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 ReuseFrameIndexVals("arm-reuse-frame-index-vals", cl::Hidden, cl::init(true),
45 cl::desc("Reuse repeated frame index values"));
50 unsigned ARMBaseRegisterInfo::getRegisterNumbering(unsigned RegEnum,
58 llvm_unreachable("Unknown ARM register!");
59 case R0: case D0: case Q0: return 0;
60 case R1: case D1: case Q1: return 1;
61 case R2: case D2: case Q2: return 2;
62 case R3: case D3: case Q3: return 3;
63 case R4: case D4: case Q4: return 4;
64 case R5: case D5: case Q5: return 5;
65 case R6: case D6: case Q6: return 6;
66 case R7: case D7: case Q7: return 7;
67 case R8: case D8: case Q8: return 8;
68 case R9: case D9: case Q9: return 9;
69 case R10: case D10: case Q10: return 10;
70 case R11: case D11: case Q11: return 11;
71 case R12: case D12: case Q12: return 12;
72 case SP: case D13: case Q13: return 13;
73 case LR: case D14: case Q14: return 14;
74 case PC: case D15: case Q15: return 15;
93 case S0: case S1: case S2: case S3:
94 case S4: case S5: case S6: case S7:
95 case S8: case S9: case S10: case S11:
96 case S12: case S13: case S14: case S15:
97 case S16: case S17: case S18: case S19:
98 case S20: case S21: case S22: case S23:
99 case S24: case S25: case S26: case S27:
100 case S28: case S29: case S30: case S31: {
104 default: return 0; // Avoid compile time warning.
142 ARMBaseRegisterInfo::ARMBaseRegisterInfo(const ARMBaseInstrInfo &tii,
143 const ARMSubtarget &sti)
144 : ARMGenRegisterInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
146 FramePtr((STI.isTargetDarwin() || STI.isThumb()) ? ARM::R7 : ARM::R11) {
150 ARMBaseRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
151 static const unsigned CalleeSavedRegs[] = {
152 ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8,
153 ARM::R7, ARM::R6, ARM::R5, ARM::R4,
155 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
156 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
160 static const unsigned DarwinCalleeSavedRegs[] = {
161 // Darwin ABI deviates from ARM standard ABI. R9 is not a callee-saved
163 ARM::LR, ARM::R7, ARM::R6, ARM::R5, ARM::R4,
164 ARM::R11, ARM::R10, ARM::R8,
166 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
167 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
170 return STI.isTargetDarwin() ? DarwinCalleeSavedRegs : CalleeSavedRegs;
173 BitVector ARMBaseRegisterInfo::
174 getReservedRegs(const MachineFunction &MF) const {
175 // FIXME: avoid re-calculating this everytime.
176 BitVector Reserved(getNumRegs());
177 Reserved.set(ARM::SP);
178 Reserved.set(ARM::PC);
179 Reserved.set(ARM::FPSCR);
180 if (STI.isTargetDarwin() || hasFP(MF))
181 Reserved.set(FramePtr);
182 // Some targets reserve R9.
183 if (STI.isR9Reserved())
184 Reserved.set(ARM::R9);
188 bool ARMBaseRegisterInfo::isReservedReg(const MachineFunction &MF,
189 unsigned Reg) const {
197 if (FramePtr == Reg && (STI.isTargetDarwin() || hasFP(MF)))
201 return STI.isR9Reserved();
207 const TargetRegisterClass *
208 ARMBaseRegisterInfo::getMatchingSuperRegClass(const TargetRegisterClass *A,
209 const TargetRegisterClass *B,
210 unsigned SubIdx) const {
218 if (A->getSize() == 8) {
219 if (B == &ARM::SPR_8RegClass)
220 return &ARM::DPR_8RegClass;
221 assert(B == &ARM::SPRRegClass && "Expecting SPR register class!");
222 if (A == &ARM::DPR_8RegClass)
224 return &ARM::DPR_VFP2RegClass;
227 if (A->getSize() == 16) {
228 if (B == &ARM::SPR_8RegClass)
229 return &ARM::QPR_8RegClass;
230 return &ARM::QPR_VFP2RegClass;
233 if (A->getSize() == 32) {
234 if (B == &ARM::SPR_8RegClass)
235 return 0; // Do not allow coalescing!
236 return &ARM::QQPR_VFP2RegClass;
239 assert(A->getSize() == 64 && "Expecting a QQQQ register class!");
240 return 0; // Do not allow coalescing!
247 if (A->getSize() == 16) {
248 if (B == &ARM::DPR_VFP2RegClass)
249 return &ARM::QPR_VFP2RegClass;
250 if (B == &ARM::DPR_8RegClass)
251 return 0; // Do not allow coalescing!
255 if (A->getSize() == 32) {
256 if (B == &ARM::DPR_VFP2RegClass)
257 return &ARM::QQPR_VFP2RegClass;
258 if (B == &ARM::DPR_8RegClass)
259 return 0; // Do not allow coalescing!
263 assert(A->getSize() == 64 && "Expecting a QQQQ register class!");
264 if (B != &ARM::DPRRegClass)
265 return 0; // Do not allow coalescing!
272 // D sub-registers of QQQQ registers.
273 if (A->getSize() == 64 && B == &ARM::DPRRegClass)
275 return 0; // Do not allow coalescing!
281 if (A->getSize() == 32) {
282 if (B == &ARM::QPR_VFP2RegClass)
283 return &ARM::QQPR_VFP2RegClass;
284 if (B == &ARM::QPR_8RegClass)
285 return 0; // Do not allow coalescing!
289 assert(A->getSize() == 64 && "Expecting a QQQQ register class!");
290 if (B == &ARM::QPRRegClass)
292 return 0; // Do not allow coalescing!
296 // Q sub-registers of QQQQ registers.
297 if (A->getSize() == 64 && B == &ARM::QPRRegClass)
299 return 0; // Do not allow coalescing!
306 ARMBaseRegisterInfo::canCombineSubRegIndices(const TargetRegisterClass *RC,
307 SmallVectorImpl<unsigned> &SubIndices,
308 unsigned &NewSubIdx) const {
310 unsigned Size = RC->getSize() * 8;
314 NewSubIdx = 0; // Whole register.
315 unsigned NumRegs = SubIndices.size();
317 // 8 D registers -> 1 QQQQ register.
