1 //===- ARMRegisterInfo.cpp - ARM Register Information -----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the "Instituto Nokia de Tecnologia" and
6 // is distributed under the University of Illinois Open Source
7 // License. See LICENSE.TXT for details.
9 //===----------------------------------------------------------------------===//
11 // This file contains the ARM implementation of the MRegisterInfo class.
13 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMInstrInfo.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMRegisterInfo.h"
20 #include "ARMSubtarget.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineLocation.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> EnableScavenging("enable-arm-reg-scavenging", cl::Hidden,
40 cl::desc("Enable register scavenging on ARM"));
42 unsigned ARMRegisterInfo::getRegisterNumbering(unsigned RegEnum) {
45 case R0: case S0: case D0: return 0;
46 case R1: case S1: case D1: return 1;
47 case R2: case S2: case D2: return 2;
48 case R3: case S3: case D3: return 3;
49 case R4: case S4: case D4: return 4;
50 case R5: case S5: case D5: return 5;
51 case R6: case S6: case D6: return 6;
52 case R7: case S7: case D7: return 7;
53 case R8: case S8: case D8: return 8;
54 case R9: case S9: case D9: return 9;
55 case R10: case S10: case D10: return 10;
56 case R11: case S11: case D11: return 11;
57 case R12: case S12: case D12: return 12;
58 case SP: case S13: case D13: return 13;
59 case LR: case S14: case D14: return 14;
60 case PC: case S15: case D15: return 15;
78 assert(0 && "Unknown ARM register!");
83 ARMRegisterInfo::ARMRegisterInfo(const TargetInstrInfo &tii,
84 const ARMSubtarget &sti)
85 : ARMGenRegisterInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
87 FramePtr(STI.useThumbBacktraces() ? ARM::R7 : ARM::R11) {
90 bool ARMRegisterInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
91 MachineBasicBlock::iterator MI,
92 const std::vector<CalleeSavedInfo> &CSI) const {
93 MachineFunction &MF = *MBB.getParent();
94 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
95 if (!AFI->isThumbFunction() || CSI.empty())
98 MachineInstrBuilder MIB = BuildMI(MBB, MI, TII.get(ARM::tPUSH));
99 for (unsigned i = CSI.size(); i != 0; --i) {
100 unsigned Reg = CSI[i-1].getReg();
101 // Add the callee-saved register as live-in. It's killed at the spill.
103 MIB.addReg(Reg, false/*isDef*/,false/*isImp*/,true/*isKill*/);
108 bool ARMRegisterInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
109 MachineBasicBlock::iterator MI,
110 const std::vector<CalleeSavedInfo> &CSI) const {
111 MachineFunction &MF = *MBB.getParent();
112 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
113 if (!AFI->isThumbFunction() || CSI.empty())
116 bool isVarArg = AFI->getVarArgsRegSaveSize() > 0;
117 MachineInstr *PopMI = new MachineInstr(TII.get(ARM::tPOP));
118 MBB.insert(MI, PopMI);
119 for (unsigned i = CSI.size(); i != 0; --i) {
120 unsigned Reg = CSI[i-1].getReg();
121 if (Reg == ARM::LR) {
122 // Special epilogue for vararg functions. See emitEpilogue
126 PopMI->setInstrDescriptor(TII.get(ARM::tPOP_RET));
129 PopMI->addRegOperand(Reg, true);
134 void ARMRegisterInfo::
135 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
136 unsigned SrcReg, int FI,
137 const TargetRegisterClass *RC) const {
138 if (RC == ARM::GPRRegisterClass) {
139 MachineFunction &MF = *MBB.getParent();
140 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
141 if (AFI->isThumbFunction())
142 BuildMI(MBB, I, TII.get(ARM::tSpill)).addReg(SrcReg, false, false, true)
143 .addFrameIndex(FI).addImm(0);
145 BuildMI(MBB, I, TII.get(ARM::STR)).addReg(SrcReg, false, false, true)
146 .addFrameIndex(FI).addReg(0).addImm(0);
147 } else if (RC == ARM::DPRRegisterClass) {
148 BuildMI(MBB, I, TII.get(ARM::FSTD)).addReg(SrcReg, false, false, true)
149 .addFrameIndex(FI).addImm(0);
151 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
152 BuildMI(MBB, I, TII.get(ARM::FSTS)).addReg(SrcReg, false, false, true)
153 .addFrameIndex(FI).addImm(0);
157 void ARMRegisterInfo::
158 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
159 unsigned DestReg, int FI,
160 const TargetRegisterClass *RC) const {
161 if (RC == ARM::GPRRegisterClass) {
162 MachineFunction &MF = *MBB.getParent();
163 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
164 if (AFI->isThumbFunction())
165 BuildMI(MBB, I, TII.get(ARM::tRestore), DestReg)
166 .addFrameIndex(FI).addImm(0);
168 BuildMI(MBB, I, TII.get(ARM::LDR), DestReg)
169 .addFrameIndex(FI).addReg(0).addImm(0);
170 } else if (RC == ARM::DPRRegisterClass) {
171 BuildMI(MBB, I, TII.get(ARM::FLDD), DestReg)
172 .addFrameIndex(FI).addImm(0);
174 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
175 BuildMI(MBB, I, TII.get(ARM::FLDS), DestReg)
176 .addFrameIndex(FI).addImm(0);
180 void ARMRegisterInfo::copyRegToReg(MachineBasicBlock &MBB,
181 MachineBasicBlock::iterator I,
182 unsigned DestReg, unsigned SrcReg,
183 const TargetRegisterClass *RC) const {
184 if (RC == ARM::GPRRegisterClass) {
185 MachineFunction &MF = *MBB.getParent();
186 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
187 BuildMI(MBB, I, TII.get(AFI->isThumbFunction() ? ARM::tMOVrr : ARM::MOVrr),
188 DestReg).addReg(SrcReg);
189 } else if (RC == ARM::SPRRegisterClass)
190 BuildMI(MBB, I, TII.get(ARM::FCPYS), DestReg).addReg(SrcReg);
191 else if (RC == ARM::DPRRegisterClass)
192 BuildMI(MBB, I, TII.get(ARM::FCPYD), DestReg).addReg(SrcReg);
197 /// isLowRegister - Returns true if the register is low register r0-r7.
