1 //===- AArch64LoadStoreOptimizer.cpp - AArch64 load/store opt. pass -------===//
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/ADT/StringRef.h"
22 #include "llvm/ADT/iterator_range.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/CodeGen/MachineBasicBlock.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineFunctionPass.h"
27 #include "llvm/CodeGen/MachineInstr.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineOperand.h"
30 #include "llvm/CodeGen/TargetRegisterInfo.h"
31 #include "llvm/IR/DebugLoc.h"
32 #include "llvm/MC/MCRegisterInfo.h"
33 #include "llvm/Pass.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/raw_ostream.h"
45 #define DEBUG_TYPE "aarch64-ldst-opt"
47 STATISTIC(NumPairCreated, "Number of load/store pair instructions generated");
48 STATISTIC(NumPostFolded, "Number of post-index updates folded");
49 STATISTIC(NumPreFolded, "Number of pre-index updates folded");
50 STATISTIC(NumUnscaledPairCreated,
51 "Number of load/store from unscaled generated");
52 STATISTIC(NumZeroStoresPromoted, "Number of narrow zero stores promoted");
53 STATISTIC(NumLoadsFromStoresPromoted, "Number of loads from stores promoted");
55 // The LdStLimit limits how far we search for load/store pairs.
56 static cl::opt<unsigned> LdStLimit("aarch64-load-store-scan-limit",
57 cl::init(20), cl::Hidden);
59 // The UpdateLimit limits how far we search for update instructions when we form
60 // pre-/post-index instructions.
61 static cl::opt<unsigned> UpdateLimit("aarch64-update-scan-limit", cl::init(100),
64 #define AARCH64_LOAD_STORE_OPT_NAME "AArch64 load / store optimization pass"
68 using LdStPairFlags = struct LdStPairFlags {
69 // If a matching instruction is found, MergeForward is set to true if the
70 // merge is to remove the first instruction and replace the second with
71 // a pair-wise insn, and false if the reverse is true.
72 bool MergeForward = false;
74 // SExtIdx gives the index of the result of the load pair that must be
75 // extended. The value of SExtIdx assumes that the paired load produces the
76 // value in this order: (I, returned iterator), i.e., -1 means no value has
77 // to be extended, 0 means I, and 1 means the returned iterator.
80 LdStPairFlags() = default;
82 void setMergeForward(bool V = true) { MergeForward = V; }
83 bool getMergeForward() const { return MergeForward; }
85 void setSExtIdx(int V) { SExtIdx = V; }
86 int getSExtIdx() const { return SExtIdx; }
89 struct AArch64LoadStoreOpt : public MachineFunctionPass {
92 AArch64LoadStoreOpt() : MachineFunctionPass(ID) {
93 initializeAArch64LoadStoreOptPass(*PassRegistry::getPassRegistry());
97 const AArch64InstrInfo *TII;
98 const TargetRegisterInfo *TRI;
99 const AArch64Subtarget *Subtarget;
101 // Track which registers have been modified and used.
102 BitVector ModifiedRegs, UsedRegs;
104 void getAnalysisUsage(AnalysisUsage &AU) const override {
105 AU.addRequired<AAResultsWrapperPass>();
106 MachineFunctionPass::getAnalysisUsage(AU);
109 // Scan the instructions looking for a load/store that can be combined
110 // with the current instruction into a load/store pair.
111 // Return the matching instruction if one is found, else MBB->end().
112 MachineBasicBlock::iterator findMatchingInsn(MachineBasicBlock::iterator I,
113 LdStPairFlags &Flags,
115 bool FindNarrowMerge);
117 // Scan the instructions looking for a store that writes to the address from
118 // which the current load instruction reads. Return true if one is found.
119 bool findMatchingStore(MachineBasicBlock::iterator I, unsigned Limit,
120 MachineBasicBlock::iterator &StoreI);
122 // Merge the two instructions indicated into a wider narrow store instruction.
123 MachineBasicBlock::iterator
124 mergeNarrowZeroStores(MachineBasicBlock::iterator I,
125 MachineBasicBlock::iterator MergeMI,
126 const LdStPairFlags &Flags);
128 // Merge the two instructions indicated into a single pair-wise instruction.
129 MachineBasicBlock::iterator
130 mergePairedInsns(MachineBasicBlock::iterator I,
131 MachineBasicBlock::iterator Paired,
132 const LdStPairFlags &Flags);
134 // Promote the load that reads directly from the address stored to.
135 MachineBasicBlock::iterator
136 promoteLoadFromStore(MachineBasicBlock::iterator LoadI,
137 MachineBasicBlock::iterator StoreI);
139 // Scan the instruction list to find a base register update that can
140 // be combined with the current instruction (a load or store) using
141 // pre or post indexed addressing with writeback. Scan forwards.
142 MachineBasicBlock::iterator
143 findMatchingUpdateInsnForward(MachineBasicBlock::iterator I,
144 int UnscaledOffset, unsigned Limit);
146 // Scan the instruction list to find a base register update that can
147 // be combined with the current instruction (a load or store) using
148 // pre or post indexed addressing with writeback. Scan backwards.
149 MachineBasicBlock::iterator
150 findMatchingUpdateInsnBackward(MachineBasicBlock::iterator I, unsigned Limit);
152 // Find an instruction that updates the base register of the ld/st
154 bool isMatchingUpdateInsn(MachineInstr &MemMI, MachineInstr &MI,
155 unsigned BaseReg, int Offset);
157 // Merge a pre- or post-index base register update into a ld/st instruction.
158 MachineBasicBlock::iterator
159 mergeUpdateInsn(MachineBasicBlock::iterator I,
160 MachineBasicBlock::iterator Update, bool IsPreIdx);
162 // Find and merge zero store instructions.
163 bool tryToMergeZeroStInst(MachineBasicBlock::iterator &MBBI);
165 // Find and pair ldr/str instructions.