318 return (Size == 512 &&
319 SubIndices[0] == ARM::dsub_0 &&
320 SubIndices[1] == ARM::dsub_1 &&
321 SubIndices[2] == ARM::dsub_2 &&
322 SubIndices[3] == ARM::dsub_3 &&
323 SubIndices[4] == ARM::dsub_4 &&
324 SubIndices[5] == ARM::dsub_5 &&
325 SubIndices[6] == ARM::dsub_6 &&
326 SubIndices[7] == ARM::dsub_7);
327 } else if (NumRegs == 4) {
328 if (SubIndices[0] == ARM::qsub_0) {
329 // 4 Q registers -> 1 QQQQ register.
330 return (Size == 512 &&
331 SubIndices[1] == ARM::qsub_1 &&
332 SubIndices[2] == ARM::qsub_2 &&
333 SubIndices[3] == ARM::qsub_3);
334 } else if (SubIndices[0] == ARM::dsub_0) {
335 // 4 D registers -> 1 QQ register.
337 SubIndices[1] == ARM::dsub_1 &&
338 SubIndices[2] == ARM::dsub_2 &&
339 SubIndices[3] == ARM::dsub_3) {
341 NewSubIdx = ARM::qqsub_0;
344 } else if (SubIndices[0] == ARM::dsub_4) {
345 // 4 D registers -> 1 QQ register (2nd).
347 SubIndices[1] == ARM::dsub_5 &&
348 SubIndices[2] == ARM::dsub_6 &&
349 SubIndices[3] == ARM::dsub_7) {
350 NewSubIdx = ARM::qqsub_1;
353 } else if (SubIndices[0] == ARM::ssub_0) {
354 // 4 S registers -> 1 Q register.
356 SubIndices[1] == ARM::ssub_1 &&
357 SubIndices[2] == ARM::ssub_2 &&
358 SubIndices[3] == ARM::ssub_3) {
360 NewSubIdx = ARM::qsub_0;
364 } else if (NumRegs == 2) {
365 if (SubIndices[0] == ARM::qsub_0) {
366 // 2 Q registers -> 1 QQ register.
367 if (Size >= 256 && SubIndices[1] == ARM::qsub_1) {
369 NewSubIdx = ARM::qqsub_0;
372 } else if (SubIndices[0] == ARM::qsub_2) {
373 // 2 Q registers -> 1 QQ register (2nd).
374 if (Size == 512 && SubIndices[1] == ARM::qsub_3) {
375 NewSubIdx = ARM::qqsub_1;
378 } else if (SubIndices[0] == ARM::dsub_0) {
379 // 2 D registers -> 1 Q register.
380 if (Size >= 128 && SubIndices[1] == ARM::dsub_1) {
382 NewSubIdx = ARM::qsub_0;
385 } else if (SubIndices[0] == ARM::dsub_2) {
386 // 2 D registers -> 1 Q register (2nd).
387 if (Size >= 256 && SubIndices[1] == ARM::dsub_3) {
388 NewSubIdx = ARM::qsub_1;
391 } else if (SubIndices[0] == ARM::dsub_4) {
392 // 2 D registers -> 1 Q register (3rd).
393 if (Size == 512 && SubIndices[1] == ARM::dsub_5) {
394 NewSubIdx = ARM::qsub_2;
397 } else if (SubIndices[0] == ARM::dsub_6) {
398 // 2 D registers -> 1 Q register (3rd).
399 if (Size == 512 && SubIndices[1] == ARM::dsub_7) {
400 NewSubIdx = ARM::qsub_3;
403 } else if (SubIndices[0] == ARM::ssub_0) {
404 // 2 S registers -> 1 D register.
405 if (SubIndices[1] == ARM::ssub_1) {
407 NewSubIdx = ARM::dsub_0;
410 } else if (SubIndices[0] == ARM::ssub_2) {
411 // 2 S registers -> 1 D register (2nd).
412 if (Size >= 128 && SubIndices[1] == ARM::ssub_3) {
413 NewSubIdx = ARM::dsub_1;
422 const TargetRegisterClass *
423 ARMBaseRegisterInfo::getPointerRegClass(unsigned Kind) const {
424 return ARM::GPRRegisterClass;
427 /// getAllocationOrder - Returns the register allocation order for a specified
428 /// register class in the form of a pair of TargetRegisterClass iterators.
429 std::pair<TargetRegisterClass::iterator,TargetRegisterClass::iterator>
430 ARMBaseRegisterInfo::getAllocationOrder(const TargetRegisterClass *RC,
431 unsigned HintType, unsigned HintReg,
432 const MachineFunction &MF) const {
433 // Alternative register allocation orders when favoring even / odd registers
434 // of register pairs.
436 // No FP, R9 is available.
437 static const unsigned GPREven1[] = {
438 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8, ARM::R10,
439 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7,
442 static const unsigned GPROdd1[] = {
443 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R9, ARM::R11,
444 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6,
448 // FP is R7, R9 is available.
449 static const unsigned GPREven2[] = {
450 ARM::R0, ARM::R2, ARM::R4, ARM::R8, ARM::R10,
451 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6,
454 static const unsigned GPROdd2[] = {
455 ARM::R1, ARM::R3, ARM::R5, ARM::R9, ARM::R11,
456 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6,
460 // FP is R11, R9 is available.
461 static const unsigned GPREven3[] = {
462 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8,
463 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7,
466 static const unsigned GPROdd3[] = {
467 ARM::R1, ARM::R3, ARM::R5, ARM::R6, ARM::R9,
468 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R7,
472 // No FP, R9 is not available.
473 static const unsigned GPREven4[] = {
474 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R10,
475 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8,
478 static const unsigned GPROdd4[] = {
479 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R11,
480 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8,
484 // FP is R7, R9 is not available.
485 static const unsigned GPREven5[] = {
486 ARM::R0, ARM::R2, ARM::R4, ARM::R10,
487 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6, ARM::R8,
490 static const unsigned GPROdd5[] = {
491 ARM::R1, ARM::R3, ARM::R5, ARM::R11,
492 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8,
496 // FP is R11, R9 is not available.
