1 //===-- ARMFrameLowering.cpp - ARM Frame Information ----------------------===//
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 TargetFrameLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "ARMFrameLowering.h"
15 #include "ARMBaseInstrInfo.h"
16 #include "ARMBaseRegisterInfo.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "MCTargetDesc/ARMAddressingModes.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineModuleInfo.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/RegisterScavenging.h"
26 #include "llvm/IR/CallingConv.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/MC/MCContext.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Target/TargetOptions.h"
35 SpillAlignedNEONRegs("align-neon-spills", cl::Hidden, cl::init(true),
36 cl::desc("Align ARM NEON spills in prolog and epilog"));
38 static MachineBasicBlock::iterator
39 skipAlignedDPRCS2Spills(MachineBasicBlock::iterator MI,
40 unsigned NumAlignedDPRCS2Regs);
42 /// hasFP - Return true if the specified function should have a dedicated frame
43 /// pointer register. This is true if the function has variable sized allocas
44 /// or if frame pointer elimination is disabled.
45 bool ARMFrameLowering::hasFP(const MachineFunction &MF) const {
46 const TargetRegisterInfo *RegInfo = MF.getTarget().getRegisterInfo();
48 // iOS requires FP not to be clobbered for backtracing purpose.
49 if (STI.isTargetIOS())
52 const MachineFrameInfo *MFI = MF.getFrameInfo();
53 // Always eliminate non-leaf frame pointers.
54 return ((MF.getTarget().Options.DisableFramePointerElim(MF) &&
56 RegInfo->needsStackRealignment(MF) ||
57 MFI->hasVarSizedObjects() ||
58 MFI->isFrameAddressTaken());
61 /// hasReservedCallFrame - Under normal circumstances, when a frame pointer is
62 /// not required, we reserve argument space for call sites in the function
63 /// immediately on entry to the current function. This eliminates the need for
64 /// add/sub sp brackets around call sites. Returns true if the call frame is
65 /// included as part of the stack frame.
66 bool ARMFrameLowering::hasReservedCallFrame(const MachineFunction &MF) const {
67 const MachineFrameInfo *FFI = MF.getFrameInfo();
68 unsigned CFSize = FFI->getMaxCallFrameSize();
69 // It's not always a good idea to include the call frame as part of the
70 // stack frame. ARM (especially Thumb) has small immediate offset to
71 // address the stack frame. So a large call frame can cause poor codegen
72 // and may even makes it impossible to scavenge a register.
73 if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12
76 return !MF.getFrameInfo()->hasVarSizedObjects();
79 /// canSimplifyCallFramePseudos - If there is a reserved call frame, the
80 /// call frame pseudos can be simplified. Unlike most targets, having a FP
81 /// is not sufficient here since we still may reference some objects via SP
82 /// even when FP is available in Thumb2 mode.
84 ARMFrameLowering::canSimplifyCallFramePseudos(const MachineFunction &MF) const {
85 return hasReservedCallFrame(MF) || MF.getFrameInfo()->hasVarSizedObjects();
88 static bool isCSRestore(MachineInstr *MI,
89 const ARMBaseInstrInfo &TII,
90 const uint16_t *CSRegs) {
91 // Integer spill area is handled with "pop".
92 if (isPopOpcode(MI->getOpcode())) {
93 // The first two operands are predicates. The last two are
94 // imp-def and imp-use of SP. Check everything in between.
95 for (int i = 5, e = MI->getNumOperands(); i != e; ++i)
96 if (!isCalleeSavedRegister(MI->getOperand(i).getReg(), CSRegs))
100 if ((MI->getOpcode() == ARM::LDR_POST_IMM ||
101 MI->getOpcode() == ARM::LDR_POST_REG ||
102 MI->getOpcode() == ARM::t2LDR_POST) &&
103 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs) &&
104 MI->getOperand(1).getReg() == ARM::SP)
110 static void emitRegPlusImmediate(bool isARM, MachineBasicBlock &MBB,
111 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
112 const ARMBaseInstrInfo &TII, unsigned DestReg,
113 unsigned SrcReg, int NumBytes,
114 unsigned MIFlags = MachineInstr::NoFlags,
115 ARMCC::CondCodes Pred = ARMCC::AL,
116 unsigned PredReg = 0) {
118 emitARMRegPlusImmediate(MBB, MBBI, dl, DestReg, SrcReg, NumBytes,
119 Pred, PredReg, TII, MIFlags);
121 emitT2RegPlusImmediate(MBB, MBBI, dl, DestReg, SrcReg, NumBytes,
122 Pred, PredReg, TII, MIFlags);
125 static void emitSPUpdate(bool isARM, MachineBasicBlock &MBB,
126 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
127 const ARMBaseInstrInfo &TII, int NumBytes,
128 unsigned MIFlags = MachineInstr::NoFlags,
129 ARMCC::CondCodes Pred = ARMCC::AL,
130 unsigned PredReg = 0) {
131 emitRegPlusImmediate(isARM, MBB, MBBI, dl, TII, ARM::SP, ARM::SP, NumBytes,
132 MIFlags, Pred, PredReg);
135 static int sizeOfSPAdjustment(const MachineInstr *MI) {
136 assert(MI->getOpcode() == ARM::VSTMDDB_UPD);
138 // ARM and Thumb2 push/pop insts have explicit "sp, sp" operands (+
139 // pred) so the list starts at 4.
140 for (int i = MI->getNumOperands() - 1; i >= 4; --i)
145 void ARMFrameLowering::emitPrologue(MachineFunction &MF) const {
146 MachineBasicBlock &MBB = MF.front();
147 MachineBasicBlock::iterator MBBI = MBB.begin();
148 MachineFrameInfo *MFI = MF.getFrameInfo();
149 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
150 MachineModuleInfo &MMI = MF.getMMI();
151 MCContext &Context = MMI.getContext();
152 const MCRegisterInfo *MRI = Context.getRegisterInfo();
153 const ARMBaseRegisterInfo *RegInfo =
154 static_cast<const ARMBaseRegisterInfo*>(MF.getTarget().getRegisterInfo());
155 const ARMBaseInstrInfo &TII =
156 *static_cast<const ARMBaseInstrInfo*>(MF.getTarget().getInstrInfo());
157 assert(!AFI->isThumb1OnlyFunction() &&
158 "This emitPrologue does not support Thumb1!");
159 bool isARM = !AFI->isThumbFunction();
160 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
161 unsigned ArgRegsSaveSize = AFI->getArgRegsSaveSize(Align);
162 unsigned NumBytes = MFI->getStackSize();
163 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
164 DebugLoc dl = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
165 unsigned FramePtr = RegInfo->getFrameRegister(MF);
168 // Determine the sizes of each callee-save spill areas and record which frame
169 // belongs to which callee-save spill areas.
170 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
171 int FramePtrSpillFI = 0;
174 // All calls are tail calls in GHC calling conv, and functions have no
175 // prologue/epilogue.
176 if (MF.getFunction()->getCallingConv() == CallingConv::GHC)
179 // Allocate the vararg register save area.
180 if (ArgRegsSaveSize) {
181 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -ArgRegsSaveSize,
182 MachineInstr::FrameSetup);
183 CFAOffset -= ArgRegsSaveSize;
184 unsigned CFIIndex = MMI.addFrameInst(
185 MCCFIInstruction::createDefCfaOffset(nullptr, CFAOffset));
186 BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
187 .addCFIIndex(CFIIndex);
190 if (!AFI->hasStackFrame()) {
191 if (NumBytes - ArgRegsSaveSize != 0) {
192 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -(NumBytes - ArgRegsSaveSize),
193 MachineInstr::FrameSetup);
194 CFAOffset -= NumBytes - ArgRegsSaveSize;
195 unsigned CFIIndex = MMI.addFrameInst(
196 MCCFIInstruction::createDefCfaOffset(nullptr, CFAOffset));
197 BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
198 .addCFIIndex(CFIIndex);
203 // Determine spill area sizes.
204 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
205 unsigned Reg = CSI[i].getReg();
206 int FI = CSI[i].getFrameIdx();
213 if (STI.isTargetMachO()) {
228 FramePtrSpillFI = FI;
232 // This is a DPR. Exclude the aligned DPRCS2 spills.
235 if (Reg < ARM::D8 || Reg >= ARM::D8 + AFI->getNumAlignedDPRCS2Regs())
241 MachineBasicBlock::iterator LastPush = MBB.end(), GPRCS1Push, GPRCS2Push,
244 GPRCS1Push = LastPush = MBBI++;
246 // Determine starting offsets of spill areas.
247 bool HasFP = hasFP(MF);
248 unsigned DPRCSOffset = NumBytes - (ArgRegsSaveSize + GPRCS1Size
249 + GPRCS2Size + DPRCSSize);
250 unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize;
251 unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size;
252 int FramePtrOffsetInPush = 0;
254 FramePtrOffsetInPush = MFI->getObjectOffset(FramePtrSpillFI)
255 + GPRCS1Size + ArgRegsSaveSize;
256 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) +
259 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
260 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
261 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
265 GPRCS2Push = LastPush = MBBI++;
270 // Since vpush register list cannot have gaps, there may be multiple vpush
271 // instructions in the prologue.
272 while (MBBI->getOpcode() == ARM::VSTMDDB_UPD)
276 // Move past the aligned DPRCS2 area.