166 bool tryToPairLdStInst(MachineBasicBlock::iterator &MBBI);
168 // Find and promote load instructions which read directly from store.
169 bool tryToPromoteLoadFromStore(MachineBasicBlock::iterator &MBBI);
171 // Find and merge a base register updates before or after a ld/st instruction.
172 bool tryToMergeLdStUpdate(MachineBasicBlock::iterator &MBBI);
174 bool optimizeBlock(MachineBasicBlock &MBB, bool EnableNarrowZeroStOpt);
176 bool runOnMachineFunction(MachineFunction &Fn) override;
178 MachineFunctionProperties getRequiredProperties() const override {
179 return MachineFunctionProperties().set(
180 MachineFunctionProperties::Property::NoVRegs);
183 StringRef getPassName() const override { return AARCH64_LOAD_STORE_OPT_NAME; }
186 char AArch64LoadStoreOpt::ID = 0;
188 } // end anonymous namespace
190 INITIALIZE_PASS(AArch64LoadStoreOpt, "aarch64-ldst-opt",
191 AARCH64_LOAD_STORE_OPT_NAME, false, false)
193 static bool isNarrowStore(unsigned Opc) {
197 case AArch64::STRBBui:
198 case AArch64::STURBBi:
199 case AArch64::STRHHui:
200 case AArch64::STURHHi:
205 // Scaling factor for unscaled load or store.
206 static int getMemScale(MachineInstr &MI) {
207 switch (MI.getOpcode()) {
209 llvm_unreachable("Opcode has unknown scale!");
210 case AArch64::LDRBBui:
211 case AArch64::LDURBBi:
212 case AArch64::LDRSBWui:
213 case AArch64::LDURSBWi:
214 case AArch64::STRBBui:
215 case AArch64::STURBBi:
217 case AArch64::LDRHHui:
218 case AArch64::LDURHHi:
219 case AArch64::LDRSHWui:
220 case AArch64::LDURSHWi:
221 case AArch64::STRHHui:
222 case AArch64::STURHHi:
224 case AArch64::LDRSui:
225 case AArch64::LDURSi:
226 case AArch64::LDRSWui:
227 case AArch64::LDURSWi:
228 case AArch64::LDRWui:
229 case AArch64::LDURWi:
230 case AArch64::STRSui:
231 case AArch64::STURSi:
232 case AArch64::STRWui:
233 case AArch64::STURWi:
235 case AArch64::LDPSWi:
240 case AArch64::LDRDui:
241 case AArch64::LDURDi:
242 case AArch64::LDRXui:
243 case AArch64::LDURXi:
244 case AArch64::STRDui:
245 case AArch64::STURDi:
246 case AArch64::STRXui:
247 case AArch64::STURXi:
253 case AArch64::LDRQui:
254 case AArch64::LDURQi:
255 case AArch64::STRQui:
256 case AArch64::STURQi:
263 static unsigned getMatchingNonSExtOpcode(unsigned Opc,
264 bool *IsValidLdStrOpc = nullptr) {
266 *IsValidLdStrOpc = true;
270 *IsValidLdStrOpc = false;
271 return std::numeric_limits<unsigned>::max();
272 case AArch64::STRDui:
273 case AArch64::STURDi:
274 case AArch64::STRQui:
275 case AArch64::STURQi:
276 case AArch64::STRBBui:
277 case AArch64::STURBBi:
278 case AArch64::STRHHui:
279 case AArch64::STURHHi:
280 case AArch64::STRWui:
281 case AArch64::STURWi:
282 case AArch64::STRXui:
283 case AArch64::STURXi:
284 case AArch64::LDRDui:
285 case AArch64::LDURDi:
286 case AArch64::LDRQui:
287 case AArch64::LDURQi:
288 case AArch64::LDRWui:
289 case AArch64::LDURWi:
290 case AArch64::LDRXui:
291 case AArch64::LDURXi:
292 case AArch64::STRSui:
293 case AArch64::STURSi:
294 case AArch64::LDRSui:
295 case AArch64::LDURSi:
297 case AArch64::LDRSWui:
298 return AArch64::LDRWui;
299 case AArch64::LDURSWi:
300 return AArch64::LDURWi;
304 static unsigned getMatchingWideOpcode(unsigned Opc) {
307 llvm_unreachable("Opcode has no wide equivalent!");
308 case AArch64::STRBBui:
309 return AArch64::STRHHui;
310 case AArch64::STRHHui:
311 return AArch64::STRWui;
312 case AArch64::STURBBi:
313 return AArch64::STURHHi;
314 case AArch64::STURHHi:
315 return AArch64::STURWi;
316 case AArch64::STURWi:
317 return AArch64::STURXi;
318 case AArch64::STRWui:
319 return AArch64::STRXui;
323 static unsigned getMatchingPairOpcode(unsigned Opc) {
326 llvm_unreachable("Opcode has no pairwise equivalent!");
327 case AArch64::STRSui:
328 case AArch64::STURSi:
329 return AArch64::STPSi;
330 case AArch64::STRDui:
331 case AArch64::STURDi:
332 return AArch64::STPDi;
333 case AArch64::STRQui:
334 case AArch64::STURQi:
335 return AArch64::STPQi;
336 case AArch64::STRWui:
337 case AArch64::STURWi:
338 return AArch64::STPWi;
339 case AArch64::STRXui:
340 case AArch64::STURXi:
341 return AArch64::STPXi;
342 case AArch64::LDRSui:
343 case AArch64::LDURSi:
344 return AArch64::LDPSi;
345 case AArch64::LDRDui:
346 case AArch64::LDURDi:
347 return AArch64::LDPDi;
348 case AArch64::LDRQui:
349 case AArch64::LDURQi:
350 return AArch64::LDPQi;
351 case AArch64::LDRWui:
352 case AArch64::LDURWi:
353 return AArch64::LDPWi;
354 case AArch64::LDRXui:
355 case AArch64::LDURXi:
356 return AArch64::LDPXi;