497 static const unsigned GPREven6[] = {
498 ARM::R0, ARM::R2, ARM::R4, ARM::R6,
499 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8
501 static const unsigned GPROdd6[] = {
502 ARM::R1, ARM::R3, ARM::R5, ARM::R7,
503 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8
507 if (HintType == ARMRI::RegPairEven) {
508 if (isPhysicalRegister(HintReg) && getRegisterPairEven(HintReg, MF) == 0)
509 // It's no longer possible to fulfill this hint. Return the default
511 return std::make_pair(RC->allocation_order_begin(MF),
512 RC->allocation_order_end(MF));
514 if (!STI.isTargetDarwin() && !hasFP(MF)) {
515 if (!STI.isR9Reserved())
516 return std::make_pair(GPREven1,
517 GPREven1 + (sizeof(GPREven1)/sizeof(unsigned)));
519 return std::make_pair(GPREven4,
520 GPREven4 + (sizeof(GPREven4)/sizeof(unsigned)));
521 } else if (FramePtr == ARM::R7) {
522 if (!STI.isR9Reserved())
523 return std::make_pair(GPREven2,
524 GPREven2 + (sizeof(GPREven2)/sizeof(unsigned)));
526 return std::make_pair(GPREven5,
527 GPREven5 + (sizeof(GPREven5)/sizeof(unsigned)));
528 } else { // FramePtr == ARM::R11
529 if (!STI.isR9Reserved())
530 return std::make_pair(GPREven3,
531 GPREven3 + (sizeof(GPREven3)/sizeof(unsigned)));
533 return std::make_pair(GPREven6,
534 GPREven6 + (sizeof(GPREven6)/sizeof(unsigned)));
536 } else if (HintType == ARMRI::RegPairOdd) {
537 if (isPhysicalRegister(HintReg) && getRegisterPairOdd(HintReg, MF) == 0)
538 // It's no longer possible to fulfill this hint. Return the default
540 return std::make_pair(RC->allocation_order_begin(MF),
541 RC->allocation_order_end(MF));
543 if (!STI.isTargetDarwin() && !hasFP(MF)) {
544 if (!STI.isR9Reserved())
545 return std::make_pair(GPROdd1,
546 GPROdd1 + (sizeof(GPROdd1)/sizeof(unsigned)));
548 return std::make_pair(GPROdd4,
549 GPROdd4 + (sizeof(GPROdd4)/sizeof(unsigned)));
550 } else if (FramePtr == ARM::R7) {
551 if (!STI.isR9Reserved())
552 return std::make_pair(GPROdd2,
553 GPROdd2 + (sizeof(GPROdd2)/sizeof(unsigned)));
555 return std::make_pair(GPROdd5,
556 GPROdd5 + (sizeof(GPROdd5)/sizeof(unsigned)));
557 } else { // FramePtr == ARM::R11
558 if (!STI.isR9Reserved())
559 return std::make_pair(GPROdd3,
560 GPROdd3 + (sizeof(GPROdd3)/sizeof(unsigned)));
562 return std::make_pair(GPROdd6,
563 GPROdd6 + (sizeof(GPROdd6)/sizeof(unsigned)));
566 return std::make_pair(RC->allocation_order_begin(MF),
567 RC->allocation_order_end(MF));
570 /// ResolveRegAllocHint - Resolves the specified register allocation hint
571 /// to a physical register. Returns the physical register if it is successful.
573 ARMBaseRegisterInfo::ResolveRegAllocHint(unsigned Type, unsigned Reg,
574 const MachineFunction &MF) const {
575 if (Reg == 0 || !isPhysicalRegister(Reg))
579 else if (Type == (unsigned)ARMRI::RegPairOdd)
581 return getRegisterPairOdd(Reg, MF);
582 else if (Type == (unsigned)ARMRI::RegPairEven)
584 return getRegisterPairEven(Reg, MF);
589 ARMBaseRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg,
590 MachineFunction &MF) const {
591 MachineRegisterInfo *MRI = &MF.getRegInfo();
592 std::pair<unsigned, unsigned> Hint = MRI->getRegAllocationHint(Reg);
593 if ((Hint.first == (unsigned)ARMRI::RegPairOdd ||
594 Hint.first == (unsigned)ARMRI::RegPairEven) &&
595 Hint.second && TargetRegisterInfo::isVirtualRegister(Hint.second)) {
596 // If 'Reg' is one of the even / odd register pair and it's now changed
597 // (e.g. coalesced) into a different register. The other register of the
598 // pair allocation hint must be updated to reflect the relationship
600 unsigned OtherReg = Hint.second;
601 Hint = MRI->getRegAllocationHint(OtherReg);
602 if (Hint.second == Reg)
603 // Make sure the pair has not already divorced.
604 MRI->setRegAllocationHint(OtherReg, Hint.first, NewReg);
608 /// hasFP - Return true if the specified function should have a dedicated frame
609 /// pointer register. This is true if the function has variable sized allocas
610 /// or if frame pointer elimination is disabled.
612 bool ARMBaseRegisterInfo::hasFP(const MachineFunction &MF) const {
613 const MachineFrameInfo *MFI = MF.getFrameInfo();
614 // Always eliminate non-leaf frame pointers.
615 return ((DisableFramePointerElim(MF) && MFI->hasCalls()) ||
616 needsStackRealignment(MF) ||
617 MFI->hasVarSizedObjects() ||
618 MFI->isFrameAddressTaken());
621 bool ARMBaseRegisterInfo::canRealignStack(const MachineFunction &MF) const {
622 const MachineFrameInfo *MFI = MF.getFrameInfo();
623 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
624 return (RealignStack &&
625 !AFI->isThumb1OnlyFunction() &&
626 !MFI->hasVarSizedObjects());
629 bool ARMBaseRegisterInfo::
630 needsStackRealignment(const MachineFunction &MF) const {
631 const MachineFrameInfo *MFI = MF.getFrameInfo();
632 const Function *F = MF.getFunction();
633 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
634 unsigned StackAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
635 bool requiresRealignment = ((MFI->getMaxAlignment() > StackAlign) ||
636 F->hasFnAttr(Attribute::StackAlignment));
638 // FIXME: Currently we don't support stack realignment for functions with
639 // variable-sized allocas.
640 // FIXME: It's more complicated than this...
641 if (0 && requiresRealignment && MFI->hasVarSizedObjects())
643 "Stack realignment in presense of dynamic allocas is not supported");
645 // FIXME: This probably isn't the right place for this.
646 if (0 && requiresRealignment && AFI->isThumb1OnlyFunction())
648 "Stack realignment in thumb1 functions is not supported");
650 return requiresRealignment && canRealignStack(MF);
653 bool ARMBaseRegisterInfo::
654 cannotEliminateFrame(const MachineFunction &MF) const {
655 const MachineFrameInfo *MFI = MF.getFrameInfo();
656 if (DisableFramePointerElim(MF) && MFI->adjustsStack())
658 return MFI->hasVarSizedObjects() || MFI->isFrameAddressTaken()
659 || needsStackRealignment(MF);
662 /// estimateStackSize - Estimate and return the size of the frame.
