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 // Merge a pre-index base register update into a ld/st instruction.
121 MachineBasicBlock::iterator
122 mergePreIdxUpdateInsn(MachineBasicBlock::iterator I,
123 MachineBasicBlock::iterator Update);
125 // Merge a post-index base register update into a ld/st instruction.
126 MachineBasicBlock::iterator
127 mergePostIdxUpdateInsn(MachineBasicBlock::iterator I,
128 MachineBasicBlock::iterator Update);
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 // Size in bytes of the data moved by an unscaled load or store
168 static int getMemSize(MachineInstr *MI) {
169 switch (MI->getOpcode()) {
171 llvm_unreachable("Opcode has unknown size!");
172 case AArch64::STRSui:
173 case AArch64::STURSi:
175 case AArch64::STRDui:
176 case AArch64::STURDi:
178 case AArch64::STRQui:
179 case AArch64::STURQi:
181 case AArch64::STRWui:
182 case AArch64::STURWi:
184 case AArch64::STRXui:
185 case AArch64::STURXi:
187 case AArch64::LDRSui:
188 case AArch64::LDURSi:
190 case AArch64::LDRDui:
191 case AArch64::LDURDi:
193 case AArch64::LDRQui:
194 case AArch64::LDURQi:
196 case AArch64::LDRWui:
197 case AArch64::LDURWi:
199 case AArch64::LDRXui:
200 case AArch64::LDURXi:
202 case AArch64::LDRSWui:
203 case AArch64::LDURSWi:
208 static unsigned getMatchingNonSExtOpcode(unsigned Opc,
209 bool *IsValidLdStrOpc = nullptr) {
211 *IsValidLdStrOpc = true;
215 *IsValidLdStrOpc = false;
217 case AArch64::STRDui:
218 case AArch64::STURDi:
219 case AArch64::STRQui:
220 case AArch64::STURQi:
221 case AArch64::STRWui:
222 case AArch64::STURWi:
223 case AArch64::STRXui:
224 case AArch64::STURXi:
225 case AArch64::LDRDui:
226 case AArch64::LDURDi:
227 case AArch64::LDRQui:
228 case AArch64::LDURQi:
229 case AArch64::LDRWui:
230 case AArch64::LDURWi:
231 case AArch64::LDRXui:
232 case AArch64::LDURXi:
233 case AArch64::STRSui:
234 case AArch64::STURSi:
235 case AArch64::LDRSui:
236 case AArch64::LDURSi:
238 case AArch64::LDRSWui:
239 return AArch64::LDRWui;
240 case AArch64::LDURSWi:
241 return AArch64::LDURWi;
245 static unsigned getMatchingPairOpcode(unsigned Opc) {
248 llvm_unreachable("Opcode has no pairwise equivalent!");
249 case AArch64::STRSui:
250 case AArch64::STURSi:
251 return AArch64::STPSi;
252 case AArch64::STRDui:
253 case AArch64::STURDi:
254 return AArch64::STPDi;
255 case AArch64::STRQui:
256 case AArch64::STURQi:
257 return AArch64::STPQi;
258 case AArch64::STRWui:
259 case AArch64::STURWi:
260 return AArch64::STPWi;
261 case AArch64::STRXui:
262 case AArch64::STURXi:
263 return AArch64::STPXi;
264 case AArch64::LDRSui:
265 case AArch64::LDURSi:
266 return AArch64::LDPSi;
267 case AArch64::LDRDui:
268 case AArch64::LDURDi:
269 return AArch64::LDPDi;
270 case AArch64::LDRQui:
271 case AArch64::LDURQi:
272 return AArch64::LDPQi;
273 case AArch64::LDRWui:
274 case AArch64::LDURWi:
275 return AArch64::LDPWi;
276 case AArch64::LDRXui:
277 case AArch64::LDURXi:
278 return AArch64::LDPXi;
279 case AArch64::LDRSWui:
280 case AArch64::LDURSWi:
281 return AArch64::LDPSWi;
285 static unsigned getPreIndexedOpcode(unsigned Opc) {
288 llvm_unreachable("Opcode has no pre-indexed equivalent!");
289 case AArch64::STRSui:
290 return AArch64::STRSpre;
291 case AArch64::STRDui:
292 return AArch64::STRDpre;
293 case AArch64::STRQui:
294 return AArch64::STRQpre;
