1 //===- ARMBaseRegisterInfo.cpp - ARM Register Information -------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the base ARM implementation of TargetRegisterInfo class.
12 //===----------------------------------------------------------------------===//
15 #include "ARMAddressingModes.h"
16 #include "ARMBaseInstrInfo.h"
17 #include "ARMBaseRegisterInfo.h"
18 #include "ARMInstrInfo.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMSubtarget.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/LLVMContext.h"
25 #include "llvm/CodeGen/MachineConstantPool.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineLocation.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/RegisterScavenging.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetFrameInfo.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/ADT/BitVector.h"
39 #include "llvm/ADT/SmallVector.h"
40 #include "llvm/Support/CommandLine.h"
44 ScavengeFrameIndexVals("arm-virtual-frame-index-vals", cl::Hidden,
46 cl::desc("Resolve frame index values via scavenging in PEI"));
49 ReuseFrameIndexVals("arm-reuse-frame-index-vals", cl::Hidden, cl::init(false),
50 cl::desc("Reuse repeated frame index values"));
53 ARMDynamicStackAlign("arm-dynamic-stack-alignment", cl::Hidden, cl::init(false),
54 cl::desc("Dynamically re-align the stack as needed"));
56 unsigned ARMBaseRegisterInfo::getRegisterNumbering(unsigned RegEnum,
64 llvm_unreachable("Unknown ARM register!");
65 case R0: case D0: case Q0: return 0;
66 case R1: case D1: case Q1: return 1;
67 case R2: case D2: case Q2: return 2;
68 case R3: case D3: case Q3: return 3;
69 case R4: case D4: case Q4: return 4;
70 case R5: case D5: case Q5: return 5;
71 case R6: case D6: case Q6: return 6;
72 case R7: case D7: case Q7: return 7;
73 case R8: case D8: case Q8: return 8;
74 case R9: case D9: case Q9: return 9;
75 case R10: case D10: case Q10: return 10;
76 case R11: case D11: case Q11: return 11;
77 case R12: case D12: case Q12: return 12;
78 case SP: case D13: case Q13: return 13;
79 case LR: case D14: case Q14: return 14;
80 case PC: case D15: case Q15: return 15;
99 case S0: case S1: case S2: case S3:
100 case S4: case S5: case S6: case S7:
101 case S8: case S9: case S10: case S11:
102 case S12: case S13: case S14: case S15:
103 case S16: case S17: case S18: case S19:
104 case S20: case S21: case S22: case S23:
105 case S24: case S25: case S26: case S27:
106 case S28: case S29: case S30: case S31: {
110 default: return 0; // Avoid compile time warning.
148 ARMBaseRegisterInfo::ARMBaseRegisterInfo(const ARMBaseInstrInfo &tii,
149 const ARMSubtarget &sti)
150 : ARMGenRegisterInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
152 FramePtr((STI.isTargetDarwin() || STI.isThumb()) ? ARM::R7 : ARM::R11) {
156 ARMBaseRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
157 static const unsigned CalleeSavedRegs[] = {
158 ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8,
159 ARM::R7, ARM::R6, ARM::R5, ARM::R4,
161 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
162 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
166 static const unsigned DarwinCalleeSavedRegs[] = {
167 // Darwin ABI deviates from ARM standard ABI. R9 is not a callee-saved
169 ARM::LR, ARM::R7, ARM::R6, ARM::R5, ARM::R4,
170 ARM::R11, ARM::R10, ARM::R8,
172 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
173 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
176 return STI.isTargetDarwin() ? DarwinCalleeSavedRegs : CalleeSavedRegs;
179 const TargetRegisterClass* const *
180 ARMBaseRegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const {
181 static const TargetRegisterClass * const CalleeSavedRegClasses[] = {
182 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
183 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
184 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
186 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
187 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
191 static const TargetRegisterClass * const ThumbCalleeSavedRegClasses[] = {
192 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
193 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::tGPRRegClass,
194 &ARM::tGPRRegClass,&ARM::tGPRRegClass,&ARM::tGPRRegClass,
196 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
197 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
201 static const TargetRegisterClass * const DarwinCalleeSavedRegClasses[] = {
202 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
203 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
204 &ARM::GPRRegClass, &ARM::GPRRegClass,
206 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
207 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
211 static const TargetRegisterClass * const DarwinThumbCalleeSavedRegClasses[] ={
212 &ARM::GPRRegClass, &ARM::tGPRRegClass, &ARM::tGPRRegClass,
213 &ARM::tGPRRegClass, &ARM::tGPRRegClass, &ARM::GPRRegClass,
214 &ARM::GPRRegClass, &ARM::GPRRegClass,
216 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
217 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
221 if (STI.isThumb1Only()) {
222 return STI.isTargetDarwin()
223 ? DarwinThumbCalleeSavedRegClasses : ThumbCalleeSavedRegClasses;
225 return STI.isTargetDarwin()
226 ? DarwinCalleeSavedRegClasses : CalleeSavedRegClasses;
229 BitVector ARMBaseRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
230 // FIXME: avoid re-calculating this everytime.
231 BitVector Reserved(getNumRegs());
232 Reserved.set(ARM::SP);
233 Reserved.set(ARM::PC);
234 if (STI.isTargetDarwin() || hasFP(MF))
235 Reserved.set(FramePtr);
236 // Some targets reserve R9.
