bit ? X86::MOV64rr : X86::MOV32rr);
812 BuildMI(MBB, MBBI, DL, TII.get(Opc), StackPtr)
815 } else if (NumBytes) {
816 // Adjust stack pointer back: ESP += numbytes.
817 emitSPUpdate(MBB, MBBI, StackPtr, NumBytes, Is64Bit, IsLP64, UseLEA,
821 // We're returning from function via eh_return.
822 if (RetOpcode == X86::EH_RETURN || RetOpcode == X86::EH_RETURN64) {
823 MBBI = MBB.getLastNonDebugInstr();
824 MachineOperand &DestAddr = MBBI->getOperand(0);
825 assert(DestAddr.isReg() && "Offset should be in register!");
826 BuildMI(MBB, MBBI, DL,
827 TII.get(Is64Bit ? X86::MOV64rr : X86::MOV32rr),
828 StackPtr).addReg(DestAddr.getReg());
829 } else if (RetOpcode == X86::TCRETURNri || RetOpcode == X86::TCRETURNdi ||
830 RetOpcode == X86::TCRETURNmi ||
831 RetOpcode == X86::TCRETURNri64 || RetOpcode == X86::TCRETURNdi64 ||
832 RetOpcode == X86::TCRETURNmi64) {
833 bool isMem = RetOpcode == X86::TCRETURNmi || RetOpcode == X86::TCRETURNmi64;
834 // Tail call return: adjust the stack pointer and jump to callee.
835 MBBI = MBB.getLastNonDebugInstr();
836 MachineOperand &JumpTarget = MBBI->getOperand(0);
837 MachineOperand &StackAdjust = MBBI->getOperand(isMem ? 5 : 1);
838 assert(StackAdjust.isImm() && "Expecting immediate value.");
840 // Adjust stack pointer.
841 int StackAdj = StackAdjust.getImm();
842 int MaxTCDelta = X86FI->getTCReturnAddrDelta();
844 assert(MaxTCDelta <= 0 && "MaxTCDelta should never be positive");
846 // Incoporate the retaddr area.
847 Offset = StackAdj-MaxTCDelta;
848 assert(Offset >= 0 && "Offset should never be negative");
851 // Check for possible merge with preceding ADD instruction.
852 Offset += mergeSPUpdates(MBB, MBBI, StackPtr, true);
853 emitSPUpdate(MBB, MBBI, StackPtr, Offset, Is64Bit, IsLP64,
854 UseLEA, TII, *RegInfo);
857 // Jump to label or value in register.
858 if (RetOpcode == X86::TCRETURNdi || RetOpcode == X86::TCRETURNdi64) {
859 MachineInstrBuilder MIB =
860 BuildMI(MBB, MBBI, DL, TII.get((RetOpcode == X86::TCRETURNdi)
861 ? X86::TAILJMPd : X86::TAILJMPd64));
862 if (JumpTarget.isGlobal())
863 MIB.addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(),
864 JumpTarget.getTargetFlags());
866 assert(JumpTarget.isSymbol());
867 MIB.addExternalSymbol(JumpTarget.getSymbolName(),
868 JumpTarget.getTargetFlags());
870 } else if (RetOpcode == X86::TCRETURNmi || RetOpcode == X86::TCRETURNmi64) {
871 MachineInstrBuilder MIB =
872 BuildMI(MBB, MBBI, DL, TII.get((RetOpcode == X86::TCRETURNmi)
873 ? X86::TAILJMPm : X86::TAILJMPm64));
874 for (unsigned i = 0; i != 5; ++i)
875 MIB.addOperand(MBBI->getOperand(i));
876 } else if (RetOpcode == X86::TCRETURNri64) {
877 BuildMI(MBB, MBBI, DL, TII.get(X86::TAILJMPr64)).
878 addReg(JumpTarget.getReg(), RegState::Kill);
880 BuildMI(MBB, MBBI, DL, TII.get(X86::TAILJMPr)).
881 addReg(JumpTarget.getReg(), RegState::Kill);
884 MachineInstr *NewMI = prior(MBBI);
885 NewMI->copyImplicitOps(MF, MBBI);
887 // Delete the pseudo instruction TCRETURN.
889 } else if ((RetOpcode == X86::RET || RetOpcode == X86::RETI) &&
890 (X86FI->getTCReturnAddrDelta() < 0)) {
891 // Add the return addr area delta back since we are not tail calling.