357 case AArch64::LDRSWui:
358 case AArch64::LDURSWi:
359 return AArch64::LDPSWi;
363 static unsigned isMatchingStore(MachineInstr &LoadInst,
364 MachineInstr &StoreInst) {
365 unsigned LdOpc = LoadInst.getOpcode();
366 unsigned StOpc = StoreInst.getOpcode();
369 llvm_unreachable("Unsupported load instruction!");
370 case AArch64::LDRBBui:
371 return StOpc == AArch64::STRBBui || StOpc == AArch64::STRHHui ||
372 StOpc == AArch64::STRWui || StOpc == AArch64::STRXui;
373 case AArch64::LDURBBi:
374 return StOpc == AArch64::STURBBi || StOpc == AArch64::STURHHi ||
375 StOpc == AArch64::STURWi || StOpc == AArch64::STURXi;
376 case AArch64::LDRHHui:
377 return StOpc == AArch64::STRHHui || StOpc == AArch64::STRWui ||
378 StOpc == AArch64::STRXui;
379 case AArch64::LDURHHi:
380 return StOpc == AArch64::STURHHi || StOpc == AArch64::STURWi ||
381 StOpc == AArch64::STURXi;
382 case AArch64::LDRWui:
383 return StOpc == AArch64::STRWui || StOpc == AArch64::STRXui;
384 case AArch64::LDURWi:
385 return StOpc == AArch64::STURWi || StOpc == AArch64::STURXi;
386 case AArch64::LDRXui:
387 return StOpc == AArch64::STRXui;
388 case AArch64::LDURXi:
389 return StOpc == AArch64::STURXi;
393 static unsigned getPreIndexedOpcode(unsigned Opc) {
394 // FIXME: We don't currently support creating pre-indexed loads/stores when
395 // the load or store is the unscaled version. If we decide to perform such an
396 // optimization in the future the cases for the unscaled loads/stores will
397 // need to be added here.
400 llvm_unreachable("Opcode has no pre-indexed equivalent!");
401 case AArch64::STRSui:
402 return AArch64::STRSpre;
403 case AArch64::STRDui:
404 return AArch64::STRDpre;
405 case AArch64::STRQui:
406 return AArch64::STRQpre;
407 case AArch64::STRBBui:
408 return AArch64::STRBBpre;
409 case AArch64::STRHHui:
410 return AArch64::STRHHpre;
411 case AArch64::STRWui:
412 return AArch64::STRWpre;
413 case AArch64::STRXui:
414 return AArch64::STRXpre;
415 case AArch64::LDRSui:
416 return AArch64::LDRSpre;
417 case AArch64::LDRDui:
418 return AArch64::LDRDpre;
419 case AArch64::LDRQui:
420 return AArch64::LDRQpre;
421 case AArch64::LDRBBui:
422 return AArch64::LDRBBpre;
423 case AArch64::LDRHHui:
424 return AArch64::LDRHHpre;
425 case AArch64::LDRWui:
426 return AArch64::LDRWpre;
427 case AArch64::LDRXui:
428 return AArch64::LDRXpre;
429 case AArch64::LDRSWui:
430 return AArch64::LDRSWpre;
432 return AArch64::LDPSpre;
433 case AArch64::LDPSWi:
434 return AArch64::LDPSWpre;
436 return AArch64::LDPDpre;
438 return AArch64::LDPQpre;
440 return AArch64::LDPWpre;
442 return AArch64::LDPXpre;
444 return AArch64::STPSpre;
446 return AArch64::STPDpre;
448 return AArch64::STPQpre;
450 return AArch64::STPWpre;
452 return AArch64::STPXpre;
456 static unsigned getPostIndexedOpcode(unsigned Opc) {
459 llvm_unreachable("Opcode has no post-indexed wise equivalent!");
460 case AArch64::STRSui:
461 case AArch64::STURSi:
462 return AArch64::STRSpost;
463 case AArch64::STRDui:
464 case AArch64::STURDi:
465 return AArch64::STRDpost;
466 case AArch64::STRQui:
467 case AArch64::STURQi:
468 return AArch64::STRQpost;
469 case AArch64::STRBBui:
470 return AArch64::STRBBpost;
471 case AArch64::STRHHui:
472 return AArch64::STRHHpost;
473 case AArch64::STRWui:
474 case AArch64::STURWi:
475 return AArch64::STRWpost;
476 case AArch64::STRXui:
477 case AArch64::STURXi:
478 return AArch64::STRXpost;
479 case AArch64::LDRSui:
480 case AArch64::LDURSi:
481 return AArch64::LDRSpost;
482 case AArch64::LDRDui:
483 case AArch64::LDURDi:
484 return AArch64::LDRDpost;
485 case AArch64::LDRQui:
486 case AArch64::LDURQi:
487 return AArch64::LDRQpost;
488 case AArch64::LDRBBui:
489 return AArch64::LDRBBpost;
490 case AArch64::LDRHHui:
491 return AArch64::LDRHHpost;
492 case AArch64::LDRWui:
493 case AArch64::LDURWi:
494 return AArch64::LDRWpost;
495 case AArch64::LDRXui:
496 case AArch64::LDURXi:
497 return AArch64::LDRXpost;
498 case AArch64::LDRSWui:
499 return AArch64::LDRSWpost;
501 return AArch64::LDPSpost;
502 case AArch64::LDPSWi:
503 return AArch64::LDPSWpost;
505 return AArch64::LDPDpost;
507 return AArch64::LDPQpost;
509 return AArch64::LDPWpost;
511 return AArch64::LDPXpost;
513 return AArch64::STPSpost;
515 return AArch64::STPDpost;
517 return AArch64::STPQpost;
519 return AArch64::STPWpost;
521 return AArch64::STPXpost;
525 static bool isPairedLdSt(const MachineInstr &MI) {
526 switch (MI.getOpcode()) {
530 case AArch64::LDPSWi:
