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 bool ARMFrameLowering::noFramePointerElim(const MachineFunction &MF) const {
47 // iOS always has a FP for backtracking, force other targets to keep their FP
48 // when doing FastISel. The emitted code is currently superior, and in cases
49 // like test-suite's lencod FastISel isn't quite correct when FP is eliminated.
50 return TargetFrameLowering::noFramePointerElim(MF) ||
51 MF.getSubtarget<ARMSubtarget>().useFastISel();
54 /// hasFP - Return true if the specified function should have a dedicated frame
55 /// pointer register. This is true if the function has variable sized allocas
56 /// or if frame pointer elimination is disabled.
57 bool ARMFrameLowering::hasFP(const MachineFunction &MF) const {
58 const TargetRegisterInfo *RegInfo = MF.getSubtarget().getRegisterInfo();
60 // iOS requires FP not to be clobbered for backtracing purpose.
61 if (STI.isTargetIOS())
64 const MachineFrameInfo *MFI = MF.getFrameInfo();
65 // Always eliminate non-leaf frame pointers.
66 return ((MF.getTarget().Options.DisableFramePointerElim(MF) &&
68 RegInfo->needsStackRealignment(MF) ||
69 MFI->hasVarSizedObjects() ||
70 MFI->isFrameAddressTaken());
73 /// hasReservedCallFrame - Under normal circumstances, when a frame pointer is
74 /// not required, we reserve argument space for call sites in the function
75 /// immediately on entry to the current function. This eliminates the need for
76 /// add/sub sp brackets around call sites. Returns true if the call frame is
77 /// included as part of the stack frame.
78 bool ARMFrameLowering::hasReservedCallFrame(const MachineFunction &MF) const {
79 const MachineFrameInfo *FFI = MF.getFrameInfo();
80 unsigned CFSize = FFI->getMaxCallFrameSize();
81 // It's not always a good idea to include the call frame as part of the
82 // stack frame. ARM (especially Thumb) has small immediate offset to
83 // address the stack frame. So a large call frame can cause poor codegen
84 // and may even makes it impossible to scavenge a register.
85 if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12
88 return !MF.getFrameInfo()->hasVarSizedObjects();
91 /// canSimplifyCallFramePseudos - If there is a reserved call frame, the
92 /// call frame pseudos can be simplified. Unlike most targets, having a FP
93 /// is not sufficient here since we still may reference some objects via SP
94 /// even when FP is available in Thumb2 mode.
96 ARMFrameLowering::canSimplifyCallFramePseudos(const MachineFunction &MF) const {
97 return hasReservedCallFrame(MF) || MF.getFrameInfo()->hasVarSizedObjects();
100 static bool isCSRestore(MachineInstr *MI,
101 const ARMBaseInstrInfo &TII,
102 const MCPhysReg *CSRegs) {
103 // Integer spill area is handled with "pop".
104 if (isPopOpcode(MI->getOpcode())) {
105 // The first two operands are predicates. The last two are
106 // imp-def and imp-use of SP. Check everything in between.
107 for (int i = 5, e = MI->getNumOperands(); i != e; ++i)
108 if (!isCalleeSavedRegister(MI->getOperand(i).getReg(), CSRegs))
112 if ((MI->getOpcode() == ARM::LDR_POST_IMM ||
113 MI->getOpcode() == ARM::LDR_POST_REG ||
114 MI->getOpcode() == ARM::t2LDR_POST) &&
115 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs) &&
116 MI->getOperand(1).getReg() == ARM::SP)
122 static void emitRegPlusImmediate(bool isARM, MachineBasicBlock &MBB,
123 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
124 const ARMBaseInstrInfo &TII, unsigned DestReg,
125 unsigned SrcReg, int NumBytes,
126 unsigned MIFlags = MachineInstr::NoFlags,
127 ARMCC::CondCodes Pred = ARMCC::AL,
128 unsigned PredReg = 0) {
130 emitARMRegPlusImmediate(MBB, MBBI, dl, DestReg, SrcReg, NumBytes,
131 Pred, PredReg, TII, MIFlags);
133 emitT2RegPlusImmediate(MBB, MBBI, dl, DestReg, SrcReg, NumBytes,
134 Pred, PredReg, TII, MIFlags);
137 static void emitSPUpdate(bool isARM, MachineBasicBlock &MBB,
138 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
139 const ARMBaseInstrInfo &TII, int NumBytes,
140 unsigned MIFlags = MachineInstr::NoFlags,
141 ARMCC::CondCodes Pred = ARMCC::AL,
142 unsigned PredReg = 0) {
143 emitRegPlusImmediate(isARM, MBB, MBBI, dl, TII, ARM::SP, ARM::SP, NumBytes,
144 MIFlags, Pred, PredReg);
147 static int sizeOfSPAdjustment(const MachineInstr *MI) {
149 switch (MI->getOpcode()) {
150 case ARM::VSTMDDB_UPD:
154 case ARM::t2STMDB_UPD:
158 case ARM::STR_PRE_IMM:
161 llvm_unreachable("Unknown push or pop like instruction");
165 // ARM and Thumb2 push/pop insts have explicit "sp, sp" operands (+
166 // pred) so the list starts at 4.
167 for (int i = MI->getNumOperands() - 1; i >= 4; --i)
172 static bool WindowsRequiresStackProbe(const MachineFunction &MF,
173 size_t StackSizeInBytes) {
174 const MachineFrameInfo *MFI = MF.getFrameInfo();
175 const Function *F = MF.getFunction();
176 unsigned StackProbeSize = (MFI->getStackProtectorIndex() > 0) ? 4080 : 4096;
177 if (F->hasFnAttribute("stack-probe-size"))
178 F->getFnAttribute("stack-probe-size")
180 .getAsInteger(0, StackProbeSize);
181 return StackSizeInBytes >= StackProbeSize;
185 struct StackAdjustingInsts {
187 MachineBasicBlock::iterator I;
192 SmallVector<InstInfo, 4> Insts;
194 void addInst(MachineBasicBlock::iterator I, unsigned SPAdjust,
195 bool BeforeFPSet = false) {
196 InstInfo Info = {I, SPAdjust, BeforeFPSet};
197 Insts.push_back(Info);
200 void addExtraBytes(const MachineBasicBlock::iterator I, unsigned ExtraBytes) {
201 auto Info = std::find_if(Insts.begin(), Insts.end(),
202 [&](InstInfo &Info) { return Info.I == I; });
203 assert(Info != Insts.end() && "invalid sp adjusting instruction");
204 Info->SPAdjust += ExtraBytes;
207 void emitDefCFAOffsets(MachineModuleInfo &MMI, MachineBasicBlock &MBB,
208 DebugLoc dl, const ARMBaseInstrInfo &TII, bool HasFP) {
209 unsigned CFAOffset = 0;
210 for (auto &Info : Insts) {
211 if (HasFP && !Info.BeforeFPSet)
214 CFAOffset -= Info.SPAdjust;
215 unsigned CFIIndex = MMI.addFrameInst(
216 MCCFIInstruction::createDefCfaOffset(nullptr, CFAOffset));
217 BuildMI(MBB, std::next(Info.I), dl,
218 TII.get(TargetOpcode::CFI_INSTRUCTION))
219 .addCFIIndex(CFIIndex)
220 .setMIFlags(MachineInstr::FrameSetup);
226 /// Emit an instruction sequence that will align the address in
227 /// register Reg by zero-ing out the lower bits. For versions of the
228 /// architecture that support Neon, this must be done in a single
229 /// instruction, since skipAlignedDPRCS2Spills assumes it is done in a
230 /// single instruction. That function only gets called when optimizing
231 /// spilling of D registers on a core with the Neon instruction set
233 static void emitAligningInstructions(MachineFunction &MF, ARMFunctionInfo *AFI,
234 const TargetInstrInfo &TII,
235 MachineBasicBlock &MBB,
236 MachineBasicBlock::iterator MBBI,
237 DebugLoc DL, const unsigned Reg,
238 const unsigned Alignment,
239 const bool MustBeSingleInstruction) {
240 const ARMSubtarget &AST =
241 static_cast<const ARMSubtarget &>(MF.getSubtarget());
242 const bool CanUseBFC = AST.hasV6T2Ops() || AST.hasV7Ops();
243 const unsigned AlignMask = Alignment - 1;
244 const unsigned NrBitsToZero = countTrailingZeros(Alignment);
245 assert(!AFI->isThumb1OnlyFunction() && "Thumb1 not supported");
246 if (!AFI->isThumbFunction()) {
247 // if the BFC instruction is available, use that to zero the lower
249 // bfc Reg, #0, log2(Alignment)
250 // otherwise use BIC, if the mask to zero the required number of bits
251 // can be encoded in the bic immediate field
252 // bic Reg, Reg, Alignment-1
254 // lsr Reg, Reg, log2(Alignment)
255 // lsl Reg, Reg, log2(Alignment)
257 AddDefaultPred(BuildMI(MBB, MBBI, DL, TII.get(ARM::BFC), Reg)
258 .addReg(Reg, RegState::Kill)
259 .addImm(~AlignMask));
260 } else if (AlignMask <= 255) {
262 AddDefaultPred(BuildMI(MBB, MBBI, DL, TII.get(ARM::BICri), Reg)
263 .addReg(Reg, RegState::Kill)
264 .addImm(AlignMask)));
266 assert(!MustBeSingleInstruction &&
267 "Shouldn't call emitAligningInstructions demanding a single "
268 "instruction to be emitted for large stack alignment for a target "
270 AddDefaultCC(AddDefaultPred(
271 BuildMI(MBB, MBBI, DL, TII.get(ARM::MOVsi), Reg)
272 .addReg(Reg, RegState::Kill)
273 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsr, NrBitsToZero))));
274 AddDefaultCC(AddDefaultPred(
275 BuildMI(MBB, MBBI, DL, TII.get(ARM::MOVsi), Reg)
276 .addReg(Reg, RegState::Kill)
277 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, NrBitsToZero))));
280 // Since this is only reached for Thumb-2 targets, the BFC instruction
281 // should always be available.