663 static unsigned estimateStackSize(MachineFunction &MF) {
664 const MachineFrameInfo *FFI = MF.getFrameInfo();
666 for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) {
667 int FixedOff = -FFI->getObjectOffset(i);
668 if (FixedOff > Offset) Offset = FixedOff;
670 for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) {
671 if (FFI->isDeadObjectIndex(i))
673 Offset += FFI->getObjectSize(i);
674 unsigned Align = FFI->getObjectAlignment(i);
675 // Adjust to alignment boundary
676 Offset = (Offset+Align-1)/Align*Align;
678 return (unsigned)Offset;
681 /// estimateRSStackSizeLimit - Look at each instruction that references stack
682 /// frames and return the stack size limit beyond which some of these
683 /// instructions will require a scratch register during their expansion later.
685 ARMBaseRegisterInfo::estimateRSStackSizeLimit(MachineFunction &MF) const {
686 unsigned Limit = (1 << 12) - 1;
687 for (MachineFunction::iterator BB = MF.begin(),E = MF.end(); BB != E; ++BB) {
688 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
690 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
691 if (!I->getOperand(i).isFI()) continue;
693 // When using ADDri to get the address of a stack object, 255 is the
694 // largest offset guaranteed to fit in the immediate offset.
695 if (I->getOpcode() == ARM::ADDri) {
696 Limit = std::min(Limit, (1U << 8) - 1);
700 // Otherwise check the addressing mode.
701 switch (I->getDesc().TSFlags & ARMII::AddrModeMask) {
702 case ARMII::AddrMode3:
703 case ARMII::AddrModeT2_i8:
704 Limit = std::min(Limit, (1U << 8) - 1);
706 case ARMII::AddrMode5:
707 case ARMII::AddrModeT2_i8s4:
708 Limit = std::min(Limit, ((1U << 8) - 1) * 4);
710 case ARMII::AddrModeT2_i12:
711 if (hasFP(MF)) Limit = std::min(Limit, (1U << 8) - 1);
713 case ARMII::AddrMode6:
714 // Addressing mode 6 (load/store) instructions can't encode an
715 // immediate offset for stack references.
720 break; // At most one FI per instruction
728 static unsigned GetFunctionSizeInBytes(const MachineFunction &MF,
729 const ARMBaseInstrInfo &TII) {
731 for (MachineFunction::const_iterator MBBI = MF.begin(), E = MF.end();
733 const MachineBasicBlock &MBB = *MBBI;
734 for (MachineBasicBlock::const_iterator I = MBB.begin(),E = MBB.end();
736 FnSize += TII.GetInstSizeInBytes(I);
742 ARMBaseRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
743 RegScavenger *RS) const {
744 // This tells PEI to spill the FP as if it is any other callee-save register
745 // to take advantage the eliminateFrameIndex machinery. This also ensures it
746 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
747 // to combine multiple loads / stores.
748 bool CanEliminateFrame = true;
749 bool CS1Spilled = false;
750 bool LRSpilled = false;
751 unsigned NumGPRSpills = 0;
752 SmallVector<unsigned, 4> UnspilledCS1GPRs;
753 SmallVector<unsigned, 4> UnspilledCS2GPRs;
754 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
755 MachineFrameInfo *MFI = MF.getFrameInfo();
757 // Spill R4 if Thumb2 function requires stack realignment - it will be used as
759 // FIXME: It will be better just to find spare register here.
760 if (needsStackRealignment(MF) &&
761 AFI->isThumb2Function())
762 MF.getRegInfo().setPhysRegUsed(ARM::R4);
764 // Spill LR if Thumb1 function uses variable length argument lists.
765 if (AFI->isThumb1OnlyFunction() && AFI->getVarArgsRegSaveSize() > 0)
766 MF.getRegInfo().setPhysRegUsed(ARM::LR);
768 // Don't spill FP if the frame can be eliminated. This is determined
769 // by scanning the callee-save registers to see if any is used.
770 const unsigned *CSRegs = getCalleeSavedRegs();
771 for (unsigned i = 0; CSRegs[i]; ++i) {
772 unsigned Reg = CSRegs[i];
773 bool Spilled = false;
774 if (MF.getRegInfo().isPhysRegUsed(Reg)) {
775 AFI->setCSRegisterIsSpilled(Reg);
777 CanEliminateFrame = false;
779 // Check alias registers too.
780 for (const unsigned *Aliases = getAliasSet(Reg); *Aliases; ++Aliases) {
781 if (MF.getRegInfo().isPhysRegUsed(*Aliases)) {
783 CanEliminateFrame = false;
788 if (!ARM::GPRRegisterClass->contains(Reg))
794 if (!STI.isTargetDarwin()) {
801 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
816 if (!STI.isTargetDarwin()) {
817 UnspilledCS1GPRs.push_back(Reg);
827 UnspilledCS1GPRs.push_back(Reg);
830 UnspilledCS2GPRs.push_back(Reg);
836 bool ForceLRSpill = false;
837 if (!LRSpilled && AFI->isThumb1OnlyFunction()) {
838 unsigned FnSize = GetFunctionSizeInBytes(MF, TII);
839 // Force LR to be spilled if the Thumb function size is > 2048. This enables
840 // use of BL to implement far jump. If it turns out that it's not needed
841 // then the branch fix up path will undo it.
842 if (FnSize >= (1 << 11)) {
843 CanEliminateFrame = false;
848 // If any of the stack slot references may be out of range of an immediate
849 // offset, make sure a register (or a spill slot) is available for the
850 // register scavenger. Note that if we're indexing off the frame pointer, the
851 // effective stack size is 4 bytes larger since the FP points to the stack
852 // slot of the previous FP. Also, if we have variable sized objects in the
853 // function, stack slot references will often be negative, and some of
854 // our instructions are positive-offset only, so conservatively consider
855 // that case to want a spill slot (or register) as well. Similarly, if
856 // the function adjusts the stack pointer during execution and the
857 // adjustments aren't already part of our stack size estimate, our offset
858 // calculations may be off, so be conservative.
859 // FIXME: We could add logic to be more precise about negative offsets
860 // and which instructions will need a scratch register for them. Is it
861 // worth the effort and added fragility?
