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/Target/TargetFrameInfo.h"
29 #include "llvm/Target/TargetMachine.h"
30 #include "llvm/Target/TargetOptions.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/ADT/STLExtras.h"
37 unsigned ARMRegisterInfo::getRegisterNumbering(unsigned RegEnum) {
40 case R0: case S0: case D0: return 0;
41 case R1: case S1: case D1: return 1;
42 case R2: case S2: case D2: return 2;
43 case R3: case S3: case D3: return 3;
44 case R4: case S4: case D4: return 4;
45 case R5: case S5: case D5: return 5;
46 case R6: case S6: case D6: return 6;
47 case R7: case S7: case D7: return 7;
48 case R8: case S8: case D8: return 8;
49 case R9: case S9: case D9: return 9;
50 case R10: case S10: case D10: return 10;
51 case R11: case S11: case D11: return 11;
52 case R12: case S12: case D12: return 12;
53 case SP: case S13: case D13: return 13;
54 case LR: case S14: case D14: return 14;
55 case PC: case S15: case D15: return 15;
73 std::cerr << "Unknown ARM register!\n";
78 ARMRegisterInfo::ARMRegisterInfo(const TargetInstrInfo &tii,
79 const ARMSubtarget &sti)
80 : ARMGenRegisterInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
82 FramePtr(STI.useThumbBacktraces() ? ARM::R7 : ARM::R11) {
85 bool ARMRegisterInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
86 MachineBasicBlock::iterator MI,
87 const std::vector<CalleeSavedInfo> &CSI) const {
88 MachineFunction &MF = *MBB.getParent();
89 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
90 if (!AFI->isThumbFunction() || CSI.empty())
93 MachineInstrBuilder MIB = BuildMI(MBB, MI, TII.get(ARM::tPUSH));
94 for (unsigned i = CSI.size(); i != 0; --i)
95 MIB.addReg(CSI[i-1].getReg());
99 bool ARMRegisterInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
100 MachineBasicBlock::iterator MI,
101 const std::vector<CalleeSavedInfo> &CSI) const {
102 MachineFunction &MF = *MBB.getParent();
103 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
104 if (!AFI->isThumbFunction() || CSI.empty())
107 bool isVarArg = AFI->getVarArgsRegSaveSize() > 0;
108 MachineInstr *PopMI = new MachineInstr(TII.get(ARM::tPOP));
109 MBB.insert(MI, PopMI);
110 for (unsigned i = CSI.size(); i != 0; --i) {
111 unsigned Reg = CSI[i-1].getReg();
112 if (Reg == ARM::LR) {
113 // Special epilogue for vararg functions. See emitEpilogue
117 PopMI->setInstrDescriptor(TII.get(ARM::tPOP_RET));
120 PopMI->addRegOperand(Reg, true);
125 void ARMRegisterInfo::
126 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
127 unsigned SrcReg, int FI,
128 const TargetRegisterClass *RC) const {
129 if (RC == ARM::GPRRegisterClass) {
130 MachineFunction &MF = *MBB.getParent();
131 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
132 if (AFI->isThumbFunction())
133 BuildMI(MBB, I, TII.get(ARM::tSpill)).addReg(SrcReg)
134 .addFrameIndex(FI).addImm(0);
136 BuildMI(MBB, I, TII.get(ARM::STR)).addReg(SrcReg)
137 .addFrameIndex(FI).addReg(0).addImm(0);
138 } else if (RC == ARM::DPRRegisterClass) {
139 BuildMI(MBB, I, TII.get(ARM::FSTD)).addReg(SrcReg)
140 .addFrameIndex(FI).addImm(0);
142 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
143 BuildMI(MBB, I, TII.get(ARM::FSTS)).addReg(SrcReg)
144 .addFrameIndex(FI).addImm(0);
148 void ARMRegisterInfo::
149 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
150 unsigned DestReg, int FI,
151 const TargetRegisterClass *RC) const {
152 if (RC == ARM::GPRRegisterClass) {
153 MachineFunction &MF = *MBB.getParent();
154 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
155 if (AFI->isThumbFunction())
156 BuildMI(MBB, I, TII.get(ARM::tRestore), DestReg)
157 .addFrameIndex(FI).addImm(0);
159 BuildMI(MBB, I, TII.get(ARM::LDR), DestReg)
160 .addFrameIndex(FI).addReg(0).addImm(0);
161 } else if (RC == ARM::DPRRegisterClass) {
162 BuildMI(MBB, I, TII.get(ARM::FLDD), DestReg)
163 .addFrameIndex(FI).addImm(0);
165 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
166 BuildMI(MBB, I, TII.get(ARM::FLDS), DestReg)
167 .addFrameIndex(FI).addImm(0);
171 void ARMRegisterInfo::copyRegToReg(MachineBasicBlock &MBB,
172 MachineBasicBlock::iterator I,
173 unsigned DestReg, unsigned SrcReg,
174 const TargetRegisterClass *RC) const {
175 if (RC == ARM::GPRRegisterClass) {
176 MachineFunction &MF = *MBB.getParent();
177 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
178 BuildMI(MBB, I, TII.get(AFI->isThumbFunction() ? ARM::tMOVrr : ARM::MOVrr),
179 DestReg).addReg(SrcReg);
180 } else if (RC == ARM::SPRRegisterClass)
181 BuildMI(MBB, I, TII.get(ARM::FCPYS), DestReg).addReg(SrcReg);
182 else if (RC == ARM::DPRRegisterClass)
183 BuildMI(MBB, I, TII.get(ARM::FCPYD), DestReg).addReg(SrcReg);
188 /// isLowRegister - Returns true if the register is low register r0-r7.
