1 //===-- PPCFrameLowering.cpp - PPC 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 PPC implementation of TargetFrameLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "PPCFrameLowering.h"
15 #include "PPCInstrBuilder.h"
16 #include "PPCInstrInfo.h"
17 #include "PPCMachineFunctionInfo.h"
18 #include "PPCSubtarget.h"
19 #include "PPCTargetMachine.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/CodeGen/RegisterScavenging.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/Target/TargetOptions.h"
31 /// VRRegNo - Map from a numbered VR register to its enum value.
33 static const MCPhysReg VRRegNo[] = {
34 PPC::V0 , PPC::V1 , PPC::V2 , PPC::V3 , PPC::V4 , PPC::V5 , PPC::V6 , PPC::V7 ,
35 PPC::V8 , PPC::V9 , PPC::V10, PPC::V11, PPC::V12, PPC::V13, PPC::V14, PPC::V15,
36 PPC::V16, PPC::V17, PPC::V18, PPC::V19, PPC::V20, PPC::V21, PPC::V22, PPC::V23,
37 PPC::V24, PPC::V25, PPC::V26, PPC::V27, PPC::V28, PPC::V29, PPC::V30, PPC::V31
40 static unsigned computeReturnSaveOffset(const PPCSubtarget &STI) {
41 if (STI.isDarwinABI())
42 return STI.isPPC64() ? 16 : 8;
44 return STI.isPPC64() ? 16 : 4;
47 static unsigned computeTOCSaveOffset(const PPCSubtarget &STI) {
48 return STI.isELFv2ABI() ? 24 : 40;
51 static unsigned computeFramePointerSaveOffset(const PPCSubtarget &STI) {
52 // For the Darwin ABI:
53 // We cannot use the TOC save slot (offset +20) in the PowerPC linkage area
54 // for saving the frame pointer (if needed.) While the published ABI has
55 // not used this slot since at least MacOSX 10.2, there is older code
56 // around that does use it, and that needs to continue to work.
57 if (STI.isDarwinABI())
58 return STI.isPPC64() ? -8U : -4U;
60 // SVR4 ABI: First slot in the general register save area.
61 return STI.isPPC64() ? -8U : -4U;
64 static unsigned computeLinkageSize(const PPCSubtarget &STI) {
65 if (STI.isDarwinABI() || STI.isPPC64())
66 return (STI.isELFv2ABI() ? 4 : 6) * (STI.isPPC64() ? 8 : 4);
72 static unsigned computeBasePointerSaveOffset(const PPCSubtarget &STI) {
73 if (STI.isDarwinABI())
74 return STI.isPPC64() ? -16U : -8U;
76 // SVR4 ABI: First slot in the general register save area.
79 : STI.getTargetMachine().isPositionIndependent() ? -12U : -8U;
82 PPCFrameLowering::PPCFrameLowering(const PPCSubtarget &STI)
83 : TargetFrameLowering(TargetFrameLowering::StackGrowsDown,
84 STI.getPlatformStackAlignment(), 0),
85 Subtarget(STI), ReturnSaveOffset(computeReturnSaveOffset(Subtarget)),
86 TOCSaveOffset(computeTOCSaveOffset(Subtarget)),
87 FramePointerSaveOffset(computeFramePointerSaveOffset(Subtarget)),
88 LinkageSize(computeLinkageSize(Subtarget)),
89 BasePointerSaveOffset(computeBasePointerSaveOffset(STI)) {}
91 // With the SVR4 ABI, callee-saved registers have fixed offsets on the stack.
92 const PPCFrameLowering::SpillSlot *PPCFrameLowering::getCalleeSavedSpillSlots(
93 unsigned &NumEntries) const {
94 if (Subtarget.isDarwinABI()) {
96 if (Subtarget.isPPC64()) {
97 static const SpillSlot darwin64Offsets = {PPC::X31, -8};
98 return &darwin64Offsets;
100 static const SpillSlot darwinOffsets = {PPC::R31, -4};
101 return &darwinOffsets;
105 // Early exit if not using the SVR4 ABI.
106 if (!Subtarget.isSVR4ABI()) {
111 // Note that the offsets here overlap, but this is fixed up in
112 // processFunctionBeforeFrameFinalized.
114 static const SpillSlot Offsets[] = {
115 // Floating-point register save area offsets.
135 // General register save area offsets.
155 // CR save area offset. We map each of the nonvolatile CR fields
156 // to the slot for CR2, which is the first of the nonvolatile CR
157 // fields to be assigned, so that we only allocate one save slot.
158 // See PPCRegisterInfo::hasReservedSpillSlot() for more information.
161 // VRSAVE save area offset.
164 // Vector register save area
178 static const SpillSlot Offsets64[] = {
179 // Floating-point register save area offsets.
199 // General register save area offsets.
219 // VRSAVE save area offset.
222 // Vector register save area
236 if (Subtarget.isPPC64()) {
237 NumEntries = array_lengthof(Offsets64);
241 NumEntries = array_lengthof(Offsets);
247 /// RemoveVRSaveCode - We have found that this function does not need any code
248 /// to manipulate the VRSAVE register, even though it uses vector registers.
249 /// This can happen when the only registers used are known to be live in or out
250 /// of the function. Remove all of the VRSAVE related code from the function.
251 /// FIXME: The removal of the code results in a compile failure at -O0 when the
252 /// function contains a function call, as the GPR containing original VRSAVE
253 /// contents is spilled and reloaded around the call. Without the prolog code,
254 /// the spill instruction refers to an undefined register. This code needs
255 /// to account for all uses of that GPR.
256 static void RemoveVRSaveCode(MachineInstr &MI) {
257 MachineBasicBlock *Entry = MI.getParent();
258 MachineFunction *MF = Entry->getParent();
260 // We know that the MTVRSAVE instruction immediately follows MI. Remove it.
261 MachineBasicBlock::iterator MBBI = MI;
263 assert(MBBI != Entry->end() && MBBI->getOpcode() == PPC::MTVRSAVE);
264 MBBI->eraseFromParent();
266 bool RemovedAllMTVRSAVEs = true;
267 // See if we can find and remove the MTVRSAVE instruction from all of the
269 for (MachineFunction::iterator I = MF->begin(), E = MF->end(); I != E; ++I) {
270 // If last instruction is a return instruction, add an epilogue
271 if (I->isReturnBlock()) {
272 bool FoundIt = false;
273 for (MBBI = I->end(); MBBI != I->begin(); ) {
275 if (MBBI->getOpcode() == PPC::MTVRSAVE) {
276 MBBI->eraseFromParent(); // remove it.
281 RemovedAllMTVRSAVEs &= FoundIt;
285 // If we found and removed all MTVRSAVE instructions, remove the read of
287 if (RemovedAllMTVRSAVEs) {
289 assert(MBBI != Entry->begin() && "UPDATE_VRSAVE is first instr in block?");
291 assert(MBBI->getOpcode() == PPC::MFVRSAVE && "VRSAVE instrs wandered?");
292 MBBI->eraseFromParent();
295 // Finally, nuke the UPDATE_VRSAVE.
296 MI.eraseFromParent();
299 // HandleVRSaveUpdate - MI is the UPDATE_VRSAVE instruction introduced by the
300 // instruction selector. Based on the vector registers that have been used,
301 // transform this into the appropriate ORI instruction.
302 static void HandleVRSaveUpdate(MachineInstr &MI, const TargetInstrInfo &TII) {
303 MachineFunction *MF = MI.getParent()->getParent();
304 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
305 DebugLoc dl = MI.getDebugLoc();
307 const MachineRegisterInfo &MRI = MF->getRegInfo();
308 unsigned UsedRegMask = 0;
309 for (unsigned i = 0; i != 32; ++i)
310 if (MRI.isPhysRegModified(VRRegNo[i]))
311 UsedRegMask |= 1 << (31-i);
313 // Live in and live out values already must be in the mask, so don't bother
315 for (MachineRegisterInfo::livein_iterator
316 I = MF->getRegInfo().livein_begin(),
317 E = MF->getRegInfo().livein_end(); I != E; ++I) {
318 unsigned RegNo = TRI->getEncodingValue(I->first);
319 if (VRRegNo[RegNo] == I->first) // If this really is a vector reg.
320 UsedRegMask &= ~(1 << (31-RegNo)); // Doesn't need to be marked.
323 // Live out registers appear as use operands on return instructions.
324 for (MachineFunction::const_iterator BI = MF->begin(), BE = MF->end();
325 UsedRegMask != 0 && BI != BE; ++BI) {
326 const MachineBasicBlock &MBB = *BI;
327 if (!MBB.isReturnBlock())
329 const MachineInstr &Ret = MBB.back();
330 for (unsigned I = 0, E = Ret.getNumOperands(); I != E; ++I) {
331 const MachineOperand &MO = Ret.getOperand(I);
332 if (!MO.isReg() || !PPC::VRRCRegClass.contains(MO.getReg()))
334 unsigned RegNo = TRI->getEncodingValue(MO.getReg());
335 UsedRegMask &= ~(1 << (31-RegNo));
339 // If no registers are used, turn this into a copy.
340 if (UsedRegMask == 0) {
341 // Remove all VRSAVE code.
342 RemoveVRSaveCode(MI);
346 unsigned SrcReg = MI.getOperand(1).getReg();
347 unsigned DstReg = MI.getOperand(0).getReg();
349 if ((UsedRegMask & 0xFFFF) == UsedRegMask) {
350 if (DstReg != SrcReg)
351 BuildMI(*MI.getParent(), MI, dl, TII.get(PPC::ORI), DstReg)
353 .addImm(UsedRegMask);
355 BuildMI(*MI.getParent(), MI, dl, TII.get(PPC::ORI), DstReg)
356 .addReg(SrcReg, RegState::Kill)
357 .addImm(UsedRegMask);
358 } else if ((UsedRegMask & 0xFFFF0000) == UsedRegMask) {
359 if (DstReg != SrcReg)
360 BuildMI(*MI.getParent(), MI, dl, TII.get(PPC::ORIS), DstReg)
362 .addImm(UsedRegMask >> 16);
364 BuildMI(*MI.getParent(), MI, dl, TII.get(PPC::ORIS), DstReg)
365 .addReg(SrcReg, RegState::Kill)
366 .addImm(UsedRegMask >> 16);
368 if (DstReg != SrcReg)
369 BuildMI(*MI.getParent(), MI, dl, TII.get(PPC::ORIS), DstReg)
371 .addImm(UsedRegMask >> 16);
373 BuildMI(*MI.getParent(), MI, dl, TII.get(PPC::ORIS), DstReg)
374 .addReg(SrcReg, RegState::Kill)
375 .addImm(UsedRegMask >> 16);
377 BuildMI(*MI.getParent(), MI, dl, TII.get(PPC::ORI), DstReg)
378 .addReg(DstReg, RegState::Kill)
379 .addImm(UsedRegMask & 0xFFFF);
382 // Remove the old UPDATE_VRSAVE instruction.
