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> 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 ARMRegisterInfo::ARMRegisterInfo(const TargetInstrInfo &tii,
85 const ARMSubtarget &sti)
86 : ARMGenRegisterInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
88 FramePtr((STI.useThumbBacktraces() || STI.isThumb()) ? ARM::R7 : ARM::R11) {
91 bool ARMRegisterInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
92 MachineBasicBlock::iterator MI,
93 const std::vector<CalleeSavedInfo> &CSI) const {
94 MachineFunction &MF = *MBB.getParent();
95 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
96 if (!AFI->isThumbFunction() || CSI.empty())
99 MachineInstrBuilder MIB = BuildMI(MBB, MI, TII.get(ARM::tPUSH));
100 for (unsigned i = CSI.size(); i != 0; --i) {
101 unsigned Reg = CSI[i-1].getReg();
102 // Add the callee-saved register as live-in. It's killed at the spill.
104 MIB.addReg(Reg, false/*isDef*/,false/*isImp*/,true/*isKill*/);
109 bool ARMRegisterInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
110 MachineBasicBlock::iterator MI,
111 const std::vector<CalleeSavedInfo> &CSI) const {
112 MachineFunction &MF = *MBB.getParent();
113 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
114 if (!AFI->isThumbFunction() || CSI.empty())
117 bool isVarArg = AFI->getVarArgsRegSaveSize() > 0;
118 MachineInstr *PopMI = new MachineInstr(TII.get(ARM::tPOP));
119 MBB.insert(MI, PopMI);
120 for (unsigned i = CSI.size(); i != 0; --i) {
121 unsigned Reg = CSI[i-1].getReg();
122 if (Reg == ARM::LR) {
123 // Special epilogue for vararg functions. See emitEpilogue
127 PopMI->setInstrDescriptor(TII.get(ARM::tPOP_RET));
130 PopMI->addRegOperand(Reg, true);
136 const MachineInstrBuilder &AddDefaultPred(const MachineInstrBuilder &MIB) {
137 return MIB.addImm((int64_t)ARMCC::AL).addReg(0);
141 const MachineInstrBuilder &AddDefaultCC(const MachineInstrBuilder &MIB) {
142 return MIB.addReg(0);
145 static const MachineInstrBuilder &ARMInstrAddOperand(MachineInstrBuilder &MIB,
146 MachineOperand &MO) {
148 MIB = MIB.addReg(MO.getReg(), MO.isDef(), MO.isImplicit());
149 else if (MO.isImmediate())
150 MIB = MIB.addImm(MO.getImm());
151 else if (MO.isFrameIndex())
152 MIB = MIB.addFrameIndex(MO.getFrameIndex());
154 assert(0 && "Unknown operand for ARMInstrAddOperand!");
159 void ARMRegisterInfo::
160 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
161 unsigned SrcReg, int FI,
162 const TargetRegisterClass *RC) const {
163 if (RC == ARM::GPRRegisterClass) {
164 MachineFunction &MF = *MBB.getParent();
165 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
166 if (AFI->isThumbFunction())
167 BuildMI(MBB, I, TII.get(ARM::tSpill)).addReg(SrcReg, false, false, true)
168 .addFrameIndex(FI).addImm(0);
170 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::STR))
171 .addReg(SrcReg, false, false, true)
172 .addFrameIndex(FI).addReg(0).addImm(0));
173 } else if (RC == ARM::DPRRegisterClass) {
174 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FSTD))
175 .addReg(SrcReg, false, false, true)
176 .addFrameIndex(FI).addImm(0));
178 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
179 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FSTS))
180 .addReg(SrcReg, false, false, true)
181 .addFrameIndex(FI).addImm(0));
185 void ARMRegisterInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
186 SmallVectorImpl<MachineOperand> &Addr,
187 const TargetRegisterClass *RC,
188 SmallVectorImpl<MachineInstr*> &NewMIs) const {
190 if (RC == ARM::GPRRegisterClass) {
191 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
192 if (AFI->isThumbFunction()) {
193 Opc = Addr[0].isFrameIndex() ? ARM::tSpill : ARM::tSTR;
194 MachineInstrBuilder MIB =
195 BuildMI(TII.get(Opc)).addReg(SrcReg, false, false, true);
196 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
197 MIB = ARMInstrAddOperand(MIB, Addr[i]);
198 NewMIs.push_back(MIB);
202 } else if (RC == ARM::DPRRegisterClass) {
205 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
209 MachineInstrBuilder MIB =
210 BuildMI(TII.get(Opc)).addReg(SrcReg, false, false, true);
211 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
212 MIB = ARMInstrAddOperand(MIB, Addr[i]);
214 NewMIs.push_back(MIB);
218 void ARMRegisterInfo::
219 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
220 unsigned DestReg, int FI,
221 const TargetRegisterClass *RC) const {
222 if (RC == ARM::GPRRegisterClass) {
223 MachineFunction &MF = *MBB.getParent();
224 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
225 if (AFI->isThumbFunction())
226 BuildMI(MBB, I, TII.get(ARM::tRestore), DestReg)
227 .addFrameIndex(FI).addImm(0);
229 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::LDR), DestReg)
230 .addFrameIndex(FI).addReg(0).addImm(0));
231 } else if (RC == ARM::DPRRegisterClass) {
232 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FLDD), DestReg)
233 .addFrameIndex(FI).addImm(0));
235 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
236 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FLDS), DestReg)
237 .addFrameIndex(FI).addImm(0));
241 void ARMRegisterInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
242 SmallVectorImpl<MachineOperand> &Addr,
243 const TargetRegisterClass *RC,
244 SmallVectorImpl<MachineInstr*> &NewMIs) const {
246 if (RC == ARM::GPRRegisterClass) {
247 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
248 if (AFI->isThumbFunction()) {
249 Opc = Addr[0].isFrameIndex() ? ARM::tRestore : ARM::tLDR;
250 MachineInstrBuilder MIB = BuildMI(TII.get(Opc), DestReg);
251 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
252 MIB = ARMInstrAddOperand(MIB, Addr[i]);
253 NewMIs.push_back(MIB);
257 } else if (RC == ARM::DPRRegisterClass) {
260 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
264 MachineInstrBuilder MIB = BuildMI(TII.get(Opc), DestReg);
265 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
266 MIB = ARMInstrAddOperand(MIB, Addr[i]);
268 NewMIs.push_back(MIB);
272 void ARMRegisterInfo::copyRegToReg(MachineBasicBlock &MBB,
273 MachineBasicBlock::iterator I,
274 unsigned DestReg, unsigned SrcReg,
275 const TargetRegisterClass *DestRC,
276 const TargetRegisterClass *SrcRC) const {
277 if (DestRC != SrcRC) {
278 cerr << "Not yet supported!";
282 if (DestRC == ARM::GPRRegisterClass) {
283 MachineFunction &MF = *MBB.getParent();
284 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
285 if (AFI->isThumbFunction())
286 BuildMI(MBB, I, TII.get(ARM::tMOVr), DestReg).addReg(SrcReg);
288 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::MOVr), DestReg)
290 } else if (DestRC == ARM::SPRRegisterClass)
291 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FCPYS), DestReg)
293 else if (DestRC == ARM::DPRRegisterClass)
294 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FCPYD), DestReg)
300 /// emitLoadConstPool - Emits a load from constpool to materialize the
301 /// specified immediate.
