1 //===- ARMRegisterInfo.cpp - ARM Register Information -----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the ARM implementation of the MRegisterInfo class.
12 //===----------------------------------------------------------------------===//
15 #include "ARMAddressingModes.h"
16 #include "ARMInstrInfo.h"
17 #include "ARMMachineFunctionInfo.h"
18 #include "ARMRegisterInfo.h"
19 #include "ARMSubtarget.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/CodeGen/MachineConstantPool.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineLocation.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/RegisterScavenging.h"
29 #include "llvm/Target/TargetFrameInfo.h"
30 #include "llvm/Target/TargetMachine.h"
31 #include "llvm/Target/TargetOptions.h"
32 #include "llvm/ADT/BitVector.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/STLExtras.h"
35 #include "llvm/Support/CommandLine.h"
39 static cl::opt<bool> ThumbRegScavenging("enable-thumb-reg-scavenging",
41 cl::desc("Enable register scavenging on Thumb"));
43 unsigned ARMRegisterInfo::getRegisterNumbering(unsigned RegEnum) {
46 case R0: case S0: case D0: return 0;
47 case R1: case S1: case D1: return 1;
48 case R2: case S2: case D2: return 2;
49 case R3: case S3: case D3: return 3;
50 case R4: case S4: case D4: return 4;
51 case R5: case S5: case D5: return 5;
52 case R6: case S6: case D6: return 6;
53 case R7: case S7: case D7: return 7;
54 case R8: case S8: case D8: return 8;
55 case R9: case S9: case D9: return 9;
56 case R10: case S10: case D10: return 10;
57 case R11: case S11: case D11: return 11;
58 case R12: case S12: case D12: return 12;
59 case SP: case S13: case D13: return 13;
60 case LR: case S14: case D14: return 14;
61 case PC: case S15: case D15: return 15;
79 assert(0 && "Unknown ARM register!");
84 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->addOperand(MachineOperand::CreateReg(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.getIndex());
154 assert(0 && "Unknown operand for ARMInstrAddOperand!");
159 void ARMRegisterInfo::
160 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
161 unsigned SrcReg, bool isKill, 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, isKill)
168 .addFrameIndex(FI).addImm(0);
170 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::STR))
171 .addReg(SrcReg, false, false, isKill)
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, isKill)
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, isKill)
181 .addFrameIndex(FI).addImm(0));
185 void ARMRegisterInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
187 SmallVectorImpl<MachineOperand> &Addr,
188 const TargetRegisterClass *RC,
189 SmallVectorImpl<MachineInstr*> &NewMIs) const {
191 if (RC == ARM::GPRRegisterClass) {
192 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
193 if (AFI->isThumbFunction()) {
194 Opc = Addr[0].isFrameIndex() ? ARM::tSpill : ARM::tSTR;
195 MachineInstrBuilder MIB =
196 BuildMI(TII.get(Opc)).addReg(SrcReg, false, false, isKill);
197 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
198 MIB = ARMInstrAddOperand(MIB, Addr[i]);
199 NewMIs.push_back(MIB);
203 } else if (RC == ARM::DPRRegisterClass) {
206 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
210 MachineInstrBuilder MIB =
211 BuildMI(TII.get(Opc)).addReg(SrcReg, false, false, isKill);
212 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
213 MIB = ARMInstrAddOperand(MIB, Addr[i]);
215 NewMIs.push_back(MIB);
219 void ARMRegisterInfo::
220 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
221 unsigned DestReg, int FI,
222 const TargetRegisterClass *RC) const {
223 if (RC == ARM::GPRRegisterClass) {
224 MachineFunction &MF = *MBB.getParent();
225 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
226 if (AFI->isThumbFunction())
227 BuildMI(MBB, I, TII.get(ARM::tRestore), DestReg)
228 .addFrameIndex(FI).addImm(0);
230 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::LDR), DestReg)
231 .addFrameIndex(FI).addReg(0).addImm(0));
232 } else if (RC == ARM::DPRRegisterClass) {
233 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FLDD), DestReg)
234 .addFrameIndex(FI).addImm(0));
236 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
237 AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FLDS), DestReg)
238 .addFrameIndex(FI).addImm(0));
242 void ARMRegisterInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
243 SmallVectorImpl<MachineOperand> &Addr,
244 const TargetRegisterClass *RC,
245 SmallVectorImpl<MachineInstr*> &NewMIs) const {
247 if (RC == ARM::GPRRegisterClass) {
248 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
249 if (AFI->isThumbFunction()) {
250 Opc = Addr[0].isFrameIndex() ? ARM::tRestore : ARM::tLDR;
251 MachineInstrBuilder MIB = BuildMI(TII.get(Opc), DestReg);
252 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
253 MIB = ARMInstrAddOperand(MIB, Addr[i]);
254 NewMIs.push_back(MIB);
258 } else if (RC == ARM::DPRRegisterClass) {
261 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
265 MachineInstrBuilder MIB = BuildMI(TII.get(Opc), DestReg);
266 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
267 MIB = ARMInstrAddOperand(MIB, Addr[i]);
269 NewMIs.push_back(MIB);
273 /// emitLoadConstPool - Emits a load from constpool to materialize the
274 /// specified immediate.
275 static void emitLoadConstPool(MachineBasicBlock &MBB,
276 MachineBasicBlock::iterator &MBBI,
277 unsigned DestReg, int Val,
278 ARMCC::CondCodes Pred, unsigned PredReg,
279 const TargetInstrInfo &TII, bool isThumb) {
280 MachineFunction &MF = *MBB.getParent();
281 MachineConstantPool *ConstantPool = MF.getConstantPool();
282 Constant *C = ConstantInt::get(Type::Int32Ty, Val);
283 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 2);
285 BuildMI(MBB, MBBI, TII.get(ARM::tLDRcp), DestReg).addConstantPoolIndex(Idx);
287 BuildMI(MBB, MBBI, TII.get(ARM::LDRcp), DestReg).addConstantPoolIndex(Idx)
288 .addReg(0).addImm(0).addImm((unsigned)Pred).addReg(PredReg);
291 void ARMRegisterInfo::reMaterialize(MachineBasicBlock &MBB,
292 MachineBasicBlock::iterator I,
294 const MachineInstr *Orig) const {
295 if (Orig->getOpcode() == ARM::MOVi2pieces) {
296 emitLoadConstPool(MBB, I, DestReg,
297 Orig->getOperand(1).getImm(),
298 (ARMCC::CondCodes)Orig->getOperand(2).getImm(),
299 Orig->getOperand(3).getReg(),
304 MachineInstr *MI = Orig->clone();
305 MI->getOperand(0).setReg(DestReg);
309 /// isLowRegister - Returns true if the register is low register r0-r7.