283 AddDefaultPred(BuildMI(MBB, MBBI, DL, TII.get(ARM::t2BFC), Reg)
284 .addReg(Reg, RegState::Kill)
285 .addImm(~AlignMask));
289 void ARMFrameLowering::emitPrologue(MachineFunction &MF,
290 MachineBasicBlock &MBB) const {
291 assert(&MBB == &MF.front() && "Shrink-wrapping not yet implemented");
292 MachineBasicBlock::iterator MBBI = MBB.begin();
293 MachineFrameInfo *MFI = MF.getFrameInfo();
294 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
295 MachineModuleInfo &MMI = MF.getMMI();
296 MCContext &Context = MMI.getContext();
297 const TargetMachine &TM = MF.getTarget();
298 const MCRegisterInfo *MRI = Context.getRegisterInfo();
299 const ARMBaseRegisterInfo *RegInfo = STI.getRegisterInfo();
300 const ARMBaseInstrInfo &TII = *STI.getInstrInfo();
301 assert(!AFI->isThumb1OnlyFunction() &&
302 "This emitPrologue does not support Thumb1!");
303 bool isARM = !AFI->isThumbFunction();
304 unsigned Align = STI.getFrameLowering()->getStackAlignment();
305 unsigned ArgRegsSaveSize = AFI->getArgRegsSaveSize();
306 unsigned NumBytes = MFI->getStackSize();
307 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
308 DebugLoc dl = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
309 unsigned FramePtr = RegInfo->getFrameRegister(MF);
311 // Determine the sizes of each callee-save spill areas and record which frame
312 // belongs to which callee-save spill areas.
313 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
314 int FramePtrSpillFI = 0;
317 // All calls are tail calls in GHC calling conv, and functions have no
318 // prologue/epilogue.
319 if (MF.getFunction()->getCallingConv() == CallingConv::GHC)
322 StackAdjustingInsts DefCFAOffsetCandidates;
323 bool HasFP = hasFP(MF);
325 // Allocate the vararg register save area.
326 if (ArgRegsSaveSize) {
327 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -ArgRegsSaveSize,
328 MachineInstr::FrameSetup);
329 DefCFAOffsetCandidates.addInst(std::prev(MBBI), ArgRegsSaveSize, true);
332 if (!AFI->hasStackFrame() &&
333 (!STI.isTargetWindows() || !WindowsRequiresStackProbe(MF, NumBytes))) {
334 if (NumBytes - ArgRegsSaveSize != 0) {
335 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -(NumBytes - ArgRegsSaveSize),
336 MachineInstr::FrameSetup);
337 DefCFAOffsetCandidates.addInst(std::prev(MBBI),
338 NumBytes - ArgRegsSaveSize, true);
340 DefCFAOffsetCandidates.emitDefCFAOffsets(MMI, MBB, dl, TII, HasFP);
344 // Determine spill area sizes.
345 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
346 unsigned Reg = CSI[i].getReg();
347 int FI = CSI[i].getFrameIdx();
354 if (STI.isTargetDarwin()) {
369 FramePtrSpillFI = FI;
373 // This is a DPR. Exclude the aligned DPRCS2 spills.
376 if (Reg < ARM::D8 || Reg >= ARM::D8 + AFI->getNumAlignedDPRCS2Regs())
382 MachineBasicBlock::iterator LastPush = MBB.end(), GPRCS1Push, GPRCS2Push;
383 if (GPRCS1Size > 0) {
384 GPRCS1Push = LastPush = MBBI++;
385 DefCFAOffsetCandidates.addInst(LastPush, GPRCS1Size, true);
388 // Determine starting offsets of spill areas.
389 unsigned GPRCS1Offset = NumBytes - ArgRegsSaveSize - GPRCS1Size;
390 unsigned GPRCS2Offset = GPRCS1Offset - GPRCS2Size;
391 unsigned DPRAlign = DPRCSSize ? std::min(8U, Align) : 4U;
392 unsigned DPRGapSize = (GPRCS1Size + GPRCS2Size + ArgRegsSaveSize) % DPRAlign;
393 unsigned DPRCSOffset = GPRCS2Offset - DPRGapSize - DPRCSSize;
394 int FramePtrOffsetInPush = 0;
396 FramePtrOffsetInPush =
397 MFI->getObjectOffset(FramePtrSpillFI) + ArgRegsSaveSize;
398 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) +
401 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
402 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
403 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
406 if (GPRCS2Size > 0) {
407 GPRCS2Push = LastPush = MBBI++;
408 DefCFAOffsetCandidates.addInst(LastPush, GPRCS2Size);
411 // Prolog/epilog inserter assumes we correctly align DPRs on the stack, so our
412 // .cfi_offset operations will reflect that.
414 assert(DPRGapSize == 4 && "unexpected alignment requirements for DPRs");
415 if (tryFoldSPUpdateIntoPushPop(STI, MF, LastPush, DPRGapSize))
416 DefCFAOffsetCandidates.addExtraBytes(LastPush, DPRGapSize);
418 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -DPRGapSize,
419 MachineInstr::FrameSetup);
420 DefCFAOffsetCandidates.addInst(std::prev(MBBI), DPRGapSize);
426 // Since vpush register list cannot have gaps, there may be multiple vpush
427 // instructions in the prologue.
428 while (MBBI->getOpcode() == ARM::VSTMDDB_UPD) {
429 DefCFAOffsetCandidates.addInst(MBBI, sizeOfSPAdjustment(MBBI));
434 // Move past the aligned DPRCS2 area.
435 if (AFI->getNumAlignedDPRCS2Regs() > 0) {
436 MBBI = skipAlignedDPRCS2Spills(MBBI, AFI->getNumAlignedDPRCS2Regs());
437 // The code inserted by emitAlignedDPRCS2Spills realigns the stack, and
438 // leaves the stack pointer pointing to the DPRCS2 area.
440 // Adjust NumBytes to represent the stack slots below the DPRCS2 area.
441 NumBytes += MFI->getObjectOffset(D8SpillFI);
443 NumBytes = DPRCSOffset;
445 if (STI.isTargetWindows() && WindowsRequiresStackProbe(MF, NumBytes)) {
446 uint32_t NumWords = NumBytes >> 2;
448 if (NumWords < 65536)
449 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVi16), ARM::R4)
451 .setMIFlags(MachineInstr::FrameSetup));
453 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVi32imm), ARM::R4)
455 .setMIFlags(MachineInstr::FrameSetup);
457 switch (TM.getCodeModel()) {
458 case CodeModel::Small:
459 case CodeModel::Medium:
460 case CodeModel::Default:
461 case CodeModel::Kernel:
462 BuildMI(MBB, MBBI, dl, TII.get(ARM::tBL))
463 .addImm((unsigned)ARMCC::AL).addReg(0)
464 .addExternalSymbol("__chkstk")
465 .addReg(ARM::R4, RegState::Implicit)
466 .setMIFlags(MachineInstr::FrameSetup);
468 case CodeModel::Large:
469 case CodeModel::JITDefault:
470 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVi32imm), ARM::R12)
471 .addExternalSymbol("__chkstk")
472 .setMIFlags(MachineInstr::FrameSetup);
474 BuildMI(MBB, MBBI, dl, TII.get(ARM::tBLXr))
475 .addImm((unsigned)ARMCC::AL).addReg(0)
476 .addReg(ARM::R12, RegState::Kill)
477 .addReg(ARM::R4, RegState::Implicit)
478 .setMIFlags(MachineInstr::FrameSetup);
482 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::t2SUBrr),
484 .addReg(ARM::SP, RegState::Define)
485 .addReg(ARM::R4, RegState::Kill)
486 .setMIFlags(MachineInstr::FrameSetup)));
491 // Adjust SP after all the callee-save spills.
492 if (tryFoldSPUpdateIntoPushPop(STI, MF, LastPush, NumBytes))
493 DefCFAOffsetCandidates.addExtraBytes(LastPush, NumBytes);
495 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes,
496 MachineInstr::FrameSetup);
497 DefCFAOffsetCandidates.addInst(std::prev(MBBI), NumBytes);
501 // Restore from fp only in ARM mode: e.g. sub sp, r7, #24
502 // Note it's not safe to do this in Thumb2 mode because it would have
503 // taken two instructions:
506 // If an interrupt is taken between the two instructions, then sp is in
507 // an inconsistent state (pointing to the middle of callee-saved area).
508 // The interrupt handler can end up clobbering the registers.
509 AFI->setShouldRestoreSPFromFP(true);
512 // Set FP to point to the stack slot that contains the previous FP.
513 // For iOS, FP is R7, which has now been stored in spill area 1.
514 // Otherwise, if this is not iOS, all the callee-saved registers go
515 // into spill area 1, including the FP in R11. In either case, it
516 // is in area one and the adjustment needs to take place just after
519 MachineBasicBlock::iterator AfterPush = std::next(GPRCS1Push);
520 unsigned PushSize = sizeOfSPAdjustment(GPRCS1Push);
521 emitRegPlusImmediate(!AFI->isThumbFunction(), MBB, AfterPush,
522 dl, TII, FramePtr, ARM::SP,
523 PushSize + FramePtrOffsetInPush,
524 MachineInstr::FrameSetup);
525 if (FramePtrOffsetInPush + PushSize != 0) {
526 unsigned CFIIndex = MMI.addFrameInst(MCCFIInstruction::createDefCfa(
527 nullptr, MRI->getDwarfRegNum(FramePtr, true),
528 -(ArgRegsSaveSize - FramePtrOffsetInPush)));
529 BuildMI(MBB, AfterPush, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
530 .addCFIIndex(CFIIndex)
531 .setMIFlags(MachineInstr::FrameSetup);
534 MMI.addFrameInst(MCCFIInstruction::createDefCfaRegister(
535 nullptr, MRI->getDwarfRegNum(FramePtr, true)));
536 BuildMI(MBB, AfterPush, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
537 .addCFIIndex(CFIIndex)
538 .setMIFlags(MachineInstr::FrameSetup);
542 // Now that the prologue's actual instructions are finalised, we can insert
543 // the necessary DWARF cf instructions to describe the situation. Start by
544 // recording where each register ended up:
545 if (GPRCS1Size > 0) {
546 MachineBasicBlock::iterator Pos = std::next(GPRCS1Push);
548 for (const auto &Entry : CSI) {
549 unsigned Reg = Entry.getReg();
550 int FI = Entry.getFrameIdx();
557 if (STI.isTargetDarwin())
569 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
570 nullptr, MRI->getDwarfRegNum(Reg, true), MFI->getObjectOffset(FI)));
571 BuildMI(MBB, Pos, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
572 .addCFIIndex(CFIIndex)
573 .setMIFlags(MachineInstr::FrameSetup);
579 if (GPRCS2Size > 0) {
580 MachineBasicBlock::iterator Pos = std::next(GPRCS2Push);
581 for (const auto &Entry : CSI) {
582 unsigned Reg = Entry.getReg();
583 int FI = Entry.getFrameIdx();
590 if (STI.isTargetDarwin()) {
591 unsigned DwarfReg = MRI->getDwarfRegNum(Reg, true);
592 unsigned Offset = MFI->getObjectOffset(FI);
593 unsigned CFIIndex = MMI.addFrameInst(
594 MCCFIInstruction::createOffset(nullptr, DwarfReg, Offset));
595 BuildMI(MBB, Pos, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
596 .addCFIIndex(CFIIndex)
597 .setMIFlags(MachineInstr::FrameSetup);
605 // Since vpush register list cannot have gaps, there may be multiple vpush
606 // instructions in the prologue.
