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 ARMFrameLowering::ARMFrameLowering(const ARMSubtarget &sti)
43 : TargetFrameLowering(StackGrowsDown, sti.getStackAlignment(), 0, 4),
46 /// hasFP - Return true if the specified function should have a dedicated frame
47 /// pointer register. This is true if the function has variable sized allocas
48 /// or if frame pointer elimination is disabled.
49 bool ARMFrameLowering::hasFP(const MachineFunction &MF) const {
50 const TargetRegisterInfo *RegInfo = MF.getSubtarget().getRegisterInfo();
52 // iOS requires FP not to be clobbered for backtracing purpose.
53 if (STI.isTargetIOS())
56 const MachineFrameInfo *MFI = MF.getFrameInfo();
57 // Always eliminate non-leaf frame pointers.
58 return ((MF.getTarget().Options.DisableFramePointerElim(MF) &&
60 RegInfo->needsStackRealignment(MF) ||
61 MFI->hasVarSizedObjects() ||
62 MFI->isFrameAddressTaken());
65 /// hasReservedCallFrame - Under normal circumstances, when a frame pointer is
66 /// not required, we reserve argument space for call sites in the function
67 /// immediately on entry to the current function. This eliminates the need for
68 /// add/sub sp brackets around call sites. Returns true if the call frame is
69 /// included as part of the stack frame.
70 bool ARMFrameLowering::hasReservedCallFrame(const MachineFunction &MF) const {
71 const MachineFrameInfo *FFI = MF.getFrameInfo();
72 unsigned CFSize = FFI->getMaxCallFrameSize();
73 // It's not always a good idea to include the call frame as part of the
74 // stack frame. ARM (especially Thumb) has small immediate offset to
75 // address the stack frame. So a large call frame can cause poor codegen
76 // and may even makes it impossible to scavenge a register.
77 if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12
80 return !MF.getFrameInfo()->hasVarSizedObjects();
83 /// canSimplifyCallFramePseudos - If there is a reserved call frame, the
84 /// call frame pseudos can be simplified. Unlike most targets, having a FP
85 /// is not sufficient here since we still may reference some objects via SP
86 /// even when FP is available in Thumb2 mode.
88 ARMFrameLowering::canSimplifyCallFramePseudos(const MachineFunction &MF) const {
89 return hasReservedCallFrame(MF) || MF.getFrameInfo()->hasVarSizedObjects();
92 static bool isCSRestore(MachineInstr *MI,
93 const ARMBaseInstrInfo &TII,
94 const MCPhysReg *CSRegs) {
95 // Integer spill area is handled with "pop".
96 if (isPopOpcode(MI->getOpcode())) {
97 // The first two operands are predicates. The last two are
98 // imp-def and imp-use of SP. Check everything in between.
99 for (int i = 5, e = MI->getNumOperands(); i != e; ++i)
100 if (!isCalleeSavedRegister(MI->getOperand(i).getReg(), CSRegs))
104 if ((MI->getOpcode() == ARM::LDR_POST_IMM ||
105 MI->getOpcode() == ARM::LDR_POST_REG ||
106 MI->getOpcode() == ARM::t2LDR_POST) &&
107 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs) &&
108 MI->getOperand(1).getReg() == ARM::SP)
114 static void emitRegPlusImmediate(bool isARM, MachineBasicBlock &MBB,
115 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
116 const ARMBaseInstrInfo &TII, unsigned DestReg,
117 unsigned SrcReg, int NumBytes,
118 unsigned MIFlags = MachineInstr::NoFlags,
119 ARMCC::CondCodes Pred = ARMCC::AL,
120 unsigned PredReg = 0) {
122 emitARMRegPlusImmediate(MBB, MBBI, dl, DestReg, SrcReg, NumBytes,
123 Pred, PredReg, TII, MIFlags);
125 emitT2RegPlusImmediate(MBB, MBBI, dl, DestReg, SrcReg, NumBytes,
126 Pred, PredReg, TII, MIFlags);
129 static void emitSPUpdate(bool isARM, MachineBasicBlock &MBB,
130 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
131 const ARMBaseInstrInfo &TII, int NumBytes,
132 unsigned MIFlags = MachineInstr::NoFlags,
133 ARMCC::CondCodes Pred = ARMCC::AL,
134 unsigned PredReg = 0) {
135 emitRegPlusImmediate(isARM, MBB, MBBI, dl, TII, ARM::SP, ARM::SP, NumBytes,
136 MIFlags, Pred, PredReg);
139 static int sizeOfSPAdjustment(const MachineInstr *MI) {
141 switch (MI->getOpcode()) {
142 case ARM::VSTMDDB_UPD:
146 case ARM::t2STMDB_UPD:
150 case ARM::STR_PRE_IMM:
153 llvm_unreachable("Unknown push or pop like instruction");
157 // ARM and Thumb2 push/pop insts have explicit "sp, sp" operands (+
158 // pred) so the list starts at 4.
159 for (int i = MI->getNumOperands() - 1; i >= 4; --i)
164 static bool WindowsRequiresStackProbe(const MachineFunction &MF,
165 size_t StackSizeInBytes) {
166 const MachineFrameInfo *MFI = MF.getFrameInfo();
167 if (MFI->getStackProtectorIndex() > 0)
168 return StackSizeInBytes >= 4080;
169 return StackSizeInBytes >= 4096;
173 struct StackAdjustingInsts {
175 MachineBasicBlock::iterator I;
180 SmallVector<InstInfo, 4> Insts;
182 void addInst(MachineBasicBlock::iterator I, unsigned SPAdjust,
183 bool BeforeFPSet = false) {
184 InstInfo Info = {I, SPAdjust, BeforeFPSet};
185 Insts.push_back(Info);
188 void addExtraBytes(const MachineBasicBlock::iterator I, unsigned ExtraBytes) {
189 auto Info = std::find_if(Insts.begin(), Insts.end(),
190 [&](InstInfo &Info) { return Info.I == I; });
191 assert(Info != Insts.end() && "invalid sp adjusting instruction");
192 Info->SPAdjust += ExtraBytes;
195 void emitDefCFAOffsets(MachineModuleInfo &MMI, MachineBasicBlock &MBB,
196 DebugLoc dl, const ARMBaseInstrInfo &TII, bool HasFP) {
197 unsigned CFAOffset = 0;
198 for (auto &Info : Insts) {
199 if (HasFP && !Info.BeforeFPSet)
202 CFAOffset -= Info.SPAdjust;
203 unsigned CFIIndex = MMI.addFrameInst(
204 MCCFIInstruction::createDefCfaOffset(nullptr, CFAOffset));
205 BuildMI(MBB, std::next(Info.I), dl,
206 TII.get(TargetOpcode::CFI_INSTRUCTION))
207 .addCFIIndex(CFIIndex)
208 .setMIFlags(MachineInstr::FrameSetup);
214 /// Emit an instruction sequence that will align the address in
215 /// register Reg by zero-ing out the lower bits. For versions of the
216 /// architecture that support Neon, this must be done in a single
217 /// instruction, since skipAlignedDPRCS2Spills assumes it is done in a
218 /// single instruction. That function only gets called when optimizing
219 /// spilling of D registers on a core with the Neon instruction set
221 static void emitAligningInstructions(MachineFunction &MF, ARMFunctionInfo *AFI,
222 const TargetInstrInfo &TII,
223 MachineBasicBlock &MBB,
224 MachineBasicBlock::iterator MBBI,
225 DebugLoc DL, const unsigned Reg,
226 const unsigned Alignment,
227 const bool MustBeSingleInstruction) {
228 const ARMSubtarget &AST =
229 static_cast<const ARMSubtarget &>(MF.getSubtarget());
230 const bool CanUseBFC = AST.hasV6T2Ops() || AST.hasV7Ops();
231 const unsigned AlignMask = Alignment - 1;
232 const unsigned NrBitsToZero = countTrailingZeros(Alignment);
233 assert(!AFI->isThumb1OnlyFunction() && "Thumb1 not supported");
234 if (!AFI->isThumbFunction()) {
235 // if the BFC instruction is available, use that to zero the lower
237 // bfc Reg, #0, log2(Alignment)
238 // otherwise use BIC, if the mask to zero the required number of bits
239 // can be encoded in the bic immediate field
240 // bic Reg, Reg, Alignment-1
242 // lsr Reg, Reg, log2(Alignment)
243 // lsl Reg, Reg, log2(Alignment)
245 AddDefaultPred(BuildMI(MBB, MBBI, DL, TII.get(ARM::BFC), Reg)
246 .addReg(Reg, RegState::Kill)
247 .addImm(~AlignMask));
248 } else if (AlignMask <= 255) {
250 AddDefaultPred(BuildMI(MBB, MBBI, DL, TII.get(ARM::BICri), Reg)
251 .addReg(Reg, RegState::Kill)
252 .addImm(AlignMask)));
254 assert(!MustBeSingleInstruction &&
255 "Shouldn't call emitAligningInstructions demanding a single "
256 "instruction to be emitted for large stack alignment for a target "
258 AddDefaultCC(AddDefaultPred(
259 BuildMI(MBB, MBBI, DL, TII.get(ARM::MOVsi), Reg)
260 .addReg(Reg, RegState::Kill)
261 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsr, NrBitsToZero))));
262 AddDefaultCC(AddDefaultPred(
263 BuildMI(MBB, MBBI, DL, TII.get(ARM::MOVsi), Reg)
264 .addReg(Reg, RegState::Kill)
265 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, NrBitsToZero))));
268 // Since this is only reached for Thumb-2 targets, the BFC instruction
269 // should always be available.
