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 "llvm/ADT/BitVector.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/RegisterScavenging.h"
29 #include "llvm/CodeGen/VirtRegMap.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Target/TargetFrameLowering.h"
38 #include "llvm/Target/TargetMachine.h"
39 #include "llvm/Target/TargetOptions.h"
41 #define DEBUG_TYPE "arm-register-info"
43 #define GET_REGINFO_TARGET_DESC
44 #include "ARMGenRegisterInfo.inc"
48 ARMBaseRegisterInfo::ARMBaseRegisterInfo()
49 : ARMGenRegisterInfo(ARM::LR, 0, 0, ARM::PC), BasePtr(ARM::R6) {}
51 static unsigned getFramePointerReg(const ARMSubtarget &STI) {
52 if (STI.isTargetMachO()) {
53 if (STI.isTargetDarwin() || STI.isThumb1Only())
57 } else if (STI.isTargetWindows())
60 return STI.isThumb() ? ARM::R7 : ARM::R11;
64 ARMBaseRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
65 const ARMSubtarget &STI = MF->getSubtarget<ARMSubtarget>();
66 const MCPhysReg *RegList =
67 STI.isTargetDarwin() ? CSR_iOS_SaveList : CSR_AAPCS_SaveList;
69 const Function *F = MF->getFunction();
70 if (F->getCallingConv() == CallingConv::GHC) {
71 // GHC set of callee saved regs is empty as all those regs are
72 // used for passing STG regs around
73 return CSR_NoRegs_SaveList;
74 } else if (F->hasFnAttribute("interrupt")) {
76 // M-class CPUs have hardware which saves the registers needed to allow a
77 // function conforming to the AAPCS to function as a handler.
78 return CSR_AAPCS_SaveList;
79 } else if (F->getFnAttribute("interrupt").getValueAsString() == "FIQ") {
80 // Fast interrupt mode gives the handler a private copy of R8-R14, so less
81 // need to be saved to restore user-mode state.
82 return CSR_FIQ_SaveList;
84 // Generally only R13-R14 (i.e. SP, LR) are automatically preserved by
85 // exception handling.
86 return CSR_GenericInt_SaveList;
90 if (STI.isTargetDarwin() && F->getCallingConv() == CallingConv::CXX_FAST_TLS)
91 return CSR_iOS_CXX_TLS_SaveList;
96 ARMBaseRegisterInfo::getCallPreservedMask(const MachineFunction &MF,
97 CallingConv::ID CC) const {
98 const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();
99 if (CC == CallingConv::GHC)
100 // This is academic becase all GHC calls are (supposed to be) tail calls
101 return CSR_NoRegs_RegMask;
102 if (STI.isTargetDarwin() && CC == CallingConv::CXX_FAST_TLS)
103 return CSR_iOS_CXX_TLS_RegMask;
104 return STI.isTargetDarwin() ? CSR_iOS_RegMask : CSR_AAPCS_RegMask;
108 ARMBaseRegisterInfo::getNoPreservedMask() const {
109 return CSR_NoRegs_RegMask;
113 ARMBaseRegisterInfo::getTLSCallPreservedMask(const MachineFunction &MF) const {
114 assert(MF.getSubtarget<ARMSubtarget>().isTargetDarwin() &&
115 "only know about special TLS call on Darwin");
116 return CSR_iOS_TLSCall_RegMask;
121 ARMBaseRegisterInfo::getThisReturnPreservedMask(const MachineFunction &MF,
122 CallingConv::ID CC) const {
123 const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();
124 // This should return a register mask that is the same as that returned by
125 // getCallPreservedMask but that additionally preserves the register used for
126 // the first i32 argument (which must also be the register used to return a
127 // single i32 return value)
129 // In case that the calling convention does not use the same register for
130 // both or otherwise does not want to enable this optimization, the function
131 // should return NULL
132 if (CC == CallingConv::GHC)
133 // This is academic becase all GHC calls are (supposed to be) tail calls
135 return STI.isTargetDarwin() ? CSR_iOS_ThisReturn_RegMask
136 : CSR_AAPCS_ThisReturn_RegMask;
139 BitVector ARMBaseRegisterInfo::
140 getReservedRegs(const MachineFunction &MF) const {
141 const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();
142 const ARMFrameLowering *TFI = getFrameLowering(MF);
144 // FIXME: avoid re-calculating this every time.