277 if (AFI->getNumAlignedDPRCS2Regs() > 0) {
278 MBBI = skipAlignedDPRCS2Spills(MBBI, AFI->getNumAlignedDPRCS2Regs());
279 // The code inserted by emitAlignedDPRCS2Spills realigns the stack, and
280 // leaves the stack pointer pointing to the DPRCS2 area.
282 // Adjust NumBytes to represent the stack slots below the DPRCS2 area.
283 NumBytes += MFI->getObjectOffset(D8SpillFI);
285 NumBytes = DPRCSOffset;
287 unsigned adjustedGPRCS1Size = GPRCS1Size;
289 // Adjust SP after all the callee-save spills.
290 if (tryFoldSPUpdateIntoPushPop(STI, MF, LastPush, NumBytes)) {
291 if (LastPush == GPRCS1Push) {
292 FramePtrOffsetInPush += NumBytes;
293 adjustedGPRCS1Size += NumBytes;
297 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes,
298 MachineInstr::FrameSetup);
301 // Restore from fp only in ARM mode: e.g. sub sp, r7, #24
302 // Note it's not safe to do this in Thumb2 mode because it would have
303 // taken two instructions:
306 // If an interrupt is taken between the two instructions, then sp is in
307 // an inconsistent state (pointing to the middle of callee-saved area).
308 // The interrupt handler can end up clobbering the registers.
309 AFI->setShouldRestoreSPFromFP(true);
312 if (adjustedGPRCS1Size > 0) {
313 CFAOffset -= adjustedGPRCS1Size;
314 unsigned CFIIndex = MMI.addFrameInst(
315 MCCFIInstruction::createDefCfaOffset(nullptr, CFAOffset));
316 MachineBasicBlock::iterator Pos = ++GPRCS1Push;
317 BuildMI(MBB, Pos, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
318 .addCFIIndex(CFIIndex);
319 for (std::vector<CalleeSavedInfo>::const_iterator I = CSI.begin(),
320 E = CSI.end(); I != E; ++I) {
321 unsigned Reg = I->getReg();
322 int FI = I->getFrameIdx();
329 if (STI.isTargetMachO())
341 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
342 nullptr, MRI->getDwarfRegNum(Reg, true), MFI->getObjectOffset(FI)));
343 BuildMI(MBB, Pos, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
344 .addCFIIndex(CFIIndex);
350 // Set FP to point to the stack slot that contains the previous FP.
351 // For iOS, FP is R7, which has now been stored in spill area 1.
352 // Otherwise, if this is not iOS, all the callee-saved registers go
353 // into spill area 1, including the FP in R11. In either case, it
354 // is in area one and the adjustment needs to take place just after
357 emitRegPlusImmediate(!AFI->isThumbFunction(), MBB, GPRCS1Push, dl, TII,
358 FramePtr, ARM::SP, FramePtrOffsetInPush,
359 MachineInstr::FrameSetup);
360 if (FramePtrOffsetInPush) {
361 CFAOffset += FramePtrOffsetInPush;
362 unsigned CFIIndex = MMI.addFrameInst(MCCFIInstruction::createDefCfa(
363 nullptr, MRI->getDwarfRegNum(FramePtr, true), CFAOffset));
364 BuildMI(MBB, GPRCS1Push, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
365 .addCFIIndex(CFIIndex);
369 MMI.addFrameInst(MCCFIInstruction::createDefCfaRegister(
370 nullptr, MRI->getDwarfRegNum(FramePtr, true)));
371 BuildMI(MBB, GPRCS1Push, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
372 .addCFIIndex(CFIIndex);
376 if (GPRCS2Size > 0) {
377 MachineBasicBlock::iterator Pos = ++GPRCS2Push;
379 CFAOffset -= GPRCS2Size;
380 unsigned CFIIndex = MMI.addFrameInst(
381 MCCFIInstruction::createDefCfaOffset(nullptr, CFAOffset));
382 BuildMI(MBB, Pos, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
383 .addCFIIndex(CFIIndex);
385 for (std::vector<CalleeSavedInfo>::const_iterator I = CSI.begin(),
386 E = CSI.end(); I != E; ++I) {
387 unsigned Reg = I->getReg();
388 int FI = I->getFrameIdx();
395 if (STI.isTargetMachO()) {
396 unsigned DwarfReg = MRI->getDwarfRegNum(Reg, true);
397 unsigned Offset = MFI->getObjectOffset(FI);
398 unsigned CFIIndex = MMI.addFrameInst(
399 MCCFIInstruction::createOffset(nullptr, DwarfReg, Offset));
400 BuildMI(MBB, Pos, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
401 .addCFIIndex(CFIIndex);
409 // Since vpush register list cannot have gaps, there may be multiple vpush
410 // instructions in the prologue.
412 MachineBasicBlock::iterator Push = DPRCSPush++;
414 CFAOffset -= sizeOfSPAdjustment(Push);;
415 unsigned CFIIndex = MMI.addFrameInst(
416 MCCFIInstruction::createDefCfaOffset(nullptr, CFAOffset));
417 BuildMI(MBB, DPRCSPush, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
418 .addCFIIndex(CFIIndex);
420 } while (DPRCSPush->getOpcode() == ARM::VSTMDDB_UPD);
422 for (std::vector<CalleeSavedInfo>::const_iterator I = CSI.begin(),
423 E = CSI.end(); I != E; ++I) {
424 unsigned Reg = I->getReg();
425 int FI = I->getFrameIdx();
426 if ((Reg >= ARM::D0 && Reg <= ARM::D31) &&
427 (Reg < ARM::D8 || Reg >= ARM::D8 + AFI->getNumAlignedDPRCS2Regs())) {
428 unsigned DwarfReg = MRI->getDwarfRegNum(Reg, true);
429 unsigned Offset = MFI->getObjectOffset(FI);
430 unsigned CFIIndex = MMI.addFrameInst(
431 MCCFIInstruction::createOffset(nullptr, DwarfReg, Offset));
432 BuildMI(MBB, DPRCSPush, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
433 .addCFIIndex(CFIIndex);
440 CFAOffset -= NumBytes;
441 unsigned CFIIndex = MMI.addFrameInst(
442 MCCFIInstruction::createDefCfaOffset(nullptr, CFAOffset));
443 BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
444 .addCFIIndex(CFIIndex);
448 if (STI.isTargetELF() && hasFP(MF))
449 MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() -
450 AFI->getFramePtrSpillOffset());
452 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
453 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
454 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
456 // If we need dynamic stack realignment, do it here. Be paranoid and make
457 // sure if we also have VLAs, we have a base pointer for frame access.
458 // If aligned NEON registers were spilled, the stack has already been
460 if (!AFI->getNumAlignedDPRCS2Regs() && RegInfo->needsStackRealignment(MF)) {
461 unsigned MaxAlign = MFI->getMaxAlignment();
462 assert (!AFI->isThumb1OnlyFunction());
463 if (!AFI->isThumbFunction()) {
464 // Emit bic sp, sp, MaxAlign
465 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl,
466 TII.get(ARM::BICri), ARM::SP)
467 .addReg(ARM::SP, RegState::Kill)
468 .addImm(MaxAlign-1)));
470 // We cannot use sp as source/dest register here, thus we're emitting the
471 // following sequence:
473 // bic r4, r4, MaxAlign
475 // FIXME: It will be better just to find spare register here.
476 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), ARM::R4)
477 .addReg(ARM::SP, RegState::Kill));
478 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl,
479 TII.get(ARM::t2BICri), ARM::R4)
480 .addReg(ARM::R4, RegState::Kill)
481 .addImm(MaxAlign-1)));
482 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), ARM::SP)
483 .addReg(ARM::R4, RegState::Kill));
486 AFI->setShouldRestoreSPFromFP(true);
489 // If we need a base pointer, set it up here. It's whatever the value
490 // of the stack pointer is at this point. Any variable size objects
491 // will be allocated after this, so we can still use the base pointer
492 // to reference locals.
493 // FIXME: Clarify FrameSetup flags here.
494 if (RegInfo->hasBasePointer(MF)) {
496 BuildMI(MBB, MBBI, dl,
497 TII.get(ARM::MOVr), RegInfo->getBaseRegister())
499 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
501 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr),
502 RegInfo->getBaseRegister())
506 // If the frame has variable sized objects then the epilogue must restore
507 // the sp from fp. We can assume there's an FP here since hasFP already
508 // checks for hasVarSizedObjects.
509 if (MFI->hasVarSizedObjects())
510 AFI->setShouldRestoreSPFromFP(true);
513 void ARMFrameLowering::emitEpilogue(MachineFunction &MF,
514 MachineBasicBlock &MBB) const {
515 MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr();
516 assert(MBBI->isReturn() && "Can only insert epilog into returning blocks");
517 unsigned RetOpcode = MBBI->getOpcode();
518 DebugLoc dl = MBBI->getDebugLoc();
519 MachineFrameInfo *MFI = MF.getFrameInfo();
520 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
521 const TargetRegisterInfo *RegInfo = MF.getTarget().getRegisterInfo();
522 const ARMBaseInstrInfo &TII =
523 *static_cast<const ARMBaseInstrInfo*>(MF.getTarget().getInstrInfo());
524 assert(!AFI->isThumb1OnlyFunction() &&
525 "This emitEpilogue does not support Thumb1!");
526 bool isARM = !AFI->isThumbFunction();
528 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
529 unsigned ArgRegsSaveSize = AFI->getArgRegsSaveSize(Align);
530 int NumBytes = (int)MFI->getStackSize();
531 unsigned FramePtr = RegInfo->getFrameRegister(MF);
533 // All calls are tail calls in GHC calling conv, and functions have no
534 // prologue/epilogue.