607 MachineBasicBlock::iterator Pos = std::next(LastPush);
608 for (const auto &Entry : CSI) {
609 unsigned Reg = Entry.getReg();
610 int FI = Entry.getFrameIdx();
611 if ((Reg >= ARM::D0 && Reg <= ARM::D31) &&
612 (Reg < ARM::D8 || Reg >= ARM::D8 + AFI->getNumAlignedDPRCS2Regs())) {
613 unsigned DwarfReg = MRI->getDwarfRegNum(Reg, true);
614 unsigned Offset = MFI->getObjectOffset(FI);
615 unsigned CFIIndex = MMI.addFrameInst(
616 MCCFIInstruction::createOffset(nullptr, DwarfReg, Offset));
617 BuildMI(MBB, Pos, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
618 .addCFIIndex(CFIIndex)
619 .setMIFlags(MachineInstr::FrameSetup);
624 // Now we can emit descriptions of where the canonical frame address was
625 // throughout the process. If we have a frame pointer, it takes over the job
626 // half-way through, so only the first few .cfi_def_cfa_offset instructions
627 // actually get emitted.
628 DefCFAOffsetCandidates.emitDefCFAOffsets(MMI, MBB, dl, TII, HasFP);
630 if (STI.isTargetELF() && hasFP(MF))
631 MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() -
632 AFI->getFramePtrSpillOffset());
634 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
635 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
636 AFI->setDPRCalleeSavedGapSize(DPRGapSize);
637 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
639 // If we need dynamic stack realignment, do it here. Be paranoid and make
640 // sure if we also have VLAs, we have a base pointer for frame access.
641 // If aligned NEON registers were spilled, the stack has already been
643 if (!AFI->getNumAlignedDPRCS2Regs() && RegInfo->needsStackRealignment(MF)) {
644 unsigned MaxAlign = MFI->getMaxAlignment();
645 assert(!AFI->isThumb1OnlyFunction());
646 if (!AFI->isThumbFunction()) {
647 emitAligningInstructions(MF, AFI, TII, MBB, MBBI, dl, ARM::SP, MaxAlign,
650 // We cannot use sp as source/dest register here, thus we're using r4 to
651 // perform the calculations. We're emitting the following sequence:
653 // -- use emitAligningInstructions to produce best sequence to zero
654 // -- out lower bits in r4
656 // FIXME: It will be better just to find spare register here.
657 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), ARM::R4)
658 .addReg(ARM::SP, RegState::Kill));
659 emitAligningInstructions(MF, AFI, TII, MBB, MBBI, dl, ARM::R4, MaxAlign,
661 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), ARM::SP)
662 .addReg(ARM::R4, RegState::Kill));
665 AFI->setShouldRestoreSPFromFP(true);
668 // If we need a base pointer, set it up here. It's whatever the value
669 // of the stack pointer is at this point. Any variable size objects
670 // will be allocated after this, so we can still use the base pointer
671 // to reference locals.
672 // FIXME: Clarify FrameSetup flags here.
673 if (RegInfo->hasBasePointer(MF)) {
675 BuildMI(MBB, MBBI, dl,
676 TII.get(ARM::MOVr), RegInfo->getBaseRegister())
678 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
680 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr),
681 RegInfo->getBaseRegister())
685 // If the frame has variable sized objects then the epilogue must restore
686 // the sp from fp. We can assume there's an FP here since hasFP already
687 // checks for hasVarSizedObjects.
688 if (MFI->hasVarSizedObjects())
689 AFI->setShouldRestoreSPFromFP(true);
692 void ARMFrameLowering::emitEpilogue(MachineFunction &MF,
693 MachineBasicBlock &MBB) const {
694 MachineFrameInfo *MFI = MF.getFrameInfo();
695 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
696 const TargetRegisterInfo *RegInfo = MF.getSubtarget().getRegisterInfo();
697 const ARMBaseInstrInfo &TII =
698 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
699 assert(!AFI->isThumb1OnlyFunction() &&
700 "This emitEpilogue does not support Thumb1!");
701 bool isARM = !AFI->isThumbFunction();
703 unsigned ArgRegsSaveSize = AFI->getArgRegsSaveSize();
704 int NumBytes = (int)MFI->getStackSize();
705 unsigned FramePtr = RegInfo->getFrameRegister(MF);
707 // All calls are tail calls in GHC calling conv, and functions have no
708 // prologue/epilogue.
709 if (MF.getFunction()->getCallingConv() == CallingConv::GHC)
712 // First put ourselves on the first (from top) terminator instructions.
713 MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator();
714 DebugLoc dl = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
716 if (!AFI->hasStackFrame()) {
717 if (NumBytes - ArgRegsSaveSize != 0)
718 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes - ArgRegsSaveSize);
720 // Unwind MBBI to point to first LDR / VLDRD.
721 const MCPhysReg *CSRegs = RegInfo->getCalleeSavedRegs(&MF);
722 if (MBBI != MBB.begin()) {
725 } while (MBBI != MBB.begin() && isCSRestore(MBBI, TII, CSRegs));
726 if (!isCSRestore(MBBI, TII, CSRegs))
730 // Move SP to start of FP callee save spill area.
731 NumBytes -= (ArgRegsSaveSize +
732 AFI->getGPRCalleeSavedArea1Size() +
733 AFI->getGPRCalleeSavedArea2Size() +
734 AFI->getDPRCalleeSavedGapSize() +
735 AFI->getDPRCalleeSavedAreaSize());
737 // Reset SP based on frame pointer only if the stack frame extends beyond
738 // frame pointer stack slot or target is ELF and the function has FP.
739 if (AFI->shouldRestoreSPFromFP()) {
740 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
743 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, FramePtr, -NumBytes,
746 // It's not possible to restore SP from FP in a single instruction.
747 // For iOS, this looks like:
750 // This is bad, if an interrupt is taken after the mov, sp is in an
751 // inconsistent state.
752 // Use the first callee-saved register as a scratch register.
753 assert(!MFI->getPristineRegs(MF).test(ARM::R4) &&
754 "No scratch register to restore SP from FP!");
755 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::R4, FramePtr, -NumBytes,
757 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr),
764 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), ARM::SP)
765 .addReg(FramePtr).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
767 AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr),
771 } else if (NumBytes &&
772 !tryFoldSPUpdateIntoPushPop(STI, MF, MBBI, NumBytes))
773 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes);
775 // Increment past our save areas.
776 if (AFI->getDPRCalleeSavedAreaSize()) {
778 // Since vpop register list cannot have gaps, there may be multiple vpop
779 // instructions in the epilogue.
780 while (MBBI->getOpcode() == ARM::VLDMDIA_UPD)
783 if (AFI->getDPRCalleeSavedGapSize()) {
784 assert(AFI->getDPRCalleeSavedGapSize() == 4 &&
785 "unexpected DPR alignment gap");
786 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getDPRCalleeSavedGapSize());
789 if (AFI->getGPRCalleeSavedArea2Size()) MBBI++;
790 if (AFI->getGPRCalleeSavedArea1Size()) MBBI++;
794 emitSPUpdate(isARM, MBB, MBBI, dl, TII, ArgRegsSaveSize);
797 /// getFrameIndexReference - Provide a base+offset reference to an FI slot for
798 /// debug info. It's the same as what we use for resolving the code-gen
799 /// references for now. FIXME: This can go wrong when references are
800 /// SP-relative and simple call frames aren't used.
802 ARMFrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI,
803 unsigned &FrameReg) const {
804 return ResolveFrameIndexReference(MF, FI, FrameReg, 0);
808 ARMFrameLowering::ResolveFrameIndexReference(const MachineFunction &MF,
809 int FI, unsigned &FrameReg,
811 const MachineFrameInfo *MFI = MF.getFrameInfo();
812 const ARMBaseRegisterInfo *RegInfo = static_cast<const ARMBaseRegisterInfo *>(
813 MF.getSubtarget().getRegisterInfo());
814 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
815 int Offset = MFI->getObjectOffset(FI) + MFI->getStackSize();
816 int FPOffset = Offset - AFI->getFramePtrSpillOffset();
817 bool isFixed = MFI->isFixedObjectIndex(FI);
822 // SP can move around if there are allocas. We may also lose track of SP
823 // when emergency spilling inside a non-reserved call frame setup.
824 bool hasMovingSP = !hasReservedCallFrame(MF);
826 // When dynamically realigning the stack, use the frame pointer for
827 // parameters, and the stack/base pointer for locals.
828 if (RegInfo->needsStackRealignment(MF)) {
829 assert (hasFP(MF) && "dynamic stack realignment without a FP!");
831 FrameReg = RegInfo->getFrameRegister(MF);
833 } else if (hasMovingSP) {
834 assert(RegInfo->hasBasePointer(MF) &&
835 "VLAs and dynamic stack alignment, but missing base pointer!");
836 FrameReg = RegInfo->getBaseRegister();
841 // If there is a frame pointer, use it when we can.
842 if (hasFP(MF) && AFI->hasStackFrame()) {
843 // Use frame pointer to reference fixed objects. Use it for locals if
844 // there are VLAs (and thus the SP isn't reliable as a base).
845 if (isFixed || (hasMovingSP && !RegInfo->hasBasePointer(MF))) {
846 FrameReg = RegInfo->getFrameRegister(MF);
848 } else if (hasMovingSP) {
849 assert(RegInfo->hasBasePointer(MF) && "missing base pointer!");
850 if (AFI->isThumb2Function()) {
851 // Try to use the frame pointer if we can, else use the base pointer
852 // since it's available. This is handy for the emergency spill slot, in
854 if (FPOffset >= -255 && FPOffset < 0) {
855 FrameReg = RegInfo->getFrameRegister(MF);
859 } else if (AFI->isThumb2Function()) {
860 // Use add <rd>, sp, #<imm8>
861 // ldr <rd>, [sp, #<imm8>]
862 // if at all possible to save space.
