1 //===-- ARMBaseRegisterInfo.cpp - ARM Register 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 base ARM implementation of TargetRegisterInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "ARMBaseRegisterInfo.h"
16 #include "ARMBaseInstrInfo.h"
17 #include "ARMFrameLowering.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMSubtarget.h"
20 #include "MCTargetDesc/ARMAddressingModes.h"
21 #include "MCTargetDesc/ARMBaseInfo.h"
22 #include "llvm/ADT/BitVector.h"
23 #include "llvm/ADT/STLExtras.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/CodeGen/MachineBasicBlock.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstr.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/MachineOperand.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/RegisterScavenging.h"
34 #include "llvm/CodeGen/TargetInstrInfo.h"
35 #include "llvm/CodeGen/TargetRegisterInfo.h"
36 #include "llvm/CodeGen/VirtRegMap.h"
37 #include "llvm/IR/Attributes.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DebugLoc.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/MC/MCInstrDesc.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Target/TargetMachine.h"
47 #include "llvm/Target/TargetOptions.h"
51 #define DEBUG_TYPE "arm-register-info"
53 #define GET_REGINFO_TARGET_DESC
54 #include "ARMGenRegisterInfo.inc"
58 ARMBaseRegisterInfo::ARMBaseRegisterInfo()
59 : ARMGenRegisterInfo(ARM::LR, 0, 0, ARM::PC) {}
61 static unsigned getFramePointerReg(const ARMSubtarget &STI) {
62 return STI.useR7AsFramePointer() ? ARM::R7 : ARM::R11;
66 ARMBaseRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
67 const ARMSubtarget &STI = MF->getSubtarget<ARMSubtarget>();
68 bool UseSplitPush = STI.splitFramePushPop(*MF);
69 const MCPhysReg *RegList =
72 : (UseSplitPush ? CSR_AAPCS_SplitPush_SaveList : CSR_AAPCS_SaveList);
74 const Function &F = MF->getFunction();
75 if (F.getCallingConv() == CallingConv::GHC) {
76 // GHC set of callee saved regs is empty as all those regs are
77 // used for passing STG regs around
78 return CSR_NoRegs_SaveList;
79 } else if (F.hasFnAttribute("interrupt")) {
81 // M-class CPUs have hardware which saves the registers needed to allow a
82 // function conforming to the AAPCS to function as a handler.
83 return UseSplitPush ? CSR_AAPCS_SplitPush_SaveList : CSR_AAPCS_SaveList;
84 } else if (F.getFnAttribute("interrupt").getValueAsString() == "FIQ") {
85 // Fast interrupt mode gives the handler a private copy of R8-R14, so less
86 // need to be saved to restore user-mode state.
87 return CSR_FIQ_SaveList;
89 // Generally only R13-R14 (i.e. SP, LR) are automatically preserved by
90 // exception handling.
91 return CSR_GenericInt_SaveList;
95 if (STI.getTargetLowering()->supportSwiftError() &&
96 F.getAttributes().hasAttrSomewhere(Attribute::SwiftError)) {
97 if (STI.isTargetDarwin())
98 return CSR_iOS_SwiftError_SaveList;
100 return UseSplitPush ? CSR_AAPCS_SplitPush_SwiftError_SaveList :
101 CSR_AAPCS_SwiftError_SaveList;
104 if (STI.isTargetDarwin() && F.getCallingConv() == CallingConv::CXX_FAST_TLS)
105 return MF->getInfo<ARMFunctionInfo>()->isSplitCSR()
106 ? CSR_iOS_CXX_TLS_PE_SaveList
107 : CSR_iOS_CXX_TLS_SaveList;