295 case AArch64::STRWui:
296 return AArch64::STRWpre;
297 case AArch64::STRXui:
298 return AArch64::STRXpre;
299 case AArch64::LDRSui:
300 return AArch64::LDRSpre;
301 case AArch64::LDRDui:
302 return AArch64::LDRDpre;
303 case AArch64::LDRQui:
304 return AArch64::LDRQpre;
305 case AArch64::LDRWui:
306 return AArch64::LDRWpre;
307 case AArch64::LDRXui:
308 return AArch64::LDRXpre;
309 case AArch64::LDRSWui:
310 return AArch64::LDRSWpre;
314 static unsigned getPostIndexedOpcode(unsigned Opc) {
317 llvm_unreachable("Opcode has no post-indexed wise equivalent!");
318 case AArch64::STRSui:
319 return AArch64::STRSpost;
320 case AArch64::STRDui:
321 return AArch64::STRDpost;
322 case AArch64::STRQui:
323 return AArch64::STRQpost;
324 case AArch64::STRWui:
325 return AArch64::STRWpost;
326 case AArch64::STRXui:
327 return AArch64::STRXpost;
328 case AArch64::LDRSui:
329 return AArch64::LDRSpost;
330 case AArch64::LDRDui:
331 return AArch64::LDRDpost;
332 case AArch64::LDRQui:
333 return AArch64::LDRQpost;
334 case AArch64::LDRWui:
335 return AArch64::LDRWpost;
336 case AArch64::LDRXui:
337 return AArch64::LDRXpost;
338 case AArch64::LDRSWui:
339 return AArch64::LDRSWpost;
343 static const MachineOperand &getLdStRegOp(const MachineInstr *MI) {
344 return MI->getOperand(0);
347 static const MachineOperand &getLdStBaseOp(const MachineInstr *MI) {
348 return MI->getOperand(1);
351 static const MachineOperand &getLdStOffsetOp(const MachineInstr *MI) {
352 return MI->getOperand(2);
355 MachineBasicBlock::iterator
356 AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I,
357 MachineBasicBlock::iterator Paired,
358 const LdStPairFlags &Flags) {
359 MachineBasicBlock::iterator NextI = I;
361 // If NextI is the second of the two instructions to be merged, we need
362 // to skip one further. Either way we merge will invalidate the iterator,
363 // and we don't need to scan the new instruction, as it's a pairwise
364 // instruction, which we're not considering for further action anyway.
368 int SExtIdx = Flags.getSExtIdx();
370 SExtIdx == -1 ? I->getOpcode() : getMatchingNonSExtOpcode(I->getOpcode());
371 bool IsUnscaled = isUnscaledLdSt(Opc);
373 IsUnscaled && EnableAArch64UnscaledMemOp ? getMemSize(I) : 1;
375 bool MergeForward = Flags.getMergeForward();
376 unsigned NewOpc = getMatchingPairOpcode(Opc);
377 // Insert our new paired instruction after whichever of the paired
378 // instructions MergeForward indicates.
379 MachineBasicBlock::iterator InsertionPoint = MergeForward ? Paired : I;
380 // Also based on MergeForward is from where we copy the base register operand
381 // so we get the flags compatible with the input code.
382 const MachineOperand &BaseRegOp =
383 MergeForward ? getLdStBaseOp(Paired) : getLdStBaseOp(I);
385 // Which register is Rt and which is Rt2 depends on the offset order.
386 MachineInstr *RtMI, *Rt2MI;
387 if (getLdStOffsetOp(I).getImm() ==
388 getLdStOffsetOp(Paired).getImm() + OffsetStride) {
391 // Here we swapped the assumption made for SExtIdx.
392 // I.e., we turn ldp I, Paired into ldp Paired, I.
393 // Update the index accordingly.
395 SExtIdx = (SExtIdx + 1) % 2;
401 int OffsetImm = getLdStOffsetOp(RtMI).getImm();
402 if (IsUnscaled && EnableAArch64UnscaledMemOp)
403 OffsetImm /= OffsetStride;
405 // Construct the new instruction.