863 if (Offset >= 0 && (Offset & 3) == 0 && Offset <= 1020)
865 // In Thumb2 mode, the negative offset is very limited. Try to avoid
866 // out of range references. ldr <rt>,[<rn>, #-<imm8>]
867 if (FPOffset >= -255 && FPOffset < 0) {
868 FrameReg = RegInfo->getFrameRegister(MF);
871 } else if (Offset > (FPOffset < 0 ? -FPOffset : FPOffset)) {
872 // Otherwise, use SP or FP, whichever is closer to the stack slot.
873 FrameReg = RegInfo->getFrameRegister(MF);
877 // Use the base pointer if we have one.
878 if (RegInfo->hasBasePointer(MF))
879 FrameReg = RegInfo->getBaseRegister();
883 int ARMFrameLowering::getFrameIndexOffset(const MachineFunction &MF,
886 return getFrameIndexReference(MF, FI, FrameReg);
889 void ARMFrameLowering::emitPushInst(MachineBasicBlock &MBB,
890 MachineBasicBlock::iterator MI,
891 const std::vector<CalleeSavedInfo> &CSI,
892 unsigned StmOpc, unsigned StrOpc,
894 bool(*Func)(unsigned, bool),
895 unsigned NumAlignedDPRCS2Regs,
896 unsigned MIFlags) const {
897 MachineFunction &MF = *MBB.getParent();
898 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
901 if (MI != MBB.end()) DL = MI->getDebugLoc();
903 SmallVector<std::pair<unsigned,bool>, 4> Regs;
904 unsigned i = CSI.size();
906 unsigned LastReg = 0;
907 for (; i != 0; --i) {
908 unsigned Reg = CSI[i-1].getReg();
909 if (!(Func)(Reg, STI.isTargetDarwin())) continue;
911 // D-registers in the aligned area DPRCS2 are NOT spilled here.
912 if (Reg >= ARM::D8 && Reg < ARM::D8 + NumAlignedDPRCS2Regs)
915 // Add the callee-saved register as live-in unless it's LR and
916 // @llvm.returnaddress is called. If LR is returned for
917 // @llvm.returnaddress then it's already added to the function and
918 // entry block live-in sets.
920 if (Reg == ARM::LR) {
921 if (MF.getFrameInfo()->isReturnAddressTaken() &&
922 MF.getRegInfo().isLiveIn(Reg))
929 // If NoGap is true, push consecutive registers and then leave the rest
930 // for other instructions. e.g.
931 // vpush {d8, d10, d11} -> vpush {d8}, vpush {d10, d11}
932 if (NoGap && LastReg && LastReg != Reg-1)
935 Regs.push_back(std::make_pair(Reg, isKill));
940 if (Regs.size() > 1 || StrOpc== 0) {
941 MachineInstrBuilder MIB =
942 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(StmOpc), ARM::SP)
943 .addReg(ARM::SP).setMIFlags(MIFlags));
944 for (unsigned i = 0, e = Regs.size(); i < e; ++i)
945 MIB.addReg(Regs[i].first, getKillRegState(Regs[i].second));
946 } else if (Regs.size() == 1) {
947 MachineInstrBuilder MIB = BuildMI(MBB, MI, DL, TII.get(StrOpc),
949 .addReg(Regs[0].first, getKillRegState(Regs[0].second))
950 .addReg(ARM::SP).setMIFlags(MIFlags)
956 // Put any subsequent vpush instructions before this one: they will refer to
957 // higher register numbers so need to be pushed first in order to preserve
959 if (MI != MBB.begin())
964 void ARMFrameLowering::emitPopInst(MachineBasicBlock &MBB,
965 MachineBasicBlock::iterator MI,
966 const std::vector<CalleeSavedInfo> &CSI,
967 unsigned LdmOpc, unsigned LdrOpc,
968 bool isVarArg, bool NoGap,
969 bool(*Func)(unsigned, bool),
970 unsigned NumAlignedDPRCS2Regs) const {
971 MachineFunction &MF = *MBB.getParent();
972 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
973 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
975 bool isTailCall = false;
976 bool isInterrupt = false;
977 if (MBB.end() != MI) {
978 DL = MI->getDebugLoc();
979 unsigned RetOpcode = MI->getOpcode();
980 isTailCall = (RetOpcode == ARM::TCRETURNdi || RetOpcode == ARM::TCRETURNri);
982 RetOpcode == ARM::SUBS_PC_LR || RetOpcode == ARM::t2SUBS_PC_LR;
985 SmallVector<unsigned, 4> Regs;
986 unsigned i = CSI.size();
988 unsigned LastReg = 0;
989 bool DeleteRet = false;
990 for (; i != 0; --i) {
991 unsigned Reg = CSI[i-1].getReg();
992 if (!(Func)(Reg, STI.isTargetDarwin())) continue;
994 // The aligned reloads from area DPRCS2 are not inserted here.
995 if (Reg >= ARM::D8 && Reg < ARM::D8 + NumAlignedDPRCS2Regs)
998 if (Reg == ARM::LR && !isTailCall && !isVarArg && !isInterrupt &&
1000 if (MBB.succ_empty()) {
1003 LdmOpc = AFI->isThumbFunction() ? ARM::t2LDMIA_RET : ARM::LDMIA_RET;
1005 LdmOpc = AFI->isThumbFunction() ? ARM::t2LDMIA_UPD : ARM::LDMIA_UPD;
1006 // Fold the return instruction into the LDM.
1009 // If NoGap is true, pop consecutive registers and then leave the rest
1010 // for other instructions. e.g.
1011 // vpop {d8, d10, d11} -> vpop {d8}, vpop {d10, d11}
1012 if (NoGap && LastReg && LastReg != Reg-1)
1016 Regs.push_back(Reg);
1021 if (Regs.size() > 1 || LdrOpc == 0) {
1022 MachineInstrBuilder MIB =
1023 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(LdmOpc), ARM::SP)
1025 for (unsigned i = 0, e = Regs.size(); i < e; ++i)
1026 MIB.addReg(Regs[i], getDefRegState(true));
1027 if (DeleteRet && MI != MBB.end()) {
1028 MIB.copyImplicitOps(&*MI);
1029 MI->eraseFromParent();
1032 } else if (Regs.size() == 1) {
1033 // If we adjusted the reg to PC from LR above, switch it back here. We
1034 // only do that for LDM.
1035 if (Regs[0] == ARM::PC)
1037 MachineInstrBuilder MIB =
1038 BuildMI(MBB, MI, DL, TII.get(LdrOpc), Regs[0])
1039 .addReg(ARM::SP, RegState::Define)
1041 // ARM mode needs an extra reg0 here due to addrmode2. Will go away once
1042 // that refactoring is complete (eventually).
1043 if (LdrOpc == ARM::LDR_POST_REG || LdrOpc == ARM::LDR_POST_IMM) {
1045 MIB.addImm(ARM_AM::getAM2Opc(ARM_AM::add, 4, ARM_AM::no_shift));
1048 AddDefaultPred(MIB);
1052 // Put any subsequent vpop instructions after this one: they will refer to
1053 // higher register numbers so need to be popped afterwards.
1054 if (MI != MBB.end())
1059 /// Emit aligned spill instructions for NumAlignedDPRCS2Regs D-registers
1060 /// starting from d8. Also insert stack realignment code and leave the stack
1061 /// pointer pointing to the d8 spill slot.
1062 static void emitAlignedDPRCS2Spills(MachineBasicBlock &MBB,
1063 MachineBasicBlock::iterator MI,
1064 unsigned NumAlignedDPRCS2Regs,
1065 const std::vector<CalleeSavedInfo> &CSI,
1066 const TargetRegisterInfo *TRI) {
1067 MachineFunction &MF = *MBB.getParent();
1068 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1069 DebugLoc DL = MI->getDebugLoc();
1070 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
1071 MachineFrameInfo &MFI = *MF.getFrameInfo();
1073 // Mark the D-register spill slots as properly aligned. Since MFI computes
1074 // stack slot layout backwards, this can actually mean that the d-reg stack
1075 // slot offsets can be wrong. The offset for d8 will always be correct.
1076 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1077 unsigned DNum = CSI[i].getReg() - ARM::D8;
1080 int FI = CSI[i].getFrameIdx();
1081 // The even-numbered registers will be 16-byte aligned, the odd-numbered
1082 // registers will be 8-byte aligned.
1083 MFI.setObjectAlignment(FI, DNum % 2 ? 8 : 16);
1085 // The stack slot for D8 needs to be maximally aligned because this is
1086 // actually the point where we align the stack pointer. MachineFrameInfo
1087 // computes all offsets relative to the incoming stack pointer which is a
1088 // bit weird when realigning the stack. Any extra padding for this
1089 // over-alignment is not realized because the code inserted below adjusts
1090 // the stack pointer by numregs * 8 before aligning the stack pointer.
1092 MFI.setObjectAlignment(FI, MFI.getMaxAlignment());
1095 // Move the stack pointer to the d8 spill slot, and align it at the same
1096 // time. Leave the stack slot address in the scratch register r4.
1098 // sub r4, sp, #numregs * 8
1099 // bic r4, r4, #align - 1
1102 bool isThumb = AFI->isThumbFunction();
1103 assert(!AFI->isThumb1OnlyFunction() && "Can't realign stack for thumb1");
1104 AFI->setShouldRestoreSPFromFP(true);
1106 // sub r4, sp, #numregs * 8
1107 // The immediate is <= 64, so it doesn't need any special encoding.
1108 unsigned Opc = isThumb ? ARM::t2SUBri : ARM::SUBri;
1109 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(Opc), ARM::R4)
1111 .addImm(8 * NumAlignedDPRCS2Regs)));
1113 unsigned MaxAlign = MF.getFrameInfo()->getMaxAlignment();
1114 // We must set parameter MustBeSingleInstruction to true, since
1115 // skipAlignedDPRCS2Spills expects exactly 3 instructions to perform
1116 // stack alignment. Luckily, this can always be done since all ARM
1117 // architecture versions that support Neon also support the BFC
1119 emitAligningInstructions(MF, AFI, TII, MBB, MI, DL, ARM::R4, MaxAlign, true);
1122 // The stack pointer must be adjusted before spilling anything, otherwise
1123 // the stack slots could be clobbered by an interrupt handler.