199 static bool isLowRegister(unsigned Reg) {
202 case R0: case R1: case R2: case R3:
203 case R4: case R5: case R6: case R7:
210 MachineInstr *ARMRegisterInfo::foldMemoryOperand(MachineInstr *MI,
211 unsigned OpNum, int FI) const {
212 unsigned Opc = MI->getOpcode();
213 MachineInstr *NewMI = NULL;
217 if (OpNum == 0) { // move -> store
218 unsigned SrcReg = MI->getOperand(1).getReg();
219 NewMI = BuildMI(TII.get(ARM::STR)).addReg(SrcReg).addFrameIndex(FI)
220 .addReg(0).addImm(0);
221 } else { // move -> load
222 unsigned DstReg = MI->getOperand(0).getReg();
223 NewMI = BuildMI(TII.get(ARM::LDR), DstReg).addFrameIndex(FI).addReg(0)
229 if (OpNum == 0) { // move -> store
230 unsigned SrcReg = MI->getOperand(1).getReg();
231 if (isPhysicalRegister(SrcReg) && !isLowRegister(SrcReg))
232 // tSpill cannot take a high register operand.
234 NewMI = BuildMI(TII.get(ARM::tSpill)).addReg(SrcReg).addFrameIndex(FI)
236 } else { // move -> load
237 unsigned DstReg = MI->getOperand(0).getReg();
238 if (isPhysicalRegister(DstReg) && !isLowRegister(DstReg))
239 // tRestore cannot target a high register operand.
241 NewMI = BuildMI(TII.get(ARM::tRestore), DstReg).addFrameIndex(FI)
247 if (OpNum == 0) { // move -> store
248 unsigned SrcReg = MI->getOperand(1).getReg();
249 NewMI = BuildMI(TII.get(ARM::FSTS)).addReg(SrcReg).addFrameIndex(FI)
251 } else { // move -> load
252 unsigned DstReg = MI->getOperand(0).getReg();
253 NewMI = BuildMI(TII.get(ARM::FLDS), DstReg).addFrameIndex(FI).addImm(0);
258 if (OpNum == 0) { // move -> store
259 unsigned SrcReg = MI->getOperand(1).getReg();
260 NewMI = BuildMI(TII.get(ARM::FSTD)).addReg(SrcReg).addFrameIndex(FI)
262 } else { // move -> load
263 unsigned DstReg = MI->getOperand(0).getReg();
264 NewMI = BuildMI(TII.get(ARM::FLDD), DstReg).addFrameIndex(FI).addImm(0);
271 NewMI->copyKillDeadInfo(MI);
275 const unsigned* ARMRegisterInfo::getCalleeSavedRegs() const {
276 static const unsigned CalleeSavedRegs[] = {
277 ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8,
278 ARM::R7, ARM::R6, ARM::R5, ARM::R4,
280 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
281 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
285 static const unsigned DarwinCalleeSavedRegs[] = {
286 ARM::LR, ARM::R7, ARM::R6, ARM::R5, ARM::R4,
287 ARM::R11, ARM::R10, ARM::R9, ARM::R8,
289 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
290 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
293 return STI.isTargetDarwin() ? DarwinCalleeSavedRegs : CalleeSavedRegs;
296 const TargetRegisterClass* const *
297 ARMRegisterInfo::getCalleeSavedRegClasses() const {
298 static const TargetRegisterClass * const CalleeSavedRegClasses[] = {
299 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
300 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
301 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
303 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
304 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
307 return CalleeSavedRegClasses;
310 BitVector ARMRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
311 // FIXME: avoid re-calculating this everytime.
312 BitVector Reserved(getNumRegs());
313 Reserved.set(ARM::SP);
314 Reserved.set(ARM::PC);
315 if (STI.isTargetDarwin() || hasFP(MF))
316 Reserved.set(FramePtr);
317 // Some targets reserve R9.
318 if (STI.isR9Reserved())
319 Reserved.set(ARM::R9);
320 // At PEI time, if LR is used, it will be spilled upon entry.
321 if (MF.getUsedPhysregs() && !MF.isPhysRegUsed((unsigned)ARM::LR))
322 Reserved.set(ARM::LR);
327 ARMRegisterInfo::isReservedReg(const MachineFunction &MF, unsigned Reg) const {
335 if (FramePtr == Reg && (STI.isTargetDarwin() || hasFP(MF)))
339 return STI.isR9Reserved();
346 ARMRegisterInfo::requiresRegisterScavenging(const MachineFunction &MF) const {
347 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
348 return EnableScavenging && !AFI->isThumbFunction();
351 /// hasFP - Return true if the specified function should have a dedicated frame
352 /// pointer register. This is true if the function has variable sized allocas
353 /// or if frame pointer elimination is disabled.
355 bool ARMRegisterInfo::hasFP(const MachineFunction &MF) const {
356 return NoFramePointerElim || MF.getFrameInfo()->hasVarSizedObjects();
359 /// emitARMRegPlusImmediate - Emits a series of instructions to materialize
360 /// a destreg = basereg + immediate in ARM code.
362 void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
363 MachineBasicBlock::iterator &MBBI,
364 unsigned DestReg, unsigned BaseReg,
365 int NumBytes, const TargetInstrInfo &TII) {
366 bool isSub = NumBytes < 0;
367 if (isSub) NumBytes = -NumBytes;
370 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
371 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
372 assert(ThisVal && "Didn't extract field correctly");
374 // We will handle these bits from offset, clear them.
375 NumBytes &= ~ThisVal;
377 // Get the properly encoded SOImmVal field.
378 int SOImmVal = ARM_AM::getSOImmVal(ThisVal);
379 assert(SOImmVal != -1 && "Bit extraction didn't work?");
381 // Build the new ADD / SUB.
