1 //=- AArch64LoadStoreOptimizer.cpp - AArch64 load/store opt. pass -*- 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 a pass that performs load / store related peephole
11 // optimizations. This pass should be run after register allocation.
13 //===----------------------------------------------------------------------===//
15 #include "AArch64InstrInfo.h"
16 #include "AArch64Subtarget.h"
17 #include "MCTargetDesc/AArch64AddressingModes.h"
18 #include "llvm/ADT/BitVector.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/CodeGen/MachineBasicBlock.h"
22 #include "llvm/CodeGen/MachineFunctionPass.h"
23 #include "llvm/CodeGen/MachineInstr.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/ErrorHandling.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include "llvm/Target/TargetInstrInfo.h"
30 #include "llvm/Target/TargetMachine.h"
31 #include "llvm/Target/TargetRegisterInfo.h"
34 #define DEBUG_TYPE "aarch64-ldst-opt"
36 /// AArch64AllocLoadStoreOpt - Post-register allocation pass to combine
37 /// load / store instructions to form ldp / stp instructions.
39 STATISTIC(NumPairCreated, "Number of load/store pair instructions generated");
40 STATISTIC(NumPostFolded, "Number of post-index updates folded");
41 STATISTIC(NumPreFolded, "Number of pre-index updates folded");
42 STATISTIC(NumUnscaledPairCreated,
43 "Number of load/store from unscaled generated");
45 static cl::opt<unsigned> ScanLimit("aarch64-load-store-scan-limit",
46 cl::init(20), cl::Hidden);
48 // Place holder while testing unscaled load/store combining
49 static cl::opt<bool> EnableAArch64UnscaledMemOp(
50 "aarch64-unscaled-mem-op", cl::Hidden,
51 cl::desc("Allow AArch64 unscaled load/store combining"), cl::init(true));
54 void initializeAArch64LoadStoreOptPass(PassRegistry &);
57 #define AARCH64_LOAD_STORE_OPT_NAME "AArch64 load / store optimization pass"
61 typedef struct LdStPairFlags {
62 // If a matching instruction is found, MergeForward is set to true if the
63 // merge is to remove the first instruction and replace the second with
64 // a pair-wise insn, and false if the reverse is true.
67 // SExtIdx gives the index of the result of the load pair that must be
68 // extended. The value of SExtIdx assumes that the paired load produces the
69 // value in this order: (I, returned iterator), i.e., -1 means no value has
70 // to be extended, 0 means I, and 1 means the returned iterator.
73 LdStPairFlags() : MergeForward(false), SExtIdx(-1) {}
75 void setMergeForward(bool V = true) { MergeForward = V; }
76 bool getMergeForward() const { return MergeForward; }
78 void setSExtIdx(int V) { SExtIdx = V; }
79 int getSExtIdx() const { return SExtIdx; }
83 struct AArch64LoadStoreOpt : public MachineFunctionPass {
85 AArch64LoadStoreOpt() : MachineFunctionPass(ID) {
86 initializeAArch64LoadStoreOptPass(*PassRegistry::getPassRegistry());
89 const AArch64InstrInfo *TII;
90 const TargetRegisterInfo *TRI;
92 // Scan the instructions looking for a load/store that can be combined
93 // with the current instruction into a load/store pair.
94 // Return the matching instruction if one is found, else MBB->end().
95 MachineBasicBlock::iterator findMatchingInsn(MachineBasicBlock::iterator I,
98 // Merge the two instructions indicated into a single pair-wise instruction.
99 // If MergeForward is true, erase the first instruction and fold its
100 // operation into the second. If false, the reverse. Return the instruction
101 // following the first instruction (which may change during processing).
102 MachineBasicBlock::iterator
103 mergePairedInsns(MachineBasicBlock::iterator I,
104 MachineBasicBlock::iterator Paired,
105 const LdStPairFlags &Flags);
107 // Scan the instruction list to find a base register update that can
108 // be combined with the current instruction (a load or store) using
109 // pre or post indexed addressing with writeback. Scan forwards.
110 MachineBasicBlock::iterator
111 findMatchingUpdateInsnForward(MachineBasicBlock::iterator I, unsigned Limit,
114 // Scan the instruction list to find a base register update that can
115 // be combined with the current instruction (a load or store) using
116 // pre or post indexed addressing with writeback. Scan backwards.
117 MachineBasicBlock::iterator
118 findMatchingUpdateInsnBackward(MachineBasicBlock::iterator I, unsigned Limit);
120 // Find an instruction that updates the base register of the ld/st
122 bool isMatchingUpdateInsn(MachineInstr *MemMI, MachineInstr *MI,
123 unsigned BaseReg, int Offset);
125 // Merge a pre- or post-index base register update into a ld/st instruction.