311 static bool isLowRegister(unsigned Reg) {
314 case R0: case R1: case R2: case R3:
315 case R4: case R5: case R6: case R7:
322 MachineInstr *ARMRegisterInfo::foldMemoryOperand(MachineInstr *MI,
323 SmallVectorImpl<unsigned> &Ops,
325 if (Ops.size() != 1) return NULL;
327 unsigned OpNum = Ops[0];
328 unsigned Opc = MI->getOpcode();
329 MachineInstr *NewMI = NULL;
333 if (MI->getOperand(4).getReg() == ARM::CPSR)
334 // If it is updating CPSR, then it cannot be foled.
336 unsigned Pred = MI->getOperand(2).getImm();
337 unsigned PredReg = MI->getOperand(3).getReg();
338 if (OpNum == 0) { // move -> store
339 unsigned SrcReg = MI->getOperand(1).getReg();
340 NewMI = BuildMI(TII.get(ARM::STR)).addReg(SrcReg).addFrameIndex(FI)
341 .addReg(0).addImm(0).addImm(Pred).addReg(PredReg);
342 } else { // move -> load
343 unsigned DstReg = MI->getOperand(0).getReg();
344 NewMI = BuildMI(TII.get(ARM::LDR), DstReg).addFrameIndex(FI).addReg(0)
345 .addImm(0).addImm(Pred).addReg(PredReg);
350 if (OpNum == 0) { // move -> store
351 unsigned SrcReg = MI->getOperand(1).getReg();
352 if (isPhysicalRegister(SrcReg) && !isLowRegister(SrcReg))
353 // tSpill cannot take a high register operand.
355 NewMI = BuildMI(TII.get(ARM::tSpill)).addReg(SrcReg).addFrameIndex(FI)
357 } else { // move -> load
358 unsigned DstReg = MI->getOperand(0).getReg();
359 if (isPhysicalRegister(DstReg) && !isLowRegister(DstReg))
360 // tRestore cannot target a high register operand.
362 NewMI = BuildMI(TII.get(ARM::tRestore), DstReg).addFrameIndex(FI)
368 unsigned Pred = MI->getOperand(2).getImm();
369 unsigned PredReg = MI->getOperand(3).getReg();
370 if (OpNum == 0) { // move -> store
371 unsigned SrcReg = MI->getOperand(1).getReg();
372 NewMI = BuildMI(TII.get(ARM::FSTS)).addReg(SrcReg).addFrameIndex(FI)
373 .addImm(0).addImm(Pred).addReg(PredReg);
374 } else { // move -> load
375 unsigned DstReg = MI->getOperand(0).getReg();
376 NewMI = BuildMI(TII.get(ARM::FLDS), DstReg).addFrameIndex(FI)
377 .addImm(0).addImm(Pred).addReg(PredReg);
382 unsigned Pred = MI->getOperand(2).getImm();
383 unsigned PredReg = MI->getOperand(3).getReg();
384 if (OpNum == 0) { // move -> store
385 unsigned SrcReg = MI->getOperand(1).getReg();
386 NewMI = BuildMI(TII.get(ARM::FSTD)).addReg(SrcReg).addFrameIndex(FI)
387 .addImm(0).addImm(Pred).addReg(PredReg);
388 } else { // move -> load
389 unsigned DstReg = MI->getOperand(0).getReg();
390 NewMI = BuildMI(TII.get(ARM::FLDD), DstReg).addFrameIndex(FI)
391 .addImm(0).addImm(Pred).addReg(PredReg);
398 NewMI->copyKillDeadInfo(MI);
402 bool ARMRegisterInfo::canFoldMemoryOperand(MachineInstr *MI,
403 SmallVectorImpl<unsigned> &Ops) const {
404 if (Ops.size() != 1) return false;
406 unsigned OpNum = Ops[0];
407 unsigned Opc = MI->getOpcode();
411 // If it is updating CPSR, then it cannot be foled.
412 return MI->getOperand(4).getReg() != ARM::CPSR;
414 if (OpNum == 0) { // move -> store
415 unsigned SrcReg = MI->getOperand(1).getReg();
416 if (isPhysicalRegister(SrcReg) && !isLowRegister(SrcReg))
417 // tSpill cannot take a high register operand.
419 } else { // move -> load
420 unsigned DstReg = MI->getOperand(0).getReg();
421 if (isPhysicalRegister(DstReg) && !isLowRegister(DstReg))
422 // tRestore cannot target a high register operand.
436 ARMRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
437 static const unsigned CalleeSavedRegs[] = {
438 ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8,
439 ARM::R7, ARM::R6, ARM::R5, ARM::R4,
441 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
442 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
446 static const unsigned DarwinCalleeSavedRegs[] = {
447 ARM::LR, ARM::R7, ARM::R6, ARM::R5, ARM::R4,
448 ARM::R11, ARM::R10, ARM::R9, ARM::R8,
450 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
451 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
454 return STI.isTargetDarwin() ? DarwinCalleeSavedRegs : CalleeSavedRegs;
457 const TargetRegisterClass* const *
458 ARMRegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const {
459 static const TargetRegisterClass * const CalleeSavedRegClasses[] = {
460 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
461 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
462 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
464 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
465 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
468 return CalleeSavedRegClasses;
471 BitVector ARMRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
472 // FIXME: avoid re-calculating this everytime.
473 BitVector Reserved(getNumRegs());
474 Reserved.set(ARM::SP);
475 Reserved.set(ARM::PC);
476 if (STI.isTargetDarwin() || hasFP(MF))
477 Reserved.set(FramePtr);
478 // Some targets reserve R9.
479 if (STI.isR9Reserved())
480 Reserved.set(ARM::R9);
485 ARMRegisterInfo::isReservedReg(const MachineFunction &MF, unsigned Reg) const {
493 if (FramePtr == Reg && (STI.isTargetDarwin() || hasFP(MF)))
497 return STI.isR9Reserved();
504 ARMRegisterInfo::requiresRegisterScavenging(const MachineFunction &MF) const {
505 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
506 return ThumbRegScavenging || !AFI->isThumbFunction();
509 /// hasFP - Return true if the specified function should have a dedicated frame
510 /// pointer register. This is true if the function has variable sized allocas
511 /// or if frame pointer elimination is disabled.
513 bool ARMRegisterInfo::hasFP(const MachineFunction &MF) const {
514 return NoFramePointerElim || MF.getFrameInfo()->hasVarSizedObjects();
517 // hasReservedCallFrame - Under normal circumstances, when a frame pointer is
518 // not required, we reserve argument space for call sites in the function
519 // immediately on entry to the current function. This eliminates the need for
520 // add/sub sp brackets around call sites. Returns true if the call frame is
521 // included as part of the stack frame.