535 if (MF.getFunction()->getCallingConv() == CallingConv::GHC)
538 if (!AFI->hasStackFrame()) {
539 if (NumBytes - ArgRegsSaveSize != 0)
540 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes - ArgRegsSaveSize);
542 // Unwind MBBI to point to first LDR / VLDRD.
543 const uint16_t *CSRegs = RegInfo->getCalleeSavedRegs(&MF);
544 if (MBBI != MBB.begin()) {
547 } while (MBBI != MBB.begin() && isCSRestore(MBBI, TII, CSRegs));
548 if (!isCSRestore(MBBI, TII, CSRegs))
552 // Move SP to start of FP callee save spill area.
553 NumBytes -= (ArgRegsSaveSize +
554 AFI->getGPRCalleeSavedArea1Size() +
555 AFI->getGPRCalleeSavedArea2Size() +
556 AFI->getDPRCalleeSavedAreaSize());
558 // Reset SP based on frame pointer only if the stack frame extends beyond
559 // frame pointer stack slot or target is ELF and the function has FP.
560 if (AFI->shouldRestoreSPFromFP()) {
561 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
564 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, FramePtr, -NumBytes,
567 // It's not possible to restore SP from FP in a single instruction.
568 // For iOS, this looks like:
571 // This is bad, if an interrupt is taken after the mov, sp is in an
572 // inconsistent state.
573 // Use the first callee-saved register as a scratch register.
574 assert(MF.getRegInfo().isPhysRegUsed(ARM::R4) &&
575 "No scratch register to restore SP from FP!");
576 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::R4, FramePtr, -NumBytes,
578 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr),
585 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), ARM::SP)
586 .addReg(FramePtr).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
588 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr),
592 } else if (NumBytes &&
593 !tryFoldSPUpdateIntoPushPop(STI, MF, MBBI, NumBytes))
594 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes);
596 // Increment past our save areas.
597 if (AFI->getDPRCalleeSavedAreaSize()) {
599 // Since vpop register list cannot have gaps, there may be multiple vpop
600 // instructions in the epilogue.
601 while (MBBI->getOpcode() == ARM::VLDMDIA_UPD)
604 if (AFI->getGPRCalleeSavedArea2Size()) MBBI++;
605 if (AFI->getGPRCalleeSavedArea1Size()) MBBI++;
608 if (RetOpcode == ARM::TCRETURNdi || RetOpcode == ARM::TCRETURNri) {
609 // Tail call return: adjust the stack pointer and jump to callee.
610 MBBI = MBB.getLastNonDebugInstr();
611 MachineOperand &JumpTarget = MBBI->getOperand(0);
613 // Jump to label or value in register.
614 if (RetOpcode == ARM::TCRETURNdi) {
615 unsigned TCOpcode = STI.isThumb() ?
616 (STI.isTargetMachO() ? ARM::tTAILJMPd : ARM::tTAILJMPdND) :
618 MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, TII.get(TCOpcode));
619 if (JumpTarget.isGlobal())
620 MIB.addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(),
621 JumpTarget.getTargetFlags());
623 assert(JumpTarget.isSymbol());
624 MIB.addExternalSymbol(JumpTarget.getSymbolName(),
625 JumpTarget.getTargetFlags());
628 // Add the default predicate in Thumb mode.
629 if (STI.isThumb()) MIB.addImm(ARMCC::AL).addReg(0);
630 } else if (RetOpcode == ARM::TCRETURNri) {
631 BuildMI(MBB, MBBI, dl,
632 TII.get(STI.isThumb() ? ARM::tTAILJMPr : ARM::TAILJMPr)).
633 addReg(JumpTarget.getReg(), RegState::Kill);
636 MachineInstr *NewMI = std::prev(MBBI);
637 for (unsigned i = 1, e = MBBI->getNumOperands(); i != e; ++i)
638 NewMI->addOperand(MBBI->getOperand(i));
640 // Delete the pseudo instruction TCRETURN.
646 emitSPUpdate(isARM, MBB, MBBI, dl, TII, ArgRegsSaveSize);
649 /// getFrameIndexReference - Provide a base+offset reference to an FI slot for
650 /// debug info. It's the same as what we use for resolving the code-gen
651 /// references for now. FIXME: This can go wrong when references are
652 /// SP-relative and simple call frames aren't used.
654 ARMFrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI,
655 unsigned &FrameReg) const {
656 return ResolveFrameIndexReference(MF, FI, FrameReg, 0);
660 ARMFrameLowering::ResolveFrameIndexReference(const MachineFunction &MF,
661 int FI, unsigned &FrameReg,
663 const MachineFrameInfo *MFI = MF.getFrameInfo();
664 const ARMBaseRegisterInfo *RegInfo =
665 static_cast<const ARMBaseRegisterInfo*>(MF.getTarget().getRegisterInfo());
666 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
667 int Offset = MFI->getObjectOffset(FI) + MFI->getStackSize();
668 int FPOffset = Offset - AFI->getFramePtrSpillOffset();
669 bool isFixed = MFI->isFixedObjectIndex(FI);
674 // SP can move around if there are allocas. We may also lose track of SP
675 // when emergency spilling inside a non-reserved call frame setup.
676 bool hasMovingSP = !hasReservedCallFrame(MF);
678 // When dynamically realigning the stack, use the frame pointer for
679 // parameters, and the stack/base pointer for locals.
680 if (RegInfo->needsStackRealignment(MF)) {
681 assert (hasFP(MF) && "dynamic stack realignment without a FP!");
683 FrameReg = RegInfo->getFrameRegister(MF);
685 } else if (hasMovingSP) {
686 assert(RegInfo->hasBasePointer(MF) &&
687 "VLAs and dynamic stack alignment, but missing base pointer!");
688 FrameReg = RegInfo->getBaseRegister();
693 // If there is a frame pointer, use it when we can.
694 if (hasFP(MF) && AFI->hasStackFrame()) {
695 // Use frame pointer to reference fixed objects. Use it for locals if
696 // there are VLAs (and thus the SP isn't reliable as a base).
697 if (isFixed || (hasMovingSP && !RegInfo->hasBasePointer(MF))) {
698 FrameReg = RegInfo->getFrameRegister(MF);
700 } else if (hasMovingSP) {
701 assert(RegInfo->hasBasePointer(MF) && "missing base pointer!");
702 if (AFI->isThumb2Function()) {
703 // Try to use the frame pointer if we can, else use the base pointer
704 // since it's available. This is handy for the emergency spill slot, in
706 if (FPOffset >= -255 && FPOffset < 0) {
707 FrameReg = RegInfo->getFrameRegister(MF);
711 } else if (AFI->isThumb2Function()) {
712 // Use add <rd>, sp, #<imm8>
713 // ldr <rd>, [sp, #<imm8>]
714 // if at all possible to save space.
715 if (Offset >= 0 && (Offset & 3) == 0 && Offset <= 1020)
717 // In Thumb2 mode, the negative offset is very limited. Try to avoid
718 // out of range references. ldr <rt>,[<rn>, #-<imm8>]
719 if (FPOffset >= -255 && FPOffset < 0) {
720 FrameReg = RegInfo->getFrameRegister(MF);
723 } else if (Offset > (FPOffset < 0 ? -FPOffset : FPOffset)) {
724 // Otherwise, use SP or FP, whichever is closer to the stack slot.
725 FrameReg = RegInfo->getFrameRegister(MF);
729 // Use the base pointer if we have one.
730 if (RegInfo->hasBasePointer(MF))
731 FrameReg = RegInfo->getBaseRegister();
735 int ARMFrameLowering::getFrameIndexOffset(const MachineFunction &MF,
738 return getFrameIndexReference(MF, FI, FrameReg);
741 void ARMFrameLowering::emitPushInst(MachineBasicBlock &MBB,
742 MachineBasicBlock::iterator MI,
743 const std::vector<CalleeSavedInfo> &CSI,
744 unsigned StmOpc, unsigned StrOpc,
746 bool(*Func)(unsigned, bool),
747 unsigned NumAlignedDPRCS2Regs,
748 unsigned MIFlags) const {
749 MachineFunction &MF = *MBB.getParent();
750 const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
753 if (MI != MBB.end()) DL = MI->getDebugLoc();
755 SmallVector<std::pair<unsigned,bool>, 4> Regs;
756 unsigned i = CSI.size();
758 unsigned LastReg = 0;
759 for (; i != 0; --i) {
760 unsigned Reg = CSI[i-1].getReg();
761 if (!(Func)(Reg, STI.isTargetMachO())) continue;
763 // D-registers in the aligned area DPRCS2 are NOT spilled here.
764 if (Reg >= ARM::D8 && Reg < ARM::D8 + NumAlignedDPRCS2Regs)
767 // Add the callee-saved register as live-in unless it's LR and
768 // @llvm.returnaddress is called. If LR is returned for
769 // @llvm.returnaddress then it's already added to the function and
770 // entry block live-in sets.
772 if (Reg == ARM::LR) {
773 if (MF.getFrameInfo()->isReturnAddressTaken() &&
774 MF.getRegInfo().isLiveIn(Reg))
781 // If NoGap is true, push consecutive registers and then leave the rest
782 // for other instructions. e.g.