126 MachineBasicBlock::iterator
127 mergeUpdateInsn(MachineBasicBlock::iterator I,
128 MachineBasicBlock::iterator Update, bool IsPreIdx);
130 bool optimizeBlock(MachineBasicBlock &MBB);
132 bool runOnMachineFunction(MachineFunction &Fn) override;
134 const char *getPassName() const override {
135 return AARCH64_LOAD_STORE_OPT_NAME;
138 char AArch64LoadStoreOpt::ID = 0;
141 INITIALIZE_PASS(AArch64LoadStoreOpt, "aarch64-ldst-opt",
142 AARCH64_LOAD_STORE_OPT_NAME, false, false)
144 static bool isUnscaledLdSt(unsigned Opc) {
148 case AArch64::STURSi:
149 case AArch64::STURDi:
150 case AArch64::STURQi:
151 case AArch64::STURWi:
152 case AArch64::STURXi:
153 case AArch64::LDURSi:
154 case AArch64::LDURDi:
155 case AArch64::LDURQi:
156 case AArch64::LDURWi:
157 case AArch64::LDURXi:
158 case AArch64::LDURSWi:
163 static bool isUnscaledLdSt(MachineInstr *MI) {
164 return isUnscaledLdSt(MI->getOpcode());
167 // Scaling factor for unscaled load or store.
168 static int getMemScale(MachineInstr *MI) {
169 switch (MI->getOpcode()) {
171 llvm_unreachable("Opcode has unknown size!");
172 case AArch64::LDRSui:
173 case AArch64::LDURSi:
174 case AArch64::LDRSWui:
175 case AArch64::LDURSWi:
176 case AArch64::LDRWui:
177 case AArch64::LDURWi:
178 case AArch64::STRSui:
179 case AArch64::STURSi:
180 case AArch64::STRWui:
181 case AArch64::STURWi:
187 case AArch64::LDRDui:
188 case AArch64::LDURDi:
189 case AArch64::LDRXui:
190 case AArch64::LDURXi:
191 case AArch64::STRDui:
192 case AArch64::STURDi:
193 case AArch64::STRXui:
194 case AArch64::STURXi:
200 case AArch64::LDRQui:
201 case AArch64::LDURQi:
202 case AArch64::STRQui:
203 case AArch64::STURQi:
210 static unsigned getMatchingNonSExtOpcode(unsigned Opc,
211 bool *IsValidLdStrOpc = nullptr) {
213 *IsValidLdStrOpc = true;
217 *IsValidLdStrOpc = false;
219 case AArch64::STRDui:
220 case AArch64::STURDi:
221 case AArch64::STRQui:
222 case AArch64::STURQi:
223 case AArch64::STRWui:
224 case AArch64::STURWi:
225 case AArch64::STRXui:
226 case AArch64::STURXi:
227 case AArch64::LDRDui:
228 case AArch64::LDURDi:
229 case AArch64::LDRQui:
230 case AArch64::LDURQi:
231 case AArch64::LDRWui:
232 case AArch64::LDURWi:
233 case AArch64::LDRXui:
234 case AArch64::LDURXi:
235 case AArch64::STRSui:
236 case AArch64::STURSi:
237 case AArch64::LDRSui:
238 case AArch64::LDURSi:
240 case AArch64::LDRSWui:
241 return AArch64::LDRWui;
242 case AArch64::LDURSWi:
243 return AArch64::LDURWi;
247 static unsigned getMatchingPairOpcode(unsigned Opc) {
250 llvm_unreachable("Opcode has no pairwise equivalent!");
251 case AArch64::STRSui:
252 case AArch64::STURSi:
253 return AArch64::STPSi;
254 case AArch64::STRDui:
255 case AArch64::STURDi:
256 return AArch64::STPDi;
257 case AArch64::STRQui:
258 case AArch64::STURQi:
259 return AArch64::STPQi;
260 case AArch64::STRWui:
261 case AArch64::STURWi:
262 return AArch64::STPWi;
263 case AArch64::STRXui:
264 case AArch64::STURXi:
265 return AArch64::STPXi;
266 case AArch64::LDRSui:
267 case AArch64::LDURSi:
268 return AArch64::LDPSi;
269 case AArch64::LDRDui:
270 case AArch64::LDURDi:
271 return AArch64::LDPDi;
272 case AArch64::LDRQui:
273 case AArch64::LDURQi:
274 return AArch64::LDPQi;
275 case AArch64::LDRWui:
276 case AArch64::LDURWi:
277 return AArch64::LDPWi;
278 case AArch64::LDRXui:
279 case AArch64::LDURXi:
280 return AArch64::LDPXi;
281 case AArch64::LDRSWui:
282 case AArch64::LDURSWi:
283 return AArch64::LDPSWi;
287 static unsigned getPreIndexedOpcode(unsigned Opc) {
290 llvm_unreachable("Opcode has no pre-indexed equivalent!");
291 case AArch64::STRSui:
292 return AArch64::STRSpre;
293 case AArch64::STRDui:
294 return AArch64::STRDpre;
295 case AArch64::STRQui:
296 return AArch64::STRQpre;
297 case AArch64::STRWui:
298 return AArch64::STRWpre;
299 case AArch64::STRXui:
300 return AArch64::STRXpre;
301 case AArch64::LDRSui:
302 return AArch64::LDRSpre;