406 MachineInstrBuilder MIB = BuildMI(*I->getParent(), InsertionPoint,
407 I->getDebugLoc(), TII->get(NewOpc))
408 .addOperand(getLdStRegOp(RtMI))
409 .addOperand(getLdStRegOp(Rt2MI))
410 .addOperand(BaseRegOp)
414 // FIXME: Do we need/want to copy the mem operands from the source
415 // instructions? Probably. What uses them after this?
417 DEBUG(dbgs() << "Creating pair load/store. Replacing instructions:\n ");
418 DEBUG(I->print(dbgs()));
419 DEBUG(dbgs() << " ");
420 DEBUG(Paired->print(dbgs()));
421 DEBUG(dbgs() << " with instruction:\n ");
424 // Generate the sign extension for the proper result of the ldp.
425 // I.e., with X1, that would be:
426 // %W1<def> = KILL %W1, %X1<imp-def>
427 // %X1<def> = SBFMXri %X1<kill>, 0, 31
428 MachineOperand &DstMO = MIB->getOperand(SExtIdx);
429 // Right now, DstMO has the extended register, since it comes from an
431 unsigned DstRegX = DstMO.getReg();
432 // Get the W variant of that register.
433 unsigned DstRegW = TRI->getSubReg(DstRegX, AArch64::sub_32);
434 // Update the result of LDP to use the W instead of the X variant.
435 DstMO.setReg(DstRegW);
436 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
437 DEBUG(dbgs() << "\n");
438 // Make the machine verifier happy by providing a definition for
440 // Insert this definition right after the generated LDP, i.e., before
442 MachineInstrBuilder MIBKill =
443 BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
444 TII->get(TargetOpcode::KILL), DstRegW)
446 .addReg(DstRegX, RegState::Define);
447 MIBKill->getOperand(2).setImplicit();
448 // Create the sign extension.
449 MachineInstrBuilder MIBSXTW =
450 BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
451 TII->get(AArch64::SBFMXri), DstRegX)
456 DEBUG(dbgs() << " Extend operand:\n ");
457 DEBUG(((MachineInstr *)MIBSXTW)->print(dbgs()));
458 DEBUG(dbgs() << "\n");
460 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
461 DEBUG(dbgs() << "\n");
464 // Erase the old instructions.
465 I->eraseFromParent();
466 Paired->eraseFromParent();
471 /// trackRegDefsUses - Remember what registers the specified instruction uses
473 static void trackRegDefsUses(const MachineInstr *MI, BitVector &ModifiedRegs,
475 const TargetRegisterInfo *TRI) {
476 for (const MachineOperand &MO : MI->operands()) {
478 ModifiedRegs.setBitsNotInMask(MO.getRegMask());
482 unsigned Reg = MO.getReg();
484 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
485 ModifiedRegs.set(*AI);
487 assert(MO.isUse() && "Reg operand not a def and not a use?!?");
488 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
494 static bool inBoundsForPair(bool IsUnscaled, int Offset, int OffsetStride) {
495 if (!IsUnscaled && (Offset > 63 || Offset < -64))
498 // Convert the byte-offset used by unscaled into an "element" offset used
499 // by the scaled pair load/store instructions.
500 int ElemOffset = Offset / OffsetStride;
501 if (ElemOffset > 63 || ElemOffset < -64)
507 // Do alignment, specialized to power of 2 and for signed ints,
508 // avoiding having to do a C-style cast from uint_64t to int when
509 // using RoundUpToAlignment from include/llvm/Support/MathExtras.h.
510 // FIXME: Move this function to include/MathExtras.h?
511 static int alignTo(int Num, int PowOf2) {
512 return (Num + PowOf2 - 1) & ~(PowOf2 - 1);
515 static bool mayAlias(MachineInstr *MIa, MachineInstr *MIb,
516 const AArch64InstrInfo *TII) {
517 // One of the instructions must modify memory.
518 if (!MIa->mayStore() && !MIb->mayStore())
521 // Both instructions must be memory operations.