237 if (STI.isR9Reserved())
238 Reserved.set(ARM::R9);
242 bool ARMBaseRegisterInfo::isReservedReg(const MachineFunction &MF,
243 unsigned Reg) const {
251 if (FramePtr == Reg && (STI.isTargetDarwin() || hasFP(MF)))
255 return STI.isR9Reserved();
261 const TargetRegisterClass *
262 ARMBaseRegisterInfo::getMatchingSuperRegClass(const TargetRegisterClass *A,
263 const TargetRegisterClass *B,
264 unsigned SubIdx) const {
272 if (A->getSize() == 8) {
273 if (A == &ARM::DPR_8RegClass)
275 return &ARM::DPR_VFP2RegClass;
278 assert(A->getSize() == 16 && "Expecting a Q register class!");
279 return &ARM::QPR_VFP2RegClass;
288 const TargetRegisterClass *
289 ARMBaseRegisterInfo::getPointerRegClass(unsigned Kind) const {
290 return ARM::GPRRegisterClass;
293 /// getAllocationOrder - Returns the register allocation order for a specified
294 /// register class in the form of a pair of TargetRegisterClass iterators.
295 std::pair<TargetRegisterClass::iterator,TargetRegisterClass::iterator>
296 ARMBaseRegisterInfo::getAllocationOrder(const TargetRegisterClass *RC,
297 unsigned HintType, unsigned HintReg,
298 const MachineFunction &MF) const {
299 // Alternative register allocation orders when favoring even / odd registers
300 // of register pairs.
302 // No FP, R9 is available.
303 static const unsigned GPREven1[] = {
304 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8, ARM::R10,
305 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7,
308 static const unsigned GPROdd1[] = {
309 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R9, ARM::R11,
310 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6,
314 // FP is R7, R9 is available.
315 static const unsigned GPREven2[] = {
316 ARM::R0, ARM::R2, ARM::R4, ARM::R8, ARM::R10,
317 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6,
320 static const unsigned GPROdd2[] = {
321 ARM::R1, ARM::R3, ARM::R5, ARM::R9, ARM::R11,
322 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6,
326 // FP is R11, R9 is available.
327 static const unsigned GPREven3[] = {
328 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8,
329 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7,
332 static const unsigned GPROdd3[] = {
333 ARM::R1, ARM::R3, ARM::R5, ARM::R6, ARM::R9,
334 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R7,
338 // No FP, R9 is not available.
339 static const unsigned GPREven4[] = {
340 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R10,
341 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8,
344 static const unsigned GPROdd4[] = {
345 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R11,
346 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8,
350 // FP is R7, R9 is not available.
351 static const unsigned GPREven5[] = {
352 ARM::R0, ARM::R2, ARM::R4, ARM::R10,
353 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6, ARM::R8,
356 static const unsigned GPROdd5[] = {
357 ARM::R1, ARM::R3, ARM::R5, ARM::R11,
358 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8,
362 // FP is R11, R9 is not available.
363 static const unsigned GPREven6[] = {
364 ARM::R0, ARM::R2, ARM::R4, ARM::R6,
365 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8
367 static const unsigned GPROdd6[] = {
368 ARM::R1, ARM::R3, ARM::R5, ARM::R7,
369 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8
373 if (HintType == ARMRI::RegPairEven) {
374 if (isPhysicalRegister(HintReg) && getRegisterPairEven(HintReg, MF) == 0)
375 // It's no longer possible to fulfill this hint. Return the default
377 return std::make_pair(RC->allocation_order_begin(MF),
378 RC->allocation_order_end(MF));
380 if (!STI.isTargetDarwin() && !hasFP(MF)) {
381 if (!STI.isR9Reserved())
382 return std::make_pair(GPREven1,
383 GPREven1 + (sizeof(GPREven1)/sizeof(unsigned)));
385 return std::make_pair(GPREven4,
386 GPREven4 + (sizeof(GPREven4)/sizeof(unsigned)));
387 } else if (FramePtr == ARM::R7) {
388 if (!STI.isR9Reserved())
389 return std::make_pair(GPREven2,
390 GPREven2 + (sizeof(GPREven2)/sizeof(unsigned)));
392 return std::make_pair(GPREven5,
393 GPREven5 + (sizeof(GPREven5)/sizeof(unsigned)));
394 } else { // FramePtr == ARM::R11
395 if (!STI.isR9Reserved())
396 return std::make_pair(GPREven3,
397 GPREven3 + (sizeof(GPREven3)/sizeof(unsigned)));
399 return std::make_pair(GPREven6,
400 GPREven6 + (sizeof(GPREven6)/sizeof(unsigned)));
402 } else if (HintType == ARMRI::RegPairOdd) {
403 if (isPhysicalRegister(HintReg) && getRegisterPairOdd(HintReg, MF) == 0)
404 // It's no longer possible to fulfill this hint. Return the default
406 return std::make_pair(RC->allocation_order_begin(MF),
407 RC->allocation_order_end(MF));
409 if (!STI.isTargetDarwin() && !hasFP(MF)) {
410 if (!STI.isR9Reserved())
411 return std::make_pair(GPROdd1,
412 GPROdd1 + (sizeof(GPROdd1)/sizeof(unsigned)));
414 return std::make_pair(GPROdd4,
415 GPROdd4 + (sizeof(GPROdd4)/sizeof(unsigned)));
416 } else if (FramePtr == ARM::R7) {
417 if (!STI.isR9Reserved())
418 return std::make_pair(GPROdd2,
419 GPROdd2 + (sizeof(GPROdd2)/sizeof(unsigned)));
421 return std::make_pair(GPROdd5,
422 GPROdd5 + (sizeof(GPROdd5)/sizeof(unsigned)));
423 } else { // FramePtr == ARM::R11
424 if (!STI.isR9Reserved())
425 return std::make_pair(GPROdd3,
426 GPROdd3 + (sizeof(GPROdd3)/sizeof(unsigned)));
428 return std::make_pair(GPROdd6,
429 GPROdd6 + (sizeof(GPROdd6)/sizeof(unsigned)));
432 return std::make_pair(RC->allocation_order_begin(MF),
433 RC->allocation_order_end(MF));
436 /// ResolveRegAllocHint - Resolves the specified register allocation hint
437 /// to a physical register. Returns the physical register if it is successful.