303 case AArch64::LDRDui:
304 return AArch64::LDRDpre;
305 case AArch64::LDRQui:
306 return AArch64::LDRQpre;
307 case AArch64::LDRWui:
308 return AArch64::LDRWpre;
309 case AArch64::LDRXui:
310 return AArch64::LDRXpre;
311 case AArch64::LDRSWui:
312 return AArch64::LDRSWpre;
314 return AArch64::LDPSpre;
316 return AArch64::LDPDpre;
318 return AArch64::LDPQpre;
320 return AArch64::LDPWpre;
322 return AArch64::LDPXpre;
324 return AArch64::STPSpre;
326 return AArch64::STPDpre;
328 return AArch64::STPQpre;
330 return AArch64::STPWpre;
332 return AArch64::STPXpre;
336 static unsigned getPostIndexedOpcode(unsigned Opc) {
339 llvm_unreachable("Opcode has no post-indexed wise equivalent!");
340 case AArch64::STRSui:
341 return AArch64::STRSpost;
342 case AArch64::STRDui:
343 return AArch64::STRDpost;
344 case AArch64::STRQui:
345 return AArch64::STRQpost;
346 case AArch64::STRWui:
347 return AArch64::STRWpost;
348 case AArch64::STRXui:
349 return AArch64::STRXpost;
350 case AArch64::LDRSui:
351 return AArch64::LDRSpost;
352 case AArch64::LDRDui:
353 return AArch64::LDRDpost;
354 case AArch64::LDRQui:
355 return AArch64::LDRQpost;
356 case AArch64::LDRWui:
357 return AArch64::LDRWpost;
358 case AArch64::LDRXui:
359 return AArch64::LDRXpost;
360 case AArch64::LDRSWui:
361 return AArch64::LDRSWpost;
363 return AArch64::LDPSpost;
365 return AArch64::LDPDpost;
367 return AArch64::LDPQpost;
369 return AArch64::LDPWpost;
371 return AArch64::LDPXpost;
373 return AArch64::STPSpost;
375 return AArch64::STPDpost;
377 return AArch64::STPQpost;
379 return AArch64::STPWpost;
381 return AArch64::STPXpost;
385 static bool isPairedLdSt(const MachineInstr *MI) {
386 switch (MI->getOpcode()) {
403 static const MachineOperand &getLdStRegOp(const MachineInstr *MI,
404 unsigned PairedRegOp = 0) {
405 assert(PairedRegOp < 2 && "Unexpected register operand idx.");
406 unsigned Idx = isPairedLdSt(MI) ? PairedRegOp : 0;
407 return MI->getOperand(Idx);
410 static const MachineOperand &getLdStBaseOp(const MachineInstr *MI) {
411 unsigned Idx = isPairedLdSt(MI) ? 2 : 1;
412 return MI->getOperand(Idx);
415 static const MachineOperand &getLdStOffsetOp(const MachineInstr *MI) {
416 unsigned Idx = isPairedLdSt(MI) ? 3 : 2;
417 return MI->getOperand(Idx);
420 MachineBasicBlock::iterator
421 AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I,
422 MachineBasicBlock::iterator Paired,
423 const LdStPairFlags &Flags) {
424 MachineBasicBlock::iterator NextI = I;
426 // If NextI is the second of the two instructions to be merged, we need
427 // to skip one further. Either way we merge will invalidate the iterator,
428 // and we don't need to scan the new instruction, as it's a pairwise
429 // instruction, which we're not considering for further action anyway.
433 int SExtIdx = Flags.getSExtIdx();
435 SExtIdx == -1 ? I->getOpcode() : getMatchingNonSExtOpcode(I->getOpcode());
436 bool IsUnscaled = isUnscaledLdSt(Opc);
438 IsUnscaled && EnableAArch64UnscaledMemOp ? getMemScale(I) : 1;
440 bool MergeForward = Flags.getMergeForward();
441 unsigned NewOpc = getMatchingPairOpcode(Opc);
442 // Insert our new paired instruction after whichever of the paired
443 // instructions MergeForward indicates.
444 MachineBasicBlock::iterator InsertionPoint = MergeForward ? Paired : I;
445 // Also based on MergeForward is from where we copy the base register operand
446 // so we get the flags compatible with the input code.
447 const MachineOperand &BaseRegOp =
448 MergeForward ? getLdStBaseOp(Paired) : getLdStBaseOp(I);
450 // Which register is Rt and which is Rt2 depends on the offset order.
451 MachineInstr *RtMI, *Rt2MI;
452 if (getLdStOffsetOp(I).getImm() ==
453 getLdStOffsetOp(Paired).getImm() + OffsetStride) {
456 // Here we swapped the assumption made for SExtIdx.
457 // I.e., we turn ldp I, Paired into ldp Paired, I.
458 // Update the index accordingly.