892 int delta = -1*X86FI->getTCReturnAddrDelta();
893 MBBI = MBB.getLastNonDebugInstr();
895 // Check for possible merge with preceding ADD instruction.
896 delta += mergeSPUpdates(MBB, MBBI, StackPtr, true);
897 emitSPUpdate(MBB, MBBI, StackPtr, delta, Is64Bit, IsLP64, UseLEA, TII,
902 int X86FrameLowering::getFrameIndexOffset(const MachineFunction &MF, int FI) const {
903 const X86RegisterInfo *RegInfo =
904 static_cast<const X86RegisterInfo*>(MF.getTarget().getRegisterInfo());
905 const MachineFrameInfo *MFI = MF.getFrameInfo();
906 int Offset = MFI->getObjectOffset(FI) - getOffsetOfLocalArea();
907 uint64_t StackSize = MFI->getStackSize();
909 if (RegInfo->hasBasePointer(MF)) {
910 assert (hasFP(MF) && "VLAs and dynamic stack realign, but no FP?!");
912 // Skip the saved EBP.
913 return Offset + RegInfo->getSlotSize();
915 assert((-(Offset + StackSize)) % MFI->getObjectAlignment(FI) == 0);
916 return Offset + StackSize;
918 } else if (RegInfo->needsStackRealignment(MF)) {
920 // Skip the saved EBP.
921 return Offset + RegInfo->getSlotSize();
923 assert((-(Offset + StackSize)) % MFI->getObjectAlignment(FI) == 0);
924 return Offset + StackSize;
926 // FIXME: Support tail calls
929 return Offset + StackSize;
931 // Skip the saved EBP.
932 Offset += RegInfo->getSlotSize();
934 // Skip the RETADDR move area
935 const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
936 int TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
937 if (TailCallReturnAddrDelta < 0)
938 Offset -= TailCallReturnAddrDelta;
944 int X86FrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI,
945 unsigned &FrameReg) const {
946 const X86RegisterInfo *RegInfo =
947 static_cast<const X86RegisterInfo*>(MF.getTarget().getRegisterInfo());
948 // We can't calculate offset from frame pointer if the stack is realigned,
949 // so enforce usage of stack/base pointer. The base pointer is used when we
950 // have dynamic allocas in addition to dynamic realignment.
951 if (RegInfo->hasBasePointer(MF))
952 FrameReg = RegInfo->getBaseRegister();
953 else if (RegInfo->needsStackRealignment(MF))
954 FrameReg = RegInfo->getStackRegister();
956 FrameReg = RegInfo->getFrameRegister(MF);
957 return getFrameIndexOffset(MF, FI);
960 bool X86FrameLowering::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
961 MachineBasicBlock::iterator MI,
962 const std::vector<CalleeSavedInfo> &CSI,
963 const TargetRegisterInfo *TRI) const {
967 DebugLoc DL = MBB.findDebugLoc(MI);
969 MachineFunction &MF = *MBB.getParent();
971 unsigned SlotSize = STI.is64Bit() ? 8 : 4;
972 unsigned FPReg = TRI->getFrameRegister(MF);
973 unsigned CalleeFrameSize = 0;
975 const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
976 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
978 // Push GPRs. It increases frame size.
979 unsigned Opc = STI.is64Bit() ? X86::PUSH64r : X86::PUSH32r;
980 for (unsigned i = CSI.size(); i != 0; --i) {
981 unsigned Reg = CSI[i-1].getReg();
982 if (!X86::GR64RegClass.contains(Reg) &&
983 !X86::GR32RegClass.contains(Reg))
985 // Add the callee-saved register as live-in. It's killed at the spill.
988 // X86RegisterInfo::emitPrologue will handle spilling of frame register.
990 CalleeFrameSize += SlotSize;
991 BuildMI(MBB, MI, DL, TII.get(Opc)).addReg(Reg, RegState::Kill)
992 .setMIFlag(MachineInstr::FrameSetup);
995 X86FI->setCalleeSavedFrameSize(CalleeFrameSize);
997 // Make XMM regs spilled. X86 does not have ability of push/pop XMM.
998 // It can be done by spilling XMMs to stack frame.
999 // Note that only Win64 ABI might spill XMMs.
1000 for (unsigned i = CSI.size(); i != 0; --i) {
1001 unsigned Reg = CSI[i-1].getReg();
1002 if (X86::GR64RegClass.contains(Reg) ||
1003 X86::GR32RegClass.contains(Reg))
1005 // Add the callee-saved register as live-in. It's killed at the spill.