522 if (!MIa->mayLoadOrStore() && !MIb->mayLoadOrStore())
525 return !TII->areMemAccessesTriviallyDisjoint(MIa, MIb);
528 static bool mayAlias(MachineInstr *MIa,
529 SmallVectorImpl<MachineInstr *> &MemInsns,
530 const AArch64InstrInfo *TII) {
531 for (auto &MIb : MemInsns)
532 if (mayAlias(MIa, MIb, TII))
538 /// findMatchingInsn - Scan the instructions looking for a load/store that can
539 /// be combined with the current instruction into a load/store pair.
540 MachineBasicBlock::iterator
541 AArch64LoadStoreOpt::findMatchingInsn(MachineBasicBlock::iterator I,
542 LdStPairFlags &Flags,
544 MachineBasicBlock::iterator E = I->getParent()->end();
545 MachineBasicBlock::iterator MBBI = I;
546 MachineInstr *FirstMI = I;
549 unsigned Opc = FirstMI->getOpcode();
550 bool MayLoad = FirstMI->mayLoad();
551 bool IsUnscaled = isUnscaledLdSt(FirstMI);
552 unsigned Reg = getLdStRegOp(FirstMI).getReg();
553 unsigned BaseReg = getLdStBaseOp(FirstMI).getReg();
554 int Offset = getLdStOffsetOp(FirstMI).getImm();
556 // Early exit if the first instruction modifies the base register.
557 // e.g., ldr x0, [x0]
558 if (FirstMI->modifiesRegister(BaseReg, TRI))
561 // Early exit if the offset if not possible to match. (6 bits of positive
562 // range, plus allow an extra one in case we find a later insn that matches
565 IsUnscaled && EnableAArch64UnscaledMemOp ? getMemSize(FirstMI) : 1;
566 if (!inBoundsForPair(IsUnscaled, Offset, OffsetStride))
569 // Track which registers have been modified and used between the first insn
570 // (inclusive) and the second insn.
571 BitVector ModifiedRegs, UsedRegs;
572 ModifiedRegs.resize(TRI->getNumRegs());
573 UsedRegs.resize(TRI->getNumRegs());
575 // Remember any instructions that read/write memory between FirstMI and MI.
576 SmallVector<MachineInstr *, 4> MemInsns;
578 for (unsigned Count = 0; MBBI != E && Count < Limit; ++MBBI) {
579 MachineInstr *MI = MBBI;
580 // Skip DBG_VALUE instructions. Otherwise debug info can affect the
581 // optimization by changing how far we scan.
582 if (MI->isDebugValue())
585 // Now that we know this is a real instruction, count it.
588 bool CanMergeOpc = Opc == MI->getOpcode();
589 Flags.setSExtIdx(-1);
591 bool IsValidLdStrOpc;
592 unsigned NonSExtOpc = getMatchingNonSExtOpcode(Opc, &IsValidLdStrOpc);
593 assert(IsValidLdStrOpc &&
594 "Given Opc should be a Load or Store with an immediate");
595 // Opc will be the first instruction in the pair.
596 Flags.setSExtIdx(NonSExtOpc == (unsigned)Opc ? 1 : 0);
597 CanMergeOpc = NonSExtOpc == getMatchingNonSExtOpcode(MI->getOpcode());
600 if (CanMergeOpc && getLdStOffsetOp(MI).isImm()) {
601 // If we've found another instruction with the same opcode, check to see
602 // if the base and offset are compatible with our starting instruction.
603 // These instructions all have scaled immediate operands, so we just
604 // check for +1/-1. Make sure to check the new instruction offset is
605 // actually an immediate and not a symbolic reference destined for
608 // Pairwise instructions have a 7-bit signed offset field. Single insns
609 // have a 12-bit unsigned offset field. To be a valid combine, the
610 // final offset must be in range.
611 unsigned MIBaseReg = getLdStBaseOp(MI).getReg();
612 int MIOffset = getLdStOffsetOp(MI).getImm();
613 if (BaseReg == MIBaseReg && ((Offset == MIOffset + OffsetStride) ||
614 (Offset + OffsetStride == MIOffset))) {
615 int MinOffset = Offset < MIOffset ? Offset : MIOffset;
616 // If this is a volatile load/store that otherwise matched, stop looking
617 // as something is going on that we don't have enough information to
618 // safely transform. Similarly, stop if we see a hint to avoid pairs.