271 AddDefaultPred(BuildMI(MBB, MBBI, DL, TII.get(ARM::t2BFC), Reg)
272 .addReg(Reg, RegState::Kill)
273 .addImm(~AlignMask));
277 void ARMFrameLowering::emitPrologue(MachineFunction &MF) const {
278 MachineBasicBlock &MBB = MF.front();
279 MachineBasicBlock::iterator MBBI = MBB.begin();
280 MachineFrameInfo *MFI = MF.getFrameInfo();
281 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
282 MachineModuleInfo &MMI = MF.getMMI();
283 MCContext &Context = MMI.getContext();
284 const TargetMachine &TM = MF.getTarget();
285 const MCRegisterInfo *MRI = Context.getRegisterInfo();
286 const ARMBaseRegisterInfo *RegInfo = STI.getRegisterInfo();
287 const ARMBaseInstrInfo &TII = *STI.getInstrInfo();
288 assert(!AFI->isThumb1OnlyFunction() &&
289 "This emitPrologue does not support Thumb1!");
290 bool isARM = !AFI->isThumbFunction();
291 unsigned Align = STI.getFrameLowering()->getStackAlignment();
292 unsigned ArgRegsSaveSize = AFI->getArgRegsSaveSize(Align);
293 unsigned NumBytes = MFI->getStackSize();
294 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
295 DebugLoc dl = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
296 unsigned FramePtr = RegInfo->getFrameRegister(MF);
298 // Determine the sizes of each callee-save spill areas and record which frame
299 // belongs to which callee-save spill areas.
300 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
301 int FramePtrSpillFI = 0;
304 // All calls are tail calls in GHC calling conv, and functions have no
305 // prologue/epilogue.
306 if (MF.getFunction()->getCallingConv() == CallingConv::GHC)
309 StackAdjustingInsts DefCFAOffsetCandidates;
311 // Allocate the vararg register save area.
312 if (ArgRegsSaveSize) {
313 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -ArgRegsSaveSize,
314 MachineInstr::FrameSetup);
315 DefCFAOffsetCandidates.addInst(std::prev(MBBI), ArgRegsSaveSize, true);
318 if (!AFI->hasStackFrame() &&
319 (!STI.isTargetWindows() || !WindowsRequiresStackProbe(MF, NumBytes))) {
320 if (NumBytes - ArgRegsSaveSize != 0) {
321 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -(NumBytes - ArgRegsSaveSize),
322 MachineInstr::FrameSetup);
323 DefCFAOffsetCandidates.addInst(std::prev(MBBI),
324 NumBytes - ArgRegsSaveSize, true);
329 // Determine spill area sizes.
330 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
331 unsigned Reg = CSI[i].getReg();
332 int FI = CSI[i].getFrameIdx();
339 if (STI.isTargetDarwin()) {
354 FramePtrSpillFI = FI;
358 // This is a DPR. Exclude the aligned DPRCS2 spills.
361 if (Reg < ARM::D8 || Reg >= ARM::D8 + AFI->getNumAlignedDPRCS2Regs())
367 MachineBasicBlock::iterator LastPush = MBB.end(), GPRCS1Push, GPRCS2Push;
368 if (GPRCS1Size > 0) {
369 GPRCS1Push = LastPush = MBBI++;
370 DefCFAOffsetCandidates.addInst(LastPush, GPRCS1Size, true);
373 // Determine starting offsets of spill areas.
374 bool HasFP = hasFP(MF);
375 unsigned GPRCS1Offset = NumBytes - ArgRegsSaveSize - GPRCS1Size;
376 unsigned GPRCS2Offset = GPRCS1Offset - GPRCS2Size;
377 unsigned DPRAlign = DPRCSSize ? std::min(8U, Align) : 4U;
378 unsigned DPRGapSize = (GPRCS1Size + GPRCS2Size + ArgRegsSaveSize) % DPRAlign;
379 unsigned DPRCSOffset = GPRCS2Offset - DPRGapSize - DPRCSSize;
380 int FramePtrOffsetInPush = 0;
382 FramePtrOffsetInPush =
383 MFI->getObjectOffset(FramePtrSpillFI) + ArgRegsSaveSize;
384 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) +
387 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
388 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
389 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
392 if (GPRCS2Size > 0) {
393 GPRCS2Push = LastPush = MBBI++;
394 DefCFAOffsetCandidates.addInst(LastPush, GPRCS2Size);
397 // Prolog/epilog inserter assumes we correctly align DPRs on the stack, so our
398 // .cfi_offset operations will reflect that.
400 assert(DPRGapSize == 4 && "unexpected alignment requirements for DPRs");
401 if (tryFoldSPUpdateIntoPushPop(STI, MF, LastPush, DPRGapSize))
402 DefCFAOffsetCandidates.addExtraBytes(LastPush, DPRGapSize);
404 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -DPRGapSize,
405 MachineInstr::FrameSetup);
406 DefCFAOffsetCandidates.addInst(std::prev(MBBI), DPRGapSize);
412 // Since vpush register list cannot have gaps, there may be multiple vpush
413 // instructions in the prologue.
414 while (MBBI->getOpcode() == ARM::VSTMDDB_UPD) {
415 DefCFAOffsetCandidates.addInst(MBBI, sizeOfSPAdjustment(MBBI));
420 // Move past the aligned DPRCS2 area.
421 if (AFI->getNumAlignedDPRCS2Regs() > 0) {
422 MBBI = skipAlignedDPRCS2Spills(MBBI, AFI->getNumAlignedDPRCS2Regs());
423 // The code inserted by emitAlignedDPRCS2Spills realigns the stack, and
424 // leaves the stack pointer pointing to the DPRCS2 area.
426 // Adjust NumBytes to represent the stack slots below the DPRCS2 area.
427 NumBytes += MFI->getObjectOffset(D8SpillFI);
429 NumBytes = DPRCSOffset;
431 if (STI.isTargetWindows() && WindowsRequiresStackProbe(MF, NumBytes)) {
432 uint32_t NumWords = NumBytes >> 2;
434 if (NumWords < 65536)
435 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVi16), ARM::R4)
437 .setMIFlags(MachineInstr::FrameSetup));
439 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVi32imm), ARM::R4)
441 .setMIFlags(MachineInstr::FrameSetup);
443 switch (TM.getCodeModel()) {
444 case CodeModel::Small:
445 case CodeModel::Medium:
446 case CodeModel::Default:
447 case CodeModel::Kernel:
448 BuildMI(MBB, MBBI, dl, TII.get(ARM::tBL))
449 .addImm((unsigned)ARMCC::AL).addReg(0)
450 .addExternalSymbol("__chkstk")
451 .addReg(ARM::R4, RegState::Implicit)
452 .setMIFlags(MachineInstr::FrameSetup);
454 case CodeModel::Large:
455 case CodeModel::JITDefault:
456 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVi32imm), ARM::R12)
457 .addExternalSymbol("__chkstk")
458 .setMIFlags(MachineInstr::FrameSetup);
460 BuildMI(MBB, MBBI, dl, TII.get(ARM::tBLXr))
461 .addImm((unsigned)ARMCC::AL).addReg(0)
462 .addReg(ARM::R12, RegState::Kill)
463 .addReg(ARM::R4, RegState::Implicit)
464 .setMIFlags(MachineInstr::FrameSetup);
468 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::t2SUBrr),
470 .addReg(ARM::SP, RegState::Define)
471 .addReg(ARM::R4, RegState::Kill)
472 .setMIFlags(MachineInstr::FrameSetup)));
477 // Adjust SP after all the callee-save spills.
478 if (tryFoldSPUpdateIntoPushPop(STI, MF, LastPush, NumBytes))
479 DefCFAOffsetCandidates.addExtraBytes(LastPush, NumBytes);
481 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes,
482 MachineInstr::FrameSetup);
483 DefCFAOffsetCandidates.addInst(std::prev(MBBI), NumBytes);
487 // Restore from fp only in ARM mode: e.g. sub sp, r7, #24
488 // Note it's not safe to do this in Thumb2 mode because it would have
489 // taken two instructions:
492 // If an interrupt is taken between the two instructions, then sp is in
493 // an inconsistent state (pointing to the middle of callee-saved area).
494 // The interrupt handler can end up clobbering the registers.
495 AFI->setShouldRestoreSPFromFP(true);
498 // Set FP to point to the stack slot that contains the previous FP.
499 // For iOS, FP is R7, which has now been stored in spill area 1.
500 // Otherwise, if this is not iOS, all the callee-saved registers go
501 // into spill area 1, including the FP in R11. In either case, it
502 // is in area one and the adjustment needs to take place just after
505 MachineBasicBlock::iterator AfterPush = std::next(GPRCS1Push);
506 unsigned PushSize = sizeOfSPAdjustment(GPRCS1Push);
507 emitRegPlusImmediate(!AFI->isThumbFunction(), MBB, AfterPush,
508 dl, TII, FramePtr, ARM::SP,
509 PushSize + FramePtrOffsetInPush,
510 MachineInstr::FrameSetup);
511 if (FramePtrOffsetInPush + PushSize != 0) {
512 unsigned CFIIndex = MMI.addFrameInst(MCCFIInstruction::createDefCfa(
513 nullptr, MRI->getDwarfRegNum(FramePtr, true),
514 -(ArgRegsSaveSize - FramePtrOffsetInPush)));
515 BuildMI(MBB, AfterPush, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
516 .addCFIIndex(CFIIndex)
517 .setMIFlags(MachineInstr::FrameSetup);
520 MMI.addFrameInst(MCCFIInstruction::createDefCfaRegister(
521 nullptr, MRI->getDwarfRegNum(FramePtr, true)));
522 BuildMI(MBB, AfterPush, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
523 .addCFIIndex(CFIIndex)
524 .setMIFlags(MachineInstr::FrameSetup);
528 // Now that the prologue's actual instructions are finalised, we can insert
529 // the necessary DWARF cf instructions to describe the situation. Start by
530 // recording where each register ended up:
531 if (GPRCS1Size > 0) {
532 MachineBasicBlock::iterator Pos = std::next(GPRCS1Push);
534 for (const auto &Entry : CSI) {
535 unsigned Reg = Entry.getReg();
536 int FI = Entry.getFrameIdx();
543 if (STI.isTargetDarwin())
555 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
556 nullptr, MRI->getDwarfRegNum(Reg, true), MFI->getObjectOffset(FI)));
557 BuildMI(MBB, Pos, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
558 .addCFIIndex(CFIIndex)
559 .setMIFlags(MachineInstr::FrameSetup);
565 if (GPRCS2Size > 0) {
566 MachineBasicBlock::iterator Pos = std::next(GPRCS2Push);
567 for (const auto &Entry : CSI) {
568 unsigned Reg = Entry.getReg();
569 int FI = Entry.getFrameIdx();
576 if (STI.isTargetDarwin()) {
577 unsigned DwarfReg = MRI->getDwarfRegNum(Reg, true);
578 unsigned Offset = MFI->getObjectOffset(FI);
579 unsigned CFIIndex = MMI.addFrameInst(
580 MCCFIInstruction::createOffset(nullptr, DwarfReg, Offset));
581 BuildMI(MBB, Pos, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
582 .addCFIIndex(CFIIndex)
583 .setMIFlags(MachineInstr::FrameSetup);
591 // Since vpush register list cannot have gaps, there may be multiple vpush
592 // instructions in the prologue.
