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 // 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::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:
183 case AArch64::LDRDui:
184 case AArch64::LDURDi:
185 case AArch64::LDRXui:
186 case AArch64::LDURXi:
187 case AArch64::STRDui:
188 case AArch64::STURDi:
189 case AArch64::STRXui:
190 case AArch64::STURXi:
192 case AArch64::LDRQui:
193 case AArch64::LDURQi:
194 case AArch64::STRQui:
195 case AArch64::STURQi:
200 static unsigned getMatchingNonSExtOpcode(unsigned Opc,
201 bool *IsValidLdStrOpc = nullptr) {
203 *IsValidLdStrOpc = true;
207 *IsValidLdStrOpc = false;
209 case AArch64::STRDui:
210 case AArch64::STURDi:
211 case AArch64::STRQui:
212 case AArch64::STURQi:
213 case AArch64::STRWui:
214 case AArch64::STURWi:
215 case AArch64::STRXui:
216 case AArch64::STURXi:
217 case AArch64::LDRDui:
218 case AArch64::LDURDi:
219 case AArch64::LDRQui:
220 case AArch64::LDURQi:
221 case AArch64::LDRWui:
222 case AArch64::LDURWi:
223 case AArch64::LDRXui:
224 case AArch64::LDURXi:
225 case AArch64::STRSui:
226 case AArch64::STURSi:
227 case AArch64::LDRSui:
228 case AArch64::LDURSi:
230 case AArch64::LDRSWui:
231 return AArch64::LDRWui;
232 case AArch64::LDURSWi:
233 return AArch64::LDURWi;
237 static unsigned getMatchingPairOpcode(unsigned Opc) {
240 llvm_unreachable("Opcode has no pairwise equivalent!");
241 case AArch64::STRSui:
242 case AArch64::STURSi:
243 return AArch64::STPSi;
244 case AArch64::STRDui:
245 case AArch64::STURDi:
246 return AArch64::STPDi;
247 case AArch64::STRQui:
248 case AArch64::STURQi:
249 return AArch64::STPQi;
250 case AArch64::STRWui:
251 case AArch64::STURWi:
252 return AArch64::STPWi;
253 case AArch64::STRXui:
254 case AArch64::STURXi:
255 return AArch64::STPXi;
256 case AArch64::LDRSui:
257 case AArch64::LDURSi:
258 return AArch64::LDPSi;
259 case AArch64::LDRDui:
260 case AArch64::LDURDi:
261 return AArch64::LDPDi;
262 case AArch64::LDRQui:
263 case AArch64::LDURQi:
264 return AArch64::LDPQi;
265 case AArch64::LDRWui:
266 case AArch64::LDURWi:
267 return AArch64::LDPWi;
268 case AArch64::LDRXui:
269 case AArch64::LDURXi:
270 return AArch64::LDPXi;
271 case AArch64::LDRSWui:
272 case AArch64::LDURSWi:
273 return AArch64::LDPSWi;
277 static unsigned getPreIndexedOpcode(unsigned Opc) {
280 llvm_unreachable("Opcode has no pre-indexed equivalent!");
281 case AArch64::STRSui:
282 return AArch64::STRSpre;
283 case AArch64::STRDui:
284 return AArch64::STRDpre;
285 case AArch64::STRQui:
286 return AArch64::STRQpre;
287 case AArch64::STRWui:
288 return AArch64::STRWpre;
289 case AArch64::STRXui:
290 return AArch64::STRXpre;
291 case AArch64::LDRSui:
292 return AArch64::LDRSpre;
293 case AArch64::LDRDui:
294 return AArch64::LDRDpre;
295 case AArch64::LDRQui:
296 return AArch64::LDRQpre;
297 case AArch64::LDRWui:
298 return AArch64::LDRWpre;
299 case AArch64::LDRXui:
300 return AArch64::LDRXpre;
301 case AArch64::LDRSWui:
302 return AArch64::LDRSWpre;
304 return AArch64::LDPSpre;
306 return AArch64::LDPDpre;
308 return AArch64::LDPQpre;
310 return AArch64::LDPWpre;
312 return AArch64::LDPXpre;
314 return AArch64::STPSpre;
316 return AArch64::STPDpre;
318 return AArch64::STPQpre;
320 return AArch64::STPWpre;
322 return AArch64::STPXpre;
326 static unsigned getPostIndexedOpcode(unsigned Opc) {
329 llvm_unreachable("Opcode has no post-indexed wise equivalent!");