619 if (MI->hasOrderedMemoryRef() || TII->isLdStPairSuppressed(MI))
621 // If the resultant immediate offset of merging these instructions
622 // is out of range for a pairwise instruction, bail and keep looking.
623 bool MIIsUnscaled = isUnscaledLdSt(MI);
624 if (!inBoundsForPair(MIIsUnscaled, MinOffset, OffsetStride)) {
625 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
626 if (MI->mayLoadOrStore())
627 MemInsns.push_back(MI);
630 // If the alignment requirements of the paired (scaled) instruction
631 // can't express the offset of the unscaled input, bail and keep
633 if (IsUnscaled && EnableAArch64UnscaledMemOp &&
634 (alignTo(MinOffset, OffsetStride) != MinOffset)) {
635 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
636 if (MI->mayLoadOrStore())
637 MemInsns.push_back(MI);
640 // If the destination register of the loads is the same register, bail
641 // and keep looking. A load-pair instruction with both destination
642 // registers the same is UNPREDICTABLE and will result in an exception.
643 if (MayLoad && Reg == getLdStRegOp(MI).getReg()) {
644 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
645 if (MI->mayLoadOrStore())
646 MemInsns.push_back(MI);
650 // If the Rt of the second instruction was not modified or used between
651 // the two instructions and none of the instructions between the second
652 // and first alias with the second, we can combine the second into the
654 if (!ModifiedRegs[getLdStRegOp(MI).getReg()] &&
655 !(MI->mayLoad() && UsedRegs[getLdStRegOp(MI).getReg()]) &&
656 !mayAlias(MI, MemInsns, TII)) {
657 Flags.setMergeForward(false);
661 // Likewise, if the Rt of the first instruction is not modified or used
662 // between the two instructions and none of the instructions between the
663 // first and the second alias with the first, we can combine the first
665 if (!ModifiedRegs[getLdStRegOp(FirstMI).getReg()] &&
666 !(FirstMI->mayLoad() && UsedRegs[getLdStRegOp(FirstMI).getReg()]) &&
667 !mayAlias(FirstMI, MemInsns, TII)) {
668 Flags.setMergeForward(true);
671 // Unable to combine these instructions due to interference in between.
676 // If the instruction wasn't a matching load or store. Stop searching if we
677 // encounter a call instruction that might modify memory.
681 // Update modified / uses register lists.
682 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
684 // Otherwise, if the base register is modified, we have no match, so
686 if (ModifiedRegs[BaseReg])
689 // Update list of instructions that read/write memory.
690 if (MI->mayLoadOrStore())
691 MemInsns.push_back(MI);
696 MachineBasicBlock::iterator
697 AArch64LoadStoreOpt::mergePreIdxUpdateInsn(MachineBasicBlock::iterator I,
698 MachineBasicBlock::iterator Update) {
699 assert((Update->getOpcode() == AArch64::ADDXri ||
700 Update->getOpcode() == AArch64::SUBXri) &&
701 "Unexpected base register update instruction to merge!");
702 MachineBasicBlock::iterator NextI = I;
703 // Return the instruction following the merged instruction, which is
704 // the instruction following our unmerged load. Unless that's the add/sub
705 // instruction we're merging, in which case it's the one after that.
706 if (++NextI == Update)
709 int Value = Update->getOperand(2).getImm();
710 assert(AArch64_AM::getShiftValue(Update->getOperand(3).getImm()) == 0 &&
711 "Can't merge 1 << 12 offset into pre-indexed load / store");
712 if (Update->getOpcode() == AArch64::SUBXri)
715 unsigned NewOpc = getPreIndexedOpcode(I->getOpcode());
716 MachineInstrBuilder MIB =
717 BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc))
718 .addOperand(getLdStRegOp(Update))
719 .addOperand(getLdStRegOp(I))
720 .addOperand(getLdStBaseOp(I))
724 DEBUG(dbgs() << "Creating pre-indexed load/store.");
725 DEBUG(dbgs() << " Replacing instructions:\n ");
726 DEBUG(I->print(dbgs()));
727 DEBUG(dbgs() << " ");
728 DEBUG(Update->print(dbgs()));
729 DEBUG(dbgs() << " with instruction:\n ");
730 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
731 DEBUG(dbgs() << "\n");
733 // Erase the old instructions for the block.