382 BuildMI(MBB, MBBI, TII.get(isSub ? ARM::SUBri : ARM::ADDri), DestReg)
383 .addReg(BaseReg, false, false, true).addImm(SOImmVal);
388 /// calcNumMI - Returns the number of instructions required to materialize
389 /// the specific add / sub r, c instruction.
390 static unsigned calcNumMI(int Opc, int ExtraOpc, unsigned Bytes,
391 unsigned NumBits, unsigned Scale) {
393 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
395 if (Opc == ARM::tADDrSPi) {
396 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
401 Chunk = ((1 << NumBits) - 1) * Scale;
404 NumMIs += Bytes / Chunk;
405 if ((Bytes % Chunk) != 0)
412 /// emitLoadConstPool - Emits a load from constpool to materialize NumBytes
414 static void emitLoadConstPool(MachineBasicBlock &MBB,
415 MachineBasicBlock::iterator &MBBI,
416 unsigned DestReg, int NumBytes,
417 const TargetInstrInfo &TII) {
418 MachineFunction &MF = *MBB.getParent();
419 MachineConstantPool *ConstantPool = MF.getConstantPool();
420 Constant *C = ConstantInt::get(Type::Int32Ty, NumBytes);
421 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 2);
422 BuildMI(MBB, MBBI, TII.get(ARM::tLDRpci), DestReg).addConstantPoolIndex(Idx);
425 /// emitThumbRegPlusImmInReg - Emits a series of instructions to materialize
426 /// a destreg = basereg + immediate in Thumb code. Materialize the immediate
427 /// in a register using mov / mvn sequences or load the immediate from a
430 void emitThumbRegPlusImmInReg(MachineBasicBlock &MBB,
431 MachineBasicBlock::iterator &MBBI,
432 unsigned DestReg, unsigned BaseReg,
433 int NumBytes, bool CanChangeCC,
434 const TargetInstrInfo &TII) {
435 bool isHigh = !isLowRegister(DestReg) ||
436 (BaseReg != 0 && !isLowRegister(BaseReg));
438 // Subtract doesn't have high register version. Load the negative value
439 // if either base or dest register is a high register. Also, if do not
440 // issue sub as part of the sequence if condition register is to be
442 if (NumBytes < 0 && !isHigh && CanChangeCC) {
444 NumBytes = -NumBytes;
446 unsigned LdReg = DestReg;
447 if (DestReg == ARM::SP) {
448 assert(BaseReg == ARM::SP && "Unexpected!");
450 BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), ARM::R12)
451 .addReg(ARM::R3, false, false, true);
454 if (NumBytes <= 255 && NumBytes >= 0)
455 BuildMI(MBB, MBBI, TII.get(ARM::tMOVri8), LdReg).addImm(NumBytes);
456 else if (NumBytes < 0 && NumBytes >= -255) {
457 BuildMI(MBB, MBBI, TII.get(ARM::tMOVri8), LdReg).addImm(NumBytes);
458 BuildMI(MBB, MBBI, TII.get(ARM::tNEG), LdReg)
459 .addReg(LdReg, false, false, true);
461 emitLoadConstPool(MBB, MBBI, LdReg, NumBytes, TII);
464 int Opc = (isSub) ? ARM::tSUBrr : (isHigh ? ARM::tADDhirr : ARM::tADDrr);
465 const MachineInstrBuilder MIB = BuildMI(MBB, MBBI, TII.get(Opc), DestReg);
466 if (DestReg == ARM::SP || isSub)
467 MIB.addReg(BaseReg).addReg(LdReg, false, false, true);
469 MIB.addReg(LdReg).addReg(BaseReg, false, false, true);
470 if (DestReg == ARM::SP)
471 BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), ARM::R3)
472 .addReg(ARM::R12, false, false, true);
475 /// emitThumbRegPlusImmediate - Emits a series of instructions to materialize
476 /// a destreg = basereg + immediate in Thumb code.
478 void emitThumbRegPlusImmediate(MachineBasicBlock &MBB,
479 MachineBasicBlock::iterator &MBBI,
480 unsigned DestReg, unsigned BaseReg,
481 int NumBytes, const TargetInstrInfo &TII) {
482 bool isSub = NumBytes < 0;
483 unsigned Bytes = (unsigned)NumBytes;
484 if (isSub) Bytes = -NumBytes;
485 bool isMul4 = (Bytes & 3) == 0;
486 bool isTwoAddr = false;
487 bool DstNotEqBase = false;
488 unsigned NumBits = 1;
493 if (DestReg == BaseReg && BaseReg == ARM::SP) {
494 assert(isMul4 && "Thumb sp inc / dec size must be multiple of 4!");
497 Opc = isSub ? ARM::tSUBspi : ARM::tADDspi;
499 } else if (!isSub && BaseReg == ARM::SP) {
502 // r1 = add sp, 100 * 4
506 ExtraOpc = ARM::tADDi3;
515 if (DestReg != BaseReg)
518 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
522 unsigned NumMIs = calcNumMI(Opc, ExtraOpc, Bytes, NumBits, Scale);
523 unsigned Threshold = (DestReg == ARM::SP) ? 3 : 2;
524 if (NumMIs > Threshold) {
525 // This will expand into too many instructions. Load the immediate from a
527 emitThumbRegPlusImmInReg(MBB, MBBI, DestReg, BaseReg, NumBytes, true, TII);
532 if (isLowRegister(DestReg) && isLowRegister(BaseReg)) {
533 // If both are low registers, emit DestReg = add BaseReg, max(Imm, 7)
534 unsigned Chunk = (1 << 3) - 1;
535 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
537 BuildMI(MBB, MBBI, TII.get(isSub ? ARM::tSUBi3 : ARM::tADDi3), DestReg)
538 .addReg(BaseReg, false, false, true).addImm(ThisVal);
540 BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), DestReg)
541 .addReg(BaseReg, false, false, true);
546 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
548 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
551 // Build the new tADD / tSUB.