145 BitVector Reserved(getNumRegs());
146 Reserved.set(ARM::SP);
147 Reserved.set(ARM::PC);
148 Reserved.set(ARM::FPSCR);
149 Reserved.set(ARM::APSR_NZCV);
151 Reserved.set(getFramePointerReg(STI));
152 if (hasBasePointer(MF))
153 Reserved.set(BasePtr);
154 // Some targets reserve R9.
155 if (STI.isR9Reserved())
156 Reserved.set(ARM::R9);
157 // Reserve D16-D31 if the subtarget doesn't support them.
158 if (!STI.hasVFP3() || STI.hasD16()) {
159 assert(ARM::D31 == ARM::D16 + 15);
160 for (unsigned i = 0; i != 16; ++i)
161 Reserved.set(ARM::D16 + i);
163 const TargetRegisterClass *RC = &ARM::GPRPairRegClass;
164 for(TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); I!=E; ++I)
165 for (MCSubRegIterator SI(*I, this); SI.isValid(); ++SI)
166 if (Reserved.test(*SI)) Reserved.set(*I);
171 const TargetRegisterClass *
172 ARMBaseRegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC,
173 const MachineFunction &) const {
174 const TargetRegisterClass *Super = RC;
175 TargetRegisterClass::sc_iterator I = RC->getSuperClasses();
177 switch (Super->getID()) {
178 case ARM::GPRRegClassID:
179 case ARM::SPRRegClassID:
180 case ARM::DPRRegClassID:
181 case ARM::QPRRegClassID:
182 case ARM::QQPRRegClassID:
183 case ARM::QQQQPRRegClassID:
184 case ARM::GPRPairRegClassID:
192 const TargetRegisterClass *
193 ARMBaseRegisterInfo::getPointerRegClass(const MachineFunction &MF, unsigned Kind)
195 return &ARM::GPRRegClass;
198 const TargetRegisterClass *
199 ARMBaseRegisterInfo::getCrossCopyRegClass(const TargetRegisterClass *RC) const {
200 if (RC == &ARM::CCRRegClass)
201 return &ARM::rGPRRegClass; // Can't copy CCR registers.
206 ARMBaseRegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC,
207 MachineFunction &MF) const {
208 const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();
209 const ARMFrameLowering *TFI = getFrameLowering(MF);
211 switch (RC->getID()) {
214 case ARM::tGPRRegClassID:
215 return TFI->hasFP(MF) ? 4 : 5;
216 case ARM::GPRRegClassID: {
217 unsigned FP = TFI->hasFP(MF) ? 1 : 0;
218 return 10 - FP - (STI.isR9Reserved() ? 1 : 0);
220 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
221 case ARM::DPRRegClassID:
226 // Get the other register in a GPRPair.
227 static unsigned getPairedGPR(unsigned Reg, bool Odd, const MCRegisterInfo *RI) {
228 for (MCSuperRegIterator Supers(Reg, RI); Supers.isValid(); ++Supers)
229 if (ARM::GPRPairRegClass.contains(*Supers))
230 return RI->getSubReg(*Supers, Odd ? ARM::gsub_1 : ARM::gsub_0);
234 // Resolve the RegPairEven / RegPairOdd register allocator hints.
236 ARMBaseRegisterInfo::getRegAllocationHints(unsigned VirtReg,
237 ArrayRef<MCPhysReg> Order,
238 SmallVectorImpl<MCPhysReg> &Hints,
239 const MachineFunction &MF,
240 const VirtRegMap *VRM,
241 const LiveRegMatrix *Matrix) const {
242 const MachineRegisterInfo &MRI = MF.getRegInfo();
243 std::pair<unsigned, unsigned> Hint = MRI.getRegAllocationHint(VirtReg);
246 switch (Hint.first) {
247 case ARMRI::RegPairEven:
250 case ARMRI::RegPairOdd:
254 TargetRegisterInfo::getRegAllocationHints(VirtReg, Order, Hints, MF, VRM);
258 // This register should preferably be even (Odd == 0) or odd (Odd == 1).
259 // Check if the other part of the pair has already been assigned, and provide
260 // the paired register as the first hint.
261 unsigned Paired = Hint.second;
265 unsigned PairedPhys = 0;
266 if (TargetRegisterInfo::isPhysicalRegister(Paired)) {
268 } else if (VRM && VRM->hasPhys(Paired)) {
269 PairedPhys = getPairedGPR(VRM->getPhys(Paired), Odd, this);
272 // First prefer the paired physreg.
