1 //===- ARMBaseRegisterInfo.cpp - ARM Register Information -------*- C++ -*-===//
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 //===----------------------------------------------------------------------===//
15 #include "ARMAddressingModes.h"
16 #include "ARMBaseInstrInfo.h"
17 #include "ARMBaseRegisterInfo.h"
18 #include "ARMInstrInfo.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMSubtarget.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/LLVMContext.h"
25 #include "llvm/CodeGen/MachineConstantPool.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineLocation.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/RegisterScavenging.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetFrameInfo.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/ADT/BitVector.h"
39 #include "llvm/ADT/SmallVector.h"
40 #include "llvm/Support/CommandLine.h"
44 ForceAllBaseRegAlloc("arm-force-base-reg-alloc", cl::Hidden, cl::init(false),
45 cl::desc("Force use of virtual base registers for stack load/store"));
47 EnableLocalStackAlloc("enable-local-stack-alloc", cl::init(true), cl::Hidden,
48 cl::desc("Enable pre-regalloc stack frame index allocation"));
54 EnableBasePointer("arm-use-base-pointer", cl::Hidden, cl::init(true),
55 cl::desc("Enable use of a base pointer for complex stack frames"));
58 ARMBaseRegisterInfo::ARMBaseRegisterInfo(const ARMBaseInstrInfo &tii,
59 const ARMSubtarget &sti)
60 : ARMGenRegisterInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
62 FramePtr((STI.isTargetDarwin() || STI.isThumb()) ? ARM::R7 : ARM::R11),
67 ARMBaseRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
68 static const unsigned CalleeSavedRegs[] = {
69 ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8,
70 ARM::R7, ARM::R6, ARM::R5, ARM::R4,
72 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
73 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
77 static const unsigned DarwinCalleeSavedRegs[] = {
78 // Darwin ABI deviates from ARM standard ABI. R9 is not a callee-saved
80 ARM::LR, ARM::R7, ARM::R6, ARM::R5, ARM::R4,
81 ARM::R11, ARM::R10, ARM::R8,
83 ARM::D15, ARM::D14, ARM::D13, ARM::D12,
84 ARM::D11, ARM::D10, ARM::D9, ARM::D8,
87 return STI.isTargetDarwin() ? DarwinCalleeSavedRegs : CalleeSavedRegs;
90 BitVector ARMBaseRegisterInfo::
91 getReservedRegs(const MachineFunction &MF) const {
92 // FIXME: avoid re-calculating this everytime.
93 BitVector Reserved(getNumRegs());
94 Reserved.set(ARM::SP);
95 Reserved.set(ARM::PC);
96 Reserved.set(ARM::FPSCR);
98 Reserved.set(FramePtr);
99 if (hasBasePointer(MF))
100 Reserved.set(BasePtr);
101 // Some targets reserve R9.
102 if (STI.isR9Reserved())
103 Reserved.set(ARM::R9);
107 bool ARMBaseRegisterInfo::isReservedReg(const MachineFunction &MF,
108 unsigned Reg) const {
115 if (hasBasePointer(MF))
120 if (FramePtr == Reg && hasFP(MF))
124 return STI.isR9Reserved();
130 const TargetRegisterClass *
131 ARMBaseRegisterInfo::getMatchingSuperRegClass(const TargetRegisterClass *A,
132 const TargetRegisterClass *B,
133 unsigned SubIdx) const {
141 if (A->getSize() == 8) {
142 if (B == &ARM::SPR_8RegClass)
143 return &ARM::DPR_8RegClass;
144 assert(B == &ARM::SPRRegClass && "Expecting SPR register class!");
145 if (A == &ARM::DPR_8RegClass)
147 return &ARM::DPR_VFP2RegClass;
150 if (A->getSize() == 16) {
151 if (B == &ARM::SPR_8RegClass)
152 return &ARM::QPR_8RegClass;
153 return &ARM::QPR_VFP2RegClass;
156 if (A->getSize() == 32) {
157 if (B == &ARM::SPR_8RegClass)
158 return 0; // Do not allow coalescing!
159 return &ARM::QQPR_VFP2RegClass;
162 assert(A->getSize() == 64 && "Expecting a QQQQ register class!");
163 return 0; // Do not allow coalescing!
170 if (A->getSize() == 16) {
171 if (B == &ARM::DPR_VFP2RegClass)
172 return &ARM::QPR_VFP2RegClass;
173 if (B == &ARM::DPR_8RegClass)
174 return 0; // Do not allow coalescing!
178 if (A->getSize() == 32) {
179 if (B == &ARM::DPR_VFP2RegClass)
180 return &ARM::QQPR_VFP2RegClass;
181 if (B == &ARM::DPR_8RegClass)
182 return 0; // Do not allow coalescing!
186 assert(A->getSize() == 64 && "Expecting a QQQQ register class!");
187 if (B != &ARM::DPRRegClass)
188 return 0; // Do not allow coalescing!
195 // D sub-registers of QQQQ registers.
196 if (A->getSize() == 64 && B == &ARM::DPRRegClass)
198 return 0; // Do not allow coalescing!
204 if (A->getSize() == 32) {
205 if (B == &ARM::QPR_VFP2RegClass)
206 return &ARM::QQPR_VFP2RegClass;
207 if (B == &ARM::QPR_8RegClass)
208 return 0; // Do not allow coalescing!
212 assert(A->getSize() == 64 && "Expecting a QQQQ register class!");
213 if (B == &ARM::QPRRegClass)
215 return 0; // Do not allow coalescing!
219 // Q sub-registers of QQQQ registers.
220 if (A->getSize() == 64 && B == &ARM::QPRRegClass)
222 return 0; // Do not allow coalescing!
229 ARMBaseRegisterInfo::canCombineSubRegIndices(const TargetRegisterClass *RC,
230 SmallVectorImpl<unsigned> &SubIndices,
231 unsigned &NewSubIdx) const {
233 unsigned Size = RC->getSize() * 8;
237 NewSubIdx = 0; // Whole register.
238 unsigned NumRegs = SubIndices.size();
240 // 8 D registers -> 1 QQQQ register.
241 return (Size == 512 &&
242 SubIndices[0] == ARM::dsub_0 &&
243 SubIndices[1] == ARM::dsub_1 &&
244 SubIndices[2] == ARM::dsub_2 &&
245 SubIndices[3] == ARM::dsub_3 &&
246 SubIndices[4] == ARM::dsub_4 &&
247 SubIndices[5] == ARM::dsub_5 &&
248 SubIndices[6] == ARM::dsub_6 &&
249 SubIndices[7] == ARM::dsub_7);
250 } else if (NumRegs == 4) {
251 if (SubIndices[0] == ARM::qsub_0) {
252 // 4 Q registers -> 1 QQQQ register.
253 return (Size == 512 &&
254 SubIndices[1] == ARM::qsub_1 &&
255 SubIndices[2] == ARM::qsub_2 &&
256 SubIndices[3] == ARM::qsub_3);
257 } else if (SubIndices[0] == ARM::dsub_0) {
258 // 4 D registers -> 1 QQ register.
260 SubIndices[1] == ARM::dsub_1 &&
261 SubIndices[2] == ARM::dsub_2 &&
262 SubIndices[3] == ARM::dsub_3) {
264 NewSubIdx = ARM::qqsub_0;
267 } else if (SubIndices[0] == ARM::dsub_4) {
268 // 4 D registers -> 1 QQ register (2nd).
270 SubIndices[1] == ARM::dsub_5 &&
271 SubIndices[2] == ARM::dsub_6 &&
272 SubIndices[3] == ARM::dsub_7) {
273 NewSubIdx = ARM::qqsub_1;
276 } else if (SubIndices[0] == ARM::ssub_0) {
277 // 4 S registers -> 1 Q register.
