1 //===- ARMRegisterInfo.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 ARM implementation of the TargetRegisterInfo class.
12 //===----------------------------------------------------------------------===//
15 #include "ARMAddressingModes.h"
16 #include "ARMInstrInfo.h"
17 #include "ARMMachineFunctionInfo.h"
18 #include "ARMRegisterInfo.h"
19 #include "ARMSubtarget.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/CodeGen/MachineConstantPool.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineLocation.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/RegisterScavenging.h"
29 #include "llvm/Target/TargetFrameInfo.h"
30 #include "llvm/Target/TargetMachine.h"
31 #include "llvm/Target/TargetOptions.h"
32 #include "llvm/ADT/BitVector.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/STLExtras.h"
35 #include "llvm/Support/CommandLine.h"
39 static cl::opt<bool> ThumbRegScavenging("enable-thumb-reg-scavenging",
41 cl::desc("Enable register scavenging on Thumb"));
43 unsigned ARMRegisterInfo::getRegisterNumbering(unsigned RegEnum) {
46 case R0: case S0: case D0: return 0;
47 case R1: case S1: case D1: return 1;
48 case R2: case S2: case D2: return 2;
49 case R3: case S3: case D3: return 3;
50 case R4: case S4: case D4: return 4;
51 case R5: case S5: case D5: return 5;
52 case R6: case S6: case D6: return 6;
53 case R7: case S7: case D7: return 7;
54 case R8: case S8: case D8: return 8;
55 case R9: case S9: case D9: return 9;
56 case R10: case S10: case D10: return 10;
57 case R11: case S11: case D11: return 11;
58 case R12: case S12: case D12: return 12;
59 case SP: case S13: case D13: return 13;
60 case LR: case S14: case D14: return 14;
61 case PC: case S15: case D15: return 15;
79 assert(0 && "Unknown ARM register!");
84 unsigned ARMRegisterInfo::getRegisterNumbering(unsigned RegEnum,
91 assert(0 && "Unknown ARM register!");
93 case R0: case D0: return 0;
94 case R1: case D1: return 1;
95 case R2: case D2: return 2;
96 case R3: case D3: return 3;
97 case R4: case D4: return 4;
98 case R5: case D5: return 5;
99 case R6: case D6: return 6;
100 case R7: case D7: return 7;
101 case R8: case D8: return 8;
102 case R9: case D9: return 9;
103 case R10: case D10: return 10;
104 case R11: case D11: return 11;
105 case R12: case D12: return 12;
106 case SP: case D13: return 13;
107 case LR: case D14: return 14;
108 case PC: case D15: return 15;
110 case S0: case S1: case S2: case S3:
111 case S4: case S5: case S6: case S7:
112 case S8: case S9: case S10: case S11:
113 case S12: case S13: case S14: case S15:
114 case S16: case S17: case S18: case S19:
115 case S20: case S21: case S22: case S23:
116 case S24: case S25: case S26: case S27:
117 case S28: case S29: case S30: case S31: {
120 default: return 0; // Avoid compile time warning.
158 ARMRegisterInfo::ARMRegisterInfo(const TargetInstrInfo &tii,
159 const ARMSubtarget &sti)
160 : ARMGenRegisterInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
162 FramePtr((STI.isTargetDarwin() || STI.isThumb()) ? ARM::R7 : ARM::R11) {
166 const MachineInstrBuilder &AddDefaultPred(const MachineInstrBuilder &MIB) {
167 return MIB.addImm((int64_t)ARMCC::AL).addReg(0);
171 const MachineInstrBuilder &AddDefaultCC(const MachineInstrBuilder &MIB) {
172 return MIB.addReg(0);
175 /// emitLoadConstPool - Emits a load from constpool to materialize the
176 /// specified immediate.
177 void ARMRegisterInfo::emitLoadConstPool(MachineBasicBlock &MBB,
178 MachineBasicBlock::iterator &MBBI,
179 unsigned DestReg, int Val,
180 unsigned Pred, unsigned PredReg,
181 const TargetInstrInfo *TII,
184 MachineFunction &MF = *MBB.getParent();
185 MachineConstantPool *ConstantPool = MF.getConstantPool();
186 Constant *C = ConstantInt::get(Type::Int32Ty, Val);
187 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 4);
189 BuildMI(MBB, MBBI, dl,
190 TII->get(ARM::tLDRcp),DestReg).addConstantPoolIndex(Idx);
192 BuildMI(MBB, MBBI, dl, TII->get(ARM::LDRcp), DestReg)
193 .addConstantPoolIndex(Idx)
194 .addReg(0).addImm(0).addImm(Pred).addReg(PredReg);
197 /// isLowRegister - Returns true if the register is low register r0-r7.
199 bool ARMRegisterInfo::isLowRegister(unsigned Reg) const {
202 case R0: case R1: case R2: case R3:
203 case R4: case R5: case R6: case R7:
210 const TargetRegisterClass*
211 ARMRegisterInfo::getPhysicalRegisterRegClass(unsigned Reg, MVT VT) const {
213 if (isLowRegister(Reg))
214 return ARM::tGPRRegisterClass;
218 case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11:
219 case ARM::R12: case ARM::SP: case ARM::LR: case ARM::PC:
220 return ARM::GPRRegisterClass;
223 return TargetRegisterInfo::getPhysicalRegisterRegClass(Reg, VT);
227 ARMRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
228 static const unsigned CalleeSavedRegs[] = {
229 ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8,
230 ARM::R7, ARM::R6, ARM::R5, ARM::R4,
232 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
233 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
237 static const unsigned DarwinCalleeSavedRegs[] = {
238 ARM::LR, ARM::R7, ARM::R6, ARM::R5, ARM::R4,
239 ARM::R11, ARM::R10, ARM::R9, ARM::R8,
241 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
242 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
245 return STI.isTargetDarwin() ? DarwinCalleeSavedRegs : CalleeSavedRegs;
248 const TargetRegisterClass* const *
249 ARMRegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const {
250 static const TargetRegisterClass * const CalleeSavedRegClasses[] = {
251 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
252 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
253 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
255 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
256 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
259 static const TargetRegisterClass * const ThumbCalleeSavedRegClasses[] = {
260 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
261 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::tGPRRegClass,
262 &ARM::tGPRRegClass,&ARM::tGPRRegClass,&ARM::tGPRRegClass,
264 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
265 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
268 return STI.isThumb() ? ThumbCalleeSavedRegClasses : CalleeSavedRegClasses;
271 BitVector ARMRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
272 // FIXME: avoid re-calculating this everytime.
273 BitVector Reserved(getNumRegs());
274 Reserved.set(ARM::SP);
275 Reserved.set(ARM::PC);
276 if (STI.isTargetDarwin() || hasFP(MF))
277 Reserved.set(FramePtr);
278 // Some targets reserve R9.
