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 MachineInstr *PopMI = new MachineInstr(TII.get(ARM::tPOP));
108 MBB.insert(MI, PopMI);
109 for (unsigned i = CSI.size(); i != 0; --i) {
110 unsigned Reg = CSI[i-1].getReg();
111 if (Reg == ARM::LR) {
113 PopMI->setInstrDescriptor(TII.get(ARM::tPOP_RET));
116 PopMI->addRegOperand(Reg, true);
121 void ARMRegisterInfo::
122 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
123 unsigned SrcReg, int FI,
124 const TargetRegisterClass *RC) const {
125 if (RC == ARM::GPRRegisterClass) {
126 MachineFunction &MF = *MBB.getParent();
127 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
128 if (AFI->isThumbFunction())
129 BuildMI(MBB, I, TII.get(ARM::tSTRspi)).addReg(SrcReg)
130 .addFrameIndex(FI).addImm(0);
132 BuildMI(MBB, I, TII.get(ARM::STR)).addReg(SrcReg)
133 .addFrameIndex(FI).addReg(0).addImm(0);
134 } else if (RC == ARM::DPRRegisterClass) {
135 BuildMI(MBB, I, TII.get(ARM::FSTD)).addReg(SrcReg)
136 .addFrameIndex(FI).addImm(0);
138 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
139 BuildMI(MBB, I, TII.get(ARM::FSTS)).addReg(SrcReg)
140 .addFrameIndex(FI).addImm(0);
144 void ARMRegisterInfo::
145 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
146 unsigned DestReg, int FI,
147 const TargetRegisterClass *RC) const {
148 if (RC == ARM::GPRRegisterClass) {
149 MachineFunction &MF = *MBB.getParent();
150 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
151 if (AFI->isThumbFunction())
152 BuildMI(MBB, I, TII.get(ARM::tLDRspi), DestReg)
153 .addFrameIndex(FI).addImm(0);
155 BuildMI(MBB, I, TII.get(ARM::LDR), DestReg)
156 .addFrameIndex(FI).addReg(0).addImm(0);
157 } else if (RC == ARM::DPRRegisterClass) {
158 BuildMI(MBB, I, TII.get(ARM::FLDD), DestReg)
159 .addFrameIndex(FI).addImm(0);
161 assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
162 BuildMI(MBB, I, TII.get(ARM::FLDS), DestReg)
163 .addFrameIndex(FI).addImm(0);
167 void ARMRegisterInfo::copyRegToReg(MachineBasicBlock &MBB,
168 MachineBasicBlock::iterator I,
169 unsigned DestReg, unsigned SrcReg,
170 const TargetRegisterClass *RC) const {
171 if (RC == ARM::GPRRegisterClass) {
172 MachineFunction &MF = *MBB.getParent();
173 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
174 BuildMI(MBB, I, TII.get(AFI->isThumbFunction() ? ARM::tMOVrr : ARM::MOVrr),
175 DestReg).addReg(SrcReg);
176 } else if (RC == ARM::SPRRegisterClass)
177 BuildMI(MBB, I, TII.get(ARM::FCPYS), DestReg).addReg(SrcReg);
178 else if (RC == ARM::DPRRegisterClass)
179 BuildMI(MBB, I, TII.get(ARM::FCPYD), DestReg).addReg(SrcReg);
184 MachineInstr *ARMRegisterInfo::foldMemoryOperand(MachineInstr *MI,
185 unsigned OpNum, int FI) const {
186 unsigned Opc = MI->getOpcode();
187 MachineInstr *NewMI = NULL;
191 if (OpNum == 0) { // move -> store
192 unsigned SrcReg = MI->getOperand(1).getReg();
193 NewMI = BuildMI(TII.get(ARM::STR)).addReg(SrcReg).addFrameIndex(FI)
194 .addReg(0).addImm(0);
195 } else { // move -> load
196 unsigned DstReg = MI->getOperand(0).getReg();
197 NewMI = BuildMI(TII.get(ARM::LDR), DstReg).addFrameIndex(FI).addReg(0)
203 if (OpNum == 0) { // move -> store
204 unsigned SrcReg = MI->getOperand(1).getReg();
205 NewMI = BuildMI(TII.get(ARM::tSTRspi)).addReg(SrcReg).addFrameIndex(FI)
207 } else { // move -> load
208 unsigned DstReg = MI->getOperand(0).getReg();
209 NewMI = BuildMI(TII.get(ARM::tLDRspi), DstReg).addFrameIndex(FI)
215 if (OpNum == 0) { // move -> store
216 unsigned SrcReg = MI->getOperand(1).getReg();
217 NewMI = BuildMI(TII.get(ARM::FSTS)).addReg(SrcReg).addFrameIndex(FI)
219 } else { // move -> load
220 unsigned DstReg = MI->getOperand(0).getReg();
221 NewMI = BuildMI(TII.get(ARM::FLDS), DstReg).addFrameIndex(FI).addImm(0);
226 if (OpNum == 0) { // move -> store
227 unsigned SrcReg = MI->getOperand(1).getReg();
228 NewMI = BuildMI(TII.get(ARM::FSTD)).addReg(SrcReg).addFrameIndex(FI)
230 } else { // move -> load
231 unsigned DstReg = MI->getOperand(0).getReg();