330 case AArch64::STRSui:
331 return AArch64::STRSpost;
332 case AArch64::STRDui:
333 return AArch64::STRDpost;
334 case AArch64::STRQui:
335 return AArch64::STRQpost;
336 case AArch64::STRWui:
337 return AArch64::STRWpost;
338 case AArch64::STRXui:
339 return AArch64::STRXpost;
340 case AArch64::LDRSui:
341 return AArch64::LDRSpost;
342 case AArch64::LDRDui:
343 return AArch64::LDRDpost;
344 case AArch64::LDRQui:
345 return AArch64::LDRQpost;
346 case AArch64::LDRWui:
347 return AArch64::LDRWpost;
348 case AArch64::LDRXui:
349 return AArch64::LDRXpost;
350 case AArch64::LDRSWui:
351 return AArch64::LDRSWpost;
353 return AArch64::LDPSpost;
355 return AArch64::LDPDpost;
357 return AArch64::LDPQpost;
359 return AArch64::LDPWpost;
361 return AArch64::LDPXpost;
363 return AArch64::STPSpost;
365 return AArch64::STPDpost;
367 return AArch64::STPQpost;
369 return AArch64::STPWpost;
371 return AArch64::STPXpost;
375 static bool isPairedLdSt(const MachineInstr *MI) {
376 switch (MI->getOpcode()) {
393 static const MachineOperand &getLdStRegOp(const MachineInstr *MI,
394 unsigned PairedRegOp = 0) {
395 assert(PairedRegOp < 2 && "Unexpected register operand idx.");
396 unsigned Idx = isPairedLdSt(MI) ? PairedRegOp : 0;
397 return MI->getOperand(Idx);
400 static const MachineOperand &getLdStBaseOp(const MachineInstr *MI) {
401 unsigned Idx = isPairedLdSt(MI) ? 2 : 1;
402 return MI->getOperand(Idx);
405 static const MachineOperand &getLdStOffsetOp(const MachineInstr *MI) {
406 unsigned Idx = isPairedLdSt(MI) ? 3 : 2;
407 return MI->getOperand(Idx);
410 MachineBasicBlock::iterator
411 AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I,
412 MachineBasicBlock::iterator Paired,
413 const LdStPairFlags &Flags) {
414 MachineBasicBlock::iterator NextI = I;
416 // If NextI is the second of the two instructions to be merged, we need
417 // to skip one further. Either way we merge will invalidate the iterator,
418 // and we don't need to scan the new instruction, as it's a pairwise
419 // instruction, which we're not considering for further action anyway.
423 int SExtIdx = Flags.getSExtIdx();
425 SExtIdx == -1 ? I->getOpcode() : getMatchingNonSExtOpcode(I->getOpcode());
426 bool IsUnscaled = isUnscaledLdSt(Opc);
428 IsUnscaled && EnableAArch64UnscaledMemOp ? getMemSize(I) : 1;
430 bool MergeForward = Flags.getMergeForward();
431 unsigned NewOpc = getMatchingPairOpcode(Opc);
432 // Insert our new paired instruction after whichever of the paired
433 // instructions MergeForward indicates.
434 MachineBasicBlock::iterator InsertionPoint = MergeForward ? Paired : I;
435 // Also based on MergeForward is from where we copy the base register operand
436 // so we get the flags compatible with the input code.
437 const MachineOperand &BaseRegOp =
438 MergeForward ? getLdStBaseOp(Paired) : getLdStBaseOp(I);
440 // Which register is Rt and which is Rt2 depends on the offset order.
441 MachineInstr *RtMI, *Rt2MI;
442 if (getLdStOffsetOp(I).getImm() ==
443 getLdStOffsetOp(Paired).getImm() + OffsetStride) {
446 // Here we swapped the assumption made for SExtIdx.
447 // I.e., we turn ldp I, Paired into ldp Paired, I.
448 // Update the index accordingly.
450 SExtIdx = (SExtIdx + 1) % 2;
456 int OffsetImm = getLdStOffsetOp(RtMI).getImm();
457 if (IsUnscaled && EnableAArch64UnscaledMemOp)
458 OffsetImm /= OffsetStride;
460 // Construct the new instruction.