279 if (STI.isR9Reserved())
280 Reserved.set(ARM::R9);
285 ARMRegisterInfo::isReservedReg(const MachineFunction &MF, unsigned Reg) const {
293 if (FramePtr == Reg && (STI.isTargetDarwin() || hasFP(MF)))
297 return STI.isR9Reserved();
303 const TargetRegisterClass *ARMRegisterInfo::getPointerRegClass() const {
304 return &ARM::GPRRegClass;
307 /// getAllocationOrder - Returns the register allocation order for a specified
308 /// register class in the form of a pair of TargetRegisterClass iterators.
309 std::pair<TargetRegisterClass::iterator,TargetRegisterClass::iterator>
310 ARMRegisterInfo::getAllocationOrder(const TargetRegisterClass *RC,
311 unsigned HintType, unsigned HintReg,
312 const MachineFunction &MF) const {
313 // Alternative register allocation orders when favoring even / odd registers
314 // of register pairs.
316 // No FP, R9 is available.
317 static const unsigned GPREven1[] = {
318 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8, ARM::R10,
319 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7,
322 static const unsigned GPROdd1[] = {
323 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R9, ARM::R11,
324 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6,
328 // FP is R7, R9 is available.
329 static const unsigned GPREven2[] = {
330 ARM::R0, ARM::R2, ARM::R4, ARM::R8, ARM::R10,
331 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6,
334 static const unsigned GPROdd2[] = {
335 ARM::R1, ARM::R3, ARM::R5, ARM::R9, ARM::R11,
336 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6,
340 // FP is R11, R9 is available.
341 static const unsigned GPREven3[] = {
342 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8,
343 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7,
346 static const unsigned GPROdd3[] = {
347 ARM::R1, ARM::R3, ARM::R5, ARM::R6, ARM::R9,
348 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R7,
352 // No FP, R9 is not available.
353 static const unsigned GPREven4[] = {
354 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R10,
355 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8,
358 static const unsigned GPROdd4[] = {
359 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R11,
360 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8,
364 // FP is R7, R9 is not available.
365 static const unsigned GPREven5[] = {
366 ARM::R0, ARM::R2, ARM::R4, ARM::R10,
367 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6, ARM::R8,
370 static const unsigned GPROdd5[] = {
371 ARM::R1, ARM::R3, ARM::R5, ARM::R11,
372 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8,
376 // FP is R11, R9 is not available.
377 static const unsigned GPREven6[] = {
378 ARM::R0, ARM::R2, ARM::R4, ARM::R6,
379 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8
381 static const unsigned GPROdd6[] = {
382 ARM::R1, ARM::R3, ARM::R5, ARM::R7,
383 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8
387 if (HintType == ARMRI::RegPairEven) {
388 if (isPhysicalRegister(HintReg) && getRegisterPairEven(HintReg, MF) == 0)
389 // It's no longer possible to fulfill this hint. Return the default
391 return std::make_pair(RC->allocation_order_begin(MF),
392 RC->allocation_order_end(MF));
394 if (!STI.isTargetDarwin() && !hasFP(MF)) {
395 if (!STI.isR9Reserved())
396 return std::make_pair(GPREven1,
397 GPREven1 + (sizeof(GPREven1)/sizeof(unsigned)));
399 return std::make_pair(GPREven4,
400 GPREven4 + (sizeof(GPREven4)/sizeof(unsigned)));
401 } else if (FramePtr == ARM::R7) {
402 if (!STI.isR9Reserved())
403 return std::make_pair(GPREven2,
404 GPREven2 + (sizeof(GPREven2)/sizeof(unsigned)));
406 return std::make_pair(GPREven5,
407 GPREven5 + (sizeof(GPREven5)/sizeof(unsigned)));
408 } else { // FramePtr == ARM::R11
409 if (!STI.isR9Reserved())
410 return std::make_pair(GPREven3,
411 GPREven3 + (sizeof(GPREven3)/sizeof(unsigned)));
413 return std::make_pair(GPREven6,
414 GPREven6 + (sizeof(GPREven6)/sizeof(unsigned)));
416 } else if (HintType == ARMRI::RegPairOdd) {
417 if (isPhysicalRegister(HintReg) && getRegisterPairOdd(HintReg, MF) == 0)
418 // It's no longer possible to fulfill this hint. Return the default
420 return std::make_pair(RC->allocation_order_begin(MF),
421 RC->allocation_order_end(MF));
423 if (!STI.isTargetDarwin() && !hasFP(MF)) {
424 if (!STI.isR9Reserved())
425 return std::make_pair(GPROdd1,
426 GPROdd1 + (sizeof(GPROdd1)/sizeof(unsigned)));
428 return std::make_pair(GPROdd4,
429 GPROdd4 + (sizeof(GPROdd4)/sizeof(unsigned)));
430 } else if (FramePtr == ARM::R7) {
431 if (!STI.isR9Reserved())
432 return std::make_pair(GPROdd2,
433 GPROdd2 + (sizeof(GPROdd2)/sizeof(unsigned)));
435 return std::make_pair(GPROdd5,
436 GPROdd5 + (sizeof(GPROdd5)/sizeof(unsigned)));
437 } else { // FramePtr == ARM::R11
438 if (!STI.isR9Reserved())
439 return std::make_pair(GPROdd3,
440 GPROdd3 + (sizeof(GPROdd3)/sizeof(unsigned)));
442 return std::make_pair(GPROdd6,
443 GPROdd6 + (sizeof(GPROdd6)/sizeof(unsigned)));
446 return std::make_pair(RC->allocation_order_begin(MF),
447 RC->allocation_order_end(MF));
450 /// ResolveRegAllocHint - Resolves the specified register allocation hint
451 /// to a physical register. Returns the physical register if it is successful.
453 ARMRegisterInfo::ResolveRegAllocHint(unsigned Type, unsigned Reg,
454 const MachineFunction &MF) const {
455 if (Reg == 0 || !isPhysicalRegister(Reg))
459 else if (Type == (unsigned)ARMRI::RegPairOdd)
461 return getRegisterPairOdd(Reg, MF);
462 else if (Type == (unsigned)ARMRI::RegPairEven)
464 return getRegisterPairEven(Reg, MF);
469 ARMRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg,
470 MachineFunction &MF) const {
471 MachineRegisterInfo *MRI = &MF.getRegInfo();
472 std::pair<unsigned, unsigned> Hint = MRI->getRegAllocationHint(Reg);
473 if ((Hint.first == (unsigned)ARMRI::RegPairOdd ||
474 Hint.first == (unsigned)ARMRI::RegPairEven) &&
475 Hint.second && TargetRegisterInfo::isVirtualRegister(Hint.second)) {
476 // If 'Reg' is one of the even / odd register pair and it's now changed
477 // (e.g. coalesced) into a different register. The other register of the
478 // pair allocation hint must be updated to reflect the relationship
480 unsigned OtherReg = Hint.second;
481 Hint = MRI->getRegAllocationHint(OtherReg);
482 if (Hint.second == Reg)
483 // Make sure the pair has not already divorced.