232 NewMI = BuildMI(TII.get(ARM::FLDD), DstReg).addFrameIndex(FI).addImm(0);
239 NewMI->copyKillDeadInfo(MI);
243 const unsigned* ARMRegisterInfo::getCalleeSavedRegs() const {
244 static const unsigned CalleeSavedRegs[] = {
245 ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8,
246 ARM::R7, ARM::R6, ARM::R5, ARM::R4,
248 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
249 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
253 static const unsigned DarwinCalleeSavedRegs[] = {
254 ARM::LR, ARM::R7, ARM::R6, ARM::R5, ARM::R4,
255 ARM::R11, ARM::R10, ARM::R9, ARM::R8,
257 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
258 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
261 return STI.isTargetDarwin() ? DarwinCalleeSavedRegs : CalleeSavedRegs;
264 const TargetRegisterClass* const *
265 ARMRegisterInfo::getCalleeSavedRegClasses() const {
266 static const TargetRegisterClass * const CalleeSavedRegClasses[] = {
267 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
268 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
269 &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
271 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
272 &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass, &ARM::DPRRegClass,
275 return CalleeSavedRegClasses;
278 /// hasFP - Return true if the specified function should have a dedicated frame
279 /// pointer register. This is true if the function has variable sized allocas
280 /// or if frame pointer elimination is disabled.
282 bool ARMRegisterInfo::hasFP(const MachineFunction &MF) const {
283 return NoFramePointerElim || MF.getFrameInfo()->hasVarSizedObjects();
286 /// emitARMRegPlusImmediate - Emits a series of instructions to materialize
287 /// a destreg = basereg + immediate in ARM code.
289 void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
290 MachineBasicBlock::iterator &MBBI,
291 unsigned DestReg, unsigned BaseReg,
292 int NumBytes, const TargetInstrInfo &TII) {
293 bool isSub = NumBytes < 0;
294 if (isSub) NumBytes = -NumBytes;
297 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
298 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
299 assert(ThisVal && "Didn't extract field correctly");
301 // We will handle these bits from offset, clear them.
302 NumBytes &= ~ThisVal;
304 // Get the properly encoded SOImmVal field.
305 int SOImmVal = ARM_AM::getSOImmVal(ThisVal);
306 assert(SOImmVal != -1 && "Bit extraction didn't work?");
308 // Build the new ADD / SUB.
309 BuildMI(MBB, MBBI, TII.get(isSub ? ARM::SUBri : ARM::ADDri), DestReg)
310 .addReg(BaseReg).addImm(SOImmVal);
315 /// isLowRegister - Returns true if the register is low register r0-r7.
317 static bool isLowRegister(unsigned Reg) {
320 case R0: case R1: case R2: case R3:
321 case R4: case R5: case R6: case R7:
328 /// calcNumMI - Returns the number of instructions required to materialize
329 /// the specific add / sub r, c instruction.
330 static unsigned calcNumMI(int Opc, int ExtraOpc, unsigned Bytes,
331 unsigned NumBits, unsigned Scale) {
333 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
335 if (Opc == ARM::tADDrSPi) {
336 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
341 Chunk = ((1 << NumBits) - 1) * Scale;
344 NumMIs += Bytes / Chunk;
345 if ((Bytes % Chunk) != 0)
352 /// emitThumbRegPlusConstPool - Emits a series of instructions to materialize
353 /// a destreg = basereg + immediate in Thumb code. Load the immediate from a
356 void emitThumbRegPlusConstPool(MachineBasicBlock &MBB,
357 MachineBasicBlock::iterator &MBBI,
358 unsigned DestReg, unsigned BaseReg,
359 int NumBytes, const TargetInstrInfo &TII) {
360 MachineFunction &MF = *MBB.getParent();
361 MachineConstantPool *ConstantPool = MF.getConstantPool();
362 bool isHigh = !isLowRegister(DestReg) || !isLowRegister(BaseReg);
364 // Subtract doesn't have high register version. Load the negative value
365 // if either base or dest register is a high register.