190 static bool isLowRegister(unsigned Reg) {
193 case R0: case R1: case R2: case R3:
194 case R4: case R5: case R6: case R7:
201 MachineInstr *ARMRegisterInfo::foldMemoryOperand(MachineInstr *MI,
202 unsigned OpNum, int FI) const {
203 unsigned Opc = MI->getOpcode();
204 MachineInstr *NewMI = NULL;
208 if (OpNum == 0) { // move -> store
209 unsigned SrcReg = MI->getOperand(1).getReg();
210 NewMI = BuildMI(TII.get(ARM::STR)).addReg(SrcReg).addFrameIndex(FI)
211 .addReg(0).addImm(0);
212 } else { // move -> load
213 unsigned DstReg = MI->getOperand(0).getReg();
214 NewMI = BuildMI(TII.get(ARM::LDR), DstReg).addFrameIndex(FI).addReg(0)
220 if (OpNum == 0) { // move -> store
221 unsigned SrcReg = MI->getOperand(1).getReg();
222 if (!isLowRegister(SrcReg))
223 // tSpill cannot take a high register operand.
225 NewMI = BuildMI(TII.get(ARM::tSpill)).addReg(SrcReg).addFrameIndex(FI)
227 } else { // move -> load
228 unsigned DstReg = MI->getOperand(0).getReg();
229 if (!isLowRegister(DstReg))
230 // tRestore cannot target a high register operand.
232 NewMI = BuildMI(TII.get(ARM::tRestore), DstReg).addFrameIndex(FI)
238 if (OpNum == 0) { // move -> store
239 unsigned SrcReg = MI->getOperand(1).getReg();
240 NewMI = BuildMI(TII.get(ARM::FSTS)).addReg(SrcReg).addFrameIndex(FI)
242 } else { // move -> load
243 unsigned DstReg = MI->getOperand(0).getReg();
244 NewMI = BuildMI(TII.get(ARM::FLDS), DstReg).addFrameIndex(FI).addImm(0);
249 if (OpNum == 0) { // move -> store
250 unsigned SrcReg = MI->getOperand(1).getReg();
251 NewMI = BuildMI(TII.get(ARM::FSTD)).addReg(SrcReg).addFrameIndex(FI)
253 } else { // move -> load
254 unsigned DstReg = MI->getOperand(0).getReg();
255 NewMI = BuildMI(TII.get(ARM::FLDD), DstReg).addFrameIndex(FI).addImm(0);
262 NewMI->copyKillDeadInfo(MI);
266 const unsigned* ARMRegisterInfo::getCalleeSavedRegs() const {
267 static const unsigned CalleeSavedRegs[] = {
268 ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8,
269 ARM::R7, ARM::R6, ARM::R5, ARM::R4,
271 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
272 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
276 static const unsigned DarwinCalleeSavedRegs[] = {
277 ARM::LR, ARM::R7, ARM::R6, ARM::R5, ARM::R4,
278 ARM::R11, ARM::R10, ARM::R9, ARM::R8,
280 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
281 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
284 return STI.isTargetDarwin() ? DarwinCalleeSavedRegs : CalleeSavedRegs;
287 const TargetRegisterClass* const *
288 ARMRegisterInfo::getCalleeSavedRegClasses() const {
289 static const TargetRegisterClass * const CalleeSavedRegClasses[] = {
290 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
291 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
292 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
294 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
295 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
298 return CalleeSavedRegClasses;
301 /// hasFP - Return true if the specified function should have a dedicated frame
302 /// pointer register. This is true if the function has variable sized allocas
303 /// or if frame pointer elimination is disabled.
305 bool ARMRegisterInfo::hasFP(const MachineFunction &MF) const {
306 return NoFramePointerElim || MF.getFrameInfo()->hasVarSizedObjects();
309 /// emitARMRegPlusImmediate - Emits a series of instructions to materialize
310 /// a destreg = basereg + immediate in ARM code.
312 void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
313 MachineBasicBlock::iterator &MBBI,
314 unsigned DestReg, unsigned BaseReg,
315 int NumBytes, const TargetInstrInfo &TII) {
316 bool isSub = NumBytes < 0;
317 if (isSub) NumBytes = -NumBytes;
320 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
321 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
322 assert(ThisVal && "Didn't extract field correctly");
324 // We will handle these bits from offset, clear them.
325 NumBytes &= ~ThisVal;
327 // Get the properly encoded SOImmVal field.