460 SExtIdx = (SExtIdx + 1) % 2;
466 int OffsetImm = getLdStOffsetOp(RtMI).getImm();
467 if (IsUnscaled && EnableAArch64UnscaledMemOp)
468 OffsetImm /= OffsetStride;
470 // Construct the new instruction.
471 MachineInstrBuilder MIB = BuildMI(*I->getParent(), InsertionPoint,
472 I->getDebugLoc(), TII->get(NewOpc))
473 .addOperand(getLdStRegOp(RtMI))
474 .addOperand(getLdStRegOp(Rt2MI))
475 .addOperand(BaseRegOp)
479 // FIXME: Do we need/want to copy the mem operands from the source
480 // instructions? Probably. What uses them after this?
482 DEBUG(dbgs() << "Creating pair load/store. Replacing instructions:\n ");
483 DEBUG(I->print(dbgs()));
484 DEBUG(dbgs() << " ");
485 DEBUG(Paired->print(dbgs()));
486 DEBUG(dbgs() << " with instruction:\n ");
489 // Generate the sign extension for the proper result of the ldp.
490 // I.e., with X1, that would be:
491 // %W1<def> = KILL %W1, %X1<imp-def>
492 // %X1<def> = SBFMXri %X1<kill>, 0, 31
493 MachineOperand &DstMO = MIB->getOperand(SExtIdx);
494 // Right now, DstMO has the extended register, since it comes from an
496 unsigned DstRegX = DstMO.getReg();
497 // Get the W variant of that register.
498 unsigned DstRegW = TRI->getSubReg(DstRegX, AArch64::sub_32);
499 // Update the result of LDP to use the W instead of the X variant.
500 DstMO.setReg(DstRegW);
501 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
502 DEBUG(dbgs() << "\n");
503 // Make the machine verifier happy by providing a definition for
505 // Insert this definition right after the generated LDP, i.e., before
507 MachineInstrBuilder MIBKill =
508 BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
509 TII->get(TargetOpcode::KILL), DstRegW)
511 .addReg(DstRegX, RegState::Define);
512 MIBKill->getOperand(2).setImplicit();
513 // Create the sign extension.
514 MachineInstrBuilder MIBSXTW =
515 BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
516 TII->get(AArch64::SBFMXri), DstRegX)
521 DEBUG(dbgs() << " Extend operand:\n ");
522 DEBUG(((MachineInstr *)MIBSXTW)->print(dbgs()));
523 DEBUG(dbgs() << "\n");
525 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
526 DEBUG(dbgs() << "\n");
529 // Erase the old instructions.
530 I->eraseFromParent();
531 Paired->eraseFromParent();
536 /// trackRegDefsUses - Remember what registers the specified instruction uses
538 static void trackRegDefsUses(const MachineInstr *MI, BitVector &ModifiedRegs,
540 const TargetRegisterInfo *TRI) {
541 for (const MachineOperand &MO : MI->operands()) {
543 ModifiedRegs.setBitsNotInMask(MO.getRegMask());
547 unsigned Reg = MO.getReg();
549 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
550 ModifiedRegs.set(*AI);
552 assert(MO.isUse() && "Reg operand not a def and not a use?!?");
553 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
559 static bool inBoundsForPair(bool IsUnscaled, int Offset, int OffsetStride) {
560 // Convert the byte-offset used by unscaled into an "element" offset used
561 // by the scaled pair load/store instructions.
563 Offset /= OffsetStride;
565 return Offset <= 63 && Offset >= -64;
568 // Do alignment, specialized to power of 2 and for signed ints,
569 // avoiding having to do a C-style cast from uint_64t to int when
570 // using RoundUpToAlignment from include/llvm/Support/MathExtras.h.
571 // FIXME: Move this function to include/MathExtras.h?
572 static int alignTo(int Num, int PowOf2) {
573 return (Num + PowOf2 - 1) & ~(PowOf2 - 1);
576 static bool mayAlias(MachineInstr *MIa, MachineInstr *MIb,
577 const AArch64InstrInfo *TII) {
578 // One of the instructions must modify memory.
579 if (!MIa->mayStore() && !MIb->mayStore())
582 // Both instructions must be memory operations.
583 if (!MIa->mayLoadOrStore() && !MIb->mayLoadOrStore())
586 return !TII->areMemAccessesTriviallyDisjoint(MIa, MIb);
589 static bool mayAlias(MachineInstr *MIa,
590 SmallVectorImpl<MachineInstr *> &MemInsns,
591 const AArch64InstrInfo *TII) {
592 for (auto &MIb : MemInsns)
593 if (mayAlias(MIa, MIb, TII))
599 /// findMatchingInsn - Scan the instructions looking for a load/store that can
600 /// be combined with the current instruction into a load/store pair.