366 if (NumBytes < 0 && !isHigh) {
368 NumBytes = -NumBytes;
370 Constant *C = ConstantInt::get(Type::Int32Ty, NumBytes);
371 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 2);
372 unsigned LdReg = DestReg;
373 if (DestReg == ARM::SP) {
374 assert(BaseReg == ARM::SP && "Unexpected!");
376 BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), ARM::R12).addReg(ARM::R3);
378 // Load the constant.
379 BuildMI(MBB, MBBI, TII.get(ARM::tLDRpci), LdReg).addConstantPoolIndex(Idx);
381 int Opc = (isSub) ? ARM::tSUBrr : (isHigh ? ARM::tADDhirr : ARM::tADDrr);
382 const MachineInstrBuilder MIB = BuildMI(MBB, MBBI, TII.get(Opc), DestReg);
383 if (DestReg == ARM::SP)
384 MIB.addReg(BaseReg).addReg(LdReg);
386 MIB.addReg(LdReg).addReg(BaseReg);
387 if (DestReg == ARM::SP)
388 BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), ARM::R3).addReg(ARM::R12);
391 /// emitThumbRegPlusImmediate - Emits a series of instructions to materialize
392 /// a destreg = basereg + immediate in Thumb code.
394 void emitThumbRegPlusImmediate(MachineBasicBlock &MBB,
395 MachineBasicBlock::iterator &MBBI,
396 unsigned DestReg, unsigned BaseReg,
397 int NumBytes, const TargetInstrInfo &TII) {
398 bool isSub = NumBytes < 0;
399 unsigned Bytes = (unsigned)NumBytes;
400 if (isSub) Bytes = -NumBytes;
401 bool isMul4 = (Bytes & 3) == 0;
402 bool isTwoAddr = false;
403 bool DstNeBase = false;
404 unsigned NumBits = 1;
409 if (DestReg == BaseReg && BaseReg == ARM::SP) {
410 assert(isMul4 && "Thumb sp inc / dec size must be multiple of 4!");
413 Opc = isSub ? ARM::tSUBspi : ARM::tADDspi;
415 } else if (!isSub && BaseReg == ARM::SP) {
418 // r1 = add sp, 100 * 4
422 ExtraOpc = ARM::tADDi3;
431 if (DestReg != BaseReg)
434 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
438 unsigned NumMIs = calcNumMI(Opc, ExtraOpc, Bytes, NumBits, Scale);
439 unsigned Threshold = (DestReg == ARM::SP) ? 4 : 3;
440 if (NumMIs > Threshold) {
441 // This will expand into too many instructions. Load the immediate from a
443 emitThumbRegPlusConstPool(MBB, MBBI, DestReg, BaseReg, NumBytes, TII);
448 if (isLowRegister(DestReg) && isLowRegister(BaseReg)) {
449 // If both are low registers, emit DestReg = add BaseReg, max(Imm, 7)
450 unsigned Chunk = (1 << 3) - 1;
451 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
453 BuildMI(MBB, MBBI, TII.get(isSub ? ARM::tSUBi3 : ARM::tADDi3), DestReg)
454 .addReg(BaseReg).addImm(ThisVal);
456 BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), DestReg).addReg(BaseReg);
461 unsigned Chunk = ((1 << NumBits) - 1) * Scale;
463 unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
466 // Build the new tADD / tSUB.