1007 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
1008 TII.storeRegToStackSlot(MBB, MI, Reg, true, CSI[i-1].getFrameIdx(),
1015 bool X86FrameLowering::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1016 MachineBasicBlock::iterator MI,
1017 const std::vector<CalleeSavedInfo> &CSI,
1018 const TargetRegisterInfo *TRI) const {
1022 DebugLoc DL = MBB.findDebugLoc(MI);
1024 MachineFunction &MF = *MBB.getParent();
1025 const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
1027 // Reload XMMs from stack frame.
1028 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1029 unsigned Reg = CSI[i].getReg();
1030 if (X86::GR64RegClass.contains(Reg) ||
1031 X86::GR32RegClass.contains(Reg))
1033 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
1034 TII.loadRegFromStackSlot(MBB, MI, Reg, CSI[i].getFrameIdx(),
1039 unsigned FPReg = TRI->getFrameRegister(MF);
1040 unsigned Opc = STI.is64Bit() ? X86::POP64r : X86::POP32r;
1041 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1042 unsigned Reg = CSI[i].getReg();
1043 if (!X86::GR64RegClass.contains(Reg) &&
1044 !X86::GR32RegClass.contains(Reg))
1047 // X86RegisterInfo::emitEpilogue will handle restoring of frame register.
1049 BuildMI(MBB, MI, DL, TII.get(Opc), Reg);
1055 X86FrameLowering::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
1056 RegScavenger *RS) const {
1057 MachineFrameInfo *MFI = MF.getFrameInfo();
1058 const X86RegisterInfo *RegInfo = TM.getRegisterInfo();
1059 unsigned SlotSize = RegInfo->getSlotSize();
1061 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1062 int64_t TailCallReturnAddrDelta = X86FI->getTCReturnAddrDelta();
1064 if (TailCallReturnAddrDelta < 0) {
1065 // create RETURNADDR area
1074 MFI->CreateFixedObject(-TailCallReturnAddrDelta,
1075 TailCallReturnAddrDelta - SlotSize, true);
1079 assert((TailCallReturnAddrDelta <= 0) &&
1080 "The Delta should always be zero or negative");
1081 const TargetFrameLowering &TFI = *MF.getTarget().getFrameLowering();
1083 // Create a frame entry for the EBP register that must be saved.
1084 int FrameIdx = MFI->CreateFixedObject(SlotSize,
1086 TFI.getOffsetOfLocalArea() +
1087 TailCallReturnAddrDelta,
1089 assert(FrameIdx == MFI->getObjectIndexBegin() &&
1090 "Slot for EBP register must be last in order to be found!");
1094 // Spill the BasePtr if it's used.
1095 if (RegInfo->hasBasePointer(MF))
1096 MF.getRegInfo().setPhysRegUsed(RegInfo->getBaseRegister());
1100 HasNestArgument(const MachineFunction *MF) {
1101 const Function *F = MF->getFunction();
1102 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
1104 if (I->hasNestAttr())
1110 /// GetScratchRegister - Get a temp register for performing work in the
1111 /// segmented stack and the Erlang/HiPE stack prologue. Depending on platform
1112 /// and the properties of the function either one or two registers will be
1113 /// needed. Set primary to true for the first register, false for the second.