302 static void emitLoadConstPool(MachineBasicBlock &MBB,
303 MachineBasicBlock::iterator &MBBI,
304 unsigned DestReg, int Val,
305 ARMCC::CondCodes Pred, unsigned PredReg,
306 const TargetInstrInfo &TII, bool isThumb) {
307 MachineFunction &MF = *MBB.getParent();
308 MachineConstantPool *ConstantPool = MF.getConstantPool();
309 Constant *C = ConstantInt::get(Type::Int32Ty, Val);
310 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 2);
312 BuildMI(MBB, MBBI, TII.get(ARM::tLDRcp), DestReg).addConstantPoolIndex(Idx);
314 BuildMI(MBB, MBBI, TII.get(ARM::LDRcp), DestReg).addConstantPoolIndex(Idx)
315 .addReg(0).addImm(0).addImm((unsigned)Pred).addReg(PredReg);
318 void ARMRegisterInfo::reMaterialize(MachineBasicBlock &MBB,
319 MachineBasicBlock::iterator I,
321 const MachineInstr *Orig) const {
322 if (Orig->getOpcode() == ARM::MOVi2pieces) {
323 emitLoadConstPool(MBB, I, DestReg,
324 Orig->getOperand(1).getImmedValue(),
325 (ARMCC::CondCodes)Orig->getOperand(2).getImmedValue(),
326 Orig->getOperand(3).getReg(),
331 MachineInstr *MI = Orig->clone();
332 MI->getOperand(0).setReg(DestReg);
336 /// isLowRegister - Returns true if the register is low register r0-r7.
338 static bool isLowRegister(unsigned Reg) {
341 case R0: case R1: case R2: case R3:
342 case R4: case R5: case R6: case R7:
349 MachineInstr *ARMRegisterInfo::foldMemoryOperand(MachineInstr *MI,
350 unsigned OpNum, int FI) const {
351 unsigned Opc = MI->getOpcode();
352 MachineInstr *NewMI = NULL;
356 if (MI->getOperand(4).getReg() == ARM::CPSR)
357 // If it is updating CPSR, then it cannot be foled.
359 unsigned Pred = MI->getOperand(2).getImmedValue();
360 unsigned PredReg = MI->getOperand(3).getReg();
361 if (OpNum == 0) { // move -> store
362 unsigned SrcReg = MI->getOperand(1).getReg();
363 NewMI = BuildMI(TII.get(ARM::STR)).addReg(SrcReg).addFrameIndex(FI)
364 .addReg(0).addImm(0).addImm(Pred).addReg(PredReg);
365 } else { // move -> load
366 unsigned DstReg = MI->getOperand(0).getReg();
367 NewMI = BuildMI(TII.get(ARM::LDR), DstReg).addFrameIndex(FI).addReg(0)
368 .addImm(0).addImm(Pred).addReg(PredReg);
373 if (OpNum == 0) { // move -> store
374 unsigned SrcReg = MI->getOperand(1).getReg();
375 if (isPhysicalRegister(SrcReg) && !isLowRegister(SrcReg))
376 // tSpill cannot take a high register operand.
378 NewMI = BuildMI(TII.get(ARM::tSpill)).addReg(SrcReg).addFrameIndex(FI)
380 } else { // move -> load
381 unsigned DstReg = MI->getOperand(0).getReg();
382 if (isPhysicalRegister(DstReg) && !isLowRegister(DstReg))
383 // tRestore cannot target a high register operand.
385 NewMI = BuildMI(TII.get(ARM::tRestore), DstReg).addFrameIndex(FI)
391 unsigned Pred = MI->getOperand(2).getImmedValue();
392 unsigned PredReg = MI->getOperand(3).getReg();
393 if (OpNum == 0) { // move -> store
394 unsigned SrcReg = MI->getOperand(1).getReg();
395 NewMI = BuildMI(TII.get(ARM::FSTS)).addReg(SrcReg).addFrameIndex(FI)
396 .addImm(0).addImm(Pred).addReg(PredReg);
397 } else { // move -> load
398 unsigned DstReg = MI->getOperand(0).getReg();
399 NewMI = BuildMI(TII.get(ARM::FLDS), DstReg).addFrameIndex(FI)
400 .addImm(0).addImm(Pred).addReg(PredReg);
405 unsigned Pred = MI->getOperand(2).getImmedValue();
406 unsigned PredReg = MI->getOperand(3).getReg();
407 if (OpNum == 0) { // move -> store
408 unsigned SrcReg = MI->getOperand(1).getReg();
409 NewMI = BuildMI(TII.get(ARM::FSTD)).addReg(SrcReg).addFrameIndex(FI)
410 .addImm(0).addImm(Pred).addReg(PredReg);
411 } else { // move -> load
412 unsigned DstReg = MI->getOperand(0).getReg();
413 NewMI = BuildMI(TII.get(ARM::FLDD), DstReg).addFrameIndex(FI)
414 .addImm(0).addImm(Pred).addReg(PredReg);
421 NewMI->copyKillDeadInfo(MI);
426 ARMRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
427 static const unsigned CalleeSavedRegs[] = {
428 ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8,
429 ARM::R7, ARM::R6, ARM::R5, ARM::R4,
431 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
432 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
436 static const unsigned DarwinCalleeSavedRegs[] = {
437 ARM::LR, ARM::R7, ARM::R6, ARM::R5, ARM::R4,
438 ARM::R11, ARM::R10, ARM::R9, ARM::R8,
440 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
441 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
444 return STI.isTargetDarwin() ? DarwinCalleeSavedRegs : CalleeSavedRegs;
447 const TargetRegisterClass* const *
448 ARMRegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const {
449 static const TargetRegisterClass * const CalleeSavedRegClasses[] = {
450 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
451 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
452 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
454 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
455 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
458 return CalleeSavedRegClasses;
461 BitVector ARMRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
462 // FIXME: avoid re-calculating this everytime.
463 BitVector Reserved(getNumRegs());
464 Reserved.set(ARM::SP);
465 Reserved.set(ARM::PC);
466 if (STI.isTargetDarwin() || hasFP(MF))
467 Reserved.set(FramePtr);
468 // Some targets reserve R9.