544 static const MachineOperand &getLdStRegOp(const MachineInstr &MI,
545 unsigned PairedRegOp = 0) {
546 assert(PairedRegOp < 2 && "Unexpected register operand idx.");
547 unsigned Idx = isPairedLdSt(MI) ? PairedRegOp : 0;
548 return MI.getOperand(Idx);
551 static const MachineOperand &getLdStBaseOp(const MachineInstr &MI) {
552 unsigned Idx = isPairedLdSt(MI) ? 2 : 1;
553 return MI.getOperand(Idx);
556 static const MachineOperand &getLdStOffsetOp(const MachineInstr &MI) {
557 unsigned Idx = isPairedLdSt(MI) ? 3 : 2;
558 return MI.getOperand(Idx);
561 static bool isLdOffsetInRangeOfSt(MachineInstr &LoadInst,
562 MachineInstr &StoreInst,
563 const AArch64InstrInfo *TII) {
564 assert(isMatchingStore(LoadInst, StoreInst) && "Expect only matched ld/st.");
565 int LoadSize = getMemScale(LoadInst);
566 int StoreSize = getMemScale(StoreInst);
567 int UnscaledStOffset = TII->isUnscaledLdSt(StoreInst)
568 ? getLdStOffsetOp(StoreInst).getImm()
569 : getLdStOffsetOp(StoreInst).getImm() * StoreSize;
570 int UnscaledLdOffset = TII->isUnscaledLdSt(LoadInst)
571 ? getLdStOffsetOp(LoadInst).getImm()
572 : getLdStOffsetOp(LoadInst).getImm() * LoadSize;
573 return (UnscaledStOffset <= UnscaledLdOffset) &&
574 (UnscaledLdOffset + LoadSize <= (UnscaledStOffset + StoreSize));
577 static bool isPromotableZeroStoreInst(MachineInstr &MI) {
578 unsigned Opc = MI.getOpcode();
579 return (Opc == AArch64::STRWui || Opc == AArch64::STURWi ||
580 isNarrowStore(Opc)) &&
581 getLdStRegOp(MI).getReg() == AArch64::WZR;
584 static bool isPromotableLoadFromStore(MachineInstr &MI) {
585 switch (MI.getOpcode()) {
588 // Scaled instructions.
589 case AArch64::LDRBBui:
590 case AArch64::LDRHHui:
591 case AArch64::LDRWui:
592 case AArch64::LDRXui:
593 // Unscaled instructions.
594 case AArch64::LDURBBi:
595 case AArch64::LDURHHi:
596 case AArch64::LDURWi:
597 case AArch64::LDURXi:
602 static bool isMergeableLdStUpdate(MachineInstr &MI) {
603 unsigned Opc = MI.getOpcode();
607 // Scaled instructions.
608 case AArch64::STRSui:
609 case AArch64::STRDui:
610 case AArch64::STRQui:
611 case AArch64::STRXui:
612 case AArch64::STRWui:
613 case AArch64::STRHHui:
614 case AArch64::STRBBui:
615 case AArch64::LDRSui:
616 case AArch64::LDRDui:
617 case AArch64::LDRQui:
618 case AArch64::LDRXui:
619 case AArch64::LDRWui:
620 case AArch64::LDRHHui:
621 case AArch64::LDRBBui:
622 // Unscaled instructions.
623 case AArch64::STURSi:
624 case AArch64::STURDi:
625 case AArch64::STURQi:
626 case AArch64::STURWi:
627 case AArch64::STURXi:
628 case AArch64::LDURSi:
629 case AArch64::LDURDi:
630 case AArch64::LDURQi:
631 case AArch64::LDURWi:
632 case AArch64::LDURXi:
633 // Paired instructions.
635 case AArch64::LDPSWi:
645 // Make sure this is a reg+imm (as opposed to an address reloc).
646 if (!getLdStOffsetOp(MI).isImm())
653 MachineBasicBlock::iterator
654 AArch64LoadStoreOpt::mergeNarrowZeroStores(MachineBasicBlock::iterator I,
655 MachineBasicBlock::iterator MergeMI,
656 const LdStPairFlags &Flags) {
657 assert(isPromotableZeroStoreInst(*I) && isPromotableZeroStoreInst(*MergeMI) &&
658 "Expected promotable zero stores.");
660 MachineBasicBlock::iterator NextI = I;
662 // If NextI is the second of the two instructions to be merged, we need
663 // to skip one further. Either way we merge will invalidate the iterator,
664 // and we don't need to scan the new instruction, as it's a pairwise
665 // instruction, which we're not considering for further action anyway.
666 if (NextI == MergeMI)
669 unsigned Opc = I->getOpcode();
670 bool IsScaled = !TII->isUnscaledLdSt(Opc);
671 int OffsetStride = IsScaled ? 1 : getMemScale(*I);
673 bool MergeForward = Flags.getMergeForward();
674 // Insert our new paired instruction after whichever of the paired
675 // instructions MergeForward indicates.
676 MachineBasicBlock::iterator InsertionPoint = MergeForward ? MergeMI : I;
677 // Also based on MergeForward is from where we copy the base register operand
678 // so we get the flags compatible with the input code.
679 const MachineOperand &BaseRegOp =
680 MergeForward ? getLdStBaseOp(*MergeMI) : getLdStBaseOp(*I);
682 // Which register is Rt and which is Rt2 depends on the offset order.
684 if (getLdStOffsetOp(*I).getImm() ==
685 getLdStOffsetOp(*MergeMI).getImm() + OffsetStride)
690 int OffsetImm = getLdStOffsetOp(*RtMI).getImm();
691 // Change the scaled offset from small to large type.
693 assert(((OffsetImm & 1) == 0) && "Unexpected offset to merge");
697 // Construct the new instruction.