468 BuildMI(MBB, MBBI, TII.get(Opc), DestReg).addReg(DestReg).addImm(ThisVal);
470 BuildMI(MBB, MBBI, TII.get(Opc), DestReg).addReg(BaseReg).addImm(ThisVal);
473 if (Opc == ARM::tADDrSPi) {
479 Chunk = ((1 << NumBits) - 1) * Scale;
480 Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
487 BuildMI(MBB, MBBI, TII.get(ExtraOpc), DestReg).addReg(DestReg)
488 .addImm(((unsigned)NumBytes) & 3);
492 void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
493 int NumBytes, bool isThumb, const TargetInstrInfo &TII) {
495 emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII);
497 emitARMRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII);
500 void ARMRegisterInfo::
501 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
502 MachineBasicBlock::iterator I) const {
504 // If we have alloca, convert as follows:
505 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
506 // ADJCALLSTACKUP -> add, sp, sp, amount
507 MachineInstr *Old = I;
508 unsigned Amount = Old->getOperand(0).getImmedValue();
510 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
511 // We need to keep the stack aligned properly. To do this, we round the
512 // amount of space needed for the outgoing arguments up to the next
513 // alignment boundary.
514 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
515 Amount = (Amount+Align-1)/Align*Align;
517 // Replace the pseudo instruction with a new instruction...
518 if (Old->getOpcode() == ARM::ADJCALLSTACKDOWN) {
519 emitSPUpdate(MBB, I, -Amount, AFI->isThumbFunction(), TII);
521 assert(Old->getOpcode() == ARM::ADJCALLSTACKUP);
522 emitSPUpdate(MBB, I, Amount, AFI->isThumbFunction(), TII);
529 /// emitThumbConstant - Emit a series of instructions to materialize a
531 static void emitThumbConstant(MachineBasicBlock &MBB,
532 MachineBasicBlock::iterator &MBBI,
533 unsigned DestReg, int Imm,
534 const TargetInstrInfo &TII) {
535 bool isSub = Imm < 0;
536 if (isSub) Imm = -Imm;
538 int Chunk = (1 << 8) - 1;
539 int ThisVal = (Imm > Chunk) ? Chunk : Imm;
541 BuildMI(MBB, MBBI, TII.get(ARM::tMOVri8), DestReg).addImm(ThisVal);
543 emitThumbRegPlusImmediate(MBB, MBBI, DestReg, DestReg, Imm, TII);
545 BuildMI(MBB, MBBI, TII.get(ARM::tNEG), DestReg).addReg(DestReg);
548 void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II) const{
550 MachineInstr &MI = *II;
551 MachineBasicBlock &MBB = *MI.getParent();
552 MachineFunction &MF = *MBB.getParent();
553 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
554 bool isThumb = AFI->isThumbFunction();
556 while (!MI.getOperand(i).isFrameIndex()) {
558 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
561 unsigned FrameReg = ARM::SP;
562 int FrameIndex = MI.getOperand(i).getFrameIndex();
563 int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex) +
564 MF.getFrameInfo()->getStackSize();
566 if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex))
567 Offset -= AFI->getGPRCalleeSavedArea1Offset();
568 else if (AFI->isGPRCalleeSavedArea2Frame(FrameIndex))
569 Offset -= AFI->getGPRCalleeSavedArea2Offset();
570 else if (AFI->isDPRCalleeSavedAreaFrame(FrameIndex))
571 Offset -= AFI->getDPRCalleeSavedAreaOffset();
572 else if (hasFP(MF)) {
573 // There is alloca()'s in this function, must reference off the frame
575 FrameReg = getFrameRegister(MF);
576 Offset -= AFI->getFramePtrSpillOffset();
579 unsigned Opcode = MI.getOpcode();
580 const TargetInstrDescriptor &Desc = TII.get(Opcode);
581 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
584 if (Opcode == ARM::ADDri) {
585 Offset += MI.getOperand(i+1).getImm();
587 // Turn it into a move.
588 MI.setInstrDescriptor(TII.get(ARM::MOVrr));
589 MI.getOperand(i).ChangeToRegister(FrameReg, false);
590 MI.RemoveOperand(i+1);
592 } else if (Offset < 0) {
595 MI.setInstrDescriptor(TII.get(ARM::SUBri));
598 // Common case: small offset, fits into instruction.
599 int ImmedOffset = ARM_AM::getSOImmVal(Offset);
600 if (ImmedOffset != -1) {
601 // Replace the FrameIndex with sp / fp
602 MI.getOperand(i).ChangeToRegister(FrameReg, false);
603 MI.getOperand(i+1).ChangeToImmediate(ImmedOffset);
607 // Otherwise, we fallback to common code below to form the imm offset with
608 // a sequence of ADDri instructions. First though, pull as much of the imm
609 // into this ADDri as possible.