461 MachineInstrBuilder MIB = BuildMI(*I->getParent(), InsertionPoint,
462 I->getDebugLoc(), TII->get(NewOpc))
463 .addOperand(getLdStRegOp(RtMI))
464 .addOperand(getLdStRegOp(Rt2MI))
465 .addOperand(BaseRegOp)
469 // FIXME: Do we need/want to copy the mem operands from the source
470 // instructions? Probably. What uses them after this?
472 DEBUG(dbgs() << "Creating pair load/store. Replacing instructions:\n ");
473 DEBUG(I->print(dbgs()));
474 DEBUG(dbgs() << " ");
475 DEBUG(Paired->print(dbgs()));
476 DEBUG(dbgs() << " with instruction:\n ");
479 // Generate the sign extension for the proper result of the ldp.
480 // I.e., with X1, that would be:
481 // %W1<def> = KILL %W1, %X1<imp-def>
482 // %X1<def> = SBFMXri %X1<kill>, 0, 31
483 MachineOperand &DstMO = MIB->getOperand(SExtIdx);
484 // Right now, DstMO has the extended register, since it comes from an
486 unsigned DstRegX = DstMO.getReg();
487 // Get the W variant of that register.
488 unsigned DstRegW = TRI->getSubReg(DstRegX, AArch64::sub_32);
489 // Update the result of LDP to use the W instead of the X variant.
490 DstMO.setReg(DstRegW);
491 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
492 DEBUG(dbgs() << "\n");
493 // Make the machine verifier happy by providing a definition for
495 // Insert this definition right after the generated LDP, i.e., before
497 MachineInstrBuilder MIBKill =
498 BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
499 TII->get(TargetOpcode::KILL), DstRegW)
501 .addReg(DstRegX, RegState::Define);
502 MIBKill->getOperand(2).setImplicit();
503 // Create the sign extension.
504 MachineInstrBuilder MIBSXTW =
505 BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
506 TII->get(AArch64::SBFMXri), DstRegX)
511 DEBUG(dbgs() << " Extend operand:\n ");
512 DEBUG(((MachineInstr *)MIBSXTW)->print(dbgs()));
513 DEBUG(dbgs() << "\n");
515 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
516 DEBUG(dbgs() << "\n");
519 // Erase the old instructions.
520 I->eraseFromParent();
521 Paired->eraseFromParent();
526 /// trackRegDefsUses - Remember what registers the specified instruction uses
528 static void trackRegDefsUses(const MachineInstr *MI, BitVector &ModifiedRegs,
530 const TargetRegisterInfo *TRI) {
531 for (const MachineOperand &MO : MI->operands()) {
533 ModifiedRegs.setBitsNotInMask(MO.getRegMask());
537 unsigned Reg = MO.getReg();
539 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
540 ModifiedRegs.set(*AI);
542 assert(MO.isUse() && "Reg operand not a def and not a use?!?");
543 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
549 static bool inBoundsForPair(bool IsUnscaled, int Offset, int OffsetStride) {
550 // Convert the byte-offset used by unscaled into an "element" offset used
551 // by the scaled pair load/store instructions.
553 Offset /= OffsetStride;
555 return Offset <= 63 && Offset >= -64;
558 // Do alignment, specialized to power of 2 and for signed ints,
559 // avoiding having to do a C-style cast from uint_64t to int when
560 // using RoundUpToAlignment from include/llvm/Support/MathExtras.h.
561 // FIXME: Move this function to include/MathExtras.h?
562 static int alignTo(int Num, int PowOf2) {
563 return (Num + PowOf2 - 1) & ~(PowOf2 - 1);
566 static bool mayAlias(MachineInstr *MIa, MachineInstr *MIb,
567 const AArch64InstrInfo *TII) {
568 // One of the instructions must modify memory.
569 if (!MIa->mayStore() && !MIb->mayStore())
572 // Both instructions must be memory operations.