111 const MCPhysReg *ARMBaseRegisterInfo::getCalleeSavedRegsViaCopy(
112 const MachineFunction *MF) const {
113 assert(MF && "Invalid MachineFunction pointer.");
114 if (MF->getFunction().getCallingConv() == CallingConv::CXX_FAST_TLS &&
115 MF->getInfo<ARMFunctionInfo>()->isSplitCSR())
116 return CSR_iOS_CXX_TLS_ViaCopy_SaveList;
121 ARMBaseRegisterInfo::getCallPreservedMask(const MachineFunction &MF,
122 CallingConv::ID CC) const {
123 const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();
124 if (CC == CallingConv::GHC)
125 // This is academic because all GHC calls are (supposed to be) tail calls
126 return CSR_NoRegs_RegMask;
128 if (STI.getTargetLowering()->supportSwiftError() &&
129 MF.getFunction().getAttributes().hasAttrSomewhere(Attribute::SwiftError))
130 return STI.isTargetDarwin() ? CSR_iOS_SwiftError_RegMask
131 : CSR_AAPCS_SwiftError_RegMask;
133 if (STI.isTargetDarwin() && CC == CallingConv::CXX_FAST_TLS)
134 return CSR_iOS_CXX_TLS_RegMask;
135 return STI.isTargetDarwin() ? CSR_iOS_RegMask : CSR_AAPCS_RegMask;
139 ARMBaseRegisterInfo::getNoPreservedMask() const {
140 return CSR_NoRegs_RegMask;
144 ARMBaseRegisterInfo::getTLSCallPreservedMask(const MachineFunction &MF) const {
145 assert(MF.getSubtarget<ARMSubtarget>().isTargetDarwin() &&
146 "only know about special TLS call on Darwin");
147 return CSR_iOS_TLSCall_RegMask;
151 ARMBaseRegisterInfo::getSjLjDispatchPreservedMask(const MachineFunction &MF) const {
152 const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();
153 if (!STI.useSoftFloat() && STI.hasVFP2() && !STI.isThumb1Only())
154 return CSR_NoRegs_RegMask;
156 return CSR_FPRegs_RegMask;
160 ARMBaseRegisterInfo::getThisReturnPreservedMask(const MachineFunction &MF,
161 CallingConv::ID CC) const {
162 const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();
163 // This should return a register mask that is the same as that returned by
164 // getCallPreservedMask but that additionally preserves the register used for
165 // the first i32 argument (which must also be the register used to return a
166 // single i32 return value)
168 // In case that the calling convention does not use the same register for
169 // both or otherwise does not want to enable this optimization, the function
170 // should return NULL
171 if (CC == CallingConv::GHC)
172 // This is academic because all GHC calls are (supposed to be) tail calls
174 return STI.isTargetDarwin() ? CSR_iOS_ThisReturn_RegMask
175 : CSR_AAPCS_ThisReturn_RegMask;
178 BitVector ARMBaseRegisterInfo::
179 getReservedRegs(const MachineFunction &MF) const {
180 const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();
181 const ARMFrameLowering *TFI = getFrameLowering(MF);
183 // FIXME: avoid re-calculating this every time.
184 BitVector Reserved(getNumRegs());
185 markSuperRegs(Reserved, ARM::SP);
186 markSuperRegs(Reserved, ARM::PC);
187 markSuperRegs(Reserved, ARM::FPSCR);
188 markSuperRegs(Reserved, ARM::APSR_NZCV);
190 markSuperRegs(Reserved, getFramePointerReg(STI));
191 if (hasBasePointer(MF))
192 markSuperRegs(Reserved, BasePtr);
193 // Some targets reserve R9.
194 if (STI.isR9Reserved())
195 markSuperRegs(Reserved, ARM::R9);
196 // Reserve D16-D31 if the subtarget doesn't support them.