484 MRI->setRegAllocationHint(OtherReg, Hint.first, NewReg);
489 ARMRegisterInfo::requiresRegisterScavenging(const MachineFunction &MF) const {
490 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
491 return ThumbRegScavenging || !AFI->isThumbFunction();
494 /// hasFP - Return true if the specified function should have a dedicated frame
495 /// pointer register. This is true if the function has variable sized allocas
496 /// or if frame pointer elimination is disabled.
498 bool ARMRegisterInfo::hasFP(const MachineFunction &MF) const {
499 const MachineFrameInfo *MFI = MF.getFrameInfo();
500 return (NoFramePointerElim ||
501 MFI->hasVarSizedObjects() ||
502 MFI->isFrameAddressTaken());
505 // hasReservedCallFrame - Under normal circumstances, when a frame pointer is
506 // not required, we reserve argument space for call sites in the function
507 // immediately on entry to the current function. This eliminates the need for
508 // add/sub sp brackets around call sites. Returns true if the call frame is
509 // included as part of the stack frame.
510 bool ARMRegisterInfo::hasReservedCallFrame(MachineFunction &MF) const {
511 const MachineFrameInfo *FFI = MF.getFrameInfo();
512 unsigned CFSize = FFI->getMaxCallFrameSize();
513 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
514 // It's not always a good idea to include the call frame as part of the
515 // stack frame. ARM (especially Thumb) has small immediate offset to
516 // address the stack frame. So a large call frame can cause poor codegen
517 // and may even makes it impossible to scavenge a register.
518 if (AFI->isThumbFunction()) {
519 if (CFSize >= ((1 << 8) - 1) * 4 / 2) // Half of imm8 * 4
522 if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12
525 return !MF.getFrameInfo()->hasVarSizedObjects();
528 /// emitARMRegPlusImmediate - Emits a series of instructions to materialize
529 /// a destreg = basereg + immediate in ARM code.
531 void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
532 MachineBasicBlock::iterator &MBBI,
533 unsigned DestReg, unsigned BaseReg, int NumBytes,
534 ARMCC::CondCodes Pred, unsigned PredReg,
535 const TargetInstrInfo &TII,
537 bool isSub = NumBytes < 0;
538 if (isSub) NumBytes = -NumBytes;
541 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
542 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
543 assert(ThisVal && "Didn't extract field correctly");
545 // We will handle these bits from offset, clear them.
546 NumBytes &= ~ThisVal;
548 // Get the properly encoded SOImmVal field.
549 int SOImmVal = ARM_AM::getSOImmVal(ThisVal);
550 assert(SOImmVal != -1 && "Bit extraction didn't work?");
552 // Build the new ADD / SUB.
553 BuildMI(MBB, MBBI, dl, TII.get(isSub ? ARM::SUBri : ARM::ADDri), DestReg)
554 .addReg(BaseReg, RegState::Kill).addImm(SOImmVal)
555 .addImm((unsigned)Pred).addReg(PredReg).addReg(0);
560 /// calcNumMI - Returns the number of instructions required to materialize
561 /// the specific add / sub r, c instruction.
562 static unsigned calcNumMI(int Opc, int ExtraOpc, unsigned Bytes,
563 unsigned NumBits, unsigned Scale) {
565 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
567 if (Opc == ARM::tADDrSPi) {
568 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
572 Scale = 1; // Followed by a number of tADDi8.
573 Chunk = ((1 << NumBits) - 1) * Scale;
576 NumMIs += Bytes / Chunk;
577 if ((Bytes % Chunk) != 0)
584 /// emitThumbRegPlusImmInReg - Emits a series of instructions to materialize
585 /// a destreg = basereg + immediate in Thumb code. Materialize the immediate
586 /// in a register using mov / mvn sequences or load the immediate from a
589 void emitThumbRegPlusImmInReg(MachineBasicBlock &MBB,
590 MachineBasicBlock::iterator &MBBI,
591 unsigned DestReg, unsigned BaseReg,
592 int NumBytes, bool CanChangeCC,
593 const TargetInstrInfo &TII,
594 const ARMRegisterInfo& MRI,
596 bool isHigh = !MRI.isLowRegister(DestReg) ||
597 (BaseReg != 0 && !MRI.isLowRegister(BaseReg));
599 // Subtract doesn't have high register version. Load the negative value
600 // if either base or dest register is a high register. Also, if do not
601 // issue sub as part of the sequence if condition register is to be
603 if (NumBytes < 0 && !isHigh && CanChangeCC) {
605 NumBytes = -NumBytes;
607 unsigned LdReg = DestReg;
608 if (DestReg == ARM::SP) {
609 assert(BaseReg == ARM::SP && "Unexpected!");
611 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVlor2hir), ARM::R12)
612 .addReg(ARM::R3, RegState::Kill);
615 if (NumBytes <= 255 && NumBytes >= 0)
616 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes);
617 else if (NumBytes < 0 && NumBytes >= -255) {
618 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes);
619 BuildMI(MBB, MBBI, dl, TII.get(ARM::tNEG), LdReg)
620 .addReg(LdReg, RegState::Kill);
622 MRI.emitLoadConstPool(MBB, MBBI, LdReg, NumBytes, ARMCC::AL, 0, &TII,
626 int Opc = (isSub) ? ARM::tSUBrr : (isHigh ? ARM::tADDhirr : ARM::tADDrr);
627 const MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl,
628 TII.get(Opc), DestReg);
629 if (DestReg == ARM::SP || isSub)
630 MIB.addReg(BaseReg).addReg(LdReg, RegState::Kill);
632 MIB.addReg(LdReg).addReg(BaseReg, RegState::Kill);
633 if (DestReg == ARM::SP)
634 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVhir2lor), ARM::R3)
635 .addReg(ARM::R12, RegState::Kill);
638 /// emitThumbRegPlusImmediate - Emits a series of instructions to materialize
639 /// a destreg = basereg + immediate in Thumb code.
641 void emitThumbRegPlusImmediate(MachineBasicBlock &MBB,
642 MachineBasicBlock::iterator &MBBI,
643 unsigned DestReg, unsigned BaseReg,
644 int NumBytes, const TargetInstrInfo &TII,
645 const ARMRegisterInfo& MRI,
647 bool isSub = NumBytes < 0;
648 unsigned Bytes = (unsigned)NumBytes;
649 if (isSub) Bytes = -NumBytes;
650 bool isMul4 = (Bytes & 3) == 0;
651 bool isTwoAddr = false;
652 bool DstNotEqBase = false;
653 unsigned NumBits = 1;
658 if (DestReg == BaseReg && BaseReg == ARM::SP) {
659 assert(isMul4 && "Thumb sp inc / dec size must be multiple of 4!");
662 Opc = isSub ? ARM::tSUBspi : ARM::tADDspi;
664 } else if (!isSub && BaseReg == ARM::SP) {
667 // r1 = add sp, 100 * 4
671 ExtraOpc = ARM::tADDi3;
680 if (DestReg != BaseReg)
683 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
687 unsigned NumMIs = calcNumMI(Opc, ExtraOpc, Bytes, NumBits, Scale);
688 unsigned Threshold = (DestReg == ARM::SP) ? 3 : 2;
689 if (NumMIs > Threshold) {
690 // This will expand into too many instructions. Load the immediate from a
692 emitThumbRegPlusImmInReg(MBB, MBBI, DestReg, BaseReg, NumBytes, true, TII,
698 if (MRI.isLowRegister(DestReg) && MRI.isLowRegister(BaseReg)) {
699 // If both are low registers, emit DestReg = add BaseReg, max(Imm, 7)
700 unsigned Chunk = (1 << 3) - 1;
701 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
703 BuildMI(MBB, MBBI, dl,TII.get(isSub ? ARM::tSUBi3 : ARM::tADDi3), DestReg)
704 .addReg(BaseReg, RegState::Kill).addImm(ThisVal);
706 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), DestReg)
707 .addReg(BaseReg, RegState::Kill);
712 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
714 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
717 // Build the new tADD / tSUB.