573 if (!MIa->mayLoadOrStore() && !MIb->mayLoadOrStore())
576 return !TII->areMemAccessesTriviallyDisjoint(MIa, MIb);
579 static bool mayAlias(MachineInstr *MIa,
580 SmallVectorImpl<MachineInstr *> &MemInsns,
581 const AArch64InstrInfo *TII) {
582 for (auto &MIb : MemInsns)
583 if (mayAlias(MIa, MIb, TII))
589 /// findMatchingInsn - Scan the instructions looking for a load/store that can
590 /// be combined with the current instruction into a load/store pair.
591 MachineBasicBlock::iterator
592 AArch64LoadStoreOpt::findMatchingInsn(MachineBasicBlock::iterator I,
593 LdStPairFlags &Flags,
595 MachineBasicBlock::iterator E = I->getParent()->end();
596 MachineBasicBlock::iterator MBBI = I;
597 MachineInstr *FirstMI = I;
600 unsigned Opc = FirstMI->getOpcode();
601 bool MayLoad = FirstMI->mayLoad();
602 bool IsUnscaled = isUnscaledLdSt(FirstMI);
603 unsigned Reg = getLdStRegOp(FirstMI).getReg();
604 unsigned BaseReg = getLdStBaseOp(FirstMI).getReg();
605 int Offset = getLdStOffsetOp(FirstMI).getImm();
607 // Early exit if the first instruction modifies the base register.
608 // e.g., ldr x0, [x0]
609 if (FirstMI->modifiesRegister(BaseReg, TRI))
612 // Early exit if the offset if not possible to match. (6 bits of positive
613 // range, plus allow an extra one in case we find a later insn that matches
616 IsUnscaled && EnableAArch64UnscaledMemOp ? getMemSize(FirstMI) : 1;
617 if (!inBoundsForPair(IsUnscaled, Offset, OffsetStride))
620 // Track which registers have been modified and used between the first insn
621 // (inclusive) and the second insn.
622 BitVector ModifiedRegs, UsedRegs;
623 ModifiedRegs.resize(TRI->getNumRegs());
624 UsedRegs.resize(TRI->getNumRegs());
626 // Remember any instructions that read/write memory between FirstMI and MI.
627 SmallVector<MachineInstr *, 4> MemInsns;
629 for (unsigned Count = 0; MBBI != E && Count < Limit; ++MBBI) {
630 MachineInstr *MI = MBBI;
631 // Skip DBG_VALUE instructions. Otherwise debug info can affect the
632 // optimization by changing how far we scan.
633 if (MI->isDebugValue())
636 // Now that we know this is a real instruction, count it.
639 bool CanMergeOpc = Opc == MI->getOpcode();
640 Flags.setSExtIdx(-1);
642 bool IsValidLdStrOpc;
643 unsigned NonSExtOpc = getMatchingNonSExtOpcode(Opc, &IsValidLdStrOpc);
644 assert(IsValidLdStrOpc &&
645 "Given Opc should be a Load or Store with an immediate");
646 // Opc will be the first instruction in the pair.
647 Flags.setSExtIdx(NonSExtOpc == (unsigned)Opc ? 1 : 0);
648 CanMergeOpc = NonSExtOpc == getMatchingNonSExtOpcode(MI->getOpcode());
651 if (CanMergeOpc && getLdStOffsetOp(MI).isImm()) {
652 assert(MI->mayLoadOrStore() && "Expected memory operation.");
653 // If we've found another instruction with the same opcode, check to see
654 // if the base and offset are compatible with our starting instruction.
655 // These instructions all have scaled immediate operands, so we just
656 // check for +1/-1. Make sure to check the new instruction offset is
657 // actually an immediate and not a symbolic reference destined for
660 // Pairwise instructions have a 7-bit signed offset field. Single insns
661 // have a 12-bit unsigned offset field. To be a valid combine, the
662 // final offset must be in range.
663 unsigned MIBaseReg = getLdStBaseOp(MI).getReg();
664 int MIOffset = getLdStOffsetOp(MI).getImm();
665 if (BaseReg == MIBaseReg && ((Offset == MIOffset + OffsetStride) ||
666 (Offset + OffsetStride == MIOffset))) {
667 int MinOffset = Offset < MIOffset ? Offset : MIOffset;
668 // If this is a volatile load/store that otherwise matched, stop looking
669 // as something is going on that we don't have enough information to
670 // safely transform. Similarly, stop if we see a hint to avoid pairs.