197 if (!STI.hasVFP3() || STI.hasD16()) {
198 static_assert(ARM::D31 == ARM::D16 + 15, "Register list not consecutive!");
199 for (unsigned R = 0; R < 16; ++R)
200 markSuperRegs(Reserved, ARM::D16 + R);
202 const TargetRegisterClass &RC = ARM::GPRPairRegClass;
203 for (unsigned Reg : RC)
204 for (MCSubRegIterator SI(Reg, this); SI.isValid(); ++SI)
205 if (Reserved.test(*SI))
206 markSuperRegs(Reserved, Reg);
208 assert(checkAllSuperRegsMarked(Reserved));
212 bool ARMBaseRegisterInfo::
213 isAsmClobberable(const MachineFunction &MF, unsigned PhysReg) const {
214 return !getReservedRegs(MF).test(PhysReg);
217 const TargetRegisterClass *
218 ARMBaseRegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC,
219 const MachineFunction &) const {
220 const TargetRegisterClass *Super = RC;
221 TargetRegisterClass::sc_iterator I = RC->getSuperClasses();
223 switch (Super->getID()) {
224 case ARM::GPRRegClassID:
225 case ARM::SPRRegClassID:
226 case ARM::DPRRegClassID:
227 case ARM::QPRRegClassID:
228 case ARM::QQPRRegClassID:
229 case ARM::QQQQPRRegClassID:
230 case ARM::GPRPairRegClassID:
238 const TargetRegisterClass *
239 ARMBaseRegisterInfo::getPointerRegClass(const MachineFunction &MF, unsigned Kind)
241 return &ARM::GPRRegClass;
244 const TargetRegisterClass *
245 ARMBaseRegisterInfo::getCrossCopyRegClass(const TargetRegisterClass *RC) const {
246 if (RC == &ARM::CCRRegClass)
247 return &ARM::rGPRRegClass; // Can't copy CCR registers.
252 ARMBaseRegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC,
253 MachineFunction &MF) const {
254 const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();
255 const ARMFrameLowering *TFI = getFrameLowering(MF);
257 switch (RC->getID()) {
260 case ARM::tGPRRegClassID: {
261 // hasFP ends up calling getMaxCallFrameComputed() which may not be
262 // available when getPressureLimit() is called as part of
263 // ScheduleDAGRRList.
264 bool HasFP = MF.getFrameInfo().isMaxCallFrameSizeComputed()
265 ? TFI->hasFP(MF) : true;
268 case ARM::GPRRegClassID: {
269 bool HasFP = MF.getFrameInfo().isMaxCallFrameSizeComputed()
270 ? TFI->hasFP(MF) : true;
271 return 10 - HasFP - (STI.isR9Reserved() ? 1 : 0);
273 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
274 case ARM::DPRRegClassID:
279 // Get the other register in a GPRPair.
280 static unsigned getPairedGPR(unsigned Reg, bool Odd, const MCRegisterInfo *RI) {
281 for (MCSuperRegIterator Supers(Reg, RI); Supers.isValid(); ++Supers)
282 if (ARM::GPRPairRegClass.contains(*Supers))
283 return RI->getSubReg(*Supers, Odd ? ARM::gsub_1 : ARM::gsub_0);
287 // Resolve the RegPairEven / RegPairOdd register allocator hints.
289 ARMBaseRegisterInfo::getRegAllocationHints(unsigned VirtReg,
290 ArrayRef<MCPhysReg> Order,
291 SmallVectorImpl<MCPhysReg> &Hints,
292 const MachineFunction &MF,
293 const VirtRegMap *VRM,
294 const LiveRegMatrix *Matrix) const {
295 const MachineRegisterInfo &MRI = MF.getRegInfo();
296 std::pair<unsigned, unsigned> Hint = MRI.getRegAllocationHint(VirtReg);
299 switch (Hint.first) {
300 case ARMRI::RegPairEven:
303 case ARMRI::RegPairOdd:
307 TargetRegisterInfo::getRegAllocationHints(VirtReg, Order, Hints, MF, VRM);
311 // This register should preferably be even (Odd == 0) or odd (Odd == 1).
312 // Check if the other part of the pair has already been assigned, and provide
313 // the paired register as the first hint.
314 unsigned Paired = Hint.second;
318 unsigned PairedPhys = 0;
319 if (TargetRegisterInfo::isPhysicalRegister(Paired)) {
321 } else if (VRM && VRM->hasPhys(Paired)) {
322 PairedPhys = getPairedGPR(VRM->getPhys(Paired), Odd, this);
325 // First prefer the paired physreg.
326 if (PairedPhys && is_contained(Order, PairedPhys))
327 Hints.push_back(PairedPhys);
329 // Then prefer even or odd registers.
330 for (unsigned Reg : Order) {
331 if (Reg == PairedPhys || (getEncodingValue(Reg) & 1) != Odd)
333 // Don't provide hints that are paired to a reserved register.