274 std::find(Order.begin(), Order.end(), PairedPhys) != Order.end())
275 Hints.push_back(PairedPhys);
277 // Then prefer even or odd registers.
278 for (unsigned I = 0, E = Order.size(); I != E; ++I) {
279 unsigned Reg = Order[I];
280 if (Reg == PairedPhys || (getEncodingValue(Reg) & 1) != Odd)
282 // Don't provide hints that are paired to a reserved register.
283 unsigned Paired = getPairedGPR(Reg, !Odd, this);
284 if (!Paired || MRI.isReserved(Paired))
286 Hints.push_back(Reg);
291 ARMBaseRegisterInfo::updateRegAllocHint(unsigned Reg, unsigned NewReg,
292 MachineFunction &MF) const {
293 MachineRegisterInfo *MRI = &MF.getRegInfo();
294 std::pair<unsigned, unsigned> Hint = MRI->getRegAllocationHint(Reg);
295 if ((Hint.first == (unsigned)ARMRI::RegPairOdd ||
296 Hint.first == (unsigned)ARMRI::RegPairEven) &&
297 TargetRegisterInfo::isVirtualRegister(Hint.second)) {
298 // If 'Reg' is one of the even / odd register pair and it's now changed
299 // (e.g. coalesced) into a different register. The other register of the
300 // pair allocation hint must be updated to reflect the relationship
302 unsigned OtherReg = Hint.second;
303 Hint = MRI->getRegAllocationHint(OtherReg);
304 // Make sure the pair has not already divorced.
305 if (Hint.second == Reg) {
306 MRI->setRegAllocationHint(OtherReg, Hint.first, NewReg);
307 if (TargetRegisterInfo::isVirtualRegister(NewReg))
308 MRI->setRegAllocationHint(NewReg,
309 Hint.first == (unsigned)ARMRI::RegPairOdd ? ARMRI::RegPairEven
310 : ARMRI::RegPairOdd, OtherReg);
315 bool ARMBaseRegisterInfo::hasBasePointer(const MachineFunction &MF) const {
316 const MachineFrameInfo *MFI = MF.getFrameInfo();
317 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
318 const ARMFrameLowering *TFI = getFrameLowering(MF);
320 // When outgoing call frames are so large that we adjust the stack pointer
321 // around the call, we can no longer use the stack pointer to reach the
322 // emergency spill slot.
323 if (needsStackRealignment(MF) && !TFI->hasReservedCallFrame(MF))
326 // Thumb has trouble with negative offsets from the FP. Thumb2 has a limited
327 // negative range for ldr/str (255), and thumb1 is positive offsets only.
328 // It's going to be better to use the SP or Base Pointer instead. When there
329 // are variable sized objects, we can't reference off of the SP, so we
330 // reserve a Base Pointer.
331 if (AFI->isThumbFunction() && MFI->hasVarSizedObjects()) {
332 // Conservatively estimate whether the negative offset from the frame
333 // pointer will be sufficient to reach. If a function has a smallish
334 // frame, it's less likely to have lots of spills and callee saved
335 // space, so it's all more likely to be within range of the frame pointer.
336 // If it's wrong, the scavenger will still enable access to work, it just
338 if (AFI->isThumb2Function() && MFI->getLocalFrameSize() < 128)
346 bool ARMBaseRegisterInfo::canRealignStack(const MachineFunction &MF) const {
347 const MachineRegisterInfo *MRI = &MF.getRegInfo();
348 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
349 const ARMFrameLowering *TFI = getFrameLowering(MF);
350 // We can't realign the stack if:
351 // 1. Dynamic stack realignment is explicitly disabled,
352 // 2. This is a Thumb1 function (it's not useful, so we don't bother), or
353 // 3. There are VLAs in the function and the base pointer is disabled.
354 if (!TargetRegisterInfo::canRealignStack(MF))
356 if (AFI->isThumb1OnlyFunction())
358 // Stack realignment requires a frame pointer. If we already started
359 // register allocation with frame pointer elimination, it is too late now.
360 if (!MRI->canReserveReg(getFramePointerReg(MF.getSubtarget<ARMSubtarget>())))
362 // We may also need a base pointer if there are dynamic allocas or stack
363 // pointer adjustments around calls.