469 if (STI.isR9Reserved())
470 Reserved.set(ARM::R9);
475 ARMRegisterInfo::isReservedReg(const MachineFunction &MF, unsigned Reg) const {
483 if (FramePtr == Reg && (STI.isTargetDarwin() || hasFP(MF)))
487 return STI.isR9Reserved();
494 ARMRegisterInfo::requiresRegisterScavenging(const MachineFunction &MF) const {
495 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
496 return ThumbRegScavenging || !AFI->isThumbFunction();
499 /// hasFP - Return true if the specified function should have a dedicated frame
500 /// pointer register. This is true if the function has variable sized allocas
501 /// or if frame pointer elimination is disabled.
503 bool ARMRegisterInfo::hasFP(const MachineFunction &MF) const {
504 return NoFramePointerElim || MF.getFrameInfo()->hasVarSizedObjects();
507 // hasReservedCallFrame - Under normal circumstances, when a frame pointer is
508 // not required, we reserve argument space for call sites in the function
509 // immediately on entry to the current function. This eliminates the need for
510 // add/sub sp brackets around call sites. Returns true if the call frame is
511 // included as part of the stack frame.
512 bool ARMRegisterInfo::hasReservedCallFrame(MachineFunction &MF) const {
513 const MachineFrameInfo *FFI = MF.getFrameInfo();
514 unsigned CFSize = FFI->getMaxCallFrameSize();
515 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
516 // It's not always a good idea to include the call frame as part of the
517 // stack frame. ARM (especially Thumb) has small immediate offset to
518 // address the stack frame. So a large call frame can cause poor codegen
519 // and may even makes it impossible to scavenge a register.
520 if (AFI->isThumbFunction()) {
521 if (CFSize >= ((1 << 8) - 1) * 4 / 2) // Half of imm8 * 4
524 if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12
527 return !MF.getFrameInfo()->hasVarSizedObjects();
530 /// emitARMRegPlusImmediate - Emits a series of instructions to materialize
531 /// a destreg = basereg + immediate in ARM code.
533 void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
534 MachineBasicBlock::iterator &MBBI,
535 unsigned DestReg, unsigned BaseReg, int NumBytes,
536 ARMCC::CondCodes Pred, unsigned PredReg,
537 const TargetInstrInfo &TII) {
538 bool isSub = NumBytes < 0;
539 if (isSub) NumBytes = -NumBytes;
542 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
543 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
544 assert(ThisVal && "Didn't extract field correctly");
546 // We will handle these bits from offset, clear them.
547 NumBytes &= ~ThisVal;
549 // Get the properly encoded SOImmVal field.
550 int SOImmVal = ARM_AM::getSOImmVal(ThisVal);
551 assert(SOImmVal != -1 && "Bit extraction didn't work?");
553 // Build the new ADD / SUB.
554 BuildMI(MBB, MBBI, TII.get(isSub ? ARM::SUBri : ARM::ADDri), DestReg)
555 .addReg(BaseReg, false, false, true).addImm(SOImmVal)
556 .addImm((unsigned)Pred).addReg(PredReg).addReg(0);
561 /// calcNumMI - Returns the number of instructions required to materialize
562 /// the specific add / sub r, c instruction.
563 static unsigned calcNumMI(int Opc, int ExtraOpc, unsigned Bytes,
564 unsigned NumBits, unsigned Scale) {
566 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
568 if (Opc == ARM::tADDrSPi) {
569 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
573 Scale = 1; // Followed by a number of tADDi8.
574 Chunk = ((1 << NumBits) - 1) * Scale;
577 NumMIs += Bytes / Chunk;
578 if ((Bytes % Chunk) != 0)
585 /// emitThumbRegPlusImmInReg - Emits a series of instructions to materialize
586 /// a destreg = basereg + immediate in Thumb code. Materialize the immediate
587 /// in a register using mov / mvn sequences or load the immediate from a
590 void emitThumbRegPlusImmInReg(MachineBasicBlock &MBB,
591 MachineBasicBlock::iterator &MBBI,
592 unsigned DestReg, unsigned BaseReg,
593 int NumBytes, bool CanChangeCC,
594 const TargetInstrInfo &TII) {
595 bool isHigh = !isLowRegister(DestReg) ||
596 (BaseReg != 0 && !isLowRegister(BaseReg));
598 // Subtract doesn't have high register version. Load the negative value
599 // if either base or dest register is a high register. Also, if do not
600 // issue sub as part of the sequence if condition register is to be
602 if (NumBytes < 0 && !isHigh && CanChangeCC) {
604 NumBytes = -NumBytes;
606 unsigned LdReg = DestReg;
607 if (DestReg == ARM::SP) {
608 assert(BaseReg == ARM::SP && "Unexpected!");
610 BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), ARM::R12)
611 .addReg(ARM::R3, false, false, true);
614 if (NumBytes <= 255 && NumBytes >= 0)
615 BuildMI(MBB, MBBI, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes);
616 else if (NumBytes < 0 && NumBytes >= -255) {
617 BuildMI(MBB, MBBI, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes);
618 BuildMI(MBB, MBBI, TII.get(ARM::tNEG), LdReg)
619 .addReg(LdReg, false, false, true);
621 emitLoadConstPool(MBB, MBBI, LdReg, NumBytes, ARMCC::AL, 0, TII, true);
624 int Opc = (isSub) ? ARM::tSUBrr : (isHigh ? ARM::tADDhirr : ARM::tADDrr);
625 const MachineInstrBuilder MIB = BuildMI(MBB, MBBI, TII.get(Opc), DestReg);
626 if (DestReg == ARM::SP || isSub)
627 MIB.addReg(BaseReg).addReg(LdReg, false, false, true);
629 MIB.addReg(LdReg).addReg(BaseReg, false, false, true);
630 if (DestReg == ARM::SP)
631 BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), ARM::R3)
632 .addReg(ARM::R12, false, false, true);
635 /// emitThumbRegPlusImmediate - Emits a series of instructions to materialize
636 /// a destreg = basereg + immediate in Thumb code.