1124 // Leave r4 live, it is used below.
1125 Opc = isThumb ? ARM::tMOVr : ARM::MOVr;
1126 MachineInstrBuilder MIB = BuildMI(MBB, MI, DL, TII.get(Opc), ARM::SP)
1128 MIB = AddDefaultPred(MIB);
1132 // Now spill NumAlignedDPRCS2Regs registers starting from d8.
1133 // r4 holds the stack slot address.
1134 unsigned NextReg = ARM::D8;
1136 // 16-byte aligned vst1.64 with 4 d-regs and address writeback.
1137 // The writeback is only needed when emitting two vst1.64 instructions.
1138 if (NumAlignedDPRCS2Regs >= 6) {
1139 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1140 &ARM::QQPRRegClass);
1141 MBB.addLiveIn(SupReg);
1142 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VST1d64Qwb_fixed),
1144 .addReg(ARM::R4, RegState::Kill).addImm(16)
1146 .addReg(SupReg, RegState::ImplicitKill));
1148 NumAlignedDPRCS2Regs -= 4;
1151 // We won't modify r4 beyond this point. It currently points to the next
1152 // register to be spilled.
1153 unsigned R4BaseReg = NextReg;
1155 // 16-byte aligned vst1.64 with 4 d-regs, no writeback.
1156 if (NumAlignedDPRCS2Regs >= 4) {
1157 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1158 &ARM::QQPRRegClass);
1159 MBB.addLiveIn(SupReg);
1160 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VST1d64Q))
1161 .addReg(ARM::R4).addImm(16).addReg(NextReg)
1162 .addReg(SupReg, RegState::ImplicitKill));
1164 NumAlignedDPRCS2Regs -= 4;
1167 // 16-byte aligned vst1.64 with 2 d-regs.
1168 if (NumAlignedDPRCS2Regs >= 2) {
1169 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1171 MBB.addLiveIn(SupReg);
1172 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VST1q64))
1173 .addReg(ARM::R4).addImm(16).addReg(SupReg));
1175 NumAlignedDPRCS2Regs -= 2;
1178 // Finally, use a vanilla vstr.64 for the odd last register.
1179 if (NumAlignedDPRCS2Regs) {
1180 MBB.addLiveIn(NextReg);
1181 // vstr.64 uses addrmode5 which has an offset scale of 4.
1182 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VSTRD))
1184 .addReg(ARM::R4).addImm((NextReg-R4BaseReg)*2));
1187 // The last spill instruction inserted should kill the scratch register r4.
1188 std::prev(MI)->addRegisterKilled(ARM::R4, TRI);
1191 /// Skip past the code inserted by emitAlignedDPRCS2Spills, and return an
1192 /// iterator to the following instruction.
1193 static MachineBasicBlock::iterator
1194 skipAlignedDPRCS2Spills(MachineBasicBlock::iterator MI,
1195 unsigned NumAlignedDPRCS2Regs) {
1196 // sub r4, sp, #numregs * 8
1197 // bic r4, r4, #align - 1
1200 assert(MI->mayStore() && "Expecting spill instruction");
1202 // These switches all fall through.
1203 switch(NumAlignedDPRCS2Regs) {
1206 assert(MI->mayStore() && "Expecting spill instruction");
1209 assert(MI->mayStore() && "Expecting spill instruction");
1213 assert(MI->killsRegister(ARM::R4) && "Missed kill flag");
1219 /// Emit aligned reload instructions for NumAlignedDPRCS2Regs D-registers
1220 /// starting from d8. These instructions are assumed to execute while the
1221 /// stack is still aligned, unlike the code inserted by emitPopInst.
1222 static void emitAlignedDPRCS2Restores(MachineBasicBlock &MBB,
1223 MachineBasicBlock::iterator MI,
1224 unsigned NumAlignedDPRCS2Regs,
1225 const std::vector<CalleeSavedInfo> &CSI,
1226 const TargetRegisterInfo *TRI) {
1227 MachineFunction &MF = *MBB.getParent();
1228 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1229 DebugLoc DL = MI->getDebugLoc();
1230 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
1232 // Find the frame index assigned to d8.
1234 for (unsigned i = 0, e = CSI.size(); i != e; ++i)
1235 if (CSI[i].getReg() == ARM::D8) {
1236 D8SpillFI = CSI[i].getFrameIdx();
1240 // Materialize the address of the d8 spill slot into the scratch register r4.
1241 // This can be fairly complicated if the stack frame is large, so just use
1242 // the normal frame index elimination mechanism to do it. This code runs as
1243 // the initial part of the epilog where the stack and base pointers haven't
1244 // been changed yet.
1245 bool isThumb = AFI->isThumbFunction();
1246 assert(!AFI->isThumb1OnlyFunction() && "Can't realign stack for thumb1");
1248 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
1249 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(Opc), ARM::R4)
1250 .addFrameIndex(D8SpillFI).addImm(0)));
1252 // Now restore NumAlignedDPRCS2Regs registers starting from d8.
1253 unsigned NextReg = ARM::D8;
1255 // 16-byte aligned vld1.64 with 4 d-regs and writeback.
1256 if (NumAlignedDPRCS2Regs >= 6) {
1257 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1258 &ARM::QQPRRegClass);
1259 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLD1d64Qwb_fixed), NextReg)
1260 .addReg(ARM::R4, RegState::Define)
1261 .addReg(ARM::R4, RegState::Kill).addImm(16)
1262 .addReg(SupReg, RegState::ImplicitDefine));
1264 NumAlignedDPRCS2Regs -= 4;
1267 // We won't modify r4 beyond this point. It currently points to the next
1268 // register to be spilled.
1269 unsigned R4BaseReg = NextReg;
1271 // 16-byte aligned vld1.64 with 4 d-regs, no writeback.
1272 if (NumAlignedDPRCS2Regs >= 4) {
1273 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1274 &ARM::QQPRRegClass);
1275 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLD1d64Q), NextReg)
1276 .addReg(ARM::R4).addImm(16)
1277 .addReg(SupReg, RegState::ImplicitDefine));
1279 NumAlignedDPRCS2Regs -= 4;
1282 // 16-byte aligned vld1.64 with 2 d-regs.
1283 if (NumAlignedDPRCS2Regs >= 2) {
1284 unsigned SupReg = TRI->getMatchingSuperReg(NextReg, ARM::dsub_0,
1286 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLD1q64), SupReg)
1287 .addReg(ARM::R4).addImm(16));
1289 NumAlignedDPRCS2Regs -= 2;
1292 // Finally, use a vanilla vldr.64 for the remaining odd register.
1293 if (NumAlignedDPRCS2Regs)
1294 AddDefaultPred(BuildMI(MBB, MI, DL, TII.get(ARM::VLDRD), NextReg)
1295 .addReg(ARM::R4).addImm(2*(NextReg-R4BaseReg)));
1297 // Last store kills r4.
1298 std::prev(MI)->addRegisterKilled(ARM::R4, TRI);
1301 bool ARMFrameLowering::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
1302 MachineBasicBlock::iterator MI,
1303 const std::vector<CalleeSavedInfo> &CSI,
1304 const TargetRegisterInfo *TRI) const {
1308 MachineFunction &MF = *MBB.getParent();
1309 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1311 unsigned PushOpc = AFI->isThumbFunction() ? ARM::t2STMDB_UPD : ARM::STMDB_UPD;
1312 unsigned PushOneOpc = AFI->isThumbFunction() ?
1313 ARM::t2STR_PRE : ARM::STR_PRE_IMM;
1314 unsigned FltOpc = ARM::VSTMDDB_UPD;
1315 unsigned NumAlignedDPRCS2Regs = AFI->getNumAlignedDPRCS2Regs();
1316 emitPushInst(MBB, MI, CSI, PushOpc, PushOneOpc, false, &isARMArea1Register, 0,
1317 MachineInstr::FrameSetup);
1318 emitPushInst(MBB, MI, CSI, PushOpc, PushOneOpc, false, &isARMArea2Register, 0,
1319 MachineInstr::FrameSetup);
1320 emitPushInst(MBB, MI, CSI, FltOpc, 0, true, &isARMArea3Register,
1321 NumAlignedDPRCS2Regs, MachineInstr::FrameSetup);
1323 // The code above does not insert spill code for the aligned DPRCS2 registers.
1324 // The stack realignment code will be inserted between the push instructions
1325 // and these spills.
1326 if (NumAlignedDPRCS2Regs)
1327 emitAlignedDPRCS2Spills(MBB, MI, NumAlignedDPRCS2Regs, CSI, TRI);
1332 bool ARMFrameLowering::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1333 MachineBasicBlock::iterator MI,
1334 const std::vector<CalleeSavedInfo> &CSI,
1335 const TargetRegisterInfo *TRI) const {
1339 MachineFunction &MF = *MBB.getParent();
1340 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1341 bool isVarArg = AFI->getArgRegsSaveSize() > 0;
1342 unsigned NumAlignedDPRCS2Regs = AFI->getNumAlignedDPRCS2Regs();
1344 // The emitPopInst calls below do not insert reloads for the aligned DPRCS2
1345 // registers. Do that here instead.
1346 if (NumAlignedDPRCS2Regs)
1347 emitAlignedDPRCS2Restores(MBB, MI, NumAlignedDPRCS2Regs, CSI, TRI);
1349 unsigned PopOpc = AFI->isThumbFunction() ? ARM::t2LDMIA_UPD : ARM::LDMIA_UPD;
1350 unsigned LdrOpc = AFI->isThumbFunction() ? ARM::t2LDR_POST :ARM::LDR_POST_IMM;
1351 unsigned FltOpc = ARM::VLDMDIA_UPD;
1352 emitPopInst(MBB, MI, CSI, FltOpc, 0, isVarArg, true, &isARMArea3Register,
1353 NumAlignedDPRCS2Regs);
1354 emitPopInst(MBB, MI, CSI, PopOpc, LdrOpc, isVarArg, false,
1355 &isARMArea2Register, 0);
1356 emitPopInst(MBB, MI, CSI, PopOpc, LdrOpc, isVarArg, false,
1357 &isARMArea1Register, 0);
1362 // FIXME: Make generic?