593 MachineBasicBlock::iterator Pos = std::next(LastPush);
594 for (const auto &Entry : CSI) {
595 unsigned Reg = Entry.getReg();
596 int FI = Entry.getFrameIdx();
597 if ((Reg >= ARM::D0 && Reg <= ARM::D31) &&
598 (Reg < ARM::D8 || Reg >= ARM::D8 + AFI->getNumAlignedDPRCS2Regs())) {
599 unsigned DwarfReg = MRI->getDwarfRegNum(Reg, true);
600 unsigned Offset = MFI->getObjectOffset(FI);
601 unsigned CFIIndex = MMI.addFrameInst(
602 MCCFIInstruction::createOffset(nullptr, DwarfReg, Offset));
603 BuildMI(MBB, Pos, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
604 .addCFIIndex(CFIIndex)
605 .setMIFlags(MachineInstr::FrameSetup);
610 // Now we can emit descriptions of where the canonical frame address was
611 // throughout the process. If we have a frame pointer, it takes over the job
612 // half-way through, so only the first few .cfi_def_cfa_offset instructions
613 // actually get emitted.
614 DefCFAOffsetCandidates.emitDefCFAOffsets(MMI, MBB, dl, TII, HasFP);
616 if (STI.isTargetELF() && hasFP(MF))
617 MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() -
618 AFI->getFramePtrSpillOffset());
620 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
621 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
622 AFI->setDPRCalleeSavedGapSize(DPRGapSize);
623 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
625 // If we need dynamic stack realignment, do it here. Be paranoid and make
626 // sure if we also have VLAs, we have a base pointer for frame access.
627 // If aligned NEON registers were spilled, the stack has already been
629 if (!AFI->getNumAlignedDPRCS2Regs() && RegInfo->needsStackRealignment(MF)) {
630 unsigned MaxAlign = MFI->getMaxAlignment();
631 assert(!AFI->isThumb1OnlyFunction());
632 if (!AFI->isThumbFunction()) {
633 emitAligningInstructions(MF, AFI, TII, MBB, MBBI, dl, ARM::SP, MaxAlign,
636 // We cannot use sp as source/dest register here, thus we're using r4 to
637 // perform the calculations. We're emitting the following sequence:
639 // -- use emitAligningInstructions to produce best sequence to zero
640 // -- out lower bits in r4
642 // FIXME: It will be better just to find spare register here.
643 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), ARM::R4)
644 .addReg(ARM::SP, RegState::Kill));
645 emitAligningInstructions(MF, AFI, TII, MBB, MBBI, dl, ARM::R4, MaxAlign,
647 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), ARM::SP)
648 .addReg(ARM::R4, RegState::Kill));
651 AFI->setShouldRestoreSPFromFP(true);
654 // If we need a base pointer, set it up here. It's whatever the value
655 // of the stack pointer is at this point. Any variable size objects
656 // will be allocated after this, so we can still use the base pointer
657 // to reference locals.
658 // FIXME: Clarify FrameSetup flags here.
659 if (RegInfo->hasBasePointer(MF)) {
661 BuildMI(MBB, MBBI, dl,
662 TII.get(ARM::MOVr), RegInfo->getBaseRegister())
664 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
666 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr),
667 RegInfo->getBaseRegister())
671 // If the frame has variable sized objects then the epilogue must restore
672 // the sp from fp. We can assume there's an FP here since hasFP already
673 // checks for hasVarSizedObjects.
674 if (MFI->hasVarSizedObjects())
675 AFI->setShouldRestoreSPFromFP(true);
678 // Resolve TCReturn pseudo-instruction
679 void ARMFrameLowering::fixTCReturn(MachineFunction &MF,
680 MachineBasicBlock &MBB) const {
681 MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr();
682 assert(MBBI->isReturn() && "Can only insert epilog into returning blocks");
683 unsigned RetOpcode = MBBI->getOpcode();
684 DebugLoc dl = MBBI->getDebugLoc();
685 const ARMBaseInstrInfo &TII =
686 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
688 if (!(RetOpcode == ARM::TCRETURNdi || RetOpcode == ARM::TCRETURNri))
691 // Tail call return: adjust the stack pointer and jump to callee.
692 MBBI = MBB.getLastNonDebugInstr();
693 MachineOperand &JumpTarget = MBBI->getOperand(0);
695 // Jump to label or value in register.
696 if (RetOpcode == ARM::TCRETURNdi) {
697 unsigned TCOpcode = STI.isThumb() ?
698 (STI.isTargetMachO() ? ARM::tTAILJMPd : ARM::tTAILJMPdND) :
700 MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, TII.get(TCOpcode));
701 if (JumpTarget.isGlobal())
702 MIB.addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(),
703 JumpTarget.getTargetFlags());
705 assert(JumpTarget.isSymbol());
706 MIB.addExternalSymbol(JumpTarget.getSymbolName(),
707 JumpTarget.getTargetFlags());
710 // Add the default predicate in Thumb mode.
711 if (STI.isThumb()) MIB.addImm(ARMCC::AL).addReg(0);
712 } else if (RetOpcode == ARM::TCRETURNri) {
713 BuildMI(MBB, MBBI, dl,
714 TII.get(STI.isThumb() ? ARM::tTAILJMPr : ARM::TAILJMPr)).
715 addReg(JumpTarget.getReg(), RegState::Kill);
718 MachineInstr *NewMI = std::prev(MBBI);
719 for (unsigned i = 1, e = MBBI->getNumOperands(); i != e; ++i)
720 NewMI->addOperand(MBBI->getOperand(i));
722 // Delete the pseudo instruction TCRETURN.
727 void ARMFrameLowering::emitEpilogue(MachineFunction &MF,
728 MachineBasicBlock &MBB) const {
729 MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr();
730 assert(MBBI->isReturn() && "Can only insert epilog into returning blocks");
731 DebugLoc dl = MBBI->getDebugLoc();
732 MachineFrameInfo *MFI = MF.getFrameInfo();
733 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
734 const TargetRegisterInfo *RegInfo = MF.getSubtarget().getRegisterInfo();
735 const ARMBaseInstrInfo &TII =
736 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
737 assert(!AFI->isThumb1OnlyFunction() &&
738 "This emitEpilogue does not support Thumb1!");
739 bool isARM = !AFI->isThumbFunction();
741 unsigned Align = STI.getFrameLowering()->getStackAlignment();
742 unsigned ArgRegsSaveSize = AFI->getArgRegsSaveSize(Align);
743 int NumBytes = (int)MFI->getStackSize();
744 unsigned FramePtr = RegInfo->getFrameRegister(MF);
746 // All calls are tail calls in GHC calling conv, and functions have no
747 // prologue/epilogue.
748 if (MF.getFunction()->getCallingConv() == CallingConv::GHC) {
749 fixTCReturn(MF, MBB);
753 if (!AFI->hasStackFrame()) {
754 if (NumBytes - ArgRegsSaveSize != 0)
755 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes - ArgRegsSaveSize);
757 // Unwind MBBI to point to first LDR / VLDRD.
758 const MCPhysReg *CSRegs = RegInfo->getCalleeSavedRegs(&MF);
759 if (MBBI != MBB.begin()) {
762 } while (MBBI != MBB.begin() && isCSRestore(MBBI, TII, CSRegs));
763 if (!isCSRestore(MBBI, TII, CSRegs))
767 // Move SP to start of FP callee save spill area.
768 NumBytes -= (ArgRegsSaveSize +
769 AFI->getGPRCalleeSavedArea1Size() +
770 AFI->getGPRCalleeSavedArea2Size() +
771 AFI->getDPRCalleeSavedGapSize() +
772 AFI->getDPRCalleeSavedAreaSize());
774 // Reset SP based on frame pointer only if the stack frame extends beyond
775 // frame pointer stack slot or target is ELF and the function has FP.
776 if (AFI->shouldRestoreSPFromFP()) {
777 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
780 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, FramePtr, -NumBytes,
783 // It's not possible to restore SP from FP in a single instruction.
784 // For iOS, this looks like:
787 // This is bad, if an interrupt is taken after the mov, sp is in an
788 // inconsistent state.
789 // Use the first callee-saved register as a scratch register.
790 assert(MF.getRegInfo().isPhysRegUsed(ARM::R4) &&
791 "No scratch register to restore SP from FP!");
792 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::R4, FramePtr, -NumBytes,
794 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr),
801 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), ARM::SP)
802 .addReg(FramePtr).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
804 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr),
808 } else if (NumBytes &&
809 !tryFoldSPUpdateIntoPushPop(STI, MF, MBBI, NumBytes))
810 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes);
812 // Increment past our save areas.
813 if (AFI->getDPRCalleeSavedAreaSize()) {
815 // Since vpop register list cannot have gaps, there may be multiple vpop
816 // instructions in the epilogue.
817 while (MBBI->getOpcode() == ARM::VLDMDIA_UPD)
820 if (AFI->getDPRCalleeSavedGapSize()) {
821 assert(AFI->getDPRCalleeSavedGapSize() == 4 &&
822 "unexpected DPR alignment gap");
823 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getDPRCalleeSavedGapSize());
826 if (AFI->getGPRCalleeSavedArea2Size()) MBBI++;
827 if (AFI->getGPRCalleeSavedArea1Size()) MBBI++;
830 fixTCReturn(MF, MBB);
833 emitSPUpdate(isARM, MBB, MBBI, dl, TII, ArgRegsSaveSize);
836 /// getFrameIndexReference - Provide a base+offset reference to an FI slot for
837 /// debug info. It's the same as what we use for resolving the code-gen
838 /// references for now. FIXME: This can go wrong when references are
839 /// SP-relative and simple call frames aren't used.