783 // vpush {d8, d10, d11} -> vpush {d8}, vpush {d10, d11}
784 if (NoGap && LastReg && LastReg != Reg-1)
787 Regs.push_back(std::make_pair(Reg, isKill));
792 if (Regs.size() > 1 || StrOpc== 0) {
793 MachineInstrBuilder MIB =
794 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(StmOpc), ARM::SP)
795 .addReg(ARM::SP).setMIFlags(MIFlags));
796 for (unsigned i = 0, e = Regs.size(); i < e; ++i)
797 MIB.addReg(Regs[i].first, getKillRegState(Regs[i].second));
798 } else if (Regs.size() == 1) {
799 MachineInstrBuilder MIB = BuildMI(MBB, MI, DL, TII.get(StrOpc),
801 .addReg(Regs[0].first, getKillRegState(Regs[0].second))
802 .addReg(ARM::SP).setMIFlags(MIFlags)
808 // Put any subsequent vpush instructions before this one: they will refer to
809 // higher register numbers so need to be pushed first in order to preserve
815 void ARMFrameLowering::emitPopInst(MachineBasicBlock &MBB,
816 MachineBasicBlock::iterator MI,
817 const std::vector<CalleeSavedInfo> &CSI,
818 unsigned LdmOpc, unsigned LdrOpc,
819 bool isVarArg, bool NoGap,
820 bool(*Func)(unsigned, bool),
821 unsigned NumAlignedDPRCS2Regs) const {
822 MachineFunction &MF = *MBB.getParent();
823 const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
824 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
825 DebugLoc DL = MI->getDebugLoc();
826 unsigned RetOpcode = MI->getOpcode();
827 bool isTailCall = (RetOpcode == ARM::TCRETURNdi ||
828 RetOpcode == ARM::TCRETURNri);
830 RetOpcode == ARM::SUBS_PC_LR || RetOpcode == ARM::t2SUBS_PC_LR;
832 SmallVector<unsigned, 4> Regs;
833 unsigned i = CSI.size();
835 unsigned LastReg = 0;
836 bool DeleteRet = false;
837 for (; i != 0; --i) {
838 unsigned Reg = CSI[i-1].getReg();
839 if (!(Func)(Reg, STI.isTargetMachO())) continue;
841 // The aligned reloads from area DPRCS2 are not inserted here.
842 if (Reg >= ARM::D8 && Reg < ARM::D8 + NumAlignedDPRCS2Regs)
845 if (Reg == ARM::LR && !isTailCall && !isVarArg && !isInterrupt &&
848 LdmOpc = AFI->isThumbFunction() ? ARM::t2LDMIA_RET : ARM::LDMIA_RET;
849 // Fold the return instruction into the LDM.
853 // If NoGap is true, pop consecutive registers and then leave the rest
854 // for other instructions. e.g.
855 // vpop {d8, d10, d11} -> vpop {d8}, vpop {d10, d11}
856 if (NoGap && LastReg && LastReg != Reg-1)
865 if (Regs.size() > 1 || LdrOpc == 0) {
866 MachineInstrBuilder MIB =
867 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(LdmOpc), ARM::SP)
869 for (unsigned i = 0, e = Regs.size(); i < e; ++i)
870 MIB.addReg(Regs[i], getDefRegState(true));
872 MIB.copyImplicitOps(&*MI);
873 MI->eraseFromParent();
876 } else if (Regs.size() == 1) {
877 // If we adjusted the reg to PC from LR above, switch it back here. We
878 // only do that for LDM.
879 if (Regs[0] == ARM::PC)
881 MachineInstrBuilder MIB =
882 BuildMI(MBB, MI, DL, TII.get(LdrOpc), Regs[0])
883 .addReg(ARM::SP, RegState::Define)
885 // ARM mode needs an extra reg0 here due to addrmode2. Will go away once
886 // that refactoring is complete (eventually).
887 if (LdrOpc == ARM::LDR_POST_REG || LdrOpc == ARM::LDR_POST_IMM) {
889 MIB.addImm(ARM_AM::getAM2Opc(ARM_AM::add, 4, ARM_AM::no_shift));
896 // Put any subsequent vpop instructions after this one: they will refer to
897 // higher register numbers so need to be popped afterwards.
902 /// Emit aligned spill instructions for NumAlignedDPRCS2Regs D-registers
903 /// starting from d8. Also insert stack realignment code and leave the stack
904 /// pointer pointing to the d8 spill slot.
905 static void emitAlignedDPRCS2Spills(MachineBasicBlock &MBB,
906 MachineBasicBlock::iterator MI,
907 unsigned NumAlignedDPRCS2Regs,
908 const std::vector<CalleeSavedInfo> &CSI,
909 const TargetRegisterInfo *TRI) {
910 MachineFunction &MF = *MBB.getParent();
911 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
912 DebugLoc DL = MI->getDebugLoc();
913 const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
914 MachineFrameInfo &MFI = *MF.getFrameInfo();
916 // Mark the D-register spill slots as properly aligned. Since MFI computes
917 // stack slot layout backwards, this can actually mean that the d-reg stack
918 // slot offsets can be wrong. The offset for d8 will always be correct.
919 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
920 unsigned DNum = CSI[i].getReg() - ARM::D8;
923 int FI = CSI[i].getFrameIdx();
924 // The even-numbered registers will be 16-byte aligned, the odd-numbered
925 // registers will be 8-byte aligned.
926 MFI.setObjectAlignment(FI, DNum % 2 ? 8 : 16);
928 // The stack slot for D8 needs to be maximally aligned because this is
929 // actually the point where we align the stack pointer. MachineFrameInfo
930 // computes all offsets relative to the incoming stack pointer which is a
931 // bit weird when realigning the stack. Any extra padding for this
932 // over-alignment is not realized because the code inserted below adjusts
933 // the stack pointer by numregs * 8 before aligning the stack pointer.
935 MFI.setObjectAlignment(FI, MFI.getMaxAlignment());
938 // Move the stack pointer to the d8 spill slot, and align it at the same
939 // time. Leave the stack slot address in the scratch register r4.
941 // sub r4, sp, #numregs * 8
942 // bic r4, r4, #align - 1
945 bool isThumb = AFI->isThumbFunction();
946 assert(!AFI->isThumb1OnlyFunction() && "Can't realign stack for thumb1");
947 AFI->setShouldRestoreSPFromFP(true);
949 // sub r4, sp, #numregs * 8
950 // The immediate is <= 64, so it doesn't need any special encoding.
951 unsigned Opc = isThumb ? ARM::t2SUBri : ARM::SUBri;
952 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(Opc), ARM::R4)
954 .addImm(8 * NumAlignedDPRCS2Regs)));
956 // bic r4, r4, #align-1
957 Opc = isThumb ? ARM::t2BICri : ARM::BICri;
958 unsigned MaxAlign = MF.getFrameInfo()->getMaxAlignment();
959 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(Opc), ARM::R4)
960 .addReg(ARM::R4, RegState::Kill)
961 .addImm(MaxAlign - 1)));
964 // The stack pointer must be adjusted before spilling anything, otherwise
965 // the stack slots could be clobbered by an interrupt handler.
966 // Leave r4 live, it is used below.
967 Opc = isThumb ? ARM::tMOVr : ARM::MOVr;
968 MachineInstrBuilder MIB = BuildMI(MBB, MI, DL, TII.get(Opc), ARM::SP)
970 MIB = AddDefaultPred(MIB);
974 // Now spill NumAlignedDPRCS2Regs registers starting from d8.
975 // r4 holds the stack slot address.
976 unsigned NextReg = ARM::D8;
978 // 16-byte aligned vst1.64 with 4 d-regs and address writeback.
979 // The writeback is only needed when emitting two vst1.64 instructions.
980 if (NumAlignedDPRCS2Regs >= 6) {
981 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
983 MBB.addLiveIn(SupReg);
984 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VST1d64Qwb_fixed),
986 .addReg(ARM::R4, RegState::Kill).addImm(16)
988 .addReg(SupReg, RegState::ImplicitKill));
990 NumAlignedDPRCS2Regs -= 4;
993 // We won't modify r4 beyond this point. It currently points to the next
994 // register to be spilled.
995 unsigned R4BaseReg = NextReg;
997 // 16-byte aligned vst1.64 with 4 d-regs, no writeback.
998 if (NumAlignedDPRCS2Regs >= 4) {
999 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1000 &ARM::QQPRRegClass);
1001 MBB.addLiveIn(SupReg);
1002 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VST1d64Q))
1003 .addReg(ARM::R4).addImm(16).addReg(NextReg)
1004 .addReg(SupReg, RegState::ImplicitKill));
1006 NumAlignedDPRCS2Regs -= 4;
1009 // 16-byte aligned vst1.64 with 2 d-regs.
1010 if (NumAlignedDPRCS2Regs >= 2) {
1011 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1013 MBB.addLiveIn(SupReg);
1014 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VST1q64))
1015 .addReg(ARM::R4).addImm(16).addReg(SupReg));
1017 NumAlignedDPRCS2Regs -= 2;
1020 // Finally, use a vanilla vstr.64 for the odd last register.
1021 if (NumAlignedDPRCS2Regs) {
1022 MBB.addLiveIn(NextReg);
1023 // vstr.64 uses addrmode5 which has an offset scale of 4.