553 BuildMI(MBB, MBBI, TII.get(Opc), DestReg).addReg(DestReg).addImm(ThisVal);
555 bool isKill = BaseReg != ARM::SP;
556 BuildMI(MBB, MBBI, TII.get(Opc), DestReg)
557 .addReg(BaseReg, false, false, isKill).addImm(ThisVal);
560 if (Opc == ARM::tADDrSPi) {
566 Chunk = ((1 << NumBits) - 1) * Scale;
567 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
574 BuildMI(MBB, MBBI, TII.get(ExtraOpc), DestReg)
575 .addReg(DestReg, false, false, true)
576 .addImm(((unsigned)NumBytes) & 3);
580 void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
581 int NumBytes, bool isThumb, const TargetInstrInfo &TII) {
583 emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII);
585 emitARMRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII);
588 void ARMRegisterInfo::
589 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
590 MachineBasicBlock::iterator I) const {
592 // If we have alloca, convert as follows:
593 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
594 // ADJCALLSTACKUP -> add, sp, sp, amount
595 MachineInstr *Old = I;
596 unsigned Amount = Old->getOperand(0).getImmedValue();
598 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
599 // We need to keep the stack aligned properly. To do this, we round the
600 // amount of space needed for the outgoing arguments up to the next
601 // alignment boundary.
602 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
603 Amount = (Amount+Align-1)/Align*Align;
605 // Replace the pseudo instruction with a new instruction...
606 if (Old->getOpcode() == ARM::ADJCALLSTACKDOWN) {
607 emitSPUpdate(MBB, I, -Amount, AFI->isThumbFunction(), TII);
609 assert(Old->getOpcode() == ARM::ADJCALLSTACKUP);
610 emitSPUpdate(MBB, I, Amount, AFI->isThumbFunction(), TII);
617 /// emitThumbConstant - Emit a series of instructions to materialize a
619 static void emitThumbConstant(MachineBasicBlock &MBB,
620 MachineBasicBlock::iterator &MBBI,
621 unsigned DestReg, int Imm,
622 const TargetInstrInfo &TII) {
623 bool isSub = Imm < 0;
624 if (isSub) Imm = -Imm;
626 int Chunk = (1 << 8) - 1;
627 int ThisVal = (Imm > Chunk) ? Chunk : Imm;
629 BuildMI(MBB, MBBI, TII.get(ARM::tMOVri8), DestReg).addImm(ThisVal);
631 emitThumbRegPlusImmediate(MBB, MBBI, DestReg, DestReg, Imm, TII);
633 BuildMI(MBB, MBBI, TII.get(ARM::tNEG), DestReg)
634 .addReg(DestReg, false, false, true);
637 /// findScratchRegister - Find a 'free' ARM register. If register scavenger
638 /// is not being used, R12 is available. Otherwise, try for a call-clobbered
639 /// register first and then a spilled callee-saved register if that fails.
641 unsigned findScratchRegister(RegScavenger *RS, const TargetRegisterClass *RC,
642 ARMFunctionInfo *AFI) {
643 unsigned Reg = RS ? RS->FindUnusedReg(RC, true) : (unsigned) ARM::R12;
645 // Try a already spilled CS register.
646 Reg = RS->FindUnusedReg(RC, AFI->getSpilledCSRegisters());
651 void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
652 RegScavenger *RS) const{
654 MachineInstr &MI = *II;
655 MachineBasicBlock &MBB = *MI.getParent();
656 MachineFunction &MF = *MBB.getParent();
657 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
658 bool isThumb = AFI->isThumbFunction();
660 while (!MI.getOperand(i).isFrameIndex()) {
662 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
665 unsigned FrameReg = ARM::SP;
666 int FrameIndex = MI.getOperand(i).getFrameIndex();
667 int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex) +
668 MF.getFrameInfo()->getStackSize();
670 if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex))
671 Offset -= AFI->getGPRCalleeSavedArea1Offset();
672 else if (AFI->isGPRCalleeSavedArea2Frame(FrameIndex))
673 Offset -= AFI->getGPRCalleeSavedArea2Offset();
674 else if (AFI->isDPRCalleeSavedAreaFrame(FrameIndex))
675 Offset -= AFI->getDPRCalleeSavedAreaOffset();
676 else if (hasFP(MF)) {
677 // There is alloca()'s in this function, must reference off the frame
679 FrameReg = getFrameRegister(MF);
680 Offset -= AFI->getFramePtrSpillOffset();
683 unsigned Opcode = MI.getOpcode();
684 const TargetInstrDescriptor &Desc = TII.get(Opcode);
685 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
688 if (Opcode == ARM::ADDri) {
689 Offset += MI.getOperand(i+1).getImm();
691 // Turn it into a move.
692 MI.setInstrDescriptor(TII.get(ARM::MOVrr));
693 MI.getOperand(i).ChangeToRegister(FrameReg, false);
694 MI.RemoveOperand(i+1);
696 } else if (Offset < 0) {
699 MI.setInstrDescriptor(TII.get(ARM::SUBri));
702 // Common case: small offset, fits into instruction.
703 int ImmedOffset = ARM_AM::getSOImmVal(Offset);
704 if (ImmedOffset != -1) {
705 // Replace the FrameIndex with sp / fp
706 MI.getOperand(i).ChangeToRegister(FrameReg, false);
707 MI.getOperand(i+1).ChangeToImmediate(ImmedOffset);
711 // Otherwise, we fallback to common code below to form the imm offset with
712 // a sequence of ADDri instructions. First though, pull as much of the imm
713 // into this ADDri as possible.
714 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
715 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, (32-RotAmt) & 31);
717 // We will handle these bits from offset, clear them.
718 Offset &= ~ThisImmVal;
720 // Get the properly encoded SOImmVal field.
721 int ThisSOImmVal = ARM_AM::getSOImmVal(ThisImmVal);
722 assert(ThisSOImmVal != -1 && "Bit extraction didn't work?");
723 MI.getOperand(i+1).ChangeToImmediate(ThisSOImmVal);
724 } else if (Opcode == ARM::tADDrSPi) {
725 Offset += MI.getOperand(i+1).getImm();
726 assert((Offset & 3) == 0 &&
727 "Thumb add/sub sp, #imm immediate must be multiple of 4!");
729 // Turn it into a move.