1115 GetScratchRegister(bool Is64Bit, const MachineFunction &MF, bool Primary) {
1116 CallingConv::ID CallingConvention = MF.getFunction()->getCallingConv();
1119 if (CallingConvention == CallingConv::HiPE) {
1121 return Primary ? X86::R14 : X86::R13;
1123 return Primary ? X86::EBX : X86::EDI;
1127 return Primary ? X86::R11 : X86::R12;
1129 bool IsNested = HasNestArgument(&MF);
1131 if (CallingConvention == CallingConv::X86_FastCall ||
1132 CallingConvention == CallingConv::Fast) {
1134 report_fatal_error("Segmented stacks does not support fastcall with "
1135 "nested function.");
1136 return Primary ? X86::EAX : X86::ECX;
1139 return Primary ? X86::EDX : X86::EAX;
1140 return Primary ? X86::ECX : X86::EAX;
1143 // The stack limit in the TCB is set to this many bytes above the actual stack
1145 static const uint64_t kSplitStackAvailable = 256;
1148 X86FrameLowering::adjustForSegmentedStacks(MachineFunction &MF) const {
1149 MachineBasicBlock &prologueMBB = MF.front();
1150 MachineFrameInfo *MFI = MF.getFrameInfo();
1151 const X86InstrInfo &TII = *TM.getInstrInfo();
1153 bool Is64Bit = STI.is64Bit();
1154 unsigned TlsReg, TlsOffset;
1157 unsigned ScratchReg = GetScratchRegister(Is64Bit, MF, true);
1158 assert(!MF.getRegInfo().isLiveIn(ScratchReg) &&
1159 "Scratch register is live-in");
1161 if (MF.getFunction()->isVarArg())
1162 report_fatal_error("Segmented stacks do not support vararg functions.");
1163 if (!STI.isTargetLinux() && !STI.isTargetDarwin() &&
1164 !STI.isTargetWin32() && !STI.isTargetFreeBSD())
1165 report_fatal_error("Segmented stacks not supported on this platform.");
1167 MachineBasicBlock *allocMBB = MF.CreateMachineBasicBlock();
1168 MachineBasicBlock *checkMBB = MF.CreateMachineBasicBlock();
1169 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1170 bool IsNested = false;
1172 // We need to know if the function has a nest argument only in 64 bit mode.
1174 IsNested = HasNestArgument(&MF);
1176 // The MOV R10, RAX needs to be in a different block, since the RET we emit in
1177 // allocMBB needs to be last (terminating) instruction.
1179 for (MachineBasicBlock::livein_iterator i = prologueMBB.livein_begin(),
1180 e = prologueMBB.livein_end(); i != e; i++) {
1181 allocMBB->addLiveIn(*i);
1182 checkMBB->addLiveIn(*i);
1186 allocMBB->addLiveIn(X86::R10);
1188 MF.push_front(allocMBB);
1189 MF.push_front(checkMBB);
1191 // Eventually StackSize will be calculated by a link-time pass; which will
1192 // also decide whether checking code needs to be injected into this particular
1194 StackSize = MFI->getStackSize();
1196 // When the frame size is less than 256 we just compare the stack
1197 // boundary directly to the value of the stack pointer, per gcc.
1198 bool CompareStackPointer = StackSize < kSplitStackAvailable;
1200 // Read the limit off the current stacklet off the stack_guard location.
1202 if (STI.isTargetLinux()) {
1205 } else if (STI.isTargetDarwin()) {
1207 TlsOffset = 0x60 + 90*8; // See pthread_machdep.h. Steal TLS slot 90.
1208 } else if (STI.isTargetFreeBSD()) {
1212 report_fatal_error("Segmented stacks not supported on this platform.");
1215 if (CompareStackPointer)
1216 ScratchReg = X86::RSP;
1218 BuildMI(checkMBB, DL, TII.get(X86::LEA64r), ScratchReg).addReg(X86::RSP)
1219 .addImm(1).addReg(0).addImm(-StackSize).addReg(0);
1221 BuildMI(checkMBB, DL, TII.get(X86::CMP64rm)).addReg(ScratchReg)
1222 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg);
1224 if (STI.isTargetLinux()) {
1227 } else if (STI.isTargetDarwin()) {
1229 TlsOffset = 0x48 + 90*4;
1230 } else if (STI.isTargetWin32()) {
1232 TlsOffset = 0x14; // pvArbitrary, reserved for application use
1233 } else if (STI.isTargetFreeBSD()) {
1234 report_fatal_error("Segmented stacks not supported on FreeBSD i386.");
1236 report_fatal_error("Segmented stacks not supported on this platform.");
1239 if (CompareStackPointer)
1240 ScratchReg = X86::ESP;
1242 BuildMI(checkMBB, DL, TII.get(X86::LEA32r), ScratchReg).addReg(X86::ESP)