601 MachineBasicBlock::iterator
602 AArch64LoadStoreOpt::findMatchingInsn(MachineBasicBlock::iterator I,
603 LdStPairFlags &Flags,
605 MachineBasicBlock::iterator E = I->getParent()->end();
606 MachineBasicBlock::iterator MBBI = I;
607 MachineInstr *FirstMI = I;
610 unsigned Opc = FirstMI->getOpcode();
611 bool MayLoad = FirstMI->mayLoad();
612 bool IsUnscaled = isUnscaledLdSt(FirstMI);
613 unsigned Reg = getLdStRegOp(FirstMI).getReg();
614 unsigned BaseReg = getLdStBaseOp(FirstMI).getReg();
615 int Offset = getLdStOffsetOp(FirstMI).getImm();
617 // Early exit if the first instruction modifies the base register.
618 // e.g., ldr x0, [x0]
619 if (FirstMI->modifiesRegister(BaseReg, TRI))
622 // Early exit if the offset if not possible to match. (6 bits of positive
623 // range, plus allow an extra one in case we find a later insn that matches
626 IsUnscaled && EnableAArch64UnscaledMemOp ? getMemScale(FirstMI) : 1;
627 if (!inBoundsForPair(IsUnscaled, Offset, OffsetStride))
630 // Track which registers have been modified and used between the first insn
631 // (inclusive) and the second insn.
632 BitVector ModifiedRegs, UsedRegs;
633 ModifiedRegs.resize(TRI->getNumRegs());
634 UsedRegs.resize(TRI->getNumRegs());
636 // Remember any instructions that read/write memory between FirstMI and MI.
637 SmallVector<MachineInstr *, 4> MemInsns;
639 for (unsigned Count = 0; MBBI != E && Count < Limit; ++MBBI) {
640 MachineInstr *MI = MBBI;
641 // Skip DBG_VALUE instructions. Otherwise debug info can affect the
642 // optimization by changing how far we scan.
643 if (MI->isDebugValue())
646 // Now that we know this is a real instruction, count it.
649 bool CanMergeOpc = Opc == MI->getOpcode();
650 Flags.setSExtIdx(-1);
652 bool IsValidLdStrOpc;
653 unsigned NonSExtOpc = getMatchingNonSExtOpcode(Opc, &IsValidLdStrOpc);
654 assert(IsValidLdStrOpc &&
655 "Given Opc should be a Load or Store with an immediate");
656 // Opc will be the first instruction in the pair.
657 Flags.setSExtIdx(NonSExtOpc == (unsigned)Opc ? 1 : 0);
658 CanMergeOpc = NonSExtOpc == getMatchingNonSExtOpcode(MI->getOpcode());
661 if (CanMergeOpc && getLdStOffsetOp(MI).isImm()) {
662 assert(MI->mayLoadOrStore() && "Expected memory operation.");
663 // If we've found another instruction with the same opcode, check to see
664 // if the base and offset are compatible with our starting instruction.
665 // These instructions all have scaled immediate operands, so we just
666 // check for +1/-1. Make sure to check the new instruction offset is
667 // actually an immediate and not a symbolic reference destined for
670 // Pairwise instructions have a 7-bit signed offset field. Single insns
671 // have a 12-bit unsigned offset field. To be a valid combine, the
672 // final offset must be in range.
673 unsigned MIBaseReg = getLdStBaseOp(MI).getReg();
674 int MIOffset = getLdStOffsetOp(MI).getImm();
675 if (BaseReg == MIBaseReg && ((Offset == MIOffset + OffsetStride) ||
676 (Offset + OffsetStride == MIOffset))) {
677 int MinOffset = Offset < MIOffset ? Offset : MIOffset;
678 // If this is a volatile load/store that otherwise matched, stop looking
679 // as something is going on that we don't have enough information to
680 // safely transform. Similarly, stop if we see a hint to avoid pairs.
681 if (MI->hasOrderedMemoryRef() || TII->isLdStPairSuppressed(MI))
683 // If the resultant immediate offset of merging these instructions
684 // is out of range for a pairwise instruction, bail and keep looking.
685 bool MIIsUnscaled = isUnscaledLdSt(MI);
686 if (!inBoundsForPair(MIIsUnscaled, MinOffset, OffsetStride)) {
687 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
688 MemInsns.push_back(MI);
691 // If the alignment requirements of the paired (scaled) instruction
692 // can't express the offset of the unscaled input, bail and keep
694 if (IsUnscaled && EnableAArch64UnscaledMemOp &&
695 (alignTo(MinOffset, OffsetStride) != MinOffset)) {
696 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
697 MemInsns.push_back(MI);
700 // If the destination register of the loads is the same register, bail
701 // and keep looking. A load-pair instruction with both destination
702 // registers the same is UNPREDICTABLE and will result in an exception.
703 if (MayLoad && Reg == getLdStRegOp(MI).getReg()) {
704 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
705 MemInsns.push_back(MI);
709 // If the Rt of the second instruction was not modified or used between
710 // the two instructions and none of the instructions between the second
711 // and first alias with the second, we can combine the second into the
713 if (!ModifiedRegs[getLdStRegOp(MI).getReg()] &&
714 !(MI->mayLoad() && UsedRegs[getLdStRegOp(MI).getReg()]) &&
715 !mayAlias(MI, MemInsns, TII)) {
716 Flags.setMergeForward(false);
720 // Likewise, if the Rt of the first instruction is not modified or used
721 // between the two instructions and none of the instructions between the
722 // first and the second alias with the first, we can combine the first
724 if (!ModifiedRegs[getLdStRegOp(FirstMI).getReg()] &&
725 !(MayLoad && UsedRegs[getLdStRegOp(FirstMI).getReg()]) &&
726 !mayAlias(FirstMI, MemInsns, TII)) {
727 Flags.setMergeForward(true);
730 // Unable to combine these instructions due to interference in between.