841 ARMFrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI,
842 unsigned &FrameReg) const {
843 return ResolveFrameIndexReference(MF, FI, FrameReg, 0);
847 ARMFrameLowering::ResolveFrameIndexReference(const MachineFunction &MF,
848 int FI, unsigned &FrameReg,
850 const MachineFrameInfo *MFI = MF.getFrameInfo();
851 const ARMBaseRegisterInfo *RegInfo = static_cast<const ARMBaseRegisterInfo *>(
852 MF.getSubtarget().getRegisterInfo());
853 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
854 int Offset = MFI->getObjectOffset(FI) + MFI->getStackSize();
855 int FPOffset = Offset - AFI->getFramePtrSpillOffset();
856 bool isFixed = MFI->isFixedObjectIndex(FI);
861 // SP can move around if there are allocas. We may also lose track of SP
862 // when emergency spilling inside a non-reserved call frame setup.
863 bool hasMovingSP = !hasReservedCallFrame(MF);
865 // When dynamically realigning the stack, use the frame pointer for
866 // parameters, and the stack/base pointer for locals.
867 if (RegInfo->needsStackRealignment(MF)) {
868 assert (hasFP(MF) && "dynamic stack realignment without a FP!");
870 FrameReg = RegInfo->getFrameRegister(MF);
872 } else if (hasMovingSP) {
873 assert(RegInfo->hasBasePointer(MF) &&
874 "VLAs and dynamic stack alignment, but missing base pointer!");
875 FrameReg = RegInfo->getBaseRegister();
880 // If there is a frame pointer, use it when we can.
881 if (hasFP(MF) && AFI->hasStackFrame()) {
882 // Use frame pointer to reference fixed objects. Use it for locals if
883 // there are VLAs (and thus the SP isn't reliable as a base).
884 if (isFixed || (hasMovingSP && !RegInfo->hasBasePointer(MF))) {
885 FrameReg = RegInfo->getFrameRegister(MF);
887 } else if (hasMovingSP) {
888 assert(RegInfo->hasBasePointer(MF) && "missing base pointer!");
889 if (AFI->isThumb2Function()) {
890 // Try to use the frame pointer if we can, else use the base pointer
891 // since it's available. This is handy for the emergency spill slot, in
893 if (FPOffset >= -255 && FPOffset < 0) {
894 FrameReg = RegInfo->getFrameRegister(MF);
898 } else if (AFI->isThumb2Function()) {
899 // Use add <rd>, sp, #<imm8>
900 // ldr <rd>, [sp, #<imm8>]
901 // if at all possible to save space.
902 if (Offset >= 0 && (Offset & 3) == 0 && Offset <= 1020)
904 // In Thumb2 mode, the negative offset is very limited. Try to avoid
905 // out of range references. ldr <rt>,[<rn>, #-<imm8>]
906 if (FPOffset >= -255 && FPOffset < 0) {
907 FrameReg = RegInfo->getFrameRegister(MF);
910 } else if (Offset > (FPOffset < 0 ? -FPOffset : FPOffset)) {
911 // Otherwise, use SP or FP, whichever is closer to the stack slot.
912 FrameReg = RegInfo->getFrameRegister(MF);
916 // Use the base pointer if we have one.
917 if (RegInfo->hasBasePointer(MF))
918 FrameReg = RegInfo->getBaseRegister();
922 int ARMFrameLowering::getFrameIndexOffset(const MachineFunction &MF,
925 return getFrameIndexReference(MF, FI, FrameReg);
928 void ARMFrameLowering::emitPushInst(MachineBasicBlock &MBB,
929 MachineBasicBlock::iterator MI,
930 const std::vector<CalleeSavedInfo> &CSI,
931 unsigned StmOpc, unsigned StrOpc,
933 bool(*Func)(unsigned, bool),
934 unsigned NumAlignedDPRCS2Regs,
935 unsigned MIFlags) const {
936 MachineFunction &MF = *MBB.getParent();
937 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
940 if (MI != MBB.end()) DL = MI->getDebugLoc();
942 SmallVector<std::pair<unsigned,bool>, 4> Regs;
943 unsigned i = CSI.size();
945 unsigned LastReg = 0;
946 for (; i != 0; --i) {
947 unsigned Reg = CSI[i-1].getReg();
948 if (!(Func)(Reg, STI.isTargetDarwin())) continue;
950 // D-registers in the aligned area DPRCS2 are NOT spilled here.
951 if (Reg >= ARM::D8 && Reg < ARM::D8 + NumAlignedDPRCS2Regs)
954 // Add the callee-saved register as live-in unless it's LR and
955 // @llvm.returnaddress is called. If LR is returned for
956 // @llvm.returnaddress then it's already added to the function and
957 // entry block live-in sets.
959 if (Reg == ARM::LR) {
960 if (MF.getFrameInfo()->isReturnAddressTaken() &&
961 MF.getRegInfo().isLiveIn(Reg))
968 // If NoGap is true, push consecutive registers and then leave the rest
969 // for other instructions. e.g.
970 // vpush {d8, d10, d11} -> vpush {d8}, vpush {d10, d11}
971 if (NoGap && LastReg && LastReg != Reg-1)
974 Regs.push_back(std::make_pair(Reg, isKill));
979 if (Regs.size() > 1 || StrOpc== 0) {
980 MachineInstrBuilder MIB =
981 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(StmOpc), ARM::SP)
982 .addReg(ARM::SP).setMIFlags(MIFlags));
983 for (unsigned i = 0, e = Regs.size(); i < e; ++i)
984 MIB.addReg(Regs[i].first, getKillRegState(Regs[i].second));
985 } else if (Regs.size() == 1) {
986 MachineInstrBuilder MIB = BuildMI(MBB, MI, DL, TII.get(StrOpc),
988 .addReg(Regs[0].first, getKillRegState(Regs[0].second))
989 .addReg(ARM::SP).setMIFlags(MIFlags)
995 // Put any subsequent vpush instructions before this one: they will refer to
996 // higher register numbers so need to be pushed first in order to preserve
1002 void ARMFrameLowering::emitPopInst(MachineBasicBlock &MBB,
1003 MachineBasicBlock::iterator MI,
1004 const std::vector<CalleeSavedInfo> &CSI,
1005 unsigned LdmOpc, unsigned LdrOpc,
1006 bool isVarArg, bool NoGap,
1007 bool(*Func)(unsigned, bool),
1008 unsigned NumAlignedDPRCS2Regs) const {
1009 MachineFunction &MF = *MBB.getParent();
1010 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
1011 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1012 DebugLoc DL = MI->getDebugLoc();
1013 unsigned RetOpcode = MI->getOpcode();
1014 bool isTailCall = (RetOpcode == ARM::TCRETURNdi ||
1015 RetOpcode == ARM::TCRETURNri);
1017 RetOpcode == ARM::SUBS_PC_LR || RetOpcode == ARM::t2SUBS_PC_LR;
1019 SmallVector<unsigned, 4> Regs;
1020 unsigned i = CSI.size();
1022 unsigned LastReg = 0;
1023 bool DeleteRet = false;
1024 for (; i != 0; --i) {
1025 unsigned Reg = CSI[i-1].getReg();
1026 if (!(Func)(Reg, STI.isTargetDarwin())) continue;
1028 // The aligned reloads from area DPRCS2 are not inserted here.
1029 if (Reg >= ARM::D8 && Reg < ARM::D8 + NumAlignedDPRCS2Regs)
1032 if (Reg == ARM::LR && !isTailCall && !isVarArg && !isInterrupt &&
1035 LdmOpc = AFI->isThumbFunction() ? ARM::t2LDMIA_RET : ARM::LDMIA_RET;
1036 // Fold the return instruction into the LDM.
1040 // If NoGap is true, pop consecutive registers and then leave the rest
1041 // for other instructions. e.g.
1042 // vpop {d8, d10, d11} -> vpop {d8}, vpop {d10, d11}
1043 if (NoGap && LastReg && LastReg != Reg-1)
1047 Regs.push_back(Reg);
1052 if (Regs.size() > 1 || LdrOpc == 0) {
1053 MachineInstrBuilder MIB =
1054 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(LdmOpc), ARM::SP)
1056 for (unsigned i = 0, e = Regs.size(); i < e; ++i)
1057 MIB.addReg(Regs[i], getDefRegState(true));
1059 MIB.copyImplicitOps(&*MI);
1060 MI->eraseFromParent();
1063 } else if (Regs.size() == 1) {
1064 // If we adjusted the reg to PC from LR above, switch it back here. We
1065 // only do that for LDM.
1066 if (Regs[0] == ARM::PC)
1068 MachineInstrBuilder MIB =
1069 BuildMI(MBB, MI, DL, TII.get(LdrOpc), Regs[0])
1070 .addReg(ARM::SP, RegState::Define)
1072 // ARM mode needs an extra reg0 here due to addrmode2. Will go away once
1073 // that refactoring is complete (eventually).
1074 if (LdrOpc == ARM::LDR_POST_REG || LdrOpc == ARM::LDR_POST_IMM) {
1076 MIB.addImm(ARM_AM::getAM2Opc(ARM_AM::add, 4, ARM_AM::no_shift));
1079 AddDefaultPred(MIB);
1083 // Put any subsequent vpop instructions after this one: they will refer to
1084 // higher register numbers so need to be popped afterwards.
1089 /// Emit aligned spill instructions for NumAlignedDPRCS2Regs D-registers
1090 /// starting from d8. Also insert stack realignment code and leave the stack
1091 /// pointer pointing to the d8 spill slot.
1092 static void emitAlignedDPRCS2Spills(MachineBasicBlock &MBB,
1093 MachineBasicBlock::iterator MI,
1094 unsigned NumAlignedDPRCS2Regs,
1095 const std::vector<CalleeSavedInfo> &CSI,
1096 const TargetRegisterInfo *TRI) {
1097 MachineFunction &MF = *MBB.getParent();
1098 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1099 DebugLoc DL = MI->getDebugLoc();
1100 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
1101 MachineFrameInfo &MFI = *MF.getFrameInfo();
1103 // Mark the D-register spill slots as properly aligned. Since MFI computes
1104 // stack slot layout backwards, this can actually mean that the d-reg stack
1105 // slot offsets can be wrong. The offset for d8 will always be correct.
1106 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1107 unsigned DNum = CSI[i].getReg() - ARM::D8;
1110 int FI = CSI[i].getFrameIdx();
1111 // The even-numbered registers will be 16-byte aligned, the odd-numbered
1112 // registers will be 8-byte aligned.