1243 .addImm(1).addReg(0).addImm(-StackSize).addReg(0);
1245 if (STI.isTargetLinux() || STI.isTargetWin32()) {
1246 BuildMI(checkMBB, DL, TII.get(X86::CMP32rm)).addReg(ScratchReg)
1247 .addReg(0).addImm(0).addReg(0).addImm(TlsOffset).addReg(TlsReg);
1248 } else if (STI.isTargetDarwin()) {
1250 // TlsOffset doesn't fit into a mod r/m byte so we need an extra register
1251 unsigned ScratchReg2;
1253 if (CompareStackPointer) {
1254 // The primary scratch register is available for holding the TLS offset
1255 ScratchReg2 = GetScratchRegister(Is64Bit, MF, true);
1256 SaveScratch2 = false;
1258 // Need to use a second register to hold the TLS offset
1259 ScratchReg2 = GetScratchRegister(Is64Bit, MF, false);
1261 // Unfortunately, with fastcc the second scratch register may hold an arg
1262 SaveScratch2 = MF.getRegInfo().isLiveIn(ScratchReg2);
1265 // If Scratch2 is live-in then it needs to be saved
1266 assert((!MF.getRegInfo().isLiveIn(ScratchReg2) || SaveScratch2) &&
1267 "Scratch register is live-in and not saved");
1270 BuildMI(checkMBB, DL, TII.get(X86::PUSH32r))
1271 .addReg(ScratchReg2, RegState::Kill);
1273 BuildMI(checkMBB, DL, TII.get(X86::MOV32ri), ScratchReg2)
1275 BuildMI(checkMBB, DL, TII.get(X86::CMP32rm))
1277 .addReg(ScratchReg2).addImm(1).addReg(0)
1282 BuildMI(checkMBB, DL, TII.get(X86::POP32r), ScratchReg2);
1286 // This jump is taken if SP >= (Stacklet Limit + Stack Space required).
1287 // It jumps to normal execution of the function body.
1288 BuildMI(checkMBB, DL, TII.get(X86::JA_4)).addMBB(&prologueMBB);
1290 // On 32 bit we first push the arguments size and then the frame size. On 64
1291 // bit, we pass the stack frame size in r10 and the argument size in r11.
1293 // Functions with nested arguments use R10, so it needs to be saved across
1294 // the call to _morestack
1297 BuildMI(allocMBB, DL, TII.get(X86::MOV64rr), X86::RAX).addReg(X86::R10);
1299 BuildMI(allocMBB, DL, TII.get(X86::MOV64ri), X86::R10)
1301 BuildMI(allocMBB, DL, TII.get(X86::MOV64ri), X86::R11)
1302 .addImm(X86FI->getArgumentStackSize());
1303 MF.getRegInfo().setPhysRegUsed(X86::R10);
1304 MF.getRegInfo().setPhysRegUsed(X86::R11);
1306 BuildMI(allocMBB, DL, TII.get(X86::PUSHi32))
1307 .addImm(X86FI->getArgumentStackSize());
1308 BuildMI(allocMBB, DL, TII.get(X86::PUSHi32))
1312 // __morestack is in libgcc
1314 BuildMI(allocMBB, DL, TII.get(X86::CALL64pcrel32))
1315 .addExternalSymbol("__morestack");
1317 BuildMI(allocMBB, DL, TII.get(X86::CALLpcrel32))
1318 .addExternalSymbol("__morestack");
1321 BuildMI(allocMBB, DL, TII.get(X86::MORESTACK_RET_RESTORE_R10));
1323 BuildMI(allocMBB, DL, TII.get(X86::MORESTACK_RET));
1325 allocMBB->addSuccessor(&prologueMBB);
1327 checkMBB->addSuccessor(allocMBB);
1328 checkMBB->addSuccessor(&prologueMBB);
1335 /// Erlang programs may need a special prologue to handle the stack size they
1336 /// might need at runtime. That is because Erlang/OTP does not implement a C
1337 /// stack but uses a custom implementation of hybrid stack/heap architecture.
1338 /// (for more information see Eric Stenman's Ph.D. thesis:
1339 /// http://publications.uu.se/uu/fulltext/nbn_se_uu_diva-2688.pdf)
1342 /// temp0 = sp - MaxStack
1343 /// if( temp0 < SP_LIMIT(P) ) goto IncStack else goto OldStart
1347 /// call inc_stack # doubles the stack space
1348 /// temp0 = sp - MaxStack
1349 /// if( temp0 < SP_LIMIT(P) ) goto IncStack else goto OldStart
1350 void X86FrameLowering::adjustForHiPEPrologue(MachineFunction &MF) const {
1351 const X86InstrInfo &TII = *TM.getInstrInfo();
1352 MachineFrameInfo *MFI = MF.getFrameInfo();
1353 const unsigned SlotSize = TM.getRegisterInfo()->getSlotSize();
1354 const bool Is64Bit = STI.is64Bit();
1356 // HiPE-specific values
1357 const unsigned HipeLeafWords = 24;
1358 const unsigned CCRegisteredArgs = Is64Bit ? 6 : 5;
1359 const unsigned Guaranteed = HipeLeafWords * SlotSize;
1360 unsigned CallerStkArity = MF.getFunction()->arg_size() > CCRegisteredArgs ?