638 void emitThumbRegPlusImmediate(MachineBasicBlock &MBB,
639 MachineBasicBlock::iterator &MBBI,
640 unsigned DestReg, unsigned BaseReg,
641 int NumBytes, const TargetInstrInfo &TII) {
642 bool isSub = NumBytes < 0;
643 unsigned Bytes = (unsigned)NumBytes;
644 if (isSub) Bytes = -NumBytes;
645 bool isMul4 = (Bytes & 3) == 0;
646 bool isTwoAddr = false;
647 bool DstNotEqBase = false;
648 unsigned NumBits = 1;
653 if (DestReg == BaseReg && BaseReg == ARM::SP) {
654 assert(isMul4 && "Thumb sp inc / dec size must be multiple of 4!");
657 Opc = isSub ? ARM::tSUBspi : ARM::tADDspi;
659 } else if (!isSub && BaseReg == ARM::SP) {
662 // r1 = add sp, 100 * 4
666 ExtraOpc = ARM::tADDi3;
675 if (DestReg != BaseReg)
678 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
682 unsigned NumMIs = calcNumMI(Opc, ExtraOpc, Bytes, NumBits, Scale);
683 unsigned Threshold = (DestReg == ARM::SP) ? 3 : 2;
684 if (NumMIs > Threshold) {
685 // This will expand into too many instructions. Load the immediate from a
687 emitThumbRegPlusImmInReg(MBB, MBBI, DestReg, BaseReg, NumBytes, true, TII);
692 if (isLowRegister(DestReg) && isLowRegister(BaseReg)) {
693 // If both are low registers, emit DestReg = add BaseReg, max(Imm, 7)
694 unsigned Chunk = (1 << 3) - 1;
695 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
697 BuildMI(MBB, MBBI, TII.get(isSub ? ARM::tSUBi3 : ARM::tADDi3), DestReg)
698 .addReg(BaseReg, false, false, true).addImm(ThisVal);
700 BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), DestReg)
701 .addReg(BaseReg, false, false, true);
706 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
708 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
711 // Build the new tADD / tSUB.
713 BuildMI(MBB, MBBI, TII.get(Opc), DestReg).addReg(DestReg).addImm(ThisVal);
715 bool isKill = BaseReg != ARM::SP;
716 BuildMI(MBB, MBBI, TII.get(Opc), DestReg)
717 .addReg(BaseReg, false, false, isKill).addImm(ThisVal);
720 if (Opc == ARM::tADDrSPi) {
726 Chunk = ((1 << NumBits) - 1) * Scale;
727 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
734 BuildMI(MBB, MBBI, TII.get(ExtraOpc), DestReg)
735 .addReg(DestReg, false, false, true)
736 .addImm(((unsigned)NumBytes) & 3);
740 void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
741 int NumBytes, ARMCC::CondCodes Pred, unsigned PredReg,
742 bool isThumb, const TargetInstrInfo &TII) {
744 emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII);
746 emitARMRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes,
750 void ARMRegisterInfo::
751 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
752 MachineBasicBlock::iterator I) const {
753 if (!hasReservedCallFrame(MF)) {
754 // If we have alloca, convert as follows:
755 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
756 // ADJCALLSTACKUP -> add, sp, sp, amount
757 MachineInstr *Old = I;
758 unsigned Amount = Old->getOperand(0).getImmedValue();
760 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
761 // We need to keep the stack aligned properly. To do this, we round the
762 // amount of space needed for the outgoing arguments up to the next
763 // alignment boundary.
764 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
765 Amount = (Amount+Align-1)/Align*Align;
767 // Replace the pseudo instruction with a new instruction...
768 unsigned Opc = Old->getOpcode();
769 bool isThumb = AFI->isThumbFunction();
770 ARMCC::CondCodes Pred = isThumb
771 ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(1).getImmedValue();
772 if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) {
773 // Note: PredReg is operand 2 for ADJCALLSTACKDOWN.
774 unsigned PredReg = isThumb ? 0 : Old->getOperand(2).getReg();
775 emitSPUpdate(MBB, I, -Amount, Pred, PredReg, isThumb, TII);
777 // Note: PredReg is operand 3 for ADJCALLSTACKUP.
778 unsigned PredReg = isThumb ? 0 : Old->getOperand(3).getReg();
779 assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP);
780 emitSPUpdate(MBB, I, Amount, Pred, PredReg, isThumb, TII);
787 /// emitThumbConstant - Emit a series of instructions to materialize a
789 static void emitThumbConstant(MachineBasicBlock &MBB,
790 MachineBasicBlock::iterator &MBBI,
791 unsigned DestReg, int Imm,
792 const TargetInstrInfo &TII) {
793 bool isSub = Imm < 0;
794 if (isSub) Imm = -Imm;
796 int Chunk = (1 << 8) - 1;
797 int ThisVal = (Imm > Chunk) ? Chunk : Imm;
799 BuildMI(MBB, MBBI, TII.get(ARM::tMOVi8), DestReg).addImm(ThisVal);
801 emitThumbRegPlusImmediate(MBB, MBBI, DestReg, DestReg, Imm, TII);
803 BuildMI(MBB, MBBI, TII.get(ARM::tNEG), DestReg)
804 .addReg(DestReg, false, false, true);
807 /// findScratchRegister - Find a 'free' ARM register. If register scavenger
808 /// is not being used, R12 is available. Otherwise, try for a call-clobbered
809 /// register first and then a spilled callee-saved register if that fails.
811 unsigned findScratchRegister(RegScavenger *RS, const TargetRegisterClass *RC,
812 ARMFunctionInfo *AFI) {
813 unsigned Reg = RS ? RS->FindUnusedReg(RC, true) : (unsigned) ARM::R12;
815 // Try a already spilled CS register.
816 Reg = RS->FindUnusedReg(RC, AFI->getSpilledCSRegisters());
821 void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
822 int SPAdj, RegScavenger *RS) const{
824 MachineInstr &MI = *II;
825 MachineBasicBlock &MBB = *MI.getParent();
826 MachineFunction &MF = *MBB.getParent();
827 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
828 bool isThumb = AFI->isThumbFunction();
830 while (!MI.getOperand(i).isFrameIndex()) {
832 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
835 unsigned FrameReg = ARM::SP;
836 int FrameIndex = MI.getOperand(i).getFrameIndex();
837 int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex) +
838 MF.getFrameInfo()->getStackSize() + SPAdj;
840 if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex))
841 Offset -= AFI->getGPRCalleeSavedArea1Offset();
842 else if (AFI->isGPRCalleeSavedArea2Frame(FrameIndex))
843 Offset -= AFI->getGPRCalleeSavedArea2Offset();
844 else if (AFI->isDPRCalleeSavedAreaFrame(FrameIndex))
845 Offset -= AFI->getDPRCalleeSavedAreaOffset();
846 else if (hasFP(MF)) {
847 assert(SPAdj == 0 && "Unexpected");
848 // There is alloca()'s in this function, must reference off the frame
850 FrameReg = getFrameRegister(MF);
851 Offset -= AFI->getFramePtrSpillOffset();
854 unsigned Opcode = MI.getOpcode();
855 const TargetInstrDescriptor &Desc = TII.get(Opcode);
856 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
859 if (Opcode == ARM::ADDri) {
860 Offset += MI.getOperand(i+1).getImm();
862 // Turn it into a move.