719 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
720 .addReg(DestReg).addImm(ThisVal);
722 bool isKill = BaseReg != ARM::SP;
723 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
724 .addReg(BaseReg, getKillRegState(isKill)).addImm(ThisVal);
727 if (Opc == ARM::tADDrSPi) {
733 Chunk = ((1 << NumBits) - 1) * Scale;
734 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
741 BuildMI(MBB, MBBI, dl, TII.get(ExtraOpc), DestReg)
742 .addReg(DestReg, RegState::Kill)
743 .addImm(((unsigned)NumBytes) & 3);
747 void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
748 int NumBytes, ARMCC::CondCodes Pred, unsigned PredReg,
749 bool isThumb, const TargetInstrInfo &TII,
750 const ARMRegisterInfo& MRI,
753 emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII,
756 emitARMRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes,
757 Pred, PredReg, TII, dl);
760 void ARMRegisterInfo::
761 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
762 MachineBasicBlock::iterator I) const {
763 if (!hasReservedCallFrame(MF)) {
764 // If we have alloca, convert as follows:
765 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
766 // ADJCALLSTACKUP -> add, sp, sp, amount
767 MachineInstr *Old = I;
768 DebugLoc dl = Old->getDebugLoc();
769 unsigned Amount = Old->getOperand(0).getImm();
771 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
772 // We need to keep the stack aligned properly. To do this, we round the
773 // amount of space needed for the outgoing arguments up to the next
774 // alignment boundary.
775 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
776 Amount = (Amount+Align-1)/Align*Align;
778 // Replace the pseudo instruction with a new instruction...
779 unsigned Opc = Old->getOpcode();
780 bool isThumb = AFI->isThumbFunction();
781 ARMCC::CondCodes Pred = isThumb
782 ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(1).getImm();
783 if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) {
784 // Note: PredReg is operand 2 for ADJCALLSTACKDOWN.
785 unsigned PredReg = isThumb ? 0 : Old->getOperand(2).getReg();
786 emitSPUpdate(MBB, I, -Amount, Pred, PredReg, isThumb, TII, *this, dl);
788 // Note: PredReg is operand 3 for ADJCALLSTACKUP.
789 unsigned PredReg = isThumb ? 0 : Old->getOperand(3).getReg();
790 assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP);
791 emitSPUpdate(MBB, I, Amount, Pred, PredReg, isThumb, TII, *this, dl);
798 /// emitThumbConstant - Emit a series of instructions to materialize a
800 static void emitThumbConstant(MachineBasicBlock &MBB,
801 MachineBasicBlock::iterator &MBBI,
802 unsigned DestReg, int Imm,
803 const TargetInstrInfo &TII,
804 const ARMRegisterInfo& MRI,
806 bool isSub = Imm < 0;
807 if (isSub) Imm = -Imm;
809 int Chunk = (1 << 8) - 1;
810 int ThisVal = (Imm > Chunk) ? Chunk : Imm;
812 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVi8), DestReg).addImm(ThisVal);
814 emitThumbRegPlusImmediate(MBB, MBBI, DestReg, DestReg, Imm, TII, MRI, dl);
816 BuildMI(MBB, MBBI, dl, TII.get(ARM::tNEG), DestReg)
817 .addReg(DestReg, RegState::Kill);
820 /// findScratchRegister - Find a 'free' ARM register. If register scavenger
821 /// is not being used, R12 is available. Otherwise, try for a call-clobbered
822 /// register first and then a spilled callee-saved register if that fails.
824 unsigned findScratchRegister(RegScavenger *RS, const TargetRegisterClass *RC,
825 ARMFunctionInfo *AFI) {
826 unsigned Reg = RS ? RS->FindUnusedReg(RC, true) : (unsigned) ARM::R12;
827 assert (!AFI->isThumbFunction());
829 // Try a already spilled CS register.
830 Reg = RS->FindUnusedReg(RC, AFI->getSpilledCSRegisters());
835 void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
836 int SPAdj, RegScavenger *RS) const{
838 MachineInstr &MI = *II;
839 MachineBasicBlock &MBB = *MI.getParent();
840 MachineFunction &MF = *MBB.getParent();
841 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
842 bool isThumb = AFI->isThumbFunction();
843 DebugLoc dl = MI.getDebugLoc();
845 while (!MI.getOperand(i).isFI()) {
847 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
850 unsigned FrameReg = ARM::SP;
851 int FrameIndex = MI.getOperand(i).getIndex();
852 int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex) +
853 MF.getFrameInfo()->getStackSize() + SPAdj;
855 if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex))
856 Offset -= AFI->getGPRCalleeSavedArea1Offset();
857 else if (AFI->isGPRCalleeSavedArea2Frame(FrameIndex))
858 Offset -= AFI->getGPRCalleeSavedArea2Offset();
859 else if (AFI->isDPRCalleeSavedAreaFrame(FrameIndex))
860 Offset -= AFI->getDPRCalleeSavedAreaOffset();
861 else if (hasFP(MF)) {
862 assert(SPAdj == 0 && "Unexpected");
863 // There is alloca()'s in this function, must reference off the frame
865 FrameReg = getFrameRegister(MF);
866 Offset -= AFI->getFramePtrSpillOffset();
869 unsigned Opcode = MI.getOpcode();
870 const TargetInstrDesc &Desc = MI.getDesc();
871 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
874 // Memory operands in inline assembly always use AddrMode2.
875 if (Opcode == ARM::INLINEASM)
876 AddrMode = ARMII::AddrMode2;
878 if (Opcode == ARM::ADDri) {
879 Offset += MI.getOperand(i+1).getImm();
881 // Turn it into a move.
882 MI.setDesc(TII.get(ARM::MOVr));
883 MI.getOperand(i).ChangeToRegister(FrameReg, false);
884 MI.RemoveOperand(i+1);
886 } else if (Offset < 0) {
889 MI.setDesc(TII.get(ARM::SUBri));
892 // Common case: small offset, fits into instruction.