698 DebugLoc DL = I->getDebugLoc();
699 MachineBasicBlock *MBB = I->getParent();
700 MachineInstrBuilder MIB;
701 MIB = BuildMI(*MBB, InsertionPoint, DL, TII->get(getMatchingWideOpcode(Opc)))
702 .addReg(isNarrowStore(Opc) ? AArch64::WZR : AArch64::XZR)
705 .setMemRefs(I->mergeMemRefsWith(*MergeMI));
708 DEBUG(dbgs() << "Creating wider store. Replacing instructions:\n ");
709 DEBUG(I->print(dbgs()));
710 DEBUG(dbgs() << " ");
711 DEBUG(MergeMI->print(dbgs()));
712 DEBUG(dbgs() << " with instruction:\n ");
713 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
714 DEBUG(dbgs() << "\n");
716 // Erase the old instructions.
717 I->eraseFromParent();
718 MergeMI->eraseFromParent();
722 MachineBasicBlock::iterator
723 AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I,
724 MachineBasicBlock::iterator Paired,
725 const LdStPairFlags &Flags) {
726 MachineBasicBlock::iterator NextI = I;
728 // If NextI is the second of the two instructions to be merged, we need
729 // to skip one further. Either way we merge will invalidate the iterator,
730 // and we don't need to scan the new instruction, as it's a pairwise
731 // instruction, which we're not considering for further action anyway.
735 int SExtIdx = Flags.getSExtIdx();
737 SExtIdx == -1 ? I->getOpcode() : getMatchingNonSExtOpcode(I->getOpcode());
738 bool IsUnscaled = TII->isUnscaledLdSt(Opc);
739 int OffsetStride = IsUnscaled ? getMemScale(*I) : 1;
741 bool MergeForward = Flags.getMergeForward();
742 // Insert our new paired instruction after whichever of the paired
743 // instructions MergeForward indicates.
744 MachineBasicBlock::iterator InsertionPoint = MergeForward ? Paired : I;
745 // Also based on MergeForward is from where we copy the base register operand
746 // so we get the flags compatible with the input code.
747 const MachineOperand &BaseRegOp =
748 MergeForward ? getLdStBaseOp(*Paired) : getLdStBaseOp(*I);
750 int Offset = getLdStOffsetOp(*I).getImm();
751 int PairedOffset = getLdStOffsetOp(*Paired).getImm();
752 bool PairedIsUnscaled = TII->isUnscaledLdSt(Paired->getOpcode());
753 if (IsUnscaled != PairedIsUnscaled) {
754 // We're trying to pair instructions that differ in how they are scaled. If
755 // I is scaled then scale the offset of Paired accordingly. Otherwise, do
756 // the opposite (i.e., make Paired's offset unscaled).
757 int MemSize = getMemScale(*Paired);
758 if (PairedIsUnscaled) {
759 // If the unscaled offset isn't a multiple of the MemSize, we can't
760 // pair the operations together.
761 assert(!(PairedOffset % getMemScale(*Paired)) &&
762 "Offset should be a multiple of the stride!");
763 PairedOffset /= MemSize;
765 PairedOffset *= MemSize;
769 // Which register is Rt and which is Rt2 depends on the offset order.
770 MachineInstr *RtMI, *Rt2MI;
771 if (Offset == PairedOffset + OffsetStride) {
774 // Here we swapped the assumption made for SExtIdx.
775 // I.e., we turn ldp I, Paired into ldp Paired, I.
776 // Update the index accordingly.
778 SExtIdx = (SExtIdx + 1) % 2;
783 int OffsetImm = getLdStOffsetOp(*RtMI).getImm();
784 // Scale the immediate offset, if necessary.
785 if (TII->isUnscaledLdSt(RtMI->getOpcode())) {
786 assert(!(OffsetImm % getMemScale(*RtMI)) &&
787 "Unscaled offset cannot be scaled.");
788 OffsetImm /= getMemScale(*RtMI);
791 // Construct the new instruction.
792 MachineInstrBuilder MIB;
793 DebugLoc DL = I->getDebugLoc();
794 MachineBasicBlock *MBB = I->getParent();
795 MachineOperand RegOp0 = getLdStRegOp(*RtMI);
796 MachineOperand RegOp1 = getLdStRegOp(*Rt2MI);
797 // Kill flags may become invalid when moving stores for pairing.
798 if (RegOp0.isUse()) {
800 // Clear kill flags on store if moving upwards. Example:
803 // STRWui kill %w1 ; need to clear kill flag when moving STRWui upwards
804 RegOp0.setIsKill(false);
805 RegOp1.setIsKill(false);
807 // Clear kill flags of the first stores register. Example:
809 // USE kill %w1 ; need to clear kill flag when moving STRWui downwards
811 unsigned Reg = getLdStRegOp(*I).getReg();
812 for (MachineInstr &MI : make_range(std::next(I), Paired))
813 MI.clearRegisterKills(Reg, TRI);
816 MIB = BuildMI(*MBB, InsertionPoint, DL, TII->get(getMatchingPairOpcode(Opc)))
821 .setMemRefs(I->mergeMemRefsWith(*Paired));
825 DEBUG(dbgs() << "Creating pair load/store. Replacing instructions:\n ");
826 DEBUG(I->print(dbgs()));
827 DEBUG(dbgs() << " ");
828 DEBUG(Paired->print(dbgs()));
829 DEBUG(dbgs() << " with instruction:\n ");
831 // Generate the sign extension for the proper result of the ldp.
832 // I.e., with X1, that would be:
833 // %w1 = KILL %w1, implicit-def %x1
834 // %x1 = SBFMXri killed %x1, 0, 31
835 MachineOperand &DstMO = MIB->getOperand(SExtIdx);
836 // Right now, DstMO has the extended register, since it comes from an
838 unsigned DstRegX = DstMO.getReg();
839 // Get the W variant of that register.
840 unsigned DstRegW = TRI->getSubReg(DstRegX, AArch64::sub_32);
841 // Update the result of LDP to use the W instead of the X variant.