671 if (MI->hasOrderedMemoryRef() || TII->isLdStPairSuppressed(MI))
673 // If the resultant immediate offset of merging these instructions
674 // is out of range for a pairwise instruction, bail and keep looking.
675 bool MIIsUnscaled = isUnscaledLdSt(MI);
676 if (!inBoundsForPair(MIIsUnscaled, MinOffset, OffsetStride)) {
677 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
678 MemInsns.push_back(MI);
681 // If the alignment requirements of the paired (scaled) instruction
682 // can't express the offset of the unscaled input, bail and keep
684 if (IsUnscaled && EnableAArch64UnscaledMemOp &&
685 (alignTo(MinOffset, OffsetStride) != MinOffset)) {
686 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
687 MemInsns.push_back(MI);
690 // If the destination register of the loads is the same register, bail
691 // and keep looking. A load-pair instruction with both destination
692 // registers the same is UNPREDICTABLE and will result in an exception.
693 if (MayLoad && Reg == getLdStRegOp(MI).getReg()) {
694 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
695 MemInsns.push_back(MI);
699 // If the Rt of the second instruction was not modified or used between
700 // the two instructions and none of the instructions between the second
701 // and first alias with the second, we can combine the second into the
703 if (!ModifiedRegs[getLdStRegOp(MI).getReg()] &&
704 !(MI->mayLoad() && UsedRegs[getLdStRegOp(MI).getReg()]) &&
705 !mayAlias(MI, MemInsns, TII)) {
706 Flags.setMergeForward(false);
710 // Likewise, if the Rt of the first instruction is not modified or used
711 // between the two instructions and none of the instructions between the
712 // first and the second alias with the first, we can combine the first
714 if (!ModifiedRegs[getLdStRegOp(FirstMI).getReg()] &&
715 !(MayLoad && UsedRegs[getLdStRegOp(FirstMI).getReg()]) &&
716 !mayAlias(FirstMI, MemInsns, TII)) {
717 Flags.setMergeForward(true);
720 // Unable to combine these instructions due to interference in between.
725 // If the instruction wasn't a matching load or store. Stop searching if we
726 // encounter a call instruction that might modify memory.
730 // Update modified / uses register lists.
731 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
733 // Otherwise, if the base register is modified, we have no match, so
735 if (ModifiedRegs[BaseReg])
738 // Update list of instructions that read/write memory.
739 if (MI->mayLoadOrStore())
740 MemInsns.push_back(MI);
745 MachineBasicBlock::iterator
746 AArch64LoadStoreOpt::mergeUpdateInsn(MachineBasicBlock::iterator I,
747 MachineBasicBlock::iterator Update,
749 assert((Update->getOpcode() == AArch64::ADDXri ||
750 Update->getOpcode() == AArch64::SUBXri) &&
751 "Unexpected base register update instruction to merge!");
752 MachineBasicBlock::iterator NextI = I;
753 // Return the instruction following the merged instruction, which is
754 // the instruction following our unmerged load. Unless that's the add/sub
755 // instruction we're merging, in which case it's the one after that.
756 if (++NextI == Update)
759 int Value = Update->getOperand(2).getImm();
760 assert(AArch64_AM::getShiftValue(Update->getOperand(3).getImm()) == 0 &&
761 "Can't merge 1 << 12 offset into pre-/post-indexed load / store");
762 if (Update->getOpcode() == AArch64::SUBXri)
765 unsigned NewOpc = IsPreIdx ? getPreIndexedOpcode(I->getOpcode())
766 : getPostIndexedOpcode(I->getOpcode());
767 MachineInstrBuilder MIB;
768 if (!isPairedLdSt(I)) {
769 // Non-paired instruction.
770 MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc))
771 .addOperand(getLdStRegOp(Update))
772 .addOperand(getLdStRegOp(I))
773 .addOperand(getLdStBaseOp(I))
776 // Paired instruction.