1363 static unsigned GetFunctionSizeInBytes(const MachineFunction &MF,
1364 const ARMBaseInstrInfo &TII) {
1365 unsigned FnSize = 0;
1366 for (auto &MBB : MF) {
1367 for (auto &MI : MBB)
1368 FnSize += TII.GetInstSizeInBytes(&MI);
1373 /// estimateRSStackSizeLimit - Look at each instruction that references stack
1374 /// frames and return the stack size limit beyond which some of these
1375 /// instructions will require a scratch register during their expansion later.
1376 // FIXME: Move to TII?
1377 static unsigned estimateRSStackSizeLimit(MachineFunction &MF,
1378 const TargetFrameLowering *TFI) {
1379 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1380 unsigned Limit = (1 << 12) - 1;
1381 for (auto &MBB : MF) {
1382 for (auto &MI : MBB) {
1383 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1384 if (!MI.getOperand(i).isFI())
1387 // When using ADDri to get the address of a stack object, 255 is the
1388 // largest offset guaranteed to fit in the immediate offset.
1389 if (MI.getOpcode() == ARM::ADDri) {
1390 Limit = std::min(Limit, (1U << 8) - 1);
1394 // Otherwise check the addressing mode.
1395 switch (MI.getDesc().TSFlags & ARMII::AddrModeMask) {
1396 case ARMII::AddrMode3:
1397 case ARMII::AddrModeT2_i8:
1398 Limit = std::min(Limit, (1U << 8) - 1);
1400 case ARMII::AddrMode5:
1401 case ARMII::AddrModeT2_i8s4:
1402 Limit = std::min(Limit, ((1U << 8) - 1) * 4);
1404 case ARMII::AddrModeT2_i12:
1405 // i12 supports only positive offset so these will be converted to
1406 // i8 opcodes. See llvm::rewriteT2FrameIndex.
1407 if (TFI->hasFP(MF) && AFI->hasStackFrame())
1408 Limit = std::min(Limit, (1U << 8) - 1);
1410 case ARMII::AddrMode4:
1411 case ARMII::AddrMode6:
1412 // Addressing modes 4 & 6 (load/store) instructions can't encode an
1413 // immediate offset for stack references.
1418 break; // At most one FI per instruction
1426 // In functions that realign the stack, it can be an advantage to spill the
1427 // callee-saved vector registers after realigning the stack. The vst1 and vld1
1428 // instructions take alignment hints that can improve performance.
1431 checkNumAlignedDPRCS2Regs(MachineFunction &MF, BitVector &SavedRegs) {
1432 MF.getInfo<ARMFunctionInfo>()->setNumAlignedDPRCS2Regs(0);
1433 if (!SpillAlignedNEONRegs)
1436 // Naked functions don't spill callee-saved registers.
1437 if (MF.getFunction()->hasFnAttribute(Attribute::Naked))
1440 // We are planning to use NEON instructions vst1 / vld1.
1441 if (!static_cast<const ARMSubtarget &>(MF.getSubtarget()).hasNEON())
1444 // Don't bother if the default stack alignment is sufficiently high.
1445 if (MF.getSubtarget().getFrameLowering()->getStackAlignment() >= 8)
1448 // Aligned spills require stack realignment.
1449 if (!static_cast<const ARMBaseRegisterInfo *>(
1450 MF.getSubtarget().getRegisterInfo())->canRealignStack(MF))
1453 // We always spill contiguous d-registers starting from d8. Count how many
1454 // needs spilling. The register allocator will almost always use the
1455 // callee-saved registers in order, but it can happen that there are holes in
1456 // the range. Registers above the hole will be spilled to the standard DPRCS
1458 unsigned NumSpills = 0;
1459 for (; NumSpills < 8; ++NumSpills)
1460 if (!SavedRegs.test(ARM::D8 + NumSpills))
1463 // Don't do this for just one d-register. It's not worth it.
1467 // Spill the first NumSpills D-registers after realigning the stack.
1468 MF.getInfo<ARMFunctionInfo>()->setNumAlignedDPRCS2Regs(NumSpills);
1470 // A scratch register is required for the vst1 / vld1 instructions.
1471 SavedRegs.set(ARM::R4);
1474 void ARMFrameLowering::determineCalleeSaves(MachineFunction &MF,
1475 BitVector &SavedRegs,
1476 RegScavenger *RS) const {
1477 TargetFrameLowering::determineCalleeSaves(MF, SavedRegs, RS);
1478 // This tells PEI to spill the FP as if it is any other callee-save register
1479 // to take advantage the eliminateFrameIndex machinery. This also ensures it
1480 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
1481 // to combine multiple loads / stores.
1482 bool CanEliminateFrame = true;
1483 bool CS1Spilled = false;
1484 bool LRSpilled = false;
1485 unsigned NumGPRSpills = 0;
1486 SmallVector<unsigned, 4> UnspilledCS1GPRs;
1487 SmallVector<unsigned, 4> UnspilledCS2GPRs;
1488 const ARMBaseRegisterInfo *RegInfo = static_cast<const ARMBaseRegisterInfo *>(
1489 MF.getSubtarget().getRegisterInfo());
1490 const ARMBaseInstrInfo &TII =
1491 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
1492 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1493 MachineFrameInfo *MFI = MF.getFrameInfo();
1494 MachineRegisterInfo &MRI = MF.getRegInfo();
1495 unsigned FramePtr = RegInfo->getFrameRegister(MF);
1497 // Spill R4 if Thumb2 function requires stack realignment - it will be used as
1498 // scratch register. Also spill R4 if Thumb2 function has varsized objects,
1499 // since it's not always possible to restore sp from fp in a single
1501 // FIXME: It will be better just to find spare register here.
1502 if (AFI->isThumb2Function() &&
1503 (MFI->hasVarSizedObjects() || RegInfo->needsStackRealignment(MF)))
1504 SavedRegs.set(ARM::R4);
1506 if (AFI->isThumb1OnlyFunction()) {
1507 // Spill LR if Thumb1 function uses variable length argument lists.
1508 if (AFI->getArgRegsSaveSize() > 0)
1509 SavedRegs.set(ARM::LR);
1511 // Spill R4 if Thumb1 epilogue has to restore SP from FP. We don't know
1512 // for sure what the stack size will be, but for this, an estimate is good
1513 // enough. If there anything changes it, it'll be a spill, which implies
1514 // we've used all the registers and so R4 is already used, so not marking
1515 // it here will be OK.
1516 // FIXME: It will be better just to find spare register here.
1517 unsigned StackSize = MFI->estimateStackSize(MF);
1518 if (MFI->hasVarSizedObjects() || StackSize > 508)
1519 SavedRegs.set(ARM::R4);
1522 // See if we can spill vector registers to aligned stack.
1523 checkNumAlignedDPRCS2Regs(MF, SavedRegs);
1525 // Spill the BasePtr if it's used.
1526 if (RegInfo->hasBasePointer(MF))
1527 SavedRegs.set(RegInfo->getBaseRegister());
1529 // Don't spill FP if the frame can be eliminated. This is determined
1530 // by scanning the callee-save registers to see if any is modified.
1531 const MCPhysReg *CSRegs = RegInfo->getCalleeSavedRegs(&MF);
1532 for (unsigned i = 0; CSRegs[i]; ++i) {
1533 unsigned Reg = CSRegs[i];
1534 bool Spilled = false;
1535 if (SavedRegs.test(Reg)) {
1537 CanEliminateFrame = false;
1540 if (!ARM::GPRRegClass.contains(Reg))
1546 if (!STI.isTargetDarwin()) {
1553 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
1558 case ARM::R0: case ARM::R1:
1559 case ARM::R2: case ARM::R3:
1560 case ARM::R4: case ARM::R5:
1561 case ARM::R6: case ARM::R7:
1568 if (!STI.isTargetDarwin()) {
1569 UnspilledCS1GPRs.push_back(Reg);
1574 case ARM::R0: case ARM::R1:
1575 case ARM::R2: case ARM::R3:
1576 case ARM::R4: case ARM::R5:
1577 case ARM::R6: case ARM::R7:
1579 UnspilledCS1GPRs.push_back(Reg);
1582 UnspilledCS2GPRs.push_back(Reg);
1588 bool ForceLRSpill = false;
1589 if (!LRSpilled && AFI->isThumb1OnlyFunction()) {
1590 unsigned FnSize = GetFunctionSizeInBytes(MF, TII);
1591 // Force LR to be spilled if the Thumb function size is > 2048. This enables
1592 // use of BL to implement far jump. If it turns out that it's not needed
1593 // then the branch fix up path will undo it.
1594 if (FnSize >= (1 << 11)) {
1595 CanEliminateFrame = false;
1596 ForceLRSpill = true;
1600 // If any of the stack slot references may be out of range of an immediate
1601 // offset, make sure a register (or a spill slot) is available for the
1602 // register scavenger. Note that if we're indexing off the frame pointer, the
1603 // effective stack size is 4 bytes larger since the FP points to the stack
1604 // slot of the previous FP. Also, if we have variable sized objects in the
1605 // function, stack slot references will often be negative, and some of
1606 // our instructions are positive-offset only, so conservatively consider
1607 // that case to want a spill slot (or register) as well. Similarly, if
1608 // the function adjusts the stack pointer during execution and the
1609 // adjustments aren't already part of our stack size estimate, our offset
1610 // calculations may be off, so be conservative.
1611 // FIXME: We could add logic to be more precise about negative offsets
1612 // and which instructions will need a scratch register for them. Is it
1613 // worth the effort and added fragility?
1616 (MFI->estimateStackSize(MF) +
1617 ((hasFP(MF) && AFI->hasStackFrame()) ? 4:0) >=
1618 estimateRSStackSizeLimit(MF, this)))
1619 || MFI->hasVarSizedObjects()
1620 || (MFI->adjustsStack() && !canSimplifyCallFramePseudos(MF));
1622 bool ExtraCSSpill = false;
1623 if (BigStack || !CanEliminateFrame || RegInfo->cannotEliminateFrame(MF)) {
1624 AFI->setHasStackFrame(true);
1626 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
1627 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
1628 if (!LRSpilled && CS1Spilled) {
1629 SavedRegs.set(ARM::LR);
1631 SmallVectorImpl<unsigned>::iterator LRPos;
1632 LRPos = std::find(UnspilledCS1GPRs.begin(), UnspilledCS1GPRs.end(),
1634 if (LRPos != UnspilledCS1GPRs.end())
1635 UnspilledCS1GPRs.erase(LRPos);
1637 ForceLRSpill = false;
1638 ExtraCSSpill = true;
1642 SavedRegs.set(FramePtr);
1643 auto FPPos = std::find(UnspilledCS1GPRs.begin(), UnspilledCS1GPRs.end(),
1645 if (FPPos != UnspilledCS1GPRs.end())
1646 UnspilledCS1GPRs.erase(FPPos);
1650 // If stack and double are 8-byte aligned and we are spilling an odd number
1651 // of GPRs, spill one extra callee save GPR so we won't have to pad between
1652 // the integer and double callee save areas.