863 (RS && (estimateStackSize(MF) + (hasFP(MF) ? 4:0) >=
864 estimateRSStackSizeLimit(MF)))
865 || MFI->hasVarSizedObjects()
866 || (MFI->adjustsStack() && !canSimplifyCallFramePseudos(MF));
868 bool ExtraCSSpill = false;
869 if (BigStack || !CanEliminateFrame || cannotEliminateFrame(MF)) {
870 AFI->setHasStackFrame(true);
872 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
873 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
874 if (!LRSpilled && CS1Spilled) {
875 MF.getRegInfo().setPhysRegUsed(ARM::LR);
876 AFI->setCSRegisterIsSpilled(ARM::LR);
878 UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(),
879 UnspilledCS1GPRs.end(), (unsigned)ARM::LR));
880 ForceLRSpill = false;
884 // Darwin ABI requires FP to point to the stack slot that contains the
886 if (STI.isTargetDarwin() || hasFP(MF)) {
887 MF.getRegInfo().setPhysRegUsed(FramePtr);
891 // If stack and double are 8-byte aligned and we are spilling an odd number
892 // of GPRs. Spill one extra callee save GPR so we won't have to pad between
893 // the integer and double callee save areas.
894 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
895 if (TargetAlign == 8 && (NumGPRSpills & 1)) {
896 if (CS1Spilled && !UnspilledCS1GPRs.empty()) {
897 for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) {
898 unsigned Reg = UnspilledCS1GPRs[i];
899 // Don't spill high register if the function is thumb1
900 if (!AFI->isThumb1OnlyFunction() ||
901 isARMLowRegister(Reg) || Reg == ARM::LR) {
902 MF.getRegInfo().setPhysRegUsed(Reg);
903 AFI->setCSRegisterIsSpilled(Reg);
904 if (!isReservedReg(MF, Reg))
909 } else if (!UnspilledCS2GPRs.empty() &&
910 !AFI->isThumb1OnlyFunction()) {
911 unsigned Reg = UnspilledCS2GPRs.front();
912 MF.getRegInfo().setPhysRegUsed(Reg);
913 AFI->setCSRegisterIsSpilled(Reg);
914 if (!isReservedReg(MF, Reg))
919 // Estimate if we might need to scavenge a register at some point in order
920 // to materialize a stack offset. If so, either spill one additional
921 // callee-saved register or reserve a special spill slot to facilitate
922 // register scavenging. Thumb1 needs a spill slot for stack pointer
923 // adjustments also, even when the frame itself is small.
924 if (BigStack && !ExtraCSSpill) {
925 // If any non-reserved CS register isn't spilled, just spill one or two
926 // extra. That should take care of it!
927 unsigned NumExtras = TargetAlign / 4;
928 SmallVector<unsigned, 2> Extras;
929 while (NumExtras && !UnspilledCS1GPRs.empty()) {
930 unsigned Reg = UnspilledCS1GPRs.back();
931 UnspilledCS1GPRs.pop_back();
932 if (!isReservedReg(MF, Reg) &&
933 (!AFI->isThumb1OnlyFunction() || isARMLowRegister(Reg) ||
935 Extras.push_back(Reg);
939 // For non-Thumb1 functions, also check for hi-reg CS registers
940 if (!AFI->isThumb1OnlyFunction()) {
941 while (NumExtras && !UnspilledCS2GPRs.empty()) {
942 unsigned Reg = UnspilledCS2GPRs.back();
943 UnspilledCS2GPRs.pop_back();
944 if (!isReservedReg(MF, Reg)) {
945 Extras.push_back(Reg);
950 if (Extras.size() && NumExtras == 0) {
951 for (unsigned i = 0, e = Extras.size(); i != e; ++i) {
952 MF.getRegInfo().setPhysRegUsed(Extras[i]);
953 AFI->setCSRegisterIsSpilled(Extras[i]);
955 } else if (!AFI->isThumb1OnlyFunction()) {
956 // note: Thumb1 functions spill to R12, not the stack. Reserve a slot
957 // closest to SP or frame pointer.
958 const TargetRegisterClass *RC = ARM::GPRRegisterClass;
959 RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
967 MF.getRegInfo().setPhysRegUsed(ARM::LR);
968 AFI->setCSRegisterIsSpilled(ARM::LR);
969 AFI->setLRIsSpilledForFarJump(true);
973 unsigned ARMBaseRegisterInfo::getRARegister() const {
978 ARMBaseRegisterInfo::getFrameRegister(const MachineFunction &MF) const {
979 if (STI.isTargetDarwin() || hasFP(MF))
984 // Provide a base+offset reference to an FI slot for debug info. It's the
985 // same as what we use for resolving the code-gen references for now.
986 // FIXME: This can go wrong when references are SP-relative and simple call
987 // frames aren't used.
989 ARMBaseRegisterInfo::getFrameIndexReference(const MachineFunction &MF, int FI,
990 unsigned &FrameReg) const {
991 return ResolveFrameIndexReference(MF, FI, FrameReg, 0);
995 ARMBaseRegisterInfo::ResolveFrameIndexReference(const MachineFunction &MF,
999 const MachineFrameInfo *MFI = MF.getFrameInfo();
1000 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1001 int Offset = MFI->getObjectOffset(FI) + MFI->getStackSize();
1002 int FPOffset = Offset - AFI->getFramePtrSpillOffset();
1003 bool isFixed = MFI->isFixedObjectIndex(FI);
1007 if (AFI->isGPRCalleeSavedArea1Frame(FI))
1008 return Offset - AFI->getGPRCalleeSavedArea1Offset();
1009 else if (AFI->isGPRCalleeSavedArea2Frame(FI))
1010 return Offset - AFI->getGPRCalleeSavedArea2Offset();
1011 else if (AFI->isDPRCalleeSavedAreaFrame(FI))
1012 return Offset - AFI->getDPRCalleeSavedAreaOffset();
1014 // When dynamically realigning the stack, use the frame pointer for
1015 // parameters, and the stack pointer for locals.
1016 if (needsStackRealignment(MF)) {
1017 assert (hasFP(MF) && "dynamic stack realignment without a FP!");
1019 FrameReg = getFrameRegister(MF);
1025 // If there is a frame pointer, use it when we can.
1026 if (hasFP(MF) && AFI->hasStackFrame()) {
1027 // Use frame pointer to reference fixed objects. Use it for locals if
1028 // there are VLAs (and thus the SP isn't reliable as a base).
1029 if (isFixed || MFI->hasVarSizedObjects()) {
1030 FrameReg = getFrameRegister(MF);
1032 } else if (AFI->isThumb2Function()) {
1033 // In Thumb2 mode, the negative offset is very limited. Try to avoid
1034 // out of range references.
1035 if (FPOffset >= -255 && FPOffset < 0) {
1036 FrameReg = getFrameRegister(MF);
1039 } else if (Offset > (FPOffset < 0 ? -FPOffset : FPOffset)) {
1040 // Otherwise, use SP or FP, whichever is closer to the stack slot.