1024 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VSTRD))
1026 .addReg(ARM::R4).addImm((NextReg-R4BaseReg)*2));
1029 // The last spill instruction inserted should kill the scratch register r4.
1030 std::prev(MI)->addRegisterKilled(ARM::R4, TRI);
1033 /// Skip past the code inserted by emitAlignedDPRCS2Spills, and return an
1034 /// iterator to the following instruction.
1035 static MachineBasicBlock::iterator
1036 skipAlignedDPRCS2Spills(MachineBasicBlock::iterator MI,
1037 unsigned NumAlignedDPRCS2Regs) {
1038 // sub r4, sp, #numregs * 8
1039 // bic r4, r4, #align - 1
1042 assert(MI->mayStore() && "Expecting spill instruction");
1044 // These switches all fall through.
1045 switch(NumAlignedDPRCS2Regs) {
1048 assert(MI->mayStore() && "Expecting spill instruction");
1051 assert(MI->mayStore() && "Expecting spill instruction");
1055 assert(MI->killsRegister(ARM::R4) && "Missed kill flag");
1061 /// Emit aligned reload instructions for NumAlignedDPRCS2Regs D-registers
1062 /// starting from d8. These instructions are assumed to execute while the
1063 /// stack is still aligned, unlike the code inserted by emitPopInst.
1064 static void emitAlignedDPRCS2Restores(MachineBasicBlock &MBB,
1065 MachineBasicBlock::iterator MI,
1066 unsigned NumAlignedDPRCS2Regs,
1067 const std::vector<CalleeSavedInfo> &CSI,
1068 const TargetRegisterInfo *TRI) {
1069 MachineFunction &MF = *MBB.getParent();
1070 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1071 DebugLoc DL = MI->getDebugLoc();
1072 const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
1074 // Find the frame index assigned to d8.
1076 for (unsigned i = 0, e = CSI.size(); i != e; ++i)
1077 if (CSI[i].getReg() == ARM::D8) {
1078 D8SpillFI = CSI[i].getFrameIdx();
1082 // Materialize the address of the d8 spill slot into the scratch register r4.
1083 // This can be fairly complicated if the stack frame is large, so just use
1084 // the normal frame index elimination mechanism to do it. This code runs as
1085 // the initial part of the epilog where the stack and base pointers haven't
1086 // been changed yet.
1087 bool isThumb = AFI->isThumbFunction();
1088 assert(!AFI->isThumb1OnlyFunction() && "Can't realign stack for thumb1");
1090 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
1091 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(Opc), ARM::R4)
1092 .addFrameIndex(D8SpillFI).addImm(0)));
1094 // Now restore NumAlignedDPRCS2Regs registers starting from d8.
1095 unsigned NextReg = ARM::D8;
1097 // 16-byte aligned vld1.64 with 4 d-regs and writeback.
1098 if (NumAlignedDPRCS2Regs >= 6) {
1099 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1100 &ARM::QQPRRegClass);
1101 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLD1d64Qwb_fixed), NextReg)
1102 .addReg(ARM::R4, RegState::Define)
1103 .addReg(ARM::R4, RegState::Kill).addImm(16)
1104 .addReg(SupReg, RegState::ImplicitDefine));
1106 NumAlignedDPRCS2Regs -= 4;
1109 // We won't modify r4 beyond this point. It currently points to the next
1110 // register to be spilled.
1111 unsigned R4BaseReg = NextReg;
1113 // 16-byte aligned vld1.64 with 4 d-regs, no writeback.
1114 if (NumAlignedDPRCS2Regs >= 4) {
1115 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1116 &ARM::QQPRRegClass);
1117 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLD1d64Q), NextReg)
1118 .addReg(ARM::R4).addImm(16)
1119 .addReg(SupReg, RegState::ImplicitDefine));
1121 NumAlignedDPRCS2Regs -= 4;
1124 // 16-byte aligned vld1.64 with 2 d-regs.
1125 if (NumAlignedDPRCS2Regs >= 2) {
1126 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1128 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLD1q64), SupReg)
1129 .addReg(ARM::R4).addImm(16));
1131 NumAlignedDPRCS2Regs -= 2;
1134 // Finally, use a vanilla vldr.64 for the remaining odd register.
1135 if (NumAlignedDPRCS2Regs)
1136 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLDRD), NextReg)
1137 .addReg(ARM::R4).addImm(2*(NextReg-R4BaseReg)));
1139 // Last store kills r4.
1140 std::prev(MI)->addRegisterKilled(ARM::R4, TRI);
1143 bool ARMFrameLowering::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
1144 MachineBasicBlock::iterator MI,
1145 const std::vector<CalleeSavedInfo> &CSI,
1146 const TargetRegisterInfo *TRI) const {
1150 MachineFunction &MF = *MBB.getParent();
1151 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1153 unsigned PushOpc = AFI->isThumbFunction() ? ARM::t2STMDB_UPD : ARM::STMDB_UPD;
1154 unsigned PushOneOpc = AFI->isThumbFunction() ?
1155 ARM::t2STR_PRE : ARM::STR_PRE_IMM;
1156 unsigned FltOpc = ARM::VSTMDDB_UPD;
1157 unsigned NumAlignedDPRCS2Regs = AFI->getNumAlignedDPRCS2Regs();
1158 emitPushInst(MBB, MI, CSI, PushOpc, PushOneOpc, false, &isARMArea1Register, 0,
1159 MachineInstr::FrameSetup);
1160 emitPushInst(MBB, MI, CSI, PushOpc, PushOneOpc, false, &isARMArea2Register, 0,
1161 MachineInstr::FrameSetup);
1162 emitPushInst(MBB, MI, CSI, FltOpc, 0, true, &isARMArea3Register,
1163 NumAlignedDPRCS2Regs, MachineInstr::FrameSetup);
1165 // The code above does not insert spill code for the aligned DPRCS2 registers.
1166 // The stack realignment code will be inserted between the push instructions
1167 // and these spills.
1168 if (NumAlignedDPRCS2Regs)
1169 emitAlignedDPRCS2Spills(MBB, MI, NumAlignedDPRCS2Regs, CSI, TRI);
1174 bool ARMFrameLowering::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1175 MachineBasicBlock::iterator MI,
1176 const std::vector<CalleeSavedInfo> &CSI,
1177 const TargetRegisterInfo *TRI) const {
1181 MachineFunction &MF = *MBB.getParent();
1182 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1183 bool isVarArg = AFI->getArgRegsSaveSize() > 0;
1184 unsigned NumAlignedDPRCS2Regs = AFI->getNumAlignedDPRCS2Regs();
1186 // The emitPopInst calls below do not insert reloads for the aligned DPRCS2
1187 // registers. Do that here instead.
1188 if (NumAlignedDPRCS2Regs)
1189 emitAlignedDPRCS2Restores(MBB, MI, NumAlignedDPRCS2Regs, CSI, TRI);
1191 unsigned PopOpc = AFI->isThumbFunction() ? ARM::t2LDMIA_UPD : ARM::LDMIA_UPD;
1192 unsigned LdrOpc = AFI->isThumbFunction() ? ARM::t2LDR_POST :ARM::LDR_POST_IMM;
1193 unsigned FltOpc = ARM::VLDMDIA_UPD;
1194 emitPopInst(MBB, MI, CSI, FltOpc, 0, isVarArg, true, &isARMArea3Register,
1195 NumAlignedDPRCS2Regs);
1196 emitPopInst(MBB, MI, CSI, PopOpc, LdrOpc, isVarArg, false,
1197 &isARMArea2Register, 0);
1198 emitPopInst(MBB, MI, CSI, PopOpc, LdrOpc, isVarArg, false,
1199 &isARMArea1Register, 0);
1204 // FIXME: Make generic?
1205 static unsigned GetFunctionSizeInBytes(const MachineFunction &MF,
1206 const ARMBaseInstrInfo &TII) {
1207 unsigned FnSize = 0;
1208 for (MachineFunction::const_iterator MBBI = MF.begin(), E = MF.end();
1209 MBBI != E; ++MBBI) {
1210 const MachineBasicBlock &MBB = *MBBI;
1211 for (MachineBasicBlock::const_iterator I = MBB.begin(),E = MBB.end();
1213 FnSize += TII.GetInstSizeInBytes(I);
1218 /// estimateRSStackSizeLimit - Look at each instruction that references stack
1219 /// frames and return the stack size limit beyond which some of these
1220 /// instructions will require a scratch register during their expansion later.
1221 // FIXME: Move to TII?
1222 static unsigned estimateRSStackSizeLimit(MachineFunction &MF,
1223 const TargetFrameLowering *TFI) {
1224 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1225 unsigned Limit = (1 << 12) - 1;
1226 for (MachineFunction::iterator BB = MF.begin(),E = MF.end(); BB != E; ++BB) {
1227 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
1229 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
1230 if (!I->getOperand(i).isFI()) continue;
1232 // When using ADDri to get the address of a stack object, 255 is the
1233 // largest offset guaranteed to fit in the immediate offset.
1234 if (I->getOpcode() == ARM::ADDri) {
1235 Limit = std::min(Limit, (1U << 8) - 1);
1239 // Otherwise check the addressing mode.