777 const MachineFunction &MF = *I->getParent()->getParent();
778 int Scale = TII->getRegClass(I->getDesc(), 0, TRI, MF)->getSize();
779 MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc))
780 .addOperand(getLdStRegOp(Update))
781 .addOperand(getLdStRegOp(I, 0))
782 .addOperand(getLdStRegOp(I, 1))
783 .addOperand(getLdStBaseOp(I))
784 .addImm(Value / Scale);
789 DEBUG(dbgs() << "Creating pre-indexed load/store.");
791 DEBUG(dbgs() << "Creating post-indexed load/store.");
792 DEBUG(dbgs() << " Replacing instructions:\n ");
793 DEBUG(I->print(dbgs()));
794 DEBUG(dbgs() << " ");
795 DEBUG(Update->print(dbgs()));
796 DEBUG(dbgs() << " with instruction:\n ");
797 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
798 DEBUG(dbgs() << "\n");
800 // Erase the old instructions for the block.
801 I->eraseFromParent();
802 Update->eraseFromParent();
807 bool AArch64LoadStoreOpt::isMatchingUpdateInsn(MachineInstr *MemMI,
809 unsigned BaseReg, int Offset) {
810 switch (MI->getOpcode()) {
813 case AArch64::SUBXri:
814 // Negate the offset for a SUB instruction.
817 case AArch64::ADDXri:
818 // Make sure it's a vanilla immediate operand, not a relocation or
819 // anything else we can't handle.
820 if (!MI->getOperand(2).isImm())
822 // Watch out for 1 << 12 shifted value.
823 if (AArch64_AM::getShiftValue(MI->getOperand(3).getImm()))
826 // The update instruction source and destination register must be the
827 // same as the load/store base register.
828 if (MI->getOperand(0).getReg() != BaseReg ||
829 MI->getOperand(1).getReg() != BaseReg)
832 bool IsPairedInsn = isPairedLdSt(MemMI);
833 int UpdateOffset = MI->getOperand(2).getImm();
834 // For non-paired load/store instructions, the immediate must fit in a
835 // signed 9-bit integer.
836 if (!IsPairedInsn && (UpdateOffset > 255 || UpdateOffset < -256))
839 // For paired load/store instructions, the immediate must be a multiple of
840 // the scaling factor. The scaled offset must also fit into a signed 7-bit
843 const MachineFunction &MF = *MemMI->getParent()->getParent();
844 int Scale = TII->getRegClass(MemMI->getDesc(), 0, TRI, MF)->getSize();
845 if (UpdateOffset % Scale != 0)
848 int ScaledOffset = UpdateOffset / Scale;
849 if (ScaledOffset > 64 || ScaledOffset < -64)
853 // If we have a non-zero Offset, we check that it matches the amount
854 // we're adding to the register.
855 if (!Offset || Offset == MI->getOperand(2).getImm())
862 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnForward(
863 MachineBasicBlock::iterator I, unsigned Limit, int Value) {
864 MachineBasicBlock::iterator E = I->getParent()->end();
865 MachineInstr *MemMI = I;
866 MachineBasicBlock::iterator MBBI = I;
867 const MachineFunction &MF = *MemMI->getParent()->getParent();
869 unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
870 int Offset = getLdStOffsetOp(MemMI).getImm() *
871 TII->getRegClass(MemMI->getDesc(), 0, TRI, MF)->getSize();
873 // If the base register overlaps a destination register, we can't
875 bool IsPairedInsn = isPairedLdSt(MemMI);
876 for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) {
877 unsigned DestReg = getLdStRegOp(MemMI, i).getReg();
878 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg))
882 // Scan forward looking for post-index opportunities.
883 // Updating instructions can't be formed if the memory insn already
884 // has an offset other than the value we're looking for.
888 // Track which registers have been modified and used between the first insn
889 // (inclusive) and the second insn.
890 BitVector ModifiedRegs, UsedRegs;
891 ModifiedRegs.resize(TRI->getNumRegs());
892 UsedRegs.resize(TRI->getNumRegs());
894 for (unsigned Count = 0; MBBI != E; ++MBBI) {
895 MachineInstr *MI = MBBI;
896 // Skip DBG_VALUE instructions. Otherwise debug info can affect the
897 // optimization by changing how far we scan.
898 if (MI->isDebugValue())
901 // Now that we know this is a real instruction, count it.
904 // If we found a match, return it.
905 if (isMatchingUpdateInsn(I, MI, BaseReg, Value))
908 // Update the status of what the instruction clobbered and used.