383 MI.eraseFromParent();
386 static bool spillsCR(const MachineFunction &MF) {
387 const PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
388 return FuncInfo->isCRSpilled();
391 static bool spillsVRSAVE(const MachineFunction &MF) {
392 const PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
393 return FuncInfo->isVRSAVESpilled();
396 static bool hasSpills(const MachineFunction &MF) {
397 const PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
398 return FuncInfo->hasSpills();
401 static bool hasNonRISpills(const MachineFunction &MF) {
402 const PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
403 return FuncInfo->hasNonRISpills();
406 /// MustSaveLR - Return true if this function requires that we save the LR
407 /// register onto the stack in the prolog and restore it in the epilog of the
409 static bool MustSaveLR(const MachineFunction &MF, unsigned LR) {
410 const PPCFunctionInfo *MFI = MF.getInfo<PPCFunctionInfo>();
412 // We need a save/restore of LR if there is any def of LR (which is
413 // defined by calls, including the PIC setup sequence), or if there is
414 // some use of the LR stack slot (e.g. for builtin_return_address).
415 // (LR comes in 32 and 64 bit versions.)
416 MachineRegisterInfo::def_iterator RI = MF.getRegInfo().def_begin(LR);
417 return RI !=MF.getRegInfo().def_end() || MFI->isLRStoreRequired();
420 /// determineFrameLayout - Determine the size of the frame and maximum call
422 unsigned PPCFrameLowering::determineFrameLayout(MachineFunction &MF,
424 bool UseEstimate) const {
425 MachineFrameInfo &MFI = MF.getFrameInfo();
427 // Get the number of bytes to allocate from the FrameInfo
429 UseEstimate ? MFI.estimateStackSize(MF) : MFI.getStackSize();
431 // Get stack alignments. The frame must be aligned to the greatest of these:
432 unsigned TargetAlign = getStackAlignment(); // alignment required per the ABI
433 unsigned MaxAlign = MFI.getMaxAlignment(); // algmt required by data in frame
434 unsigned AlignMask = std::max(MaxAlign, TargetAlign) - 1;
436 const PPCRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
438 // If we are a leaf function, and use up to 224 bytes of stack space,
439 // don't have a frame pointer, calls, or dynamic alloca then we do not need
440 // to adjust the stack pointer (we fit in the Red Zone).
441 // The 32-bit SVR4 ABI has no Red Zone. However, it can still generate
442 // stackless code if all local vars are reg-allocated.
443 bool DisableRedZone = MF.getFunction()->hasFnAttribute(Attribute::NoRedZone);
444 unsigned LR = RegInfo->getRARegister();
445 if (!DisableRedZone &&
446 (Subtarget.isPPC64() || // 32-bit SVR4, no stack-
447 !Subtarget.isSVR4ABI() || // allocated locals.
449 FrameSize <= 224 && // Fits in red zone.
450 !MFI.hasVarSizedObjects() && // No dynamic alloca.
451 !MFI.adjustsStack() && // No calls.
452 !MustSaveLR(MF, LR) &&
453 !RegInfo->hasBasePointer(MF)) { // No special alignment.
460 // Get the maximum call frame size of all the calls.
461 unsigned maxCallFrameSize = MFI.getMaxCallFrameSize();
463 // Maximum call frame needs to be at least big enough for linkage area.
464 unsigned minCallFrameSize = getLinkageSize();
465 maxCallFrameSize = std::max(maxCallFrameSize, minCallFrameSize);
467 // If we have dynamic alloca then maxCallFrameSize needs to be aligned so
468 // that allocations will be aligned.
469 if (MFI.hasVarSizedObjects())
470 maxCallFrameSize = (maxCallFrameSize + AlignMask) & ~AlignMask;
472 // Update maximum call frame size.
474 MFI.setMaxCallFrameSize(maxCallFrameSize);
476 // Include call frame size in total.
477 FrameSize += maxCallFrameSize;
479 // Make sure the frame is aligned.
480 FrameSize = (FrameSize + AlignMask) & ~AlignMask;
482 // Update frame info.
484 MFI.setStackSize(FrameSize);
489 // hasFP - Return true if the specified function actually has a dedicated frame
491 bool PPCFrameLowering::hasFP(const MachineFunction &MF) const {
492 const MachineFrameInfo &MFI = MF.getFrameInfo();
493 // FIXME: This is pretty much broken by design: hasFP() might be called really
494 // early, before the stack layout was calculated and thus hasFP() might return
495 // true or false here depending on the time of call.
496 return (MFI.getStackSize()) && needsFP(MF);
499 // needsFP - Return true if the specified function should have a dedicated frame
500 // pointer register. This is true if the function has variable sized allocas or
501 // if frame pointer elimination is disabled.
502 bool PPCFrameLowering::needsFP(const MachineFunction &MF) const {
503 const MachineFrameInfo &MFI = MF.getFrameInfo();
505 // Naked functions have no stack frame pushed, so we don't have a frame
507 if (MF.getFunction()->hasFnAttribute(Attribute::Naked))
510 return MF.getTarget().Options.DisableFramePointerElim(MF) ||
511 MFI.hasVarSizedObjects() || MFI.hasStackMap() || MFI.hasPatchPoint() ||
512 (MF.getTarget().Options.GuaranteedTailCallOpt &&
513 MF.getInfo<PPCFunctionInfo>()->hasFastCall());
516 void PPCFrameLowering::replaceFPWithRealFP(MachineFunction &MF) const {
517 bool is31 = needsFP(MF);
518 unsigned FPReg = is31 ? PPC::R31 : PPC::R1;
519 unsigned FP8Reg = is31 ? PPC::X31 : PPC::X1;
521 const PPCRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
522 bool HasBP = RegInfo->hasBasePointer(MF);
523 unsigned BPReg = HasBP ? (unsigned) RegInfo->getBaseRegister(MF) : FPReg;
524 unsigned BP8Reg = HasBP ? (unsigned) PPC::X30 : FPReg;
526 for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
528 for (MachineBasicBlock::iterator MBBI = BI->end(); MBBI != BI->begin(); ) {
530 for (unsigned I = 0, E = MBBI->getNumOperands(); I != E; ++I) {
531 MachineOperand &MO = MBBI->getOperand(I);
535 switch (MO.getReg()) {
554 /* This function will do the following:
555 - If MBB is an entry or exit block, set SR1 and SR2 to R0 and R12
556 respectively (defaults recommended by the ABI) and return true
557 - If MBB is not an entry block, initialize the register scavenger and look
558 for available registers.
559 - If the defaults (R0/R12) are available, return true
560 - If TwoUniqueRegsRequired is set to true, it looks for two unique
561 registers. Otherwise, look for a single available register.
562 - If the required registers are found, set SR1 and SR2 and return true.
563 - If the required registers are not found, set SR2 or both SR1 and SR2 to
564 PPC::NoRegister and return false.
566 Note that if both SR1 and SR2 are valid parameters and TwoUniqueRegsRequired
567 is not set, this function will attempt to find two different registers, but
568 still return true if only one register is available (and set SR1 == SR2).
571 PPCFrameLowering::findScratchRegister(MachineBasicBlock *MBB,
573 bool TwoUniqueRegsRequired,
575 unsigned *SR2) const {
577 unsigned R0 = Subtarget.isPPC64() ? PPC::X0 : PPC::R0;
578 unsigned R12 = Subtarget.isPPC64() ? PPC::X12 : PPC::R12;
580 // Set the defaults for the two scratch registers.
585 assert (SR1 && "Asking for the second scratch register but not the first?");
589 // If MBB is an entry or exit block, use R0 and R12 as the scratch registers.
590 if ((UseAtEnd && MBB->isReturnBlock()) ||
591 (!UseAtEnd && (&MBB->getParent()->front() == MBB)))
594 RS.enterBasicBlock(*MBB);
596 if (UseAtEnd && !MBB->empty()) {
597 // The scratch register will be used at the end of the block, so must
598 // consider all registers used within the block
600 MachineBasicBlock::iterator MBBI = MBB->getFirstTerminator();
601 // If no terminator, back iterator up to previous instruction.
602 if (MBBI == MBB->end())
603 MBBI = std::prev(MBBI);
605 if (MBBI != MBB->begin())
609 // If the two registers are available, we're all good.
610 // Note that we only return here if both R0 and R12 are available because
611 // although the function may not require two unique registers, it may benefit
612 // from having two so we should try to provide them.
613 if (!RS.isRegUsed(R0) && !RS.isRegUsed(R12))
616 // Get the list of callee-saved registers for the target.
617 const PPCRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
618 const MCPhysReg *CSRegs = RegInfo->getCalleeSavedRegs(MBB->getParent());
620 // Get all the available registers in the block.
621 BitVector BV = RS.getRegsAvailable(Subtarget.isPPC64() ? &PPC::G8RCRegClass :
624 // We shouldn't use callee-saved registers as scratch registers as they may be
625 // available when looking for a candidate block for shrink wrapping but not
626 // available when the actual prologue/epilogue is being emitted because they
627 // were added as live-in to the prologue block by PrologueEpilogueInserter.
628 for (int i = 0; CSRegs[i]; ++i)
631 // Set the first scratch register to the first available one.
633 int FirstScratchReg = BV.find_first();
634 *SR1 = FirstScratchReg == -1 ? (unsigned)PPC::NoRegister : FirstScratchReg;
637 // If there is another one available, set the second scratch register to that.