364 if (TFI->hasReservedCallFrame(MF))
366 // A base pointer is required and allowed. Check that it isn't too late to
368 return MRI->canReserveReg(BasePtr);
371 bool ARMBaseRegisterInfo::
372 cannotEliminateFrame(const MachineFunction &MF) const {
373 const MachineFrameInfo *MFI = MF.getFrameInfo();
374 if (MF.getTarget().Options.DisableFramePointerElim(MF) && MFI->adjustsStack())
376 return MFI->hasVarSizedObjects() || MFI->isFrameAddressTaken()
377 || needsStackRealignment(MF);
381 ARMBaseRegisterInfo::getFrameRegister(const MachineFunction &MF) const {
382 const ARMSubtarget &STI = MF.getSubtarget<ARMSubtarget>();
383 const ARMFrameLowering *TFI = getFrameLowering(MF);
386 return getFramePointerReg(STI);
390 /// emitLoadConstPool - Emits a load from constpool to materialize the
391 /// specified immediate.
392 void ARMBaseRegisterInfo::
393 emitLoadConstPool(MachineBasicBlock &MBB,
394 MachineBasicBlock::iterator &MBBI,
396 unsigned DestReg, unsigned SubIdx, int Val,
397 ARMCC::CondCodes Pred,
398 unsigned PredReg, unsigned MIFlags) const {
399 MachineFunction &MF = *MBB.getParent();
400 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
401 MachineConstantPool *ConstantPool = MF.getConstantPool();
403 ConstantInt::get(Type::getInt32Ty(MF.getFunction()->getContext()), Val);
404 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 4);
406 BuildMI(MBB, MBBI, dl, TII.get(ARM::LDRcp))
407 .addReg(DestReg, getDefRegState(true), SubIdx)
408 .addConstantPoolIndex(Idx)
409 .addImm(0).addImm(Pred).addReg(PredReg)
410 .setMIFlags(MIFlags);
413 bool ARMBaseRegisterInfo::
414 requiresRegisterScavenging(const MachineFunction &MF) const {
418 bool ARMBaseRegisterInfo::
419 trackLivenessAfterRegAlloc(const MachineFunction &MF) const {
423 bool ARMBaseRegisterInfo::
424 requiresFrameIndexScavenging(const MachineFunction &MF) const {
428 bool ARMBaseRegisterInfo::
429 requiresVirtualBaseRegisters(const MachineFunction &MF) const {
433 int64_t ARMBaseRegisterInfo::
434 getFrameIndexInstrOffset(const MachineInstr *MI, int Idx) const {
435 const MCInstrDesc &Desc = MI->getDesc();
436 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
437 int64_t InstrOffs = 0;
441 case ARMII::AddrModeT2_i8:
442 case ARMII::AddrModeT2_i12:
443 case ARMII::AddrMode_i12:
444 InstrOffs = MI->getOperand(Idx+1).getImm();
447 case ARMII::AddrMode5: {
449 const MachineOperand &OffOp = MI->getOperand(Idx+1);
450 InstrOffs = ARM_AM::getAM5Offset(OffOp.getImm());
451 if (ARM_AM::getAM5Op(OffOp.getImm()) == ARM_AM::sub)
452 InstrOffs = -InstrOffs;
456 case ARMII::AddrMode2: {
458 InstrOffs = ARM_AM::getAM2Offset(MI->getOperand(ImmIdx).getImm());
459 if (ARM_AM::getAM2Op(MI->getOperand(ImmIdx).getImm()) == ARM_AM::sub)
460 InstrOffs = -InstrOffs;
463 case ARMII::AddrMode3: {
465 InstrOffs = ARM_AM::getAM3Offset(MI->getOperand(ImmIdx).getImm());
466 if (ARM_AM::getAM3Op(MI->getOperand(ImmIdx).getImm()) == ARM_AM::sub)
467 InstrOffs = -InstrOffs;
470 case ARMII::AddrModeT1_s: {
472 InstrOffs = MI->getOperand(ImmIdx).getImm();
477 llvm_unreachable("Unsupported addressing mode!");
480 return InstrOffs * Scale;
483 /// needsFrameBaseReg - Returns true if the instruction's frame index
484 /// reference would be better served by a base register other than FP
485 /// or SP. Used by LocalStackFrameAllocation to determine which frame index
486 /// references it should create new base registers for.