909 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
911 // Otherwise, if the base register is used or modified, we have no match, so
913 if (ModifiedRegs[BaseReg] || UsedRegs[BaseReg])
919 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnBackward(
920 MachineBasicBlock::iterator I, unsigned Limit) {
921 MachineBasicBlock::iterator B = I->getParent()->begin();
922 MachineBasicBlock::iterator E = I->getParent()->end();
923 MachineInstr *MemMI = I;
924 MachineBasicBlock::iterator MBBI = I;
925 const MachineFunction &MF = *MemMI->getParent()->getParent();
927 unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
928 int Offset = getLdStOffsetOp(MemMI).getImm();
929 unsigned RegSize = TII->getRegClass(MemMI->getDesc(), 0, TRI, MF)->getSize();
931 // If the load/store is the first instruction in the block, there's obviously
932 // not any matching update. Ditto if the memory offset isn't zero.
933 if (MBBI == B || Offset != 0)
935 // If the base register overlaps a destination register, we can't
937 bool IsPairedInsn = isPairedLdSt(MemMI);
938 for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) {
939 unsigned DestReg = getLdStRegOp(MemMI, i).getReg();
940 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg))
944 // Track which registers have been modified and used between the first insn
945 // (inclusive) and the second insn.
946 BitVector ModifiedRegs, UsedRegs;
947 ModifiedRegs.resize(TRI->getNumRegs());
948 UsedRegs.resize(TRI->getNumRegs());
950 for (unsigned Count = 0; MBBI != B; --MBBI) {
951 MachineInstr *MI = MBBI;
952 // Skip DBG_VALUE instructions. Otherwise debug info can affect the
953 // optimization by changing how far we scan.
954 if (MI->isDebugValue())
957 // Now that we know this is a real instruction, count it.
960 // If we found a match, return it.
961 if (isMatchingUpdateInsn(I, MI, BaseReg, RegSize))
964 // Update the status of what the instruction clobbered and used.
965 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
967 // Otherwise, if the base register is used or modified, we have no match, so
969 if (ModifiedRegs[BaseReg] || UsedRegs[BaseReg])
975 bool AArch64LoadStoreOpt::optimizeBlock(MachineBasicBlock &MBB) {
976 bool Modified = false;
977 // Two tranformations to do here:
978 // 1) Find loads and stores that can be merged into a single load or store
985 // 2) Find base register updates that can be merged into the load or store
986 // as a base-reg writeback.
993 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
995 MachineInstr *MI = MBBI;
996 switch (MI->getOpcode()) {
998 // Just move on to the next instruction.
1001 // Scaled instructions.
1002 case AArch64::STRSui:
1003 case AArch64::STRDui:
1004 case AArch64::STRQui:
1005 case AArch64::STRXui:
1006 case AArch64::STRWui:
1007 case AArch64::LDRSui:
1008 case AArch64::LDRDui:
1009 case AArch64::LDRQui:
1010 case AArch64::LDRXui:
1011 case AArch64::LDRWui:
1012 case AArch64::LDRSWui:
1013 // Unscaled instructions.
1014 case AArch64::STURSi:
1015 case AArch64::STURDi:
1016 case AArch64::STURQi:
1017 case AArch64::STURWi:
1018 case AArch64::STURXi:
1019 case AArch64::LDURSi:
1020 case AArch64::LDURDi:
1021 case AArch64::LDURQi:
1022 case AArch64::LDURWi:
1023 case AArch64::LDURXi:
1024 case AArch64::LDURSWi: {
1025 // If this is a volatile load/store, don't mess with it.
1026 if (MI->hasOrderedMemoryRef()) {
1030 // Make sure this is a reg+imm (as opposed to an address reloc).
1031 if (!getLdStOffsetOp(MI).isImm()) {
1035 // Check if this load/store has a hint to avoid pair formation.
1036 // MachineMemOperands hints are set by the AArch64StorePairSuppress pass.
1037 if (TII->isLdStPairSuppressed(MI)) {
1041 // Look ahead up to ScanLimit instructions for a pairable instruction.
1042 LdStPairFlags Flags;
1043 MachineBasicBlock::iterator Paired =
1044 findMatchingInsn(MBBI, Flags, ScanLimit);
1047 if (isUnscaledLdSt(MI))
1048 ++NumUnscaledPairCreated;
1050 // Merge the loads into a pair. Keeping the iterator straight is a
1051 // pain, so we let the merge routine tell us what the next instruction
1052 // is after it's done mucking about.