334 unsigned Paired = getPairedGPR(Reg, !Odd, this);
335 if (!Paired || MRI.isReserved(Paired))
337 Hints.push_back(Reg);
343 ARMBaseRegisterInfo::updateRegAllocHint(unsigned Reg, unsigned NewReg,
344 MachineFunction &MF) const {
345 MachineRegisterInfo *MRI = &MF.getRegInfo();
346 std::pair<unsigned, unsigned> Hint = MRI->getRegAllocationHint(Reg);
347 if ((Hint.first == (unsigned)ARMRI::RegPairOdd ||
348 Hint.first == (unsigned)ARMRI::RegPairEven) &&
349 TargetRegisterInfo::isVirtualRegister(Hint.second)) {
350 // If 'Reg' is one of the even / odd register pair and it's now changed
351 // (e.g. coalesced) into a different register. The other register of the
352 // pair allocation hint must be updated to reflect the relationship
354 unsigned OtherReg = Hint.second;
355 Hint = MRI->getRegAllocationHint(OtherReg);
356 // Make sure the pair has not already divorced.
357 if (Hint.second == Reg) {
358 MRI->setRegAllocationHint(OtherReg, Hint.first, NewReg);
359 if (TargetRegisterInfo::isVirtualRegister(NewReg))
360 MRI->setRegAllocationHint(NewReg,
361 Hint.first == (unsigned)ARMRI::RegPairOdd ? ARMRI::RegPairEven
362 : ARMRI::RegPairOdd, OtherReg);
367 bool ARMBaseRegisterInfo::hasBasePointer(const MachineFunction &MF) const {
368 const MachineFrameInfo &MFI = MF.getFrameInfo();
369 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
370 const ARMFrameLowering *TFI = getFrameLowering(MF);
372 // When outgoing call frames are so large that we adjust the stack pointer
373 // around the call, we can no longer use the stack pointer to reach the
374 // emergency spill slot.
375 if (needsStackRealignment(MF) && !TFI->hasReservedCallFrame(MF))
378 // Thumb has trouble with negative offsets from the FP. Thumb2 has a limited
379 // negative range for ldr/str (255), and thumb1 is positive offsets only.
380 // It's going to be better to use the SP or Base Pointer instead. When there
381 // are variable sized objects, we can't reference off of the SP, so we
382 // reserve a Base Pointer.
383 if (AFI->isThumbFunction() && MFI.hasVarSizedObjects()) {
384 // Conservatively estimate whether the negative offset from the frame
385 // pointer will be sufficient to reach. If a function has a smallish
386 // frame, it's less likely to have lots of spills and callee saved
387 // space, so it's all more likely to be within range of the frame pointer.
388 // If it's wrong, the scavenger will still enable access to work, it just
390 if (AFI->isThumb2Function() && MFI.getLocalFrameSize() < 128)
398 bool ARMBaseRegisterInfo::canRealignStack(const MachineFunction &MF) const {
399 const MachineRegisterInfo *MRI = &MF.getRegInfo();
400 const ARMFrameLowering *TFI = getFrameLowering(MF);
401 // We can't realign the stack if:
402 // 1. Dynamic stack realignment is explicitly disabled,
403 // 2. There are VLAs in the function and the base pointer is disabled.
404 if (!TargetRegisterInfo::canRealignStack(MF))
406 // Stack realignment requires a frame pointer. If we already started
407 // register allocation with frame pointer elimination, it is too late now.
408 if (!MRI->canReserveReg(getFramePointerReg(MF.getSubtarget<ARMSubtarget>())))
410 // We may also need a base pointer if there are dynamic allocas or stack
411 // pointer adjustments around calls.
412 if (TFI->hasReservedCallFrame(MF))
414 // A base pointer is required and allowed. Check that it isn't too late to
416 return MRI->canReserveReg(BasePtr);
419 bool ARMBaseRegisterInfo::
420 cannotEliminateFrame(const MachineFunction &MF) const {
421 const MachineFrameInfo &MFI = MF.getFrameInfo();
422 if (MF.getTarget().Options.DisableFramePointerElim(MF) && MFI.adjustsStack())
424 return MFI.hasVarSizedObjects() || MFI.isFrameAddressTaken()
425 || needsStackRealignment(MF);
429 ARMBaseRegisterInfo::getFrameRegister(const MachineFunction &MF) const {
430 const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();
431 const ARMFrameLowering *TFI = getFrameLowering(MF);
434 return getFramePointerReg(STI);
438 /// emitLoadConstPool - Emits a load from constpool to materialize the
439 /// specified immediate.