522 bool ARMRegisterInfo::hasReservedCallFrame(MachineFunction &MF) const {
523 const MachineFrameInfo *FFI = MF.getFrameInfo();
524 unsigned CFSize = FFI->getMaxCallFrameSize();
525 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
526 // It's not always a good idea to include the call frame as part of the
527 // stack frame. ARM (especially Thumb) has small immediate offset to
528 // address the stack frame. So a large call frame can cause poor codegen
529 // and may even makes it impossible to scavenge a register.
530 if (AFI->isThumbFunction()) {
531 if (CFSize >= ((1 << 8) - 1) * 4 / 2) // Half of imm8 * 4
534 if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12
537 return !MF.getFrameInfo()->hasVarSizedObjects();
540 /// emitARMRegPlusImmediate - Emits a series of instructions to materialize
541 /// a destreg = basereg + immediate in ARM code.
543 void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
544 MachineBasicBlock::iterator &MBBI,
545 unsigned DestReg, unsigned BaseReg, int NumBytes,
546 ARMCC::CondCodes Pred, unsigned PredReg,
547 const TargetInstrInfo &TII) {
548 bool isSub = NumBytes < 0;
549 if (isSub) NumBytes = -NumBytes;
552 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
553 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
554 assert(ThisVal && "Didn't extract field correctly");
556 // We will handle these bits from offset, clear them.
557 NumBytes &= ~ThisVal;
559 // Get the properly encoded SOImmVal field.
560 int SOImmVal = ARM_AM::getSOImmVal(ThisVal);
561 assert(SOImmVal != -1 && "Bit extraction didn't work?");
563 // Build the new ADD / SUB.
564 BuildMI(MBB, MBBI, TII.get(isSub ? ARM::SUBri : ARM::ADDri), DestReg)
565 .addReg(BaseReg, false, false, true).addImm(SOImmVal)
566 .addImm((unsigned)Pred).addReg(PredReg).addReg(0);
571 /// calcNumMI - Returns the number of instructions required to materialize
572 /// the specific add / sub r, c instruction.
573 static unsigned calcNumMI(int Opc, int ExtraOpc, unsigned Bytes,
574 unsigned NumBits, unsigned Scale) {
576 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
578 if (Opc == ARM::tADDrSPi) {
579 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
583 Scale = 1; // Followed by a number of tADDi8.
584 Chunk = ((1 << NumBits) - 1) * Scale;
587 NumMIs += Bytes / Chunk;
588 if ((Bytes % Chunk) != 0)
595 /// emitThumbRegPlusImmInReg - Emits a series of instructions to materialize
596 /// a destreg = basereg + immediate in Thumb code. Materialize the immediate
597 /// in a register using mov / mvn sequences or load the immediate from a
600 void emitThumbRegPlusImmInReg(MachineBasicBlock &MBB,
601 MachineBasicBlock::iterator &MBBI,
602 unsigned DestReg, unsigned BaseReg,
603 int NumBytes, bool CanChangeCC,
604 const TargetInstrInfo &TII) {
605 bool isHigh = !isLowRegister(DestReg) ||
606 (BaseReg != 0 && !isLowRegister(BaseReg));
608 // Subtract doesn't have high register version. Load the negative value
609 // if either base or dest register is a high register. Also, if do not
610 // issue sub as part of the sequence if condition register is to be
612 if (NumBytes < 0 && !isHigh && CanChangeCC) {
614 NumBytes = -NumBytes;
616 unsigned LdReg = DestReg;
617 if (DestReg == ARM::SP) {
618 assert(BaseReg == ARM::SP && "Unexpected!");
620 BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), ARM::R12)
621 .addReg(ARM::R3, false, false, true);
624 if (NumBytes <= 255 && NumBytes >= 0)
625 BuildMI(MBB, MBBI, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes);
626 else if (NumBytes < 0 && NumBytes >= -255) {
627 BuildMI(MBB, MBBI, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes);
628 BuildMI(MBB, MBBI, TII.get(ARM::tNEG), LdReg)
629 .addReg(LdReg, false, false, true);
631 emitLoadConstPool(MBB, MBBI, LdReg, NumBytes, ARMCC::AL, 0, TII, true);
634 int Opc = (isSub) ? ARM::tSUBrr : (isHigh ? ARM::tADDhirr : ARM::tADDrr);
635 const MachineInstrBuilder MIB = BuildMI(MBB, MBBI, TII.get(Opc), DestReg);
636 if (DestReg == ARM::SP || isSub)
637 MIB.addReg(BaseReg).addReg(LdReg, false, false, true);
639 MIB.addReg(LdReg).addReg(BaseReg, false, false, true);
640 if (DestReg == ARM::SP)
641 BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), ARM::R3)
642 .addReg(ARM::R12, false, false, true);
645 /// emitThumbRegPlusImmediate - Emits a series of instructions to materialize
646 /// a destreg = basereg + immediate in Thumb code.
648 void emitThumbRegPlusImmediate(MachineBasicBlock &MBB,
649 MachineBasicBlock::iterator &MBBI,
650 unsigned DestReg, unsigned BaseReg,
651 int NumBytes, const TargetInstrInfo &TII) {
652 bool isSub = NumBytes < 0;
653 unsigned Bytes = (unsigned)NumBytes;
654 if (isSub) Bytes = -NumBytes;
655 bool isMul4 = (Bytes & 3) == 0;
656 bool isTwoAddr = false;
657 bool DstNotEqBase = false;
658 unsigned NumBits = 1;
663 if (DestReg == BaseReg && BaseReg == ARM::SP) {
664 assert(isMul4 && "Thumb sp inc / dec size must be multiple of 4!");
667 Opc = isSub ? ARM::tSUBspi : ARM::tADDspi;
669 } else if (!isSub && BaseReg == ARM::SP) {
672 // r1 = add sp, 100 * 4
676 ExtraOpc = ARM::tADDi3;
685 if (DestReg != BaseReg)
688 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
692 unsigned NumMIs = calcNumMI(Opc, ExtraOpc, Bytes, NumBits, Scale);
693 unsigned Threshold = (DestReg == ARM::SP) ? 3 : 2;
694 if (NumMIs > Threshold) {
695 // This will expand into too many instructions. Load the immediate from a
697 emitThumbRegPlusImmInReg(MBB, MBBI, DestReg, BaseReg, NumBytes, true, TII);
702 if (isLowRegister(DestReg) && isLowRegister(BaseReg)) {
703 // If both are low registers, emit DestReg = add BaseReg, max(Imm, 7)
704 unsigned Chunk = (1 << 3) - 1;
705 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
707 BuildMI(MBB, MBBI, TII.get(isSub ? ARM::tSUBi3 : ARM::tADDi3), DestReg)
708 .addReg(BaseReg, false, false, true).addImm(ThisVal);
710 BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), DestReg)
711 .addReg(BaseReg, false, false, true);
716 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
718 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
721 // Build the new tADD / tSUB.