610 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
611 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, (32-RotAmt) & 31);
613 // We will handle these bits from offset, clear them.
614 Offset &= ~ThisImmVal;
616 // Get the properly encoded SOImmVal field.
617 int ThisSOImmVal = ARM_AM::getSOImmVal(ThisImmVal);
618 assert(ThisSOImmVal != -1 && "Bit extraction didn't work?");
619 MI.getOperand(i+1).ChangeToImmediate(ThisSOImmVal);
620 } else if (Opcode == ARM::tADDrSPi) {
621 Offset += MI.getOperand(i+1).getImm();
622 assert((Offset & 3) == 0 &&
623 "Thumb add/sub sp, #imm immediate must be multiple of 4!");
626 // Turn it into a move.
627 MI.setInstrDescriptor(TII.get(ARM::tMOVrr));
628 MI.getOperand(i).ChangeToRegister(FrameReg, false);
629 MI.RemoveOperand(i+1);
633 // Common case: small offset, fits into instruction.
634 if ((Offset & ~255U) == 0) {
635 // Replace the FrameIndex with sp / fp
636 MI.getOperand(i).ChangeToRegister(FrameReg, false);
637 MI.getOperand(i+1).ChangeToImmediate(Offset);
641 unsigned DestReg = MI.getOperand(0).getReg();
643 // Translate r0 = add sp, imm to
644 // r0 = add sp, 255*4
645 // r0 = add r0, (imm - 255*4)
646 MI.getOperand(i).ChangeToRegister(FrameReg, false);
647 MI.getOperand(i+1).ChangeToImmediate(255);
648 Offset = (Offset - 255) << 2;
649 MachineBasicBlock::iterator NII = next(II);
650 emitThumbRegPlusImmediate(MBB, NII, DestReg, DestReg, Offset, TII);
652 // Translate r0 = add sp, -imm to
653 // r0 = -imm (this is then translated into a series of instructons)
656 emitThumbConstant(MBB, II, DestReg, Offset, TII);
657 MI.setInstrDescriptor(TII.get(ARM::tADDhirr));
658 MI.getOperand(i).ChangeToRegister(DestReg, false);
659 MI.getOperand(i+1).ChangeToRegister(FrameReg, false);
665 unsigned NumBits = 0;
668 case ARMII::AddrMode2: {
670 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
671 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
676 case ARMII::AddrMode3: {
678 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
679 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
684 case ARMII::AddrMode5: {
686 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
687 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
693 case ARMII::AddrModeTs: {
695 InstrOffs = MI.getOperand(ImmIdx).getImm();
701 std::cerr << "Unsupported addressing mode!\n";
706 Offset += InstrOffs * Scale;
707 assert((Scale == 1 || (Offset & (Scale-1)) == 0) &&
708 "Can't encode this offset!");
714 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
715 int ImmedOffset = Offset / Scale;
716 unsigned Mask = (1 << NumBits) - 1;
717 if ((unsigned)Offset <= Mask * Scale) {
718 // Replace the FrameIndex with sp
719 MI.getOperand(i).ChangeToRegister(FrameReg, false);
721 ImmedOffset |= 1 << NumBits;
722 ImmOp.ChangeToImmediate(ImmedOffset);
726 // Otherwise, it didn't fit. Pull in what we can to simplify the immediate.
727 if (AddrMode == ARMII::AddrModeTs) {
728 // Thumb tLDRspi, tSTRspi. These will change to instructions that use a
729 // different base register.
731 Mask = (1 << NumBits) - 1;
734 ImmedOffset = ImmedOffset & Mask;
736 ImmedOffset |= 1 << NumBits;
737 ImmOp.ChangeToImmediate(ImmedOffset);
738 Offset &= ~(Mask*Scale);
741 // If we get here, the immediate doesn't fit into the instruction. We folded
742 // as much as possible above, handle the rest, providing a register that is
744 assert(Offset && "This code isn't needed if offset already handled!");
747 if (TII.isLoad(Opcode)) {
748 // Use the destination register to materialize sp + offset.
749 unsigned TmpReg = MI.getOperand(0).getReg();
750 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg,
751 isSub ? -Offset : Offset, TII);
752 MI.setInstrDescriptor(TII.get(ARM::tLDR));
753 MI.getOperand(i).ChangeToRegister(TmpReg, false);
754 MI.addRegOperand(0, false); // tLDR has an extra register operand.
755 } else if (TII.isStore(Opcode)) {
756 // FIXME! This is horrific!!! We need register scavenging.