439 ARMBaseRegisterInfo::ResolveRegAllocHint(unsigned Type, unsigned Reg,
440 const MachineFunction &MF) const {
441 if (Reg == 0 || !isPhysicalRegister(Reg))
445 else if (Type == (unsigned)ARMRI::RegPairOdd)
447 return getRegisterPairOdd(Reg, MF);
448 else if (Type == (unsigned)ARMRI::RegPairEven)
450 return getRegisterPairEven(Reg, MF);
455 ARMBaseRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg,
456 MachineFunction &MF) const {
457 MachineRegisterInfo *MRI = &MF.getRegInfo();
458 std::pair<unsigned, unsigned> Hint = MRI->getRegAllocationHint(Reg);
459 if ((Hint.first == (unsigned)ARMRI::RegPairOdd ||
460 Hint.first == (unsigned)ARMRI::RegPairEven) &&
461 Hint.second && TargetRegisterInfo::isVirtualRegister(Hint.second)) {
462 // If 'Reg' is one of the even / odd register pair and it's now changed
463 // (e.g. coalesced) into a different register. The other register of the
464 // pair allocation hint must be updated to reflect the relationship
466 unsigned OtherReg = Hint.second;
467 Hint = MRI->getRegAllocationHint(OtherReg);
468 if (Hint.second == Reg)
469 // Make sure the pair has not already divorced.
470 MRI->setRegAllocationHint(OtherReg, Hint.first, NewReg);
474 static unsigned calculateMaxStackAlignment(const MachineFrameInfo *FFI) {
475 unsigned MaxAlign = 0;
477 for (int i = FFI->getObjectIndexBegin(),
478 e = FFI->getObjectIndexEnd(); i != e; ++i) {
479 if (FFI->isDeadObjectIndex(i))
482 unsigned Align = FFI->getObjectAlignment(i);
483 MaxAlign = std::max(MaxAlign, Align);
489 /// hasFP - Return true if the specified function should have a dedicated frame
490 /// pointer register. This is true if the function has variable sized allocas
491 /// or if frame pointer elimination is disabled.
493 bool ARMBaseRegisterInfo::hasFP(const MachineFunction &MF) const {
494 const MachineFrameInfo *MFI = MF.getFrameInfo();
495 return (NoFramePointerElim ||
496 needsStackRealignment(MF) ||
497 MFI->hasVarSizedObjects() ||
498 MFI->isFrameAddressTaken());
501 bool ARMBaseRegisterInfo::
502 needsStackRealignment(const MachineFunction &MF) const {
503 // Only do this for ARM if explicitly enabled
504 // FIXME: Once it's passing all the tests, enable by default
505 if (!ARMDynamicStackAlign)
508 const MachineFrameInfo *MFI = MF.getFrameInfo();
509 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
510 unsigned StackAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
511 return (RealignStack &&
512 !AFI->isThumb1OnlyFunction() &&
513 (MFI->getMaxAlignment() > StackAlign) &&
514 !MFI->hasVarSizedObjects());
518 bool ARMBaseRegisterInfo::cannotEliminateFrame(const MachineFunction &MF) const {
519 const MachineFrameInfo *MFI = MF.getFrameInfo();
520 if (NoFramePointerElim && MFI->hasCalls())
522 return MFI->hasVarSizedObjects() || MFI->isFrameAddressTaken();
525 /// estimateStackSize - Estimate and return the size of the frame.
526 static unsigned estimateStackSize(MachineFunction &MF, MachineFrameInfo *MFI) {
527 const MachineFrameInfo *FFI = MF.getFrameInfo();
529 for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) {
530 int FixedOff = -FFI->getObjectOffset(i);
531 if (FixedOff > Offset) Offset = FixedOff;
533 for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) {
534 if (FFI->isDeadObjectIndex(i))
536 Offset += FFI->getObjectSize(i);
537 unsigned Align = FFI->getObjectAlignment(i);
538 // Adjust to alignment boundary
539 Offset = (Offset+Align-1)/Align*Align;
541 return (unsigned)Offset;
544 /// estimateRSStackSizeLimit - Look at each instruction that references stack
545 /// frames and return the stack size limit beyond which some of these
546 /// instructions will require scratch register during their expansion later.