487 bool ARMBaseRegisterInfo::
488 needsFrameBaseReg(MachineInstr *MI, int64_t Offset) const {
489 for (unsigned i = 0; !MI->getOperand(i).isFI(); ++i) {
490 assert(i < MI->getNumOperands() &&"Instr doesn't have FrameIndex operand!");
493 // It's the load/store FI references that cause issues, as it can be difficult
494 // to materialize the offset if it won't fit in the literal field. Estimate
495 // based on the size of the local frame and some conservative assumptions
496 // about the rest of the stack frame (note, this is pre-regalloc, so
497 // we don't know everything for certain yet) whether this offset is likely
498 // to be out of range of the immediate. Return true if so.
500 // We only generate virtual base registers for loads and stores, so
501 // return false for everything else.
502 unsigned Opc = MI->getOpcode();
504 case ARM::LDRi12: case ARM::LDRH: case ARM::LDRBi12:
505 case ARM::STRi12: case ARM::STRH: case ARM::STRBi12:
506 case ARM::t2LDRi12: case ARM::t2LDRi8:
507 case ARM::t2STRi12: case ARM::t2STRi8:
508 case ARM::VLDRS: case ARM::VLDRD:
509 case ARM::VSTRS: case ARM::VSTRD:
510 case ARM::tSTRspi: case ARM::tLDRspi:
516 // Without a virtual base register, if the function has variable sized
517 // objects, all fixed-size local references will be via the frame pointer,
518 // Approximate the offset and see if it's legal for the instruction.
519 // Note that the incoming offset is based on the SP value at function entry,
520 // so it'll be negative.
521 MachineFunction &MF = *MI->getParent()->getParent();
522 const ARMFrameLowering *TFI = getFrameLowering(MF);
523 MachineFrameInfo *MFI = MF.getFrameInfo();
524 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
526 // Estimate an offset from the frame pointer.
527 // Conservatively assume all callee-saved registers get pushed. R4-R6
528 // will be earlier than the FP, so we ignore those.
530 int64_t FPOffset = Offset - 8;
531 // ARM and Thumb2 functions also need to consider R8-R11 and D8-D15
532 if (!AFI->isThumbFunction() || !AFI->isThumb1OnlyFunction())
534 // Estimate an offset from the stack pointer.
535 // The incoming offset is relating to the SP at the start of the function,
536 // but when we access the local it'll be relative to the SP after local
537 // allocation, so adjust our SP-relative offset by that allocation size.
538 Offset += MFI->getLocalFrameSize();
539 // Assume that we'll have at least some spill slots allocated.
540 // FIXME: This is a total SWAG number. We should run some statistics
541 // and pick a real one.
542 Offset += 128; // 128 bytes of spill slots
544 // If there's a frame pointer and the addressing mode allows it, try using it.
545 // The FP is only available if there is no dynamic realignment. We
546 // don't know for sure yet whether we'll need that, so we guess based
547 // on whether there are any local variables that would trigger it.
548 unsigned StackAlign = TFI->getStackAlignment();
549 if (TFI->hasFP(MF) &&
550 !((MFI->getLocalFrameMaxAlign() > StackAlign) && canRealignStack(MF))) {
551 if (isFrameOffsetLegal(MI, getFrameRegister(MF), FPOffset))
554 // If we can reference via the stack pointer, try that.
555 // FIXME: This (and the code that resolves the references) can be improved
556 // to only disallow SP relative references in the live range of
557 // the VLA(s). In practice, it's unclear how much difference that
558 // would make, but it may be worth doing.
559 if (!MFI->hasVarSizedObjects() && isFrameOffsetLegal(MI, ARM::SP, Offset))
562 // The offset likely isn't legal, we want to allocate a virtual base register.
566 /// materializeFrameBaseRegister - Insert defining instruction(s) for BaseReg to
567 /// be a pointer to FrameIdx at the beginning of the basic block.