328 int SOImmVal = ARM_AM::getSOImmVal(ThisVal);
329 assert(SOImmVal != -1 && "Bit extraction didn't work?");
331 // Build the new ADD / SUB.
332 BuildMI(MBB, MBBI, TII.get(isSub ? ARM::SUBri : ARM::ADDri), DestReg)
333 .addReg(BaseReg).addImm(SOImmVal);
338 /// calcNumMI - Returns the number of instructions required to materialize
339 /// the specific add / sub r, c instruction.
340 static unsigned calcNumMI(int Opc, int ExtraOpc, unsigned Bytes,
341 unsigned NumBits, unsigned Scale) {
343 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
345 if (Opc == ARM::tADDrSPi) {
346 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
351 Chunk = ((1 << NumBits) - 1) * Scale;
354 NumMIs += Bytes / Chunk;
355 if ((Bytes % Chunk) != 0)
362 /// emitLoadConstPool - Emits a load from constpool to materialize NumBytes
364 static void emitLoadConstPool(MachineBasicBlock &MBB,
365 MachineBasicBlock::iterator &MBBI,
366 unsigned DestReg, int NumBytes,
367 const TargetInstrInfo &TII) {
368 MachineFunction &MF = *MBB.getParent();
369 MachineConstantPool *ConstantPool = MF.getConstantPool();
370 Constant *C = ConstantInt::get(Type::Int32Ty, NumBytes);
371 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 2);
372 BuildMI(MBB, MBBI, TII.get(ARM::tLDRpci), DestReg).addConstantPoolIndex(Idx);
375 /// emitThumbRegPlusConstPool - Emits a series of instructions to materialize
376 /// a destreg = basereg + immediate in Thumb code. Load the immediate from a
379 void emitThumbRegPlusConstPool(MachineBasicBlock &MBB,
380 MachineBasicBlock::iterator &MBBI,
381 unsigned DestReg, unsigned BaseReg,
382 int NumBytes, bool CanChangeCC,
383 const TargetInstrInfo &TII) {
384 bool isHigh = !isLowRegister(DestReg) ||
385 (BaseReg != 0 && !isLowRegister(BaseReg));
387 // Subtract doesn't have high register version. Load the negative value
388 // if either base or dest register is a high register. Also, if do not
389 // issue sub as part of the sequence if condition register is to be
391 if (NumBytes < 0 && !isHigh && CanChangeCC) {
393 NumBytes = -NumBytes;
395 unsigned LdReg = DestReg;
396 if (DestReg == ARM::SP) {
397 assert(BaseReg == ARM::SP && "Unexpected!");
399 BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), ARM::R12).addReg(ARM::R3);
402 if (NumBytes <= 255 && NumBytes >= 0)
403 BuildMI(MBB, MBBI, TII.get(ARM::tMOVri8), LdReg).addImm(NumBytes);
405 emitLoadConstPool(MBB, MBBI, LdReg, NumBytes, TII);
408 int Opc = (isSub) ? ARM::tSUBrr : (isHigh ? ARM::tADDhirr : ARM::tADDrr);
409 const MachineInstrBuilder MIB = BuildMI(MBB, MBBI, TII.get(Opc), DestReg);
410 if (DestReg == ARM::SP)
411 MIB.addReg(BaseReg).addReg(LdReg);
413 MIB.addReg(BaseReg).addReg(LdReg);
415 MIB.addReg(LdReg).addReg(BaseReg);
416 if (DestReg == ARM::SP)
417 BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), ARM::R3).addReg(ARM::R12);
420 /// emitThumbRegPlusImmediate - Emits a series of instructions to materialize
421 /// a destreg = basereg + immediate in Thumb code.
423 void emitThumbRegPlusImmediate(MachineBasicBlock &MBB,
424 MachineBasicBlock::iterator &MBBI,
425 unsigned DestReg, unsigned BaseReg,
426 int NumBytes, const TargetInstrInfo &TII) {
427 bool isSub = NumBytes < 0;
428 unsigned Bytes = (unsigned)NumBytes;
429 if (isSub) Bytes = -NumBytes;
430 bool isMul4 = (Bytes & 3) == 0;
431 bool isTwoAddr = false;
432 bool DstNotEqBase = false;
433 unsigned NumBits = 1;
438 if (DestReg == BaseReg && BaseReg == ARM::SP) {
439 assert(isMul4 && "Thumb sp inc / dec size must be multiple of 4!");
442 Opc = isSub ? ARM::tSUBspi : ARM::tADDspi;
444 } else if (!isSub && BaseReg == ARM::SP) {
447 // r1 = add sp, 100 * 4
451 ExtraOpc = ARM::tADDi3;
460 if (DestReg != BaseReg)
463 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
467 unsigned NumMIs = calcNumMI(Opc, ExtraOpc, Bytes, NumBits, Scale);
468 unsigned Threshold = (DestReg == ARM::SP) ? 3 : 2;
469 if (NumMIs > Threshold) {
470 // This will expand into too many instructions. Load the immediate from a
472 emitThumbRegPlusConstPool(MBB, MBBI, DestReg, BaseReg, NumBytes, true, TII);
477 if (isLowRegister(DestReg) && isLowRegister(BaseReg)) {
478 // If both are low registers, emit DestReg = add BaseReg, max(Imm, 7)
479 unsigned Chunk = (1 << 3) - 1;
480 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
482 BuildMI(MBB, MBBI, TII.get(isSub ? ARM::tSUBi3 : ARM::tADDi3), DestReg)
483 .addReg(BaseReg).addImm(ThisVal);
485 BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), DestReg).addReg(BaseReg);
490 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
492 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
495 // Build the new tADD / tSUB.