548 ARMBaseRegisterInfo::estimateRSStackSizeLimit(MachineFunction &MF) const {
549 unsigned Limit = (1 << 12) - 1;
550 for (MachineFunction::iterator BB = MF.begin(),E = MF.end(); BB != E; ++BB) {
551 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
553 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
554 if (!I->getOperand(i).isFI()) continue;
556 const TargetInstrDesc &Desc = TII.get(I->getOpcode());
557 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
558 if (AddrMode == ARMII::AddrMode3 ||
559 AddrMode == ARMII::AddrModeT2_i8)
562 if (AddrMode == ARMII::AddrMode5 ||
563 AddrMode == ARMII::AddrModeT2_i8s4)
564 Limit = std::min(Limit, ((1U << 8) - 1) * 4);
566 if (AddrMode == ARMII::AddrModeT2_i12 && hasFP(MF))
567 // When the stack offset is negative, we will end up using
568 // the i8 instructions instead.
570 break; // At most one FI per instruction
579 ARMBaseRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
580 RegScavenger *RS) const {
581 // This tells PEI to spill the FP as if it is any other callee-save register
582 // to take advantage the eliminateFrameIndex machinery. This also ensures it
583 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
584 // to combine multiple loads / stores.
585 bool CanEliminateFrame = true;
586 bool CS1Spilled = false;
587 bool LRSpilled = false;
588 unsigned NumGPRSpills = 0;
589 SmallVector<unsigned, 4> UnspilledCS1GPRs;
590 SmallVector<unsigned, 4> UnspilledCS2GPRs;
591 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
593 MachineFrameInfo *MFI = MF.getFrameInfo();
595 // Calculate and set max stack object alignment early, so we can decide
596 // whether we will need stack realignment (and thus FP).
597 if (ARMDynamicStackAlign) {
598 unsigned MaxAlign = std::max(MFI->getMaxAlignment(),
599 calculateMaxStackAlignment(MFI));
600 MFI->setMaxAlignment(MaxAlign);
603 // Don't spill FP if the frame can be eliminated. This is determined
604 // by scanning the callee-save registers to see if any is used.
605 const unsigned *CSRegs = getCalleeSavedRegs();
606 const TargetRegisterClass* const *CSRegClasses = getCalleeSavedRegClasses();
607 for (unsigned i = 0; CSRegs[i]; ++i) {
608 unsigned Reg = CSRegs[i];
609 bool Spilled = false;
610 if (MF.getRegInfo().isPhysRegUsed(Reg)) {
611 AFI->setCSRegisterIsSpilled(Reg);
613 CanEliminateFrame = false;
615 // Check alias registers too.
616 for (const unsigned *Aliases = getAliasSet(Reg); *Aliases; ++Aliases) {
617 if (MF.getRegInfo().isPhysRegUsed(*Aliases)) {
619 CanEliminateFrame = false;
624 if (CSRegClasses[i] == ARM::GPRRegisterClass ||
625 CSRegClasses[i] == ARM::tGPRRegisterClass) {
629 if (!STI.isTargetDarwin()) {
636 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
651 if (!STI.isTargetDarwin()) {
652 UnspilledCS1GPRs.push_back(Reg);
662 UnspilledCS1GPRs.push_back(Reg);
665 UnspilledCS2GPRs.push_back(Reg);
672 bool ForceLRSpill = false;
673 if (!LRSpilled && AFI->isThumb1OnlyFunction()) {
674 unsigned FnSize = TII.GetFunctionSizeInBytes(MF);
675 // Force LR to be spilled if the Thumb function size is > 2048. This enables
676 // use of BL to implement far jump. If it turns out that it's not needed
677 // then the branch fix up path will undo it.
678 if (FnSize >= (1 << 11)) {
679 CanEliminateFrame = false;
684 bool ExtraCSSpill = false;
685 if (!CanEliminateFrame || cannotEliminateFrame(MF)) {
686 AFI->setHasStackFrame(true);
688 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
689 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
690 if (!LRSpilled && CS1Spilled) {
691 MF.getRegInfo().setPhysRegUsed(ARM::LR);
692 AFI->setCSRegisterIsSpilled(ARM::LR);
694 UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(),
695 UnspilledCS1GPRs.end(), (unsigned)ARM::LR));
696 ForceLRSpill = false;
700 // Darwin ABI requires FP to point to the stack slot that contains the
702 if (STI.isTargetDarwin() || hasFP(MF)) {
703 MF.getRegInfo().setPhysRegUsed(FramePtr);
707 // If stack and double are 8-byte aligned and we are spilling an odd number
708 // of GPRs. Spill one extra callee save GPR so we won't have to pad between
709 // the integer and double callee save areas.
710 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
711 if (TargetAlign == 8 && (NumGPRSpills & 1)) {
712 if (CS1Spilled && !UnspilledCS1GPRs.empty()) {
713 for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) {
714 unsigned Reg = UnspilledCS1GPRs[i];
715 // Don't spill high register if the function is thumb1
716 if (!AFI->isThumb1OnlyFunction() ||
717 isARMLowRegister(Reg) || Reg == ARM::LR) {
718 MF.getRegInfo().setPhysRegUsed(Reg);
719 AFI->setCSRegisterIsSpilled(Reg);
720 if (!isReservedReg(MF, Reg))
725 } else if (!UnspilledCS2GPRs.empty() &&
726 !AFI->isThumb1OnlyFunction()) {
727 unsigned Reg = UnspilledCS2GPRs.front();
728 MF.getRegInfo().setPhysRegUsed(Reg);
729 AFI->setCSRegisterIsSpilled(Reg);
730 if (!isReservedReg(MF, Reg))
735 // Estimate if we might need to scavenge a register at some point in order
736 // to materialize a stack offset. If so, either spill one additional
737 // callee-saved register or reserve a special spill slot to facilitate
738 // register scavenging. Thumb1 needs a spill slot for stack pointer
739 // adjustments also, even when the frame itself is small.