734 I->eraseFromParent();
735 Update->eraseFromParent();
740 MachineBasicBlock::iterator AArch64LoadStoreOpt::mergePostIdxUpdateInsn(
741 MachineBasicBlock::iterator I, MachineBasicBlock::iterator Update) {
742 assert((Update->getOpcode() == AArch64::ADDXri ||
743 Update->getOpcode() == AArch64::SUBXri) &&
744 "Unexpected base register update instruction to merge!");
745 MachineBasicBlock::iterator NextI = I;
746 // Return the instruction following the merged instruction, which is
747 // the instruction following our unmerged load. Unless that's the add/sub
748 // instruction we're merging, in which case it's the one after that.
749 if (++NextI == Update)
752 int Value = Update->getOperand(2).getImm();
753 assert(AArch64_AM::getShiftValue(Update->getOperand(3).getImm()) == 0 &&
754 "Can't merge 1 << 12 offset into post-indexed load / store");
755 if (Update->getOpcode() == AArch64::SUBXri)
758 unsigned NewOpc = getPostIndexedOpcode(I->getOpcode());
759 MachineInstrBuilder MIB =
760 BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc))
761 .addOperand(getLdStRegOp(Update))
762 .addOperand(getLdStRegOp(I))
763 .addOperand(getLdStBaseOp(I))
767 DEBUG(dbgs() << "Creating post-indexed load/store.");
768 DEBUG(dbgs() << " Replacing instructions:\n ");
769 DEBUG(I->print(dbgs()));
770 DEBUG(dbgs() << " ");
771 DEBUG(Update->print(dbgs()));
772 DEBUG(dbgs() << " with instruction:\n ");
773 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
774 DEBUG(dbgs() << "\n");
776 // Erase the old instructions for the block.
777 I->eraseFromParent();
778 Update->eraseFromParent();
783 static bool isMatchingUpdateInsn(MachineInstr *MI, unsigned BaseReg,
785 switch (MI->getOpcode()) {
788 case AArch64::SUBXri:
789 // Negate the offset for a SUB instruction.
792 case AArch64::ADDXri:
793 // Make sure it's a vanilla immediate operand, not a relocation or
794 // anything else we can't handle.
795 if (!MI->getOperand(2).isImm())
797 // Watch out for 1 << 12 shifted value.
798 if (AArch64_AM::getShiftValue(MI->getOperand(3).getImm()))
800 // If the instruction has the base register as source and dest and the
801 // immediate will fit in a signed 9-bit integer, then we have a match.
802 if (getLdStRegOp(MI).getReg() == BaseReg &&
803 getLdStBaseOp(MI).getReg() == BaseReg &&
804 getLdStOffsetOp(MI).getImm() <= 255 &&
805 getLdStOffsetOp(MI).getImm() >= -256) {
806 // If we have a non-zero Offset, we check that it matches the amount
807 // we're adding to the register.
808 if (!Offset || Offset == MI->getOperand(2).getImm())
816 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnForward(
817 MachineBasicBlock::iterator I, unsigned Limit, int Value) {
818 MachineBasicBlock::iterator E = I->getParent()->end();
819 MachineInstr *MemMI = I;
820 MachineBasicBlock::iterator MBBI = I;
821 const MachineFunction &MF = *MemMI->getParent()->getParent();
823 unsigned DestReg = getLdStRegOp(MemMI).getReg();
824 unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
825 int Offset = getLdStOffsetOp(MemMI).getImm() *
826 TII->getRegClass(MemMI->getDesc(), 0, TRI, MF)->getSize();
828 // If the base register overlaps the destination register, we can't
830 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg))
833 // Scan forward looking for post-index opportunities.
834 // Updating instructions can't be formed if the memory insn already
835 // has an offset other than the value we're looking for.
839 // Track which registers have been modified and used between the first insn
840 // (inclusive) and the second insn.