497 BuildMI(MBB, MBBI, TII.get(Opc), DestReg).addReg(DestReg).addImm(ThisVal);
499 BuildMI(MBB, MBBI, TII.get(Opc), DestReg).addReg(BaseReg).addImm(ThisVal);
502 if (Opc == ARM::tADDrSPi) {
508 Chunk = ((1 << NumBits) - 1) * Scale;
509 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
516 BuildMI(MBB, MBBI, TII.get(ExtraOpc), DestReg).addReg(DestReg)
517 .addImm(((unsigned)NumBytes) & 3);
521 void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
522 int NumBytes, bool isThumb, const TargetInstrInfo &TII) {
524 emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII);
526 emitARMRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII);
529 void ARMRegisterInfo::
530 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
531 MachineBasicBlock::iterator I) const {
533 // If we have alloca, convert as follows:
534 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
535 // ADJCALLSTACKUP -> add, sp, sp, amount
536 MachineInstr *Old = I;
537 unsigned Amount = Old->getOperand(0).getImmedValue();
539 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
540 // We need to keep the stack aligned properly. To do this, we round the
541 // amount of space needed for the outgoing arguments up to the next
542 // alignment boundary.
543 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
544 Amount = (Amount+Align-1)/Align*Align;
546 // Replace the pseudo instruction with a new instruction...
547 if (Old->getOpcode() == ARM::ADJCALLSTACKDOWN) {
548 emitSPUpdate(MBB, I, -Amount, AFI->isThumbFunction(), TII);
550 assert(Old->getOpcode() == ARM::ADJCALLSTACKUP);
551 emitSPUpdate(MBB, I, Amount, AFI->isThumbFunction(), TII);
558 /// emitThumbConstant - Emit a series of instructions to materialize a
560 static void emitThumbConstant(MachineBasicBlock &MBB,
561 MachineBasicBlock::iterator &MBBI,
562 unsigned DestReg, int Imm,
563 const TargetInstrInfo &TII) {
564 bool isSub = Imm < 0;
565 if (isSub) Imm = -Imm;
567 int Chunk = (1 << 8) - 1;
568 int ThisVal = (Imm > Chunk) ? Chunk : Imm;
570 BuildMI(MBB, MBBI, TII.get(ARM::tMOVri8), DestReg).addImm(ThisVal);
572 emitThumbRegPlusImmediate(MBB, MBBI, DestReg, DestReg, Imm, TII);
574 BuildMI(MBB, MBBI, TII.get(ARM::tNEG), DestReg).addReg(DestReg);
577 void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II) const{
579 MachineInstr &MI = *II;
580 MachineBasicBlock &MBB = *MI.getParent();
581 MachineFunction &MF = *MBB.getParent();
582 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
583 bool isThumb = AFI->isThumbFunction();
585 while (!MI.getOperand(i).isFrameIndex()) {
587 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
590 unsigned FrameReg = ARM::SP;
591 int FrameIndex = MI.getOperand(i).getFrameIndex();
592 int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex) +
593 MF.getFrameInfo()->getStackSize();
595 if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex))
596 Offset -= AFI->getGPRCalleeSavedArea1Offset();
597 else if (AFI->isGPRCalleeSavedArea2Frame(FrameIndex))
598 Offset -= AFI->getGPRCalleeSavedArea2Offset();
599 else if (AFI->isDPRCalleeSavedAreaFrame(FrameIndex))
600 Offset -= AFI->getDPRCalleeSavedAreaOffset();
601 else if (hasFP(MF)) {
602 // There is alloca()'s in this function, must reference off the frame
604 FrameReg = getFrameRegister(MF);
605 Offset -= AFI->getFramePtrSpillOffset();
608 unsigned Opcode = MI.getOpcode();
609 const TargetInstrDescriptor &Desc = TII.get(Opcode);
610 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
613 if (Opcode == ARM::ADDri) {
614 Offset += MI.getOperand(i+1).getImm();
616 // Turn it into a move.
617 MI.setInstrDescriptor(TII.get(ARM::MOVrr));
618 MI.getOperand(i).ChangeToRegister(FrameReg, false);
619 MI.RemoveOperand(i+1);
621 } else if (Offset < 0) {
624 MI.setInstrDescriptor(TII.get(ARM::SUBri));
627 // Common case: small offset, fits into instruction.
628 int ImmedOffset = ARM_AM::getSOImmVal(Offset);
629 if (ImmedOffset != -1) {
630 // Replace the FrameIndex with sp / fp
631 MI.getOperand(i).ChangeToRegister(FrameReg, false);
632 MI.getOperand(i+1).ChangeToImmediate(ImmedOffset);
636 // Otherwise, we fallback to common code below to form the imm offset with
637 // a sequence of ADDri instructions. First though, pull as much of the imm
638 // into this ADDri as possible.