638 // Otherwise, set it to either PPC::NoRegister if this function requires two
639 // or to whatever SR1 is set to if this function doesn't require two.
641 int SecondScratchReg = BV.find_next(*SR1);
642 if (SecondScratchReg != -1)
643 *SR2 = SecondScratchReg;
645 *SR2 = TwoUniqueRegsRequired ? (unsigned)PPC::NoRegister : *SR1;
648 // Now that we've done our best to provide both registers, double check
649 // whether we were unable to provide enough.
650 if (BV.count() < (TwoUniqueRegsRequired ? 2U : 1U))
656 // We need a scratch register for spilling LR and for spilling CR. By default,
657 // we use two scratch registers to hide latency. However, if only one scratch
658 // register is available, we can adjust for that by not overlapping the spill
659 // code. However, if we need to realign the stack (i.e. have a base pointer)
660 // and the stack frame is large, we need two scratch registers.
662 PPCFrameLowering::twoUniqueScratchRegsRequired(MachineBasicBlock *MBB) const {
663 const PPCRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
664 MachineFunction &MF = *(MBB->getParent());
665 bool HasBP = RegInfo->hasBasePointer(MF);
666 unsigned FrameSize = determineFrameLayout(MF, false);
667 int NegFrameSize = -FrameSize;
668 bool IsLargeFrame = !isInt<16>(NegFrameSize);
669 MachineFrameInfo &MFI = MF.getFrameInfo();
670 unsigned MaxAlign = MFI.getMaxAlignment();
671 bool HasRedZone = Subtarget.isPPC64() || !Subtarget.isSVR4ABI();
673 return (IsLargeFrame || !HasRedZone) && HasBP && MaxAlign > 1;
676 bool PPCFrameLowering::canUseAsPrologue(const MachineBasicBlock &MBB) const {
677 MachineBasicBlock *TmpMBB = const_cast<MachineBasicBlock *>(&MBB);
679 return findScratchRegister(TmpMBB, false,
680 twoUniqueScratchRegsRequired(TmpMBB));
683 bool PPCFrameLowering::canUseAsEpilogue(const MachineBasicBlock &MBB) const {
684 MachineBasicBlock *TmpMBB = const_cast<MachineBasicBlock *>(&MBB);
686 return findScratchRegister(TmpMBB, true);
689 void PPCFrameLowering::emitPrologue(MachineFunction &MF,
690 MachineBasicBlock &MBB) const {
691 MachineBasicBlock::iterator MBBI = MBB.begin();
692 MachineFrameInfo &MFI = MF.getFrameInfo();
693 const PPCInstrInfo &TII = *Subtarget.getInstrInfo();
694 const PPCRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
696 MachineModuleInfo &MMI = MF.getMMI();
697 const MCRegisterInfo *MRI = MMI.getContext().getRegisterInfo();
699 bool needsCFI = MMI.hasDebugInfo() ||
700 MF.getFunction()->needsUnwindTableEntry();
702 // Get processor type.
703 bool isPPC64 = Subtarget.isPPC64();
705 bool isSVR4ABI = Subtarget.isSVR4ABI();
706 bool isELFv2ABI = Subtarget.isELFv2ABI();
707 assert((Subtarget.isDarwinABI() || isSVR4ABI) &&
708 "Currently only Darwin and SVR4 ABIs are supported for PowerPC.");
710 // Scan the prolog, looking for an UPDATE_VRSAVE instruction. If we find it,
713 for (unsigned i = 0; MBBI != MBB.end(); ++i, ++MBBI) {
714 if (MBBI->getOpcode() == PPC::UPDATE_VRSAVE) {
715 HandleVRSaveUpdate(*MBBI, TII);
720 // Move MBBI back to the beginning of the prologue block.
723 // Work out frame sizes.
724 unsigned FrameSize = determineFrameLayout(MF);
725 int NegFrameSize = -FrameSize;
726 if (!isInt<32>(NegFrameSize))
727 llvm_unreachable("Unhandled stack size!");
729 if (MFI.isFrameAddressTaken())
730 replaceFPWithRealFP(MF);
732 // Check if the link register (LR) must be saved.
733 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
734 bool MustSaveLR = FI->mustSaveLR();
735 const SmallVectorImpl<unsigned> &MustSaveCRs = FI->getMustSaveCRs();
736 bool MustSaveCR = !MustSaveCRs.empty();
737 // Do we have a frame pointer and/or base pointer for this function?
738 bool HasFP = hasFP(MF);
739 bool HasBP = RegInfo->hasBasePointer(MF);
740 bool HasRedZone = isPPC64 || !isSVR4ABI;
742 unsigned SPReg = isPPC64 ? PPC::X1 : PPC::R1;
743 unsigned BPReg = RegInfo->getBaseRegister(MF);
744 unsigned FPReg = isPPC64 ? PPC::X31 : PPC::R31;
745 unsigned LRReg = isPPC64 ? PPC::LR8 : PPC::LR;
746 unsigned ScratchReg = 0;
747 unsigned TempReg = isPPC64 ? PPC::X12 : PPC::R12; // another scratch reg
748 // ...(R12/X12 is volatile in both Darwin & SVR4, & can't be a function arg.)
749 const MCInstrDesc& MFLRInst = TII.get(isPPC64 ? PPC::MFLR8
751 const MCInstrDesc& StoreInst = TII.get(isPPC64 ? PPC::STD
753 const MCInstrDesc& StoreUpdtInst = TII.get(isPPC64 ? PPC::STDU
755 const MCInstrDesc& StoreUpdtIdxInst = TII.get(isPPC64 ? PPC::STDUX
757 const MCInstrDesc& LoadImmShiftedInst = TII.get(isPPC64 ? PPC::LIS8
759 const MCInstrDesc& OrImmInst = TII.get(isPPC64 ? PPC::ORI8
761 const MCInstrDesc& OrInst = TII.get(isPPC64 ? PPC::OR8
763 const MCInstrDesc& SubtractCarryingInst = TII.get(isPPC64 ? PPC::SUBFC8
765 const MCInstrDesc& SubtractImmCarryingInst = TII.get(isPPC64 ? PPC::SUBFIC8
768 // Regarding this assert: Even though LR is saved in the caller's frame (i.e.,
769 // LROffset is positive), that slot is callee-owned. Because PPC32 SVR4 has no
770 // Red Zone, an asynchronous event (a form of "callee") could claim a frame &
771 // overwrite it, so PPC32 SVR4 must claim at least a minimal frame to save LR.
772 assert((isPPC64 || !isSVR4ABI || !(!FrameSize && (MustSaveLR || HasFP))) &&
773 "FrameSize must be >0 to save/restore the FP or LR for 32-bit SVR4.");
775 // Using the same bool variable as below to suppress compiler warnings.
776 bool SingleScratchReg =
777 findScratchRegister(&MBB, false, twoUniqueScratchRegsRequired(&MBB),
778 &ScratchReg, &TempReg);
779 assert(SingleScratchReg &&
780 "Required number of registers not available in this block");
782 SingleScratchReg = ScratchReg == TempReg;
784 int LROffset = getReturnSaveOffset();
789 MachineFrameInfo &MFI = MF.getFrameInfo();
790 int FPIndex = FI->getFramePointerSaveIndex();
791 assert(FPIndex && "No Frame Pointer Save Slot!");
792 FPOffset = MFI.getObjectOffset(FPIndex);
794 FPOffset = getFramePointerSaveOffset();
801 MachineFrameInfo &MFI = MF.getFrameInfo();
802 int BPIndex = FI->getBasePointerSaveIndex();
803 assert(BPIndex && "No Base Pointer Save Slot!");
804 BPOffset = MFI.getObjectOffset(BPIndex);
806 BPOffset = getBasePointerSaveOffset();
811 if (FI->usesPICBase()) {
812 MachineFrameInfo &MFI = MF.getFrameInfo();
813 int PBPIndex = FI->getPICBasePointerSaveIndex();
814 assert(PBPIndex && "No PIC Base Pointer Save Slot!");
815 PBPOffset = MFI.getObjectOffset(PBPIndex);
818 // Get stack alignments.
819 unsigned MaxAlign = MFI.getMaxAlignment();
820 if (HasBP && MaxAlign > 1)
821 assert(isPowerOf2_32(MaxAlign) && isInt<16>(MaxAlign) &&
822 "Invalid alignment!");
824 // Frames of 32KB & larger require special handling because they cannot be
825 // indexed into with a simple STDU/STWU/STD/STW immediate offset operand.
826 bool isLargeFrame = !isInt<16>(NegFrameSize);
828 assert((isPPC64 || !MustSaveCR) &&
829 "Prologue CR saving supported only in 64-bit mode");
831 // If we need to spill the CR and the LR but we don't have two separate
832 // registers available, we must spill them one at a time
833 if (MustSaveCR && SingleScratchReg && MustSaveLR) {
834 // In the ELFv2 ABI, we are not required to save all CR fields.
835 // If only one or two CR fields are clobbered, it is more efficient to use
836 // mfocrf to selectively save just those fields, because mfocrf has short
837 // latency compares to mfcr.
838 unsigned MfcrOpcode = PPC::MFCR8;
839 unsigned CrState = RegState::ImplicitKill;
840 if (isELFv2ABI && MustSaveCRs.size() == 1) {
841 MfcrOpcode = PPC::MFOCRF8;
842 CrState = RegState::Kill;
844 MachineInstrBuilder MIB =
845 BuildMI(MBB, MBBI, dl, TII.get(MfcrOpcode), TempReg);
846 for (unsigned i = 0, e = MustSaveCRs.size(); i != e; ++i)
847 MIB.addReg(MustSaveCRs[i], CrState);
848 BuildMI(MBB, MBBI, dl, TII.get(PPC::STW8))
849 .addReg(TempReg, getKillRegState(true))
855 BuildMI(MBB, MBBI, dl, MFLRInst, ScratchReg);
858 !(SingleScratchReg && MustSaveLR)) { // will only occur for PPC64
859 // In the ELFv2 ABI, we are not required to save all CR fields.