1113 MFI.setObjectAlignment(FI, DNum % 2 ? 8 : 16);
1115 // The stack slot for D8 needs to be maximally aligned because this is
1116 // actually the point where we align the stack pointer. MachineFrameInfo
1117 // computes all offsets relative to the incoming stack pointer which is a
1118 // bit weird when realigning the stack. Any extra padding for this
1119 // over-alignment is not realized because the code inserted below adjusts
1120 // the stack pointer by numregs * 8 before aligning the stack pointer.
1122 MFI.setObjectAlignment(FI, MFI.getMaxAlignment());
1125 // Move the stack pointer to the d8 spill slot, and align it at the same
1126 // time. Leave the stack slot address in the scratch register r4.
1128 // sub r4, sp, #numregs * 8
1129 // bic r4, r4, #align - 1
1132 bool isThumb = AFI->isThumbFunction();
1133 assert(!AFI->isThumb1OnlyFunction() && "Can't realign stack for thumb1");
1134 AFI->setShouldRestoreSPFromFP(true);
1136 // sub r4, sp, #numregs * 8
1137 // The immediate is <= 64, so it doesn't need any special encoding.
1138 unsigned Opc = isThumb ? ARM::t2SUBri : ARM::SUBri;
1139 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(Opc), ARM::R4)
1141 .addImm(8 * NumAlignedDPRCS2Regs)));
1143 unsigned MaxAlign = MF.getFrameInfo()->getMaxAlignment();
1144 // We must set parameter MustBeSingleInstruction to true, since
1145 // skipAlignedDPRCS2Spills expects exactly 3 instructions to perform
1146 // stack alignment. Luckily, this can always be done since all ARM
1147 // architecture versions that support Neon also support the BFC
1149 emitAligningInstructions(MF, AFI, TII, MBB, MI, DL, ARM::R4, MaxAlign, true);
1152 // The stack pointer must be adjusted before spilling anything, otherwise
1153 // the stack slots could be clobbered by an interrupt handler.
1154 // Leave r4 live, it is used below.
1155 Opc = isThumb ? ARM::tMOVr : ARM::MOVr;
1156 MachineInstrBuilder MIB = BuildMI(MBB, MI, DL, TII.get(Opc), ARM::SP)
1158 MIB = AddDefaultPred(MIB);
1162 // Now spill NumAlignedDPRCS2Regs registers starting from d8.
1163 // r4 holds the stack slot address.
1164 unsigned NextReg = ARM::D8;
1166 // 16-byte aligned vst1.64 with 4 d-regs and address writeback.
1167 // The writeback is only needed when emitting two vst1.64 instructions.
1168 if (NumAlignedDPRCS2Regs >= 6) {
1169 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1170 &ARM::QQPRRegClass);
1171 MBB.addLiveIn(SupReg);
1172 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VST1d64Qwb_fixed),
1174 .addReg(ARM::R4, RegState::Kill).addImm(16)
1176 .addReg(SupReg, RegState::ImplicitKill));
1178 NumAlignedDPRCS2Regs -= 4;
1181 // We won't modify r4 beyond this point. It currently points to the next
1182 // register to be spilled.
1183 unsigned R4BaseReg = NextReg;
1185 // 16-byte aligned vst1.64 with 4 d-regs, no writeback.
1186 if (NumAlignedDPRCS2Regs >= 4) {
1187 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1188 &ARM::QQPRRegClass);
1189 MBB.addLiveIn(SupReg);
1190 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VST1d64Q))
1191 .addReg(ARM::R4).addImm(16).addReg(NextReg)
1192 .addReg(SupReg, RegState::ImplicitKill));
1194 NumAlignedDPRCS2Regs -= 4;
1197 // 16-byte aligned vst1.64 with 2 d-regs.
1198 if (NumAlignedDPRCS2Regs >= 2) {
1199 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1201 MBB.addLiveIn(SupReg);
1202 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VST1q64))
1203 .addReg(ARM::R4).addImm(16).addReg(SupReg));
1205 NumAlignedDPRCS2Regs -= 2;
1208 // Finally, use a vanilla vstr.64 for the odd last register.
1209 if (NumAlignedDPRCS2Regs) {
1210 MBB.addLiveIn(NextReg);
1211 // vstr.64 uses addrmode5 which has an offset scale of 4.
1212 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VSTRD))
1214 .addReg(ARM::R4).addImm((NextReg-R4BaseReg)*2));
1217 // The last spill instruction inserted should kill the scratch register r4.
1218 std::prev(MI)->addRegisterKilled(ARM::R4, TRI);
1221 /// Skip past the code inserted by emitAlignedDPRCS2Spills, and return an
1222 /// iterator to the following instruction.
1223 static MachineBasicBlock::iterator
1224 skipAlignedDPRCS2Spills(MachineBasicBlock::iterator MI,
1225 unsigned NumAlignedDPRCS2Regs) {
1226 // sub r4, sp, #numregs * 8
1227 // bic r4, r4, #align - 1
1230 assert(MI->mayStore() && "Expecting spill instruction");
1232 // These switches all fall through.
1233 switch(NumAlignedDPRCS2Regs) {
1236 assert(MI->mayStore() && "Expecting spill instruction");
1239 assert(MI->mayStore() && "Expecting spill instruction");
1243 assert(MI->killsRegister(ARM::R4) && "Missed kill flag");
1249 /// Emit aligned reload instructions for NumAlignedDPRCS2Regs D-registers
1250 /// starting from d8. These instructions are assumed to execute while the
1251 /// stack is still aligned, unlike the code inserted by emitPopInst.
1252 static void emitAlignedDPRCS2Restores(MachineBasicBlock &MBB,
1253 MachineBasicBlock::iterator MI,
1254 unsigned NumAlignedDPRCS2Regs,
1255 const std::vector<CalleeSavedInfo> &CSI,
1256 const TargetRegisterInfo *TRI) {
1257 MachineFunction &MF = *MBB.getParent();
1258 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1259 DebugLoc DL = MI->getDebugLoc();
1260 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
1262 // Find the frame index assigned to d8.
1264 for (unsigned i = 0, e = CSI.size(); i != e; ++i)
1265 if (CSI[i].getReg() == ARM::D8) {
1266 D8SpillFI = CSI[i].getFrameIdx();
1270 // Materialize the address of the d8 spill slot into the scratch register r4.
1271 // This can be fairly complicated if the stack frame is large, so just use
1272 // the normal frame index elimination mechanism to do it. This code runs as
1273 // the initial part of the epilog where the stack and base pointers haven't
1274 // been changed yet.
1275 bool isThumb = AFI->isThumbFunction();
1276 assert(!AFI->isThumb1OnlyFunction() && "Can't realign stack for thumb1");
1278 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
1279 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(Opc), ARM::R4)
1280 .addFrameIndex(D8SpillFI).addImm(0)));
1282 // Now restore NumAlignedDPRCS2Regs registers starting from d8.
1283 unsigned NextReg = ARM::D8;
1285 // 16-byte aligned vld1.64 with 4 d-regs and writeback.
1286 if (NumAlignedDPRCS2Regs >= 6) {
1287 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1288 &ARM::QQPRRegClass);
1289 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLD1d64Qwb_fixed), NextReg)
1290 .addReg(ARM::R4, RegState::Define)
1291 .addReg(ARM::R4, RegState::Kill).addImm(16)
1292 .addReg(SupReg, RegState::ImplicitDefine));
1294 NumAlignedDPRCS2Regs -= 4;
1297 // We won't modify r4 beyond this point. It currently points to the next
1298 // register to be spilled.
1299 unsigned R4BaseReg = NextReg;
1301 // 16-byte aligned vld1.64 with 4 d-regs, no writeback.
1302 if (NumAlignedDPRCS2Regs >= 4) {
1303 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1304 &ARM::QQPRRegClass);
1305 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLD1d64Q), NextReg)
1306 .addReg(ARM::R4).addImm(16)
1307 .addReg(SupReg, RegState::ImplicitDefine));
1309 NumAlignedDPRCS2Regs -= 4;
1312 // 16-byte aligned vld1.64 with 2 d-regs.
1313 if (NumAlignedDPRCS2Regs >= 2) {
1314 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1316 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLD1q64), SupReg)
1317 .addReg(ARM::R4).addImm(16));
1319 NumAlignedDPRCS2Regs -= 2;
1322 // Finally, use a vanilla vldr.64 for the remaining odd register.
1323 if (NumAlignedDPRCS2Regs)
1324 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLDRD), NextReg)
1325 .addReg(ARM::R4).addImm(2*(NextReg-R4BaseReg)));
1327 // Last store kills r4.
1328 std::prev(MI)->addRegisterKilled(ARM::R4, TRI);
1331 bool ARMFrameLowering::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
1332 MachineBasicBlock::iterator MI,
1333 const std::vector<CalleeSavedInfo> &CSI,
1334 const TargetRegisterInfo *TRI) const {
1338 MachineFunction &MF = *MBB.getParent();
1339 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1341 unsigned PushOpc = AFI->isThumbFunction() ? ARM::t2STMDB_UPD : ARM::STMDB_UPD;
1342 unsigned PushOneOpc = AFI->isThumbFunction() ?
1343 ARM::t2STR_PRE : ARM::STR_PRE_IMM;
1344 unsigned FltOpc = ARM::VSTMDDB_UPD;
1345 unsigned NumAlignedDPRCS2Regs = AFI->getNumAlignedDPRCS2Regs();
1346 emitPushInst(MBB, MI, CSI, PushOpc, PushOneOpc, false, &isARMArea1Register, 0,
1347 MachineInstr::FrameSetup);
1348 emitPushInst(MBB, MI, CSI, PushOpc, PushOneOpc, false, &isARMArea2Register, 0,
1349 MachineInstr::FrameSetup);
1350 emitPushInst(MBB, MI, CSI, FltOpc, 0, true, &isARMArea3Register,
1351 NumAlignedDPRCS2Regs, MachineInstr::FrameSetup);
1353 // The code above does not insert spill code for the aligned DPRCS2 registers.
1354 // The stack realignment code will be inserted between the push instructions
1355 // and these spills.