735 // If the instruction wasn't a matching load or store. Stop searching if we
736 // encounter a call instruction that might modify memory.
740 // Update modified / uses register lists.
741 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
743 // Otherwise, if the base register is modified, we have no match, so
745 if (ModifiedRegs[BaseReg])
748 // Update list of instructions that read/write memory.
749 if (MI->mayLoadOrStore())
750 MemInsns.push_back(MI);
755 MachineBasicBlock::iterator
756 AArch64LoadStoreOpt::mergeUpdateInsn(MachineBasicBlock::iterator I,
757 MachineBasicBlock::iterator Update,
759 assert((Update->getOpcode() == AArch64::ADDXri ||
760 Update->getOpcode() == AArch64::SUBXri) &&
761 "Unexpected base register update instruction to merge!");
762 MachineBasicBlock::iterator NextI = I;
763 // Return the instruction following the merged instruction, which is
764 // the instruction following our unmerged load. Unless that's the add/sub
765 // instruction we're merging, in which case it's the one after that.
766 if (++NextI == Update)
769 int Value = Update->getOperand(2).getImm();
770 assert(AArch64_AM::getShiftValue(Update->getOperand(3).getImm()) == 0 &&
771 "Can't merge 1 << 12 offset into pre-/post-indexed load / store");
772 if (Update->getOpcode() == AArch64::SUBXri)
775 unsigned NewOpc = IsPreIdx ? getPreIndexedOpcode(I->getOpcode())
776 : getPostIndexedOpcode(I->getOpcode());
777 MachineInstrBuilder MIB;
778 if (!isPairedLdSt(I)) {
779 // Non-paired instruction.
780 MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc))
781 .addOperand(getLdStRegOp(Update))
782 .addOperand(getLdStRegOp(I))
783 .addOperand(getLdStBaseOp(I))
786 // Paired instruction.
787 int Scale = getMemScale(I);
788 MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc))
789 .addOperand(getLdStRegOp(Update))
790 .addOperand(getLdStRegOp(I, 0))
791 .addOperand(getLdStRegOp(I, 1))
792 .addOperand(getLdStBaseOp(I))
793 .addImm(Value / Scale);
798 DEBUG(dbgs() << "Creating pre-indexed load/store.");
800 DEBUG(dbgs() << "Creating post-indexed load/store.");
801 DEBUG(dbgs() << " Replacing instructions:\n ");
802 DEBUG(I->print(dbgs()));
803 DEBUG(dbgs() << " ");
804 DEBUG(Update->print(dbgs()));
805 DEBUG(dbgs() << " with instruction:\n ");
806 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
807 DEBUG(dbgs() << "\n");
809 // Erase the old instructions for the block.
810 I->eraseFromParent();
811 Update->eraseFromParent();
816 bool AArch64LoadStoreOpt::isMatchingUpdateInsn(MachineInstr *MemMI,
818 unsigned BaseReg, int Offset) {
819 switch (MI->getOpcode()) {
822 case AArch64::SUBXri:
823 // Negate the offset for a SUB instruction.
826 case AArch64::ADDXri:
827 // Make sure it's a vanilla immediate operand, not a relocation or
828 // anything else we can't handle.
829 if (!MI->getOperand(2).isImm())
831 // Watch out for 1 << 12 shifted value.
832 if (AArch64_AM::getShiftValue(MI->getOperand(3).getImm()))
835 // The update instruction source and destination register must be the
836 // same as the load/store base register.
837 if (MI->getOperand(0).getReg() != BaseReg ||
838 MI->getOperand(1).getReg() != BaseReg)
841 bool IsPairedInsn = isPairedLdSt(MemMI);
842 int UpdateOffset = MI->getOperand(2).getImm();
843 // For non-paired load/store instructions, the immediate must fit in a
844 // signed 9-bit integer.
845 if (!IsPairedInsn && (UpdateOffset > 255 || UpdateOffset < -256))
848 // For paired load/store instructions, the immediate must be a multiple of
849 // the scaling factor. The scaled offset must also fit into a signed 7-bit
852 int Scale = getMemScale(MemMI);
853 if (UpdateOffset % Scale != 0)
856 int ScaledOffset = UpdateOffset / Scale;
857 if (ScaledOffset > 64 || ScaledOffset < -64)
861 // If we have a non-zero Offset, we check that it matches the amount
862 // we're adding to the register.