730 MI.setInstrDescriptor(TII.get(ARM::tMOVrr));
731 MI.getOperand(i).ChangeToRegister(FrameReg, false);
732 MI.RemoveOperand(i+1);
736 // Common case: small offset, fits into instruction.
737 if (((Offset >> 2) & ~255U) == 0) {
738 // Replace the FrameIndex with sp / fp
739 MI.getOperand(i).ChangeToRegister(FrameReg, false);
740 MI.getOperand(i+1).ChangeToImmediate(Offset >> 2);
744 unsigned DestReg = MI.getOperand(0).getReg();
745 unsigned Bytes = (Offset > 0) ? Offset : -Offset;
746 unsigned NumMIs = calcNumMI(Opcode, 0, Bytes, 8, 1);
747 // MI would expand into a large number of instructions. Don't try to
748 // simplify the immediate.
750 emitThumbRegPlusImmediate(MBB, II, DestReg, FrameReg, Offset, TII);
756 // Translate r0 = add sp, imm to
757 // r0 = add sp, 255*4
758 // r0 = add r0, (imm - 255*4)
759 MI.getOperand(i).ChangeToRegister(FrameReg, false);
760 MI.getOperand(i+1).ChangeToImmediate(255);
761 Offset = (Offset - 255 * 4);
762 MachineBasicBlock::iterator NII = next(II);
763 emitThumbRegPlusImmediate(MBB, NII, DestReg, DestReg, Offset, TII);
765 // Translate r0 = add sp, -imm to
766 // r0 = -imm (this is then translated into a series of instructons)
768 emitThumbConstant(MBB, II, DestReg, Offset, TII);
769 MI.setInstrDescriptor(TII.get(ARM::tADDhirr));
770 MI.getOperand(i).ChangeToRegister(DestReg, false, false, true);
771 MI.getOperand(i+1).ChangeToRegister(FrameReg, false);
777 unsigned NumBits = 0;
780 case ARMII::AddrMode2: {
782 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
783 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
788 case ARMII::AddrMode3: {
790 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
791 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
796 case ARMII::AddrMode5: {
798 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
799 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
805 case ARMII::AddrModeTs: {
807 InstrOffs = MI.getOperand(ImmIdx).getImm();
808 NumBits = (FrameReg == ARM::SP) ? 8 : 5;
813 assert(0 && "Unsupported addressing mode!");
818 Offset += InstrOffs * Scale;
819 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
820 if (Offset < 0 && !isThumb) {
825 // Common case: small offset, fits into instruction.
826 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
827 int ImmedOffset = Offset / Scale;
828 unsigned Mask = (1 << NumBits) - 1;
829 if ((unsigned)Offset <= Mask * Scale) {
830 // Replace the FrameIndex with sp
831 MI.getOperand(i).ChangeToRegister(FrameReg, false);
833 ImmedOffset |= 1 << NumBits;
834 ImmOp.ChangeToImmediate(ImmedOffset);
838 bool isThumSpillRestore = Opcode == ARM::tRestore || Opcode == ARM::tSpill;
839 if (AddrMode == ARMII::AddrModeTs) {
840 // Thumb tLDRspi, tSTRspi. These will change to instructions that use
841 // a different base register.
843 Mask = (1 << NumBits) - 1;
845 // If this is a thumb spill / restore, we will be using a constpool load to
846 // materialize the offset.
847 if (AddrMode == ARMII::AddrModeTs && isThumSpillRestore)
848 ImmOp.ChangeToImmediate(0);
850 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
851 ImmedOffset = ImmedOffset & Mask;
853 ImmedOffset |= 1 << NumBits;
854 ImmOp.ChangeToImmediate(ImmedOffset);
855 Offset &= ~(Mask*Scale);
859 // If we get here, the immediate doesn't fit into the instruction. We folded
860 // as much as possible above, handle the rest, providing a register that is
862 assert(Offset && "This code isn't needed if offset already handled!");
865 if (TII.isLoad(Opcode)) {
866 // Use the destination register to materialize sp + offset.
867 unsigned TmpReg = MI.getOperand(0).getReg();
869 if (Opcode == ARM::tRestore) {
870 if (FrameReg == ARM::SP)
871 emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg,Offset,false,TII);
873 emitLoadConstPool(MBB, II, TmpReg, Offset, TII);
877 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII);
878 MI.setInstrDescriptor(TII.get(ARM::tLDR));
879 MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true);
881 MI.addRegOperand(FrameReg, false); // Use [reg, reg] addrmode.
883 MI.addRegOperand(0, false); // tLDR has an extra register operand.
884 } else if (TII.isStore(Opcode)) {
885 // FIXME! This is horrific!!! We need register scavenging.
886 // Our temporary workaround has marked r3 unavailable. Of course, r3 is
887 // also a ABI register so it's possible that is is the register that is
888 // being storing here. If that's the case, we do the following:
890 // Use r2 to materialize sp + offset
893 unsigned ValReg = MI.getOperand(0).getReg();
894 unsigned TmpReg = ARM::R3;
896 if (ValReg == ARM::R3) {
897 BuildMI(MBB, II, TII.get(ARM::tMOVrr), ARM::R12)
898 .addReg(ARM::R2, false, false, true);
901 if (TmpReg == ARM::R3 && AFI->isR3LiveIn())
902 BuildMI(MBB, II, TII.get(ARM::tMOVrr), ARM::R12)
903 .addReg(ARM::R3, false, false, true);
904 if (Opcode == ARM::tSpill) {
905 if (FrameReg == ARM::SP)
906 emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg,Offset,false,TII);
908 emitLoadConstPool(MBB, II, TmpReg, Offset, TII);
912 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII);
913 MI.setInstrDescriptor(TII.get(ARM::tSTR));
914 MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true);
916 MI.addRegOperand(FrameReg, false); // Use [reg, reg] addrmode.
918 MI.addRegOperand(0, false); // tSTR has an extra register operand.
920 MachineBasicBlock::iterator NII = next(II);
921 if (ValReg == ARM::R3)
922 BuildMI(MBB, NII, TII.get(ARM::tMOVrr), ARM::R2)
923 .addReg(ARM::R12, false, false, true);
924 if (TmpReg == ARM::R3 && AFI->isR3LiveIn())
925 BuildMI(MBB, NII, TII.get(ARM::tMOVrr), ARM::R3)
926 .addReg(ARM::R12, false, false, true);
928 assert(false && "Unexpected opcode!");
930 // Insert a set of r12 with the full address: r12 = sp + offset
931 // If the offset we have is too large to fit into the instruction, we need
932 // to form it with a series of ADDri's. Do this by taking 8-bit chunks
934 unsigned ScratchReg = findScratchRegister(RS, &ARM::GPRRegClass, AFI);
936 // No register is "free". Scavenge a register.