723 BuildMI(MBB, MBBI, TII.get(Opc), DestReg).addReg(DestReg).addImm(ThisVal);
725 bool isKill = BaseReg != ARM::SP;
726 BuildMI(MBB, MBBI, TII.get(Opc), DestReg)
727 .addReg(BaseReg, false, false, isKill).addImm(ThisVal);
730 if (Opc == ARM::tADDrSPi) {
736 Chunk = ((1 << NumBits) - 1) * Scale;
737 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
744 BuildMI(MBB, MBBI, TII.get(ExtraOpc), DestReg)
745 .addReg(DestReg, false, false, true)
746 .addImm(((unsigned)NumBytes) & 3);
750 void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
751 int NumBytes, ARMCC::CondCodes Pred, unsigned PredReg,
752 bool isThumb, const TargetInstrInfo &TII) {
754 emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII);
756 emitARMRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes,
760 void ARMRegisterInfo::
761 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
762 MachineBasicBlock::iterator I) const {
763 if (!hasReservedCallFrame(MF)) {
764 // If we have alloca, convert as follows:
765 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
766 // ADJCALLSTACKUP -> add, sp, sp, amount
767 MachineInstr *Old = I;
768 unsigned Amount = Old->getOperand(0).getImm();
770 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
771 // We need to keep the stack aligned properly. To do this, we round the
772 // amount of space needed for the outgoing arguments up to the next
773 // alignment boundary.
774 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
775 Amount = (Amount+Align-1)/Align*Align;
777 // Replace the pseudo instruction with a new instruction...
778 unsigned Opc = Old->getOpcode();
779 bool isThumb = AFI->isThumbFunction();
780 ARMCC::CondCodes Pred = isThumb
781 ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(1).getImm();
782 if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) {
783 // Note: PredReg is operand 2 for ADJCALLSTACKDOWN.
784 unsigned PredReg = isThumb ? 0 : Old->getOperand(2).getReg();
785 emitSPUpdate(MBB, I, -Amount, Pred, PredReg, isThumb, TII);
787 // Note: PredReg is operand 3 for ADJCALLSTACKUP.
788 unsigned PredReg = isThumb ? 0 : Old->getOperand(3).getReg();
789 assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP);
790 emitSPUpdate(MBB, I, Amount, Pred, PredReg, isThumb, TII);
797 /// emitThumbConstant - Emit a series of instructions to materialize a
799 static void emitThumbConstant(MachineBasicBlock &MBB,
800 MachineBasicBlock::iterator &MBBI,
801 unsigned DestReg, int Imm,
802 const TargetInstrInfo &TII) {
803 bool isSub = Imm < 0;
804 if (isSub) Imm = -Imm;
806 int Chunk = (1 << 8) - 1;
807 int ThisVal = (Imm > Chunk) ? Chunk : Imm;
809 BuildMI(MBB, MBBI, TII.get(ARM::tMOVi8), DestReg).addImm(ThisVal);
811 emitThumbRegPlusImmediate(MBB, MBBI, DestReg, DestReg, Imm, TII);
813 BuildMI(MBB, MBBI, TII.get(ARM::tNEG), DestReg)
814 .addReg(DestReg, false, false, true);
817 /// findScratchRegister - Find a 'free' ARM register. If register scavenger
818 /// is not being used, R12 is available. Otherwise, try for a call-clobbered
819 /// register first and then a spilled callee-saved register if that fails.
821 unsigned findScratchRegister(RegScavenger *RS, const TargetRegisterClass *RC,
822 ARMFunctionInfo *AFI) {
823 unsigned Reg = RS ? RS->FindUnusedReg(RC, true) : (unsigned) ARM::R12;
825 // Try a already spilled CS register.
826 Reg = RS->FindUnusedReg(RC, AFI->getSpilledCSRegisters());
831 void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
832 int SPAdj, RegScavenger *RS) const{
834 MachineInstr &MI = *II;
835 MachineBasicBlock &MBB = *MI.getParent();
836 MachineFunction &MF = *MBB.getParent();
837 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
838 bool isThumb = AFI->isThumbFunction();
840 while (!MI.getOperand(i).isFrameIndex()) {
842 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
845 unsigned FrameReg = ARM::SP;
846 int FrameIndex = MI.getOperand(i).getIndex();
847 int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex) +
848 MF.getFrameInfo()->getStackSize() + SPAdj;
850 if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex))
851 Offset -= AFI->getGPRCalleeSavedArea1Offset();
852 else if (AFI->isGPRCalleeSavedArea2Frame(FrameIndex))
853 Offset -= AFI->getGPRCalleeSavedArea2Offset();
854 else if (AFI->isDPRCalleeSavedAreaFrame(FrameIndex))
855 Offset -= AFI->getDPRCalleeSavedAreaOffset();
856 else if (hasFP(MF)) {
857 assert(SPAdj == 0 && "Unexpected");
858 // There is alloca()'s in this function, must reference off the frame
860 FrameReg = getFrameRegister(MF);
861 Offset -= AFI->getFramePtrSpillOffset();
864 unsigned Opcode = MI.getOpcode();
865 const TargetInstrDescriptor &Desc = TII.get(Opcode);
866 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
869 if (Opcode == ARM::ADDri) {
870 Offset += MI.getOperand(i+1).getImm();
872 // Turn it into a move.
873 MI.setInstrDescriptor(TII.get(ARM::MOVr));
874 MI.getOperand(i).ChangeToRegister(FrameReg, false);
875 MI.RemoveOperand(i+1);
877 } else if (Offset < 0) {
880 MI.setInstrDescriptor(TII.get(ARM::SUBri));
883 // Common case: small offset, fits into instruction.
884 int ImmedOffset = ARM_AM::getSOImmVal(Offset);
885 if (ImmedOffset != -1) {
886 // Replace the FrameIndex with sp / fp
887 MI.getOperand(i).ChangeToRegister(FrameReg, false);
888 MI.getOperand(i+1).ChangeToImmediate(ImmedOffset);
892 // Otherwise, we fallback to common code below to form the imm offset with
893 // a sequence of ADDri instructions. First though, pull as much of the imm
894 // into this ADDri as possible.