440 void ARMBaseRegisterInfo::emitLoadConstPool(
441 MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
442 const DebugLoc &dl, unsigned DestReg, unsigned SubIdx, int Val,
443 ARMCC::CondCodes Pred, unsigned PredReg, unsigned MIFlags) const {
444 MachineFunction &MF = *MBB.getParent();
445 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
446 MachineConstantPool *ConstantPool = MF.getConstantPool();
448 ConstantInt::get(Type::getInt32Ty(MF.getFunction().getContext()), Val);
449 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 4);
451 BuildMI(MBB, MBBI, dl, TII.get(ARM::LDRcp))
452 .addReg(DestReg, getDefRegState(true), SubIdx)
453 .addConstantPoolIndex(Idx)
455 .add(predOps(Pred, PredReg))
456 .setMIFlags(MIFlags);
459 bool ARMBaseRegisterInfo::
460 requiresRegisterScavenging(const MachineFunction &MF) const {
464 bool ARMBaseRegisterInfo::
465 trackLivenessAfterRegAlloc(const MachineFunction &MF) const {
469 bool ARMBaseRegisterInfo::
470 requiresFrameIndexScavenging(const MachineFunction &MF) const {
474 bool ARMBaseRegisterInfo::
475 requiresVirtualBaseRegisters(const MachineFunction &MF) const {
479 int64_t ARMBaseRegisterInfo::
480 getFrameIndexInstrOffset(const MachineInstr *MI, int Idx) const {
481 const MCInstrDesc &Desc = MI->getDesc();
482 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
483 int64_t InstrOffs = 0;
487 case ARMII::AddrModeT2_i8:
488 case ARMII::AddrModeT2_i12:
489 case ARMII::AddrMode_i12:
490 InstrOffs = MI->getOperand(Idx+1).getImm();
493 case ARMII::AddrMode5: {
495 const MachineOperand &OffOp = MI->getOperand(Idx+1);
496 InstrOffs = ARM_AM::getAM5Offset(OffOp.getImm());
497 if (ARM_AM::getAM5Op(OffOp.getImm()) == ARM_AM::sub)
498 InstrOffs = -InstrOffs;
502 case ARMII::AddrMode2:
504 InstrOffs = ARM_AM::getAM2Offset(MI->getOperand(ImmIdx).getImm());
505 if (ARM_AM::getAM2Op(MI->getOperand(ImmIdx).getImm()) == ARM_AM::sub)
506 InstrOffs = -InstrOffs;
508 case ARMII::AddrMode3:
510 InstrOffs = ARM_AM::getAM3Offset(MI->getOperand(ImmIdx).getImm());
511 if (ARM_AM::getAM3Op(MI->getOperand(ImmIdx).getImm()) == ARM_AM::sub)
512 InstrOffs = -InstrOffs;
514 case ARMII::AddrModeT1_s:
516 InstrOffs = MI->getOperand(ImmIdx).getImm();
520 llvm_unreachable("Unsupported addressing mode!");
523 return InstrOffs * Scale;
526 /// needsFrameBaseReg - Returns true if the instruction's frame index
527 /// reference would be better served by a base register other than FP
528 /// or SP. Used by LocalStackFrameAllocation to determine which frame index
529 /// references it should create new base registers for.
530 bool ARMBaseRegisterInfo::
531 needsFrameBaseReg(MachineInstr *MI, int64_t Offset) const {
532 for (unsigned i = 0; !MI->getOperand(i).isFI(); ++i) {
533 assert(i < MI->getNumOperands() &&"Instr doesn't have FrameIndex operand!");
536 // It's the load/store FI references that cause issues, as it can be difficult
537 // to materialize the offset if it won't fit in the literal field. Estimate
538 // based on the size of the local frame and some conservative assumptions
539 // about the rest of the stack frame (note, this is pre-regalloc, so
540 // we don't know everything for certain yet) whether this offset is likely
541 // to be out of range of the immediate. Return true if so.
543 // We only generate virtual base registers for loads and stores, so
544 // return false for everything else.