893 int ImmedOffset = ARM_AM::getSOImmVal(Offset);
894 if (ImmedOffset != -1) {
895 // Replace the FrameIndex with sp / fp
896 MI.getOperand(i).ChangeToRegister(FrameReg, false);
897 MI.getOperand(i+1).ChangeToImmediate(ImmedOffset);
901 // Otherwise, we fallback to common code below to form the imm offset with
902 // a sequence of ADDri instructions. First though, pull as much of the imm
903 // into this ADDri as possible.
904 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
905 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
907 // We will handle these bits from offset, clear them.
908 Offset &= ~ThisImmVal;
910 // Get the properly encoded SOImmVal field.
911 int ThisSOImmVal = ARM_AM::getSOImmVal(ThisImmVal);
912 assert(ThisSOImmVal != -1 && "Bit extraction didn't work?");
913 MI.getOperand(i+1).ChangeToImmediate(ThisSOImmVal);
914 } else if (Opcode == ARM::tADDrSPi) {
915 Offset += MI.getOperand(i+1).getImm();
917 // Can't use tADDrSPi if it's based off the frame pointer.
918 unsigned NumBits = 0;
920 if (FrameReg != ARM::SP) {
921 Opcode = ARM::tADDi3;
922 MI.setDesc(TII.get(ARM::tADDi3));
927 assert((Offset & 3) == 0 &&
928 "Thumb add/sub sp, #imm immediate must be multiple of 4!");
932 // Turn it into a move.
933 MI.setDesc(TII.get(ARM::tMOVhir2lor));
934 MI.getOperand(i).ChangeToRegister(FrameReg, false);
935 MI.RemoveOperand(i+1);
939 // Common case: small offset, fits into instruction.
940 unsigned Mask = (1 << NumBits) - 1;
941 if (((Offset / Scale) & ~Mask) == 0) {
942 // Replace the FrameIndex with sp / fp
943 MI.getOperand(i).ChangeToRegister(FrameReg, false);
944 MI.getOperand(i+1).ChangeToImmediate(Offset / Scale);
948 unsigned DestReg = MI.getOperand(0).getReg();
949 unsigned Bytes = (Offset > 0) ? Offset : -Offset;
950 unsigned NumMIs = calcNumMI(Opcode, 0, Bytes, NumBits, Scale);
951 // MI would expand into a large number of instructions. Don't try to
952 // simplify the immediate.
954 emitThumbRegPlusImmediate(MBB, II, DestReg, FrameReg, Offset, TII,
961 // Translate r0 = add sp, imm to
962 // r0 = add sp, 255*4
963 // r0 = add r0, (imm - 255*4)
964 MI.getOperand(i).ChangeToRegister(FrameReg, false);
965 MI.getOperand(i+1).ChangeToImmediate(Mask);
966 Offset = (Offset - Mask * Scale);
967 MachineBasicBlock::iterator NII = next(II);
968 emitThumbRegPlusImmediate(MBB, NII, DestReg, DestReg, Offset, TII,
971 // Translate r0 = add sp, -imm to
972 // r0 = -imm (this is then translated into a series of instructons)
974 emitThumbConstant(MBB, II, DestReg, Offset, TII, *this, dl);
975 MI.setDesc(TII.get(ARM::tADDhirr));
976 MI.getOperand(i).ChangeToRegister(DestReg, false, false, true);
977 MI.getOperand(i+1).ChangeToRegister(FrameReg, false);
983 unsigned NumBits = 0;
986 case ARMII::AddrMode2: {
988 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
989 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
994 case ARMII::AddrMode3: {
996 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
997 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1002 case ARMII::AddrMode5: {
1004 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
1005 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1011 case ARMII::AddrModeTs: {
1013 InstrOffs = MI.getOperand(ImmIdx).getImm();
1014 NumBits = (FrameReg == ARM::SP) ? 8 : 5;
1019 assert(0 && "Unsupported addressing mode!");
1024 Offset += InstrOffs * Scale;
1025 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1026 if (Offset < 0 && !isThumb) {
1031 // Common case: small offset, fits into instruction.
1032 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1033 int ImmedOffset = Offset / Scale;
1034 unsigned Mask = (1 << NumBits) - 1;
1035 if ((unsigned)Offset <= Mask * Scale) {
1036 // Replace the FrameIndex with sp
1037 MI.getOperand(i).ChangeToRegister(FrameReg, false);
1039 ImmedOffset |= 1 << NumBits;
1040 ImmOp.ChangeToImmediate(ImmedOffset);
1044 bool isThumSpillRestore = Opcode == ARM::tRestore || Opcode == ARM::tSpill;
1045 if (AddrMode == ARMII::AddrModeTs) {
1046 // Thumb tLDRspi, tSTRspi. These will change to instructions that use
1047 // a different base register.
1049 Mask = (1 << NumBits) - 1;
1051 // If this is a thumb spill / restore, we will be using a constpool load to
1052 // materialize the offset.
1053 if (AddrMode == ARMII::AddrModeTs && isThumSpillRestore)
1054 ImmOp.ChangeToImmediate(0);
1056 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1057 ImmedOffset = ImmedOffset & Mask;
1059 ImmedOffset |= 1 << NumBits;
1060 ImmOp.ChangeToImmediate(ImmedOffset);
1061 Offset &= ~(Mask*Scale);
1065 // If we get here, the immediate doesn't fit into the instruction. We folded
1066 // as much as possible above, handle the rest, providing a register that is
1068 assert(Offset && "This code isn't needed if offset already handled!");
1071 if (Desc.mayLoad()) {
1072 // Use the destination register to materialize sp + offset.
1073 unsigned TmpReg = MI.getOperand(0).getReg();
1075 if (Opcode == ARM::tRestore) {
1076 if (FrameReg == ARM::SP)
1077 emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg,
1078 Offset, false, TII, *this, dl);
1080 emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, &TII,
1085 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII,
1087 MI.setDesc(TII.get(ARM::tLDR));
1088 MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true);
1090 // Use [reg, reg] addrmode.
1091 MI.addOperand(MachineOperand::CreateReg(FrameReg, false));
1092 else // tLDR has an extra register operand.
1093 MI.addOperand(MachineOperand::CreateReg(0, false));
1094 } else if (Desc.mayStore()) {
1095 // FIXME! This is horrific!!! We need register scavenging.
1096 // Our temporary workaround has marked r3 unavailable. Of course, r3 is
1097 // also a ABI register so it's possible that is is the register that is
1098 // being storing here. If that's the case, we do the following:
1100 // Use r2 to materialize sp + offset
1103 unsigned ValReg = MI.getOperand(0).getReg();
1104 unsigned TmpReg = ARM::R3;
1106 if (ValReg == ARM::R3) {
1107 BuildMI(MBB, II, dl, TII.get(ARM::tMOVlor2hir), ARM::R12)
1108 .addReg(ARM::R2, RegState::Kill);
1111 if (TmpReg == ARM::R3 && AFI->isR3LiveIn())
1112 BuildMI(MBB, II, dl, TII.get(ARM::tMOVlor2hir), ARM::R12)
1113 .addReg(ARM::R3, RegState::Kill);
1114 if (Opcode == ARM::tSpill) {
1115 if (FrameReg == ARM::SP)
1116 emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg,
1117 Offset, false, TII, *this, dl);
1119 emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, &TII,
1124 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII,
1126 MI.setDesc(TII.get(ARM::tSTR));
1127 MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true);
1128 if (UseRR) // Use [reg, reg] addrmode.