863 if (!Offset || Offset == MI->getOperand(2).getImm())
870 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnForward(
871 MachineBasicBlock::iterator I, unsigned Limit, int Value) {
872 MachineBasicBlock::iterator E = I->getParent()->end();
873 MachineInstr *MemMI = I;
874 MachineBasicBlock::iterator MBBI = I;
876 unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
877 int Offset = getLdStOffsetOp(MemMI).getImm() * getMemScale(MemMI);
879 // If the base register overlaps a destination register, we can't
881 bool IsPairedInsn = isPairedLdSt(MemMI);
882 for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) {
883 unsigned DestReg = getLdStRegOp(MemMI, i).getReg();
884 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg))
888 // Scan forward looking for post-index opportunities.
889 // Updating instructions can't be formed if the memory insn already
890 // has an offset other than the value we're looking for.
894 // Track which registers have been modified and used between the first insn
895 // (inclusive) and the second insn.
896 BitVector ModifiedRegs, UsedRegs;
897 ModifiedRegs.resize(TRI->getNumRegs());
898 UsedRegs.resize(TRI->getNumRegs());
900 for (unsigned Count = 0; MBBI != E; ++MBBI) {
901 MachineInstr *MI = MBBI;
902 // Skip DBG_VALUE instructions. Otherwise debug info can affect the
903 // optimization by changing how far we scan.
904 if (MI->isDebugValue())
907 // Now that we know this is a real instruction, count it.
910 // If we found a match, return it.
911 if (isMatchingUpdateInsn(I, MI, BaseReg, Value))
914 // Update the status of what the instruction clobbered and used.
915 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
917 // Otherwise, if the base register is used or modified, we have no match, so
919 if (ModifiedRegs[BaseReg] || UsedRegs[BaseReg])
925 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnBackward(
926 MachineBasicBlock::iterator I, unsigned Limit) {
927 MachineBasicBlock::iterator B = I->getParent()->begin();
928 MachineBasicBlock::iterator E = I->getParent()->end();
929 MachineInstr *MemMI = I;
930 MachineBasicBlock::iterator MBBI = I;
932 unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
933 int Offset = getLdStOffsetOp(MemMI).getImm();
934 unsigned MemSize = getMemScale(MemMI);
936 // If the load/store is the first instruction in the block, there's obviously
937 // not any matching update. Ditto if the memory offset isn't zero.
938 if (MBBI == B || Offset != 0)
940 // If the base register overlaps a destination register, we can't
942 bool IsPairedInsn = isPairedLdSt(MemMI);
943 for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) {
944 unsigned DestReg = getLdStRegOp(MemMI, i).getReg();
945 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg))
949 // Track which registers have been modified and used between the first insn
950 // (inclusive) and the second insn.
951 BitVector ModifiedRegs, UsedRegs;
952 ModifiedRegs.resize(TRI->getNumRegs());
953 UsedRegs.resize(TRI->getNumRegs());
955 for (unsigned Count = 0; MBBI != B; --MBBI) {
956 MachineInstr *MI = MBBI;
957 // Skip DBG_VALUE instructions. Otherwise debug info can affect the
958 // optimization by changing how far we scan.
959 if (MI->isDebugValue())
962 // Now that we know this is a real instruction, count it.
965 // If we found a match, return it.
966 if (isMatchingUpdateInsn(I, MI, BaseReg, MemSize))
969 // Update the status of what the instruction clobbered and used.
970 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
972 // Otherwise, if the base register is used or modified, we have no match, so
974 if (ModifiedRegs[BaseReg] || UsedRegs[BaseReg])
980 bool AArch64LoadStoreOpt::optimizeBlock(MachineBasicBlock &MBB) {
981 bool Modified = false;
982 // Two tranformations to do here:
983 // 1) Find loads and stores that can be merged into a single load or store
990 // 2) Find base register updates that can be merged into the load or store
991 // as a base-reg writeback.
998 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1000 MachineInstr *MI = MBBI;
1001 switch (MI->getOpcode()) {
1003 // Just move on to the next instruction.
1006 // Scaled instructions.
1007 case AArch64::STRSui:
1008 case AArch64::STRDui:
1009 case AArch64::STRQui:
1010 case AArch64::STRXui:
1011 case AArch64::STRWui:
1012 case AArch64::LDRSui:
1013 case AArch64::LDRDui:
1014 case AArch64::LDRQui:
1015 case AArch64::LDRXui:
1016 case AArch64::LDRWui:
1017 case AArch64::LDRSWui:
1018 // Unscaled instructions.
1019 case AArch64::STURSi:
1020 case AArch64::STURDi:
1021 case AArch64::STURQi:
1022 case AArch64::STURWi:
1023 case AArch64::STURXi:
1024 case AArch64::LDURSi:
1025 case AArch64::LDURDi:
1026 case AArch64::LDURQi:
1027 case AArch64::LDURWi:
1028 case AArch64::LDURXi:
1029 case AArch64::LDURSWi: {
1030 // If this is a volatile load/store, don't mess with it.
1031 if (MI->hasOrderedMemoryRef()) {
1035 // Make sure this is a reg+imm (as opposed to an address reloc).