937 ScratchReg = RS->scavengeRegister(&ARM::GPRRegClass, II);
938 emitARMRegPlusImmediate(MBB, II, ScratchReg, FrameReg,
939 isSub ? -Offset : Offset, TII);
940 MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true);
944 static unsigned estimateStackSize(MachineFunction &MF, MachineFrameInfo *MFI) {
945 const MachineFrameInfo *FFI = MF.getFrameInfo();
947 for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) {
948 int FixedOff = -FFI->getObjectOffset(i);
949 if (FixedOff > Offset) Offset = FixedOff;
951 for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) {
952 Offset += FFI->getObjectSize(i);
953 unsigned Align = FFI->getObjectAlignment(i);
954 // Adjust to alignment boundary
955 Offset = (Offset+Align-1)/Align*Align;
957 return (unsigned)Offset;
961 ARMRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
962 RegScavenger *RS) const {
963 // This tells PEI to spill the FP as if it is any other callee-save register
964 // to take advantage the eliminateFrameIndex machinery. This also ensures it
965 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
966 // to combine multiple loads / stores.
967 bool CanEliminateFrame = true;
968 bool CS1Spilled = false;
969 bool LRSpilled = false;
970 unsigned NumGPRSpills = 0;
971 SmallVector<unsigned, 4> UnspilledCS1GPRs;
972 SmallVector<unsigned, 4> UnspilledCS2GPRs;
973 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
975 // Don't spill FP if the frame can be eliminated. This is determined
976 // by scanning the callee-save registers to see if any is used.
977 const unsigned *CSRegs = getCalleeSavedRegs();
978 const TargetRegisterClass* const *CSRegClasses = getCalleeSavedRegClasses();
979 for (unsigned i = 0; CSRegs[i]; ++i) {
980 unsigned Reg = CSRegs[i];
981 bool Spilled = false;
982 if (MF.isPhysRegUsed(Reg)) {
983 AFI->setCSRegisterIsSpilled(Reg);
985 CanEliminateFrame = false;
987 // Check alias registers too.
988 for (const unsigned *Aliases = getAliasSet(Reg); *Aliases; ++Aliases) {
989 if (MF.isPhysRegUsed(*Aliases)) {
991 CanEliminateFrame = false;
996 if (CSRegClasses[i] == &ARM::GPRRegClass) {
1000 if (!STI.isTargetDarwin()) {
1008 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
1023 if (!STI.isTargetDarwin()) {
1024 UnspilledCS1GPRs.push_back(Reg);
1034 UnspilledCS1GPRs.push_back(Reg);
1037 UnspilledCS2GPRs.push_back(Reg);
1044 bool ForceLRSpill = false;
1045 if (!LRSpilled && AFI->isThumbFunction()) {
1046 unsigned FnSize = ARM::GetFunctionSize(MF);
1047 // Force LR to be spilled if the Thumb function size is > 2048. This enables
1048 // use of BL to implement far jump. If it turns out that it's not needed
1049 // then the branch fix up path will undo it.
1050 if (FnSize >= (1 << 11)) {
1051 CanEliminateFrame = false;
1052 ForceLRSpill = true;
1056 bool ExtraCSSpill = false;
1057 if (!CanEliminateFrame || hasFP(MF)) {
1058 AFI->setHasStackFrame(true);
1060 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
1061 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
1062 if (!LRSpilled && CS1Spilled) {
1063 MF.changePhyRegUsed(ARM::LR, true);
1064 AFI->setCSRegisterIsSpilled(ARM::LR);
1066 UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(),
1067 UnspilledCS1GPRs.end(), (unsigned)ARM::LR));
1068 ForceLRSpill = false;
1069 ExtraCSSpill = true;
1072 // Darwin ABI requires FP to point to the stack slot that contains the
1074 if (STI.isTargetDarwin() || hasFP(MF)) {
1075 MF.changePhyRegUsed(FramePtr, true);
1079 // If stack and double are 8-byte aligned and we are spilling an odd number
1080 // of GPRs. Spill one extra callee save GPR so we won't have to pad between
1081 // the integer and double callee save areas.
1082 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
1083 if (TargetAlign == 8 && (NumGPRSpills & 1)) {
1084 if (CS1Spilled && !UnspilledCS1GPRs.empty()) {
1085 unsigned Reg = UnspilledCS1GPRs.front();
1086 MF.changePhyRegUsed(Reg, true);
1087 AFI->setCSRegisterIsSpilled(Reg);
1088 if (!isReservedReg(MF, Reg))
1089 ExtraCSSpill = true;
1090 } else if (!UnspilledCS2GPRs.empty()) {
1091 unsigned Reg = UnspilledCS2GPRs.front();
1092 MF.changePhyRegUsed(Reg, true);
1093 AFI->setCSRegisterIsSpilled(Reg);
1094 if (!isReservedReg(MF, Reg))
1095 ExtraCSSpill = true;
1099 // Estimate if we might need to scavenge a register at some point in order
1100 // to materialize a stack offset. If so, either spill one additiona
1101 // callee-saved register or reserve a special spill slot to facilitate
1102 // register scavenging.