740 if (RS && !ExtraCSSpill) {
741 MachineFrameInfo *MFI = MF.getFrameInfo();
742 // If any of the stack slot references may be out of range of an
743 // immediate offset, make sure a register (or a spill slot) is
744 // available for the register scavenger. Note that if we're indexing
745 // off the frame pointer, the effective stack size is 4 bytes larger
746 // since the FP points to the stack slot of the previous FP.
747 if (estimateStackSize(MF, MFI) + (hasFP(MF) ? 4 : 0)
748 >= estimateRSStackSizeLimit(MF)) {
749 // If any non-reserved CS register isn't spilled, just spill one or two
750 // extra. That should take care of it!
751 unsigned NumExtras = TargetAlign / 4;
752 SmallVector<unsigned, 2> Extras;
753 while (NumExtras && !UnspilledCS1GPRs.empty()) {
754 unsigned Reg = UnspilledCS1GPRs.back();
755 UnspilledCS1GPRs.pop_back();
756 if (!isReservedReg(MF, Reg)) {
757 Extras.push_back(Reg);
761 // For non-Thumb1 functions, also check for hi-reg CS registers
762 if (!AFI->isThumb1OnlyFunction()) {
763 while (NumExtras && !UnspilledCS2GPRs.empty()) {
764 unsigned Reg = UnspilledCS2GPRs.back();
765 UnspilledCS2GPRs.pop_back();
766 if (!isReservedReg(MF, Reg)) {
767 Extras.push_back(Reg);
772 if (Extras.size() && NumExtras == 0) {
773 for (unsigned i = 0, e = Extras.size(); i != e; ++i) {
774 MF.getRegInfo().setPhysRegUsed(Extras[i]);
775 AFI->setCSRegisterIsSpilled(Extras[i]);
777 } else if (!AFI->isThumb1OnlyFunction()) {
778 // note: Thumb1 functions spill to R12, not the stack.
779 // Reserve a slot closest to SP or frame pointer.
780 const TargetRegisterClass *RC = ARM::GPRRegisterClass;
781 RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
782 RC->getAlignment()));
789 MF.getRegInfo().setPhysRegUsed(ARM::LR);
790 AFI->setCSRegisterIsSpilled(ARM::LR);
791 AFI->setLRIsSpilledForFarJump(true);
795 unsigned ARMBaseRegisterInfo::getRARegister() const {
799 unsigned ARMBaseRegisterInfo::getFrameRegister(MachineFunction &MF) const {
800 if (STI.isTargetDarwin() || hasFP(MF))
805 unsigned ARMBaseRegisterInfo::getEHExceptionRegister() const {
806 llvm_unreachable("What is the exception register");
810 unsigned ARMBaseRegisterInfo::getEHHandlerRegister() const {
811 llvm_unreachable("What is the exception handler register");
815 int ARMBaseRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const {
816 return ARMGenRegisterInfo::getDwarfRegNumFull(RegNum, 0);
819 unsigned ARMBaseRegisterInfo::getRegisterPairEven(unsigned Reg,
820 const MachineFunction &MF) const {
823 // Return 0 if either register of the pair is a special register.
832 return isReservedReg(MF, ARM::R7) ? 0 : ARM::R6;
834 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R8;
836 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R10;
908 unsigned ARMBaseRegisterInfo::getRegisterPairOdd(unsigned Reg,
909 const MachineFunction &MF) const {
912 // Return 0 if either register of the pair is a special register.
921 return isReservedReg(MF, ARM::R7) ? 0 : ARM::R7;
923 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R9;
925 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R11;
997 /// emitLoadConstPool - Emits a load from constpool to materialize the
998 /// specified immediate.
999 void ARMBaseRegisterInfo::
1000 emitLoadConstPool(MachineBasicBlock &MBB,
1001 MachineBasicBlock::iterator &MBBI,
1003 unsigned DestReg, unsigned SubIdx, int Val,
1004 ARMCC::CondCodes Pred,
1005 unsigned PredReg) const {
1006 MachineFunction &MF = *MBB.getParent();
1007 MachineConstantPool *ConstantPool = MF.getConstantPool();
1009 ConstantInt::get(Type::getInt32Ty(MF.getFunction()->getContext()), Val);
1010 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 4);
1012 BuildMI(MBB, MBBI, dl, TII.get(ARM::LDRcp))
1013 .addReg(DestReg, getDefRegState(true), SubIdx)
1014 .addConstantPoolIndex(Idx)
1015 .addReg(0).addImm(0).addImm(Pred).addReg(PredReg);
1018 bool ARMBaseRegisterInfo::
1019 requiresRegisterScavenging(const MachineFunction &MF) const {
1023 bool ARMBaseRegisterInfo::
1024 requiresFrameIndexScavenging(const MachineFunction &MF) const {
1025 return ScavengeFrameIndexVals;
1028 // hasReservedCallFrame - Under normal circumstances, when a frame pointer is
1029 // not required, we reserve argument space for call sites in the function
1030 // immediately on entry to the current function. This eliminates the need for
1031 // add/sub sp brackets around call sites. Returns true if the call frame is
1032 // included as part of the stack frame.