863 MI.setInstrDescriptor(TII.get(ARM::MOVr));
864 MI.getOperand(i).ChangeToRegister(FrameReg, false);
865 MI.RemoveOperand(i+1);
867 } else if (Offset < 0) {
870 MI.setInstrDescriptor(TII.get(ARM::SUBri));
873 // Common case: small offset, fits into instruction.
874 int ImmedOffset = ARM_AM::getSOImmVal(Offset);
875 if (ImmedOffset != -1) {
876 // Replace the FrameIndex with sp / fp
877 MI.getOperand(i).ChangeToRegister(FrameReg, false);
878 MI.getOperand(i+1).ChangeToImmediate(ImmedOffset);
882 // Otherwise, we fallback to common code below to form the imm offset with
883 // a sequence of ADDri instructions. First though, pull as much of the imm
884 // into this ADDri as possible.
885 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
886 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
888 // We will handle these bits from offset, clear them.
889 Offset &= ~ThisImmVal;
891 // Get the properly encoded SOImmVal field.
892 int ThisSOImmVal = ARM_AM::getSOImmVal(ThisImmVal);
893 assert(ThisSOImmVal != -1 && "Bit extraction didn't work?");
894 MI.getOperand(i+1).ChangeToImmediate(ThisSOImmVal);
895 } else if (Opcode == ARM::tADDrSPi) {
896 Offset += MI.getOperand(i+1).getImm();
898 // Can't use tADDrSPi if it's based off the frame pointer.
899 unsigned NumBits = 0;
901 if (FrameReg != ARM::SP) {
902 Opcode = ARM::tADDi3;
903 MI.setInstrDescriptor(TII.get(ARM::tADDi3));
908 assert((Offset & 3) == 0 &&
909 "Thumb add/sub sp, #imm immediate must be multiple of 4!");
913 // Turn it into a move.
914 MI.setInstrDescriptor(TII.get(ARM::tMOVr));
915 MI.getOperand(i).ChangeToRegister(FrameReg, false);
916 MI.RemoveOperand(i+1);
920 // Common case: small offset, fits into instruction.
921 unsigned Mask = (1 << NumBits) - 1;
922 if (((Offset / Scale) & ~Mask) == 0) {
923 // Replace the FrameIndex with sp / fp
924 MI.getOperand(i).ChangeToRegister(FrameReg, false);
925 MI.getOperand(i+1).ChangeToImmediate(Offset / Scale);
929 unsigned DestReg = MI.getOperand(0).getReg();
930 unsigned Bytes = (Offset > 0) ? Offset : -Offset;
931 unsigned NumMIs = calcNumMI(Opcode, 0, Bytes, NumBits, Scale);
932 // MI would expand into a large number of instructions. Don't try to
933 // simplify the immediate.
935 emitThumbRegPlusImmediate(MBB, II, DestReg, FrameReg, Offset, TII);
941 // Translate r0 = add sp, imm to
942 // r0 = add sp, 255*4
943 // r0 = add r0, (imm - 255*4)
944 MI.getOperand(i).ChangeToRegister(FrameReg, false);
945 MI.getOperand(i+1).ChangeToImmediate(Mask);
946 Offset = (Offset - Mask * Scale);
947 MachineBasicBlock::iterator NII = next(II);
948 emitThumbRegPlusImmediate(MBB, NII, DestReg, DestReg, Offset, TII);
950 // Translate r0 = add sp, -imm to
951 // r0 = -imm (this is then translated into a series of instructons)
953 emitThumbConstant(MBB, II, DestReg, Offset, TII);
954 MI.setInstrDescriptor(TII.get(ARM::tADDhirr));
955 MI.getOperand(i).ChangeToRegister(DestReg, false, false, true);
956 MI.getOperand(i+1).ChangeToRegister(FrameReg, false);
962 unsigned NumBits = 0;
965 case ARMII::AddrMode2: {
967 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
968 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
973 case ARMII::AddrMode3: {
975 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
976 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
981 case ARMII::AddrMode5: {
983 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
984 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
990 case ARMII::AddrModeTs: {
992 InstrOffs = MI.getOperand(ImmIdx).getImm();
993 NumBits = (FrameReg == ARM::SP) ? 8 : 5;
998 assert(0 && "Unsupported addressing mode!");
1003 Offset += InstrOffs * Scale;
1004 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1005 if (Offset < 0 && !isThumb) {
1010 // Common case: small offset, fits into instruction.
1011 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1012 int ImmedOffset = Offset / Scale;
1013 unsigned Mask = (1 << NumBits) - 1;
1014 if ((unsigned)Offset <= Mask * Scale) {
1015 // Replace the FrameIndex with sp
1016 MI.getOperand(i).ChangeToRegister(FrameReg, false);
1018 ImmedOffset |= 1 << NumBits;
1019 ImmOp.ChangeToImmediate(ImmedOffset);
1023 bool isThumSpillRestore = Opcode == ARM::tRestore || Opcode == ARM::tSpill;
1024 if (AddrMode == ARMII::AddrModeTs) {
1025 // Thumb tLDRspi, tSTRspi. These will change to instructions that use
1026 // a different base register.
1028 Mask = (1 << NumBits) - 1;
1030 // If this is a thumb spill / restore, we will be using a constpool load to
1031 // materialize the offset.
1032 if (AddrMode == ARMII::AddrModeTs && isThumSpillRestore)
1033 ImmOp.ChangeToImmediate(0);
1035 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1036 ImmedOffset = ImmedOffset & Mask;
1038 ImmedOffset |= 1 << NumBits;
1039 ImmOp.ChangeToImmediate(ImmedOffset);
1040 Offset &= ~(Mask*Scale);
1044 // If we get here, the immediate doesn't fit into the instruction. We folded
1045 // as much as possible above, handle the rest, providing a register that is
1047 assert(Offset && "This code isn't needed if offset already handled!");
1050 if (TII.isLoad(Opcode)) {
1051 // Use the destination register to materialize sp + offset.
1052 unsigned TmpReg = MI.getOperand(0).getReg();
1054 if (Opcode == ARM::tRestore) {
1055 if (FrameReg == ARM::SP)
1056 emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg,Offset,false,TII);
1058 emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, TII, true);
1062 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII);
1063 MI.setInstrDescriptor(TII.get(ARM::tLDR));
1064 MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true);
1066 MI.addRegOperand(FrameReg, false); // Use [reg, reg] addrmode.
1068 MI.addRegOperand(0, false); // tLDR has an extra register operand.
1069 } else if (TII.isStore(Opcode)) {
1070 // FIXME! This is horrific!!! We need register scavenging.