895 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
896 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
898 // We will handle these bits from offset, clear them.
899 Offset &= ~ThisImmVal;
901 // Get the properly encoded SOImmVal field.
902 int ThisSOImmVal = ARM_AM::getSOImmVal(ThisImmVal);
903 assert(ThisSOImmVal != -1 && "Bit extraction didn't work?");
904 MI.getOperand(i+1).ChangeToImmediate(ThisSOImmVal);
905 } else if (Opcode == ARM::tADDrSPi) {
906 Offset += MI.getOperand(i+1).getImm();
908 // Can't use tADDrSPi if it's based off the frame pointer.
909 unsigned NumBits = 0;
911 if (FrameReg != ARM::SP) {
912 Opcode = ARM::tADDi3;
913 MI.setInstrDescriptor(TII.get(ARM::tADDi3));
918 assert((Offset & 3) == 0 &&
919 "Thumb add/sub sp, #imm immediate must be multiple of 4!");
923 // Turn it into a move.
924 MI.setInstrDescriptor(TII.get(ARM::tMOVr));
925 MI.getOperand(i).ChangeToRegister(FrameReg, false);
926 MI.RemoveOperand(i+1);
930 // Common case: small offset, fits into instruction.
931 unsigned Mask = (1 << NumBits) - 1;
932 if (((Offset / Scale) & ~Mask) == 0) {
933 // Replace the FrameIndex with sp / fp
934 MI.getOperand(i).ChangeToRegister(FrameReg, false);
935 MI.getOperand(i+1).ChangeToImmediate(Offset / Scale);
939 unsigned DestReg = MI.getOperand(0).getReg();
940 unsigned Bytes = (Offset > 0) ? Offset : -Offset;
941 unsigned NumMIs = calcNumMI(Opcode, 0, Bytes, NumBits, Scale);
942 // MI would expand into a large number of instructions. Don't try to
943 // simplify the immediate.
945 emitThumbRegPlusImmediate(MBB, II, DestReg, FrameReg, Offset, TII);
951 // Translate r0 = add sp, imm to
952 // r0 = add sp, 255*4
953 // r0 = add r0, (imm - 255*4)
954 MI.getOperand(i).ChangeToRegister(FrameReg, false);
955 MI.getOperand(i+1).ChangeToImmediate(Mask);
956 Offset = (Offset - Mask * Scale);
957 MachineBasicBlock::iterator NII = next(II);
958 emitThumbRegPlusImmediate(MBB, NII, DestReg, DestReg, Offset, TII);
960 // Translate r0 = add sp, -imm to
961 // r0 = -imm (this is then translated into a series of instructons)
963 emitThumbConstant(MBB, II, DestReg, Offset, TII);
964 MI.setInstrDescriptor(TII.get(ARM::tADDhirr));
965 MI.getOperand(i).ChangeToRegister(DestReg, false, false, true);
966 MI.getOperand(i+1).ChangeToRegister(FrameReg, false);
972 unsigned NumBits = 0;
975 case ARMII::AddrMode2: {
977 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
978 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
983 case ARMII::AddrMode3: {
985 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
986 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
991 case ARMII::AddrMode5: {
993 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
994 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1000 case ARMII::AddrModeTs: {
1002 InstrOffs = MI.getOperand(ImmIdx).getImm();
1003 NumBits = (FrameReg == ARM::SP) ? 8 : 5;
1008 assert(0 && "Unsupported addressing mode!");
1013 Offset += InstrOffs * Scale;
1014 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1015 if (Offset < 0 && !isThumb) {
1020 // Common case: small offset, fits into instruction.
1021 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1022 int ImmedOffset = Offset / Scale;
1023 unsigned Mask = (1 << NumBits) - 1;
1024 if ((unsigned)Offset <= Mask * Scale) {
1025 // Replace the FrameIndex with sp
1026 MI.getOperand(i).ChangeToRegister(FrameReg, false);
1028 ImmedOffset |= 1 << NumBits;
1029 ImmOp.ChangeToImmediate(ImmedOffset);
1033 bool isThumSpillRestore = Opcode == ARM::tRestore || Opcode == ARM::tSpill;
1034 if (AddrMode == ARMII::AddrModeTs) {
1035 // Thumb tLDRspi, tSTRspi. These will change to instructions that use
1036 // a different base register.
1038 Mask = (1 << NumBits) - 1;
1040 // If this is a thumb spill / restore, we will be using a constpool load to
1041 // materialize the offset.
1042 if (AddrMode == ARMII::AddrModeTs && isThumSpillRestore)
1043 ImmOp.ChangeToImmediate(0);
1045 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1046 ImmedOffset = ImmedOffset & Mask;
1048 ImmedOffset |= 1 << NumBits;
1049 ImmOp.ChangeToImmediate(ImmedOffset);
1050 Offset &= ~(Mask*Scale);
1054 // If we get here, the immediate doesn't fit into the instruction. We folded
1055 // as much as possible above, handle the rest, providing a register that is
1057 assert(Offset && "This code isn't needed if offset already handled!");
1060 if (TII.isLoad(Opcode)) {
1061 // Use the destination register to materialize sp + offset.
1062 unsigned TmpReg = MI.getOperand(0).getReg();
1064 if (Opcode == ARM::tRestore) {
1065 if (FrameReg == ARM::SP)
1066 emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg,Offset,false,TII);
1068 emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, TII, true);
1072 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII);
1073 MI.setInstrDescriptor(TII.get(ARM::tLDR));
1074 MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true);
1076 // Use [reg, reg] addrmode.
1077 MI.addOperand(MachineOperand::CreateReg(FrameReg, false));
1078 else // tLDR has an extra register operand.
1079 MI.addOperand(MachineOperand::CreateReg(0, false));
1080 } else if (TII.isStore(Opcode)) {
1081 // FIXME! This is horrific!!! We need register scavenging.
1082 // Our temporary workaround has marked r3 unavailable. Of course, r3 is
1083 // also a ABI register so it's possible that is is the register that is
1084 // being storing here. If that's the case, we do the following:
1086 // Use r2 to materialize sp + offset
1089 unsigned ValReg = MI.getOperand(0).getReg();
1090 unsigned TmpReg = ARM::R3;
1092 if (ValReg == ARM::R3) {
1093 BuildMI(MBB, II, TII.get(ARM::tMOVr), ARM::R12)
1094 .addReg(ARM::R2, false, false, true);
1097 if (TmpReg == ARM::R3 && AFI->isR3LiveIn())
1098 BuildMI(MBB, II, TII.get(ARM::tMOVr), ARM::R12)
1099 .addReg(ARM::R3, false, false, true);
1100 if (Opcode == ARM::tSpill) {
1101 if (FrameReg == ARM::SP)
1102 emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg,Offset,false,TII);
1104 emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, TII, true);
1108 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII);
1109 MI.setInstrDescriptor(TII.get(ARM::tSTR));
1110 MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true);
1111 if (UseRR) // Use [reg, reg] addrmode.