1033 bool ARMBaseRegisterInfo::
1034 hasReservedCallFrame(MachineFunction &MF) const {
1035 const MachineFrameInfo *FFI = MF.getFrameInfo();
1036 unsigned CFSize = FFI->getMaxCallFrameSize();
1037 // It's not always a good idea to include the call frame as part of the
1038 // stack frame. ARM (especially Thumb) has small immediate offset to
1039 // address the stack frame. So a large call frame can cause poor codegen
1040 // and may even makes it impossible to scavenge a register.
1041 if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12
1044 return !MF.getFrameInfo()->hasVarSizedObjects();
1048 emitSPUpdate(bool isARM,
1049 MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
1050 DebugLoc dl, const ARMBaseInstrInfo &TII,
1052 ARMCC::CondCodes Pred = ARMCC::AL, unsigned PredReg = 0) {
1054 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, ARM::SP, NumBytes,
1055 Pred, PredReg, TII);
1057 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::SP, ARM::SP, NumBytes,
1058 Pred, PredReg, TII);
1062 void ARMBaseRegisterInfo::
1063 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
1064 MachineBasicBlock::iterator I) const {
1065 if (!hasReservedCallFrame(MF)) {
1066 // If we have alloca, convert as follows:
1067 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
1068 // ADJCALLSTACKUP -> add, sp, sp, amount
1069 MachineInstr *Old = I;
1070 DebugLoc dl = Old->getDebugLoc();
1071 unsigned Amount = Old->getOperand(0).getImm();
1073 // We need to keep the stack aligned properly. To do this, we round the
1074 // amount of space needed for the outgoing arguments up to the next
1075 // alignment boundary.
1076 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1077 Amount = (Amount+Align-1)/Align*Align;
1079 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1080 assert(!AFI->isThumb1OnlyFunction() &&
1081 "This eliminateCallFramePseudoInstr does not suppor Thumb1!");
1082 bool isARM = !AFI->isThumbFunction();
1084 // Replace the pseudo instruction with a new instruction...
1085 unsigned Opc = Old->getOpcode();
1086 ARMCC::CondCodes Pred = (ARMCC::CondCodes)Old->getOperand(1).getImm();
1087 // FIXME: Thumb2 version of ADJCALLSTACKUP and ADJCALLSTACKDOWN?
1088 if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) {
1089 // Note: PredReg is operand 2 for ADJCALLSTACKDOWN.
1090 unsigned PredReg = Old->getOperand(2).getReg();
1091 emitSPUpdate(isARM, MBB, I, dl, TII, -Amount, Pred, PredReg);
1093 // Note: PredReg is operand 3 for ADJCALLSTACKUP.
1094 unsigned PredReg = Old->getOperand(3).getReg();
1095 assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP);
1096 emitSPUpdate(isARM, MBB, I, dl, TII, Amount, Pred, PredReg);
1103 /// findScratchRegister - Find a 'free' ARM register. If register scavenger
1104 /// is not being used, R12 is available. Otherwise, try for a call-clobbered
1105 /// register first and then a spilled callee-saved register if that fails.
1107 unsigned findScratchRegister(RegScavenger *RS, const TargetRegisterClass *RC,
1108 ARMFunctionInfo *AFI) {
1109 unsigned Reg = RS ? RS->FindUnusedReg(RC) : (unsigned) ARM::R12;
1110 assert(!AFI->isThumb1OnlyFunction());
1115 ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
1116 int SPAdj, int *Value,
1117 RegScavenger *RS) const {
1119 MachineInstr &MI = *II;
1120 MachineBasicBlock &MBB = *MI.getParent();
1121 MachineFunction &MF = *MBB.getParent();
1122 const MachineFrameInfo *MFI = MF.getFrameInfo();
1123 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1124 assert(!AFI->isThumb1OnlyFunction() &&
1125 "This eliminateFrameIndex does not support Thumb1!");
1127 while (!MI.getOperand(i).isFI()) {
1129 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
1132 unsigned FrameReg = ARM::SP;
1133 int FrameIndex = MI.getOperand(i).getIndex();
1134 int Offset = MFI->getObjectOffset(FrameIndex) + MFI->getStackSize() + SPAdj;
1136 // When doing dynamic stack realignment, all of these need to change(?)
1137 if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex))
1138 Offset -= AFI->getGPRCalleeSavedArea1Offset();
1139 else if (AFI->isGPRCalleeSavedArea2Frame(FrameIndex))
1140 Offset -= AFI->getGPRCalleeSavedArea2Offset();
1141 else if (AFI->isDPRCalleeSavedAreaFrame(FrameIndex))
1142 Offset -= AFI->getDPRCalleeSavedAreaOffset();
1143 else if (needsStackRealignment(MF)) {
1144 // When dynamically realigning the stack, use the frame pointer for
1145 // parameters, and the stack pointer for locals.
1146 assert (hasFP(MF) && "dynamic stack realignment without a FP!");
1147 if (FrameIndex < 0) {
1148 FrameReg = getFrameRegister(MF);
1149 Offset -= AFI->getFramePtrSpillOffset();
1150 // When referencing from the frame pointer, stack pointer adjustments
1154 } else if (hasFP(MF) && AFI->hasStackFrame()) {
1155 assert(SPAdj == 0 && "Unexpected stack offset!");
1156 // Use frame pointer to reference fixed objects unless this is a
1157 // frameless function.