1053 MBBI = mergePairedInsns(MBBI, Paired, Flags);
1060 // FIXME: Do the other instructions.
1064 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1066 MachineInstr *MI = MBBI;
1067 // Do update merging. It's simpler to keep this separate from the above
1068 // switch, though not strictly necessary.
1069 unsigned Opc = MI->getOpcode();
1072 // Just move on to the next instruction.
1075 // Scaled instructions.
1076 case AArch64::STRSui:
1077 case AArch64::STRDui:
1078 case AArch64::STRQui:
1079 case AArch64::STRXui:
1080 case AArch64::STRWui:
1081 case AArch64::LDRSui:
1082 case AArch64::LDRDui:
1083 case AArch64::LDRQui:
1084 case AArch64::LDRXui:
1085 case AArch64::LDRWui:
1086 // Unscaled instructions.
1087 case AArch64::STURSi:
1088 case AArch64::STURDi:
1089 case AArch64::STURQi:
1090 case AArch64::STURWi:
1091 case AArch64::STURXi:
1092 case AArch64::LDURSi:
1093 case AArch64::LDURDi:
1094 case AArch64::LDURQi:
1095 case AArch64::LDURWi:
1096 case AArch64::LDURXi:
1097 // Paired instructions.
1098 case AArch64::LDPSi:
1099 case AArch64::LDPDi:
1100 case AArch64::LDPQi:
1101 case AArch64::LDPWi:
1102 case AArch64::LDPXi:
1103 case AArch64::STPSi:
1104 case AArch64::STPDi:
1105 case AArch64::STPQi:
1106 case AArch64::STPWi:
1107 case AArch64::STPXi: {
1108 // Make sure this is a reg+imm (as opposed to an address reloc).
1109 if (!getLdStOffsetOp(MI).isImm()) {
1113 // Look forward to try to form a post-index instruction. For example,
1115 // add x20, x20, #32
1117 // ldr x0, [x20], #32
1118 MachineBasicBlock::iterator Update =
1119 findMatchingUpdateInsnForward(MBBI, ScanLimit, 0);
1121 // Merge the update into the ld/st.
1122 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/false);
1127 // Don't know how to handle pre/post-index versions, so move to the next
1129 if (isUnscaledLdSt(Opc)) {
1134 // Look back to try to find a pre-index instruction. For example,
1138 // ldr x1, [x0, #8]!
1139 Update = findMatchingUpdateInsnBackward(MBBI, ScanLimit);
1141 // Merge the update into the ld/st.
1142 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true);
1147 // The immediate in the load/store is scaled by the size of the register
1148 // being loaded. The immediate in the add we're looking for,
1149 // however, is not, so adjust here.
1150 int Value = MI->getOperand(isPairedLdSt(MI) ? 3 : 2).getImm() *
1151 TII->getRegClass(MI->getDesc(), 0, TRI, *(MBB.getParent()))
1154 // FIXME: The immediate in the load/store should be scaled by the size of
1155 // the memory operation, not the size of the register being loaded/stored.
1156 // This works in general, but does not work for the LDPSW instruction,
1157 // which defines two 64-bit registers, but loads 32-bit values.
1159 // Look forward to try to find a post-index instruction. For example,
1160 // ldr x1, [x0, #64]
1163 // ldr x1, [x0, #64]!
1164 Update = findMatchingUpdateInsnForward(MBBI, ScanLimit, Value);
1166 // Merge the update into the ld/st.
1167 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true);
1173 // Nothing found. Just move to the next instruction.
1177 // FIXME: Do the other instructions.
1184 bool AArch64LoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
1185 TII = static_cast<const AArch64InstrInfo *>(Fn.getSubtarget().getInstrInfo());
1186 TRI = Fn.getSubtarget().getRegisterInfo();
1188 bool Modified = false;
1189 for (auto &MBB : Fn)
1190 Modified |= optimizeBlock(MBB);
1195 // FIXME: Do we need/want a pre-alloc pass like ARM has to try to keep
1196 // loads and stores near one another?
1198 /// createAArch64LoadStoreOptimizationPass - returns an instance of the
1199 /// load / store optimization pass.
1200 FunctionPass *llvm::createAArch64LoadStoreOptimizationPass() {
1201 return new AArch64LoadStoreOpt();