860 // If only one or two CR fields are clobbered, it is more efficient to use
861 // mfocrf to selectively save just those fields, because mfocrf has short
862 // latency compares to mfcr.
863 unsigned MfcrOpcode = PPC::MFCR8;
864 unsigned CrState = RegState::ImplicitKill;
865 if (isELFv2ABI && MustSaveCRs.size() == 1) {
866 MfcrOpcode = PPC::MFOCRF8;
867 CrState = RegState::Kill;
869 MachineInstrBuilder MIB =
870 BuildMI(MBB, MBBI, dl, TII.get(MfcrOpcode), TempReg);
871 for (unsigned i = 0, e = MustSaveCRs.size(); i != e; ++i)
872 MIB.addReg(MustSaveCRs[i], CrState);
877 BuildMI(MBB, MBBI, dl, StoreInst)
881 if (FI->usesPICBase())
882 BuildMI(MBB, MBBI, dl, StoreInst)
887 BuildMI(MBB, MBBI, dl, StoreInst)
894 BuildMI(MBB, MBBI, dl, StoreInst)
895 .addReg(ScratchReg, getKillRegState(true))
900 !(SingleScratchReg && MustSaveLR)) { // will only occur for PPC64
901 assert(HasRedZone && "A red zone is always available on PPC64");
902 BuildMI(MBB, MBBI, dl, TII.get(PPC::STW8))
903 .addReg(TempReg, getKillRegState(true))
908 // Skip the rest if this is a leaf function & all spills fit in the Red Zone.
912 // Adjust stack pointer: r1 += NegFrameSize.
913 // If there is a preferred stack alignment, align R1 now
915 if (HasBP && HasRedZone) {
916 // Save a copy of r1 as the base pointer.
917 BuildMI(MBB, MBBI, dl, OrInst, BPReg)
922 // Have we generated a STUX instruction to claim stack frame? If so,
923 // the negated frame size will be placed in ScratchReg.
924 bool HasSTUX = false;
926 // This condition must be kept in sync with canUseAsPrologue.
927 if (HasBP && MaxAlign > 1) {
929 BuildMI(MBB, MBBI, dl, TII.get(PPC::RLDICL), ScratchReg)
932 .addImm(64 - Log2_32(MaxAlign));
934 BuildMI(MBB, MBBI, dl, TII.get(PPC::RLWINM), ScratchReg)
937 .addImm(32 - Log2_32(MaxAlign))
940 BuildMI(MBB, MBBI, dl, SubtractImmCarryingInst, ScratchReg)
941 .addReg(ScratchReg, RegState::Kill)
942 .addImm(NegFrameSize);
944 assert(!SingleScratchReg && "Only a single scratch reg available");
945 BuildMI(MBB, MBBI, dl, LoadImmShiftedInst, TempReg)
946 .addImm(NegFrameSize >> 16);
947 BuildMI(MBB, MBBI, dl, OrImmInst, TempReg)
948 .addReg(TempReg, RegState::Kill)
949 .addImm(NegFrameSize & 0xFFFF);
950 BuildMI(MBB, MBBI, dl, SubtractCarryingInst, ScratchReg)
951 .addReg(ScratchReg, RegState::Kill)
952 .addReg(TempReg, RegState::Kill);
955 BuildMI(MBB, MBBI, dl, StoreUpdtIdxInst, SPReg)
956 .addReg(SPReg, RegState::Kill)
961 } else if (!isLargeFrame) {
962 BuildMI(MBB, MBBI, dl, StoreUpdtInst, SPReg)
964 .addImm(NegFrameSize)
968 BuildMI(MBB, MBBI, dl, LoadImmShiftedInst, ScratchReg)
969 .addImm(NegFrameSize >> 16);
970 BuildMI(MBB, MBBI, dl, OrImmInst, ScratchReg)
971 .addReg(ScratchReg, RegState::Kill)
972 .addImm(NegFrameSize & 0xFFFF);
973 BuildMI(MBB, MBBI, dl, StoreUpdtIdxInst, SPReg)
974 .addReg(SPReg, RegState::Kill)
981 assert(!isPPC64 && "A red zone is always available on PPC64");
983 // The negated frame size is in ScratchReg, and the SPReg has been
984 // decremented by the frame size: SPReg = old SPReg + ScratchReg.
985 // Since FPOffset, PBPOffset, etc. are relative to the beginning of
986 // the stack frame (i.e. the old SP), ideally, we would put the old
987 // SP into a register and use it as the base for the stores. The
988 // problem is that the only available register may be ScratchReg,
989 // which could be R0, and R0 cannot be used as a base address.
991 // First, set ScratchReg to the old SP. This may need to be modified
993 BuildMI(MBB, MBBI, dl, TII.get(PPC::SUBF), ScratchReg)
994 .addReg(ScratchReg, RegState::Kill)
997 if (ScratchReg == PPC::R0) {
998 // R0 cannot be used as a base register, but it can be used as an
999 // index in a store-indexed.
1002 // R0 += (FPOffset-LastOffset).
1003 // Need addic, since addi treats R0 as 0.
1004 BuildMI(MBB, MBBI, dl, TII.get(PPC::ADDIC), ScratchReg)
1006 .addImm(FPOffset-LastOffset);
1007 LastOffset = FPOffset;
1008 // Store FP into *R0.
1009 BuildMI(MBB, MBBI, dl, TII.get(PPC::STWX))
1010 .addReg(FPReg, RegState::Kill) // Save FP.
1012 .addReg(ScratchReg); // This will be the index (R0 is ok here).
1014 if (FI->usesPICBase()) {
1015 // R0 += (PBPOffset-LastOffset).
1016 BuildMI(MBB, MBBI, dl, TII.get(PPC::ADDIC), ScratchReg)
1018 .addImm(PBPOffset-LastOffset);
1019 LastOffset = PBPOffset;
1020 BuildMI(MBB, MBBI, dl, TII.get(PPC::STWX))
1021 .addReg(PPC::R30, RegState::Kill) // Save PIC base pointer.
1023 .addReg(ScratchReg); // This will be the index (R0 is ok here).
1026 // R0 += (BPOffset-LastOffset).
1027 BuildMI(MBB, MBBI, dl, TII.get(PPC::ADDIC), ScratchReg)
1029 .addImm(BPOffset-LastOffset);
1030 LastOffset = BPOffset;
1031 BuildMI(MBB, MBBI, dl, TII.get(PPC::STWX))
1032 .addReg(BPReg, RegState::Kill) // Save BP.
1034 .addReg(ScratchReg); // This will be the index (R0 is ok here).
1035 // BP = R0-LastOffset
1036 BuildMI(MBB, MBBI, dl, TII.get(PPC::ADDIC), BPReg)
1037 .addReg(ScratchReg, RegState::Kill)
1038 .addImm(-LastOffset);
1041 // ScratchReg is not R0, so use it as the base register. It is
1042 // already set to the old SP, so we can use the offsets directly.
1044 // Now that the stack frame has been allocated, save all the necessary
1045 // registers using ScratchReg as the base address.
1047 BuildMI(MBB, MBBI, dl, StoreInst)
1050 .addReg(ScratchReg);
1051 if (FI->usesPICBase())
1052 BuildMI(MBB, MBBI, dl, StoreInst)
1055 .addReg(ScratchReg);
1057 BuildMI(MBB, MBBI, dl, StoreInst)
1060 .addReg(ScratchReg);
1061 BuildMI(MBB, MBBI, dl, OrInst, BPReg)
1062 .addReg(ScratchReg, RegState::Kill)
1063 .addReg(ScratchReg);
1067 // The frame size is a known 16-bit constant (fitting in the immediate
1068 // field of STWU). To be here we have to be compiling for PPC32.
1069 // Since the SPReg has been decreased by FrameSize, add it back to each
1072 BuildMI(MBB, MBBI, dl, StoreInst)
1074 .addImm(FrameSize + FPOffset)
1076 if (FI->usesPICBase())
1077 BuildMI(MBB, MBBI, dl, StoreInst)
1079 .addImm(FrameSize + PBPOffset)
1082 BuildMI(MBB, MBBI, dl, StoreInst)
1084 .addImm(FrameSize + BPOffset)
1086 BuildMI(MBB, MBBI, dl, TII.get(PPC::ADDI), BPReg)
1093 // Add Call Frame Information for the instructions we generated above.
1098 // Define CFA in terms of BP. Do this in preference to using FP/SP,
1099 // because if the stack needed aligning then CFA won't be at a fixed
1100 // offset from FP/SP.
1101 unsigned Reg = MRI->getDwarfRegNum(BPReg, true);
1102 CFIIndex = MF.addFrameInst(
1103 MCCFIInstruction::createDefCfaRegister(nullptr, Reg));
1105 // Adjust the definition of CFA to account for the change in SP.
1106 assert(NegFrameSize);
1107 CFIIndex = MF.addFrameInst(
1108 MCCFIInstruction::createDefCfaOffset(nullptr, NegFrameSize));
1110 BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
1111 .addCFIIndex(CFIIndex);
1114 // Describe where FP was saved, at a fixed offset from CFA.
1115 unsigned Reg = MRI->getDwarfRegNum(FPReg, true);
1116 CFIIndex = MF.addFrameInst(
1117 MCCFIInstruction::createOffset(nullptr, Reg, FPOffset));
1118 BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
1119 .addCFIIndex(CFIIndex);
1122 if (FI->usesPICBase()) {
1123 // Describe where FP was saved, at a fixed offset from CFA.
1124 unsigned Reg = MRI->getDwarfRegNum(PPC::R30, true);
1125 CFIIndex = MF.addFrameInst(
1126 MCCFIInstruction::createOffset(nullptr, Reg, PBPOffset));
1127 BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
1128 .addCFIIndex(CFIIndex);
1132 // Describe where BP was saved, at a fixed offset from CFA.
1133 unsigned Reg = MRI->getDwarfRegNum(BPReg, true);
1134 CFIIndex = MF.addFrameInst(
1135 MCCFIInstruction::createOffset(nullptr, Reg, BPOffset));
1136 BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
1137 .addCFIIndex(CFIIndex);
1141 // Describe where LR was saved, at a fixed offset from CFA.