757 // Our temporary workaround has marked r3 unavailable. Of course, r3 is
758 // also a ABI register so it's possible that is is the register that is
759 // being storing here. If that's the case, we do the following:
761 // Use r2 to materialize sp + offset
764 unsigned ValReg = MI.getOperand(0).getReg();
765 unsigned TmpReg = ARM::R3;
766 if (ValReg == ARM::R3) {
767 BuildMI(MBB, II, TII.get(ARM::tMOVrr), ARM::R12).addReg(ARM::R2);
770 emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg,
771 isSub ? -Offset : Offset, TII);
772 MI.setInstrDescriptor(TII.get(ARM::tSTR));
773 MI.getOperand(i).ChangeToRegister(TmpReg, false);
774 MI.addRegOperand(0, false); // tSTR has an extra register operand.
775 if (ValReg == ARM::R3)
776 BuildMI(MBB, II, TII.get(ARM::tMOVrr), ARM::R2).addReg(ARM::R12);
778 assert(false && "Unexpected opcode!");
780 // Insert a set of r12 with the full address: r12 = sp + offset
781 // If the offset we have is too large to fit into the instruction, we need
782 // to form it with a series of ADDri's. Do this by taking 8-bit chunks
784 emitARMRegPlusImmediate(MBB, II, ARM::R12, FrameReg,
785 isSub ? -Offset : Offset, TII);
786 MI.getOperand(i).ChangeToRegister(ARM::R12, false);
790 void ARMRegisterInfo::
791 processFunctionBeforeCalleeSavedScan(MachineFunction &MF) const {
792 // This tells PEI to spill the FP as if it is any other callee-save register
793 // to take advantage the eliminateFrameIndex machinery. This also ensures it
794 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
795 // to combine multiple loads / stores.
796 bool CanEliminateFrame = true;
797 bool CS1Spilled = false;
798 bool LRSpilled = false;
799 unsigned NumGPRSpills = 0;
800 SmallVector<unsigned, 4> UnspilledCS1GPRs;
801 SmallVector<unsigned, 4> UnspilledCS2GPRs;
803 // Don't spill FP if the frame can be eliminated. This is determined
804 // by scanning the callee-save registers to see if any is used.
805 const unsigned *CSRegs = getCalleeSavedRegs();
806 const TargetRegisterClass* const *CSRegClasses = getCalleeSavedRegClasses();
807 for (unsigned i = 0; CSRegs[i]; ++i) {
808 unsigned Reg = CSRegs[i];
809 bool Spilled = false;
810 if (MF.isPhysRegUsed(Reg)) {
812 CanEliminateFrame = false;
814 // Check alias registers too.
815 for (const unsigned *Aliases = getAliasSet(Reg); *Aliases; ++Aliases) {
816 if (MF.isPhysRegUsed(*Aliases)) {
818 CanEliminateFrame = false;
823 if (CSRegClasses[i] == &ARM::GPRRegClass) {
827 if (!STI.isTargetDarwin()) {
835 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
850 if (!STI.isTargetDarwin()) {
851 UnspilledCS1GPRs.push_back(Reg);
861 UnspilledCS1GPRs.push_back(Reg);
864 UnspilledCS2GPRs.push_back(Reg);
871 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
872 bool ForceLRSpill = false;
873 if (!LRSpilled && AFI->isThumbFunction()) {
874 unsigned FnSize = ARM::GetFunctionSize(MF);
875 // Force LR spill if the Thumb function size is > 2048. This enables the
876 // use of BL to implement far jump. If it turns out that it's not needed
877 // the branch fix up path will undo it.
878 if (FnSize >= (1 << 11)) {
879 CanEliminateFrame = false;
884 if (!CanEliminateFrame || hasFP(MF)) {
885 AFI->setHasStackFrame(true);
887 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
888 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
889 if (!LRSpilled && CS1Spilled) {
890 MF.changePhyRegUsed(ARM::LR, true);
892 UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(),
893 UnspilledCS1GPRs.end(), (unsigned)ARM::LR));
894 ForceLRSpill = false;
897 // Darwin ABI requires FP to point to the stack slot that contains the
899 if (STI.isTargetDarwin() || hasFP(MF)) {
900 MF.changePhyRegUsed(FramePtr, true);
904 // If stack and double are 8-byte aligned and we are spilling an odd number
905 // of GPRs. Spill one extra callee save GPR so we won't have to pad between
906 // the integer and double callee save areas.