1361 MF.getFunction()->arg_size() - CCRegisteredArgs : 0;
1362 unsigned MaxStack = MFI->getStackSize() + CallerStkArity*SlotSize + SlotSize;
1364 assert(STI.isTargetLinux() &&
1365 "HiPE prologue is only supported on Linux operating systems.");
1367 // Compute the largest caller's frame that is needed to fit the callees'
1368 // frames. This 'MaxStack' is computed from:
1370 // a) the fixed frame size, which is the space needed for all spilled temps,
1371 // b) outgoing on-stack parameter areas, and
1372 // c) the minimum stack space this function needs to make available for the
1373 // functions it calls (a tunable ABI property).
1374 if (MFI->hasCalls()) {
1375 unsigned MoreStackForCalls = 0;
1377 for (MachineFunction::iterator MBBI = MF.begin(), MBBE = MF.end();
1378 MBBI != MBBE; ++MBBI)
1379 for (MachineBasicBlock::iterator MI = MBBI->begin(), ME = MBBI->end();
1384 // Get callee operand.
1385 const MachineOperand &MO = MI->getOperand(0);
1387 // Only take account of global function calls (no closures etc.).
1391 const Function *F = dyn_cast<Function>(MO.getGlobal());
1395 // Do not update 'MaxStack' for primitive and built-in functions
1396 // (encoded with names either starting with "erlang."/"bif_" or not
1397 // having a ".", such as a simple <Module>.<Function>.<Arity>, or an
1398 // "_", such as the BIF "suspend_0") as they are executed on another
1400 if (F->getName().find("erlang.") != StringRef::npos ||
1401 F->getName().find("bif_") != StringRef::npos ||
1402 F->getName().find_first_of("._") == StringRef::npos)
1405 unsigned CalleeStkArity =
1406 F->arg_size() > CCRegisteredArgs ? F->arg_size()-CCRegisteredArgs : 0;
1407 if (HipeLeafWords - 1 > CalleeStkArity)
1408 MoreStackForCalls = std::max(MoreStackForCalls,
1409 (HipeLeafWords - 1 - CalleeStkArity) * SlotSize);
1411 MaxStack += MoreStackForCalls;
1414 // If the stack frame needed is larger than the guaranteed then runtime checks
1415 // and calls to "inc_stack_0" BIF should be inserted in the assembly prologue.
1416 if (MaxStack > Guaranteed) {
1417 MachineBasicBlock &prologueMBB = MF.front();
1418 MachineBasicBlock *stackCheckMBB = MF.CreateMachineBasicBlock();
1419 MachineBasicBlock *incStackMBB = MF.CreateMachineBasicBlock();
1421 for (MachineBasicBlock::livein_iterator I = prologueMBB.livein_begin(),
1422 E = prologueMBB.livein_end(); I != E; I++) {
1423 stackCheckMBB->addLiveIn(*I);
1424 incStackMBB->addLiveIn(*I);
1427 MF.push_front(incStackMBB);
1428 MF.push_front(stackCheckMBB);
1430 unsigned ScratchReg, SPReg, PReg, SPLimitOffset;
1431 unsigned LEAop, CMPop, CALLop;
1435 LEAop = X86::LEA64r;
1436 CMPop = X86::CMP64rm;
1437 CALLop = X86::CALL64pcrel32;
1438 SPLimitOffset = 0x90;
1442 LEAop = X86::LEA32r;
1443 CMPop = X86::CMP32rm;
1444 CALLop = X86::CALLpcrel32;
1445 SPLimitOffset = 0x4c;
1448 ScratchReg = GetScratchRegister(Is64Bit, MF, true);
1449 assert(!MF.getRegInfo().isLiveIn(ScratchReg) &&
1450 "HiPE prologue scratch register is live-in");
1452 // Create new MBB for StackCheck:
1453 addRegOffset(BuildMI(stackCheckMBB, DL, TII.get(LEAop), ScratchReg),
1454 SPReg, false, -MaxStack);
1455 // SPLimitOffset is in a fixed heap location (pointed by BP).