279 SubIndices[1] == ARM::ssub_1 &&
280 SubIndices[2] == ARM::ssub_2 &&
281 SubIndices[3] == ARM::ssub_3) {
283 NewSubIdx = ARM::qsub_0;
287 } else if (NumRegs == 2) {
288 if (SubIndices[0] == ARM::qsub_0) {
289 // 2 Q registers -> 1 QQ register.
290 if (Size >= 256 && SubIndices[1] == ARM::qsub_1) {
292 NewSubIdx = ARM::qqsub_0;
295 } else if (SubIndices[0] == ARM::qsub_2) {
296 // 2 Q registers -> 1 QQ register (2nd).
297 if (Size == 512 && SubIndices[1] == ARM::qsub_3) {
298 NewSubIdx = ARM::qqsub_1;
301 } else if (SubIndices[0] == ARM::dsub_0) {
302 // 2 D registers -> 1 Q register.
303 if (Size >= 128 && SubIndices[1] == ARM::dsub_1) {
305 NewSubIdx = ARM::qsub_0;
308 } else if (SubIndices[0] == ARM::dsub_2) {
309 // 2 D registers -> 1 Q register (2nd).
310 if (Size >= 256 && SubIndices[1] == ARM::dsub_3) {
311 NewSubIdx = ARM::qsub_1;
314 } else if (SubIndices[0] == ARM::dsub_4) {
315 // 2 D registers -> 1 Q register (3rd).
316 if (Size == 512 && SubIndices[1] == ARM::dsub_5) {
317 NewSubIdx = ARM::qsub_2;
320 } else if (SubIndices[0] == ARM::dsub_6) {
321 // 2 D registers -> 1 Q register (3rd).
322 if (Size == 512 && SubIndices[1] == ARM::dsub_7) {
323 NewSubIdx = ARM::qsub_3;
326 } else if (SubIndices[0] == ARM::ssub_0) {
327 // 2 S registers -> 1 D register.
328 if (SubIndices[1] == ARM::ssub_1) {
330 NewSubIdx = ARM::dsub_0;
333 } else if (SubIndices[0] == ARM::ssub_2) {
334 // 2 S registers -> 1 D register (2nd).
335 if (Size >= 128 && SubIndices[1] == ARM::ssub_3) {
336 NewSubIdx = ARM::dsub_1;
345 const TargetRegisterClass *
346 ARMBaseRegisterInfo::getPointerRegClass(unsigned Kind) const {
347 return ARM::GPRRegisterClass;
350 /// getAllocationOrder - Returns the register allocation order for a specified
351 /// register class in the form of a pair of TargetRegisterClass iterators.
352 std::pair<TargetRegisterClass::iterator,TargetRegisterClass::iterator>
353 ARMBaseRegisterInfo::getAllocationOrder(const TargetRegisterClass *RC,
354 unsigned HintType, unsigned HintReg,
355 const MachineFunction &MF) const {
356 // Alternative register allocation orders when favoring even / odd registers
357 // of register pairs.
359 // No FP, R9 is available.
360 static const unsigned GPREven1[] = {
361 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8, ARM::R10,
362 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7,
365 static const unsigned GPROdd1[] = {
366 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R9, ARM::R11,
367 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6,
371 // FP is R7, R9 is available.
372 static const unsigned GPREven2[] = {
373 ARM::R0, ARM::R2, ARM::R4, ARM::R8, ARM::R10,
374 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6,
377 static const unsigned GPROdd2[] = {
378 ARM::R1, ARM::R3, ARM::R5, ARM::R9, ARM::R11,
379 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6,
383 // FP is R11, R9 is available.
384 static const unsigned GPREven3[] = {
385 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8,
386 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7,
389 static const unsigned GPROdd3[] = {
390 ARM::R1, ARM::R3, ARM::R5, ARM::R6, ARM::R9,
391 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R7,
395 // No FP, R9 is not available.
396 static const unsigned GPREven4[] = {
397 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R10,
398 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8,
401 static const unsigned GPROdd4[] = {
402 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R11,
403 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8,
407 // FP is R7, R9 is not available.
408 static const unsigned GPREven5[] = {
409 ARM::R0, ARM::R2, ARM::R4, ARM::R10,
410 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6, ARM::R8,
413 static const unsigned GPROdd5[] = {
414 ARM::R1, ARM::R3, ARM::R5, ARM::R11,
415 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8,
419 // FP is R11, R9 is not available.
420 static const unsigned GPREven6[] = {
421 ARM::R0, ARM::R2, ARM::R4, ARM::R6,
422 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8
424 static const unsigned GPROdd6[] = {
425 ARM::R1, ARM::R3, ARM::R5, ARM::R7,
426 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8
430 if (HintType == ARMRI::RegPairEven) {
431 if (isPhysicalRegister(HintReg) && getRegisterPairEven(HintReg, MF) == 0)
432 // It's no longer possible to fulfill this hint. Return the default
434 return std::make_pair(RC->allocation_order_begin(MF),
435 RC->allocation_order_end(MF));
438 if (!STI.isR9Reserved())
439 return std::make_pair(GPREven1,
440 GPREven1 + (sizeof(GPREven1)/sizeof(unsigned)));
442 return std::make_pair(GPREven4,
443 GPREven4 + (sizeof(GPREven4)/sizeof(unsigned)));
444 } else if (FramePtr == ARM::R7) {
445 if (!STI.isR9Reserved())
446 return std::make_pair(GPREven2,
447 GPREven2 + (sizeof(GPREven2)/sizeof(unsigned)));
449 return std::make_pair(GPREven5,
450 GPREven5 + (sizeof(GPREven5)/sizeof(unsigned)));
451 } else { // FramePtr == ARM::R11
452 if (!STI.isR9Reserved())
453 return std::make_pair(GPREven3,
454 GPREven3 + (sizeof(GPREven3)/sizeof(unsigned)));
456 return std::make_pair(GPREven6,
457 GPREven6 + (sizeof(GPREven6)/sizeof(unsigned)));
459 } else if (HintType == ARMRI::RegPairOdd) {
460 if (isPhysicalRegister(HintReg) && getRegisterPairOdd(HintReg, MF) == 0)
461 // It's no longer possible to fulfill this hint. Return the default
463 return std::make_pair(RC->allocation_order_begin(MF),
464 RC->allocation_order_end(MF));
467 if (!STI.isR9Reserved())
468 return std::make_pair(GPROdd1,
469 GPROdd1 + (sizeof(GPROdd1)/sizeof(unsigned)));
471 return std::make_pair(GPROdd4,
472 GPROdd4 + (sizeof(GPROdd4)/sizeof(unsigned)));
473 } else if (FramePtr == ARM::R7) {
474 if (!STI.isR9Reserved())
475 return std::make_pair(GPROdd2,
476 GPROdd2 + (sizeof(GPROdd2)/sizeof(unsigned)));
478 return std::make_pair(GPROdd5,
479 GPROdd5 + (sizeof(GPROdd5)/sizeof(unsigned)));
480 } else { // FramePtr == ARM::R11
481 if (!STI.isR9Reserved())
482 return std::make_pair(GPROdd3,
483 GPROdd3 + (sizeof(GPROdd3)/sizeof(unsigned)));
485 return std::make_pair(GPROdd6,
486 GPROdd6 + (sizeof(GPROdd6)/sizeof(unsigned)));
489 return std::make_pair(RC->allocation_order_begin(MF),
490 RC->allocation_order_end(MF));
493 /// ResolveRegAllocHint - Resolves the specified register allocation hint
494 /// to a physical register. Returns the physical register if it is successful.