842 DstMO.setReg(DstRegW);
843 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
844 DEBUG(dbgs() << "\n");
845 // Make the machine verifier happy by providing a definition for
847 // Insert this definition right after the generated LDP, i.e., before
849 MachineInstrBuilder MIBKill =
850 BuildMI(*MBB, InsertionPoint, DL, TII->get(TargetOpcode::KILL), DstRegW)
852 .addReg(DstRegX, RegState::Define);
853 MIBKill->getOperand(2).setImplicit();
854 // Create the sign extension.
855 MachineInstrBuilder MIBSXTW =
856 BuildMI(*MBB, InsertionPoint, DL, TII->get(AArch64::SBFMXri), DstRegX)
861 DEBUG(dbgs() << " Extend operand:\n ");
862 DEBUG(((MachineInstr *)MIBSXTW)->print(dbgs()));
864 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
866 DEBUG(dbgs() << "\n");
868 // Erase the old instructions.
869 I->eraseFromParent();
870 Paired->eraseFromParent();
875 MachineBasicBlock::iterator
876 AArch64LoadStoreOpt::promoteLoadFromStore(MachineBasicBlock::iterator LoadI,
877 MachineBasicBlock::iterator StoreI) {
878 MachineBasicBlock::iterator NextI = LoadI;
881 int LoadSize = getMemScale(*LoadI);
882 int StoreSize = getMemScale(*StoreI);
883 unsigned LdRt = getLdStRegOp(*LoadI).getReg();
884 const MachineOperand &StMO = getLdStRegOp(*StoreI);
885 unsigned StRt = getLdStRegOp(*StoreI).getReg();
886 bool IsStoreXReg = TRI->getRegClass(AArch64::GPR64RegClassID)->contains(StRt);
888 assert((IsStoreXReg ||
889 TRI->getRegClass(AArch64::GPR32RegClassID)->contains(StRt)) &&
890 "Unexpected RegClass");
892 MachineInstr *BitExtMI;
893 if (LoadSize == StoreSize && (LoadSize == 4 || LoadSize == 8)) {
894 // Remove the load, if the destination register of the loads is the same
895 // register for stored value.
896 if (StRt == LdRt && LoadSize == 8) {
897 for (MachineInstr &MI : make_range(StoreI->getIterator(),
898 LoadI->getIterator())) {
899 if (MI.killsRegister(StRt, TRI)) {
900 MI.clearRegisterKills(StRt, TRI);
904 DEBUG(dbgs() << "Remove load instruction:\n ");
905 DEBUG(LoadI->print(dbgs()));
906 DEBUG(dbgs() << "\n");
907 LoadI->eraseFromParent();
910 // Replace the load with a mov if the load and store are in the same size.
912 BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(),
913 TII->get(IsStoreXReg ? AArch64::ORRXrs : AArch64::ORRWrs), LdRt)
914 .addReg(IsStoreXReg ? AArch64::XZR : AArch64::WZR)
916 .addImm(AArch64_AM::getShifterImm(AArch64_AM::LSL, 0));
918 // FIXME: Currently we disable this transformation in big-endian targets as
919 // performance and correctness are verified only in little-endian.
920 if (!Subtarget->isLittleEndian())
922 bool IsUnscaled = TII->isUnscaledLdSt(*LoadI);
923 assert(IsUnscaled == TII->isUnscaledLdSt(*StoreI) &&
924 "Unsupported ld/st match");
925 assert(LoadSize <= StoreSize && "Invalid load size");
926 int UnscaledLdOffset = IsUnscaled
927 ? getLdStOffsetOp(*LoadI).getImm()
928 : getLdStOffsetOp(*LoadI).getImm() * LoadSize;
929 int UnscaledStOffset = IsUnscaled
930 ? getLdStOffsetOp(*StoreI).getImm()
931 : getLdStOffsetOp(*StoreI).getImm() * StoreSize;
932 int Width = LoadSize * 8;
933 int Immr = 8 * (UnscaledLdOffset - UnscaledStOffset);
934 int Imms = Immr + Width - 1;
935 unsigned DestReg = IsStoreXReg
936 ? TRI->getMatchingSuperReg(LdRt, AArch64::sub_32,
937 &AArch64::GPR64RegClass)
940 assert((UnscaledLdOffset >= UnscaledStOffset &&
941 (UnscaledLdOffset + LoadSize) <= UnscaledStOffset + StoreSize) &&
944 Immr = 8 * (UnscaledLdOffset - UnscaledStOffset);
945 Imms = Immr + Width - 1;
946 if (UnscaledLdOffset == UnscaledStOffset) {
947 uint32_t AndMaskEncoded = ((IsStoreXReg ? 1 : 0) << 12) // N
948 | ((Immr) << 6) // immr
949 | ((Imms) << 0) // imms
953 BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(),
954 TII->get(IsStoreXReg ? AArch64::ANDXri : AArch64::ANDWri),
957 .addImm(AndMaskEncoded);
960 BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(),
961 TII->get(IsStoreXReg ? AArch64::UBFMXri : AArch64::UBFMWri),
969 // Clear kill flags between store and load.
970 for (MachineInstr &MI : make_range(StoreI->getIterator(),
971 BitExtMI->getIterator()))
972 if (MI.killsRegister(StRt, TRI)) {
973 MI.clearRegisterKills(StRt, TRI);
977 DEBUG(dbgs() << "Promoting load by replacing :\n ");
978 DEBUG(StoreI->print(dbgs()));
979 DEBUG(dbgs() << " ");
980 DEBUG(LoadI->print(dbgs()));
981 DEBUG(dbgs() << " with instructions:\n ");
982 DEBUG(StoreI->print(dbgs()));
983 DEBUG(dbgs() << " ");
984 DEBUG((BitExtMI)->print(dbgs()));
985 DEBUG(dbgs() << "\n");
987 // Erase the old instructions.