1456 addRegOffset(BuildMI(stackCheckMBB, DL, TII.get(CMPop))
1457 .addReg(ScratchReg), PReg, false, SPLimitOffset);
1458 BuildMI(stackCheckMBB, DL, TII.get(X86::JAE_4)).addMBB(&prologueMBB);
1460 // Create new MBB for IncStack:
1461 BuildMI(incStackMBB, DL, TII.get(CALLop)).
1462 addExternalSymbol("inc_stack_0");
1463 addRegOffset(BuildMI(incStackMBB, DL, TII.get(LEAop), ScratchReg),
1464 SPReg, false, -MaxStack);
1465 addRegOffset(BuildMI(incStackMBB, DL, TII.get(CMPop))
1466 .addReg(ScratchReg), PReg, false, SPLimitOffset);
1467 BuildMI(incStackMBB, DL, TII.get(X86::JLE_4)).addMBB(incStackMBB);
1469 stackCheckMBB->addSuccessor(&prologueMBB, 99);
1470 stackCheckMBB->addSuccessor(incStackMBB, 1);
1471 incStackMBB->addSuccessor(&prologueMBB, 99);
1472 incStackMBB->addSuccessor(incStackMBB, 1);
1479 void X86FrameLowering::
1480 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
1481 MachineBasicBlock::iterator I) const {
1482 const X86InstrInfo &TII = *TM.getInstrInfo();
1483 const X86RegisterInfo &RegInfo = *TM.getRegisterInfo();
1484 unsigned StackPtr = RegInfo.getStackRegister();
1485 bool reseveCallFrame = hasReservedCallFrame(MF);
1486 int Opcode = I->getOpcode();
1487 bool isDestroy = Opcode == TII.getCallFrameDestroyOpcode();
1488 bool IsLP64 = STI.isTarget64BitLP64();
1489 DebugLoc DL = I->getDebugLoc();
1490 uint64_t Amount = !reseveCallFrame ? I->getOperand(0).getImm() : 0;
1491 uint64_t CalleeAmt = isDestroy ? I->getOperand(1).getImm() : 0;
1494 if (!reseveCallFrame) {
1495 // If the stack pointer can be changed after prologue, turn the
1496 // adjcallstackup instruction into a 'sub ESP, <amt>' and the
1497 // adjcallstackdown instruction into 'add ESP, <amt>'
1498 // TODO: consider using push / pop instead of sub + store / add
1502 // We need to keep the stack aligned properly. To do this, we round the
1503 // amount of space needed for the outgoing arguments up to the next
1504 // alignment boundary.
1505 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
1506 Amount = (Amount + StackAlign - 1) / StackAlign * StackAlign;
1508 MachineInstr *New = 0;
1509 if (Opcode == TII.getCallFrameSetupOpcode()) {
1510 New = BuildMI(MF, DL, TII.get(getSUBriOpcode(IsLP64, Amount)),
1515 assert(Opcode == TII.getCallFrameDestroyOpcode());
1517 // Factor out the amount the callee already popped.
1518 Amount -= CalleeAmt;
1521 unsigned Opc = getADDriOpcode(IsLP64, Amount);
1522 New = BuildMI(MF, DL, TII.get(Opc), StackPtr)
1523 .addReg(StackPtr).addImm(Amount);
1528 // The EFLAGS implicit def is dead.
1529 New->getOperand(3).setIsDead();
1531 // Replace the pseudo instruction with a new instruction.
1538 if (Opcode == TII.getCallFrameDestroyOpcode() && CalleeAmt) {
1539 // If we are performing frame pointer elimination and if the callee pops
1540 // something off the stack pointer, add it back. We do this until we have
1541 // more advanced stack pointer tracking ability.
1542 unsigned Opc = getSUBriOpcode(IsLP64, CalleeAmt);
1543 MachineInstr *New = BuildMI(MF, DL, TII.get(Opc), StackPtr)
1544 .addReg(StackPtr).addImm(CalleeAmt);
1546 // The EFLAGS implicit def is dead.
1547 New->getOperand(3).setIsDead();
1549 // We are not tracking the stack pointer adjustment by the callee, so make
1550 // sure we restore the stack pointer immediately after the call, there may
1551 // be spill code inserted between the CALL and ADJCALLSTACKUP instructions.
1552 MachineBasicBlock::iterator B = MBB.begin();
1553 while (I != B && !llvm::prior(I)->isCall())