1142 unsigned Reg = MRI->getDwarfRegNum(LRReg, true);
1143 CFIIndex = MF.addFrameInst(
1144 MCCFIInstruction::createOffset(nullptr, Reg, LROffset));
1145 BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
1146 .addCFIIndex(CFIIndex);
1150 // If there is a frame pointer, copy R1 into R31
1152 BuildMI(MBB, MBBI, dl, OrInst, FPReg)
1156 if (!HasBP && needsCFI) {
1157 // Change the definition of CFA from SP+offset to FP+offset, because SP
1158 // will change at every alloca.
1159 unsigned Reg = MRI->getDwarfRegNum(FPReg, true);
1160 unsigned CFIIndex = MF.addFrameInst(
1161 MCCFIInstruction::createDefCfaRegister(nullptr, Reg));
1163 BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
1164 .addCFIIndex(CFIIndex);
1169 // Describe where callee saved registers were saved, at fixed offsets from
1171 const std::vector<CalleeSavedInfo> &CSI = MFI.getCalleeSavedInfo();
1172 for (unsigned I = 0, E = CSI.size(); I != E; ++I) {
1173 unsigned Reg = CSI[I].getReg();
1174 if (Reg == PPC::LR || Reg == PPC::LR8 || Reg == PPC::RM) continue;
1176 // This is a bit of a hack: CR2LT, CR2GT, CR2EQ and CR2UN are just
1177 // subregisters of CR2. We just need to emit a move of CR2.
1178 if (PPC::CRBITRCRegClass.contains(Reg))
1181 // For SVR4, don't emit a move for the CR spill slot if we haven't
1183 if (isSVR4ABI && (PPC::CR2 <= Reg && Reg <= PPC::CR4)
1187 // For 64-bit SVR4 when we have spilled CRs, the spill location
1188 // is SP+8, not a frame-relative slot.
1189 if (isSVR4ABI && isPPC64 && (PPC::CR2 <= Reg && Reg <= PPC::CR4)) {
1190 // In the ELFv1 ABI, only CR2 is noted in CFI and stands in for
1191 // the whole CR word. In the ELFv2 ABI, every CR that was
1192 // actually saved gets its own CFI record.
1193 unsigned CRReg = isELFv2ABI? Reg : (unsigned) PPC::CR2;
1194 unsigned CFIIndex = MF.addFrameInst(MCCFIInstruction::createOffset(
1195 nullptr, MRI->getDwarfRegNum(CRReg, true), 8));
1196 BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
1197 .addCFIIndex(CFIIndex);
1201 int Offset = MFI.getObjectOffset(CSI[I].getFrameIdx());
1202 unsigned CFIIndex = MF.addFrameInst(MCCFIInstruction::createOffset(
1203 nullptr, MRI->getDwarfRegNum(Reg, true), Offset));
1204 BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
1205 .addCFIIndex(CFIIndex);
1210 void PPCFrameLowering::emitEpilogue(MachineFunction &MF,
1211 MachineBasicBlock &MBB) const {
1212 MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator();
1215 if (MBBI != MBB.end())
1216 dl = MBBI->getDebugLoc();
1218 const PPCInstrInfo &TII = *Subtarget.getInstrInfo();
1219 const PPCRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
1221 // Get alignment info so we know how to restore the SP.
1222 const MachineFrameInfo &MFI = MF.getFrameInfo();
1224 // Get the number of bytes allocated from the FrameInfo.
1225 int FrameSize = MFI.getStackSize();
1227 // Get processor type.
1228 bool isPPC64 = Subtarget.isPPC64();
1230 bool isSVR4ABI = Subtarget.isSVR4ABI();
1232 // Check if the link register (LR) has been saved.
1233 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1234 bool MustSaveLR = FI->mustSaveLR();
1235 const SmallVectorImpl<unsigned> &MustSaveCRs = FI->getMustSaveCRs();
1236 bool MustSaveCR = !MustSaveCRs.empty();
1237 // Do we have a frame pointer and/or base pointer for this function?
1238 bool HasFP = hasFP(MF);
1239 bool HasBP = RegInfo->hasBasePointer(MF);
1240 bool HasRedZone = Subtarget.isPPC64() || !Subtarget.isSVR4ABI();
1242 unsigned SPReg = isPPC64 ? PPC::X1 : PPC::R1;
1243 unsigned BPReg = RegInfo->getBaseRegister(MF);
1244 unsigned FPReg = isPPC64 ? PPC::X31 : PPC::R31;
1245 unsigned ScratchReg = 0;
1246 unsigned TempReg = isPPC64 ? PPC::X12 : PPC::R12; // another scratch reg
1247 const MCInstrDesc& MTLRInst = TII.get( isPPC64 ? PPC::MTLR8
1249 const MCInstrDesc& LoadInst = TII.get( isPPC64 ? PPC::LD
1251 const MCInstrDesc& LoadImmShiftedInst = TII.get( isPPC64 ? PPC::LIS8
1253 const MCInstrDesc& OrInst = TII.get(isPPC64 ? PPC::OR8
1255 const MCInstrDesc& OrImmInst = TII.get( isPPC64 ? PPC::ORI8
1257 const MCInstrDesc& AddImmInst = TII.get( isPPC64 ? PPC::ADDI8
1259 const MCInstrDesc& AddInst = TII.get( isPPC64 ? PPC::ADD8
1262 int LROffset = getReturnSaveOffset();
1266 // Using the same bool variable as below to suppress compiler warnings.
1267 bool SingleScratchReg = findScratchRegister(&MBB, true, false, &ScratchReg,
1269 assert(SingleScratchReg &&
1270 "Could not find an available scratch register");
1272 SingleScratchReg = ScratchReg == TempReg;
1276 int FPIndex = FI->getFramePointerSaveIndex();
1277 assert(FPIndex && "No Frame Pointer Save Slot!");
1278 FPOffset = MFI.getObjectOffset(FPIndex);
1280 FPOffset = getFramePointerSaveOffset();
1287 int BPIndex = FI->getBasePointerSaveIndex();
1288 assert(BPIndex && "No Base Pointer Save Slot!");
1289 BPOffset = MFI.getObjectOffset(BPIndex);
1291 BPOffset = getBasePointerSaveOffset();
1296 if (FI->usesPICBase()) {
1297 int PBPIndex = FI->getPICBasePointerSaveIndex();
1298 assert(PBPIndex && "No PIC Base Pointer Save Slot!");
1299 PBPOffset = MFI.getObjectOffset(PBPIndex);
1302 bool IsReturnBlock = (MBBI != MBB.end() && MBBI->isReturn());
1304 if (IsReturnBlock) {
1305 unsigned RetOpcode = MBBI->getOpcode();
1306 bool UsesTCRet = RetOpcode == PPC::TCRETURNri ||
1307 RetOpcode == PPC::TCRETURNdi ||
1308 RetOpcode == PPC::TCRETURNai ||
1309 RetOpcode == PPC::TCRETURNri8 ||
1310 RetOpcode == PPC::TCRETURNdi8 ||
1311 RetOpcode == PPC::TCRETURNai8;
1314 int MaxTCRetDelta = FI->getTailCallSPDelta();
1315 MachineOperand &StackAdjust = MBBI->getOperand(1);
1316 assert(StackAdjust.isImm() && "Expecting immediate value.");
1317 // Adjust stack pointer.
1318 int StackAdj = StackAdjust.getImm();
1319 int Delta = StackAdj - MaxTCRetDelta;
1320 assert((Delta >= 0) && "Delta must be positive");
1321 if (MaxTCRetDelta>0)
1322 FrameSize += (StackAdj +Delta);
1324 FrameSize += StackAdj;
1328 // Frames of 32KB & larger require special handling because they cannot be
1329 // indexed into with a simple LD/LWZ immediate offset operand.
1330 bool isLargeFrame = !isInt<16>(FrameSize);
1332 // On targets without red zone, the SP needs to be restored last, so that
1333 // all live contents of the stack frame are upwards of the SP. This means
1334 // that we cannot restore SP just now, since there may be more registers
1335 // to restore from the stack frame (e.g. R31). If the frame size is not
1336 // a simple immediate value, we will need a spare register to hold the
1337 // restored SP. If the frame size is known and small, we can simply adjust
1338 // the offsets of the registers to be restored, and still use SP to restore
1339 // them. In such case, the final update of SP will be to add the frame
1341 // To simplify the code, set RBReg to the base register used to restore
1342 // values from the stack, and set SPAdd to the value that needs to be added
1343 // to the SP at the end. The default values are as if red zone was present.
1344 unsigned RBReg = SPReg;
1348 // In the prologue, the loaded (or persistent) stack pointer value is
1349 // offset by the STDU/STDUX/STWU/STWUX instruction. For targets with red
1350 // zone add this offset back now.
1352 // If this function contained a fastcc call and GuaranteedTailCallOpt is
1353 // enabled (=> hasFastCall()==true) the fastcc call might contain a tail
1354 // call which invalidates the stack pointer value in SP(0). So we use the
1355 // value of R31 in this case.
1356 if (FI->hasFastCall()) {
1357 assert(HasFP && "Expecting a valid frame pointer.");
1360 if (!isLargeFrame) {
1361 BuildMI(MBB, MBBI, dl, AddImmInst, RBReg)
1362 .addReg(FPReg).addImm(FrameSize);
1364 BuildMI(MBB, MBBI, dl, LoadImmShiftedInst, ScratchReg)
1365 .addImm(FrameSize >> 16);
1366 BuildMI(MBB, MBBI, dl, OrImmInst, ScratchReg)
1367 .addReg(ScratchReg, RegState::Kill)
1368 .addImm(FrameSize & 0xFFFF);
1369 BuildMI(MBB, MBBI, dl, AddInst)
1372 .addReg(ScratchReg);
1374 } else if (!isLargeFrame && !HasBP && !MFI.hasVarSizedObjects()) {
1376 BuildMI(MBB, MBBI, dl, AddImmInst, SPReg)
1380 // Make sure that adding FrameSize will not overflow the max offset
1382 assert(FPOffset <= 0 && BPOffset <= 0 && PBPOffset <= 0 &&
1383 "Local offsets should be negative");
1385 FPOffset += FrameSize;
1386 BPOffset += FrameSize;
1387 PBPOffset += FrameSize;
1390 // We don't want to use ScratchReg as a base register, because it
1391 // could happen to be R0. Use FP instead, but make sure to preserve it.