1071 // Our temporary workaround has marked r3 unavailable. Of course, r3 is
1072 // also a ABI register so it's possible that is is the register that is
1073 // being storing here. If that's the case, we do the following:
1075 // Use r2 to materialize sp + offset
1078 unsigned ValReg = MI.getOperand(0).getReg();
1079 unsigned TmpReg = ARM::R3;
1081 if (ValReg == ARM::R3) {
1082 BuildMI(MBB, II, TII.get(ARM::tMOVr), ARM::R12)
1083 .addReg(ARM::R2, false, false, true);
1086 if (TmpReg == ARM::R3 && AFI->isR3LiveIn())
1087 BuildMI(MBB, II, TII.get(ARM::tMOVr), ARM::R12)
1088 .addReg(ARM::R3, false, false, true);
1089 if (Opcode == ARM::tSpill) {
1090 if (FrameReg == ARM::SP)
1091 emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg,Offset,false,TII);
1093 emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, TII, true);
1097 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII);
1098 MI.setInstrDescriptor(TII.get(ARM::tSTR));
1099 MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true);
1101 MI.addRegOperand(FrameReg, false); // Use [reg, reg] addrmode.
1103 MI.addRegOperand(0, false); // tSTR has an extra register operand.
1105 MachineBasicBlock::iterator NII = next(II);
1106 if (ValReg == ARM::R3)
1107 BuildMI(MBB, NII, TII.get(ARM::tMOVr), ARM::R2)
1108 .addReg(ARM::R12, false, false, true);
1109 if (TmpReg == ARM::R3 && AFI->isR3LiveIn())
1110 BuildMI(MBB, NII, TII.get(ARM::tMOVr), ARM::R3)
1111 .addReg(ARM::R12, false, false, true);
1113 assert(false && "Unexpected opcode!");
1115 // Insert a set of r12 with the full address: r12 = sp + offset
1116 // If the offset we have is too large to fit into the instruction, we need
1117 // to form it with a series of ADDri's. Do this by taking 8-bit chunks
1119 unsigned ScratchReg = findScratchRegister(RS, &ARM::GPRRegClass, AFI);
1120 if (ScratchReg == 0)
1121 // No register is "free". Scavenge a register.
1122 ScratchReg = RS->scavengeRegister(&ARM::GPRRegClass, II, SPAdj);
1123 int PIdx = MI.findFirstPredOperandIdx();
1124 ARMCC::CondCodes Pred = (PIdx == -1)
1125 ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImmedValue();
1126 unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg();
1127 emitARMRegPlusImmediate(MBB, II, ScratchReg, FrameReg,
1128 isSub ? -Offset : Offset, Pred, PredReg, TII);
1129 MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true);
1133 static unsigned estimateStackSize(MachineFunction &MF, MachineFrameInfo *MFI) {
1134 const MachineFrameInfo *FFI = MF.getFrameInfo();
1136 for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) {
1137 int FixedOff = -FFI->getObjectOffset(i);
1138 if (FixedOff > Offset) Offset = FixedOff;
1140 for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) {
1141 Offset += FFI->getObjectSize(i);
1142 unsigned Align = FFI->getObjectAlignment(i);
1143 // Adjust to alignment boundary
1144 Offset = (Offset+Align-1)/Align*Align;
1146 return (unsigned)Offset;
1150 ARMRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
1151 RegScavenger *RS) const {
1152 // This tells PEI to spill the FP as if it is any other callee-save register
1153 // to take advantage the eliminateFrameIndex machinery. This also ensures it
1154 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
1155 // to combine multiple loads / stores.
1156 bool CanEliminateFrame = true;
1157 bool CS1Spilled = false;
1158 bool LRSpilled = false;
1159 unsigned NumGPRSpills = 0;
1160 SmallVector<unsigned, 4> UnspilledCS1GPRs;
1161 SmallVector<unsigned, 4> UnspilledCS2GPRs;
1162 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1164 // Don't spill FP if the frame can be eliminated. This is determined
1165 // by scanning the callee-save registers to see if any is used.
1166 const unsigned *CSRegs = getCalleeSavedRegs();
1167 const TargetRegisterClass* const *CSRegClasses = getCalleeSavedRegClasses();
1168 for (unsigned i = 0; CSRegs[i]; ++i) {
1169 unsigned Reg = CSRegs[i];
1170 bool Spilled = false;
1171 if (MF.isPhysRegUsed(Reg)) {
1172 AFI->setCSRegisterIsSpilled(Reg);
1174 CanEliminateFrame = false;
1176 // Check alias registers too.
1177 for (const unsigned *Aliases = getAliasSet(Reg); *Aliases; ++Aliases) {
1178 if (MF.isPhysRegUsed(*Aliases)) {
1180 CanEliminateFrame = false;
1185 if (CSRegClasses[i] == &ARM::GPRRegClass) {
1189 if (!STI.isTargetDarwin()) {
1196 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
1211 if (!STI.isTargetDarwin()) {
1212 UnspilledCS1GPRs.push_back(Reg);
1222 UnspilledCS1GPRs.push_back(Reg);
1225 UnspilledCS2GPRs.push_back(Reg);
1232 bool ForceLRSpill = false;
1233 if (!LRSpilled && AFI->isThumbFunction()) {
1234 unsigned FnSize = ARM::GetFunctionSize(MF);
1235 // Force LR to be spilled if the Thumb function size is > 2048. This enables
1236 // use of BL to implement far jump. If it turns out that it's not needed
1237 // then the branch fix up path will undo it.
1238 if (FnSize >= (1 << 11)) {
1239 CanEliminateFrame = false;
1240 ForceLRSpill = true;
1244 bool ExtraCSSpill = false;
1245 if (!CanEliminateFrame || hasFP(MF)) {
1246 AFI->setHasStackFrame(true);
1248 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
1249 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
1250 if (!LRSpilled && CS1Spilled) {
1251 MF.setPhysRegUsed(ARM::LR);
1252 AFI->setCSRegisterIsSpilled(ARM::LR);
1254 UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(),
1255 UnspilledCS1GPRs.end(), (unsigned)ARM::LR));
1256 ForceLRSpill = false;
1257 ExtraCSSpill = true;
1260 // Darwin ABI requires FP to point to the stack slot that contains the
1262 if (STI.isTargetDarwin() || hasFP(MF)) {
1263 MF.setPhysRegUsed(FramePtr);
1267 // If stack and double are 8-byte aligned and we are spilling an odd number
1268 // of GPRs. Spill one extra callee save GPR so we won't have to pad between
1269 // the integer and double callee save areas.