545 unsigned Opc = MI->getOpcode();
547 case ARM::LDRi12: case ARM::LDRH: case ARM::LDRBi12:
548 case ARM::STRi12: case ARM::STRH: case ARM::STRBi12:
549 case ARM::t2LDRi12: case ARM::t2LDRi8:
550 case ARM::t2STRi12: case ARM::t2STRi8:
551 case ARM::VLDRS: case ARM::VLDRD:
552 case ARM::VSTRS: case ARM::VSTRD:
553 case ARM::tSTRspi: case ARM::tLDRspi:
559 // Without a virtual base register, if the function has variable sized
560 // objects, all fixed-size local references will be via the frame pointer,
561 // Approximate the offset and see if it's legal for the instruction.
562 // Note that the incoming offset is based on the SP value at function entry,
563 // so it'll be negative.
564 MachineFunction &MF = *MI->getParent()->getParent();
565 const ARMFrameLowering *TFI = getFrameLowering(MF);
566 MachineFrameInfo &MFI = MF.getFrameInfo();
567 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
569 // Estimate an offset from the frame pointer.
570 // Conservatively assume all callee-saved registers get pushed. R4-R6
571 // will be earlier than the FP, so we ignore those.
573 int64_t FPOffset = Offset - 8;
574 // ARM and Thumb2 functions also need to consider R8-R11 and D8-D15
575 if (!AFI->isThumbFunction() || !AFI->isThumb1OnlyFunction())
577 // Estimate an offset from the stack pointer.
578 // The incoming offset is relating to the SP at the start of the function,
579 // but when we access the local it'll be relative to the SP after local
580 // allocation, so adjust our SP-relative offset by that allocation size.
581 Offset += MFI.getLocalFrameSize();
582 // Assume that we'll have at least some spill slots allocated.
583 // FIXME: This is a total SWAG number. We should run some statistics
584 // and pick a real one.
585 Offset += 128; // 128 bytes of spill slots
587 // If there's a frame pointer and the addressing mode allows it, try using it.
588 // The FP is only available if there is no dynamic realignment. We
589 // don't know for sure yet whether we'll need that, so we guess based
590 // on whether there are any local variables that would trigger it.
591 unsigned StackAlign = TFI->getStackAlignment();
592 if (TFI->hasFP(MF) &&
593 !((MFI.getLocalFrameMaxAlign() > StackAlign) && canRealignStack(MF))) {
594 if (isFrameOffsetLegal(MI, getFrameRegister(MF), FPOffset))
597 // If we can reference via the stack pointer, try that.
598 // FIXME: This (and the code that resolves the references) can be improved
599 // to only disallow SP relative references in the live range of
600 // the VLA(s). In practice, it's unclear how much difference that
601 // would make, but it may be worth doing.
602 if (!MFI.hasVarSizedObjects() && isFrameOffsetLegal(MI, ARM::SP, Offset))
605 // The offset likely isn't legal, we want to allocate a virtual base register.
609 /// materializeFrameBaseRegister - Insert defining instruction(s) for BaseReg to
610 /// be a pointer to FrameIdx at the beginning of the basic block.