1041 FrameReg = getFrameRegister(MF);
1049 ARMBaseRegisterInfo::getFrameIndexOffset(const MachineFunction &MF,
1052 return getFrameIndexReference(MF, FI, FrameReg);
1055 unsigned ARMBaseRegisterInfo::getEHExceptionRegister() const {
1056 llvm_unreachable("What is the exception register");
1060 unsigned ARMBaseRegisterInfo::getEHHandlerRegister() const {
1061 llvm_unreachable("What is the exception handler register");
1065 int ARMBaseRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const {
1066 return ARMGenRegisterInfo::getDwarfRegNumFull(RegNum, 0);
1069 unsigned ARMBaseRegisterInfo::getRegisterPairEven(unsigned Reg,
1070 const MachineFunction &MF) const {
1073 // Return 0 if either register of the pair is a special register.
1082 return isReservedReg(MF, ARM::R7) ? 0 : ARM::R6;
1084 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R8;
1086 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R10;
1158 unsigned ARMBaseRegisterInfo::getRegisterPairOdd(unsigned Reg,
1159 const MachineFunction &MF) const {
1162 // Return 0 if either register of the pair is a special register.
1171 return isReservedReg(MF, ARM::R7) ? 0 : ARM::R7;
1173 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R9;
1175 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R11;
1247 /// emitLoadConstPool - Emits a load from constpool to materialize the
1248 /// specified immediate.
1249 void ARMBaseRegisterInfo::
1250 emitLoadConstPool(MachineBasicBlock &MBB,
1251 MachineBasicBlock::iterator &MBBI,
1253 unsigned DestReg, unsigned SubIdx, int Val,
1254 ARMCC::CondCodes Pred,
1255 unsigned PredReg) const {
1256 MachineFunction &MF = *MBB.getParent();
1257 MachineConstantPool *ConstantPool = MF.getConstantPool();
1259 ConstantInt::get(Type::getInt32Ty(MF.getFunction()->getContext()), Val);
1260 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 4);
1262 BuildMI(MBB, MBBI, dl, TII.get(ARM::LDRcp))
1263 .addReg(DestReg, getDefRegState(true), SubIdx)
1264 .addConstantPoolIndex(Idx)
1265 .addReg(0).addImm(0).addImm(Pred).addReg(PredReg);
1268 bool ARMBaseRegisterInfo::
1269 requiresRegisterScavenging(const MachineFunction &MF) const {
1273 bool ARMBaseRegisterInfo::
1274 requiresFrameIndexScavenging(const MachineFunction &MF) const {
1278 // hasReservedCallFrame - Under normal circumstances, when a frame pointer is
1279 // not required, we reserve argument space for call sites in the function
1280 // immediately on entry to the current function. This eliminates the need for
1281 // add/sub sp brackets around call sites. Returns true if the call frame is
1282 // included as part of the stack frame.
1283 bool ARMBaseRegisterInfo::
1284 hasReservedCallFrame(const MachineFunction &MF) const {
1285 const MachineFrameInfo *FFI = MF.getFrameInfo();
1286 unsigned CFSize = FFI->getMaxCallFrameSize();
1287 // It's not always a good idea to include the call frame as part of the
1288 // stack frame. ARM (especially Thumb) has small immediate offset to
1289 // address the stack frame. So a large call frame can cause poor codegen
1290 // and may even makes it impossible to scavenge a register.
1291 if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12
1294 return !MF.getFrameInfo()->hasVarSizedObjects();
1297 // canSimplifyCallFramePseudos - If there is a reserved call frame, the
1298 // call frame pseudos can be simplified. Unlike most targets, having a FP
1299 // is not sufficient here since we still may reference some objects via SP
1300 // even when FP is available in Thumb2 mode.
1301 bool ARMBaseRegisterInfo::
1302 canSimplifyCallFramePseudos(const MachineFunction &MF) const {
1303 return hasReservedCallFrame(MF) || MF.getFrameInfo()->hasVarSizedObjects();
1307 emitSPUpdate(bool isARM,
1308 MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
1309 DebugLoc dl, const ARMBaseInstrInfo &TII,
1311 ARMCC::CondCodes Pred = ARMCC::AL, unsigned PredReg = 0) {
1313 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, ARM::SP, NumBytes,
1314 Pred, PredReg, TII);
1316 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::SP, ARM::SP, NumBytes,
1317 Pred, PredReg, TII);
1321 void ARMBaseRegisterInfo::
1322 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
1323 MachineBasicBlock::iterator I) const {
1324 if (!hasReservedCallFrame(MF)) {
1325 // If we have alloca, convert as follows:
1326 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
1327 // ADJCALLSTACKUP -> add, sp, sp, amount
1328 MachineInstr *Old = I;
1329 DebugLoc dl = Old->getDebugLoc();
1330 unsigned Amount = Old->getOperand(0).getImm();
1332 // We need to keep the stack aligned properly. To do this, we round the
1333 // amount of space needed for the outgoing arguments up to the next
1334 // alignment boundary.
1335 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1336 Amount = (Amount+Align-1)/Align*Align;
1338 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1339 assert(!AFI->isThumb1OnlyFunction() &&
1340 "This eliminateCallFramePseudoInstr does not support Thumb1!");
1341 bool isARM = !AFI->isThumbFunction();
1343 // Replace the pseudo instruction with a new instruction...
1344 unsigned Opc = Old->getOpcode();
1345 int PIdx = Old->findFirstPredOperandIdx();
1346 ARMCC::CondCodes Pred = (PIdx == -1)
1347 ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(PIdx).getImm();
1348 if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) {
1349 // Note: PredReg is operand 2 for ADJCALLSTACKDOWN.
1350 unsigned PredReg = Old->getOperand(2).getReg();
1351 emitSPUpdate(isARM, MBB, I, dl, TII, -Amount, Pred, PredReg);
1353 // Note: PredReg is operand 3 for ADJCALLSTACKUP.
1354 unsigned PredReg = Old->getOperand(3).getReg();
1355 assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP);
1356 emitSPUpdate(isARM, MBB, I, dl, TII, Amount, Pred, PredReg);
1364 ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
1365 int SPAdj, FrameIndexValue *Value,
1366 RegScavenger *RS) const {
1368 MachineInstr &MI = *II;
1369 MachineBasicBlock &MBB = *MI.getParent();
1370 MachineFunction &MF = *MBB.getParent();
1371 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1372 assert(!AFI->isThumb1OnlyFunction() &&
1373 "This eliminateFrameIndex does not support Thumb1!");
1375 while (!MI.getOperand(i).isFI()) {
1377 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
1380 int FrameIndex = MI.getOperand(i).getIndex();
1383 int Offset = ResolveFrameIndexReference(MF, FrameIndex, FrameReg, SPAdj);
1385 // Special handling of dbg_value instructions.