1103 if (RS && !ExtraCSSpill && !AFI->isThumbFunction()) {
1104 MachineFrameInfo *MFI = MF.getFrameInfo();
1105 unsigned Size = estimateStackSize(MF, MFI);
1106 unsigned Limit = (1 << 12) - 1;
1107 for (MachineFunction::iterator BB = MF.begin(),E = MF.end();BB != E; ++BB)
1108 for (MachineBasicBlock::iterator I= BB->begin(); I != BB->end(); ++I) {
1109 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1110 if (I->getOperand(i).isFrameIndex()) {
1111 unsigned Opcode = I->getOpcode();
1112 const TargetInstrDescriptor &Desc = TII.get(Opcode);
1113 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1114 if (AddrMode == ARMII::AddrMode3) {
1115 Limit = (1 << 8) - 1;
1116 goto DoneEstimating;
1117 } else if (AddrMode == ARMII::AddrMode5) {
1118 Limit = ((1 << 8) - 1) * 4;
1119 goto DoneEstimating;
1124 if (Size >= Limit) {
1125 // If any non-reserved CS register isn't spilled, just spill one or two
1126 // extra. That should take care of it!
1127 unsigned NumExtras = TargetAlign / 4;
1128 SmallVector<unsigned, 2> Extras;
1129 while (NumExtras && !UnspilledCS1GPRs.empty()) {
1130 unsigned Reg = UnspilledCS1GPRs.back();
1131 UnspilledCS1GPRs.pop_back();
1132 if (!isReservedReg(MF, Reg)) {
1133 Extras.push_back(Reg);
1137 while (NumExtras && !UnspilledCS2GPRs.empty()) {
1138 unsigned Reg = UnspilledCS2GPRs.back();
1139 UnspilledCS2GPRs.pop_back();
1140 if (!isReservedReg(MF, Reg)) {
1141 Extras.push_back(Reg);
1145 if (Extras.size() && NumExtras == 0) {
1146 for (unsigned i = 0, e = Extras.size(); i != e; ++i) {
1147 MF.changePhyRegUsed(Extras[i], true);
1148 AFI->setCSRegisterIsSpilled(Extras[i]);
1151 // Reserve a slot closest to SP or frame pointer.
1152 const TargetRegisterClass *RC = &ARM::GPRRegClass;
1153 RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
1154 RC->getAlignment()));
1161 MF.changePhyRegUsed(ARM::LR, true);
1162 AFI->setCSRegisterIsSpilled(ARM::LR);
1163 AFI->setLRIsSpilledForFarJump(true);
1167 /// Move iterator pass the next bunch of callee save load / store ops for
1168 /// the particular spill area (1: integer area 1, 2: integer area 2,
1169 /// 3: fp area, 0: don't care).
1170 static void movePastCSLoadStoreOps(MachineBasicBlock &MBB,
1171 MachineBasicBlock::iterator &MBBI,
1172 int Opc, unsigned Area,
1173 const ARMSubtarget &STI) {
1174 while (MBBI != MBB.end() &&
1175 MBBI->getOpcode() == Opc && MBBI->getOperand(1).isFrameIndex()) {
1178 unsigned Category = 0;
1179 switch (MBBI->getOperand(0).getReg()) {
1180 case ARM::R4: case ARM::R5: case ARM::R6: case ARM::R7:
1184 case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11:
1185 Category = STI.isTargetDarwin() ? 2 : 1;
1187 case ARM::D8: case ARM::D9: case ARM::D10: case ARM::D11:
1188 case ARM::D12: case ARM::D13: case ARM::D14: case ARM::D15:
1195 if (Done || Category != Area)
1203 void ARMRegisterInfo::emitPrologue(MachineFunction &MF) const {
1204 MachineBasicBlock &MBB = MF.front();
1205 MachineBasicBlock::iterator MBBI = MBB.begin();
1206 MachineFrameInfo *MFI = MF.getFrameInfo();
1207 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1208 bool isThumb = AFI->isThumbFunction();
1209 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1210 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1211 unsigned NumBytes = MFI->getStackSize();
1212 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
1215 // Check if R3 is live in. It might have to be used as a scratch register.
1216 for (MachineFunction::livein_iterator I=MF.livein_begin(),E=MF.livein_end();
1218 if ((*I).first == ARM::R3) {
1219 AFI->setR3IsLiveIn(true);
1224 // Thumb add/sub sp, imm8 instructions implicitly multiply the offset by 4.
1225 NumBytes = (NumBytes + 3) & ~3;
1226 MFI->setStackSize(NumBytes);
1229 // Determine the sizes of each callee-save spill areas and record which frame
1230 // belongs to which callee-save spill areas.
1231 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
1232 int FramePtrSpillFI = 0;
1235 emitSPUpdate(MBB, MBBI, -VARegSaveSize, isThumb, TII);
1237 if (!AFI->hasStackFrame()) {
1239 emitSPUpdate(MBB, MBBI, -NumBytes, isThumb, TII);
1243 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1244 unsigned Reg = CSI[i].getReg();
1245 int FI = CSI[i].getFrameIdx();
1252 if (Reg == FramePtr)
1253 FramePtrSpillFI = FI;
1254 AFI->addGPRCalleeSavedArea1Frame(FI);
1261 if (Reg == FramePtr)
1262 FramePtrSpillFI = FI;
1263 if (STI.isTargetDarwin()) {
1264 AFI->addGPRCalleeSavedArea2Frame(FI);
1267 AFI->addGPRCalleeSavedArea1Frame(FI);
1272 AFI->addDPRCalleeSavedAreaFrame(FI);
1277 if (Align == 8 && (GPRCS1Size & 7) != 0)
1278 // Pad CS1 to ensure proper alignment.
1282 // Build the new SUBri to adjust SP for integer callee-save spill area 1.
1283 emitSPUpdate(MBB, MBBI, -GPRCS1Size, isThumb, TII);
1284 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 1, STI);
1285 } else if (MBBI != MBB.end() && MBBI->getOpcode() == ARM::tPUSH)
1288 // Darwin ABI requires FP to point to the stack slot that contains the
1290 if (STI.isTargetDarwin() || hasFP(MF))
1291 BuildMI(MBB, MBBI, TII.get(isThumb ? ARM::tADDrSPi : ARM::ADDri), FramePtr)
1292 .addFrameIndex(FramePtrSpillFI).addImm(0);
1295 // Build the new SUBri to adjust SP for integer callee-save spill area 2.