907 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
908 if (TargetAlign == 8 && (NumGPRSpills & 1)) {
909 if (CS1Spilled && !UnspilledCS1GPRs.empty())
910 MF.changePhyRegUsed(UnspilledCS1GPRs.front(), true);
911 else if (!UnspilledCS2GPRs.empty())
912 MF.changePhyRegUsed(UnspilledCS2GPRs.front(), true);
917 MF.changePhyRegUsed(ARM::LR, true);
918 AFI->setLRIsForceSpilled(true);
922 /// Move iterator pass the next bunch of callee save load / store ops for
923 /// the particular spill area (1: integer area 1, 2: integer area 2,
924 /// 3: fp area, 0: don't care).
925 static void movePastCSLoadStoreOps(MachineBasicBlock &MBB,
926 MachineBasicBlock::iterator &MBBI,
927 int Opc, unsigned Area,
928 const ARMSubtarget &STI) {
929 while (MBBI != MBB.end() &&
930 MBBI->getOpcode() == Opc && MBBI->getOperand(1).isFrameIndex()) {
933 unsigned Category = 0;
934 switch (MBBI->getOperand(0).getReg()) {
935 case ARM::R4: case ARM::R5: case ARM::R6: case ARM::R7:
939 case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11:
940 Category = STI.isTargetDarwin() ? 2 : 1;
942 case ARM::D8: case ARM::D9: case ARM::D10: case ARM::D11:
943 case ARM::D12: case ARM::D13: case ARM::D14: case ARM::D15:
950 if (Done || Category != Area)
958 void ARMRegisterInfo::emitPrologue(MachineFunction &MF) const {
959 MachineBasicBlock &MBB = MF.front();
960 MachineBasicBlock::iterator MBBI = MBB.begin();
961 MachineFrameInfo *MFI = MF.getFrameInfo();
962 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
963 bool isThumb = AFI->isThumbFunction();
964 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
965 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
966 unsigned NumBytes = MFI->getStackSize();
967 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
970 // Thumb add/sub sp, imm8 instructions implicitly multiply the offset by 4.
971 NumBytes = (NumBytes + 3) & ~3;
972 MFI->setStackSize(NumBytes);
975 // Determine the sizes of each callee-save spill areas and record which frame
976 // belongs to which callee-save spill areas.
977 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
978 int FramePtrSpillFI = 0;
979 if (!AFI->hasStackFrame()) {
981 emitSPUpdate(MBB, MBBI, -NumBytes, isThumb, TII);
986 emitSPUpdate(MBB, MBBI, -VARegSaveSize, isThumb, TII);
988 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
989 unsigned Reg = CSI[i].getReg();
990 int FI = CSI[i].getFrameIdx();
998 FramePtrSpillFI = FI;
999 AFI->addGPRCalleeSavedArea1Frame(FI);
1006 if (Reg == FramePtr)
1007 FramePtrSpillFI = FI;
1008 if (STI.isTargetDarwin()) {
1009 AFI->addGPRCalleeSavedArea2Frame(FI);
1012 AFI->addGPRCalleeSavedArea1Frame(FI);
1017 AFI->addDPRCalleeSavedAreaFrame(FI);
1022 if (Align == 8 && (GPRCS1Size & 7) != 0)
1023 // Pad CS1 to ensure proper alignment.
1027 // Build the new SUBri to adjust SP for integer callee-save spill area 1.
1028 emitSPUpdate(MBB, MBBI, -GPRCS1Size, isThumb, TII);
1029 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 1, STI);
1030 } else if (MBBI != MBB.end() && MBBI->getOpcode() == ARM::tPUSH)
1033 // Darwin ABI requires FP to point to the stack slot that contains the
1035 if (STI.isTargetDarwin() || hasFP(MF))
1036 BuildMI(MBB, MBBI, TII.get(isThumb ? ARM::tADDrSPi : ARM::ADDri), FramePtr)
1037 .addFrameIndex(FramePtrSpillFI).addImm(0);
1040 // Build the new SUBri to adjust SP for integer callee-save spill area 2.
1041 emitSPUpdate(MBB, MBBI, -GPRCS2Size, false, TII);
1043 // Build the new SUBri to adjust SP for FP callee-save spill area.