1386 if (MI.isDebugValue()) {
1387 MI.getOperand(i). ChangeToRegister(FrameReg, false /*isDef*/);
1388 MI.getOperand(i+1).ChangeToImmediate(Offset);
1392 // Modify MI as necessary to handle as much of 'Offset' as possible
1394 if (!AFI->isThumbFunction())
1395 Done = rewriteARMFrameIndex(MI, i, FrameReg, Offset, TII);
1397 assert(AFI->isThumb2Function());
1398 Done = rewriteT2FrameIndex(MI, i, FrameReg, Offset, TII);
1403 // If we get here, the immediate doesn't fit into the instruction. We folded
1404 // as much as possible above, handle the rest, providing a register that is
1407 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode4 ||
1408 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode6) &&
1409 "This code isn't needed if offset already handled!");
1411 unsigned ScratchReg = 0;
1412 int PIdx = MI.findFirstPredOperandIdx();
1413 ARMCC::CondCodes Pred = (PIdx == -1)
1414 ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm();
1415 unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg();
1417 // Must be addrmode4/6.
1418 MI.getOperand(i).ChangeToRegister(FrameReg, false, false, false);
1420 ScratchReg = MF.getRegInfo().createVirtualRegister(ARM::GPRRegisterClass);
1422 Value->first = FrameReg; // use the frame register as a kind indicator
1423 Value->second = Offset;
1425 if (!AFI->isThumbFunction())
1426 emitARMRegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg,
1427 Offset, Pred, PredReg, TII);
1429 assert(AFI->isThumb2Function());
1430 emitT2RegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg,
1431 Offset, Pred, PredReg, TII);
1433 MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true);
1434 if (!ReuseFrameIndexVals)
1440 /// Move iterator past the next bunch of callee save load / store ops for
1441 /// the particular spill area (1: integer area 1, 2: integer area 2,
1442 /// 3: fp area, 0: don't care).
1443 static void movePastCSLoadStoreOps(MachineBasicBlock &MBB,
1444 MachineBasicBlock::iterator &MBBI,
1445 int Opc1, int Opc2, unsigned Area,
1446 const ARMSubtarget &STI) {
1447 while (MBBI != MBB.end() &&
1448 ((MBBI->getOpcode() == Opc1) || (MBBI->getOpcode() == Opc2)) &&
1449 MBBI->getOperand(1).isFI()) {
1452 unsigned Category = 0;
1453 switch (MBBI->getOperand(0).getReg()) {
1454 case ARM::R4: case ARM::R5: case ARM::R6: case ARM::R7:
1458 case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11:
1459 Category = STI.isTargetDarwin() ? 2 : 1;
1461 case ARM::D8: case ARM::D9: case ARM::D10: case ARM::D11:
1462 case ARM::D12: case ARM::D13: case ARM::D14: case ARM::D15:
1469 if (Done || Category != Area)
1477 void ARMBaseRegisterInfo::
1478 emitPrologue(MachineFunction &MF) const {
1479 MachineBasicBlock &MBB = MF.front();
1480 MachineBasicBlock::iterator MBBI = MBB.begin();
1481 MachineFrameInfo *MFI = MF.getFrameInfo();
1482 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1483 assert(!AFI->isThumb1OnlyFunction() &&
1484 "This emitPrologue does not support Thumb1!");
1485 bool isARM = !AFI->isThumbFunction();
1486 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1487 unsigned NumBytes = MFI->getStackSize();
1488 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
1489 DebugLoc dl = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
1491 // Determine the sizes of each callee-save spill areas and record which frame
1492 // belongs to which callee-save spill areas.
1493 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
1494 int FramePtrSpillFI = 0;
1496 // Allocate the vararg register save area. This is not counted in NumBytes.
1498 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -VARegSaveSize);
1500 if (!AFI->hasStackFrame()) {
1502 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes);
1506 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1507 unsigned Reg = CSI[i].getReg();
1508 int FI = CSI[i].getFrameIdx();
1515 if (Reg == FramePtr)
1516 FramePtrSpillFI = FI;
1517 AFI->addGPRCalleeSavedArea1Frame(FI);
1524 if (Reg == FramePtr)
1525 FramePtrSpillFI = FI;
1526 if (STI.isTargetDarwin()) {
1527 AFI->addGPRCalleeSavedArea2Frame(FI);
1530 AFI->addGPRCalleeSavedArea1Frame(FI);
1535 AFI->addDPRCalleeSavedAreaFrame(FI);
1540 // Build the new SUBri to adjust SP for integer callee-save spill area 1.
1541 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -GPRCS1Size);
1542 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, ARM::t2STRi12, 1, STI);
1544 // Set FP to point to the stack slot that contains the previous FP.
1545 // For Darwin, FP is R7, which has now been stored in spill area 1.
1546 // Otherwise, if this is not Darwin, all the callee-saved registers go
1547 // into spill area 1, including the FP in R11. In either case, it is
1548 // now safe to emit this assignment.
1549 if (STI.isTargetDarwin() || hasFP(MF)) {
1550 unsigned ADDriOpc = !AFI->isThumbFunction() ? ARM::ADDri : ARM::t2ADDri;
1551 MachineInstrBuilder MIB =
1552 BuildMI(MBB, MBBI, dl, TII.get(ADDriOpc), FramePtr)
1553 .addFrameIndex(FramePtrSpillFI).addImm(0);
1554 AddDefaultCC(AddDefaultPred(MIB));
1557 // Build the new SUBri to adjust SP for integer callee-save spill area 2.
1558 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -GPRCS2Size);
1560 // Build the new SUBri to adjust SP for FP callee-save spill area.
1561 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, ARM::t2STRi12, 2, STI);
1562 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -DPRCSSize);
1564 // Determine starting offsets of spill areas.
1565 unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize);
1566 unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize;
1567 unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size;
1568 if (STI.isTargetDarwin() || hasFP(MF))
1569 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) +
1571 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
1572 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
1573 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
1575 movePastCSLoadStoreOps(MBB, MBBI, ARM::VSTRD, 0, 3, STI);
1576 NumBytes = DPRCSOffset;
1578 // Adjust SP after all the callee-save spills.
1579 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes);
1582 if (STI.isTargetELF() && hasFP(MF)) {
1583 MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() -
1584 AFI->getFramePtrSpillOffset());
1587 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
1588 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
1589 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
1591 // If we need dynamic stack realignment, do it here.