988 LoadI->eraseFromParent();
992 /// trackRegDefsUses - Remember what registers the specified instruction uses
994 static void trackRegDefsUses(const MachineInstr &MI, BitVector &ModifiedRegs,
996 const TargetRegisterInfo *TRI) {
997 for (const MachineOperand &MO : MI.operands()) {
999 ModifiedRegs.setBitsNotInMask(MO.getRegMask());
1003 unsigned Reg = MO.getReg();
1007 // WZR/XZR are not modified even when used as a destination register.
1008 if (Reg != AArch64::WZR && Reg != AArch64::XZR)
1009 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1010 ModifiedRegs.set(*AI);
1012 assert(MO.isUse() && "Reg operand not a def and not a use?!?");
1013 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
1019 static bool inBoundsForPair(bool IsUnscaled, int Offset, int OffsetStride) {
1020 // Convert the byte-offset used by unscaled into an "element" offset used
1021 // by the scaled pair load/store instructions.
1023 // If the byte-offset isn't a multiple of the stride, there's no point
1024 // trying to match it.
1025 if (Offset % OffsetStride)
1027 Offset /= OffsetStride;
1029 return Offset <= 63 && Offset >= -64;
1032 // Do alignment, specialized to power of 2 and for signed ints,
1033 // avoiding having to do a C-style cast from uint_64t to int when
1034 // using alignTo from include/llvm/Support/MathExtras.h.
1035 // FIXME: Move this function to include/MathExtras.h?
1036 static int alignTo(int Num, int PowOf2) {
1037 return (Num + PowOf2 - 1) & ~(PowOf2 - 1);
1040 static bool mayAlias(MachineInstr &MIa, MachineInstr &MIb,
1041 AliasAnalysis *AA) {
1042 // One of the instructions must modify memory.
1043 if (!MIa.mayStore() && !MIb.mayStore())
1046 // Both instructions must be memory operations.
1047 if (!MIa.mayLoadOrStore() && !MIb.mayLoadOrStore())
1050 return MIa.mayAlias(AA, MIb, /*UseTBAA*/false);
1053 static bool mayAlias(MachineInstr &MIa,
1054 SmallVectorImpl<MachineInstr *> &MemInsns,
1055 AliasAnalysis *AA) {
1056 for (MachineInstr *MIb : MemInsns)
1057 if (mayAlias(MIa, *MIb, AA))
1063 bool AArch64LoadStoreOpt::findMatchingStore(
1064 MachineBasicBlock::iterator I, unsigned Limit,
1065 MachineBasicBlock::iterator &StoreI) {
1066 MachineBasicBlock::iterator B = I->getParent()->begin();
1067 MachineBasicBlock::iterator MBBI = I;
1068 MachineInstr &LoadMI = *I;
1069 unsigned BaseReg = getLdStBaseOp(LoadMI).getReg();
1071 // If the load is the first instruction in the block, there's obviously
1072 // not any matching store.
1076 // Track which registers have been modified and used between the first insn
1077 // and the second insn.
1078 ModifiedRegs.reset();
1084 MachineInstr &MI = *MBBI;
1086 // Don't count transient instructions towards the search limit since there
1087 // may be different numbers of them if e.g. debug information is present.
1088 if (!MI.isTransient())
1091 // If the load instruction reads directly from the address to which the
1092 // store instruction writes and the stored value is not modified, we can
1093 // promote the load. Since we do not handle stores with pre-/post-index,
1094 // it's unnecessary to check if BaseReg is modified by the store itself.
1095 if (MI.mayStore() && isMatchingStore(LoadMI, MI) &&
1096 BaseReg == getLdStBaseOp(MI).getReg() &&
1097 isLdOffsetInRangeOfSt(LoadMI, MI, TII) &&
1098 !ModifiedRegs[getLdStRegOp(MI).getReg()]) {
1106 // Update modified / uses register lists.
1107 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1109 // Otherwise, if the base register is modified, we have no match, so
1111 if (ModifiedRegs[BaseReg])
1114 // If we encounter a store aliased with the load, return early.
1115 if (MI.mayStore() && mayAlias(LoadMI, MI, AA))
1117 } while (MBBI != B && Count < Limit);
1121 // Returns true if FirstMI and MI are candidates for merging or pairing.
1122 // Otherwise, returns false.
1123 static bool areCandidatesToMergeOrPair(MachineInstr &FirstMI, MachineInstr &MI,
1124 LdStPairFlags &Flags,
1125 const AArch64InstrInfo *TII) {
1126 // If this is volatile or if pairing is suppressed, not a candidate.
1127 if (MI.hasOrderedMemoryRef() || TII->isLdStPairSuppressed(MI))
1130 // We should have already checked FirstMI for pair suppression and volatility.
1131 assert(!FirstMI.hasOrderedMemoryRef() &&
1132 !TII->isLdStPairSuppressed(FirstMI) &&
1133 "FirstMI shouldn't get here if either of these checks are true.");
1135 unsigned OpcA = FirstMI.getOpcode();
1136 unsigned OpcB = MI.getOpcode();
1138 // Opcodes match: nothing more to check.
1142 // Try to match a sign-extended load/store with a zero-extended load/store.
1143 bool IsValidLdStrOpc, PairIsValidLdStrOpc;
1144 unsigned NonSExtOpc = getMatchingNonSExtOpcode(OpcA, &IsValidLdStrOpc);
1145 assert(IsValidLdStrOpc &&
1146 "Given Opc should be a Load or Store with an immediate");
1147 // OpcA will be the first instruction in the pair.
1148 if (NonSExtOpc == getMatchingNonSExtOpcode(OpcB, &PairIsValidLdStrOpc)) {
1149 Flags.setSExtIdx(NonSExtOpc == (unsigned)OpcA ? 1 : 0);
1153 // If the second instruction isn't even a mergable/pairable load/store, bail
1155 if (!PairIsValidLdStrOpc)
1158 // FIXME: We don't support merging narrow stores with mixed scaled/unscaled
1160 if (isNarrowStore(OpcA) || isNarrowStore(OpcB))
1163 // Try to match an unscaled load/store with a scaled load/store.