1112 MI.addOperand(MachineOperand::CreateReg(FrameReg, false));
1113 else // tSTR has an extra register operand.
1114 MI.addOperand(MachineOperand::CreateReg(0, false));
1116 MachineBasicBlock::iterator NII = next(II);
1117 if (ValReg == ARM::R3)
1118 BuildMI(MBB, NII, TII.get(ARM::tMOVr), ARM::R2)
1119 .addReg(ARM::R12, false, false, true);
1120 if (TmpReg == ARM::R3 && AFI->isR3LiveIn())
1121 BuildMI(MBB, NII, TII.get(ARM::tMOVr), ARM::R3)
1122 .addReg(ARM::R12, false, false, true);
1124 assert(false && "Unexpected opcode!");
1126 // Insert a set of r12 with the full address: r12 = sp + offset
1127 // If the offset we have is too large to fit into the instruction, we need
1128 // to form it with a series of ADDri's. Do this by taking 8-bit chunks
1130 unsigned ScratchReg = findScratchRegister(RS, &ARM::GPRRegClass, AFI);
1131 if (ScratchReg == 0)
1132 // No register is "free". Scavenge a register.
1133 ScratchReg = RS->scavengeRegister(&ARM::GPRRegClass, II, SPAdj);
1134 int PIdx = MI.findFirstPredOperandIdx();
1135 ARMCC::CondCodes Pred = (PIdx == -1)
1136 ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm();
1137 unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg();
1138 emitARMRegPlusImmediate(MBB, II, ScratchReg, FrameReg,
1139 isSub ? -Offset : Offset, Pred, PredReg, TII);
1140 MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true);
1144 static unsigned estimateStackSize(MachineFunction &MF, MachineFrameInfo *MFI) {
1145 const MachineFrameInfo *FFI = MF.getFrameInfo();
1147 for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) {
1148 int FixedOff = -FFI->getObjectOffset(i);
1149 if (FixedOff > Offset) Offset = FixedOff;
1151 for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) {
1152 Offset += FFI->getObjectSize(i);
1153 unsigned Align = FFI->getObjectAlignment(i);
1154 // Adjust to alignment boundary
1155 Offset = (Offset+Align-1)/Align*Align;
1157 return (unsigned)Offset;
1161 ARMRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
1162 RegScavenger *RS) const {
1163 // This tells PEI to spill the FP as if it is any other callee-save register
1164 // to take advantage the eliminateFrameIndex machinery. This also ensures it
1165 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
1166 // to combine multiple loads / stores.
1167 bool CanEliminateFrame = true;
1168 bool CS1Spilled = false;
1169 bool LRSpilled = false;
1170 unsigned NumGPRSpills = 0;
1171 SmallVector<unsigned, 4> UnspilledCS1GPRs;
1172 SmallVector<unsigned, 4> UnspilledCS2GPRs;
1173 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1175 // Don't spill FP if the frame can be eliminated. This is determined
1176 // by scanning the callee-save registers to see if any is used.
1177 const unsigned *CSRegs = getCalleeSavedRegs();
1178 const TargetRegisterClass* const *CSRegClasses = getCalleeSavedRegClasses();
1179 for (unsigned i = 0; CSRegs[i]; ++i) {
1180 unsigned Reg = CSRegs[i];
1181 bool Spilled = false;
1182 if (MF.getRegInfo().isPhysRegUsed(Reg)) {
1183 AFI->setCSRegisterIsSpilled(Reg);
1185 CanEliminateFrame = false;
1187 // Check alias registers too.
1188 for (const unsigned *Aliases = getAliasSet(Reg); *Aliases; ++Aliases) {
1189 if (MF.getRegInfo().isPhysRegUsed(*Aliases)) {
1191 CanEliminateFrame = false;
1196 if (CSRegClasses[i] == &ARM::GPRRegClass) {
1200 if (!STI.isTargetDarwin()) {
1207 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
1222 if (!STI.isTargetDarwin()) {
1223 UnspilledCS1GPRs.push_back(Reg);
1233 UnspilledCS1GPRs.push_back(Reg);
1236 UnspilledCS2GPRs.push_back(Reg);
1243 bool ForceLRSpill = false;
1244 if (!LRSpilled && AFI->isThumbFunction()) {
1245 unsigned FnSize = ARM::GetFunctionSize(MF);
1246 // Force LR to be spilled if the Thumb function size is > 2048. This enables
1247 // use of BL to implement far jump. If it turns out that it's not needed
1248 // then the branch fix up path will undo it.
1249 if (FnSize >= (1 << 11)) {
1250 CanEliminateFrame = false;
1251 ForceLRSpill = true;
1255 bool ExtraCSSpill = false;
1256 if (!CanEliminateFrame || hasFP(MF)) {
1257 AFI->setHasStackFrame(true);
1259 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
1260 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
1261 if (!LRSpilled && CS1Spilled) {
1262 MF.getRegInfo().setPhysRegUsed(ARM::LR);
1263 AFI->setCSRegisterIsSpilled(ARM::LR);
1265 UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(),
1266 UnspilledCS1GPRs.end(), (unsigned)ARM::LR));
1267 ForceLRSpill = false;
1268 ExtraCSSpill = true;
1271 // Darwin ABI requires FP to point to the stack slot that contains the
1273 if (STI.isTargetDarwin() || hasFP(MF)) {
1274 MF.getRegInfo().setPhysRegUsed(FramePtr);
1278 // If stack and double are 8-byte aligned and we are spilling an odd number
1279 // of GPRs. Spill one extra callee save GPR so we won't have to pad between
1280 // the integer and double callee save areas.
1281 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
1282 if (TargetAlign == 8 && (NumGPRSpills & 1)) {
1283 if (CS1Spilled && !UnspilledCS1GPRs.empty()) {
1284 for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) {
1285 unsigned Reg = UnspilledCS1GPRs[i];
1286 // Don't spiil high register if the function is thumb
1287 if (!AFI->isThumbFunction() || isLowRegister(Reg) || Reg == ARM::LR) {
1288 MF.getRegInfo().setPhysRegUsed(Reg);
1289 AFI->setCSRegisterIsSpilled(Reg);
1290 if (!isReservedReg(MF, Reg))
1291 ExtraCSSpill = true;
1295 } else if (!UnspilledCS2GPRs.empty() &&
1296 !AFI->isThumbFunction()) {
1297 unsigned Reg = UnspilledCS2GPRs.front();
1298 MF.getRegInfo().setPhysRegUsed(Reg);
1299 AFI->setCSRegisterIsSpilled(Reg);
1300 if (!isReservedReg(MF, Reg))
1301 ExtraCSSpill = true;
1305 // Estimate if we might need to scavenge a register at some point in order
1306 // to materialize a stack offset. If so, either spill one additiona
1307 // callee-saved register or reserve a special spill slot to facilitate
1308 // register scavenging.