1044 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 2, STI);
1045 emitSPUpdate(MBB, MBBI, -DPRCSSize, false, TII);
1048 // Determine starting offsets of spill areas.
1049 unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize);
1050 unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize;
1051 unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size;
1052 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) + NumBytes);
1053 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
1054 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
1055 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
1057 NumBytes = DPRCSOffset;
1059 // Insert it after all the callee-save spills.
1061 movePastCSLoadStoreOps(MBB, MBBI, ARM::FSTD, 3, STI);
1062 emitSPUpdate(MBB, MBBI, -NumBytes, isThumb, TII);
1065 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
1066 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
1067 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
1070 static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) {
1071 for (unsigned i = 0; CSRegs[i]; ++i)
1072 if (Reg == CSRegs[i])
1077 static bool isCSRestore(MachineInstr *MI, const unsigned *CSRegs) {
1078 return ((MI->getOpcode() == ARM::FLDD ||
1079 MI->getOpcode() == ARM::LDR ||
1080 MI->getOpcode() == ARM::tLDRspi) &&
1081 MI->getOperand(1).isFrameIndex() &&
1082 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs));
1085 void ARMRegisterInfo::emitEpilogue(MachineFunction &MF,
1086 MachineBasicBlock &MBB) const {
1087 MachineBasicBlock::iterator MBBI = prior(MBB.end());
1088 assert((MBBI->getOpcode() == ARM::BX_RET ||
1089 MBBI->getOpcode() == ARM::tBX_RET ||
1090 MBBI->getOpcode() == ARM::tPOP_RET) &&
1091 "Can only insert epilog into returning blocks");
1093 MachineFrameInfo *MFI = MF.getFrameInfo();
1094 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1095 bool isThumb = AFI->isThumbFunction();
1096 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1097 int NumBytes = (int)MFI->getStackSize();
1098 if (!AFI->hasStackFrame()) {
1100 emitSPUpdate(MBB, MBBI, NumBytes, isThumb, TII);
1104 // Unwind MBBI to point to first LDR / FLDD.
1105 const unsigned *CSRegs = getCalleeSavedRegs();
1106 if (MBBI != MBB.begin()) {
1109 while (MBBI != MBB.begin() && isCSRestore(MBBI, CSRegs));
1110 if (!isCSRestore(MBBI, CSRegs))
1114 // Move SP to start of FP callee save spill area.
1115 NumBytes -= (AFI->getGPRCalleeSavedArea1Size() +
1116 AFI->getGPRCalleeSavedArea2Size() +
1117 AFI->getDPRCalleeSavedAreaSize());
1119 emitSPUpdate(MBB, MBBI, NumBytes, isThumb, TII);
1121 // Darwin ABI requires FP to point to the stack slot that contains the
1123 if (STI.isTargetDarwin() || hasFP(MF)) {
1124 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1125 // Reset SP based on frame pointer only if the stack frame extends beyond
1126 // frame pointer stack slot.
1127 if (AFI->getGPRCalleeSavedArea2Size() ||
1128 AFI->getDPRCalleeSavedAreaSize() ||
1129 AFI->getDPRCalleeSavedAreaOffset())
1131 BuildMI(MBB, MBBI, TII.get(ARM::SUBri), ARM::SP).addReg(FramePtr)
1134 BuildMI(MBB, MBBI, TII.get(ARM::MOVrr), ARM::SP).addReg(FramePtr);
1135 } else if (NumBytes) {
1136 emitSPUpdate(MBB, MBBI, NumBytes, false, TII);
1139 // Move SP to start of integer callee save spill area 2.
1140 movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 3, STI);
1141 emitSPUpdate(MBB, MBBI, AFI->getDPRCalleeSavedAreaSize(), false, TII);
1143 // Move SP to start of integer callee save spill area 1.
1144 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 2, STI);
1145 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea2Size(), false, TII);
1147 // Move SP to SP upon entry to the function.
1148 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 1, STI);
1149 emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea1Size(), false, TII);
1153 emitSPUpdate(MBB, MBBI, VARegSaveSize, isThumb, TII);
1156 unsigned ARMRegisterInfo::getRARegister() const {
1160 unsigned ARMRegisterInfo::getFrameRegister(MachineFunction &MF) const {
1161 return STI.useThumbBacktraces() ? ARM::R7 : ARM::R11;
1164 #include "ARMGenRegisterInfo.inc"