639 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
640 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, (32-RotAmt) & 31);
642 // We will handle these bits from offset, clear them.
643 Offset &= ~ThisImmVal;
645 // Get the properly encoded SOImmVal field.
646 int ThisSOImmVal = ARM_AM::getSOImmVal(ThisImmVal);
647 assert(ThisSOImmVal != -1 && "Bit extraction didn't work?");
648 MI.getOperand(i+1).ChangeToImmediate(ThisSOImmVal);
649 } else if (Opcode == ARM::tADDrSPi) {
650 Offset += MI.getOperand(i+1).getImm();
651 assert((Offset & 3) == 0 &&
652 "Thumb add/sub sp, #imm immediate must be multiple of 4!");
654 // Turn it into a move.
655 MI.setInstrDescriptor(TII.get(ARM::tMOVrr));
656 MI.getOperand(i).ChangeToRegister(FrameReg, false);
657 MI.RemoveOperand(i+1);
661 // Common case: small offset, fits into instruction.
662 if (((Offset >> 2) & ~255U) == 0) {
663 // Replace the FrameIndex with sp / fp
664 MI.getOperand(i).ChangeToRegister(FrameReg, false);
665 MI.getOperand(i+1).ChangeToImmediate(Offset >> 2);
669 unsigned DestReg = MI.getOperand(0).getReg();
670 unsigned Bytes = (Offset > 0) ? Offset : -Offset;
671 unsigned NumMIs = calcNumMI(Opcode, 0, Bytes, 8, 1);
672 // MI would expand into a large number of instructions. Don't try to
673 // simplify the immediate.
675 emitThumbRegPlusImmediate(MBB, II, DestReg, FrameReg, Offset, TII);
681 // Translate r0 = add sp, imm to
682 // r0 = add sp, 255*4
683 // r0 = add r0, (imm - 255*4)
684 MI.getOperand(i).ChangeToRegister(FrameReg, false);
685 MI.getOperand(i+1).ChangeToImmediate(255);
686 Offset = (Offset - 255 * 4);
687 MachineBasicBlock::iterator NII = next(II);
688 emitThumbRegPlusImmediate(MBB, NII, DestReg, DestReg, Offset, TII);
690 // Translate r0 = add sp, -imm to
691 // r0 = -imm (this is then translated into a series of instructons)
693 emitThumbConstant(MBB, II, DestReg, Offset, TII);
694 MI.setInstrDescriptor(TII.get(ARM::tADDhirr));
695 MI.getOperand(i).ChangeToRegister(DestReg, false);
696 MI.getOperand(i+1).ChangeToRegister(FrameReg, false);
702 unsigned NumBits = 0;
705 case ARMII::AddrMode2: {
707 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
708 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
713 case ARMII::AddrMode3: {
715 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
716 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
721 case ARMII::AddrMode5: {
723 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
724 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
730 case ARMII::AddrModeTs: {
732 InstrOffs = MI.getOperand(ImmIdx).getImm();
733 NumBits = (FrameReg == ARM::SP) ? 8 : 5;
738 std::cerr << "Unsupported addressing mode!\n";
743 Offset += InstrOffs * Scale;
744 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
745 if (Offset < 0 && !isThumb) {
750 // Common case: small offset, fits into instruction.
751 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
752 int ImmedOffset = Offset / Scale;
753 unsigned Mask = (1 << NumBits) - 1;
754 if ((unsigned)Offset <= Mask * Scale) {
755 // Replace the FrameIndex with sp
756 MI.getOperand(i).ChangeToRegister(FrameReg, false);
758 ImmedOffset |= 1 << NumBits;
759 ImmOp.ChangeToImmediate(ImmedOffset);
763 // If this is a thumb spill / restore, we will be using a constpool load to
764 // materialize the offset.
765 bool isThumSpillRestore = Opcode == ARM::tRestore || Opcode == ARM::tSpill;
766 if (AddrMode == ARMII::AddrModeTs && !isThumSpillRestore) {
767 if (AddrMode == ARMII::AddrModeTs) {
768 // Thumb tLDRspi, tSTRspi. These will change to instructions that use
769 // a different base register.
771 Mask = (1 << NumBits) - 1;
773 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
774 ImmedOffset = ImmedOffset & Mask;
776 ImmedOffset |= 1 << NumBits;
777 ImmOp.ChangeToImmediate(ImmedOffset);
778 Offset &= ~(Mask*Scale);
782 // If we get here, the immediate doesn't fit into the instruction. We folded
783 // as much as possible above, handle the rest, providing a register that is
785 assert(Offset && "This code isn't needed if offset already handled!");
788 if (TII.isLoad(Opcode)) {
789 // Use the destination register to materialize sp + offset.