1592 if (needsStackRealignment(MF)) {
1593 unsigned MaxAlign = MFI->getMaxAlignment();
1594 assert (!AFI->isThumb1OnlyFunction());
1595 if (!AFI->isThumbFunction()) {
1596 // Emit bic sp, sp, MaxAlign
1597 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl,
1598 TII.get(ARM::BICri), ARM::SP)
1599 .addReg(ARM::SP, RegState::Kill)
1600 .addImm(MaxAlign-1)));
1602 // We cannot use sp as source/dest register here, thus we're emitting the
1603 // following sequence:
1605 // bic r4, r4, MaxAlign
1607 // FIXME: It will be better just to find spare register here.
1608 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVgpr2tgpr), ARM::R4)
1609 .addReg(ARM::SP, RegState::Kill);
1610 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl,
1611 TII.get(ARM::t2BICri), ARM::R4)
1612 .addReg(ARM::R4, RegState::Kill)
1613 .addImm(MaxAlign-1)));
1614 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVtgpr2gpr), ARM::SP)
1615 .addReg(ARM::R4, RegState::Kill);
1620 static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) {
1621 for (unsigned i = 0; CSRegs[i]; ++i)
1622 if (Reg == CSRegs[i])
1627 static bool isCSRestore(MachineInstr *MI,
1628 const ARMBaseInstrInfo &TII,
1629 const unsigned *CSRegs) {
1630 return ((MI->getOpcode() == (int)ARM::VLDRD ||
1631 MI->getOpcode() == (int)ARM::LDR ||
1632 MI->getOpcode() == (int)ARM::t2LDRi12) &&
1633 MI->getOperand(1).isFI() &&
1634 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs));
1637 void ARMBaseRegisterInfo::
1638 emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const {
1639 MachineBasicBlock::iterator MBBI = prior(MBB.end());
1640 assert(MBBI->getDesc().isReturn() &&
1641 "Can only insert epilog into returning blocks");
1642 unsigned RetOpcode = MBBI->getOpcode();
1643 DebugLoc dl = MBBI->getDebugLoc();
1644 MachineFrameInfo *MFI = MF.getFrameInfo();
1645 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1646 assert(!AFI->isThumb1OnlyFunction() &&
1647 "This emitEpilogue does not support Thumb1!");
1648 bool isARM = !AFI->isThumbFunction();
1650 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1651 int NumBytes = (int)MFI->getStackSize();
1653 if (!AFI->hasStackFrame()) {
1655 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes);
1657 // Unwind MBBI to point to first LDR / VLDRD.
1658 const unsigned *CSRegs = getCalleeSavedRegs();
1659 if (MBBI != MBB.begin()) {
1662 while (MBBI != MBB.begin() && isCSRestore(MBBI, TII, CSRegs));
1663 if (!isCSRestore(MBBI, TII, CSRegs))
1667 // Move SP to start of FP callee save spill area.
1668 NumBytes -= (AFI->getGPRCalleeSavedArea1Size() +
1669 AFI->getGPRCalleeSavedArea2Size() +
1670 AFI->getDPRCalleeSavedAreaSize());
1672 // Darwin ABI requires FP to point to the stack slot that contains the
1674 bool HasFP = hasFP(MF);
1675 if ((STI.isTargetDarwin() && NumBytes) || HasFP) {
1676 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1677 // Reset SP based on frame pointer only if the stack frame extends beyond
1678 // frame pointer stack slot or target is ELF and the function has FP.
1680 AFI->getGPRCalleeSavedArea2Size() ||
1681 AFI->getDPRCalleeSavedAreaSize() ||
1682 AFI->getDPRCalleeSavedAreaOffset()) {
1685 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, FramePtr, -NumBytes,
1688 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::SP, FramePtr, -NumBytes,
1693 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), ARM::SP)
1695 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
1697 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVgpr2gpr), ARM::SP)
1701 } else if (NumBytes)
1702 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes);
1704 // Move SP to start of integer callee save spill area 2.
1705 movePastCSLoadStoreOps(MBB, MBBI, ARM::VLDRD, 0, 3, STI);
1706 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getDPRCalleeSavedAreaSize());
1708 // Move SP to start of integer callee save spill area 1.
1709 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, ARM::t2LDRi12, 2, STI);
1710 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getGPRCalleeSavedArea2Size());
1712 // Move SP to SP upon entry to the function.
1713 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, ARM::t2LDRi12, 1, STI);
1714 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getGPRCalleeSavedArea1Size());
1717 if (RetOpcode == ARM::TCRETURNdi || RetOpcode == ARM::TCRETURNdiND ||
1718 RetOpcode == ARM::TCRETURNri || RetOpcode == ARM::TCRETURNriND) {
1719 // Tail call return: adjust the stack pointer and jump to callee.
1720 MBBI = prior(MBB.end());
1721 MachineOperand &JumpTarget = MBBI->getOperand(0);
1723 // Jump to label or value in register.
1724 if (RetOpcode == ARM::TCRETURNdi) {
1725 BuildMI(MBB, MBBI, dl,
1726 TII.get(STI.isThumb() ? ARM::TAILJMPdt : ARM::TAILJMPd)).
1727 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(),
1728 JumpTarget.getTargetFlags());
1729 } else if (RetOpcode == ARM::TCRETURNdiND) {
1730 BuildMI(MBB, MBBI, dl,
1731 TII.get(STI.isThumb() ? ARM::TAILJMPdNDt : ARM::TAILJMPdND)).
1732 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(),
1733 JumpTarget.getTargetFlags());
1734 } else if (RetOpcode == ARM::TCRETURNri) {
1735 BuildMI(MBB, MBBI, dl, TII.get(ARM::TAILJMPr)).
1736 addReg(JumpTarget.getReg(), RegState::Kill);
1737 } else if (RetOpcode == ARM::TCRETURNriND) {
1738 BuildMI(MBB, MBBI, dl, TII.get(ARM::TAILJMPrND)).
1739 addReg(JumpTarget.getReg(), RegState::Kill);
1742 MachineInstr *NewMI = prior(MBBI);
1743 for (unsigned i = 1, e = MBBI->getNumOperands(); i != e; ++i)
1744 NewMI->addOperand(MBBI->getOperand(i));
1746 // Delete the pseudo instruction TCRETURN.
1751 emitSPUpdate(isARM, MBB, MBBI, dl, TII, VARegSaveSize);
1754 #include "ARMGenRegisterInfo.inc"