1129 MI.addOperand(MachineOperand::CreateReg(FrameReg, false));
1130 else // tSTR has an extra register operand.
1131 MI.addOperand(MachineOperand::CreateReg(0, false));
1133 MachineBasicBlock::iterator NII = next(II);
1134 if (ValReg == ARM::R3)
1135 BuildMI(MBB, NII, dl, TII.get(ARM::tMOVhir2lor), ARM::R2)
1136 .addReg(ARM::R12, RegState::Kill);
1137 if (TmpReg == ARM::R3 && AFI->isR3LiveIn())
1138 BuildMI(MBB, NII, dl, TII.get(ARM::tMOVhir2lor), ARM::R3)
1139 .addReg(ARM::R12, RegState::Kill);
1141 assert(false && "Unexpected opcode!");
1143 // Insert a set of r12 with the full address: r12 = sp + offset
1144 // If the offset we have is too large to fit into the instruction, we need
1145 // to form it with a series of ADDri's. Do this by taking 8-bit chunks
1147 unsigned ScratchReg = findScratchRegister(RS, &ARM::GPRRegClass, AFI);
1148 if (ScratchReg == 0)
1149 // No register is "free". Scavenge a register.
1150 ScratchReg = RS->scavengeRegister(&ARM::GPRRegClass, II, SPAdj);
1151 int PIdx = MI.findFirstPredOperandIdx();
1152 ARMCC::CondCodes Pred = (PIdx == -1)
1153 ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm();
1154 unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg();
1155 emitARMRegPlusImmediate(MBB, II, ScratchReg, FrameReg,
1156 isSub ? -Offset : Offset, Pred, PredReg, TII, dl);
1157 MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true);
1161 static unsigned estimateStackSize(MachineFunction &MF, MachineFrameInfo *MFI) {
1162 const MachineFrameInfo *FFI = MF.getFrameInfo();
1164 for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) {
1165 int FixedOff = -FFI->getObjectOffset(i);
1166 if (FixedOff > Offset) Offset = FixedOff;
1168 for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) {
1169 if (FFI->isDeadObjectIndex(i))
1171 Offset += FFI->getObjectSize(i);
1172 unsigned Align = FFI->getObjectAlignment(i);
1173 // Adjust to alignment boundary
1174 Offset = (Offset+Align-1)/Align*Align;
1176 return (unsigned)Offset;
1180 ARMRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
1181 RegScavenger *RS) const {
1182 // This tells PEI to spill the FP as if it is any other callee-save register
1183 // to take advantage the eliminateFrameIndex machinery. This also ensures it
1184 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
1185 // to combine multiple loads / stores.
1186 bool CanEliminateFrame = true;
1187 bool CS1Spilled = false;
1188 bool LRSpilled = false;
1189 unsigned NumGPRSpills = 0;
1190 SmallVector<unsigned, 4> UnspilledCS1GPRs;
1191 SmallVector<unsigned, 4> UnspilledCS2GPRs;
1192 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1194 // Don't spill FP if the frame can be eliminated. This is determined
1195 // by scanning the callee-save registers to see if any is used.
1196 const unsigned *CSRegs = getCalleeSavedRegs();
1197 const TargetRegisterClass* const *CSRegClasses = getCalleeSavedRegClasses();
1198 for (unsigned i = 0; CSRegs[i]; ++i) {
1199 unsigned Reg = CSRegs[i];
1200 bool Spilled = false;
1201 if (MF.getRegInfo().isPhysRegUsed(Reg)) {
1202 AFI->setCSRegisterIsSpilled(Reg);
1204 CanEliminateFrame = false;
1206 // Check alias registers too.
1207 for (const unsigned *Aliases = getAliasSet(Reg); *Aliases; ++Aliases) {
1208 if (MF.getRegInfo().isPhysRegUsed(*Aliases)) {
1210 CanEliminateFrame = false;
1215 if (CSRegClasses[i] == &ARM::GPRRegClass) {
1219 if (!STI.isTargetDarwin()) {
1226 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
1241 if (!STI.isTargetDarwin()) {
1242 UnspilledCS1GPRs.push_back(Reg);
1252 UnspilledCS1GPRs.push_back(Reg);
1255 UnspilledCS2GPRs.push_back(Reg);
1262 bool ForceLRSpill = false;
1263 if (!LRSpilled && AFI->isThumbFunction()) {
1264 unsigned FnSize = TII.GetFunctionSizeInBytes(MF);
1265 // Force LR to be spilled if the Thumb function size is > 2048. This enables
1266 // use of BL to implement far jump. If it turns out that it's not needed
1267 // then the branch fix up path will undo it.
1268 if (FnSize >= (1 << 11)) {
1269 CanEliminateFrame = false;
1270 ForceLRSpill = true;
1274 bool ExtraCSSpill = false;
1275 if (!CanEliminateFrame || hasFP(MF)) {
1276 AFI->setHasStackFrame(true);
1278 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
1279 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
1280 if (!LRSpilled && CS1Spilled) {
1281 MF.getRegInfo().setPhysRegUsed(ARM::LR);
1282 AFI->setCSRegisterIsSpilled(ARM::LR);
1284 UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(),
1285 UnspilledCS1GPRs.end(), (unsigned)ARM::LR));
1286 ForceLRSpill = false;
1287 ExtraCSSpill = true;
1290 // Darwin ABI requires FP to point to the stack slot that contains the
1292 if (STI.isTargetDarwin() || hasFP(MF)) {
1293 MF.getRegInfo().setPhysRegUsed(FramePtr);
1297 // If stack and double are 8-byte aligned and we are spilling an odd number
1298 // of GPRs. Spill one extra callee save GPR so we won't have to pad between
1299 // the integer and double callee save areas.
1300 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
1301 if (TargetAlign == 8 && (NumGPRSpills & 1)) {
1302 if (CS1Spilled && !UnspilledCS1GPRs.empty()) {
1303 for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) {
1304 unsigned Reg = UnspilledCS1GPRs[i];
1305 // Don't spiil high register if the function is thumb
1306 if (!AFI->isThumbFunction() || isLowRegister(Reg) || Reg == ARM::LR) {
1307 MF.getRegInfo().setPhysRegUsed(Reg);
1308 AFI->setCSRegisterIsSpilled(Reg);
1309 if (!isReservedReg(MF, Reg))
1310 ExtraCSSpill = true;
1314 } else if (!UnspilledCS2GPRs.empty() &&
1315 !AFI->isThumbFunction()) {
1316 unsigned Reg = UnspilledCS2GPRs.front();
1317 MF.getRegInfo().setPhysRegUsed(Reg);
1318 AFI->setCSRegisterIsSpilled(Reg);
1319 if (!isReservedReg(MF, Reg))
1320 ExtraCSSpill = true;
1324 // Estimate if we might need to scavenge a register at some point in order
1325 // to materialize a stack offset. If so, either spill one additiona
1326 // callee-saved register or reserve a special spill slot to facilitate
1327 // register scavenging.