1240 switch (I->getDesc().TSFlags & ARMII::AddrModeMask) {
1241 case ARMII::AddrMode3:
1242 case ARMII::AddrModeT2_i8:
1243 Limit = std::min(Limit, (1U << 8) - 1);
1245 case ARMII::AddrMode5:
1246 case ARMII::AddrModeT2_i8s4:
1247 Limit = std::min(Limit, ((1U << 8) - 1) * 4);
1249 case ARMII::AddrModeT2_i12:
1250 // i12 supports only positive offset so these will be converted to
1251 // i8 opcodes. See llvm::rewriteT2FrameIndex.
1252 if (TFI->hasFP(MF) && AFI->hasStackFrame())
1253 Limit = std::min(Limit, (1U << 8) - 1);
1255 case ARMII::AddrMode4:
1256 case ARMII::AddrMode6:
1257 // Addressing modes 4 & 6 (load/store) instructions can't encode an
1258 // immediate offset for stack references.
1263 break; // At most one FI per instruction
1271 // In functions that realign the stack, it can be an advantage to spill the
1272 // callee-saved vector registers after realigning the stack. The vst1 and vld1
1273 // instructions take alignment hints that can improve performance.
1275 static void checkNumAlignedDPRCS2Regs(MachineFunction &MF) {
1276 MF.getInfo<ARMFunctionInfo>()->setNumAlignedDPRCS2Regs(0);
1277 if (!SpillAlignedNEONRegs)
1280 // Naked functions don't spill callee-saved registers.
1281 if (MF.getFunction()->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1285 // We are planning to use NEON instructions vst1 / vld1.
1286 if (!MF.getTarget().getSubtarget<ARMSubtarget>().hasNEON())
1289 // Don't bother if the default stack alignment is sufficiently high.
1290 if (MF.getTarget().getFrameLowering()->getStackAlignment() >= 8)
1293 // Aligned spills require stack realignment.
1294 const ARMBaseRegisterInfo *RegInfo =
1295 static_cast<const ARMBaseRegisterInfo*>(MF.getTarget().getRegisterInfo());
1296 if (!RegInfo->canRealignStack(MF))
1299 // We always spill contiguous d-registers starting from d8. Count how many
1300 // needs spilling. The register allocator will almost always use the
1301 // callee-saved registers in order, but it can happen that there are holes in
1302 // the range. Registers above the hole will be spilled to the standard DPRCS
1304 MachineRegisterInfo &MRI = MF.getRegInfo();
1305 unsigned NumSpills = 0;
1306 for (; NumSpills < 8; ++NumSpills)
1307 if (!MRI.isPhysRegUsed(ARM::D8 + NumSpills))
1310 // Don't do this for just one d-register. It's not worth it.
1314 // Spill the first NumSpills D-registers after realigning the stack.
1315 MF.getInfo<ARMFunctionInfo>()->setNumAlignedDPRCS2Regs(NumSpills);
1317 // A scratch register is required for the vst1 / vld1 instructions.
1318 MF.getRegInfo().setPhysRegUsed(ARM::R4);
1322 ARMFrameLowering::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
1323 RegScavenger *RS) const {
1324 // This tells PEI to spill the FP as if it is any other callee-save register
1325 // to take advantage the eliminateFrameIndex machinery. This also ensures it
1326 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
1327 // to combine multiple loads / stores.
1328 bool CanEliminateFrame = true;
1329 bool CS1Spilled = false;
1330 bool LRSpilled = false;
1331 unsigned NumGPRSpills = 0;
1332 SmallVector<unsigned, 4> UnspilledCS1GPRs;
1333 SmallVector<unsigned, 4> UnspilledCS2GPRs;
1334 const ARMBaseRegisterInfo *RegInfo =
1335 static_cast<const ARMBaseRegisterInfo*>(MF.getTarget().getRegisterInfo());
1336 const ARMBaseInstrInfo &TII =
1337 *static_cast<const ARMBaseInstrInfo*>(MF.getTarget().getInstrInfo());
1338 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1339 MachineFrameInfo *MFI = MF.getFrameInfo();
1340 MachineRegisterInfo &MRI = MF.getRegInfo();
1341 unsigned FramePtr = RegInfo->getFrameRegister(MF);
1343 // Spill R4 if Thumb2 function requires stack realignment - it will be used as
1344 // scratch register. Also spill R4 if Thumb2 function has varsized objects,
1345 // since it's not always possible to restore sp from fp in a single
1347 // FIXME: It will be better just to find spare register here.
1348 if (AFI->isThumb2Function() &&
1349 (MFI->hasVarSizedObjects() || RegInfo->needsStackRealignment(MF)))
1350 MRI.setPhysRegUsed(ARM::R4);
1352 if (AFI->isThumb1OnlyFunction()) {
1353 // Spill LR if Thumb1 function uses variable length argument lists.
1354 if (AFI->getArgRegsSaveSize() > 0)
1355 MRI.setPhysRegUsed(ARM::LR);
1357 // Spill R4 if Thumb1 epilogue has to restore SP from FP. We don't know
1358 // for sure what the stack size will be, but for this, an estimate is good
1359 // enough. If there anything changes it, it'll be a spill, which implies
1360 // we've used all the registers and so R4 is already used, so not marking
1361 // it here will be OK.
1362 // FIXME: It will be better just to find spare register here.
1363 unsigned StackSize = MFI->estimateStackSize(MF);
1364 if (MFI->hasVarSizedObjects() || StackSize > 508)
1365 MRI.setPhysRegUsed(ARM::R4);
1368 // See if we can spill vector registers to aligned stack.
1369 checkNumAlignedDPRCS2Regs(MF);
1371 // Spill the BasePtr if it's used.
1372 if (RegInfo->hasBasePointer(MF))
1373 MRI.setPhysRegUsed(RegInfo->getBaseRegister());
1375 // Don't spill FP if the frame can be eliminated. This is determined
1376 // by scanning the callee-save registers to see if any is used.
1377 const uint16_t *CSRegs = RegInfo->getCalleeSavedRegs(&MF);
1378 for (unsigned i = 0; CSRegs[i]; ++i) {
1379 unsigned Reg = CSRegs[i];
1380 bool Spilled = false;
1381 if (MRI.isPhysRegUsed(Reg)) {
1383 CanEliminateFrame = false;
1386 if (!ARM::GPRRegClass.contains(Reg))
1392 if (!STI.isTargetMachO()) {
1399 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
1404 case ARM::R0: case ARM::R1:
1405 case ARM::R2: case ARM::R3:
1406 case ARM::R4: case ARM::R5:
1407 case ARM::R6: case ARM::R7:
1414 if (!STI.isTargetMachO()) {
1415 UnspilledCS1GPRs.push_back(Reg);
1420 case ARM::R0: case ARM::R1:
1421 case ARM::R2: case ARM::R3:
1422 case ARM::R4: case ARM::R5:
1423 case ARM::R6: case ARM::R7:
1425 UnspilledCS1GPRs.push_back(Reg);
1428 UnspilledCS2GPRs.push_back(Reg);
1434 bool ForceLRSpill = false;
1435 if (!LRSpilled && AFI->isThumb1OnlyFunction()) {
1436 unsigned FnSize = GetFunctionSizeInBytes(MF, TII);
1437 // Force LR to be spilled if the Thumb function size is > 2048. This enables
1438 // use of BL to implement far jump. If it turns out that it's not needed
1439 // then the branch fix up path will undo it.
1440 if (FnSize >= (1 << 11)) {
1441 CanEliminateFrame = false;
1442 ForceLRSpill = true;
1446 // If any of the stack slot references may be out of range of an immediate
1447 // offset, make sure a register (or a spill slot) is available for the
1448 // register scavenger. Note that if we're indexing off the frame pointer, the
1449 // effective stack size is 4 bytes larger since the FP points to the stack
1450 // slot of the previous FP. Also, if we have variable sized objects in the
1451 // function, stack slot references will often be negative, and some of
1452 // our instructions are positive-offset only, so conservatively consider
1453 // that case to want a spill slot (or register) as well. Similarly, if
1454 // the function adjusts the stack pointer during execution and the
1455 // adjustments aren't already part of our stack size estimate, our offset
1456 // calculations may be off, so be conservative.
1457 // FIXME: We could add logic to be more precise about negative offsets
1458 // and which instructions will need a scratch register for them. Is it
1459 // worth the effort and added fragility?
1462 (MFI->estimateStackSize(MF) +
1463 ((hasFP(MF) && AFI->hasStackFrame()) ? 4:0) >=
1464 estimateRSStackSizeLimit(MF, this)))
1465 || MFI->hasVarSizedObjects()
1466 || (MFI->adjustsStack() && !canSimplifyCallFramePseudos(MF));
1468 bool ExtraCSSpill = false;
1469 if (BigStack || !CanEliminateFrame || RegInfo->cannotEliminateFrame(MF)) {
1470 AFI->setHasStackFrame(true);
1472 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
1473 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
1474 if (!LRSpilled && CS1Spilled) {
1475 MRI.setPhysRegUsed(ARM::LR);
1477 SmallVectorImpl<unsigned>::iterator LRPos;
1478 LRPos = std::find(UnspilledCS1GPRs.begin(), UnspilledCS1GPRs.end(),
1480 if (LRPos != UnspilledCS1GPRs.end())
1481 UnspilledCS1GPRs.erase(LRPos);
1483 ForceLRSpill = false;
1484 ExtraCSSpill = true;
1488 MRI.setPhysRegUsed(FramePtr);
1492 // If stack and double are 8-byte aligned and we are spilling an odd number
1493 // of GPRs, spill one extra callee save GPR so we won't have to pad between
1494 // the integer and double callee save areas.