1393 // If FP is not saved, copy it to ScratchReg.
1395 BuildMI(MBB, MBBI, dl, OrInst, ScratchReg)
1400 BuildMI(MBB, MBBI, dl, LoadInst, RBReg)
1405 assert(RBReg != ScratchReg && "Should have avoided ScratchReg");
1406 // If there is no red zone, ScratchReg may be needed for holding a useful
1407 // value (although not the base register). Make sure it is not overwritten
1410 assert((isPPC64 || !MustSaveCR) &&
1411 "Epilogue CR restoring supported only in 64-bit mode");
1413 // If we need to restore both the LR and the CR and we only have one
1414 // available scratch register, we must do them one at a time.
1415 if (MustSaveCR && SingleScratchReg && MustSaveLR) {
1416 // Here TempReg == ScratchReg, and in the absence of red zone ScratchReg
1418 assert(HasRedZone && "Expecting red zone");
1419 BuildMI(MBB, MBBI, dl, TII.get(PPC::LWZ8), TempReg)
1422 for (unsigned i = 0, e = MustSaveCRs.size(); i != e; ++i)
1423 BuildMI(MBB, MBBI, dl, TII.get(PPC::MTOCRF8), MustSaveCRs[i])
1424 .addReg(TempReg, getKillRegState(i == e-1));
1427 // Delay restoring of the LR if ScratchReg is needed. This is ok, since
1428 // LR is stored in the caller's stack frame. ScratchReg will be needed
1429 // if RBReg is anything other than SP. We shouldn't use ScratchReg as
1430 // a base register anyway, because it may happen to be R0.
1431 bool LoadedLR = false;
1432 if (MustSaveLR && RBReg == SPReg && isInt<16>(LROffset+SPAdd)) {
1433 BuildMI(MBB, MBBI, dl, LoadInst, ScratchReg)
1434 .addImm(LROffset+SPAdd)
1439 if (MustSaveCR && !(SingleScratchReg && MustSaveLR)) {
1440 // This will only occur for PPC64.
1441 assert(isPPC64 && "Expecting 64-bit mode");
1442 assert(RBReg == SPReg && "Should be using SP as a base register");
1443 BuildMI(MBB, MBBI, dl, TII.get(PPC::LWZ8), TempReg)
1449 // If there is red zone, restore FP directly, since SP has already been
1450 // restored. Otherwise, restore the value of FP into ScratchReg.
1451 if (HasRedZone || RBReg == SPReg)
1452 BuildMI(MBB, MBBI, dl, LoadInst, FPReg)
1456 BuildMI(MBB, MBBI, dl, LoadInst, ScratchReg)
1461 if (FI->usesPICBase())
1462 BuildMI(MBB, MBBI, dl, LoadInst, PPC::R30)
1467 BuildMI(MBB, MBBI, dl, LoadInst, BPReg)
1471 // There is nothing more to be loaded from the stack, so now we can
1472 // restore SP: SP = RBReg + SPAdd.
1473 if (RBReg != SPReg || SPAdd != 0) {
1474 assert(!HasRedZone && "This should not happen with red zone");
1475 // If SPAdd is 0, generate a copy.
1477 BuildMI(MBB, MBBI, dl, OrInst, SPReg)
1481 BuildMI(MBB, MBBI, dl, AddImmInst, SPReg)
1485 assert(RBReg != ScratchReg && "Should be using FP or SP as base register");
1487 BuildMI(MBB, MBBI, dl, OrInst, FPReg)
1489 .addReg(ScratchReg);
1491 // Now load the LR from the caller's stack frame.
1492 if (MustSaveLR && !LoadedLR)
1493 BuildMI(MBB, MBBI, dl, LoadInst, ScratchReg)
1499 !(SingleScratchReg && MustSaveLR)) // will only occur for PPC64
1500 for (unsigned i = 0, e = MustSaveCRs.size(); i != e; ++i)
1501 BuildMI(MBB, MBBI, dl, TII.get(PPC::MTOCRF8), MustSaveCRs[i])
1502 .addReg(TempReg, getKillRegState(i == e-1));
1505 BuildMI(MBB, MBBI, dl, MTLRInst).addReg(ScratchReg);
1507 // Callee pop calling convention. Pop parameter/linkage area. Used for tail
1508 // call optimization
1509 if (IsReturnBlock) {
1510 unsigned RetOpcode = MBBI->getOpcode();
1511 if (MF.getTarget().Options.GuaranteedTailCallOpt &&
1512 (RetOpcode == PPC::BLR || RetOpcode == PPC::BLR8) &&
1513 MF.getFunction()->getCallingConv() == CallingConv::Fast) {
1514 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1515 unsigned CallerAllocatedAmt = FI->getMinReservedArea();
1517 if (CallerAllocatedAmt && isInt<16>(CallerAllocatedAmt)) {
1518 BuildMI(MBB, MBBI, dl, AddImmInst, SPReg)
1519 .addReg(SPReg).addImm(CallerAllocatedAmt);
1521 BuildMI(MBB, MBBI, dl, LoadImmShiftedInst, ScratchReg)
1522 .addImm(CallerAllocatedAmt >> 16);
1523 BuildMI(MBB, MBBI, dl, OrImmInst, ScratchReg)
1524 .addReg(ScratchReg, RegState::Kill)
1525 .addImm(CallerAllocatedAmt & 0xFFFF);
1526 BuildMI(MBB, MBBI, dl, AddInst)
1529 .addReg(ScratchReg);
1532 createTailCallBranchInstr(MBB);
1537 void PPCFrameLowering::createTailCallBranchInstr(MachineBasicBlock &MBB) const {
1538 MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator();
1541 if (MBBI != MBB.end())
1542 dl = MBBI->getDebugLoc();
1544 const PPCInstrInfo &TII = *Subtarget.getInstrInfo();
1546 // Create branch instruction for pseudo tail call return instruction
1547 unsigned RetOpcode = MBBI->getOpcode();
1548 if (RetOpcode == PPC::TCRETURNdi) {
1549 MBBI = MBB.getLastNonDebugInstr();
1550 MachineOperand &JumpTarget = MBBI->getOperand(0);
1551 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILB)).
1552 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset());
1553 } else if (RetOpcode == PPC::TCRETURNri) {
1554 MBBI = MBB.getLastNonDebugInstr();
1555 assert(MBBI->getOperand(0).isReg() && "Expecting register operand.");
1556 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILBCTR));
1557 } else if (RetOpcode == PPC::TCRETURNai) {
1558 MBBI = MBB.getLastNonDebugInstr();
1559 MachineOperand &JumpTarget = MBBI->getOperand(0);
1560 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILBA)).addImm(JumpTarget.getImm());
1561 } else if (RetOpcode == PPC::TCRETURNdi8) {
1562 MBBI = MBB.getLastNonDebugInstr();
1563 MachineOperand &JumpTarget = MBBI->getOperand(0);
1564 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILB8)).
1565 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset());
1566 } else if (RetOpcode == PPC::TCRETURNri8) {
1567 MBBI = MBB.getLastNonDebugInstr();
1568 assert(MBBI->getOperand(0).isReg() && "Expecting register operand.");
1569 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILBCTR8));
1570 } else if (RetOpcode == PPC::TCRETURNai8) {
1571 MBBI = MBB.getLastNonDebugInstr();
1572 MachineOperand &JumpTarget = MBBI->getOperand(0);
1573 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILBA8)).addImm(JumpTarget.getImm());
1577 void PPCFrameLowering::determineCalleeSaves(MachineFunction &MF,
1578 BitVector &SavedRegs,
1579 RegScavenger *RS) const {
1580 TargetFrameLowering::determineCalleeSaves(MF, SavedRegs, RS);
1582 const PPCRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
1584 // Save and clear the LR state.
1585 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1586 unsigned LR = RegInfo->getRARegister();
1587 FI->setMustSaveLR(MustSaveLR(MF, LR));
1588 SavedRegs.reset(LR);
1590 // Save R31 if necessary
1591 int FPSI = FI->getFramePointerSaveIndex();
1592 bool isPPC64 = Subtarget.isPPC64();
1593 bool isDarwinABI = Subtarget.isDarwinABI();
1594 MachineFrameInfo &MFI = MF.getFrameInfo();
1596 // If the frame pointer save index hasn't been defined yet.
1597 if (!FPSI && needsFP(MF)) {
1598 // Find out what the fix offset of the frame pointer save area.
1599 int FPOffset = getFramePointerSaveOffset();
1600 // Allocate the frame index for frame pointer save area.
1601 FPSI = MFI.CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
1603 FI->setFramePointerSaveIndex(FPSI);
1606 int BPSI = FI->getBasePointerSaveIndex();
1607 if (!BPSI && RegInfo->hasBasePointer(MF)) {
1608 int BPOffset = getBasePointerSaveOffset();
1609 // Allocate the frame index for the base pointer save area.
1610 BPSI = MFI.CreateFixedObject(isPPC64? 8 : 4, BPOffset, true);
1612 FI->setBasePointerSaveIndex(BPSI);
1615 // Reserve stack space for the PIC Base register (R30).
1616 // Only used in SVR4 32-bit.
1617 if (FI->usesPICBase()) {
1618 int PBPSI = MFI.CreateFixedObject(4, -8, true);
1619 FI->setPICBasePointerSaveIndex(PBPSI);
1622 // Make sure we don't explicitly spill r31, because, for example, we have
1623 // some inline asm which explicity clobbers it, when we otherwise have a
1624 // frame pointer and are using r31's spill slot for the prologue/epilogue
1625 // code. Same goes for the base pointer and the PIC base register.
1627 SavedRegs.reset(isPPC64 ? PPC::X31 : PPC::R31);
1628 if (RegInfo->hasBasePointer(MF))
1629 SavedRegs.reset(RegInfo->getBaseRegister(MF));
1630 if (FI->usesPICBase())
1631 SavedRegs.reset(PPC::R30);
1633 // Reserve stack space to move the linkage area to in case of a tail call.