1356 if (NumAlignedDPRCS2Regs)
1357 emitAlignedDPRCS2Spills(MBB, MI, NumAlignedDPRCS2Regs, CSI, TRI);
1362 bool ARMFrameLowering::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1363 MachineBasicBlock::iterator MI,
1364 const std::vector<CalleeSavedInfo> &CSI,
1365 const TargetRegisterInfo *TRI) const {
1369 MachineFunction &MF = *MBB.getParent();
1370 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1371 bool isVarArg = AFI->getArgRegsSaveSize() > 0;
1372 unsigned NumAlignedDPRCS2Regs = AFI->getNumAlignedDPRCS2Regs();
1374 // The emitPopInst calls below do not insert reloads for the aligned DPRCS2
1375 // registers. Do that here instead.
1376 if (NumAlignedDPRCS2Regs)
1377 emitAlignedDPRCS2Restores(MBB, MI, NumAlignedDPRCS2Regs, CSI, TRI);
1379 unsigned PopOpc = AFI->isThumbFunction() ? ARM::t2LDMIA_UPD : ARM::LDMIA_UPD;
1380 unsigned LdrOpc = AFI->isThumbFunction() ? ARM::t2LDR_POST :ARM::LDR_POST_IMM;
1381 unsigned FltOpc = ARM::VLDMDIA_UPD;
1382 emitPopInst(MBB, MI, CSI, FltOpc, 0, isVarArg, true, &isARMArea3Register,
1383 NumAlignedDPRCS2Regs);
1384 emitPopInst(MBB, MI, CSI, PopOpc, LdrOpc, isVarArg, false,
1385 &isARMArea2Register, 0);
1386 emitPopInst(MBB, MI, CSI, PopOpc, LdrOpc, isVarArg, false,
1387 &isARMArea1Register, 0);
1392 // FIXME: Make generic?
1393 static unsigned GetFunctionSizeInBytes(const MachineFunction &MF,
1394 const ARMBaseInstrInfo &TII) {
1395 unsigned FnSize = 0;
1396 for (auto &MBB : MF) {
1397 for (auto &MI : MBB)
1398 FnSize += TII.GetInstSizeInBytes(&MI);
1403 /// estimateRSStackSizeLimit - Look at each instruction that references stack
1404 /// frames and return the stack size limit beyond which some of these
1405 /// instructions will require a scratch register during their expansion later.
1406 // FIXME: Move to TII?
1407 static unsigned estimateRSStackSizeLimit(MachineFunction &MF,
1408 const TargetFrameLowering *TFI) {
1409 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1410 unsigned Limit = (1 << 12) - 1;
1411 for (auto &MBB : MF) {
1412 for (auto &MI : MBB) {
1413 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1414 if (!MI.getOperand(i).isFI())
1417 // When using ADDri to get the address of a stack object, 255 is the
1418 // largest offset guaranteed to fit in the immediate offset.
1419 if (MI.getOpcode() == ARM::ADDri) {
1420 Limit = std::min(Limit, (1U << 8) - 1);
1424 // Otherwise check the addressing mode.
1425 switch (MI.getDesc().TSFlags & ARMII::AddrModeMask) {
1426 case ARMII::AddrMode3:
1427 case ARMII::AddrModeT2_i8:
1428 Limit = std::min(Limit, (1U << 8) - 1);
1430 case ARMII::AddrMode5:
1431 case ARMII::AddrModeT2_i8s4:
1432 Limit = std::min(Limit, ((1U << 8) - 1) * 4);
1434 case ARMII::AddrModeT2_i12:
1435 // i12 supports only positive offset so these will be converted to
1436 // i8 opcodes. See llvm::rewriteT2FrameIndex.
1437 if (TFI->hasFP(MF) && AFI->hasStackFrame())
1438 Limit = std::min(Limit, (1U << 8) - 1);
1440 case ARMII::AddrMode4:
1441 case ARMII::AddrMode6:
1442 // Addressing modes 4 & 6 (load/store) instructions can't encode an
1443 // immediate offset for stack references.
1448 break; // At most one FI per instruction
1456 // In functions that realign the stack, it can be an advantage to spill the
1457 // callee-saved vector registers after realigning the stack. The vst1 and vld1
1458 // instructions take alignment hints that can improve performance.
1460 static void checkNumAlignedDPRCS2Regs(MachineFunction &MF) {
1461 MF.getInfo<ARMFunctionInfo>()->setNumAlignedDPRCS2Regs(0);
1462 if (!SpillAlignedNEONRegs)
1465 // Naked functions don't spill callee-saved registers.
1466 if (MF.getFunction()->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1470 // We are planning to use NEON instructions vst1 / vld1.
1471 if (!static_cast<const ARMSubtarget &>(MF.getSubtarget()).hasNEON())
1474 // Don't bother if the default stack alignment is sufficiently high.
1475 if (MF.getSubtarget().getFrameLowering()->getStackAlignment() >= 8)
1478 // Aligned spills require stack realignment.
1479 if (!static_cast<const ARMBaseRegisterInfo *>(
1480 MF.getSubtarget().getRegisterInfo())->canRealignStack(MF))
1483 // We always spill contiguous d-registers starting from d8. Count how many
1484 // needs spilling. The register allocator will almost always use the
1485 // callee-saved registers in order, but it can happen that there are holes in
1486 // the range. Registers above the hole will be spilled to the standard DPRCS
1488 MachineRegisterInfo &MRI = MF.getRegInfo();
1489 unsigned NumSpills = 0;
1490 for (; NumSpills < 8; ++NumSpills)
1491 if (!MRI.isPhysRegUsed(ARM::D8 + NumSpills))
1494 // Don't do this for just one d-register. It's not worth it.
1498 // Spill the first NumSpills D-registers after realigning the stack.
1499 MF.getInfo<ARMFunctionInfo>()->setNumAlignedDPRCS2Regs(NumSpills);
1501 // A scratch register is required for the vst1 / vld1 instructions.
1502 MF.getRegInfo().setPhysRegUsed(ARM::R4);
1506 ARMFrameLowering::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
1507 RegScavenger *RS) const {
1508 // This tells PEI to spill the FP as if it is any other callee-save register
1509 // to take advantage the eliminateFrameIndex machinery. This also ensures it
1510 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
1511 // to combine multiple loads / stores.
1512 bool CanEliminateFrame = true;
1513 bool CS1Spilled = false;
1514 bool LRSpilled = false;
1515 unsigned NumGPRSpills = 0;
1516 SmallVector<unsigned, 4> UnspilledCS1GPRs;
1517 SmallVector<unsigned, 4> UnspilledCS2GPRs;
1518 const ARMBaseRegisterInfo *RegInfo = static_cast<const ARMBaseRegisterInfo *>(
1519 MF.getSubtarget().getRegisterInfo());
1520 const ARMBaseInstrInfo &TII =
1521 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
1522 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1523 MachineFrameInfo *MFI = MF.getFrameInfo();
1524 MachineRegisterInfo &MRI = MF.getRegInfo();
1525 unsigned FramePtr = RegInfo->getFrameRegister(MF);
1527 // Spill R4 if Thumb2 function requires stack realignment - it will be used as
1528 // scratch register. Also spill R4 if Thumb2 function has varsized objects,
1529 // since it's not always possible to restore sp from fp in a single
1531 // FIXME: It will be better just to find spare register here.
1532 if (AFI->isThumb2Function() &&
1533 (MFI->hasVarSizedObjects() || RegInfo->needsStackRealignment(MF)))
1534 MRI.setPhysRegUsed(ARM::R4);
1536 if (AFI->isThumb1OnlyFunction()) {
1537 // Spill LR if Thumb1 function uses variable length argument lists.
1538 if (AFI->getArgRegsSaveSize() > 0)
1539 MRI.setPhysRegUsed(ARM::LR);
1541 // Spill R4 if Thumb1 epilogue has to restore SP from FP. We don't know
1542 // for sure what the stack size will be, but for this, an estimate is good
1543 // enough. If there anything changes it, it'll be a spill, which implies
1544 // we've used all the registers and so R4 is already used, so not marking
1545 // it here will be OK.
1546 // FIXME: It will be better just to find spare register here.
1547 unsigned StackSize = MFI->estimateStackSize(MF);
1548 if (MFI->hasVarSizedObjects() || StackSize > 508)
1549 MRI.setPhysRegUsed(ARM::R4);
1552 // See if we can spill vector registers to aligned stack.
1553 checkNumAlignedDPRCS2Regs(MF);
1555 // Spill the BasePtr if it's used.
1556 if (RegInfo->hasBasePointer(MF))
1557 MRI.setPhysRegUsed(RegInfo->getBaseRegister());
1559 // Don't spill FP if the frame can be eliminated. This is determined
1560 // by scanning the callee-save registers to see if any is used.
1561 const MCPhysReg *CSRegs = RegInfo->getCalleeSavedRegs(&MF);
1562 for (unsigned i = 0; CSRegs[i]; ++i) {
1563 unsigned Reg = CSRegs[i];
1564 bool Spilled = false;
1565 if (MRI.isPhysRegUsed(Reg)) {
1567 CanEliminateFrame = false;
1570 if (!ARM::GPRRegClass.contains(Reg))
1576 if (!STI.isTargetDarwin()) {
1583 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
1588 case ARM::R0: case ARM::R1:
1589 case ARM::R2: case ARM::R3:
1590 case ARM::R4: case ARM::R5:
1591 case ARM::R6: case ARM::R7:
1598 if (!STI.isTargetDarwin()) {
1599 UnspilledCS1GPRs.push_back(Reg);
1604 case ARM::R0: case ARM::R1:
1605 case ARM::R2: case ARM::R3:
1606 case ARM::R4: case ARM::R5:
1607 case ARM::R6: case ARM::R7:
1609 UnspilledCS1GPRs.push_back(Reg);
1612 UnspilledCS2GPRs.push_back(Reg);
1618 bool ForceLRSpill = false;
1619 if (!LRSpilled && AFI->isThumb1OnlyFunction()) {
1620 unsigned FnSize = GetFunctionSizeInBytes(MF, TII);
1621 // Force LR to be spilled if the Thumb function size is > 2048. This enables
1622 // use of BL to implement far jump. If it turns out that it's not needed
1623 // then the branch fix up path will undo it.