1653 unsigned TargetAlign = getStackAlignment();
1654 if (TargetAlign >= 8 && (NumGPRSpills & 1)) {
1655 if (CS1Spilled && !UnspilledCS1GPRs.empty()) {
1656 for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) {
1657 unsigned Reg = UnspilledCS1GPRs[i];
1658 // Don't spill high register if the function is thumb
1659 if (!AFI->isThumbFunction() ||
1660 isARMLowRegister(Reg) || Reg == ARM::LR) {
1662 if (!MRI.isReserved(Reg))
1663 ExtraCSSpill = true;
1667 } else if (!UnspilledCS2GPRs.empty() && !AFI->isThumb1OnlyFunction()) {
1668 unsigned Reg = UnspilledCS2GPRs.front();
1670 if (!MRI.isReserved(Reg))
1671 ExtraCSSpill = true;
1675 // Estimate if we might need to scavenge a register at some point in order
1676 // to materialize a stack offset. If so, either spill one additional
1677 // callee-saved register or reserve a special spill slot to facilitate
1678 // register scavenging. Thumb1 needs a spill slot for stack pointer
1679 // adjustments also, even when the frame itself is small.
1680 if (BigStack && !ExtraCSSpill) {
1681 // If any non-reserved CS register isn't spilled, just spill one or two
1682 // extra. That should take care of it!
1683 unsigned NumExtras = TargetAlign / 4;
1684 SmallVector<unsigned, 2> Extras;
1685 while (NumExtras && !UnspilledCS1GPRs.empty()) {
1686 unsigned Reg = UnspilledCS1GPRs.back();
1687 UnspilledCS1GPRs.pop_back();
1688 if (!MRI.isReserved(Reg) &&
1689 (!AFI->isThumb1OnlyFunction() || isARMLowRegister(Reg) ||
1691 Extras.push_back(Reg);
1695 // For non-Thumb1 functions, also check for hi-reg CS registers
1696 if (!AFI->isThumb1OnlyFunction()) {
1697 while (NumExtras && !UnspilledCS2GPRs.empty()) {
1698 unsigned Reg = UnspilledCS2GPRs.back();
1699 UnspilledCS2GPRs.pop_back();
1700 if (!MRI.isReserved(Reg)) {
1701 Extras.push_back(Reg);
1706 if (Extras.size() && NumExtras == 0) {
1707 for (unsigned i = 0, e = Extras.size(); i != e; ++i) {
1708 SavedRegs.set(Extras[i]);
1710 } else if (!AFI->isThumb1OnlyFunction()) {
1711 // note: Thumb1 functions spill to R12, not the stack. Reserve a slot
1712 // closest to SP or frame pointer.
1713 const TargetRegisterClass *RC = &ARM::GPRRegClass;
1714 RS->addScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
1722 SavedRegs.set(ARM::LR);
1723 AFI->setLRIsSpilledForFarJump(true);
1728 void ARMFrameLowering::
1729 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
1730 MachineBasicBlock::iterator I) const {
1731 const ARMBaseInstrInfo &TII =
1732 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
1733 if (!hasReservedCallFrame(MF)) {
1734 // If we have alloca, convert as follows:
1735 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
1736 // ADJCALLSTACKUP -> add, sp, sp, amount
1737 MachineInstr *Old = I;
1738 DebugLoc dl = Old->getDebugLoc();
1739 unsigned Amount = Old->getOperand(0).getImm();
1741 // We need to keep the stack aligned properly. To do this, we round the
1742 // amount of space needed for the outgoing arguments up to the next
1743 // alignment boundary.
1744 unsigned Align = getStackAlignment();
1745 Amount = (Amount+Align-1)/Align*Align;
1747 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1748 assert(!AFI->isThumb1OnlyFunction() &&
1749 "This eliminateCallFramePseudoInstr does not support Thumb1!");
1750 bool isARM = !AFI->isThumbFunction();
1752 // Replace the pseudo instruction with a new instruction...
1753 unsigned Opc = Old->getOpcode();
1754 int PIdx = Old->findFirstPredOperandIdx();
1755 ARMCC::CondCodes Pred = (PIdx == -1)
1756 ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(PIdx).getImm();
1757 if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) {
1758 // Note: PredReg is operand 2 for ADJCALLSTACKDOWN.
1759 unsigned PredReg = Old->getOperand(2).getReg();
1760 emitSPUpdate(isARM, MBB, I, dl, TII, -Amount, MachineInstr::NoFlags,
1763 // Note: PredReg is operand 3 for ADJCALLSTACKUP.
1764 unsigned PredReg = Old->getOperand(3).getReg();
1765 assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP);
1766 emitSPUpdate(isARM, MBB, I, dl, TII, Amount, MachineInstr::NoFlags,
1774 /// Get the minimum constant for ARM that is greater than or equal to the
1775 /// argument. In ARM, constants can have any value that can be produced by
1776 /// rotating an 8-bit value to the right by an even number of bits within a
1778 static uint32_t alignToARMConstant(uint32_t Value) {
1779 unsigned Shifted = 0;
1784 while (!(Value & 0xC0000000)) {
1789 bool Carry = (Value & 0x00FFFFFF);
1790 Value = ((Value & 0xFF000000) >> 24) + Carry;
1792 if (Value & 0x0000100)
1793 Value = Value & 0x000001FC;
1796 Value = Value >> (Shifted - 24);
1798 Value = Value << (24 - Shifted);
1803 // The stack limit in the TCB is set to this many bytes above the actual
1805 static const uint64_t kSplitStackAvailable = 256;
1807 // Adjust the function prologue to enable split stacks. This currently only
1808 // supports android and linux.
1810 // The ABI of the segmented stack prologue is a little arbitrarily chosen, but
1811 // must be well defined in order to allow for consistent implementations of the
1812 // __morestack helper function. The ABI is also not a normal ABI in that it
1813 // doesn't follow the normal calling conventions because this allows the
1814 // prologue of each function to be optimized further.
1816 // Currently, the ABI looks like (when calling __morestack)
1818 // * r4 holds the minimum stack size requested for this function call
1819 // * r5 holds the stack size of the arguments to the function
1820 // * the beginning of the function is 3 instructions after the call to
1823 // Implementations of __morestack should use r4 to allocate a new stack, r5 to
1824 // place the arguments on to the new stack, and the 3-instruction knowledge to
1825 // jump directly to the body of the function when working on the new stack.
1827 // An old (and possibly no longer compatible) implementation of __morestack for
1828 // ARM can be found at [1].
1830 // [1] - https://github.com/mozilla/rust/blob/86efd9/src/rt/arch/arm/morestack.S
1831 void ARMFrameLowering::adjustForSegmentedStacks(
1832 MachineFunction &MF, MachineBasicBlock &PrologueMBB) const {
1835 const ARMSubtarget *ST = &MF.getSubtarget<ARMSubtarget>();
1836 bool Thumb = ST->isThumb();
1838 // Sadly, this currently doesn't support varargs, platforms other than
1839 // android/linux. Note that thumb1/thumb2 are support for android/linux.
1840 if (MF.getFunction()->isVarArg())
1841 report_fatal_error("Segmented stacks do not support vararg functions.");
1842 if (!ST->isTargetAndroid() && !ST->isTargetLinux())
1843 report_fatal_error("Segmented stacks not supported on this platform.");
1845 assert(&PrologueMBB == &MF.front() && "Shrink-wrapping not yet implemented");
1846 MachineFrameInfo *MFI = MF.getFrameInfo();
1847 MachineModuleInfo &MMI = MF.getMMI();
1848 MCContext &Context = MMI.getContext();
1849 const MCRegisterInfo *MRI = Context.getRegisterInfo();
1850 const ARMBaseInstrInfo &TII =
1851 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
1852 ARMFunctionInfo *ARMFI = MF.getInfo<ARMFunctionInfo>();
1855 uint64_t StackSize = MFI->getStackSize();
1857 // Do not generate a prologue for functions with a stack of size zero
1861 // Use R4 and R5 as scratch registers.
1862 // We save R4 and R5 before use and restore them before leaving the function.
1863 unsigned ScratchReg0 = ARM::R4;
1864 unsigned ScratchReg1 = ARM::R5;
1865 uint64_t AlignedStackSize;
1867 MachineBasicBlock *PrevStackMBB = MF.CreateMachineBasicBlock();
1868 MachineBasicBlock *PostStackMBB = MF.CreateMachineBasicBlock();
1869 MachineBasicBlock *AllocMBB = MF.CreateMachineBasicBlock();
1870 MachineBasicBlock *GetMBB = MF.CreateMachineBasicBlock();
1871 MachineBasicBlock *McrMBB = MF.CreateMachineBasicBlock();
1873 // Grab everything that reaches PrologueMBB to update there liveness as well.
1874 SmallPtrSet<MachineBasicBlock *, 8> BeforePrologueRegion;
1875 SmallVector<MachineBasicBlock *, 2> WalkList;
1876 WalkList.push_back(&PrologueMBB);
1879 MachineBasicBlock *CurMBB = WalkList.pop_back_val();
1880 for (MachineBasicBlock *PredBB : CurMBB->predecessors()) {
1881 if (BeforePrologueRegion.insert(PredBB).second)
1882 WalkList.push_back(PredBB);
1884 } while (!WalkList.empty());
1886 // The order in that list is important.
1887 // The blocks will all be inserted before PrologueMBB using that order.
1888 // Therefore the block that should appear first in the CFG should appear
1889 // first in the list.
1890 MachineBasicBlock *AddedBlocks[] = {PrevStackMBB, McrMBB, GetMBB, AllocMBB,
1892 const int NbAddedBlocks = sizeof(AddedBlocks) / sizeof(AddedBlocks[0]);
1894 for (int Idx = 0; Idx < NbAddedBlocks; ++Idx)
1895 BeforePrologueRegion.insert(AddedBlocks[Idx]);
1897 for (MachineBasicBlock::livein_iterator i = PrologueMBB.livein_begin(),
1898 e = PrologueMBB.livein_end();
1900 for (MachineBasicBlock *PredBB : BeforePrologueRegion)
1901 PredBB->addLiveIn(*i);
1904 // Remove the newly added blocks from the list, since we know
1905 // we do not have to do the following updates for them.