1158 FrameReg = getFrameRegister(MF);
1159 Offset -= AFI->getFramePtrSpillOffset();
1162 // modify MI as necessary to handle as much of 'Offset' as possible
1164 if (!AFI->isThumbFunction())
1165 Done = rewriteARMFrameIndex(MI, i, FrameReg, Offset, TII);
1167 assert(AFI->isThumb2Function());
1168 Done = rewriteT2FrameIndex(MI, i, FrameReg, Offset, TII);
1173 // If we get here, the immediate doesn't fit into the instruction. We folded
1174 // as much as possible above, handle the rest, providing a register that is
1177 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode4) &&
1178 "This code isn't needed if offset already handled!");
1180 unsigned ScratchReg = 0;
1181 int PIdx = MI.findFirstPredOperandIdx();
1182 ARMCC::CondCodes Pred = (PIdx == -1)
1183 ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm();
1184 unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg();
1186 // Must be addrmode4.
1187 MI.getOperand(i).ChangeToRegister(FrameReg, false, false, false);
1189 if (!ScavengeFrameIndexVals) {
1190 // Insert a set of r12 with the full address: r12 = sp + offset
1191 // If the offset we have is too large to fit into the instruction, we need
1192 // to form it with a series of ADDri's. Do this by taking 8-bit chunks
1194 ScratchReg = findScratchRegister(RS, ARM::GPRRegisterClass, AFI);
1195 if (ScratchReg == 0)
1196 // No register is "free". Scavenge a register.
1197 ScratchReg = RS->scavengeRegister(ARM::GPRRegisterClass, II, SPAdj);
1199 ScratchReg = MF.getRegInfo().createVirtualRegister(ARM::GPRRegisterClass);
1202 if (!AFI->isThumbFunction())
1203 emitARMRegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg,
1204 Offset, Pred, PredReg, TII);
1206 assert(AFI->isThumb2Function());
1207 emitT2RegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg,
1208 Offset, Pred, PredReg, TII);
1210 MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true);
1211 if (!ReuseFrameIndexVals || !ScavengeFrameIndexVals)
1217 /// Move iterator pass the next bunch of callee save load / store ops for
1218 /// the particular spill area (1: integer area 1, 2: integer area 2,
1219 /// 3: fp area, 0: don't care).
1220 static void movePastCSLoadStoreOps(MachineBasicBlock &MBB,
1221 MachineBasicBlock::iterator &MBBI,
1222 int Opc1, int Opc2, unsigned Area,
1223 const ARMSubtarget &STI) {
1224 while (MBBI != MBB.end() &&
1225 ((MBBI->getOpcode() == Opc1) || (MBBI->getOpcode() == Opc2)) &&
1226 MBBI->getOperand(1).isFI()) {
1229 unsigned Category = 0;
1230 switch (MBBI->getOperand(0).getReg()) {
1231 case ARM::R4: case ARM::R5: case ARM::R6: case ARM::R7:
1235 case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11:
1236 Category = STI.isTargetDarwin() ? 2 : 1;
1238 case ARM::D8: case ARM::D9: case ARM::D10: case ARM::D11:
1239 case ARM::D12: case ARM::D13: case ARM::D14: case ARM::D15:
1246 if (Done || Category != Area)
1254 void ARMBaseRegisterInfo::
1255 emitPrologue(MachineFunction &MF) const {
1256 MachineBasicBlock &MBB = MF.front();
1257 MachineBasicBlock::iterator MBBI = MBB.begin();
1258 MachineFrameInfo *MFI = MF.getFrameInfo();
1259 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1260 assert(!AFI->isThumb1OnlyFunction() &&
1261 "This emitPrologue does not suppor Thumb1!");
1262 bool isARM = !AFI->isThumbFunction();
1263 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1264 unsigned NumBytes = MFI->getStackSize();
1265 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
1266 DebugLoc dl = (MBBI != MBB.end() ?
1267 MBBI->getDebugLoc() : DebugLoc::getUnknownLoc());
1269 // Determine the sizes of each callee-save spill areas and record which frame
1270 // belongs to which callee-save spill areas.
1271 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
1272 int FramePtrSpillFI = 0;
1274 // Allocate the vararg register save area. This is not counted in NumBytes.
1276 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -VARegSaveSize);
1278 if (!AFI->hasStackFrame()) {
1280 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes);
1284 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1285 unsigned Reg = CSI[i].getReg();
1286 int FI = CSI[i].getFrameIdx();
1293 if (Reg == FramePtr)
1294 FramePtrSpillFI = FI;
1295 AFI->addGPRCalleeSavedArea1Frame(FI);
1302 if (Reg == FramePtr)
1303 FramePtrSpillFI = FI;
1304 if (STI.isTargetDarwin()) {
1305 AFI->addGPRCalleeSavedArea2Frame(FI);
1308 AFI->addGPRCalleeSavedArea1Frame(FI);
1313 AFI->addDPRCalleeSavedAreaFrame(FI);
1318 // Build the new SUBri to adjust SP for integer callee-save spill area 1.
1319 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -GPRCS1Size);
1320 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, ARM::t2STRi12, 1, STI);
1322 // Set FP to point to the stack slot that contains the previous FP.
1323 // For Darwin, FP is R7, which has now been stored in spill area 1.