496 ARMBaseRegisterInfo::ResolveRegAllocHint(unsigned Type, unsigned Reg,
497 const MachineFunction &MF) const {
498 if (Reg == 0 || !isPhysicalRegister(Reg))
502 else if (Type == (unsigned)ARMRI::RegPairOdd)
504 return getRegisterPairOdd(Reg, MF);
505 else if (Type == (unsigned)ARMRI::RegPairEven)
507 return getRegisterPairEven(Reg, MF);
512 ARMBaseRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg,
513 MachineFunction &MF) const {
514 MachineRegisterInfo *MRI = &MF.getRegInfo();
515 std::pair<unsigned, unsigned> Hint = MRI->getRegAllocationHint(Reg);
516 if ((Hint.first == (unsigned)ARMRI::RegPairOdd ||
517 Hint.first == (unsigned)ARMRI::RegPairEven) &&
518 Hint.second && TargetRegisterInfo::isVirtualRegister(Hint.second)) {
519 // If 'Reg' is one of the even / odd register pair and it's now changed
520 // (e.g. coalesced) into a different register. The other register of the
521 // pair allocation hint must be updated to reflect the relationship
523 unsigned OtherReg = Hint.second;
524 Hint = MRI->getRegAllocationHint(OtherReg);
525 if (Hint.second == Reg)
526 // Make sure the pair has not already divorced.
527 MRI->setRegAllocationHint(OtherReg, Hint.first, NewReg);
531 /// hasFP - Return true if the specified function should have a dedicated frame
532 /// pointer register. This is true if the function has variable sized allocas
533 /// or if frame pointer elimination is disabled.
535 bool ARMBaseRegisterInfo::hasFP(const MachineFunction &MF) const {
536 // Mac OS X requires FP not to be clobbered for backtracing purpose.
537 if (STI.isTargetDarwin())
540 const MachineFrameInfo *MFI = MF.getFrameInfo();
541 // Always eliminate non-leaf frame pointers.
542 return ((DisableFramePointerElim(MF) && MFI->hasCalls()) ||
543 needsStackRealignment(MF) ||
544 MFI->hasVarSizedObjects() ||
545 MFI->isFrameAddressTaken());
548 bool ARMBaseRegisterInfo::hasBasePointer(const MachineFunction &MF) const {
549 const MachineFrameInfo *MFI = MF.getFrameInfo();
550 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
552 if (!EnableBasePointer)
555 if (needsStackRealignment(MF) && MFI->hasVarSizedObjects())
558 // Thumb has trouble with negative offsets from the FP. Thumb2 has a limited
559 // negative range for ldr/str (255), and thumb1 is positive offsets only.
560 // It's going to be better to use the SP or Base Pointer instead. When there
561 // are variable sized objects, we can't reference off of the SP, so we
562 // reserve a Base Pointer.
563 if (AFI->isThumbFunction() && MFI->hasVarSizedObjects()) {
564 // Conservatively estimate whether the negative offset from the frame
565 // pointer will be sufficient to reach. If a function has a smallish
566 // frame, it's less likely to have lots of spills and callee saved
567 // space, so it's all more likely to be within range of the frame pointer.
568 // If it's wrong, the scavenger will still enable access to work, it just
570 if (AFI->isThumb2Function() && MFI->getLocalFrameSize() < 128)
578 bool ARMBaseRegisterInfo::canRealignStack(const MachineFunction &MF) const {
579 const MachineFrameInfo *MFI = MF.getFrameInfo();
580 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
581 // We can't realign the stack if:
582 // 1. Dynamic stack realignment is explicitly disabled,
583 // 2. This is a Thumb1 function (it's not useful, so we don't bother), or
584 // 3. There are VLAs in the function and the base pointer is disabled.
585 return (RealignStack && !AFI->isThumb1OnlyFunction() &&
586 (!MFI->hasVarSizedObjects() || EnableBasePointer));
589 bool ARMBaseRegisterInfo::
590 needsStackRealignment(const MachineFunction &MF) const {
591 const MachineFrameInfo *MFI = MF.getFrameInfo();
592 const Function *F = MF.getFunction();
593 unsigned StackAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
594 bool requiresRealignment = ((MFI->getLocalFrameMaxAlign() > StackAlign) ||
595 F->hasFnAttr(Attribute::StackAlignment));
597 return requiresRealignment && canRealignStack(MF);
600 bool ARMBaseRegisterInfo::
601 cannotEliminateFrame(const MachineFunction &MF) const {
602 const MachineFrameInfo *MFI = MF.getFrameInfo();
603 if (DisableFramePointerElim(MF) && MFI->adjustsStack())
605 return MFI->hasVarSizedObjects() || MFI->isFrameAddressTaken()
606 || needsStackRealignment(MF);
609 /// estimateStackSize - Estimate and return the size of the frame.
610 static unsigned estimateStackSize(MachineFunction &MF) {
611 const MachineFrameInfo *FFI = MF.getFrameInfo();
613 for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) {
614 int FixedOff = -FFI->getObjectOffset(i);
615 if (FixedOff > Offset) Offset = FixedOff;
617 for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) {
618 if (FFI->isDeadObjectIndex(i))
620 Offset += FFI->getObjectSize(i);
621 unsigned Align = FFI->getObjectAlignment(i);
622 // Adjust to alignment boundary
623 Offset = (Offset+Align-1)/Align*Align;
625 return (unsigned)Offset;
628 /// estimateRSStackSizeLimit - Look at each instruction that references stack
629 /// frames and return the stack size limit beyond which some of these
630 /// instructions will require a scratch register during their expansion later.
632 ARMBaseRegisterInfo::estimateRSStackSizeLimit(MachineFunction &MF) const {
633 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
634 unsigned Limit = (1 << 12) - 1;
635 for (MachineFunction::iterator BB = MF.begin(),E = MF.end(); BB != E; ++BB) {
636 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
638 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
639 if (!I->getOperand(i).isFI()) continue;
641 // When using ADDri to get the address of a stack object, 255 is the
642 // largest offset guaranteed to fit in the immediate offset.
643 if (I->getOpcode() == ARM::ADDri) {
644 Limit = std::min(Limit, (1U << 8) - 1);
648 // Otherwise check the addressing mode.
649 switch (I->getDesc().TSFlags & ARMII::AddrModeMask) {
650 case ARMII::AddrMode3:
651 case ARMII::AddrModeT2_i8:
652 Limit = std::min(Limit, (1U << 8) - 1);
654 case ARMII::AddrMode5:
655 case ARMII::AddrModeT2_i8s4:
656 Limit = std::min(Limit, ((1U << 8) - 1) * 4);
658 case ARMII::AddrModeT2_i12:
659 // i12 supports only positive offset so these will be converted to
660 // i8 opcodes. See llvm::rewriteT2FrameIndex.
661 if (hasFP(MF) && AFI->hasStackFrame())
662 Limit = std::min(Limit, (1U << 8) - 1);
664 case ARMII::AddrMode6:
665 // Addressing mode 6 (load/store) instructions can't encode an
666 // immediate offset for stack references.
671 break; // At most one FI per instruction
679 static unsigned GetFunctionSizeInBytes(const MachineFunction &MF,
680 const ARMBaseInstrInfo &TII) {
682 for (MachineFunction::const_iterator MBBI = MF.begin(), E = MF.end();
684 const MachineBasicBlock &MBB = *MBBI;
685 for (MachineBasicBlock::const_iterator I = MBB.begin(),E = MBB.end();
687 FnSize += TII.GetInstSizeInBytes(I);
693 ARMBaseRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
694 RegScavenger *RS) const {
695 // This tells PEI to spill the FP as if it is any other callee-save register
696 // to take advantage the eliminateFrameIndex machinery. This also ensures it
697 // is spilled in the order specified by getCalleeSavedRegs() to make it easier
698 // to combine multiple loads / stores.
699 bool CanEliminateFrame = true;
700 bool CS1Spilled = false;
701 bool LRSpilled = false;
702 unsigned NumGPRSpills = 0;
703 SmallVector<unsigned, 4> UnspilledCS1GPRs;
704 SmallVector<unsigned, 4> UnspilledCS2GPRs;
705 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
706 MachineFrameInfo *MFI = MF.getFrameInfo();
708 // Spill R4 if Thumb2 function requires stack realignment - it will be used as
710 // FIXME: It will be better just to find spare register here.