568 void ARMBaseRegisterInfo::
569 materializeFrameBaseRegister(MachineBasicBlock *MBB,
570 unsigned BaseReg, int FrameIdx,
571 int64_t Offset) const {
572 ARMFunctionInfo *AFI = MBB->getParent()->getInfo<ARMFunctionInfo>();
573 unsigned ADDriOpc = !AFI->isThumbFunction() ? ARM::ADDri :
574 (AFI->isThumb1OnlyFunction() ? ARM::tADDframe : ARM::t2ADDri);
576 MachineBasicBlock::iterator Ins = MBB->begin();
577 DebugLoc DL; // Defaults to "unknown"
578 if (Ins != MBB->end())
579 DL = Ins->getDebugLoc();
581 const MachineFunction &MF = *MBB->getParent();
582 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
583 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
584 const MCInstrDesc &MCID = TII.get(ADDriOpc);
585 MRI.constrainRegClass(BaseReg, TII.getRegClass(MCID, 0, this, MF));
587 MachineInstrBuilder MIB = BuildMI(*MBB, Ins, DL, MCID, BaseReg)
588 .addFrameIndex(FrameIdx).addImm(Offset);
590 if (!AFI->isThumb1OnlyFunction())
591 AddDefaultCC(AddDefaultPred(MIB));
594 void ARMBaseRegisterInfo::resolveFrameIndex(MachineInstr &MI, unsigned BaseReg,
595 int64_t Offset) const {
596 MachineBasicBlock &MBB = *MI.getParent();
597 MachineFunction &MF = *MBB.getParent();
598 const ARMBaseInstrInfo &TII =
599 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
600 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
601 int Off = Offset; // ARM doesn't need the general 64-bit offsets
604 assert(!AFI->isThumb1OnlyFunction() &&
605 "This resolveFrameIndex does not support Thumb1!");
607 while (!MI.getOperand(i).isFI()) {
609 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
612 if (!AFI->isThumbFunction())
613 Done = rewriteARMFrameIndex(MI, i, BaseReg, Off, TII);
615 assert(AFI->isThumb2Function());
616 Done = rewriteT2FrameIndex(MI, i, BaseReg, Off, TII);
618 assert (Done && "Unable to resolve frame index!");
622 bool ARMBaseRegisterInfo::isFrameOffsetLegal(const MachineInstr *MI, unsigned BaseReg,
623 int64_t Offset) const {
624 const MCInstrDesc &Desc = MI->getDesc();
625 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
628 while (!MI->getOperand(i).isFI()) {
630 assert(i < MI->getNumOperands() &&"Instr doesn't have FrameIndex operand!");
633 // AddrMode4 and AddrMode6 cannot handle any offset.
634 if (AddrMode == ARMII::AddrMode4 || AddrMode == ARMII::AddrMode6)
637 unsigned NumBits = 0;
639 bool isSigned = true;
641 case ARMII::AddrModeT2_i8:
642 case ARMII::AddrModeT2_i12:
643 // i8 supports only negative, and i12 supports only positive, so
644 // based on Offset sign, consider the appropriate instruction
653 case ARMII::AddrMode5:
658 case ARMII::AddrMode_i12:
659 case ARMII::AddrMode2:
662 case ARMII::AddrMode3:
665 case ARMII::AddrModeT1_s:
666 NumBits = (BaseReg == ARM::SP ? 8 : 5);
671 llvm_unreachable("Unsupported addressing mode!");
674 Offset += getFrameIndexInstrOffset(MI, i);
675 // Make sure the offset is encodable for instructions that scale the
677 if ((Offset & (Scale-1)) != 0)
680 if (isSigned && Offset < 0)
683 unsigned Mask = (1 << NumBits) - 1;
684 if ((unsigned)Offset <= Mask * Scale)
691 ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
692 int SPAdj, unsigned FIOperandNum,
693 RegScavenger *RS) const {
694 MachineInstr &MI = *II;
695 MachineBasicBlock &MBB = *MI.getParent();
696 MachineFunction &MF = *MBB.getParent();
697 const ARMBaseInstrInfo &TII =
698 *static_cast<const ARMBaseInstrInfo *>(MF.getSubtarget().getInstrInfo());
699 const ARMFrameLowering *TFI = getFrameLowering(MF);
700 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
701 assert(!AFI->isThumb1OnlyFunction() &&
702 "This eliminateFrameIndex does not support Thumb1!");
703 int FrameIndex = MI.getOperand(FIOperandNum).getIndex();
706 int Offset = TFI->ResolveFrameIndexReference(MF, FrameIndex, FrameReg, SPAdj);
708 // PEI::scavengeFrameVirtualRegs() cannot accurately track SPAdj because the
709 // call frame setup/destroy instructions have already been eliminated. That
710 // means the stack pointer cannot be used to access the emergency spill slot
711 // when !hasReservedCallFrame().