611 void ARMBaseRegisterInfo::
612 materializeFrameBaseRegister(MachineBasicBlock *MBB,
613 unsigned BaseReg, int FrameIdx,
614 int64_t Offset) const {
615 ARMFunctionInfo *AFI = MBB->getParent()->getInfo<ARMFunctionInfo>();
616 unsigned ADDriOpc = !AFI->isThumbFunction() ? ARM::ADDri :
617 (AFI->isThumb1OnlyFunction() ? ARM::tADDframe : ARM::t2ADDri);
619 MachineBasicBlock::iterator Ins = MBB->begin();
620 DebugLoc DL; // Defaults to "unknown"
621 if (Ins != MBB->end())
622 DL = Ins->getDebugLoc();
624 const MachineFunction &MF = *MBB->getParent();
625 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
626 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
627 const MCInstrDesc &MCID = TII.get(ADDriOpc);
628 MRI.constrainRegClass(BaseReg, TII.getRegClass(MCID, 0, this, MF));
630 MachineInstrBuilder MIB = BuildMI(*MBB, Ins, DL, MCID, BaseReg)
631 .addFrameIndex(FrameIdx).addImm(Offset);
633 if (!AFI->isThumb1OnlyFunction())
634 MIB.add(predOps(ARMCC::AL)).add(condCodeOp());
637 void ARMBaseRegisterInfo::resolveFrameIndex(MachineInstr &MI, unsigned BaseReg,
638 int64_t Offset) const {
639 MachineBasicBlock &MBB = *MI.getParent();
640 MachineFunction &MF = *MBB.getParent();
641 const ARMBaseInstrInfo &TII =
642 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
643 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
644 int Off = Offset; // ARM doesn't need the general 64-bit offsets
647 assert(!AFI->isThumb1OnlyFunction() &&
648 "This resolveFrameIndex does not support Thumb1!");
650 while (!MI.getOperand(i).isFI()) {
652 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
655 if (!AFI->isThumbFunction())
656 Done = rewriteARMFrameIndex(MI, i, BaseReg, Off, TII);
658 assert(AFI->isThumb2Function());
659 Done = rewriteT2FrameIndex(MI, i, BaseReg, Off, TII);
661 assert(Done && "Unable to resolve frame index!");
665 bool ARMBaseRegisterInfo::isFrameOffsetLegal(const MachineInstr *MI, unsigned BaseReg,
666 int64_t Offset) const {
667 const MCInstrDesc &Desc = MI->getDesc();
668 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
670 for (; !MI->getOperand(i).isFI(); ++i)
671 assert(i+1 < MI->getNumOperands() && "Instr doesn't have FrameIndex operand!");
673 // AddrMode4 and AddrMode6 cannot handle any offset.
674 if (AddrMode == ARMII::AddrMode4 || AddrMode == ARMII::AddrMode6)
677 unsigned NumBits = 0;
679 bool isSigned = true;
681 case ARMII::AddrModeT2_i8:
682 case ARMII::AddrModeT2_i12:
683 // i8 supports only negative, and i12 supports only positive, so
684 // based on Offset sign, consider the appropriate instruction
693 case ARMII::AddrMode5:
698 case ARMII::AddrMode_i12:
699 case ARMII::AddrMode2:
702 case ARMII::AddrMode3:
705 case ARMII::AddrModeT1_s:
706 NumBits = (BaseReg == ARM::SP ? 8 : 5);
711 llvm_unreachable("Unsupported addressing mode!");
714 Offset += getFrameIndexInstrOffset(MI, i);
715 // Make sure the offset is encodable for instructions that scale the
717 if ((Offset & (Scale-1)) != 0)
720 if (isSigned && Offset < 0)
723 unsigned Mask = (1 << NumBits) - 1;
724 if ((unsigned)Offset <= Mask * Scale)
731 ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
732 int SPAdj, unsigned FIOperandNum,
733 RegScavenger *RS) const {
734 MachineInstr &MI = *II;
735 MachineBasicBlock &MBB = *MI.getParent();
736 MachineFunction &MF = *MBB.getParent();
737 const ARMBaseInstrInfo &TII =
738 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
739 const ARMFrameLowering *TFI = getFrameLowering(MF);
740 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
741 assert(!AFI->isThumb1OnlyFunction() &&
742 "This eliminateFrameIndex does not support Thumb1!");
743 int FrameIndex = MI.getOperand(FIOperandNum).getIndex();
746 int Offset = TFI->ResolveFrameIndexReference(MF, FrameIndex, FrameReg, SPAdj);
748 // PEI::scavengeFrameVirtualRegs() cannot accurately track SPAdj because the
749 // call frame setup/destroy instructions have already been eliminated. That
750 // means the stack pointer cannot be used to access the emergency spill slot
751 // when !hasReservedCallFrame().