711 if (needsStackRealignment(MF) &&
712 AFI->isThumb2Function())
713 MF.getRegInfo().setPhysRegUsed(ARM::R4);
715 // Spill LR if Thumb1 function uses variable length argument lists.
716 if (AFI->isThumb1OnlyFunction() && AFI->getVarArgsRegSaveSize() > 0)
717 MF.getRegInfo().setPhysRegUsed(ARM::LR);
719 // Spill the BasePtr if it's used.
720 if (hasBasePointer(MF))
721 MF.getRegInfo().setPhysRegUsed(BasePtr);
723 // Don't spill FP if the frame can be eliminated. This is determined
724 // by scanning the callee-save registers to see if any is used.
725 const unsigned *CSRegs = getCalleeSavedRegs();
726 for (unsigned i = 0; CSRegs[i]; ++i) {
727 unsigned Reg = CSRegs[i];
728 bool Spilled = false;
729 if (MF.getRegInfo().isPhysRegUsed(Reg)) {
730 AFI->setCSRegisterIsSpilled(Reg);
732 CanEliminateFrame = false;
734 // Check alias registers too.
735 for (const unsigned *Aliases = getAliasSet(Reg); *Aliases; ++Aliases) {
736 if (MF.getRegInfo().isPhysRegUsed(*Aliases)) {
738 CanEliminateFrame = false;
743 if (!ARM::GPRRegisterClass->contains(Reg))
749 if (!STI.isTargetDarwin()) {
756 // Keep track if LR and any of R4, R5, R6, and R7 is spilled.
771 if (!STI.isTargetDarwin()) {
772 UnspilledCS1GPRs.push_back(Reg);
782 UnspilledCS1GPRs.push_back(Reg);
785 UnspilledCS2GPRs.push_back(Reg);
791 bool ForceLRSpill = false;
792 if (!LRSpilled && AFI->isThumb1OnlyFunction()) {
793 unsigned FnSize = GetFunctionSizeInBytes(MF, TII);
794 // Force LR to be spilled if the Thumb function size is > 2048. This enables
795 // use of BL to implement far jump. If it turns out that it's not needed
796 // then the branch fix up path will undo it.
797 if (FnSize >= (1 << 11)) {
798 CanEliminateFrame = false;
803 // If any of the stack slot references may be out of range of an immediate
804 // offset, make sure a register (or a spill slot) is available for the
805 // register scavenger. Note that if we're indexing off the frame pointer, the
806 // effective stack size is 4 bytes larger since the FP points to the stack
807 // slot of the previous FP. Also, if we have variable sized objects in the
808 // function, stack slot references will often be negative, and some of
809 // our instructions are positive-offset only, so conservatively consider
810 // that case to want a spill slot (or register) as well. Similarly, if
811 // the function adjusts the stack pointer during execution and the
812 // adjustments aren't already part of our stack size estimate, our offset
813 // calculations may be off, so be conservative.
814 // FIXME: We could add logic to be more precise about negative offsets
815 // and which instructions will need a scratch register for them. Is it
816 // worth the effort and added fragility?
819 (estimateStackSize(MF) + ((hasFP(MF) && AFI->hasStackFrame()) ? 4:0) >=
820 estimateRSStackSizeLimit(MF)))
821 || MFI->hasVarSizedObjects()
822 || (MFI->adjustsStack() && !canSimplifyCallFramePseudos(MF));
824 bool ExtraCSSpill = false;
825 if (BigStack || !CanEliminateFrame || cannotEliminateFrame(MF)) {
826 AFI->setHasStackFrame(true);
828 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
829 // Spill LR as well so we can fold BX_RET to the registers restore (LDM).
830 if (!LRSpilled && CS1Spilled) {
831 MF.getRegInfo().setPhysRegUsed(ARM::LR);
832 AFI->setCSRegisterIsSpilled(ARM::LR);
834 UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(),
835 UnspilledCS1GPRs.end(), (unsigned)ARM::LR));
836 ForceLRSpill = false;
841 MF.getRegInfo().setPhysRegUsed(FramePtr);
845 // If stack and double are 8-byte aligned and we are spilling an odd number
846 // of GPRs. Spill one extra callee save GPR so we won't have to pad between
847 // the integer and double callee save areas.
848 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
849 if (TargetAlign == 8 && (NumGPRSpills & 1)) {
850 if (CS1Spilled && !UnspilledCS1GPRs.empty()) {
851 for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) {
852 unsigned Reg = UnspilledCS1GPRs[i];
853 // Don't spill high register if the function is thumb1
854 if (!AFI->isThumb1OnlyFunction() ||
855 isARMLowRegister(Reg) || Reg == ARM::LR) {
856 MF.getRegInfo().setPhysRegUsed(Reg);
857 AFI->setCSRegisterIsSpilled(Reg);
858 if (!isReservedReg(MF, Reg))
863 } else if (!UnspilledCS2GPRs.empty() &&
864 !AFI->isThumb1OnlyFunction()) {
865 unsigned Reg = UnspilledCS2GPRs.front();
866 MF.getRegInfo().setPhysRegUsed(Reg);
867 AFI->setCSRegisterIsSpilled(Reg);
868 if (!isReservedReg(MF, Reg))
873 // Estimate if we might need to scavenge a register at some point in order
874 // to materialize a stack offset. If so, either spill one additional
875 // callee-saved register or reserve a special spill slot to facilitate
876 // register scavenging. Thumb1 needs a spill slot for stack pointer
877 // adjustments also, even when the frame itself is small.
878 if (BigStack && !ExtraCSSpill) {
879 // If any non-reserved CS register isn't spilled, just spill one or two
880 // extra. That should take care of it!
881 unsigned NumExtras = TargetAlign / 4;
882 SmallVector<unsigned, 2> Extras;
883 while (NumExtras && !UnspilledCS1GPRs.empty()) {
884 unsigned Reg = UnspilledCS1GPRs.back();
885 UnspilledCS1GPRs.pop_back();
886 if (!isReservedReg(MF, Reg) &&
887 (!AFI->isThumb1OnlyFunction() || isARMLowRegister(Reg) ||
889 Extras.push_back(Reg);
893 // For non-Thumb1 functions, also check for hi-reg CS registers
894 if (!AFI->isThumb1OnlyFunction()) {
895 while (NumExtras && !UnspilledCS2GPRs.empty()) {
896 unsigned Reg = UnspilledCS2GPRs.back();
897 UnspilledCS2GPRs.pop_back();
898 if (!isReservedReg(MF, Reg)) {
899 Extras.push_back(Reg);
904 if (Extras.size() && NumExtras == 0) {
905 for (unsigned i = 0, e = Extras.size(); i != e; ++i) {
906 MF.getRegInfo().setPhysRegUsed(Extras[i]);
907 AFI->setCSRegisterIsSpilled(Extras[i]);
909 } else if (!AFI->isThumb1OnlyFunction()) {
910 // note: Thumb1 functions spill to R12, not the stack. Reserve a slot
911 // closest to SP or frame pointer.
912 const TargetRegisterClass *RC = ARM::GPRRegisterClass;
913 RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
921 MF.getRegInfo().setPhysRegUsed(ARM::LR);
922 AFI->setCSRegisterIsSpilled(ARM::LR);
923 AFI->setLRIsSpilledForFarJump(true);
927 unsigned ARMBaseRegisterInfo::getRARegister() const {
932 ARMBaseRegisterInfo::getFrameRegister(const MachineFunction &MF) const {
938 // Provide a base+offset reference to an FI slot for debug info. It's the
939 // same as what we use for resolving the code-gen references for now.
940 // FIXME: This can go wrong when references are SP-relative and simple call
941 // frames aren't used.