790 unsigned TmpReg = MI.getOperand(0).getReg();
792 if (Opcode == ARM::tRestore) {
793 if (FrameReg == ARM::SP)
794 emitThumbRegPlusConstPool(MBB, II, TmpReg, FrameReg,Offset,false,TII);
796 emitLoadConstPool(MBB, II, TmpReg, Offset, TII);
800 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII);
801 MI.setInstrDescriptor(TII.get(ARM::tLDR));
802 MI.getOperand(i).ChangeToRegister(TmpReg, false);
804 MI.addRegOperand(FrameReg, false); // Use [reg, reg] addrmode.
806 MI.addRegOperand(0, false); // tLDR has an extra register operand.
807 } else if (TII.isStore(Opcode)) {
808 // FIXME! This is horrific!!! We need register scavenging.
809 // Our temporary workaround has marked r3 unavailable. Of course, r3 is
810 // also a ABI register so it's possible that is is the register that is
811 // being storing here. If that's the case, we do the following:
813 // Use r2 to materialize sp + offset
816 unsigned ValReg = MI.getOperand(0).getReg();
817 unsigned TmpReg = ARM::R3;
819 if (ValReg == ARM::R3) {
820 BuildMI(MBB, II, TII.get(ARM::tMOVrr), ARM::R12).addReg(ARM::R2);
823 if (Opcode == ARM::tSpill) {
824 if (FrameReg == ARM::SP)
825 emitThumbRegPlusConstPool(MBB, II, TmpReg, FrameReg,Offset,false,TII);
827 emitLoadConstPool(MBB, II, TmpReg, Offset, TII);
831 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII);
832 MI.setInstrDescriptor(TII.get(ARM::tSTR));
833 MI.getOperand(i).ChangeToRegister(TmpReg, false);
835 MI.addRegOperand(FrameReg, false); // Use [reg, reg] addrmode.
837 MI.addRegOperand(0, false); // tSTR has an extra register operand.
838 if (ValReg == ARM::R3) {
839 MachineBasicBlock::iterator NII = next(II);
840 BuildMI(MBB, NII, TII.get(ARM::tMOVrr), ARM::R2).addReg(ARM::R12);
843 assert(false && "Unexpected opcode!");
845 // Insert a set of r12 with the full address: r12 = sp + offset
846 // If the offset we have is too large to fit into the instruction, we need
847 // to form it with a series of ADDri's. Do this by taking 8-bit chunks
849 emitARMRegPlusImmediate(MBB, II, ARM::R12, FrameReg,
850 isSub ? -Offset : Offset, TII);
851 MI.getOperand(i).ChangeToRegister(ARM::R12, false);
855 void ARMRegisterInfo::
856 processFunctionBeforeCalleeSavedScan(MachineFunction &MF) const {
857 // This tells PEI to spill the FP as if it is any other callee-save register
858 // to take advantage the eliminateFrameIndex machinery. This also ensures it
859 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
860 // to combine multiple loads / stores.
861 bool CanEliminateFrame = true;
862 bool CS1Spilled = false;
863 bool LRSpilled = false;
864 unsigned NumGPRSpills = 0;
865 SmallVector<unsigned, 4> UnspilledCS1GPRs;
866 SmallVector<unsigned, 4> UnspilledCS2GPRs;
868 // Don't spill FP if the frame can be eliminated. This is determined
869 // by scanning the callee-save registers to see if any is used.
870 const unsigned *CSRegs = getCalleeSavedRegs();
871 const TargetRegisterClass* const *CSRegClasses = getCalleeSavedRegClasses();
872 for (unsigned i = 0; CSRegs[i]; ++i) {
873 unsigned Reg = CSRegs[i];
874 bool Spilled = false;
875 if (MF.isPhysRegUsed(Reg)) {
877 CanEliminateFrame = false;
879 // Check alias registers too.
880 for (const unsigned *Aliases = getAliasSet(Reg); *Aliases; ++Aliases) {
881 if (MF.isPhysRegUsed(*Aliases)) {
883 CanEliminateFrame = false;
888 if (CSRegClasses[i] == &ARM::GPRRegClass) {
892 if (!STI.isTargetDarwin()) {
900 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
915 if (!STI.isTargetDarwin()) {
916 UnspilledCS1GPRs.push_back(Reg);
926 UnspilledCS1GPRs.push_back(Reg);
929 UnspilledCS2GPRs.push_back(Reg);
936 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
937 bool ForceLRSpill = false;
938 if (!LRSpilled && AFI->isThumbFunction()) {
939 unsigned FnSize = ARM::GetFunctionSize(MF);
940 // Force LR spill if the Thumb function size is > 2048. This enables the
941 // use of BL to implement far jump. If it turns out that it's not needed
942 // the branch fix up path will undo it.
943 if (FnSize >= (1 << 11)) {
944 CanEliminateFrame = false;
949 if (!CanEliminateFrame || hasFP(MF)) {
950 AFI->setHasStackFrame(true);
952 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
953 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
954 if (!LRSpilled && CS1Spilled) {
955 MF.changePhyRegUsed(ARM::LR, true);
957 UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(),
958 UnspilledCS1GPRs.end(), (unsigned)ARM::LR));
959 ForceLRSpill = false;
962 // Darwin ABI requires FP to point to the stack slot that contains the
964 if (STI.isTargetDarwin() || hasFP(MF)) {
965 MF.changePhyRegUsed(FramePtr, true);
969 // If stack and double are 8-byte aligned and we are spilling an odd number
970 // of GPRs. Spill one extra callee save GPR so we won't have to pad between
971 // the integer and double callee save areas.