1270 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
1271 if (TargetAlign == 8 && (NumGPRSpills & 1)) {
1272 if (CS1Spilled && !UnspilledCS1GPRs.empty()) {
1273 for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) {
1274 unsigned Reg = UnspilledCS1GPRs[i];
1275 // Don't spiil high register if the function is thumb
1276 if (!AFI->isThumbFunction() || isLowRegister(Reg) || Reg == ARM::LR) {
1277 MF.setPhysRegUsed(Reg);
1278 AFI->setCSRegisterIsSpilled(Reg);
1279 if (!isReservedReg(MF, Reg))
1280 ExtraCSSpill = true;
1284 } else if (!UnspilledCS2GPRs.empty() &&
1285 !AFI->isThumbFunction()) {
1286 unsigned Reg = UnspilledCS2GPRs.front();
1287 MF.setPhysRegUsed(Reg);
1288 AFI->setCSRegisterIsSpilled(Reg);
1289 if (!isReservedReg(MF, Reg))
1290 ExtraCSSpill = true;
1294 // Estimate if we might need to scavenge a register at some point in order
1295 // to materialize a stack offset. If so, either spill one additiona
1296 // callee-saved register or reserve a special spill slot to facilitate
1297 // register scavenging.
1298 if (RS && !ExtraCSSpill && !AFI->isThumbFunction()) {
1299 MachineFrameInfo *MFI = MF.getFrameInfo();
1300 unsigned Size = estimateStackSize(MF, MFI);
1301 unsigned Limit = (1 << 12) - 1;
1302 for (MachineFunction::iterator BB = MF.begin(),E = MF.end();BB != E; ++BB)
1303 for (MachineBasicBlock::iterator I= BB->begin(); I != BB->end(); ++I) {
1304 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1305 if (I->getOperand(i).isFrameIndex()) {
1306 unsigned Opcode = I->getOpcode();
1307 const TargetInstrDescriptor &Desc = TII.get(Opcode);
1308 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1309 if (AddrMode == ARMII::AddrMode3) {
1310 Limit = (1 << 8) - 1;
1311 goto DoneEstimating;
1312 } else if (AddrMode == ARMII::AddrMode5) {
1313 unsigned ThisLimit = ((1 << 8) - 1) * 4;
1314 if (ThisLimit < Limit)
1320 if (Size >= Limit) {
1321 // If any non-reserved CS register isn't spilled, just spill one or two
1322 // extra. That should take care of it!
1323 unsigned NumExtras = TargetAlign / 4;
1324 SmallVector<unsigned, 2> Extras;
1325 while (NumExtras && !UnspilledCS1GPRs.empty()) {
1326 unsigned Reg = UnspilledCS1GPRs.back();
1327 UnspilledCS1GPRs.pop_back();
1328 if (!isReservedReg(MF, Reg)) {
1329 Extras.push_back(Reg);
1333 while (NumExtras && !UnspilledCS2GPRs.empty()) {
1334 unsigned Reg = UnspilledCS2GPRs.back();
1335 UnspilledCS2GPRs.pop_back();
1336 if (!isReservedReg(MF, Reg)) {
1337 Extras.push_back(Reg);
1341 if (Extras.size() && NumExtras == 0) {
1342 for (unsigned i = 0, e = Extras.size(); i != e; ++i) {
1343 MF.setPhysRegUsed(Extras[i]);
1344 AFI->setCSRegisterIsSpilled(Extras[i]);
1347 // Reserve a slot closest to SP or frame pointer.
1348 const TargetRegisterClass *RC = &ARM::GPRRegClass;
1349 RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
1350 RC->getAlignment()));
1357 MF.setPhysRegUsed(ARM::LR);
1358 AFI->setCSRegisterIsSpilled(ARM::LR);
1359 AFI->setLRIsSpilledForFarJump(true);
1363 /// Move iterator pass the next bunch of callee save load / store ops for
1364 /// the particular spill area (1: integer area 1, 2: integer area 2,
1365 /// 3: fp area, 0: don't care).
1366 static void movePastCSLoadStoreOps(MachineBasicBlock &MBB,
1367 MachineBasicBlock::iterator &MBBI,
1368 int Opc, unsigned Area,
1369 const ARMSubtarget &STI) {
1370 while (MBBI != MBB.end() &&
1371 MBBI->getOpcode() == Opc && MBBI->getOperand(1).isFrameIndex()) {
1374 unsigned Category = 0;
1375 switch (MBBI->getOperand(0).getReg()) {
1376 case ARM::R4: case ARM::R5: case ARM::R6: case ARM::R7:
1380 case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11:
1381 Category = STI.isTargetDarwin() ? 2 : 1;
1383 case ARM::D8: case ARM::D9: case ARM::D10: case ARM::D11:
1384 case ARM::D12: case ARM::D13: case ARM::D14: case ARM::D15:
1391 if (Done || Category != Area)
1399 void ARMRegisterInfo::emitPrologue(MachineFunction &MF) const {
1400 MachineBasicBlock &MBB = MF.front();
1401 MachineBasicBlock::iterator MBBI = MBB.begin();
1402 MachineFrameInfo *MFI = MF.getFrameInfo();
1403 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1404 bool isThumb = AFI->isThumbFunction();
1405 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1406 unsigned NumBytes = MFI->getStackSize();
1407 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
1410 // Check if R3 is live in. It might have to be used as a scratch register.
1411 for (MachineFunction::livein_iterator I=MF.livein_begin(),E=MF.livein_end();
1413 if ((*I).first == ARM::R3) {
1414 AFI->setR3IsLiveIn(true);
1419 // Thumb add/sub sp, imm8 instructions implicitly multiply the offset by 4.
1420 NumBytes = (NumBytes + 3) & ~3;
1421 MFI->setStackSize(NumBytes);
1424 // Determine the sizes of each callee-save spill areas and record which frame
1425 // belongs to which callee-save spill areas.
1426 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
1427 int FramePtrSpillFI = 0;
1430 emitSPUpdate(MBB, MBBI, -VARegSaveSize, ARMCC::AL, 0, isThumb, TII);
1432 if (!AFI->hasStackFrame()) {
1434 emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, isThumb, TII);
1438 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1439 unsigned Reg = CSI[i].getReg();
1440 int FI = CSI[i].getFrameIdx();
1447 if (Reg == FramePtr)
1448 FramePtrSpillFI = FI;
1449 AFI->addGPRCalleeSavedArea1Frame(FI);
1456 if (Reg == FramePtr)
1457 FramePtrSpillFI = FI;
1458 if (STI.isTargetDarwin()) {
1459 AFI->addGPRCalleeSavedArea2Frame(FI);
1462 AFI->addGPRCalleeSavedArea1Frame(FI);
1467 AFI->addDPRCalleeSavedAreaFrame(FI);
1473 // Build the new SUBri to adjust SP for integer callee-save spill area 1.