1624 if (FnSize >= (1 << 11)) {
1625 CanEliminateFrame = false;
1626 ForceLRSpill = true;
1630 // If any of the stack slot references may be out of range of an immediate
1631 // offset, make sure a register (or a spill slot) is available for the
1632 // register scavenger. Note that if we're indexing off the frame pointer, the
1633 // effective stack size is 4 bytes larger since the FP points to the stack
1634 // slot of the previous FP. Also, if we have variable sized objects in the
1635 // function, stack slot references will often be negative, and some of
1636 // our instructions are positive-offset only, so conservatively consider
1637 // that case to want a spill slot (or register) as well. Similarly, if
1638 // the function adjusts the stack pointer during execution and the
1639 // adjustments aren't already part of our stack size estimate, our offset
1640 // calculations may be off, so be conservative.
1641 // FIXME: We could add logic to be more precise about negative offsets
1642 // and which instructions will need a scratch register for them. Is it
1643 // worth the effort and added fragility?
1646 (MFI->estimateStackSize(MF) +
1647 ((hasFP(MF) && AFI->hasStackFrame()) ? 4:0) >=
1648 estimateRSStackSizeLimit(MF, this)))
1649 || MFI->hasVarSizedObjects()
1650 || (MFI->adjustsStack() && !canSimplifyCallFramePseudos(MF));
1652 bool ExtraCSSpill = false;
1653 if (BigStack || !CanEliminateFrame || RegInfo->cannotEliminateFrame(MF)) {
1654 AFI->setHasStackFrame(true);
1656 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
1657 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
1658 if (!LRSpilled && CS1Spilled) {
1659 MRI.setPhysRegUsed(ARM::LR);
1661 SmallVectorImpl<unsigned>::iterator LRPos;
1662 LRPos = std::find(UnspilledCS1GPRs.begin(), UnspilledCS1GPRs.end(),
1664 if (LRPos != UnspilledCS1GPRs.end())
1665 UnspilledCS1GPRs.erase(LRPos);
1667 ForceLRSpill = false;
1668 ExtraCSSpill = true;
1672 MRI.setPhysRegUsed(FramePtr);
1673 auto FPPos = std::find(UnspilledCS1GPRs.begin(), UnspilledCS1GPRs.end(),
1675 if (FPPos != UnspilledCS1GPRs.end())
1676 UnspilledCS1GPRs.erase(FPPos);
1680 // If stack and double are 8-byte aligned and we are spilling an odd number
1681 // of GPRs, spill one extra callee save GPR so we won't have to pad between
1682 // the integer and double callee save areas.
1683 unsigned TargetAlign = getStackAlignment();
1684 if (TargetAlign >= 8 && (NumGPRSpills & 1)) {
1685 if (CS1Spilled && !UnspilledCS1GPRs.empty()) {
1686 for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) {
1687 unsigned Reg = UnspilledCS1GPRs[i];
1688 // Don't spill high register if the function is thumb1
1689 if (!AFI->isThumb1OnlyFunction() ||
1690 isARMLowRegister(Reg) || Reg == ARM::LR) {
1691 MRI.setPhysRegUsed(Reg);
1692 if (!MRI.isReserved(Reg))
1693 ExtraCSSpill = true;
1697 } else if (!UnspilledCS2GPRs.empty() && !AFI->isThumb1OnlyFunction()) {
1698 unsigned Reg = UnspilledCS2GPRs.front();
1699 MRI.setPhysRegUsed(Reg);
1700 if (!MRI.isReserved(Reg))
1701 ExtraCSSpill = true;
1705 // Estimate if we might need to scavenge a register at some point in order
1706 // to materialize a stack offset. If so, either spill one additional
1707 // callee-saved register or reserve a special spill slot to facilitate
1708 // register scavenging. Thumb1 needs a spill slot for stack pointer
1709 // adjustments also, even when the frame itself is small.
1710 if (BigStack && !ExtraCSSpill) {
1711 // If any non-reserved CS register isn't spilled, just spill one or two
1712 // extra. That should take care of it!
1713 unsigned NumExtras = TargetAlign / 4;
1714 SmallVector<unsigned, 2> Extras;
1715 while (NumExtras && !UnspilledCS1GPRs.empty()) {
1716 unsigned Reg = UnspilledCS1GPRs.back();
1717 UnspilledCS1GPRs.pop_back();
1718 if (!MRI.isReserved(Reg) &&
1719 (!AFI->isThumb1OnlyFunction() || isARMLowRegister(Reg) ||
1721 Extras.push_back(Reg);
1725 // For non-Thumb1 functions, also check for hi-reg CS registers
1726 if (!AFI->isThumb1OnlyFunction()) {
1727 while (NumExtras && !UnspilledCS2GPRs.empty()) {
1728 unsigned Reg = UnspilledCS2GPRs.back();
1729 UnspilledCS2GPRs.pop_back();
1730 if (!MRI.isReserved(Reg)) {
1731 Extras.push_back(Reg);
1736 if (Extras.size() && NumExtras == 0) {
1737 for (unsigned i = 0, e = Extras.size(); i != e; ++i) {
1738 MRI.setPhysRegUsed(Extras[i]);
1740 } else if (!AFI->isThumb1OnlyFunction()) {
1741 // note: Thumb1 functions spill to R12, not the stack. Reserve a slot
1742 // closest to SP or frame pointer.
1743 const TargetRegisterClass *RC = &ARM::GPRRegClass;
1744 RS->addScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
1752 MRI.setPhysRegUsed(ARM::LR);
1753 AFI->setLRIsSpilledForFarJump(true);
1758 void ARMFrameLowering::
1759 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
1760 MachineBasicBlock::iterator I) const {
1761 const ARMBaseInstrInfo &TII =
1762 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
1763 if (!hasReservedCallFrame(MF)) {
1764 // If we have alloca, convert as follows:
1765 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
1766 // ADJCALLSTACKUP -> add, sp, sp, amount
1767 MachineInstr *Old = I;
1768 DebugLoc dl = Old->getDebugLoc();
1769 unsigned Amount = Old->getOperand(0).getImm();
1771 // We need to keep the stack aligned properly. To do this, we round the
1772 // amount of space needed for the outgoing arguments up to the next
1773 // alignment boundary.
1774 unsigned Align = getStackAlignment();
1775 Amount = (Amount+Align-1)/Align*Align;
1777 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1778 assert(!AFI->isThumb1OnlyFunction() &&
1779 "This eliminateCallFramePseudoInstr does not support Thumb1!");
1780 bool isARM = !AFI->isThumbFunction();
1782 // Replace the pseudo instruction with a new instruction...
1783 unsigned Opc = Old->getOpcode();
1784 int PIdx = Old->findFirstPredOperandIdx();
1785 ARMCC::CondCodes Pred = (PIdx == -1)
1786 ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(PIdx).getImm();
1787 if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) {
1788 // Note: PredReg is operand 2 for ADJCALLSTACKDOWN.
1789 unsigned PredReg = Old->getOperand(2).getReg();
1790 emitSPUpdate(isARM, MBB, I, dl, TII, -Amount, MachineInstr::NoFlags,
1793 // Note: PredReg is operand 3 for ADJCALLSTACKUP.
1794 unsigned PredReg = Old->getOperand(3).getReg();
1795 assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP);
1796 emitSPUpdate(isARM, MBB, I, dl, TII, Amount, MachineInstr::NoFlags,
1804 /// Get the minimum constant for ARM that is greater than or equal to the
1805 /// argument. In ARM, constants can have any value that can be produced by
1806 /// rotating an 8-bit value to the right by an even number of bits within a
1808 static uint32_t alignToARMConstant(uint32_t Value) {
1809 unsigned Shifted = 0;
1814 while (!(Value & 0xC0000000)) {
1819 bool Carry = (Value & 0x00FFFFFF);
1820 Value = ((Value & 0xFF000000) >> 24) + Carry;
1822 if (Value & 0x0000100)
1823 Value = Value & 0x000001FC;
1826 Value = Value >> (Shifted - 24);
1828 Value = Value << (24 - Shifted);
1833 // The stack limit in the TCB is set to this many bytes above the actual
1835 static const uint64_t kSplitStackAvailable = 256;
1837 // Adjust the function prologue to enable split stacks. This currently only
1838 // supports android and linux.
1840 // The ABI of the segmented stack prologue is a little arbitrarily chosen, but
1841 // must be well defined in order to allow for consistent implementations of the
1842 // __morestack helper function. The ABI is also not a normal ABI in that it
1843 // doesn't follow the normal calling conventions because this allows the
1844 // prologue of each function to be optimized further.
1846 // Currently, the ABI looks like (when calling __morestack)
1848 // * r4 holds the minimum stack size requested for this function call
1849 // * r5 holds the stack size of the arguments to the function
1850 // * the beginning of the function is 3 instructions after the call to
1853 // Implementations of __morestack should use r4 to allocate a new stack, r5 to
1854 // place the arguments on to the new stack, and the 3-instruction knowledge to
1855 // jump directly to the body of the function when working on the new stack.
1857 // An old (and possibly no longer compatible) implementation of __morestack for
1858 // ARM can be found at [1].
1860 // [1] - https://github.com/mozilla/rust/blob/86efd9/src/rt/arch/arm/morestack.S
1861 void ARMFrameLowering::adjustForSegmentedStacks(MachineFunction &MF) const {
1864 const ARMSubtarget *ST = &MF.getTarget().getSubtarget<ARMSubtarget>();
1865 bool Thumb = ST->isThumb();
1867 // Sadly, this currently doesn't support varargs, platforms other than
1868 // android/linux. Note that thumb1/thumb2 are support for android/linux.
1869 if (MF.getFunction()->isVarArg())
1870 report_fatal_error("Segmented stacks do not support vararg functions.");
1871 if (!ST->isTargetAndroid() && !ST->isTargetLinux())
1872 report_fatal_error("Segmented stacks not supported on this platform.");
1874 MachineBasicBlock &prologueMBB = MF.front();
1875 MachineFrameInfo *MFI = MF.getFrameInfo();
1876 MachineModuleInfo &MMI = MF.getMMI();
1877 MCContext &Context = MMI.getContext();
1878 const MCRegisterInfo *MRI = Context.getRegisterInfo();
1879 const ARMBaseInstrInfo &TII =
1880 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
1881 ARMFunctionInfo *ARMFI = MF.getInfo<ARMFunctionInfo>();
1884 uint64_t StackSize = MFI->getStackSize();
1886 // Do not generate a prologue for functions with a stack of size zero
1890 // Use R4 and R5 as scratch registers.
1891 // We save R4 and R5 before use and restore them before leaving the function.