713 if (RS && FrameReg == ARM::SP && RS->isScavengingFrameIndex(FrameIndex)){
714 assert(TFI->hasReservedCallFrame(MF) &&
715 "Cannot use SP to access the emergency spill slot in "
716 "functions without a reserved call frame");
717 assert(!MF.getFrameInfo()->hasVarSizedObjects() &&
718 "Cannot use SP to access the emergency spill slot in "
719 "functions with variable sized frame objects");
723 assert(!MI.isDebugValue() && "DBG_VALUEs should be handled in target-independent code");
725 // Modify MI as necessary to handle as much of 'Offset' as possible
727 if (!AFI->isThumbFunction())
728 Done = rewriteARMFrameIndex(MI, FIOperandNum, FrameReg, Offset, TII);
730 assert(AFI->isThumb2Function());
731 Done = rewriteT2FrameIndex(MI, FIOperandNum, FrameReg, Offset, TII);
736 // If we get here, the immediate doesn't fit into the instruction. We folded
737 // as much as possible above, handle the rest, providing a register that is
740 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode4 ||
741 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode6) &&
742 "This code isn't needed if offset already handled!");
744 unsigned ScratchReg = 0;
745 int PIdx = MI.findFirstPredOperandIdx();
746 ARMCC::CondCodes Pred = (PIdx == -1)
747 ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm();
748 unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg();
750 // Must be addrmode4/6.
751 MI.getOperand(FIOperandNum).ChangeToRegister(FrameReg, false, false, false);
753 ScratchReg = MF.getRegInfo().createVirtualRegister(&ARM::GPRRegClass);
754 if (!AFI->isThumbFunction())
755 emitARMRegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg,
756 Offset, Pred, PredReg, TII);
758 assert(AFI->isThumb2Function());
759 emitT2RegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg,
760 Offset, Pred, PredReg, TII);
762 // Update the original instruction to use the scratch register.
763 MI.getOperand(FIOperandNum).ChangeToRegister(ScratchReg, false, false,true);
767 bool ARMBaseRegisterInfo::shouldCoalesce(MachineInstr *MI,
768 const TargetRegisterClass *SrcRC,
770 const TargetRegisterClass *DstRC,
772 const TargetRegisterClass *NewRC) const {
773 auto MBB = MI->getParent();
774 auto MF = MBB->getParent();
775 const MachineRegisterInfo &MRI = MF->getRegInfo();
776 // If not copying into a sub-register this should be ok because we shouldn't
777 // need to split the reg.
780 // Small registers don't frequently cause a problem, so we can coalesce them.
781 if (NewRC->getSize() < 32 && DstRC->getSize() < 32 && SrcRC->getSize() < 32)
785 MRI.getTargetRegisterInfo()->getRegClassWeight(NewRC);
787 MRI.getTargetRegisterInfo()->getRegClassWeight(SrcRC);
789 MRI.getTargetRegisterInfo()->getRegClassWeight(DstRC);
790 // If the source register class is more expensive than the destination, the
791 // coalescing is probably profitable.
792 if (SrcRCWeight.RegWeight > NewRCWeight.RegWeight)
794 if (DstRCWeight.RegWeight > NewRCWeight.RegWeight)
797 // If the register allocator isn't constrained, we can always allow coalescing
798 // unfortunately we don't know yet if we will be constrained.
799 // The goal of this heuristic is to restrict how many expensive registers
800 // we allow to coalesce in a given basic block.
801 auto AFI = MF->getInfo<ARMFunctionInfo>();
802 auto It = AFI->getCoalescedWeight(MBB);
804 DEBUG(dbgs() << "\tARM::shouldCoalesce - Coalesced Weight: "
805 << It->second << "\n");
806 DEBUG(dbgs() << "\tARM::shouldCoalesce - Reg Weight: "
807 << NewRCWeight.RegWeight << "\n");
809 // This number is the largest round number that which meets the criteria:
810 // (1) addresses PR18825
811 // (2) generates better code in some test cases (like vldm-shed-a9.ll)
812 // (3) Doesn't regress any test cases (in-tree, test-suite, and SPEC)
813 // In practice the SizeMultiplier will only factor in for straight line code
814 // that uses a lot of NEON vectors, which isn't terribly common.
815 unsigned SizeMultiplier = MBB->size()/100;
816 SizeMultiplier = SizeMultiplier ? SizeMultiplier : 1;
817 if (It->second < NewRCWeight.WeightLimit * SizeMultiplier) {
818 It->second += NewRCWeight.RegWeight;