1474 emitSPUpdate(MBB, MBBI, -GPRCS1Size, ARMCC::AL, 0, isThumb, TII);
1475 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 1, STI);
1476 } else if (MBBI != MBB.end() && MBBI->getOpcode() == ARM::tPUSH)
1479 // Darwin ABI requires FP to point to the stack slot that contains the
1481 if (STI.isTargetDarwin() || hasFP(MF)) {
1482 MachineInstrBuilder MIB =
1483 BuildMI(MBB, MBBI, TII.get(isThumb ? ARM::tADDrSPi : ARM::ADDri),FramePtr)
1484 .addFrameIndex(FramePtrSpillFI).addImm(0);
1485 if (!isThumb) AddDefaultCC(AddDefaultPred(MIB));
1489 // Build the new SUBri to adjust SP for integer callee-save spill area 2.
1490 emitSPUpdate(MBB, MBBI, -GPRCS2Size, ARMCC::AL, 0, false, TII);
1492 // Build the new SUBri to adjust SP for FP callee-save spill area.
1493 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 2, STI);
1494 emitSPUpdate(MBB, MBBI, -DPRCSSize, ARMCC::AL, 0, false, TII);
1497 // Determine starting offsets of spill areas.
1498 unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize);
1499 unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize;
1500 unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size;
1501 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) + NumBytes);
1502 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
1503 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
1504 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
1506 NumBytes = DPRCSOffset;
1508 // Insert it after all the callee-save spills.
1510 movePastCSLoadStoreOps(MBB, MBBI, ARM::FSTD, 3, STI);
1511 emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, isThumb, TII);
1514 if(STI.isTargetELF() && hasFP(MF)) {
1515 MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() -
1516 AFI->getFramePtrSpillOffset());
1519 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
1520 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
1521 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
1524 static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) {
1525 for (unsigned i = 0; CSRegs[i]; ++i)
1526 if (Reg == CSRegs[i])
1531 static bool isCSRestore(MachineInstr *MI, const unsigned *CSRegs) {
1532 return ((MI->getOpcode() == ARM::FLDD ||
1533 MI->getOpcode() == ARM::LDR ||
1534 MI->getOpcode() == ARM::tRestore) &&
1535 MI->getOperand(1).isFrameIndex() &&
1536 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs));
1539 void ARMRegisterInfo::emitEpilogue(MachineFunction &MF,
1540 MachineBasicBlock &MBB) const {
1541 MachineBasicBlock::iterator MBBI = prior(MBB.end());
1542 assert((MBBI->getOpcode() == ARM::BX_RET ||
1543 MBBI->getOpcode() == ARM::tBX_RET ||
1544 MBBI->getOpcode() == ARM::tPOP_RET) &&
1545 "Can only insert epilog into returning blocks");
1547 MachineFrameInfo *MFI = MF.getFrameInfo();
1548 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1549 bool isThumb = AFI->isThumbFunction();
1550 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1551 int NumBytes = (int)MFI->getStackSize();
1552 if (!AFI->hasStackFrame()) {
1554 emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, isThumb, TII);
1556 // Unwind MBBI to point to first LDR / FLDD.
1557 const unsigned *CSRegs = getCalleeSavedRegs();
1558 if (MBBI != MBB.begin()) {
1561 while (MBBI != MBB.begin() && isCSRestore(MBBI, CSRegs));
1562 if (!isCSRestore(MBBI, CSRegs))
1566 // Move SP to start of FP callee save spill area.
1567 NumBytes -= (AFI->getGPRCalleeSavedArea1Size() +
1568 AFI->getGPRCalleeSavedArea2Size() +
1569 AFI->getDPRCalleeSavedAreaSize());
1572 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1573 // Reset SP based on frame pointer only if the stack frame extends beyond
1574 // frame pointer stack slot or target is ELF and the function has FP.
1576 emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, FramePtr, -NumBytes, TII);
1578 BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), ARM::SP).addReg(FramePtr);
1580 if (MBBI->getOpcode() == ARM::tBX_RET &&
1581 &MBB.front() != MBBI &&
1582 prior(MBBI)->getOpcode() == ARM::tPOP) {
1583 MachineBasicBlock::iterator PMBBI = prior(MBBI);
1584 emitSPUpdate(MBB, PMBBI, NumBytes, ARMCC::AL, 0, isThumb, TII);
1586 emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, isThumb, TII);
1589 // Darwin ABI requires FP to point to the stack slot that contains the
1591 if ((STI.isTargetDarwin() && NumBytes) || hasFP(MF)) {
1592 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1593 // Reset SP based on frame pointer only if the stack frame extends beyond
1594 // frame pointer stack slot or target is ELF and the function has FP.
1595 if (AFI->getGPRCalleeSavedArea2Size() ||
1596 AFI->getDPRCalleeSavedAreaSize() ||
1597 AFI->getDPRCalleeSavedAreaOffset()||
1600 BuildMI(MBB, MBBI, TII.get(ARM::SUBri), ARM::SP).addReg(FramePtr)
1602 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
1604 BuildMI(MBB, MBBI, TII.get(ARM::MOVr), ARM::SP).addReg(FramePtr)
1605 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
1606 } else if (NumBytes) {
1607 emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, false, TII);
1610 // Move SP to start of integer callee save spill area 2.
1611 movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 3, STI);
1612 emitSPUpdate(MBB, MBBI, AFI->getDPRCalleeSavedAreaSize(), ARMCC::AL, 0,
1615 // Move SP to start of integer callee save spill area 1.
1616 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 2, STI);
1617 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea2Size(), ARMCC::AL, 0,
1620 // Move SP to SP upon entry to the function.
1621 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 1, STI);
1622 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea1Size(), ARMCC::AL, 0,
1627 if (VARegSaveSize) {
1629 // Epilogue for vararg functions: pop LR to R3 and branch off it.
1630 // FIXME: Verify this is still ok when R3 is no longer being reserved.
1631 BuildMI(MBB, MBBI, TII.get(ARM::tPOP)).addReg(ARM::R3);
1633 emitSPUpdate(MBB, MBBI, VARegSaveSize, ARMCC::AL, 0, isThumb, TII);
1636 BuildMI(MBB, MBBI, TII.get(ARM::tBX_RET_vararg)).addReg(ARM::R3);
1642 unsigned ARMRegisterInfo::getRARegister() const {
1646 unsigned ARMRegisterInfo::getFrameRegister(MachineFunction &MF) const {
1647 if (STI.isTargetDarwin() || hasFP(MF))
1648 return (STI.useThumbBacktraces() || STI.isThumb()) ? ARM::R7 : ARM::R11;
1653 unsigned ARMRegisterInfo::getEHExceptionRegister() const {
1654 assert(0 && "What is the exception register");
1658 unsigned ARMRegisterInfo::getEHHandlerRegister() const {
1659 assert(0 && "What is the exception handler register");
1663 int ARMRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const {
1664 assert(0 && "What is the dwarf register number");
1668 #include "ARMGenRegisterInfo.inc"