972 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
973 if (TargetAlign == 8 && (NumGPRSpills & 1)) {
974 if (CS1Spilled && !UnspilledCS1GPRs.empty())
975 MF.changePhyRegUsed(UnspilledCS1GPRs.front(), true);
976 else if (!UnspilledCS2GPRs.empty())
977 MF.changePhyRegUsed(UnspilledCS2GPRs.front(), true);
982 MF.changePhyRegUsed(ARM::LR, true);
983 AFI->setLRIsForceSpilled(true);
987 /// Move iterator pass the next bunch of callee save load / store ops for
988 /// the particular spill area (1: integer area 1, 2: integer area 2,
989 /// 3: fp area, 0: don't care).
990 static void movePastCSLoadStoreOps(MachineBasicBlock &MBB,
991 MachineBasicBlock::iterator &MBBI,
992 int Opc, unsigned Area,
993 const ARMSubtarget &STI) {
994 while (MBBI != MBB.end() &&
995 MBBI->getOpcode() == Opc && MBBI->getOperand(1).isFrameIndex()) {
998 unsigned Category = 0;
999 switch (MBBI->getOperand(0).getReg()) {
1000 case ARM::R4: case ARM::R5: case ARM::R6: case ARM::R7:
1004 case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11:
1005 Category = STI.isTargetDarwin() ? 2 : 1;
1007 case ARM::D8: case ARM::D9: case ARM::D10: case ARM::D11:
1008 case ARM::D12: case ARM::D13: case ARM::D14: case ARM::D15:
1015 if (Done || Category != Area)
1023 void ARMRegisterInfo::emitPrologue(MachineFunction &MF) const {
1024 MachineBasicBlock &MBB = MF.front();
1025 MachineBasicBlock::iterator MBBI = MBB.begin();
1026 MachineFrameInfo *MFI = MF.getFrameInfo();
1027 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1028 bool isThumb = AFI->isThumbFunction();
1029 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1030 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1031 unsigned NumBytes = MFI->getStackSize();
1032 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
1035 // Thumb add/sub sp, imm8 instructions implicitly multiply the offset by 4.
1036 NumBytes = (NumBytes + 3) & ~3;
1037 MFI->setStackSize(NumBytes);
1040 // Determine the sizes of each callee-save spill areas and record which frame
1041 // belongs to which callee-save spill areas.
1042 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
1043 int FramePtrSpillFI = 0;
1044 if (!AFI->hasStackFrame()) {
1046 emitSPUpdate(MBB, MBBI, -NumBytes, isThumb, TII);
1051 emitSPUpdate(MBB, MBBI, -VARegSaveSize, isThumb, TII);
1053 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1054 unsigned Reg = CSI[i].getReg();
1055 int FI = CSI[i].getFrameIdx();
1062 if (Reg == FramePtr)
1063 FramePtrSpillFI = FI;
1064 AFI->addGPRCalleeSavedArea1Frame(FI);
1071 if (Reg == FramePtr)
1072 FramePtrSpillFI = FI;
1073 if (STI.isTargetDarwin()) {
1074 AFI->addGPRCalleeSavedArea2Frame(FI);
1077 AFI->addGPRCalleeSavedArea1Frame(FI);
1082 AFI->addDPRCalleeSavedAreaFrame(FI);
1087 if (Align == 8 && (GPRCS1Size & 7) != 0)
1088 // Pad CS1 to ensure proper alignment.
1092 // Build the new SUBri to adjust SP for integer callee-save spill area 1.
1093 emitSPUpdate(MBB, MBBI, -GPRCS1Size, isThumb, TII);
1094 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 1, STI);
1095 } else if (MBBI != MBB.end() && MBBI->getOpcode() == ARM::tPUSH)
1098 // Darwin ABI requires FP to point to the stack slot that contains the
1100 if (STI.isTargetDarwin() || hasFP(MF))
1101 BuildMI(MBB, MBBI, TII.get(isThumb ? ARM::tADDrSPi : ARM::ADDri), FramePtr)
1102 .addFrameIndex(FramePtrSpillFI).addImm(0);
1105 // Build the new SUBri to adjust SP for integer callee-save spill area 2.
1106 emitSPUpdate(MBB, MBBI, -GPRCS2Size, false, TII);
1108 // Build the new SUBri to adjust SP for FP callee-save spill area.
1109 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 2, STI);
1110 emitSPUpdate(MBB, MBBI, -DPRCSSize, false, TII);
1113 // Determine starting offsets of spill areas.
1114 unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize);
1115 unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize;
1116 unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size;
1117 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) + NumBytes);
1118 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
1119 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
1120 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
1122 NumBytes = DPRCSOffset;
1124 // Insert it after all the callee-save spills.