943 ARMBaseRegisterInfo::getFrameIndexReference(const MachineFunction &MF, int FI,
944 unsigned &FrameReg) const {
945 return ResolveFrameIndexReference(MF, FI, FrameReg, 0);
949 ARMBaseRegisterInfo::ResolveFrameIndexReference(const MachineFunction &MF,
953 const MachineFrameInfo *MFI = MF.getFrameInfo();
954 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
955 int Offset = MFI->getObjectOffset(FI) + MFI->getStackSize();
956 int FPOffset = Offset - AFI->getFramePtrSpillOffset();
957 bool isFixed = MFI->isFixedObjectIndex(FI);
961 if (AFI->isGPRCalleeSavedArea1Frame(FI))
962 return Offset - AFI->getGPRCalleeSavedArea1Offset();
963 else if (AFI->isGPRCalleeSavedArea2Frame(FI))
964 return Offset - AFI->getGPRCalleeSavedArea2Offset();
965 else if (AFI->isDPRCalleeSavedAreaFrame(FI))
966 return Offset - AFI->getDPRCalleeSavedAreaOffset();
968 // When dynamically realigning the stack, use the frame pointer for
969 // parameters, and the stack/base pointer for locals.
970 if (needsStackRealignment(MF)) {
971 assert (hasFP(MF) && "dynamic stack realignment without a FP!");
973 FrameReg = getFrameRegister(MF);
975 } else if (MFI->hasVarSizedObjects()) {
976 assert(hasBasePointer(MF) &&
977 "VLAs and dynamic stack alignment, but missing base pointer!");
983 // If there is a frame pointer, use it when we can.
984 if (hasFP(MF) && AFI->hasStackFrame()) {
985 // Use frame pointer to reference fixed objects. Use it for locals if
986 // there are VLAs (and thus the SP isn't reliable as a base).
987 if (isFixed || (MFI->hasVarSizedObjects() && !hasBasePointer(MF))) {
988 FrameReg = getFrameRegister(MF);
990 } else if (MFI->hasVarSizedObjects()) {
991 assert(hasBasePointer(MF) && "missing base pointer!");
992 // Use the base register since we have it.
994 } else if (AFI->isThumb2Function()) {
995 // In Thumb2 mode, the negative offset is very limited. Try to avoid
996 // out of range references.
997 if (FPOffset >= -255 && FPOffset < 0) {
998 FrameReg = getFrameRegister(MF);
1001 } else if (Offset > (FPOffset < 0 ? -FPOffset : FPOffset)) {
1002 // Otherwise, use SP or FP, whichever is closer to the stack slot.
1003 FrameReg = getFrameRegister(MF);
1007 // Use the base pointer if we have one.
1008 if (hasBasePointer(MF))
1014 ARMBaseRegisterInfo::getFrameIndexOffset(const MachineFunction &MF,
1017 return getFrameIndexReference(MF, FI, FrameReg);
1020 unsigned ARMBaseRegisterInfo::getEHExceptionRegister() const {
1021 llvm_unreachable("What is the exception register");
1025 unsigned ARMBaseRegisterInfo::getEHHandlerRegister() const {
1026 llvm_unreachable("What is the exception handler register");
1030 int ARMBaseRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const {
1031 return ARMGenRegisterInfo::getDwarfRegNumFull(RegNum, 0);
1034 unsigned ARMBaseRegisterInfo::getRegisterPairEven(unsigned Reg,
1035 const MachineFunction &MF) const {
1038 // Return 0 if either register of the pair is a special register.
1047 return (isReservedReg(MF, ARM::R7) || isReservedReg(MF, ARM::R6))
1050 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R8;
1052 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R10;
1124 unsigned ARMBaseRegisterInfo::getRegisterPairOdd(unsigned Reg,
1125 const MachineFunction &MF) const {
1128 // Return 0 if either register of the pair is a special register.
1137 return (isReservedReg(MF, ARM::R7) || isReservedReg(MF, ARM::R6))
1140 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R9;
1142 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R11;
1214 /// emitLoadConstPool - Emits a load from constpool to materialize the
1215 /// specified immediate.
1216 void ARMBaseRegisterInfo::
1217 emitLoadConstPool(MachineBasicBlock &MBB,
1218 MachineBasicBlock::iterator &MBBI,
1220 unsigned DestReg, unsigned SubIdx, int Val,
1221 ARMCC::CondCodes Pred,
1222 unsigned PredReg) const {
1223 MachineFunction &MF = *MBB.getParent();
1224 MachineConstantPool *ConstantPool = MF.getConstantPool();
1226 ConstantInt::get(Type::getInt32Ty(MF.getFunction()->getContext()), Val);
1227 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 4);
1229 BuildMI(MBB, MBBI, dl, TII.get(ARM::LDRcp))
1230 .addReg(DestReg, getDefRegState(true), SubIdx)
1231 .addConstantPoolIndex(Idx)
1232 .addReg(0).addImm(0).addImm(Pred).addReg(PredReg);
1235 bool ARMBaseRegisterInfo::
1236 requiresRegisterScavenging(const MachineFunction &MF) const {
1240 bool ARMBaseRegisterInfo::
1241 requiresFrameIndexScavenging(const MachineFunction &MF) const {
1245 bool ARMBaseRegisterInfo::
1246 requiresVirtualBaseRegisters(const MachineFunction &MF) const {
1247 return EnableLocalStackAlloc;
1250 // hasReservedCallFrame - Under normal circumstances, when a frame pointer is
1251 // not required, we reserve argument space for call sites in the function
1252 // immediately on entry to the current function. This eliminates the need for
1253 // add/sub sp brackets around call sites. Returns true if the call frame is
1254 // included as part of the stack frame.
1255 bool ARMBaseRegisterInfo::
1256 hasReservedCallFrame(const MachineFunction &MF) const {
1257 const MachineFrameInfo *FFI = MF.getFrameInfo();
1258 unsigned CFSize = FFI->getMaxCallFrameSize();
1259 // It's not always a good idea to include the call frame as part of the
1260 // stack frame. ARM (especially Thumb) has small immediate offset to
1261 // address the stack frame. So a large call frame can cause poor codegen
1262 // and may even makes it impossible to scavenge a register.
1263 if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12
1266 return !MF.getFrameInfo()->hasVarSizedObjects();
1269 // canSimplifyCallFramePseudos - If there is a reserved call frame, the
1270 // call frame pseudos can be simplified. Unlike most targets, having a FP
1271 // is not sufficient here since we still may reference some objects via SP
1272 // even when FP is available in Thumb2 mode.
1273 bool ARMBaseRegisterInfo::
1274 canSimplifyCallFramePseudos(const MachineFunction &MF) const {
1275 return hasReservedCallFrame(MF) || MF.getFrameInfo()->hasVarSizedObjects();
1279 emitSPUpdate(bool isARM,
1280 MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
1281 DebugLoc dl, const ARMBaseInstrInfo &TII,
1283 ARMCC::CondCodes Pred = ARMCC::AL, unsigned PredReg = 0) {
1285 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, ARM::SP, NumBytes,
1286 Pred, PredReg, TII);
1288 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::SP, ARM::SP, NumBytes,
1289 Pred, PredReg, TII);
1293 void ARMBaseRegisterInfo::
1294 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
1295 MachineBasicBlock::iterator I) const {
1296 if (!hasReservedCallFrame(MF)) {
1297 // If we have alloca, convert as follows:
1298 // ADJCALLSTACKDOWN -> sub, sp, sp, amount
1299 // ADJCALLSTACKUP -> add, sp, sp, amount
1300 MachineInstr *Old = I;
1301 DebugLoc dl = Old->getDebugLoc();
1302 unsigned Amount = Old->getOperand(0).getImm();
1304 // We need to keep the stack aligned properly. To do this, we round the
1305 // amount of space needed for the outgoing arguments up to the next
1306 // alignment boundary.
1307 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1308 Amount = (Amount+Align-1)/Align*Align;
1310 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1311 assert(!AFI->isThumb1OnlyFunction() &&
1312 "This eliminateCallFramePseudoInstr does not support Thumb1!");
1313 bool isARM = !AFI->isThumbFunction();
1315 // Replace the pseudo instruction with a new instruction...