1328 if (RS && !ExtraCSSpill && !AFI->isThumbFunction()) {
1329 MachineFrameInfo *MFI = MF.getFrameInfo();
1330 unsigned Size = estimateStackSize(MF, MFI);
1331 unsigned Limit = (1 << 12) - 1;
1332 for (MachineFunction::iterator BB = MF.begin(),E = MF.end();BB != E; ++BB)
1333 for (MachineBasicBlock::iterator I= BB->begin(); I != BB->end(); ++I) {
1334 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1335 if (I->getOperand(i).isFI()) {
1336 unsigned Opcode = I->getOpcode();
1337 const TargetInstrDesc &Desc = TII.get(Opcode);
1338 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1339 if (AddrMode == ARMII::AddrMode3) {
1340 Limit = (1 << 8) - 1;
1341 goto DoneEstimating;
1342 } else if (AddrMode == ARMII::AddrMode5) {
1343 unsigned ThisLimit = ((1 << 8) - 1) * 4;
1344 if (ThisLimit < Limit)
1350 if (Size >= Limit) {
1351 // If any non-reserved CS register isn't spilled, just spill one or two
1352 // extra. That should take care of it!
1353 unsigned NumExtras = TargetAlign / 4;
1354 SmallVector<unsigned, 2> Extras;
1355 while (NumExtras && !UnspilledCS1GPRs.empty()) {
1356 unsigned Reg = UnspilledCS1GPRs.back();
1357 UnspilledCS1GPRs.pop_back();
1358 if (!isReservedReg(MF, Reg)) {
1359 Extras.push_back(Reg);
1363 while (NumExtras && !UnspilledCS2GPRs.empty()) {
1364 unsigned Reg = UnspilledCS2GPRs.back();
1365 UnspilledCS2GPRs.pop_back();
1366 if (!isReservedReg(MF, Reg)) {
1367 Extras.push_back(Reg);
1371 if (Extras.size() && NumExtras == 0) {
1372 for (unsigned i = 0, e = Extras.size(); i != e; ++i) {
1373 MF.getRegInfo().setPhysRegUsed(Extras[i]);
1374 AFI->setCSRegisterIsSpilled(Extras[i]);
1377 // Reserve a slot closest to SP or frame pointer.
1378 const TargetRegisterClass *RC = &ARM::GPRRegClass;
1379 RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
1380 RC->getAlignment()));
1387 MF.getRegInfo().setPhysRegUsed(ARM::LR);
1388 AFI->setCSRegisterIsSpilled(ARM::LR);
1389 AFI->setLRIsSpilledForFarJump(true);
1393 /// Move iterator pass the next bunch of callee save load / store ops for
1394 /// the particular spill area (1: integer area 1, 2: integer area 2,
1395 /// 3: fp area, 0: don't care).
1396 static void movePastCSLoadStoreOps(MachineBasicBlock &MBB,
1397 MachineBasicBlock::iterator &MBBI,
1398 int Opc, unsigned Area,
1399 const ARMSubtarget &STI) {
1400 while (MBBI != MBB.end() &&
1401 MBBI->getOpcode() == Opc && MBBI->getOperand(1).isFI()) {
1404 unsigned Category = 0;
1405 switch (MBBI->getOperand(0).getReg()) {
1406 case ARM::R4: case ARM::R5: case ARM::R6: case ARM::R7:
1410 case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11:
1411 Category = STI.isTargetDarwin() ? 2 : 1;
1413 case ARM::D8: case ARM::D9: case ARM::D10: case ARM::D11:
1414 case ARM::D12: case ARM::D13: case ARM::D14: case ARM::D15:
1421 if (Done || Category != Area)
1429 void ARMRegisterInfo::emitPrologue(MachineFunction &MF) const {
1430 MachineBasicBlock &MBB = MF.front();
1431 MachineBasicBlock::iterator MBBI = MBB.begin();
1432 MachineFrameInfo *MFI = MF.getFrameInfo();
1433 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1434 bool isThumb = AFI->isThumbFunction();
1435 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1436 unsigned NumBytes = MFI->getStackSize();
1437 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
1438 DebugLoc dl = (MBBI != MBB.end() ?
1439 MBBI->getDebugLoc() : DebugLoc::getUnknownLoc());
1442 // Check if R3 is live in. It might have to be used as a scratch register.
1443 for (MachineRegisterInfo::livein_iterator I =MF.getRegInfo().livein_begin(),
1444 E = MF.getRegInfo().livein_end(); I != E; ++I) {
1445 if (I->first == ARM::R3) {
1446 AFI->setR3IsLiveIn(true);
1451 // Thumb add/sub sp, imm8 instructions implicitly multiply the offset by 4.
1452 NumBytes = (NumBytes + 3) & ~3;
1453 MFI->setStackSize(NumBytes);
1456 // Determine the sizes of each callee-save spill areas and record which frame
1457 // belongs to which callee-save spill areas.
1458 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
1459 int FramePtrSpillFI = 0;
1462 emitSPUpdate(MBB, MBBI, -VARegSaveSize, ARMCC::AL, 0, isThumb, TII,
1465 if (!AFI->hasStackFrame()) {
1467 emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, isThumb, TII, *this, dl);
1471 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1472 unsigned Reg = CSI[i].getReg();
1473 int FI = CSI[i].getFrameIdx();
1480 if (Reg == FramePtr)
1481 FramePtrSpillFI = FI;
1482 AFI->addGPRCalleeSavedArea1Frame(FI);
1489 if (Reg == FramePtr)
1490 FramePtrSpillFI = FI;
1491 if (STI.isTargetDarwin()) {
1492 AFI->addGPRCalleeSavedArea2Frame(FI);
1495 AFI->addGPRCalleeSavedArea1Frame(FI);
1500 AFI->addDPRCalleeSavedAreaFrame(FI);
1506 // Build the new SUBri to adjust SP for integer callee-save spill area 1.
1507 emitSPUpdate(MBB, MBBI, -GPRCS1Size, ARMCC::AL, 0, isThumb, TII, *this, dl);
1508 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 1, STI);
1509 } else if (MBBI != MBB.end() && MBBI->getOpcode() == ARM::tPUSH) {
1511 if (MBBI != MBB.end())
1512 dl = MBBI->getDebugLoc();
1515 // Darwin ABI requires FP to point to the stack slot that contains the
1517 if (STI.isTargetDarwin() || hasFP(MF)) {
1518 MachineInstrBuilder MIB =
1519 BuildMI(MBB, MBBI, dl, TII.get(isThumb ? ARM::tADDrSPi : ARM::ADDri),
1521 .addFrameIndex(FramePtrSpillFI).addImm(0);
1522 if (!isThumb) AddDefaultCC(AddDefaultPred(MIB));
1526 // Build the new SUBri to adjust SP for integer callee-save spill area 2.