1296 emitSPUpdate(MBB, MBBI, -GPRCS2Size, false, TII);
1298 // Build the new SUBri to adjust SP for FP callee-save spill area.
1299 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 2, STI);
1300 emitSPUpdate(MBB, MBBI, -DPRCSSize, false, TII);
1303 // Determine starting offsets of spill areas.
1304 unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize);
1305 unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize;
1306 unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size;
1307 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) + NumBytes);
1308 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
1309 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
1310 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
1312 NumBytes = DPRCSOffset;
1314 // Insert it after all the callee-save spills.
1316 movePastCSLoadStoreOps(MBB, MBBI, ARM::FSTD, 3, STI);
1317 emitSPUpdate(MBB, MBBI, -NumBytes, isThumb, TII);
1320 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
1321 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
1322 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
1325 static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) {
1326 for (unsigned i = 0; CSRegs[i]; ++i)
1327 if (Reg == CSRegs[i])
1332 static bool isCSRestore(MachineInstr *MI, const unsigned *CSRegs) {
1333 return ((MI->getOpcode() == ARM::FLDD ||
1334 MI->getOpcode() == ARM::LDR ||
1335 MI->getOpcode() == ARM::tRestore) &&
1336 MI->getOperand(1).isFrameIndex() &&
1337 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs));
1340 void ARMRegisterInfo::emitEpilogue(MachineFunction &MF,
1341 MachineBasicBlock &MBB) const {
1342 MachineBasicBlock::iterator MBBI = prior(MBB.end());
1343 assert((MBBI->getOpcode() == ARM::BX_RET ||
1344 MBBI->getOpcode() == ARM::tBX_RET ||
1345 MBBI->getOpcode() == ARM::tPOP_RET) &&
1346 "Can only insert epilog into returning blocks");
1348 MachineFrameInfo *MFI = MF.getFrameInfo();
1349 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1350 bool isThumb = AFI->isThumbFunction();
1351 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1352 int NumBytes = (int)MFI->getStackSize();
1353 if (!AFI->hasStackFrame()) {
1355 emitSPUpdate(MBB, MBBI, NumBytes, isThumb, TII);
1357 // Unwind MBBI to point to first LDR / FLDD.
1358 const unsigned *CSRegs = getCalleeSavedRegs();
1359 if (MBBI != MBB.begin()) {
1362 while (MBBI != MBB.begin() && isCSRestore(MBBI, CSRegs));
1363 if (!isCSRestore(MBBI, CSRegs))
1367 // Move SP to start of FP callee save spill area.
1368 NumBytes -= (AFI->getGPRCalleeSavedArea1Size() +
1369 AFI->getGPRCalleeSavedArea2Size() +
1370 AFI->getDPRCalleeSavedAreaSize());
1373 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1374 // Reset SP based on frame pointer only if the stack frame extends beyond
1375 // frame pointer stack slot or target is ELF and the function has FP.
1377 emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, FramePtr, -NumBytes, TII);
1379 BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), ARM::SP).addReg(FramePtr);
1381 if (MBBI->getOpcode() == ARM::tBX_RET &&
1382 &MBB.front() != MBBI &&
1383 prior(MBBI)->getOpcode() == ARM::tPOP) {
1384 MachineBasicBlock::iterator PMBBI = prior(MBBI);
1385 emitSPUpdate(MBB, PMBBI, NumBytes, isThumb, TII);
1387 emitSPUpdate(MBB, MBBI, NumBytes, isThumb, TII);
1390 // Darwin ABI requires FP to point to the stack slot that contains the
1392 if ((STI.isTargetDarwin() && NumBytes) || hasFP(MF)) {
1393 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1394 // Reset SP based on frame pointer only if the stack frame extends beyond
1395 // frame pointer stack slot or target is ELF and the function has FP.
1396 if (AFI->getGPRCalleeSavedArea2Size() ||
1397 AFI->getDPRCalleeSavedAreaSize() ||
1398 AFI->getDPRCalleeSavedAreaOffset()||
1401 BuildMI(MBB, MBBI, TII.get(ARM::SUBri), ARM::SP).addReg(FramePtr)
1404 BuildMI(MBB, MBBI, TII.get(ARM::MOVrr), ARM::SP).addReg(FramePtr);
1405 } else if (NumBytes) {
1406 emitSPUpdate(MBB, MBBI, NumBytes, false, TII);
1409 // Move SP to start of integer callee save spill area 2.
1410 movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 3, STI);
1411 emitSPUpdate(MBB, MBBI, AFI->getDPRCalleeSavedAreaSize(), false, TII);
1413 // Move SP to start of integer callee save spill area 1.
1414 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 2, STI);
1415 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea2Size(), false, TII);
1417 // Move SP to SP upon entry to the function.
1418 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 1, STI);
1419 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea1Size(), false, TII);
1423 if (VARegSaveSize) {
1425 // Epilogue for vararg functions: pop LR to R3 and branch off it.
1426 // FIXME: Verify this is still ok when R3 is no longer being reserved.
1427 BuildMI(MBB, MBBI, TII.get(ARM::tPOP)).addReg(ARM::R3);
1429 emitSPUpdate(MBB, MBBI, VARegSaveSize, isThumb, TII);
1432 BuildMI(MBB, MBBI, TII.get(ARM::tBX_RET_vararg)).addReg(ARM::R3);
1438 unsigned ARMRegisterInfo::getRARegister() const {
1442 unsigned ARMRegisterInfo::getFrameRegister(MachineFunction &MF) const {
1443 return STI.useThumbBacktraces() ? ARM::R7 : ARM::R11;
1446 unsigned ARMRegisterInfo::getEHExceptionRegister() const {
1447 assert(0 && "What is the exception register");
1451 unsigned ARMRegisterInfo::getEHHandlerRegister() const {
1452 assert(0 && "What is the exception handler register");
1456 #include "ARMGenRegisterInfo.inc"