841 BitVector ModifiedRegs, UsedRegs;
842 ModifiedRegs.resize(TRI->getNumRegs());
843 UsedRegs.resize(TRI->getNumRegs());
845 for (unsigned Count = 0; MBBI != E; ++MBBI) {
846 MachineInstr *MI = MBBI;
847 // Skip DBG_VALUE instructions. Otherwise debug info can affect the
848 // optimization by changing how far we scan.
849 if (MI->isDebugValue())
852 // Now that we know this is a real instruction, count it.
855 // If we found a match, return it.
856 if (isMatchingUpdateInsn(MI, BaseReg, Value))
859 // Update the status of what the instruction clobbered and used.
860 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
862 // Otherwise, if the base register is used or modified, we have no match, so
864 if (ModifiedRegs[BaseReg] || UsedRegs[BaseReg])
870 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnBackward(
871 MachineBasicBlock::iterator I, unsigned Limit) {
872 MachineBasicBlock::iterator B = I->getParent()->begin();
873 MachineBasicBlock::iterator E = I->getParent()->end();
874 MachineInstr *MemMI = I;
875 MachineBasicBlock::iterator MBBI = I;
876 const MachineFunction &MF = *MemMI->getParent()->getParent();
878 unsigned DestReg = getLdStRegOp(MemMI).getReg();
879 unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
880 int Offset = getLdStOffsetOp(MemMI).getImm();
881 unsigned RegSize = TII->getRegClass(MemMI->getDesc(), 0, TRI, MF)->getSize();
883 // If the load/store is the first instruction in the block, there's obviously
884 // not any matching update. Ditto if the memory offset isn't zero.
885 if (MBBI == B || Offset != 0)
887 // If the base register overlaps the destination register, we can't
889 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg))
892 // Track which registers have been modified and used between the first insn
893 // (inclusive) and the second insn.
894 BitVector ModifiedRegs, UsedRegs;
895 ModifiedRegs.resize(TRI->getNumRegs());
896 UsedRegs.resize(TRI->getNumRegs());
898 for (unsigned Count = 0; MBBI != B; --MBBI) {
899 MachineInstr *MI = MBBI;
900 // Skip DBG_VALUE instructions. Otherwise debug info can affect the
901 // optimization by changing how far we scan.
902 if (MI->isDebugValue())
905 // Now that we know this is a real instruction, count it.
908 // If we found a match, return it.
909 if (isMatchingUpdateInsn(MI, BaseReg, RegSize))
912 // Update the status of what the instruction clobbered and used.
913 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
915 // Otherwise, if the base register is used or modified, we have no match, so
917 if (ModifiedRegs[BaseReg] || UsedRegs[BaseReg])
923 bool AArch64LoadStoreOpt::optimizeBlock(MachineBasicBlock &MBB) {
924 bool Modified = false;
925 // Two tranformations to do here:
926 // 1) Find loads and stores that can be merged into a single load or store
933 // 2) Find base register updates that can be merged into the load or store
934 // as a base-reg writeback.
941 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
943 MachineInstr *MI = MBBI;
944 switch (MI->getOpcode()) {
946 // Just move on to the next instruction.
949 case AArch64::STRSui:
950 case AArch64::STRDui:
951 case AArch64::STRQui:
952 case AArch64::STRXui:
953 case AArch64::STRWui:
954 case AArch64::LDRSui:
955 case AArch64::LDRDui:
956 case AArch64::LDRQui:
957 case AArch64::LDRXui:
958 case AArch64::LDRWui:
959 case AArch64::LDRSWui:
960 // do the unscaled versions as well
961 case AArch64::STURSi:
962 case AArch64::STURDi:
963 case AArch64::STURQi:
964 case AArch64::STURWi:
965 case AArch64::STURXi:
966 case AArch64::LDURSi:
967 case AArch64::LDURDi:
968 case AArch64::LDURQi:
969 case AArch64::LDURWi:
970 case AArch64::LDURXi:
971 case AArch64::LDURSWi: {
972 // If this is a volatile load/store, don't mess with it.
973 if (MI->hasOrderedMemoryRef()) {
977 // Make sure this is a reg+imm (as opposed to an address reloc).
978 if (!getLdStOffsetOp(MI).isImm()) {
982 // Check if this load/store has a hint to avoid pair formation.
983 // MachineMemOperands hints are set by the AArch64StorePairSuppress pass.