1892 unsigned ScratchReg0 = ARM::R4;
1893 unsigned ScratchReg1 = ARM::R5;
1894 uint64_t AlignedStackSize;
1896 MachineBasicBlock *PrevStackMBB = MF.CreateMachineBasicBlock();
1897 MachineBasicBlock *PostStackMBB = MF.CreateMachineBasicBlock();
1898 MachineBasicBlock *AllocMBB = MF.CreateMachineBasicBlock();
1899 MachineBasicBlock *GetMBB = MF.CreateMachineBasicBlock();
1900 MachineBasicBlock *McrMBB = MF.CreateMachineBasicBlock();
1902 for (MachineBasicBlock::livein_iterator i = prologueMBB.livein_begin(),
1903 e = prologueMBB.livein_end();
1905 AllocMBB->addLiveIn(*i);
1906 GetMBB->addLiveIn(*i);
1907 McrMBB->addLiveIn(*i);
1908 PrevStackMBB->addLiveIn(*i);
1909 PostStackMBB->addLiveIn(*i);
1912 MF.push_front(PostStackMBB);
1913 MF.push_front(AllocMBB);
1914 MF.push_front(GetMBB);
1915 MF.push_front(McrMBB);
1916 MF.push_front(PrevStackMBB);
1918 // The required stack size that is aligned to ARM constant criterion.
1919 AlignedStackSize = alignToARMConstant(StackSize);
1921 // When the frame size is less than 256 we just compare the stack
1922 // boundary directly to the value of the stack pointer, per gcc.
1923 bool CompareStackPointer = AlignedStackSize < kSplitStackAvailable;
1925 // We will use two of the callee save registers as scratch registers so we
1926 // need to save those registers onto the stack.
1927 // We will use SR0 to hold stack limit and SR1 to hold the stack size
1928 // requested and arguments for __morestack().
1929 // SR0: Scratch Register #0
1930 // SR1: Scratch Register #1
1933 AddDefaultPred(BuildMI(PrevStackMBB, DL, TII.get(ARM::tPUSH)))
1934 .addReg(ScratchReg0).addReg(ScratchReg1);
1936 AddDefaultPred(BuildMI(PrevStackMBB, DL, TII.get(ARM::STMDB_UPD))
1937 .addReg(ARM::SP, RegState::Define).addReg(ARM::SP))
1938 .addReg(ScratchReg0).addReg(ScratchReg1);
1941 // Emit the relevant DWARF information about the change in stack pointer as
1942 // well as where to find both r4 and r5 (the callee-save registers)
1944 MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, -8));
1945 BuildMI(PrevStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1946 .addCFIIndex(CFIIndex);
1947 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
1948 nullptr, MRI->getDwarfRegNum(ScratchReg1, true), -4));
1949 BuildMI(PrevStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1950 .addCFIIndex(CFIIndex);
1951 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
1952 nullptr, MRI->getDwarfRegNum(ScratchReg0, true), -8));
1953 BuildMI(PrevStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1954 .addCFIIndex(CFIIndex);
1958 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::tMOVr), ScratchReg1)
1960 } else if (CompareStackPointer) {
1961 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::MOVr), ScratchReg1)
1962 .addReg(ARM::SP)).addReg(0);
1965 // sub SR1, sp, #StackSize
1966 if (!CompareStackPointer && Thumb) {
1968 AddDefaultCC(BuildMI(McrMBB, DL, TII.get(ARM::tSUBi8), ScratchReg1))
1969 .addReg(ScratchReg1).addImm(AlignedStackSize));
1970 } else if (!CompareStackPointer) {
1971 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::SUBri), ScratchReg1)
1972 .addReg(ARM::SP).addImm(AlignedStackSize)).addReg(0);
1975 if (Thumb && ST->isThumb1Only()) {
1976 unsigned PCLabelId = ARMFI->createPICLabelUId();
1977 ARMConstantPoolValue *NewCPV = ARMConstantPoolSymbol::Create(
1978 MF.getFunction()->getContext(), "__STACK_LIMIT", PCLabelId, 0);
1979 MachineConstantPool *MCP = MF.getConstantPool();
1980 unsigned CPI = MCP->getConstantPoolIndex(NewCPV, MF.getAlignment());
1982 // ldr SR0, [pc, offset(STACK_LIMIT)]
1983 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(ARM::tLDRpci), ScratchReg0)
1984 .addConstantPoolIndex(CPI));
1987 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(ARM::tLDRi), ScratchReg0)
1988 .addReg(ScratchReg0).addImm(0));
1990 // Get TLS base address from the coprocessor
1991 // mrc p15, #0, SR0, c13, c0, #3
1992 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::MRC), ScratchReg0)
1999 // Use the last tls slot on android and a private field of the TCP on linux.
2000 assert(ST->isTargetAndroid() || ST->isTargetLinux());
2001 unsigned TlsOffset = ST->isTargetAndroid() ? 63 : 1;
2003 // Get the stack limit from the right offset
2004 // ldr SR0, [sr0, #4 * TlsOffset]
2005 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(ARM::LDRi12), ScratchReg0)
2006 .addReg(ScratchReg0).addImm(4 * TlsOffset));
2009 // Compare stack limit with stack size requested.
2011 Opcode = Thumb ? ARM::tCMPr : ARM::CMPrr;
2012 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(Opcode))
2013 .addReg(ScratchReg0)
2014 .addReg(ScratchReg1));
2016 // This jump is taken if StackLimit < SP - stack required.
2017 Opcode = Thumb ? ARM::tBcc : ARM::Bcc;
2018 BuildMI(GetMBB, DL, TII.get(Opcode)).addMBB(PostStackMBB)
2023 // Calling __morestack(StackSize, Size of stack arguments).
2024 // __morestack knows that the stack size requested is in SR0(r4)
2025 // and amount size of stack arguments is in SR1(r5).
2027 // Pass first argument for the __morestack by Scratch Register #0.
2028 // The amount size of stack required
2030 AddDefaultPred(AddDefaultCC(BuildMI(AllocMBB, DL, TII.get(ARM::tMOVi8),
2031 ScratchReg0)).addImm(AlignedStackSize));
2033 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::MOVi), ScratchReg0)
2034 .addImm(AlignedStackSize)).addReg(0);
2036 // Pass second argument for the __morestack by Scratch Register #1.
2037 // The amount size of stack consumed to save function arguments.
2040 AddDefaultCC(BuildMI(AllocMBB, DL, TII.get(ARM::tMOVi8), ScratchReg1))
2041 .addImm(alignToARMConstant(ARMFI->getArgumentStackSize())));
2043 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::MOVi), ScratchReg1)
2044 .addImm(alignToARMConstant(ARMFI->getArgumentStackSize())))
2048 // push {lr} - Save return address of this function.
2050 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tPUSH)))
2053 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::STMDB_UPD))
2054 .addReg(ARM::SP, RegState::Define)
2059 // Emit the DWARF info about the change in stack as well as where to find the
2060 // previous link register
2062 MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, -12));
2063 BuildMI(AllocMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2064 .addCFIIndex(CFIIndex);
2065 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
2066 nullptr, MRI->getDwarfRegNum(ARM::LR, true), -12));
2067 BuildMI(AllocMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2068 .addCFIIndex(CFIIndex);
2070 // Call __morestack().
2072 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tBL)))
2073 .addExternalSymbol("__morestack");
2075 BuildMI(AllocMBB, DL, TII.get(ARM::BL))
2076 .addExternalSymbol("__morestack");
2079 // pop {lr} - Restore return address of this original function.
2081 if (ST->isThumb1Only()) {
2082 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tPOP)))
2083 .addReg(ScratchReg0);
2084 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tMOVr), ARM::LR)
2085 .addReg(ScratchReg0));
2087 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::t2LDR_POST))
2088 .addReg(ARM::LR, RegState::Define)
2089 .addReg(ARM::SP, RegState::Define)
2094 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::LDMIA_UPD))
2095 .addReg(ARM::SP, RegState::Define)
2100 // Restore SR0 and SR1 in case of __morestack() was called.
2101 // __morestack() will skip PostStackMBB block so we need to restore
2102 // scratch registers from here.
2105 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tPOP)))
2106 .addReg(ScratchReg0)
2107 .addReg(ScratchReg1);
2109 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::LDMIA_UPD))
2110 .addReg(ARM::SP, RegState::Define)
2112 .addReg(ScratchReg0)
2113 .addReg(ScratchReg1);
2116 // Update the CFA offset now that we've popped
2117 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, 0));
2118 BuildMI(AllocMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2119 .addCFIIndex(CFIIndex);
2121 // bx lr - Return from this function.
2122 Opcode = Thumb ? ARM::tBX_RET : ARM::BX_RET;
2123 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(Opcode)));
2125 // Restore SR0 and SR1 in case of __morestack() was not called.
2128 AddDefaultPred(BuildMI(PostStackMBB, DL, TII.get(ARM::tPOP)))
2129 .addReg(ScratchReg0)
2130 .addReg(ScratchReg1);
2132 AddDefaultPred(BuildMI(PostStackMBB, DL, TII.get(ARM::LDMIA_UPD))
2133 .addReg(ARM::SP, RegState::Define)
2135 .addReg(ScratchReg0)
2136 .addReg(ScratchReg1);
2139 // Update the CFA offset now that we've popped
2140 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, 0));
2141 BuildMI(PostStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2142 .addCFIIndex(CFIIndex);
2144 // Tell debuggers that r4 and r5 are now the same as they were in the
2145 // previous function, that they're the "Same Value".
2146 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createSameValue(
2147 nullptr, MRI->getDwarfRegNum(ScratchReg0, true)));
2148 BuildMI(PostStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2149 .addCFIIndex(CFIIndex);
2150 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createSameValue(
2151 nullptr, MRI->getDwarfRegNum(ScratchReg1, true)));
2152 BuildMI(PostStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2153 .addCFIIndex(CFIIndex);
2155 // Organizing MBB lists
2156 PostStackMBB->addSuccessor(&prologueMBB);
2158 AllocMBB->addSuccessor(PostStackMBB);
2160 GetMBB->addSuccessor(PostStackMBB);
2161 GetMBB->addSuccessor(AllocMBB);
2163 McrMBB->addSuccessor(GetMBB);
2165 PrevStackMBB->addSuccessor(McrMBB);