1309 if (RS && !ExtraCSSpill && !AFI->isThumbFunction()) {
1310 MachineFrameInfo *MFI = MF.getFrameInfo();
1311 unsigned Size = estimateStackSize(MF, MFI);
1312 unsigned Limit = (1 << 12) - 1;
1313 for (MachineFunction::iterator BB = MF.begin(),E = MF.end();BB != E; ++BB)
1314 for (MachineBasicBlock::iterator I= BB->begin(); I != BB->end(); ++I) {
1315 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1316 if (I->getOperand(i).isFrameIndex()) {
1317 unsigned Opcode = I->getOpcode();
1318 const TargetInstrDescriptor &Desc = TII.get(Opcode);
1319 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1320 if (AddrMode == ARMII::AddrMode3) {
1321 Limit = (1 << 8) - 1;
1322 goto DoneEstimating;
1323 } else if (AddrMode == ARMII::AddrMode5) {
1324 unsigned ThisLimit = ((1 << 8) - 1) * 4;
1325 if (ThisLimit < Limit)
1331 if (Size >= Limit) {
1332 // If any non-reserved CS register isn't spilled, just spill one or two
1333 // extra. That should take care of it!
1334 unsigned NumExtras = TargetAlign / 4;
1335 SmallVector<unsigned, 2> Extras;
1336 while (NumExtras && !UnspilledCS1GPRs.empty()) {
1337 unsigned Reg = UnspilledCS1GPRs.back();
1338 UnspilledCS1GPRs.pop_back();
1339 if (!isReservedReg(MF, Reg)) {
1340 Extras.push_back(Reg);
1344 while (NumExtras && !UnspilledCS2GPRs.empty()) {
1345 unsigned Reg = UnspilledCS2GPRs.back();
1346 UnspilledCS2GPRs.pop_back();
1347 if (!isReservedReg(MF, Reg)) {
1348 Extras.push_back(Reg);
1352 if (Extras.size() && NumExtras == 0) {
1353 for (unsigned i = 0, e = Extras.size(); i != e; ++i) {
1354 MF.getRegInfo().setPhysRegUsed(Extras[i]);
1355 AFI->setCSRegisterIsSpilled(Extras[i]);
1358 // Reserve a slot closest to SP or frame pointer.
1359 const TargetRegisterClass *RC = &ARM::GPRRegClass;
1360 RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
1361 RC->getAlignment()));
1368 MF.getRegInfo().setPhysRegUsed(ARM::LR);
1369 AFI->setCSRegisterIsSpilled(ARM::LR);
1370 AFI->setLRIsSpilledForFarJump(true);
1374 /// Move iterator pass the next bunch of callee save load / store ops for
1375 /// the particular spill area (1: integer area 1, 2: integer area 2,
1376 /// 3: fp area, 0: don't care).
1377 static void movePastCSLoadStoreOps(MachineBasicBlock &MBB,
1378 MachineBasicBlock::iterator &MBBI,
1379 int Opc, unsigned Area,
1380 const ARMSubtarget &STI) {
1381 while (MBBI != MBB.end() &&
1382 MBBI->getOpcode() == Opc && MBBI->getOperand(1).isFrameIndex()) {
1385 unsigned Category = 0;
1386 switch (MBBI->getOperand(0).getReg()) {
1387 case ARM::R4: case ARM::R5: case ARM::R6: case ARM::R7:
1391 case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11:
1392 Category = STI.isTargetDarwin() ? 2 : 1;
1394 case ARM::D8: case ARM::D9: case ARM::D10: case ARM::D11:
1395 case ARM::D12: case ARM::D13: case ARM::D14: case ARM::D15:
1402 if (Done || Category != Area)
1410 void ARMRegisterInfo::emitPrologue(MachineFunction &MF) const {
1411 MachineBasicBlock &MBB = MF.front();
1412 MachineBasicBlock::iterator MBBI = MBB.begin();
1413 MachineFrameInfo *MFI = MF.getFrameInfo();
1414 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1415 bool isThumb = AFI->isThumbFunction();
1416 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1417 unsigned NumBytes = MFI->getStackSize();
1418 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
1421 // Check if R3 is live in. It might have to be used as a scratch register.
1422 for (MachineRegisterInfo::livein_iterator I =MF.getRegInfo().livein_begin(),
1423 E = MF.getRegInfo().livein_end(); I != E; ++I) {
1424 if (I->first == ARM::R3) {
1425 AFI->setR3IsLiveIn(true);
1430 // Thumb add/sub sp, imm8 instructions implicitly multiply the offset by 4.
1431 NumBytes = (NumBytes + 3) & ~3;
1432 MFI->setStackSize(NumBytes);
1435 // Determine the sizes of each callee-save spill areas and record which frame
1436 // belongs to which callee-save spill areas.
1437 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
1438 int FramePtrSpillFI = 0;
1441 emitSPUpdate(MBB, MBBI, -VARegSaveSize, ARMCC::AL, 0, isThumb, TII);
1443 if (!AFI->hasStackFrame()) {
1445 emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, isThumb, TII);
1449 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1450 unsigned Reg = CSI[i].getReg();
1451 int FI = CSI[i].getFrameIdx();
1458 if (Reg == FramePtr)
1459 FramePtrSpillFI = FI;
1460 AFI->addGPRCalleeSavedArea1Frame(FI);
1467 if (Reg == FramePtr)
1468 FramePtrSpillFI = FI;
1469 if (STI.isTargetDarwin()) {
1470 AFI->addGPRCalleeSavedArea2Frame(FI);
1473 AFI->addGPRCalleeSavedArea1Frame(FI);
1478 AFI->addDPRCalleeSavedAreaFrame(FI);
1484 // Build the new SUBri to adjust SP for integer callee-save spill area 1.