1906 for (int Idx = 0; Idx < NbAddedBlocks; ++Idx) {
1907 BeforePrologueRegion.erase(AddedBlocks[Idx]);
1908 MF.insert(&PrologueMBB, AddedBlocks[Idx]);
1911 for (MachineBasicBlock *MBB : BeforePrologueRegion) {
1912 // Make sure the LiveIns are still sorted and unique.
1913 MBB->sortUniqueLiveIns();
1914 // Replace the edges to PrologueMBB by edges to the sequences
1915 // we are about to add.
1916 MBB->ReplaceUsesOfBlockWith(&PrologueMBB, AddedBlocks[0]);
1919 // The required stack size that is aligned to ARM constant criterion.
1920 AlignedStackSize = alignToARMConstant(StackSize);
1922 // When the frame size is less than 256 we just compare the stack
1923 // boundary directly to the value of the stack pointer, per gcc.
1924 bool CompareStackPointer = AlignedStackSize < kSplitStackAvailable;
1926 // We will use two of the callee save registers as scratch registers so we
1927 // need to save those registers onto the stack.
1928 // We will use SR0 to hold stack limit and SR1 to hold the stack size
1929 // requested and arguments for __morestack().
1930 // SR0: Scratch Register #0
1931 // SR1: Scratch Register #1
1934 AddDefaultPred(BuildMI(PrevStackMBB, DL, TII.get(ARM::tPUSH)))
1935 .addReg(ScratchReg0).addReg(ScratchReg1);
1937 AddDefaultPred(BuildMI(PrevStackMBB, DL, TII.get(ARM::STMDB_UPD))
1938 .addReg(ARM::SP, RegState::Define).addReg(ARM::SP))
1939 .addReg(ScratchReg0).addReg(ScratchReg1);
1942 // Emit the relevant DWARF information about the change in stack pointer as
1943 // well as where to find both r4 and r5 (the callee-save registers)
1945 MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, -8));
1946 BuildMI(PrevStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1947 .addCFIIndex(CFIIndex);
1948 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
1949 nullptr, MRI->getDwarfRegNum(ScratchReg1, true), -4));
1950 BuildMI(PrevStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1951 .addCFIIndex(CFIIndex);
1952 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
1953 nullptr, MRI->getDwarfRegNum(ScratchReg0, true), -8));
1954 BuildMI(PrevStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
1955 .addCFIIndex(CFIIndex);
1959 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::tMOVr), ScratchReg1)
1961 } else if (CompareStackPointer) {
1962 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::MOVr), ScratchReg1)
1963 .addReg(ARM::SP)).addReg(0);
1966 // sub SR1, sp, #StackSize
1967 if (!CompareStackPointer && Thumb) {
1969 AddDefaultCC(BuildMI(McrMBB, DL, TII.get(ARM::tSUBi8), ScratchReg1))
1970 .addReg(ScratchReg1).addImm(AlignedStackSize));
1971 } else if (!CompareStackPointer) {
1972 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::SUBri), ScratchReg1)
1973 .addReg(ARM::SP).addImm(AlignedStackSize)).addReg(0);
1976 if (Thumb && ST->isThumb1Only()) {
1977 unsigned PCLabelId = ARMFI->createPICLabelUId();
1978 ARMConstantPoolValue *NewCPV = ARMConstantPoolSymbol::Create(
1979 MF.getFunction()->getContext(), "__STACK_LIMIT", PCLabelId, 0);
1980 MachineConstantPool *MCP = MF.getConstantPool();
1981 unsigned CPI = MCP->getConstantPoolIndex(NewCPV, MF.getAlignment());
1983 // ldr SR0, [pc, offset(STACK_LIMIT)]
1984 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(ARM::tLDRpci), ScratchReg0)
1985 .addConstantPoolIndex(CPI));
1988 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(ARM::tLDRi), ScratchReg0)
1989 .addReg(ScratchReg0).addImm(0));
1991 // Get TLS base address from the coprocessor
1992 // mrc p15, #0, SR0, c13, c0, #3
1993 AddDefaultPred(BuildMI(McrMBB, DL, TII.get(ARM::MRC), ScratchReg0)
2000 // Use the last tls slot on android and a private field of the TCP on linux.
2001 assert(ST->isTargetAndroid() || ST->isTargetLinux());
2002 unsigned TlsOffset = ST->isTargetAndroid() ? 63 : 1;
2004 // Get the stack limit from the right offset
2005 // ldr SR0, [sr0, #4 * TlsOffset]
2006 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(ARM::LDRi12), ScratchReg0)
2007 .addReg(ScratchReg0).addImm(4 * TlsOffset));
2010 // Compare stack limit with stack size requested.
2012 Opcode = Thumb ? ARM::tCMPr : ARM::CMPrr;
2013 AddDefaultPred(BuildMI(GetMBB, DL, TII.get(Opcode))
2014 .addReg(ScratchReg0)
2015 .addReg(ScratchReg1));
2017 // This jump is taken if StackLimit < SP - stack required.
2018 Opcode = Thumb ? ARM::tBcc : ARM::Bcc;
2019 BuildMI(GetMBB, DL, TII.get(Opcode)).addMBB(PostStackMBB)
2024 // Calling __morestack(StackSize, Size of stack arguments).
2025 // __morestack knows that the stack size requested is in SR0(r4)
2026 // and amount size of stack arguments is in SR1(r5).
2028 // Pass first argument for the __morestack by Scratch Register #0.
2029 // The amount size of stack required
2031 AddDefaultPred(AddDefaultCC(BuildMI(AllocMBB, DL, TII.get(ARM::tMOVi8),
2032 ScratchReg0)).addImm(AlignedStackSize));
2034 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::MOVi), ScratchReg0)
2035 .addImm(AlignedStackSize)).addReg(0);
2037 // Pass second argument for the __morestack by Scratch Register #1.
2038 // The amount size of stack consumed to save function arguments.
2041 AddDefaultCC(BuildMI(AllocMBB, DL, TII.get(ARM::tMOVi8), ScratchReg1))
2042 .addImm(alignToARMConstant(ARMFI->getArgumentStackSize())));
2044 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::MOVi), ScratchReg1)
2045 .addImm(alignToARMConstant(ARMFI->getArgumentStackSize())))
2049 // push {lr} - Save return address of this function.
2051 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tPUSH)))
2054 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::STMDB_UPD))
2055 .addReg(ARM::SP, RegState::Define)
2060 // Emit the DWARF info about the change in stack as well as where to find the
2061 // previous link register
2063 MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, -12));
2064 BuildMI(AllocMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2065 .addCFIIndex(CFIIndex);
2066 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset(
2067 nullptr, MRI->getDwarfRegNum(ARM::LR, true), -12));
2068 BuildMI(AllocMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2069 .addCFIIndex(CFIIndex);
2071 // Call __morestack().
2073 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tBL)))
2074 .addExternalSymbol("__morestack");
2076 BuildMI(AllocMBB, DL, TII.get(ARM::BL))
2077 .addExternalSymbol("__morestack");
2080 // pop {lr} - Restore return address of this original function.
2082 if (ST->isThumb1Only()) {
2083 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tPOP)))
2084 .addReg(ScratchReg0);
2085 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tMOVr), ARM::LR)
2086 .addReg(ScratchReg0));
2088 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::t2LDR_POST))
2089 .addReg(ARM::LR, RegState::Define)
2090 .addReg(ARM::SP, RegState::Define)
2095 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::LDMIA_UPD))
2096 .addReg(ARM::SP, RegState::Define)
2101 // Restore SR0 and SR1 in case of __morestack() was called.
2102 // __morestack() will skip PostStackMBB block so we need to restore
2103 // scratch registers from here.
2106 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::tPOP)))
2107 .addReg(ScratchReg0)
2108 .addReg(ScratchReg1);
2110 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(ARM::LDMIA_UPD))
2111 .addReg(ARM::SP, RegState::Define)
2113 .addReg(ScratchReg0)
2114 .addReg(ScratchReg1);
2117 // Update the CFA offset now that we've popped
2118 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, 0));
2119 BuildMI(AllocMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2120 .addCFIIndex(CFIIndex);
2122 // bx lr - Return from this function.
2123 Opcode = Thumb ? ARM::tBX_RET : ARM::BX_RET;
2124 AddDefaultPred(BuildMI(AllocMBB, DL, TII.get(Opcode)));
2126 // Restore SR0 and SR1 in case of __morestack() was not called.
2129 AddDefaultPred(BuildMI(PostStackMBB, DL, TII.get(ARM::tPOP)))
2130 .addReg(ScratchReg0)
2131 .addReg(ScratchReg1);
2133 AddDefaultPred(BuildMI(PostStackMBB, DL, TII.get(ARM::LDMIA_UPD))
2134 .addReg(ARM::SP, RegState::Define)
2136 .addReg(ScratchReg0)
2137 .addReg(ScratchReg1);
2140 // Update the CFA offset now that we've popped
2141 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createDefCfaOffset(nullptr, 0));
2142 BuildMI(PostStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2143 .addCFIIndex(CFIIndex);
2145 // Tell debuggers that r4 and r5 are now the same as they were in the
2146 // previous function, that they're the "Same Value".
2147 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createSameValue(
2148 nullptr, MRI->getDwarfRegNum(ScratchReg0, true)));
2149 BuildMI(PostStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2150 .addCFIIndex(CFIIndex);
2151 CFIIndex = MMI.addFrameInst(MCCFIInstruction::createSameValue(
2152 nullptr, MRI->getDwarfRegNum(ScratchReg1, true)));
2153 BuildMI(PostStackMBB, DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
2154 .addCFIIndex(CFIIndex);
2156 // Organizing MBB lists
2157 PostStackMBB->addSuccessor(&PrologueMBB);
2159 AllocMBB->addSuccessor(PostStackMBB);
2161 GetMBB->addSuccessor(PostStackMBB);
2162 GetMBB->addSuccessor(AllocMBB);
2164 McrMBB->addSuccessor(GetMBB);
2166 PrevStackMBB->addSuccessor(McrMBB);