1316 unsigned Opc = Old->getOpcode();
1317 int PIdx = Old->findFirstPredOperandIdx();
1318 ARMCC::CondCodes Pred = (PIdx == -1)
1319 ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(PIdx).getImm();
1320 if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) {
1321 // Note: PredReg is operand 2 for ADJCALLSTACKDOWN.
1322 unsigned PredReg = Old->getOperand(2).getReg();
1323 emitSPUpdate(isARM, MBB, I, dl, TII, -Amount, Pred, PredReg);
1325 // Note: PredReg is operand 3 for ADJCALLSTACKUP.
1326 unsigned PredReg = Old->getOperand(3).getReg();
1327 assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP);
1328 emitSPUpdate(isARM, MBB, I, dl, TII, Amount, Pred, PredReg);
1335 int64_t ARMBaseRegisterInfo::
1336 getFrameIndexInstrOffset(const MachineInstr *MI, int Idx) const {
1337 const TargetInstrDesc &Desc = MI->getDesc();
1338 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1339 int64_t InstrOffs = 0;;
1341 unsigned ImmIdx = 0;
1343 case ARMII::AddrModeT2_i8:
1344 case ARMII::AddrModeT2_i12:
1345 // i8 supports only negative, and i12 supports only positive, so
1346 // based on Offset sign, consider the appropriate instruction
1347 InstrOffs = MI->getOperand(Idx+1).getImm();
1350 case ARMII::AddrMode5: {
1351 // VFP address mode.
1352 const MachineOperand &OffOp = MI->getOperand(Idx+1);
1353 InstrOffs = ARM_AM::getAM5Offset(OffOp.getImm());
1354 if (ARM_AM::getAM5Op(OffOp.getImm()) == ARM_AM::sub)
1355 InstrOffs = -InstrOffs;
1359 case ARMII::AddrMode2: {
1361 InstrOffs = ARM_AM::getAM2Offset(MI->getOperand(ImmIdx).getImm());
1362 if (ARM_AM::getAM2Op(MI->getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1363 InstrOffs = -InstrOffs;
1366 case ARMII::AddrMode3: {
1368 InstrOffs = ARM_AM::getAM3Offset(MI->getOperand(ImmIdx).getImm());
1369 if (ARM_AM::getAM3Op(MI->getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1370 InstrOffs = -InstrOffs;
1373 case ARMII::AddrModeT1_s: {
1375 InstrOffs = MI->getOperand(ImmIdx).getImm();
1380 llvm_unreachable("Unsupported addressing mode!");
1384 return InstrOffs * Scale;
1387 /// needsFrameBaseReg - Returns true if the instruction's frame index
1388 /// reference would be better served by a base register other than FP
1389 /// or SP. Used by LocalStackFrameAllocation to determine which frame index
1390 /// references it should create new base registers for.
1391 bool ARMBaseRegisterInfo::
1392 needsFrameBaseReg(MachineInstr *MI, int64_t Offset) const {
1393 for (unsigned i = 0; !MI->getOperand(i).isFI(); ++i) {
1394 assert(i < MI->getNumOperands() &&"Instr doesn't have FrameIndex operand!");
1397 // It's the load/store FI references that cause issues, as it can be difficult
1398 // to materialize the offset if it won't fit in the literal field. Estimate
1399 // based on the size of the local frame and some conservative assumptions
1400 // about the rest of the stack frame (note, this is pre-regalloc, so
1401 // we don't know everything for certain yet) whether this offset is likely
1402 // to be out of range of the immediate. Return true if so.
1404 // We only generate virtual base registers for loads and stores, so
1405 // return false for everything else.
1406 unsigned Opc = MI->getOpcode();
1408 case ARM::LDR: case ARM::LDRH: case ARM::LDRB:
1409 case ARM::STR: case ARM::STRH: case ARM::STRB:
1410 case ARM::t2LDRi12: case ARM::t2LDRi8:
1411 case ARM::t2STRi12: case ARM::t2STRi8:
1412 case ARM::VLDRS: case ARM::VLDRD:
1413 case ARM::VSTRS: case ARM::VSTRD:
1414 case ARM::tSTRspi: case ARM::tLDRspi:
1415 if (ForceAllBaseRegAlloc)
1422 // Without a virtual base register, if the function has variable sized
1423 // objects, all fixed-size local references will be via the frame pointer,
1424 // Approximate the offset and see if it's legal for the instruction.
1425 // Note that the incoming offset is based on the SP value at function entry,
1426 // so it'll be negative.
1427 MachineFunction &MF = *MI->getParent()->getParent();
1428 MachineFrameInfo *MFI = MF.getFrameInfo();
1429 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1431 // Estimate an offset from the frame pointer.
1432 // Conservatively assume all callee-saved registers get pushed. R4-R6
1433 // will be earlier than the FP, so we ignore those.
1435 int64_t FPOffset = Offset - 8;
1436 // ARM and Thumb2 functions also need to consider R8-R11 and D8-D15
1437 if (!AFI->isThumbFunction() || !AFI->isThumb1OnlyFunction())
1439 // Estimate an offset from the stack pointer.
1440 // The incoming offset is relating to the SP at the start of the function,
1441 // but when we access the local it'll be relative to the SP after local
1442 // allocation, so adjust our SP-relative offset by that allocation size.
1444 Offset += MFI->getLocalFrameSize();
1445 // Assume that we'll have at least some spill slots allocated.
1446 // FIXME: This is a total SWAG number. We should run some statistics
1447 // and pick a real one.
1448 Offset += 128; // 128 bytes of spill slots
1450 // If there is a frame pointer, try using it.
1451 // The FP is only available if there is no dynamic realignment. We
1452 // don't know for sure yet whether we'll need that, so we guess based
1453 // on whether there are any local variables that would trigger it.
1454 unsigned StackAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
1456 !((MFI->getLocalFrameMaxAlign() > StackAlign) && canRealignStack(MF))) {
1457 if (isFrameOffsetLegal(MI, FPOffset))
1460 // If we can reference via the stack pointer, try that.
1461 // FIXME: This (and the code that resolves the references) can be improved
1462 // to only disallow SP relative references in the live range of
1463 // the VLA(s). In practice, it's unclear how much difference that
1464 // would make, but it may be worth doing.
1465 if (!MFI->hasVarSizedObjects() && isFrameOffsetLegal(MI, Offset))
1468 // The offset likely isn't legal, we want to allocate a virtual base register.
1472 /// materializeFrameBaseRegister - Insert defining instruction(s) for
1473 /// BaseReg to be a pointer to FrameIdx before insertion point I.
1474 void ARMBaseRegisterInfo::
1475 materializeFrameBaseRegister(MachineBasicBlock::iterator I, unsigned BaseReg,
1476 int FrameIdx, int64_t Offset) const {
1477 ARMFunctionInfo *AFI =
1478 I->getParent()->getParent()->getInfo<ARMFunctionInfo>();
1479 unsigned ADDriOpc = !AFI->isThumbFunction() ? ARM::ADDri :
1480 (AFI->isThumb1OnlyFunction() ? ARM::tADDrSPi : ARM::t2ADDri);
1482 MachineInstrBuilder MIB =
1483 BuildMI(*I->getParent(), I, I->getDebugLoc(), TII.get(ADDriOpc), BaseReg)
1484 .addFrameIndex(FrameIdx).addImm(Offset);
1485 if (!AFI->isThumb1OnlyFunction())
1486 AddDefaultCC(AddDefaultPred(MIB));
1490 ARMBaseRegisterInfo::resolveFrameIndex(MachineBasicBlock::iterator I,
1491 unsigned BaseReg, int64_t Offset) const {
1492 MachineInstr &MI = *I;
1493 MachineBasicBlock &MBB = *MI.getParent();
1494 MachineFunction &MF = *MBB.getParent();
1495 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1496 int Off = Offset; // ARM doesn't need the general 64-bit offsets
1499 assert(!AFI->isThumb1OnlyFunction() &&
1500 "This resolveFrameIndex does not support Thumb1!");
1502 while (!MI.getOperand(i).isFI()) {
1504 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
1507 if (!AFI->isThumbFunction())
1508 Done = rewriteARMFrameIndex(MI, i, BaseReg, Off, TII);
1510 assert(AFI->isThumb2Function());
1511 Done = rewriteT2FrameIndex(MI, i, BaseReg, Off, TII);
1513 assert (Done && "Unable to resolve frame index!");
1516 bool ARMBaseRegisterInfo::isFrameOffsetLegal(const MachineInstr *MI,
1517 int64_t Offset) const {
1518 const TargetInstrDesc &Desc = MI->getDesc();
1519 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1522 while (!MI->getOperand(i).isFI()) {
1524 assert(i < MI->getNumOperands() &&"Instr doesn't have FrameIndex operand!");
1527 // AddrMode4 and AddrMode6 cannot handle any offset.