1485 emitSPUpdate(MBB, MBBI, -GPRCS1Size, ARMCC::AL, 0, isThumb, TII);
1486 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 1, STI);
1487 } else if (MBBI != MBB.end() && MBBI->getOpcode() == ARM::tPUSH)
1490 // Darwin ABI requires FP to point to the stack slot that contains the
1492 if (STI.isTargetDarwin() || hasFP(MF)) {
1493 MachineInstrBuilder MIB =
1494 BuildMI(MBB, MBBI, TII.get(isThumb ? ARM::tADDrSPi : ARM::ADDri),FramePtr)
1495 .addFrameIndex(FramePtrSpillFI).addImm(0);
1496 if (!isThumb) AddDefaultCC(AddDefaultPred(MIB));
1500 // Build the new SUBri to adjust SP for integer callee-save spill area 2.
1501 emitSPUpdate(MBB, MBBI, -GPRCS2Size, ARMCC::AL, 0, false, TII);
1503 // Build the new SUBri to adjust SP for FP callee-save spill area.
1504 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 2, STI);
1505 emitSPUpdate(MBB, MBBI, -DPRCSSize, ARMCC::AL, 0, false, TII);
1508 // Determine starting offsets of spill areas.
1509 unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize);
1510 unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize;
1511 unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size;
1512 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) + NumBytes);
1513 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
1514 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
1515 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
1517 NumBytes = DPRCSOffset;
1519 // Insert it after all the callee-save spills.
1521 movePastCSLoadStoreOps(MBB, MBBI, ARM::FSTD, 3, STI);
1522 emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, isThumb, TII);
1525 if(STI.isTargetELF() && hasFP(MF)) {
1526 MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() -
1527 AFI->getFramePtrSpillOffset());
1530 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
1531 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
1532 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
1535 static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) {
1536 for (unsigned i = 0; CSRegs[i]; ++i)
1537 if (Reg == CSRegs[i])
1542 static bool isCSRestore(MachineInstr *MI, const unsigned *CSRegs) {
1543 return ((MI->getOpcode() == ARM::FLDD ||
1544 MI->getOpcode() == ARM::LDR ||
1545 MI->getOpcode() == ARM::tRestore) &&
1546 MI->getOperand(1).isFrameIndex() &&
1547 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs));
1550 void ARMRegisterInfo::emitEpilogue(MachineFunction &MF,
1551 MachineBasicBlock &MBB) const {
1552 MachineBasicBlock::iterator MBBI = prior(MBB.end());
1553 assert((MBBI->getOpcode() == ARM::BX_RET ||
1554 MBBI->getOpcode() == ARM::tBX_RET ||
1555 MBBI->getOpcode() == ARM::tPOP_RET) &&
1556 "Can only insert epilog into returning blocks");
1558 MachineFrameInfo *MFI = MF.getFrameInfo();
1559 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1560 bool isThumb = AFI->isThumbFunction();
1561 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1562 int NumBytes = (int)MFI->getStackSize();
1563 if (!AFI->hasStackFrame()) {
1565 emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, isThumb, TII);
1567 // Unwind MBBI to point to first LDR / FLDD.
1568 const unsigned *CSRegs = getCalleeSavedRegs();
1569 if (MBBI != MBB.begin()) {
1572 while (MBBI != MBB.begin() && isCSRestore(MBBI, CSRegs));
1573 if (!isCSRestore(MBBI, CSRegs))
1577 // Move SP to start of FP callee save spill area.
1578 NumBytes -= (AFI->getGPRCalleeSavedArea1Size() +
1579 AFI->getGPRCalleeSavedArea2Size() +
1580 AFI->getDPRCalleeSavedAreaSize());
1583 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1584 // Reset SP based on frame pointer only if the stack frame extends beyond
1585 // frame pointer stack slot or target is ELF and the function has FP.
1587 emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, FramePtr, -NumBytes, TII);
1589 BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), ARM::SP).addReg(FramePtr);
1591 if (MBBI->getOpcode() == ARM::tBX_RET &&
1592 &MBB.front() != MBBI &&
1593 prior(MBBI)->getOpcode() == ARM::tPOP) {
1594 MachineBasicBlock::iterator PMBBI = prior(MBBI);
1595 emitSPUpdate(MBB, PMBBI, NumBytes, ARMCC::AL, 0, isThumb, TII);
1597 emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, isThumb, TII);
1600 // Darwin ABI requires FP to point to the stack slot that contains the
1602 if ((STI.isTargetDarwin() && NumBytes) || hasFP(MF)) {
1603 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1604 // Reset SP based on frame pointer only if the stack frame extends beyond
1605 // frame pointer stack slot or target is ELF and the function has FP.
1606 if (AFI->getGPRCalleeSavedArea2Size() ||
1607 AFI->getDPRCalleeSavedAreaSize() ||
1608 AFI->getDPRCalleeSavedAreaOffset()||
1611 BuildMI(MBB, MBBI, TII.get(ARM::SUBri), ARM::SP).addReg(FramePtr)
1613 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
1615 BuildMI(MBB, MBBI, TII.get(ARM::MOVr), ARM::SP).addReg(FramePtr)
1616 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
1617 } else if (NumBytes) {
1618 emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, false, TII);
1621 // Move SP to start of integer callee save spill area 2.
1622 movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 3, STI);
1623 emitSPUpdate(MBB, MBBI, AFI->getDPRCalleeSavedAreaSize(), ARMCC::AL, 0,
1626 // Move SP to start of integer callee save spill area 1.
1627 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 2, STI);
1628 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea2Size(), ARMCC::AL, 0,
1631 // Move SP to SP upon entry to the function.
1632 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 1, STI);
1633 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea1Size(), ARMCC::AL, 0,
1638 if (VARegSaveSize) {
1640 // Epilogue for vararg functions: pop LR to R3 and branch off it.
1641 // FIXME: Verify this is still ok when R3 is no longer being reserved.
1642 BuildMI(MBB, MBBI, TII.get(ARM::tPOP)).addReg(ARM::R3);
1644 emitSPUpdate(MBB, MBBI, VARegSaveSize, ARMCC::AL, 0, isThumb, TII);
1647 BuildMI(MBB, MBBI, TII.get(ARM::tBX_RET_vararg)).addReg(ARM::R3);
1653 unsigned ARMRegisterInfo::getRARegister() const {
1657 unsigned ARMRegisterInfo::getFrameRegister(MachineFunction &MF) const {
1658 if (STI.isTargetDarwin() || hasFP(MF))
1659 return (STI.useThumbBacktraces() || STI.isThumb()) ? ARM::R7 : ARM::R11;
1664 unsigned ARMRegisterInfo::getEHExceptionRegister() const {
1665 assert(0 && "What is the exception register");
1669 unsigned ARMRegisterInfo::getEHHandlerRegister() const {
1670 assert(0 && "What is the exception handler register");
1674 int ARMRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const {
1675 assert(0 && "What is the dwarf register number");
1679 #include "ARMGenRegisterInfo.inc"