1635 if (MF.getTarget().Options.GuaranteedTailCallOpt &&
1636 (TCSPDelta = FI->getTailCallSPDelta()) < 0) {
1637 MFI.CreateFixedObject(-1 * TCSPDelta, TCSPDelta, true);
1640 // For 32-bit SVR4, allocate the nonvolatile CR spill slot iff the
1641 // function uses CR 2, 3, or 4.
1642 if (!isPPC64 && !isDarwinABI &&
1643 (SavedRegs.test(PPC::CR2) ||
1644 SavedRegs.test(PPC::CR3) ||
1645 SavedRegs.test(PPC::CR4))) {
1646 int FrameIdx = MFI.CreateFixedObject((uint64_t)4, (int64_t)-4, true);
1647 FI->setCRSpillFrameIndex(FrameIdx);
1651 void PPCFrameLowering::processFunctionBeforeFrameFinalized(MachineFunction &MF,
1652 RegScavenger *RS) const {
1653 // Early exit if not using the SVR4 ABI.
1654 if (!Subtarget.isSVR4ABI()) {
1655 addScavengingSpillSlot(MF, RS);
1659 // Get callee saved register information.
1660 MachineFrameInfo &MFI = MF.getFrameInfo();
1661 const std::vector<CalleeSavedInfo> &CSI = MFI.getCalleeSavedInfo();
1663 // If the function is shrink-wrapped, and if the function has a tail call, the
1664 // tail call might not be in the new RestoreBlock, so real branch instruction
1665 // won't be generated by emitEpilogue(), because shrink-wrap has chosen new
1666 // RestoreBlock. So we handle this case here.
1667 if (MFI.getSavePoint() && MFI.hasTailCall()) {
1668 MachineBasicBlock *RestoreBlock = MFI.getRestorePoint();
1669 for (MachineBasicBlock &MBB : MF) {
1670 if (MBB.isReturnBlock() && (&MBB) != RestoreBlock)
1671 createTailCallBranchInstr(MBB);
1675 // Early exit if no callee saved registers are modified!
1676 if (CSI.empty() && !needsFP(MF)) {
1677 addScavengingSpillSlot(MF, RS);
1681 unsigned MinGPR = PPC::R31;
1682 unsigned MinG8R = PPC::X31;
1683 unsigned MinFPR = PPC::F31;
1684 unsigned MinVR = PPC::V31;
1686 bool HasGPSaveArea = false;
1687 bool HasG8SaveArea = false;
1688 bool HasFPSaveArea = false;
1689 bool HasVRSAVESaveArea = false;
1690 bool HasVRSaveArea = false;
1692 SmallVector<CalleeSavedInfo, 18> GPRegs;
1693 SmallVector<CalleeSavedInfo, 18> G8Regs;
1694 SmallVector<CalleeSavedInfo, 18> FPRegs;
1695 SmallVector<CalleeSavedInfo, 18> VRegs;
1697 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1698 unsigned Reg = CSI[i].getReg();
1699 if (PPC::GPRCRegClass.contains(Reg)) {
1700 HasGPSaveArea = true;
1702 GPRegs.push_back(CSI[i]);
1707 } else if (PPC::G8RCRegClass.contains(Reg)) {
1708 HasG8SaveArea = true;
1710 G8Regs.push_back(CSI[i]);
1715 } else if (PPC::F8RCRegClass.contains(Reg)) {
1716 HasFPSaveArea = true;
1718 FPRegs.push_back(CSI[i]);
1723 } else if (PPC::CRBITRCRegClass.contains(Reg) ||
1724 PPC::CRRCRegClass.contains(Reg)) {
1725 ; // do nothing, as we already know whether CRs are spilled
1726 } else if (PPC::VRSAVERCRegClass.contains(Reg)) {
1727 HasVRSAVESaveArea = true;
1728 } else if (PPC::VRRCRegClass.contains(Reg)) {
1729 HasVRSaveArea = true;
1731 VRegs.push_back(CSI[i]);
1737 llvm_unreachable("Unknown RegisterClass!");
1741 PPCFunctionInfo *PFI = MF.getInfo<PPCFunctionInfo>();
1742 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
1744 int64_t LowerBound = 0;
1746 // Take into account stack space reserved for tail calls.
1748 if (MF.getTarget().Options.GuaranteedTailCallOpt &&
1749 (TCSPDelta = PFI->getTailCallSPDelta()) < 0) {
1750 LowerBound = TCSPDelta;
1753 // The Floating-point register save area is right below the back chain word
1754 // of the previous stack frame.
1755 if (HasFPSaveArea) {
1756 for (unsigned i = 0, e = FPRegs.size(); i != e; ++i) {
1757 int FI = FPRegs[i].getFrameIdx();
1759 MFI.setObjectOffset(FI, LowerBound + MFI.getObjectOffset(FI));
1762 LowerBound -= (31 - TRI->getEncodingValue(MinFPR) + 1) * 8;
1765 // Check whether the frame pointer register is allocated. If so, make sure it
1766 // is spilled to the correct offset.
1768 HasGPSaveArea = true;
1770 int FI = PFI->getFramePointerSaveIndex();
1771 assert(FI && "No Frame Pointer Save Slot!");
1773 MFI.setObjectOffset(FI, LowerBound + MFI.getObjectOffset(FI));
1776 if (PFI->usesPICBase()) {
1777 HasGPSaveArea = true;
1779 int FI = PFI->getPICBasePointerSaveIndex();
1780 assert(FI && "No PIC Base Pointer Save Slot!");
1782 MFI.setObjectOffset(FI, LowerBound + MFI.getObjectOffset(FI));
1785 const PPCRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
1786 if (RegInfo->hasBasePointer(MF)) {
1787 HasGPSaveArea = true;
1789 int FI = PFI->getBasePointerSaveIndex();
1790 assert(FI && "No Base Pointer Save Slot!");
1792 MFI.setObjectOffset(FI, LowerBound + MFI.getObjectOffset(FI));
1795 // General register save area starts right below the Floating-point
1796 // register save area.
1797 if (HasGPSaveArea || HasG8SaveArea) {
1798 // Move general register save area spill slots down, taking into account
1799 // the size of the Floating-point register save area.
1800 for (unsigned i = 0, e = GPRegs.size(); i != e; ++i) {
1801 int FI = GPRegs[i].getFrameIdx();
1803 MFI.setObjectOffset(FI, LowerBound + MFI.getObjectOffset(FI));
1806 // Move general register save area spill slots down, taking into account
1807 // the size of the Floating-point register save area.
1808 for (unsigned i = 0, e = G8Regs.size(); i != e; ++i) {
1809 int FI = G8Regs[i].getFrameIdx();
1811 MFI.setObjectOffset(FI, LowerBound + MFI.getObjectOffset(FI));
1815 std::min<unsigned>(TRI->getEncodingValue(MinGPR),
1816 TRI->getEncodingValue(MinG8R));
1818 if (Subtarget.isPPC64()) {
1819 LowerBound -= (31 - MinReg + 1) * 8;
1821 LowerBound -= (31 - MinReg + 1) * 4;
1825 // For 32-bit only, the CR save area is below the general register
1826 // save area. For 64-bit SVR4, the CR save area is addressed relative
1827 // to the stack pointer and hence does not need an adjustment here.
1828 // Only CR2 (the first nonvolatile spilled) has an associated frame
1829 // index so that we have a single uniform save area.
1830 if (spillsCR(MF) && !(Subtarget.isPPC64() && Subtarget.isSVR4ABI())) {
1831 // Adjust the frame index of the CR spill slot.
1832 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1833 unsigned Reg = CSI[i].getReg();
1835 if ((Subtarget.isSVR4ABI() && Reg == PPC::CR2)
1836 // Leave Darwin logic as-is.
1837 || (!Subtarget.isSVR4ABI() &&
1838 (PPC::CRBITRCRegClass.contains(Reg) ||
1839 PPC::CRRCRegClass.contains(Reg)))) {
1840 int FI = CSI[i].getFrameIdx();
1842 MFI.setObjectOffset(FI, LowerBound + MFI.getObjectOffset(FI));
1846 LowerBound -= 4; // The CR save area is always 4 bytes long.
1849 if (HasVRSAVESaveArea) {
1850 // FIXME SVR4: Is it actually possible to have multiple elements in CSI
1851 // which have the VRSAVE register class?
1852 // Adjust the frame index of the VRSAVE spill slot.
1853 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1854 unsigned Reg = CSI[i].getReg();
1856 if (PPC::VRSAVERCRegClass.contains(Reg)) {
1857 int FI = CSI[i].getFrameIdx();
1859 MFI.setObjectOffset(FI, LowerBound + MFI.getObjectOffset(FI));
1863 LowerBound -= 4; // The VRSAVE save area is always 4 bytes long.
1866 if (HasVRSaveArea) {
1867 // Insert alignment padding, we need 16-byte alignment.
1868 LowerBound = (LowerBound - 15) & ~(15);
1870 for (unsigned i = 0, e = VRegs.size(); i != e; ++i) {
1871 int FI = VRegs[i].getFrameIdx();
1873 MFI.setObjectOffset(FI, LowerBound + MFI.getObjectOffset(FI));
1877 addScavengingSpillSlot(MF, RS);
1881 PPCFrameLowering::addScavengingSpillSlot(MachineFunction &MF,
1882 RegScavenger *RS) const {
1883 // Reserve a slot closest to SP or frame pointer if we have a dynalloc or
1884 // a large stack, which will require scavenging a register to materialize a
1887 // We need to have a scavenger spill slot for spills if the frame size is
1888 // large. In case there is no free register for large-offset addressing,
1889 // this slot is used for the necessary emergency spill. Also, we need the
1890 // slot for dynamic stack allocations.
1892 // The scavenger might be invoked if the frame offset does not fit into
1893 // the 16-bit immediate. We don't know the complete frame size here
1894 // because we've not yet computed callee-saved register spills or the
1895 // needed alignment padding.