1527 emitSPUpdate(MBB, MBBI, -GPRCS2Size, ARMCC::AL, 0, false, TII, *this, dl);
1529 // Build the new SUBri to adjust SP for FP callee-save spill area.
1530 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 2, STI);
1531 emitSPUpdate(MBB, MBBI, -DPRCSSize, ARMCC::AL, 0, false, TII, *this, dl);
1534 // Determine starting offsets of spill areas.
1535 unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize);
1536 unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize;
1537 unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size;
1538 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) + NumBytes);
1539 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
1540 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
1541 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
1543 NumBytes = DPRCSOffset;
1545 // Insert it after all the callee-save spills.
1547 movePastCSLoadStoreOps(MBB, MBBI, ARM::FSTD, 3, STI);
1548 emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, isThumb, TII, *this, dl);
1551 if(STI.isTargetELF() && hasFP(MF)) {
1552 MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() -
1553 AFI->getFramePtrSpillOffset());
1556 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
1557 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
1558 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
1561 static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) {
1562 for (unsigned i = 0; CSRegs[i]; ++i)
1563 if (Reg == CSRegs[i])
1568 static bool isCSRestore(MachineInstr *MI, const unsigned *CSRegs) {
1569 return ((MI->getOpcode() == ARM::FLDD ||
1570 MI->getOpcode() == ARM::LDR ||
1571 MI->getOpcode() == ARM::tRestore) &&
1572 MI->getOperand(1).isFI() &&
1573 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs));
1576 void ARMRegisterInfo::emitEpilogue(MachineFunction &MF,
1577 MachineBasicBlock &MBB) const {
1578 MachineBasicBlock::iterator MBBI = prior(MBB.end());
1579 assert((MBBI->getOpcode() == ARM::BX_RET ||
1580 MBBI->getOpcode() == ARM::tBX_RET ||
1581 MBBI->getOpcode() == ARM::tPOP_RET) &&
1582 "Can only insert epilog into returning blocks");
1583 DebugLoc dl = MBBI->getDebugLoc();
1584 MachineFrameInfo *MFI = MF.getFrameInfo();
1585 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1586 bool isThumb = AFI->isThumbFunction();
1587 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1588 int NumBytes = (int)MFI->getStackSize();
1590 if (!AFI->hasStackFrame()) {
1592 emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, isThumb, TII, *this, dl);
1594 // Unwind MBBI to point to first LDR / FLDD.
1595 const unsigned *CSRegs = getCalleeSavedRegs();
1596 if (MBBI != MBB.begin()) {
1599 while (MBBI != MBB.begin() && isCSRestore(MBBI, CSRegs));
1600 if (!isCSRestore(MBBI, CSRegs))
1604 // Move SP to start of FP callee save spill area.
1605 NumBytes -= (AFI->getGPRCalleeSavedArea1Size() +
1606 AFI->getGPRCalleeSavedArea2Size() +
1607 AFI->getDPRCalleeSavedAreaSize());
1610 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1611 // Reset SP based on frame pointer only if the stack frame extends beyond
1612 // frame pointer stack slot or target is ELF and the function has FP.
1614 emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, FramePtr, -NumBytes,
1617 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVlor2hir), ARM::SP)
1620 if (MBBI->getOpcode() == ARM::tBX_RET &&
1621 &MBB.front() != MBBI &&
1622 prior(MBBI)->getOpcode() == ARM::tPOP) {
1623 MachineBasicBlock::iterator PMBBI = prior(MBBI);
1624 emitSPUpdate(MBB, PMBBI, NumBytes, ARMCC::AL, 0, isThumb, TII,
1627 emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, isThumb, TII,
1631 // Darwin ABI requires FP to point to the stack slot that contains the
1633 if ((STI.isTargetDarwin() && NumBytes) || hasFP(MF)) {
1634 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1635 // Reset SP based on frame pointer only if the stack frame extends beyond
1636 // frame pointer stack slot or target is ELF and the function has FP.
1637 if (AFI->getGPRCalleeSavedArea2Size() ||
1638 AFI->getDPRCalleeSavedAreaSize() ||
1639 AFI->getDPRCalleeSavedAreaOffset()||
1642 BuildMI(MBB, MBBI, dl, TII.get(ARM::SUBri), ARM::SP).addReg(FramePtr)
1644 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
1646 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), ARM::SP).addReg(FramePtr)
1647 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
1649 } else if (NumBytes) {
1650 emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, false, TII, *this, dl);
1653 // Move SP to start of integer callee save spill area 2.
1654 movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 3, STI);
1655 emitSPUpdate(MBB, MBBI, AFI->getDPRCalleeSavedAreaSize(), ARMCC::AL, 0,
1656 false, TII, *this, dl);
1658 // Move SP to start of integer callee save spill area 1.
1659 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 2, STI);
1660 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea2Size(), ARMCC::AL, 0,
1661 false, TII, *this, dl);
1663 // Move SP to SP upon entry to the function.
1664 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 1, STI);
1665 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea1Size(), ARMCC::AL, 0,
1666 false, TII, *this, dl);
1670 if (VARegSaveSize) {
1672 // Epilogue for vararg functions: pop LR to R3 and branch off it.
1673 // FIXME: Verify this is still ok when R3 is no longer being reserved.
1674 BuildMI(MBB, MBBI, dl, TII.get(ARM::tPOP)).addReg(ARM::R3);
1676 emitSPUpdate(MBB, MBBI, VARegSaveSize, ARMCC::AL, 0, isThumb, TII,
1680 BuildMI(MBB, MBBI, dl, TII.get(ARM::tBX_RET_vararg)).addReg(ARM::R3);
1686 unsigned ARMRegisterInfo::getRARegister() const {
1690 unsigned ARMRegisterInfo::getFrameRegister(MachineFunction &MF) const {
1691 if (STI.isTargetDarwin() || hasFP(MF))
1696 unsigned ARMRegisterInfo::getEHExceptionRegister() const {
1697 assert(0 && "What is the exception register");
1701 unsigned ARMRegisterInfo::getEHHandlerRegister() const {
1702 assert(0 && "What is the exception handler register");
1706 int ARMRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const {
1707 return ARMGenRegisterInfo::getDwarfRegNumFull(RegNum, 0);
1710 unsigned ARMRegisterInfo::getRegisterPairEven(unsigned Reg,
1711 const MachineFunction &MF) const {
1714 // Return 0 if either register of the pair is a special register.
1720 return STI.isThumb() ? 0 : ARM::R2;
1724 return isReservedReg(MF, ARM::R7) ? 0 : ARM::R6;
1726 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R8;
1728 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R10;
1784 unsigned ARMRegisterInfo::getRegisterPairOdd(unsigned Reg,
1785 const MachineFunction &MF) const {
1788 // Return 0 if either register of the pair is a special register.
1794 return STI.isThumb() ? 0 : ARM::R3;
1798 return isReservedReg(MF, ARM::R7) ? 0 : ARM::R7;
1800 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R9;
1802 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R11;
1858 #include "ARMGenRegisterInfo.inc"