753 if (RS && FrameReg == ARM::SP && RS->isScavengingFrameIndex(FrameIndex)){
754 assert(TFI->hasReservedCallFrame(MF) &&
755 "Cannot use SP to access the emergency spill slot in "
756 "functions without a reserved call frame");
757 assert(!MF.getFrameInfo().hasVarSizedObjects() &&
758 "Cannot use SP to access the emergency spill slot in "
759 "functions with variable sized frame objects");
763 assert(!MI.isDebugValue() && "DBG_VALUEs should be handled in target-independent code");
765 // Modify MI as necessary to handle as much of 'Offset' as possible
767 if (!AFI->isThumbFunction())
768 Done = rewriteARMFrameIndex(MI, FIOperandNum, FrameReg, Offset, TII);
770 assert(AFI->isThumb2Function());
771 Done = rewriteT2FrameIndex(MI, FIOperandNum, FrameReg, Offset, TII);
776 // If we get here, the immediate doesn't fit into the instruction. We folded
777 // as much as possible above, handle the rest, providing a register that is
780 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode4 ||
781 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode6) &&
782 "This code isn't needed if offset already handled!");
784 unsigned ScratchReg = 0;
785 int PIdx = MI.findFirstPredOperandIdx();
786 ARMCC::CondCodes Pred = (PIdx == -1)
787 ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm();
788 unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg();
790 // Must be addrmode4/6.
791 MI.getOperand(FIOperandNum).ChangeToRegister(FrameReg, false, false, false);
793 ScratchReg = MF.getRegInfo().createVirtualRegister(&ARM::GPRRegClass);
794 if (!AFI->isThumbFunction())
795 emitARMRegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg,
796 Offset, Pred, PredReg, TII);
798 assert(AFI->isThumb2Function());
799 emitT2RegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg,
800 Offset, Pred, PredReg, TII);
802 // Update the original instruction to use the scratch register.
803 MI.getOperand(FIOperandNum).ChangeToRegister(ScratchReg, false, false,true);
807 bool ARMBaseRegisterInfo::shouldCoalesce(MachineInstr *MI,
808 const TargetRegisterClass *SrcRC,
810 const TargetRegisterClass *DstRC,
812 const TargetRegisterClass *NewRC,
813 LiveIntervals &LIS) const {
814 auto MBB = MI->getParent();
815 auto MF = MBB->getParent();
816 const MachineRegisterInfo &MRI = MF->getRegInfo();
817 // If not copying into a sub-register this should be ok because we shouldn't
818 // need to split the reg.
821 // Small registers don't frequently cause a problem, so we can coalesce them.
822 if (getRegSizeInBits(*NewRC) < 256 && getRegSizeInBits(*DstRC) < 256 &&
823 getRegSizeInBits(*SrcRC) < 256)
827 MRI.getTargetRegisterInfo()->getRegClassWeight(NewRC);
829 MRI.getTargetRegisterInfo()->getRegClassWeight(SrcRC);
831 MRI.getTargetRegisterInfo()->getRegClassWeight(DstRC);
832 // If the source register class is more expensive than the destination, the
833 // coalescing is probably profitable.
834 if (SrcRCWeight.RegWeight > NewRCWeight.RegWeight)
836 if (DstRCWeight.RegWeight > NewRCWeight.RegWeight)
839 // If the register allocator isn't constrained, we can always allow coalescing
840 // unfortunately we don't know yet if we will be constrained.
841 // The goal of this heuristic is to restrict how many expensive registers
842 // we allow to coalesce in a given basic block.
843 auto AFI = MF->getInfo<ARMFunctionInfo>();
844 auto It = AFI->getCoalescedWeight(MBB);
846 LLVM_DEBUG(dbgs() << "\tARM::shouldCoalesce - Coalesced Weight: "
847 << It->second << "\n");
848 LLVM_DEBUG(dbgs() << "\tARM::shouldCoalesce - Reg Weight: "
849 << NewRCWeight.RegWeight << "\n");
851 // This number is the largest round number that which meets the criteria:
852 // (1) addresses PR18825
853 // (2) generates better code in some test cases (like vldm-shed-a9.ll)
854 // (3) Doesn't regress any test cases (in-tree, test-suite, and SPEC)
855 // In practice the SizeMultiplier will only factor in for straight line code
856 // that uses a lot of NEON vectors, which isn't terribly common.
857 unsigned SizeMultiplier = MBB->size()/100;
858 SizeMultiplier = SizeMultiplier ? SizeMultiplier : 1;
859 if (It->second < NewRCWeight.WeightLimit * SizeMultiplier) {
860 It->second += NewRCWeight.RegWeight;