1324 // Otherwise, if this is not Darwin, all the callee-saved registers go
1325 // into spill area 1, including the FP in R11. In either case, it is
1326 // now safe to emit this assignment.
1327 if (STI.isTargetDarwin() || hasFP(MF)) {
1328 unsigned ADDriOpc = !AFI->isThumbFunction() ? ARM::ADDri : ARM::t2ADDri;
1329 MachineInstrBuilder MIB =
1330 BuildMI(MBB, MBBI, dl, TII.get(ADDriOpc), FramePtr)
1331 .addFrameIndex(FramePtrSpillFI).addImm(0);
1332 AddDefaultCC(AddDefaultPred(MIB));
1335 // Build the new SUBri to adjust SP for integer callee-save spill area 2.
1336 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -GPRCS2Size);
1338 // Build the new SUBri to adjust SP for FP callee-save spill area.
1339 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, ARM::t2STRi12, 2, STI);
1340 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -DPRCSSize);
1342 // Determine starting offsets of spill areas.
1343 unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize);
1344 unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize;
1345 unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size;
1346 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) + NumBytes);
1347 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
1348 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
1349 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
1351 NumBytes = DPRCSOffset;
1353 // Insert it after all the callee-save spills.
1354 movePastCSLoadStoreOps(MBB, MBBI, ARM::FSTD, 0, 3, STI);
1355 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes);
1358 if (STI.isTargetELF() && hasFP(MF)) {
1359 MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() -
1360 AFI->getFramePtrSpillOffset());
1363 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
1364 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
1365 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
1367 // If we need dynamic stack realignment, do it here.
1368 if (needsStackRealignment(MF)) {
1370 unsigned MaxAlign = MFI->getMaxAlignment();
1371 assert (!AFI->isThumb1OnlyFunction());
1372 Opc = AFI->isThumbFunction() ? ARM::t2BICri : ARM::BICri;
1374 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(Opc), ARM::SP)
1375 .addReg(ARM::SP, RegState::Kill)
1376 .addImm(MaxAlign-1)));
1380 static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) {
1381 for (unsigned i = 0; CSRegs[i]; ++i)
1382 if (Reg == CSRegs[i])
1387 static bool isCSRestore(MachineInstr *MI,
1388 const ARMBaseInstrInfo &TII,
1389 const unsigned *CSRegs) {
1390 return ((MI->getOpcode() == (int)ARM::FLDD ||
1391 MI->getOpcode() == (int)ARM::LDR ||
1392 MI->getOpcode() == (int)ARM::t2LDRi12) &&
1393 MI->getOperand(1).isFI() &&
1394 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs));
1397 void ARMBaseRegisterInfo::
1398 emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const {
1399 MachineBasicBlock::iterator MBBI = prior(MBB.end());
1400 assert(MBBI->getDesc().isReturn() &&
1401 "Can only insert epilog into returning blocks");
1402 DebugLoc dl = MBBI->getDebugLoc();
1403 MachineFrameInfo *MFI = MF.getFrameInfo();
1404 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1405 assert(!AFI->isThumb1OnlyFunction() &&
1406 "This emitEpilogue does not suppor Thumb1!");
1407 bool isARM = !AFI->isThumbFunction();
1409 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1410 int NumBytes = (int)MFI->getStackSize();
1412 if (!AFI->hasStackFrame()) {
1414 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes);
1416 // Unwind MBBI to point to first LDR / FLDD.
1417 const unsigned *CSRegs = getCalleeSavedRegs();
1418 if (MBBI != MBB.begin()) {
1421 while (MBBI != MBB.begin() && isCSRestore(MBBI, TII, CSRegs));
1422 if (!isCSRestore(MBBI, TII, CSRegs))
1426 // Move SP to start of FP callee save spill area.
1427 NumBytes -= (AFI->getGPRCalleeSavedArea1Size() +
1428 AFI->getGPRCalleeSavedArea2Size() +
1429 AFI->getDPRCalleeSavedAreaSize());
1431 // Darwin ABI requires FP to point to the stack slot that contains the
1433 bool HasFP = hasFP(MF);
1434 if ((STI.isTargetDarwin() && NumBytes) || HasFP) {
1435 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1436 // Reset SP based on frame pointer only if the stack frame extends beyond
1437 // frame pointer stack slot or target is ELF and the function has FP.
1439 AFI->getGPRCalleeSavedArea2Size() ||
1440 AFI->getDPRCalleeSavedAreaSize() ||
1441 AFI->getDPRCalleeSavedAreaOffset()) {
1444 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, FramePtr, -NumBytes,
1447 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::SP, FramePtr, -NumBytes,
1452 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), ARM::SP)
1454 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
1456 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVgpr2gpr), ARM::SP)
1460 } else if (NumBytes)
1461 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes);
1463 // Move SP to start of integer callee save spill area 2.
1464 movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 0, 3, STI);
1465 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getDPRCalleeSavedAreaSize());
1467 // Move SP to start of integer callee save spill area 1.
1468 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, ARM::t2LDRi12, 2, STI);
1469 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getGPRCalleeSavedArea2Size());
1471 // Move SP to SP upon entry to the function.
1472 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, ARM::t2LDRi12, 1, STI);
1473 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getGPRCalleeSavedArea1Size());
1477 emitSPUpdate(isARM, MBB, MBBI, dl, TII, VARegSaveSize);
1480 #include "ARMGenRegisterInfo.inc"