1036 if (!getLdStOffsetOp(MI).isImm()) {
1040 // Check if this load/store has a hint to avoid pair formation.
1041 // MachineMemOperands hints are set by the AArch64StorePairSuppress pass.
1042 if (TII->isLdStPairSuppressed(MI)) {
1046 // Look ahead up to ScanLimit instructions for a pairable instruction.
1047 LdStPairFlags Flags;
1048 MachineBasicBlock::iterator Paired =
1049 findMatchingInsn(MBBI, Flags, ScanLimit);
1052 if (isUnscaledLdSt(MI))
1053 ++NumUnscaledPairCreated;
1055 // Merge the loads into a pair. Keeping the iterator straight is a
1056 // pain, so we let the merge routine tell us what the next instruction
1057 // is after it's done mucking about.
1058 MBBI = mergePairedInsns(MBBI, Paired, Flags);
1065 // FIXME: Do the other instructions.
1069 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1071 MachineInstr *MI = MBBI;
1072 // Do update merging. It's simpler to keep this separate from the above
1073 // switch, though not strictly necessary.
1074 unsigned Opc = MI->getOpcode();
1077 // Just move on to the next instruction.
1080 // Scaled instructions.
1081 case AArch64::STRSui:
1082 case AArch64::STRDui:
1083 case AArch64::STRQui:
1084 case AArch64::STRXui:
1085 case AArch64::STRWui:
1086 case AArch64::LDRSui:
1087 case AArch64::LDRDui:
1088 case AArch64::LDRQui:
1089 case AArch64::LDRXui:
1090 case AArch64::LDRWui:
1091 // Unscaled instructions.
1092 case AArch64::STURSi:
1093 case AArch64::STURDi:
1094 case AArch64::STURQi:
1095 case AArch64::STURWi:
1096 case AArch64::STURXi:
1097 case AArch64::LDURSi:
1098 case AArch64::LDURDi:
1099 case AArch64::LDURQi:
1100 case AArch64::LDURWi:
1101 case AArch64::LDURXi:
1102 // Paired instructions.
1103 case AArch64::LDPSi:
1104 case AArch64::LDPDi:
1105 case AArch64::LDPQi:
1106 case AArch64::LDPWi:
1107 case AArch64::LDPXi:
1108 case AArch64::STPSi:
1109 case AArch64::STPDi:
1110 case AArch64::STPQi:
1111 case AArch64::STPWi:
1112 case AArch64::STPXi: {
1113 // Make sure this is a reg+imm (as opposed to an address reloc).
1114 if (!getLdStOffsetOp(MI).isImm()) {
1118 // Look forward to try to form a post-index instruction. For example,
1120 // add x20, x20, #32
1122 // ldr x0, [x20], #32
1123 MachineBasicBlock::iterator Update =
1124 findMatchingUpdateInsnForward(MBBI, ScanLimit, 0);
1126 // Merge the update into the ld/st.
1127 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/false);
1132 // Don't know how to handle pre/post-index versions, so move to the next
1134 if (isUnscaledLdSt(Opc)) {
1139 // Look back to try to find a pre-index instruction. For example,
1143 // ldr x1, [x0, #8]!
1144 Update = findMatchingUpdateInsnBackward(MBBI, ScanLimit);
1146 // Merge the update into the ld/st.
1147 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true);
1152 // The immediate in the load/store is scaled by the size of the register
1153 // being loaded. The immediate in the add we're looking for,
1154 // however, is not, so adjust here.
1156 MI->getOperand(isPairedLdSt(MI) ? 3 : 2).getImm() * getMemScale(MI);
1158 // FIXME: The immediate in the load/store should be scaled by the size of
1159 // the memory operation, not the size of the register being loaded/stored.
1160 // This works in general, but does not work for the LDPSW instruction,
1161 // which defines two 64-bit registers, but loads 32-bit values.
1163 // Look forward to try to find a post-index instruction. For example,
1164 // ldr x1, [x0, #64]
1167 // ldr x1, [x0, #64]!
1168 Update = findMatchingUpdateInsnForward(MBBI, ScanLimit, Value);
1170 // Merge the update into the ld/st.
1171 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true);
1177 // Nothing found. Just move to the next instruction.
1181 // FIXME: Do the other instructions.
1188 bool AArch64LoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
1189 TII = static_cast<const AArch64InstrInfo *>(Fn.getSubtarget().getInstrInfo());
1190 TRI = Fn.getSubtarget().getRegisterInfo();
1192 bool Modified = false;
1193 for (auto &MBB : Fn)
1194 Modified |= optimizeBlock(MBB);
1199 // FIXME: Do we need/want a pre-alloc pass like ARM has to try to keep
1200 // loads and stores near one another?
1202 /// createAArch64LoadStoreOptimizationPass - returns an instance of the
1203 /// load / store optimization pass.
1204 FunctionPass *llvm::createAArch64LoadStoreOptimizationPass() {
1205 return new AArch64LoadStoreOpt();