1495 unsigned TargetAlign = getStackAlignment();
1496 if (TargetAlign == 8 && (NumGPRSpills & 1)) {
1497 if (CS1Spilled && !UnspilledCS1GPRs.empty()) {
1498 for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) {
1499 unsigned Reg = UnspilledCS1GPRs[i];
1500 // Don't spill high register if the function is thumb1
1501 if (!AFI->isThumb1OnlyFunction() ||
1502 isARMLowRegister(Reg) || Reg == ARM::LR) {
1503 MRI.setPhysRegUsed(Reg);
1504 if (!MRI.isReserved(Reg))
1505 ExtraCSSpill = true;
1509 } else if (!UnspilledCS2GPRs.empty() && !AFI->isThumb1OnlyFunction()) {
1510 unsigned Reg = UnspilledCS2GPRs.front();
1511 MRI.setPhysRegUsed(Reg);
1512 if (!MRI.isReserved(Reg))
1513 ExtraCSSpill = true;
1517 // Estimate if we might need to scavenge a register at some point in order
1518 // to materialize a stack offset. If so, either spill one additional
1519 // callee-saved register or reserve a special spill slot to facilitate
1520 // register scavenging. Thumb1 needs a spill slot for stack pointer
1521 // adjustments also, even when the frame itself is small.
1522 if (BigStack && !ExtraCSSpill) {
1523 // If any non-reserved CS register isn't spilled, just spill one or two
1524 // extra. That should take care of it!
1525 unsigned NumExtras = TargetAlign / 4;
1526 SmallVector<unsigned, 2> Extras;
1527 while (NumExtras && !UnspilledCS1GPRs.empty()) {
1528 unsigned Reg = UnspilledCS1GPRs.back();
1529 UnspilledCS1GPRs.pop_back();
1530 if (!MRI.isReserved(Reg) &&
1531 (!AFI->isThumb1OnlyFunction() || isARMLowRegister(Reg) ||
1533 Extras.push_back(Reg);
1537 // For non-Thumb1 functions, also check for hi-reg CS registers
1538 if (!AFI->isThumb1OnlyFunction()) {
1539 while (NumExtras && !UnspilledCS2GPRs.empty()) {
1540 unsigned Reg = UnspilledCS2GPRs.back();
1541 UnspilledCS2GPRs.pop_back();
1542 if (!MRI.isReserved(Reg)) {
1543 Extras.push_back(Reg);
1548 if (Extras.size() && NumExtras == 0) {
1549 for (unsigned i = 0, e = Extras.size(); i != e; ++i) {
1550 MRI.setPhysRegUsed(Extras[i]);
1552 } else if (!AFI->isThumb1OnlyFunction()) {
1553 // note: Thumb1 functions spill to R12, not the stack. Reserve a slot
1554 // closest to SP or frame pointer.
1555 const TargetRegisterClass *RC = &ARM::GPRRegClass;
1556 RS->addScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
1564 MRI.setPhysRegUsed(ARM::LR);
1565 AFI->setLRIsSpilledForFarJump(true);
1570 void ARMFrameLowering::
1571 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
1572 MachineBasicBlock::iterator I) const {
1573 const ARMBaseInstrInfo &TII =
1574 *static_cast<const ARMBaseInstrInfo*>(MF.getTarget().getInstrInfo());
1575 if (!hasReservedCallFrame(MF)) {
1576 // If we have alloca, convert as follows:
1577 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
1578 // ADJCALLSTACKUP -> add, sp, sp, amount
1579 MachineInstr *Old = I;
1580 DebugLoc dl = Old->getDebugLoc();
1581 unsigned Amount = Old->getOperand(0).getImm();
1583 // We need to keep the stack aligned properly. To do this, we round the
1584 // amount of space needed for the outgoing arguments up to the next
1585 // alignment boundary.
1586 unsigned Align = getStackAlignment();
1587 Amount = (Amount+Align-1)/Align*Align;
1589 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1590 assert(!AFI->isThumb1OnlyFunction() &&
1591 "This eliminateCallFramePseudoInstr does not support Thumb1!");
1592 bool isARM = !AFI->isThumbFunction();
1594 // Replace the pseudo instruction with a new instruction...
1595 unsigned Opc = Old->getOpcode();
1596 int PIdx = Old->findFirstPredOperandIdx();
1597 ARMCC::CondCodes Pred = (PIdx == -1)
1598 ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(PIdx).getImm();
1599 if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) {
1600 // Note: PredReg is operand 2 for ADJCALLSTACKDOWN.
1601 unsigned PredReg = Old->getOperand(2).getReg();
1602 emitSPUpdate(isARM, MBB, I, dl, TII, -Amount, MachineInstr::NoFlags,
1605 // Note: PredReg is operand 3 for ADJCALLSTACKUP.
1606 unsigned PredReg = Old->getOperand(3).getReg();
1607 assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP);
1608 emitSPUpdate(isARM, MBB, I, dl, TII, Amount, MachineInstr::NoFlags,
1616 /// Get the minimum constant for ARM that is greater than or equal to the
1617 /// argument. In ARM, constants can have any value that can be produced by
1618 /// rotating an 8-bit value to the right by an even number of bits within a
1620 static uint32_t alignToARMConstant(uint32_t Value) {
1621 unsigned Shifted = 0;
1626 while (!(Value & 0xC0000000)) {
1631 bool Carry = (Value & 0x00FFFFFF);
1632 Value = ((Value & 0xFF000000) >> 24) + Carry;
1634 if (Value & 0x0000100)
1635 Value = Value & 0x000001FC;
1638 Value = Value >> (Shifted - 24);
1640 Value = Value << (24 - Shifted);
1645 // The stack limit in the TCB is set to this many bytes above the actual
1647 static const uint64_t kSplitStackAvailable = 256;
1649 // Adjust the function prologue to enable split stacks. This currently only
1650 // supports android and linux.
1652 // The ABI of the segmented stack prologue is a little arbitrarily chosen, but
1653 // must be well defined in order to allow for consistent implementations of the
1654 // __morestack helper function. The ABI is also not a normal ABI in that it
1655 // doesn't follow the normal calling conventions because this allows the
1656 // prologue of each function to be optimized further.
1658 // Currently, the ABI looks like (when calling __morestack)
1660 // * r4 holds the minimum stack size requested for this function call
1661 // * r5 holds the stack size of the arguments to the function
1662 // * the beginning of the function is 3 instructions after the call to
1665 // Implementations of __morestack should use r4 to allocate a new stack, r5 to
1666 // place the arguments on to the new stack, and the 3-instruction knowledge to
1667 // jump directly to the body of the function when working on the new stack.
1669 // An old (and possibly no longer compatible) implementation of __morestack for
1670 // ARM can be found at [1].
1672 // [1] - https://github.com/mozilla/rust/blob/86efd9/src/rt/arch/arm/morestack.S
1673 void ARMFrameLowering::adjustForSegmentedStacks(MachineFunction &MF) const {
1676 const ARMSubtarget *ST = &MF.getTarget().getSubtarget<ARMSubtarget>();
1677 bool Thumb = ST->isThumb();
1679 // Sadly, this currently doesn't support varargs, platforms other than
1680 // android/linux. Note that thumb1/thumb2 are support for android/linux.
1681 if (MF.getFunction()->isVarArg())
1682 report_fatal_error("Segmented stacks do not support vararg functions.");
1683 if (!ST->isTargetAndroid() && !ST->isTargetLinux())
1684 report_fatal_error("Segmented stacks not supported on this platfrom.");
1686 MachineBasicBlock &prologueMBB = MF.front();
1687 MachineFrameInfo *MFI = MF.getFrameInfo();
1688 MachineModuleInfo &MMI = MF.getMMI();
1689 MCContext &Context = MMI.getContext();
1690 const MCRegisterInfo *MRI = Context.getRegisterInfo();
1691 const ARMBaseInstrInfo &TII =
1692 *static_cast<const ARMBaseInstrInfo*>(MF.getTarget().getInstrInfo());
1693 ARMFunctionInfo *ARMFI = MF.getInfo<ARMFunctionInfo>();
1696 // Use R4 and R5 as scratch registers.
1697 // We save R4 and R5 before use and restore them before leaving the function.
1698 unsigned ScratchReg0 = ARM::R4;
1699 unsigned ScratchReg1 = ARM::R5;
1700 uint64_t AlignedStackSize;
1702 MachineBasicBlock *PrevStackMBB = MF.CreateMachineBasicBlock();
1703 MachineBasicBlock *PostStackMBB = MF.CreateMachineBasicBlock();
1704 MachineBasicBlock *AllocMBB = MF.CreateMachineBasicBlock();
1705 MachineBasicBlock *GetMBB = MF.CreateMachineBasicBlock();
1706 MachineBasicBlock *McrMBB = MF.CreateMachineBasicBlock();
1708 for (MachineBasicBlock::livein_iterator i = prologueMBB.livein_begin(),
1709 e = prologueMBB.livein_end();
1711 AllocMBB->addLiveIn(*i);
1712 GetMBB->addLiveIn(*i);
1713 McrMBB->addLiveIn(*i);
1714 PrevStackMBB->addLiveIn(*i);
1715 PostStackMBB->addLiveIn(*i);
1718 MF.push_front(PostStackMBB);
1719 MF.push_front(AllocMBB);
1720 MF.push_front(GetMBB);
1721 MF.push_front(McrMBB);
1722 MF.push_front(PrevStackMBB);
1724 // The required stack size that is aligned to ARM constant criterion.