1528 if (AddrMode == ARMII::AddrMode4 || AddrMode == ARMII::AddrMode6)
1531 unsigned NumBits = 0;
1533 bool isSigned = true;
1535 case ARMII::AddrModeT2_i8:
1536 case ARMII::AddrModeT2_i12:
1537 // i8 supports only negative, and i12 supports only positive, so
1538 // based on Offset sign, consider the appropriate instruction
1547 case ARMII::AddrMode5:
1548 // VFP address mode.
1552 case ARMII::AddrMode2:
1555 case ARMII::AddrMode3:
1558 case ARMII::AddrModeT1_s:
1564 llvm_unreachable("Unsupported addressing mode!");
1568 Offset += getFrameIndexInstrOffset(MI, i);
1569 // Make sure the offset is encodable for instructions that scale the
1571 if ((Offset & (Scale-1)) != 0)
1574 if (isSigned && Offset < 0)
1577 unsigned Mask = (1 << NumBits) - 1;
1578 if ((unsigned)Offset <= Mask * Scale)
1585 ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
1586 int SPAdj, RegScavenger *RS) const {
1588 MachineInstr &MI = *II;
1589 MachineBasicBlock &MBB = *MI.getParent();
1590 MachineFunction &MF = *MBB.getParent();
1591 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1592 assert(!AFI->isThumb1OnlyFunction() &&
1593 "This eliminateFrameIndex does not support Thumb1!");
1595 while (!MI.getOperand(i).isFI()) {
1597 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
1600 int FrameIndex = MI.getOperand(i).getIndex();
1603 int Offset = ResolveFrameIndexReference(MF, FrameIndex, FrameReg, SPAdj);
1605 // Special handling of dbg_value instructions.
1606 if (MI.isDebugValue()) {
1607 MI.getOperand(i). ChangeToRegister(FrameReg, false /*isDef*/);
1608 MI.getOperand(i+1).ChangeToImmediate(Offset);
1612 // Modify MI as necessary to handle as much of 'Offset' as possible
1614 if (!AFI->isThumbFunction())
1615 Done = rewriteARMFrameIndex(MI, i, FrameReg, Offset, TII);
1617 assert(AFI->isThumb2Function());
1618 Done = rewriteT2FrameIndex(MI, i, FrameReg, Offset, TII);
1623 // If we get here, the immediate doesn't fit into the instruction. We folded
1624 // as much as possible above, handle the rest, providing a register that is
1627 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode4 ||
1628 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode6) &&
1629 "This code isn't needed if offset already handled!");
1631 unsigned ScratchReg = 0;
1632 int PIdx = MI.findFirstPredOperandIdx();
1633 ARMCC::CondCodes Pred = (PIdx == -1)
1634 ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm();
1635 unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg();
1637 // Must be addrmode4/6.
1638 MI.getOperand(i).ChangeToRegister(FrameReg, false, false, false);
1640 ScratchReg = MF.getRegInfo().createVirtualRegister(ARM::GPRRegisterClass);
1641 if (!AFI->isThumbFunction())
1642 emitARMRegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg,
1643 Offset, Pred, PredReg, TII);
1645 assert(AFI->isThumb2Function());
1646 emitT2RegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg,
1647 Offset, Pred, PredReg, TII);
1649 MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true);
1653 /// Move iterator past the next bunch of callee save load / store ops for
1654 /// the particular spill area (1: integer area 1, 2: integer area 2,
1655 /// 3: fp area, 0: don't care).
1656 static void movePastCSLoadStoreOps(MachineBasicBlock &MBB,
1657 MachineBasicBlock::iterator &MBBI,
1658 int Opc1, int Opc2, unsigned Area,
1659 const ARMSubtarget &STI) {
1660 while (MBBI != MBB.end() &&
1661 ((MBBI->getOpcode() == Opc1) || (MBBI->getOpcode() == Opc2)) &&
1662 MBBI->getOperand(1).isFI()) {
1665 unsigned Category = 0;
1666 switch (MBBI->getOperand(0).getReg()) {
1667 case ARM::R4: case ARM::R5: case ARM::R6: case ARM::R7:
1671 case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11:
1672 Category = STI.isTargetDarwin() ? 2 : 1;
1674 case ARM::D8: case ARM::D9: case ARM::D10: case ARM::D11:
1675 case ARM::D12: case ARM::D13: case ARM::D14: case ARM::D15:
1682 if (Done || Category != Area)
1690 void ARMBaseRegisterInfo::
1691 emitPrologue(MachineFunction &MF) const {
1692 MachineBasicBlock &MBB = MF.front();
1693 MachineBasicBlock::iterator MBBI = MBB.begin();
1694 MachineFrameInfo *MFI = MF.getFrameInfo();
1695 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1696 assert(!AFI->isThumb1OnlyFunction() &&
1697 "This emitPrologue does not support Thumb1!");
1698 bool isARM = !AFI->isThumbFunction();
1699 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1700 unsigned NumBytes = MFI->getStackSize();
1701 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
1702 DebugLoc dl = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
1704 // Determine the sizes of each callee-save spill areas and record which frame
1705 // belongs to which callee-save spill areas.
1706 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
1707 int FramePtrSpillFI = 0;
1709 // Allocate the vararg register save area. This is not counted in NumBytes.
1711 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -VARegSaveSize);
1713 if (!AFI->hasStackFrame()) {
1715 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes);
1719 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1720 unsigned Reg = CSI[i].getReg();
1721 int FI = CSI[i].getFrameIdx();
1728 if (Reg == FramePtr)
1729 FramePtrSpillFI = FI;
1730 AFI->addGPRCalleeSavedArea1Frame(FI);
1737 if (Reg == FramePtr)
1738 FramePtrSpillFI = FI;
1739 if (STI.isTargetDarwin()) {
1740 AFI->addGPRCalleeSavedArea2Frame(FI);
1743 AFI->addGPRCalleeSavedArea1Frame(FI);
1748 AFI->addDPRCalleeSavedAreaFrame(FI);
1753 // Build the new SUBri to adjust SP for integer callee-save spill area 1.
1754 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -GPRCS1Size);
1755 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, ARM::t2STRi12, 1, STI);
1757 // Set FP to point to the stack slot that contains the previous FP.
1758 // For Darwin, FP is R7, which has now been stored in spill area 1.
1759 // Otherwise, if this is not Darwin, all the callee-saved registers go
1760 // into spill area 1, including the FP in R11. In either case, it is
1761 // now safe to emit this assignment.
1762 bool HasFP = hasFP(MF);
1764 unsigned ADDriOpc = !AFI->isThumbFunction() ? ARM::ADDri : ARM::t2ADDri;
1765 MachineInstrBuilder MIB =
1766 BuildMI(MBB, MBBI, dl, TII.get(ADDriOpc), FramePtr)
1767 .addFrameIndex(FramePtrSpillFI).addImm(0);
1768 AddDefaultCC(AddDefaultPred(MIB));
1771 // Build the new SUBri to adjust SP for integer callee-save spill area 2.