1126 movePastCSLoadStoreOps(MBB, MBBI, ARM::FSTD, 3, STI);
1127 emitSPUpdate(MBB, MBBI, -NumBytes, isThumb, TII);
1130 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
1131 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
1132 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
1135 static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) {
1136 for (unsigned i = 0; CSRegs[i]; ++i)
1137 if (Reg == CSRegs[i])
1142 static bool isCSRestore(MachineInstr *MI, const unsigned *CSRegs) {
1143 return ((MI->getOpcode() == ARM::FLDD ||
1144 MI->getOpcode() == ARM::LDR ||
1145 MI->getOpcode() == ARM::tRestore) &&
1146 MI->getOperand(1).isFrameIndex() &&
1147 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs));
1150 void ARMRegisterInfo::emitEpilogue(MachineFunction &MF,
1151 MachineBasicBlock &MBB) const {
1152 MachineBasicBlock::iterator MBBI = prior(MBB.end());
1153 assert((MBBI->getOpcode() == ARM::BX_RET ||
1154 MBBI->getOpcode() == ARM::tBX_RET ||
1155 MBBI->getOpcode() == ARM::tPOP_RET) &&
1156 "Can only insert epilog into returning blocks");
1158 MachineFrameInfo *MFI = MF.getFrameInfo();
1159 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1160 bool isThumb = AFI->isThumbFunction();
1161 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1162 int NumBytes = (int)MFI->getStackSize();
1163 if (!AFI->hasStackFrame()) {
1165 emitSPUpdate(MBB, MBBI, NumBytes, isThumb, TII);
1169 // Unwind MBBI to point to first LDR / FLDD.
1170 const unsigned *CSRegs = getCalleeSavedRegs();
1171 if (MBBI != MBB.begin()) {
1174 while (MBBI != MBB.begin() && isCSRestore(MBBI, CSRegs));
1175 if (!isCSRestore(MBBI, CSRegs))
1179 // Move SP to start of FP callee save spill area.
1180 NumBytes -= (AFI->getGPRCalleeSavedArea1Size() +
1181 AFI->getGPRCalleeSavedArea2Size() +
1182 AFI->getDPRCalleeSavedAreaSize());
1185 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1186 // Reset SP based on frame pointer only if the stack frame extends beyond
1187 // frame pointer stack slot or target is ELF and the function has FP.
1189 emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, FramePtr, -NumBytes, TII);
1191 BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), ARM::SP).addReg(FramePtr);
1193 if (MBBI->getOpcode() == ARM::tBX_RET &&
1194 &MBB.front() != MBBI &&
1195 prior(MBBI)->getOpcode() == ARM::tPOP) {
1196 MachineBasicBlock::iterator PMBBI = prior(MBBI);
1197 emitSPUpdate(MBB, PMBBI, NumBytes, isThumb, TII);
1199 emitSPUpdate(MBB, MBBI, NumBytes, isThumb, TII);
1202 // Darwin ABI requires FP to point to the stack slot that contains the
1204 if (STI.isTargetDarwin() || hasFP(MF)) {
1205 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1206 // Reset SP based on frame pointer only if the stack frame extends beyond
1207 // frame pointer stack slot or target is ELF and the function has FP.
1208 if (AFI->getGPRCalleeSavedArea2Size() ||
1209 AFI->getDPRCalleeSavedAreaSize() ||
1210 AFI->getDPRCalleeSavedAreaOffset()||
1213 BuildMI(MBB, MBBI, TII.get(ARM::SUBri), ARM::SP).addReg(FramePtr)
1216 BuildMI(MBB, MBBI, TII.get(ARM::MOVrr), ARM::SP).addReg(FramePtr);
1217 } else if (NumBytes) {
1218 emitSPUpdate(MBB, MBBI, NumBytes, false, TII);
1221 // Move SP to start of integer callee save spill area 2.
1222 movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 3, STI);
1223 emitSPUpdate(MBB, MBBI, AFI->getDPRCalleeSavedAreaSize(), false, TII);
1225 // Move SP to start of integer callee save spill area 1.
1226 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 2, STI);
1227 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea2Size(), false, TII);
1229 // Move SP to SP upon entry to the function.
1230 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 1, STI);
1231 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea1Size(), false, TII);
1234 if (VARegSaveSize) {
1236 // Epilogue for vararg functions: pop LR to R3 and branch off it.
1237 // FIXME: Verify this is still ok when R3 is no longer being reserved.
1238 BuildMI(MBB, MBBI, TII.get(ARM::tPOP)).addReg(ARM::R3);
1240 emitSPUpdate(MBB, MBBI, VARegSaveSize, isThumb, TII);
1243 BuildMI(MBB, MBBI, TII.get(ARM::tBX_RET_vararg)).addReg(ARM::R3);
1249 unsigned ARMRegisterInfo::getRARegister() const {
1253 unsigned ARMRegisterInfo::getFrameRegister(MachineFunction &MF) const {
1254 return STI.useThumbBacktraces() ? ARM::R7 : ARM::R11;
1257 #include "ARMGenRegisterInfo.inc"