984 if (TII->isLdStPairSuppressed(MI)) {
988 // Look ahead up to ScanLimit instructions for a pairable instruction.
990 MachineBasicBlock::iterator Paired =
991 findMatchingInsn(MBBI, Flags, ScanLimit);
993 // Merge the loads into a pair. Keeping the iterator straight is a
994 // pain, so we let the merge routine tell us what the next instruction
995 // is after it's done mucking about.
996 MBBI = mergePairedInsns(MBBI, Paired, Flags);
1000 if (isUnscaledLdSt(MI))
1001 ++NumUnscaledPairCreated;
1007 // FIXME: Do the other instructions.
1011 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1013 MachineInstr *MI = MBBI;
1014 // Do update merging. It's simpler to keep this separate from the above
1015 // switch, though not strictly necessary.
1016 unsigned Opc = MI->getOpcode();
1019 // Just move on to the next instruction.
1022 case AArch64::STRSui:
1023 case AArch64::STRDui:
1024 case AArch64::STRQui:
1025 case AArch64::STRXui:
1026 case AArch64::STRWui:
1027 case AArch64::LDRSui:
1028 case AArch64::LDRDui:
1029 case AArch64::LDRQui:
1030 case AArch64::LDRXui:
1031 case AArch64::LDRWui:
1032 // do the unscaled versions as well
1033 case AArch64::STURSi:
1034 case AArch64::STURDi:
1035 case AArch64::STURQi:
1036 case AArch64::STURWi:
1037 case AArch64::STURXi:
1038 case AArch64::LDURSi:
1039 case AArch64::LDURDi:
1040 case AArch64::LDURQi:
1041 case AArch64::LDURWi:
1042 case AArch64::LDURXi: {
1043 // Make sure this is a reg+imm (as opposed to an address reloc).
1044 if (!getLdStOffsetOp(MI).isImm()) {
1048 // Look ahead up to ScanLimit instructions for a mergable instruction.
1049 MachineBasicBlock::iterator Update =
1050 findMatchingUpdateInsnForward(MBBI, ScanLimit, 0);
1052 // Merge the update into the ld/st.
1053 MBBI = mergePostIdxUpdateInsn(MBBI, Update);
1058 // Don't know how to handle pre/post-index versions, so move to the next
1060 if (isUnscaledLdSt(Opc)) {
1065 // Look back to try to find a pre-index instruction. For example,
1069 // ldr x1, [x0, #8]!
1070 Update = findMatchingUpdateInsnBackward(MBBI, ScanLimit);
1072 // Merge the update into the ld/st.
1073 MBBI = mergePreIdxUpdateInsn(MBBI, Update);
1079 // Look forward to try to find a post-index instruction. For example,
1080 // ldr x1, [x0, #64]
1083 // ldr x1, [x0, #64]!
1085 // The immediate in the load/store is scaled by the size of the register
1086 // being loaded. The immediate in the add we're looking for,
1087 // however, is not, so adjust here.
1088 int Value = MI->getOperand(2).getImm() *
1089 TII->getRegClass(MI->getDesc(), 0, TRI, *(MBB.getParent()))
1091 Update = findMatchingUpdateInsnForward(MBBI, ScanLimit, Value);
1093 // Merge the update into the ld/st.
1094 MBBI = mergePreIdxUpdateInsn(MBBI, Update);
1100 // Nothing found. Just move to the next instruction.
1104 // FIXME: Do the other instructions.
1111 bool AArch64LoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
1112 TII = static_cast<const AArch64InstrInfo *>(Fn.getSubtarget().getInstrInfo());
1113 TRI = Fn.getSubtarget().getRegisterInfo();
1115 bool Modified = false;
1116 for (auto &MBB : Fn)
1117 Modified |= optimizeBlock(MBB);
1122 // FIXME: Do we need/want a pre-alloc pass like ARM has to try to keep
1123 // loads and stores near one another?
1125 /// createAArch64LoadStoreOptimizationPass - returns an instance of the
1126 /// load / store optimization pass.
1127 FunctionPass *llvm::createAArch64LoadStoreOptimizationPass() {
1128 return new AArch64LoadStoreOpt();