1896 unsigned StackSize = determineFrameLayout(MF, false, true);
1897 MachineFrameInfo &MFI = MF.getFrameInfo();
1898 if (MFI.hasVarSizedObjects() || spillsCR(MF) || spillsVRSAVE(MF) ||
1899 hasNonRISpills(MF) || (hasSpills(MF) && !isInt<16>(StackSize))) {
1900 const TargetRegisterClass &GPRC = PPC::GPRCRegClass;
1901 const TargetRegisterClass &G8RC = PPC::G8RCRegClass;
1902 const TargetRegisterClass &RC = Subtarget.isPPC64() ? G8RC : GPRC;
1903 const TargetRegisterInfo &TRI = *Subtarget.getRegisterInfo();
1904 unsigned Size = TRI.getSpillSize(RC);
1905 unsigned Align = TRI.getSpillAlignment(RC);
1906 RS->addScavengingFrameIndex(MFI.CreateStackObject(Size, Align, false));
1908 // Might we have over-aligned allocas?
1909 bool HasAlVars = MFI.hasVarSizedObjects() &&
1910 MFI.getMaxAlignment() > getStackAlignment();
1912 // These kinds of spills might need two registers.
1913 if (spillsCR(MF) || spillsVRSAVE(MF) || HasAlVars)
1914 RS->addScavengingFrameIndex(MFI.CreateStackObject(Size, Align, false));
1920 PPCFrameLowering::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
1921 MachineBasicBlock::iterator MI,
1922 const std::vector<CalleeSavedInfo> &CSI,
1923 const TargetRegisterInfo *TRI) const {
1925 // Currently, this function only handles SVR4 32- and 64-bit ABIs.
1926 // Return false otherwise to maintain pre-existing behavior.
1927 if (!Subtarget.isSVR4ABI())
1930 MachineFunction *MF = MBB.getParent();
1931 const PPCInstrInfo &TII = *Subtarget.getInstrInfo();
1933 bool CRSpilled = false;
1934 MachineInstrBuilder CRMIB;
1936 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1937 unsigned Reg = CSI[i].getReg();
1938 // Only Darwin actually uses the VRSAVE register, but it can still appear
1939 // here if, for example, @llvm.eh.unwind.init() is used. If we're not on
1940 // Darwin, ignore it.
1941 if (Reg == PPC::VRSAVE && !Subtarget.isDarwinABI())
1944 // CR2 through CR4 are the nonvolatile CR fields.
1945 bool IsCRField = PPC::CR2 <= Reg && Reg <= PPC::CR4;
1947 // Add the callee-saved register as live-in; it's killed at the spill.
1950 if (CRSpilled && IsCRField) {
1951 CRMIB.addReg(Reg, RegState::ImplicitKill);
1955 // Insert the spill to the stack frame.
1957 PPCFunctionInfo *FuncInfo = MF->getInfo<PPCFunctionInfo>();
1958 if (Subtarget.isPPC64()) {
1959 // The actual spill will happen at the start of the prologue.
1960 FuncInfo->addMustSaveCR(Reg);
1963 FuncInfo->setSpillsCR();
1965 // 32-bit: FP-relative. Note that we made sure CR2-CR4 all have
1966 // the same frame index in PPCRegisterInfo::hasReservedSpillSlot.
1967 CRMIB = BuildMI(*MF, DL, TII.get(PPC::MFCR), PPC::R12)
1968 .addReg(Reg, RegState::ImplicitKill);
1970 MBB.insert(MI, CRMIB);
1971 MBB.insert(MI, addFrameReference(BuildMI(*MF, DL, TII.get(PPC::STW))
1973 getKillRegState(true)),
1974 CSI[i].getFrameIdx()));
1977 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
1978 TII.storeRegToStackSlot(MBB, MI, Reg, true,
1979 CSI[i].getFrameIdx(), RC, TRI);
1986 restoreCRs(bool isPPC64, bool is31,
1987 bool CR2Spilled, bool CR3Spilled, bool CR4Spilled,
1988 MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
1989 const std::vector<CalleeSavedInfo> &CSI, unsigned CSIIndex) {
1991 MachineFunction *MF = MBB.getParent();
1992 const PPCInstrInfo &TII = *MF->getSubtarget<PPCSubtarget>().getInstrInfo();
1994 unsigned RestoreOp, MoveReg;
1997 // This is handled during epilogue generation.
2000 // 32-bit: FP-relative
2001 MBB.insert(MI, addFrameReference(BuildMI(*MF, DL, TII.get(PPC::LWZ),
2003 CSI[CSIIndex].getFrameIdx()));
2004 RestoreOp = PPC::MTOCRF;
2009 MBB.insert(MI, BuildMI(*MF, DL, TII.get(RestoreOp), PPC::CR2)
2010 .addReg(MoveReg, getKillRegState(!CR3Spilled && !CR4Spilled)));
2013 MBB.insert(MI, BuildMI(*MF, DL, TII.get(RestoreOp), PPC::CR3)
2014 .addReg(MoveReg, getKillRegState(!CR4Spilled)));
2017 MBB.insert(MI, BuildMI(*MF, DL, TII.get(RestoreOp), PPC::CR4)
2018 .addReg(MoveReg, getKillRegState(true)));
2021 MachineBasicBlock::iterator PPCFrameLowering::
2022 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
2023 MachineBasicBlock::iterator I) const {
2024 const TargetInstrInfo &TII = *Subtarget.getInstrInfo();
2025 if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2026 I->getOpcode() == PPC::ADJCALLSTACKUP) {
2027 // Add (actually subtract) back the amount the callee popped on return.
2028 if (int CalleeAmt = I->getOperand(1).getImm()) {
2029 bool is64Bit = Subtarget.isPPC64();
2031 unsigned StackReg = is64Bit ? PPC::X1 : PPC::R1;
2032 unsigned TmpReg = is64Bit ? PPC::X0 : PPC::R0;
2033 unsigned ADDIInstr = is64Bit ? PPC::ADDI8 : PPC::ADDI;
2034 unsigned ADDInstr = is64Bit ? PPC::ADD8 : PPC::ADD4;
2035 unsigned LISInstr = is64Bit ? PPC::LIS8 : PPC::LIS;
2036 unsigned ORIInstr = is64Bit ? PPC::ORI8 : PPC::ORI;
2037 const DebugLoc &dl = I->getDebugLoc();
2039 if (isInt<16>(CalleeAmt)) {
2040 BuildMI(MBB, I, dl, TII.get(ADDIInstr), StackReg)
2041 .addReg(StackReg, RegState::Kill)
2044 MachineBasicBlock::iterator MBBI = I;
2045 BuildMI(MBB, MBBI, dl, TII.get(LISInstr), TmpReg)
2046 .addImm(CalleeAmt >> 16);
2047 BuildMI(MBB, MBBI, dl, TII.get(ORIInstr), TmpReg)
2048 .addReg(TmpReg, RegState::Kill)
2049 .addImm(CalleeAmt & 0xFFFF);
2050 BuildMI(MBB, MBBI, dl, TII.get(ADDInstr), StackReg)
2051 .addReg(StackReg, RegState::Kill)
2056 // Simply discard ADJCALLSTACKDOWN, ADJCALLSTACKUP instructions.
2057 return MBB.erase(I);
2061 PPCFrameLowering::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
2062 MachineBasicBlock::iterator MI,
2063 const std::vector<CalleeSavedInfo> &CSI,
2064 const TargetRegisterInfo *TRI) const {
2066 // Currently, this function only handles SVR4 32- and 64-bit ABIs.
2067 // Return false otherwise to maintain pre-existing behavior.
2068 if (!Subtarget.isSVR4ABI())
2071 MachineFunction *MF = MBB.getParent();
2072 const PPCInstrInfo &TII = *Subtarget.getInstrInfo();
2073 bool CR2Spilled = false;
2074 bool CR3Spilled = false;
2075 bool CR4Spilled = false;
2076 unsigned CSIIndex = 0;
2078 // Initialize insertion-point logic; we will be restoring in reverse
2080 MachineBasicBlock::iterator I = MI, BeforeI = I;
2081 bool AtStart = I == MBB.begin();
2086 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
2087 unsigned Reg = CSI[i].getReg();
2089 // Only Darwin actually uses the VRSAVE register, but it can still appear
2090 // here if, for example, @llvm.eh.unwind.init() is used. If we're not on
2091 // Darwin, ignore it.
2092 if (Reg == PPC::VRSAVE && !Subtarget.isDarwinABI())
2095 if (Reg == PPC::CR2) {
2097 // The spill slot is associated only with CR2, which is the
2098 // first nonvolatile spilled. Save it here.
2101 } else if (Reg == PPC::CR3) {
2104 } else if (Reg == PPC::CR4) {
2108 // When we first encounter a non-CR register after seeing at
2109 // least one CR register, restore all spilled CRs together.
2110 if ((CR2Spilled || CR3Spilled || CR4Spilled)
2111 && !(PPC::CR2 <= Reg && Reg <= PPC::CR4)) {
2112 bool is31 = needsFP(*MF);
2113 restoreCRs(Subtarget.isPPC64(), is31,
2114 CR2Spilled, CR3Spilled, CR4Spilled,
2115 MBB, I, CSI, CSIIndex);
2116 CR2Spilled = CR3Spilled = CR4Spilled = false;
2119 // Default behavior for non-CR saves.
2120 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
2121 TII.loadRegFromStackSlot(MBB, I, Reg, CSI[i].getFrameIdx(),
2123 assert(I != MBB.begin() &&
2124 "loadRegFromStackSlot didn't insert any code!");
2127 // Insert in reverse order.
2136 // If we haven't yet spilled the CRs, do so now.
2137 if (CR2Spilled || CR3Spilled || CR4Spilled) {
2138 bool is31 = needsFP(*MF);
2139 restoreCRs(Subtarget.isPPC64(), is31, CR2Spilled, CR3Spilled, CR4Spilled,
2140 MBB, I, CSI, CSIIndex);
2146 bool PPCFrameLowering::enableShrinkWrapping(const MachineFunction &MF) const {
2147 return (MF.getSubtarget<PPCSubtarget>().isSVR4ABI() &&
2148 MF.getSubtarget<PPCSubtarget>().isPPC64());