1725 uint64_t StackSize = MFI->getStackSize();
1727 AlignedStackSize = alignToARMConstant(StackSize);
1729 // When the frame size is less than 256 we just compare the stack
1730 // boundary directly to the value of the stack pointer, per gcc.
1731 bool CompareStackPointer = AlignedStackSize < kSplitStackAvailable;
1733 // We will use two of the callee save registers as scratch registers so we
1734 // need to save those registers onto the stack.
1735 // We will use SR0 to hold stack limit and SR1 to hold the stack size
1736 // requested and arguments for __morestack().
1737 // SR0: Scratch Register #0
1738 // SR1: Scratch Register #1
1741 AddDefaultPred(BuildMI(PrevStackMBB, DL, TII.get(ARM::tPUSH)))
1742 .addReg(ScratchReg0).addReg(ScratchReg1);
1744 AddDefaultPred(BuildMI(PrevStackMBB, DL, TII.get(ARM::STMDB_UPD))
1745 .addReg(ARM::SP, RegState::Define).addReg(ARM::SP))
1746 .addReg(ScratchReg0).addReg(ScratchReg1);
1749 // Emit the relevant DWARF information about the change in stack pointer as
1750 // well as where to find both r4 and r5 (the callee-save registers)
1752 MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, -8));
1753 BuildMI(PrevStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1754 .addCFIIndex(CFIIndex);
1755 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
1756 nullptr, MRI->getDwarfRegNum(ScratchReg1, true), -4));
1757 BuildMI(PrevStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1758 .addCFIIndex(CFIIndex);
1759 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
1760 nullptr, MRI->getDwarfRegNum(ScratchReg0, true), -8));
1761 BuildMI(PrevStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1762 .addCFIIndex(CFIIndex);
1766 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::tMOVr), ScratchReg1)
1768 } else if (CompareStackPointer) {
1769 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::MOVr), ScratchReg1)
1770 .addReg(ARM::SP)).addReg(0);
1773 // sub SR1, sp, #StackSize
1774 if (!CompareStackPointer && Thumb) {
1776 AddDefaultCC(BuildMI(McrMBB, DL, TII.get(ARM::tSUBi8), ScratchReg1))
1777 .addReg(ScratchReg1).addImm(AlignedStackSize));
1778 } else if (!CompareStackPointer) {
1779 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::SUBri), ScratchReg1)
1780 .addReg(ARM::SP).addImm(AlignedStackSize)).addReg(0);
1783 if (Thumb && ST->isThumb1Only()) {
1784 unsigned PCLabelId = ARMFI->createPICLabelUId();
1785 ARMConstantPoolValue *NewCPV = ARMConstantPoolSymbol::Create(
1786 MF.getFunction()->getContext(), "__STACK_LIMIT", PCLabelId, 0);
1787 MachineConstantPool *MCP = MF.getConstantPool();
1788 unsigned CPI = MCP->getConstantPoolIndex(NewCPV, MF.getAlignment());
1790 // ldr SR0, [pc, offset(STACK_LIMIT)]
1791 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(ARM::tLDRpci), ScratchReg0)
1792 .addConstantPoolIndex(CPI));
1795 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(ARM::tLDRi), ScratchReg0)
1796 .addReg(ScratchReg0).addImm(0));
1798 // Get TLS base address from the coprocessor
1799 // mrc p15, #0, SR0, c13, c0, #3
1800 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::MRC), ScratchReg0)
1807 // Use the last tls slot on android and a private field of the TCP on linux.
1808 assert(ST->isTargetAndroid() || ST->isTargetLinux());
1809 unsigned TlsOffset = ST->isTargetAndroid() ? 63 : 1;
1811 // Get the stack limit from the right offset
1812 // ldr SR0, [sr0, #4 * TlsOffset]
1813 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(ARM::LDRi12), ScratchReg0)
1814 .addReg(ScratchReg0).addImm(4 * TlsOffset));
1817 // Compare stack limit with stack size requested.
1819 Opcode = Thumb ? ARM::tCMPr : ARM::CMPrr;
1820 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(Opcode))
1821 .addReg(ScratchReg0)
1822 .addReg(ScratchReg1));
1824 // This jump is taken if StackLimit < SP - stack required.
1825 Opcode = Thumb ? ARM::tBcc : ARM::Bcc;
1826 BuildMI(GetMBB, DL, TII.get(Opcode)).addMBB(PostStackMBB)
1831 // Calling __morestack(StackSize, Size of stack arguments).
1832 // __morestack knows that the stack size requested is in SR0(r4)
1833 // and amount size of stack arguments is in SR1(r5).
1835 // Pass first argument for the __morestack by Scratch Register #0.
1836 // The amount size of stack required
1838 AddDefaultPred(AddDefaultCC(BuildMI(AllocMBB, DL, TII.get(ARM::tMOVi8),
1839 ScratchReg0)).addImm(AlignedStackSize));
1841 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::MOVi), ScratchReg0)
1842 .addImm(AlignedStackSize)).addReg(0);
1844 // Pass second argument for the __morestack by Scratch Register #1.
1845 // The amount size of stack consumed to save function arguments.
1848 AddDefaultCC(BuildMI(AllocMBB, DL, TII.get(ARM::tMOVi8), ScratchReg1))
1849 .addImm(alignToARMConstant(ARMFI->getArgumentStackSize())));
1851 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::MOVi), ScratchReg1)
1852 .addImm(alignToARMConstant(ARMFI->getArgumentStackSize())))
1856 // push {lr} - Save return address of this function.
1858 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tPUSH)))
1861 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::STMDB_UPD))
1862 .addReg(ARM::SP, RegState::Define)
1867 // Emit the DWARF info about the change in stack as well as where to find the
1868 // previous link register
1870 MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, -12));
1871 BuildMI(AllocMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1872 .addCFIIndex(CFIIndex);
1873 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
1874 nullptr, MRI->getDwarfRegNum(ARM::LR, true), -12));
1875 BuildMI(AllocMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1876 .addCFIIndex(CFIIndex);
1878 // Call __morestack().
1880 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tBL)))
1881 .addExternalSymbol("__morestack");
1883 BuildMI(AllocMBB, DL, TII.get(ARM::BL))
1884 .addExternalSymbol("__morestack");
1887 // pop {lr} - Restore return address of this original function.
1889 if (ST->isThumb1Only()) {
1890 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tPOP)))
1891 .addReg(ScratchReg0);
1892 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tMOVr), ARM::LR)
1893 .addReg(ScratchReg0));
1895 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::t2LDR_POST))
1896 .addReg(ARM::LR, RegState::Define)
1897 .addReg(ARM::SP, RegState::Define)
1902 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::LDMIA_UPD))
1903 .addReg(ARM::SP, RegState::Define)
1908 // Restore SR0 and SR1 in case of __morestack() was called.
1909 // __morestack() will skip PostStackMBB block so we need to restore
1910 // scratch registers from here.
1913 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tPOP)))
1914 .addReg(ScratchReg0)
1915 .addReg(ScratchReg1);
1917 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::LDMIA_UPD))
1918 .addReg(ARM::SP, RegState::Define)
1920 .addReg(ScratchReg0)
1921 .addReg(ScratchReg1);
1924 // Update the CFA offset now that we've popped
1925 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, 0));
1926 BuildMI(AllocMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1927 .addCFIIndex(CFIIndex);
1929 // bx lr - Return from this function.
1930 Opcode = Thumb ? ARM::tBX_RET : ARM::BX_RET;
1931 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(Opcode)));
1933 // Restore SR0 and SR1 in case of __morestack() was not called.
1936 AddDefaultPred(BuildMI(PostStackMBB, DL, TII.get(ARM::tPOP)))
1937 .addReg(ScratchReg0)
1938 .addReg(ScratchReg1);
1940 AddDefaultPred(BuildMI(PostStackMBB, DL, TII.get(ARM::LDMIA_UPD))
1941 .addReg(ARM::SP, RegState::Define)
1943 .addReg(ScratchReg0)
1944 .addReg(ScratchReg1);
1947 // Update the CFA offset now that we've popped
1948 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, 0));
1949 BuildMI(PostStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1950 .addCFIIndex(CFIIndex);
1952 // Tell debuggers that r4 and r5 are now the same as they were in the
1953 // previous function, that they're the "Same Value".
1954 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createSameValue(
1955 nullptr, MRI->getDwarfRegNum(ScratchReg0, true)));
1956 BuildMI(PostStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1957 .addCFIIndex(CFIIndex);
1958 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createSameValue(
1959 nullptr, MRI->getDwarfRegNum(ScratchReg1, true)));
1960 BuildMI(PostStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1961 .addCFIIndex(CFIIndex);
1963 // Organizing MBB lists
1964 PostStackMBB->addSuccessor(&prologueMBB);
1966 AllocMBB->addSuccessor(PostStackMBB);
1968 GetMBB->addSuccessor(PostStackMBB);
1969 GetMBB->addSuccessor(AllocMBB);
1971 McrMBB->addSuccessor(GetMBB);
1973 PrevStackMBB->addSuccessor(McrMBB);