1772 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -GPRCS2Size);
1774 // Build the new SUBri to adjust SP for FP callee-save spill area.
1775 movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, ARM::t2STRi12, 2, STI);
1776 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -DPRCSSize);
1778 // Determine starting offsets of spill areas.
1779 unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize);
1780 unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize;
1781 unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size;
1783 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) +
1785 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset);
1786 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
1787 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
1789 movePastCSLoadStoreOps(MBB, MBBI, ARM::VSTRD, 0, 3, STI);
1790 NumBytes = DPRCSOffset;
1792 // Adjust SP after all the callee-save spills.
1793 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes);
1795 AFI->setShouldRestoreSPFromFP(true);
1798 if (STI.isTargetELF() && hasFP(MF)) {
1799 MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() -
1800 AFI->getFramePtrSpillOffset());
1801 AFI->setShouldRestoreSPFromFP(true);
1804 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
1805 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
1806 AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
1808 // If we need dynamic stack realignment, do it here. Be paranoid and make
1809 // sure if we also have VLAs, we have a base pointer for frame access.
1810 if (needsStackRealignment(MF)) {
1811 unsigned MaxAlign = MFI->getMaxAlignment();
1812 assert (!AFI->isThumb1OnlyFunction());
1813 if (!AFI->isThumbFunction()) {
1814 // Emit bic sp, sp, MaxAlign
1815 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl,
1816 TII.get(ARM::BICri), ARM::SP)
1817 .addReg(ARM::SP, RegState::Kill)
1818 .addImm(MaxAlign-1)));
1820 // We cannot use sp as source/dest register here, thus we're emitting the
1821 // following sequence:
1823 // bic r4, r4, MaxAlign
1825 // FIXME: It will be better just to find spare register here.
1826 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVgpr2tgpr), ARM::R4)
1827 .addReg(ARM::SP, RegState::Kill);
1828 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl,
1829 TII.get(ARM::t2BICri), ARM::R4)
1830 .addReg(ARM::R4, RegState::Kill)
1831 .addImm(MaxAlign-1)));
1832 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVtgpr2gpr), ARM::SP)
1833 .addReg(ARM::R4, RegState::Kill);
1836 AFI->setShouldRestoreSPFromFP(true);
1839 // If we need a base pointer, set it up here. It's whatever the value
1840 // of the stack pointer is at this point. Any variable size objects
1841 // will be allocated after this, so we can still use the base pointer
1842 // to reference locals.
1843 if (hasBasePointer(MF)) {
1845 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), BasePtr)
1847 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
1849 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVgpr2gpr), BasePtr)
1853 // If the frame has variable sized objects then the epilogue must restore
1855 if (!AFI->shouldRestoreSPFromFP() && MFI->hasVarSizedObjects())
1856 AFI->setShouldRestoreSPFromFP(true);
1859 static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) {
1860 for (unsigned i = 0; CSRegs[i]; ++i)
1861 if (Reg == CSRegs[i])
1866 static bool isCSRestore(MachineInstr *MI,
1867 const ARMBaseInstrInfo &TII,
1868 const unsigned *CSRegs) {
1869 return ((MI->getOpcode() == (int)ARM::VLDRD ||
1870 MI->getOpcode() == (int)ARM::LDR ||
1871 MI->getOpcode() == (int)ARM::t2LDRi12) &&
1872 MI->getOperand(1).isFI() &&
1873 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs));
1876 void ARMBaseRegisterInfo::
1877 emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const {
1878 MachineBasicBlock::iterator MBBI = prior(MBB.end());
1879 assert(MBBI->getDesc().isReturn() &&
1880 "Can only insert epilog into returning blocks");
1881 unsigned RetOpcode = MBBI->getOpcode();
1882 DebugLoc dl = MBBI->getDebugLoc();
1883 MachineFrameInfo *MFI = MF.getFrameInfo();
1884 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1885 assert(!AFI->isThumb1OnlyFunction() &&
1886 "This emitEpilogue does not support Thumb1!");
1887 bool isARM = !AFI->isThumbFunction();
1889 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
1890 int NumBytes = (int)MFI->getStackSize();
1892 if (!AFI->hasStackFrame()) {
1894 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes);
1896 // Unwind MBBI to point to first LDR / VLDRD.
1897 const unsigned *CSRegs = getCalleeSavedRegs();
1898 if (MBBI != MBB.begin()) {
1901 while (MBBI != MBB.begin() && isCSRestore(MBBI, TII, CSRegs));
1902 if (!isCSRestore(MBBI, TII, CSRegs))
1906 // Move SP to start of FP callee save spill area.
1907 NumBytes -= (AFI->getGPRCalleeSavedArea1Size() +
1908 AFI->getGPRCalleeSavedArea2Size() +
1909 AFI->getDPRCalleeSavedAreaSize());
1911 // Reset SP based on frame pointer only if the stack frame extends beyond
1912 // frame pointer stack slot or target is ELF and the function has FP.
1913 if (AFI->shouldRestoreSPFromFP()) {
1914 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
1917 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, FramePtr, -NumBytes,
1920 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::SP, FramePtr, -NumBytes,
1925 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), ARM::SP)
1926 .addReg(FramePtr).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
1928 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVgpr2gpr), ARM::SP)
1931 } else if (NumBytes)
1932 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes);
1934 // Move SP to start of integer callee save spill area 2.
1935 movePastCSLoadStoreOps(MBB, MBBI, ARM::VLDRD, 0, 3, STI);
1936 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getDPRCalleeSavedAreaSize());
1938 // Move SP to start of integer callee save spill area 1.
1939 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, ARM::t2LDRi12, 2, STI);
1940 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getGPRCalleeSavedArea2Size());
1942 // Move SP to SP upon entry to the function.
1943 movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, ARM::t2LDRi12, 1, STI);
1944 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getGPRCalleeSavedArea1Size());
1947 if (RetOpcode == ARM::TCRETURNdi || RetOpcode == ARM::TCRETURNdiND ||
1948 RetOpcode == ARM::TCRETURNri || RetOpcode == ARM::TCRETURNriND) {
1949 // Tail call return: adjust the stack pointer and jump to callee.
1950 MBBI = prior(MBB.end());
1951 MachineOperand &JumpTarget = MBBI->getOperand(0);
1953 // Jump to label or value in register.
1954 if (RetOpcode == ARM::TCRETURNdi) {
1955 BuildMI(MBB, MBBI, dl,
1956 TII.get(STI.isThumb() ? ARM::TAILJMPdt : ARM::TAILJMPd)).
1957 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(),
1958 JumpTarget.getTargetFlags());
1959 } else if (RetOpcode == ARM::TCRETURNdiND) {
1960 BuildMI(MBB, MBBI, dl,
1961 TII.get(STI.isThumb() ? ARM::TAILJMPdNDt : ARM::TAILJMPdND)).
1962 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(),
1963 JumpTarget.getTargetFlags());
1964 } else if (RetOpcode == ARM::TCRETURNri) {
1965 BuildMI(MBB, MBBI, dl, TII.get(ARM::TAILJMPr)).
1966 addReg(JumpTarget.getReg(), RegState::Kill);
1967 } else if (RetOpcode == ARM::TCRETURNriND) {
1968 BuildMI(MBB, MBBI, dl, TII.get(ARM::TAILJMPrND)).
1969 addReg(JumpTarget.getReg(), RegState::Kill);
1972 MachineInstr *NewMI = prior(MBBI);
1973 for (unsigned i = 1, e = MBBI->getNumOperands(); i != e; ++i)
1974 NewMI->addOperand(MBBI->getOperand(i));
1976 // Delete the pseudo instruction TCRETURN.
1981 emitSPUpdate(isARM, MBB, MBBI, dl, TII, VARegSaveSize);
1984 #include "ARMGenRegisterInfo.inc"