1164 return TII->isUnscaledLdSt(OpcA) != TII->isUnscaledLdSt(OpcB) &&
1165 getMatchingPairOpcode(OpcA) == getMatchingPairOpcode(OpcB);
1167 // FIXME: Can we also match a mixed sext/zext unscaled/scaled pair?
1170 /// Scan the instructions looking for a load/store that can be combined with the
1171 /// current instruction into a wider equivalent or a load/store pair.
1172 MachineBasicBlock::iterator
1173 AArch64LoadStoreOpt::findMatchingInsn(MachineBasicBlock::iterator I,
1174 LdStPairFlags &Flags, unsigned Limit,
1175 bool FindNarrowMerge) {
1176 MachineBasicBlock::iterator E = I->getParent()->end();
1177 MachineBasicBlock::iterator MBBI = I;
1178 MachineInstr &FirstMI = *I;
1181 bool MayLoad = FirstMI.mayLoad();
1182 bool IsUnscaled = TII->isUnscaledLdSt(FirstMI);
1183 unsigned Reg = getLdStRegOp(FirstMI).getReg();
1184 unsigned BaseReg = getLdStBaseOp(FirstMI).getReg();
1185 int Offset = getLdStOffsetOp(FirstMI).getImm();
1186 int OffsetStride = IsUnscaled ? getMemScale(FirstMI) : 1;
1187 bool IsPromotableZeroStore = isPromotableZeroStoreInst(FirstMI);
1189 // Track which registers have been modified and used between the first insn
1190 // (inclusive) and the second insn.
1191 ModifiedRegs.reset();
1194 // Remember any instructions that read/write memory between FirstMI and MI.
1195 SmallVector<MachineInstr *, 4> MemInsns;
1197 for (unsigned Count = 0; MBBI != E && Count < Limit; ++MBBI) {
1198 MachineInstr &MI = *MBBI;
1200 // Don't count transient instructions towards the search limit since there
1201 // may be different numbers of them if e.g. debug information is present.
1202 if (!MI.isTransient())
1205 Flags.setSExtIdx(-1);
1206 if (areCandidatesToMergeOrPair(FirstMI, MI, Flags, TII) &&
1207 getLdStOffsetOp(MI).isImm()) {
1208 assert(MI.mayLoadOrStore() && "Expected memory operation.");
1209 // If we've found another instruction with the same opcode, check to see
1210 // if the base and offset are compatible with our starting instruction.
1211 // These instructions all have scaled immediate operands, so we just
1212 // check for +1/-1. Make sure to check the new instruction offset is
1213 // actually an immediate and not a symbolic reference destined for
1215 unsigned MIBaseReg = getLdStBaseOp(MI).getReg();
1216 int MIOffset = getLdStOffsetOp(MI).getImm();
1217 bool MIIsUnscaled = TII->isUnscaledLdSt(MI);
1218 if (IsUnscaled != MIIsUnscaled) {
1219 // We're trying to pair instructions that differ in how they are scaled.
1220 // If FirstMI is scaled then scale the offset of MI accordingly.
1221 // Otherwise, do the opposite (i.e., make MI's offset unscaled).
1222 int MemSize = getMemScale(MI);
1224 // If the unscaled offset isn't a multiple of the MemSize, we can't
1225 // pair the operations together: bail and keep looking.
1226 if (MIOffset % MemSize) {
1227 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1228 MemInsns.push_back(&MI);
1231 MIOffset /= MemSize;
1233 MIOffset *= MemSize;
1237 if (BaseReg == MIBaseReg && ((Offset == MIOffset + OffsetStride) ||
1238 (Offset + OffsetStride == MIOffset))) {
1239 int MinOffset = Offset < MIOffset ? Offset : MIOffset;
1240 if (FindNarrowMerge) {
1241 // If the alignment requirements of the scaled wide load/store
1242 // instruction can't express the offset of the scaled narrow input,
1243 // bail and keep looking. For promotable zero stores, allow only when
1244 // the stored value is the same (i.e., WZR).
1245 if ((!IsUnscaled && alignTo(MinOffset, 2) != MinOffset) ||
1246 (IsPromotableZeroStore && Reg != getLdStRegOp(MI).getReg())) {
1247 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1248 MemInsns.push_back(&MI);
1252 // Pairwise instructions have a 7-bit signed offset field. Single
1253 // insns have a 12-bit unsigned offset field. If the resultant
1254 // immediate offset of merging these instructions is out of range for
1255 // a pairwise instruction, bail and keep looking.
1256 if (!inBoundsForPair(IsUnscaled, MinOffset, OffsetStride)) {
1257 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1258 MemInsns.push_back(&MI);
1261 // If the alignment requirements of the paired (scaled) instruction
1262 // can't express the offset of the unscaled input, bail and keep
1264 if (IsUnscaled && (alignTo(MinOffset, OffsetStride) != MinOffset)) {
1265 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1266 MemInsns.push_back(&MI);
1270 // If the destination register of the loads is the same register, bail
1271 // and keep looking. A load-pair instruction with both destination
1272 // registers the same is UNPREDICTABLE and will result in an exception.
1273 if (MayLoad && Reg == getLdStRegOp(MI).getReg()) {
1274 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1275 MemInsns.push_back(&MI);
1279 // If the Rt of the second instruction was not modified or used between
1280 // the two instructions and none of the instructions between the second
1281 // and first alias with the second, we can combine the second into the
1283 if (!ModifiedRegs[getLdStRegOp(MI).getReg()] &&
1284 !(MI.mayLoad() && UsedRegs[getLdStRegOp(MI).getReg()]) &&
1285 !mayAlias(MI, MemInsns, AA)) {
1286 Flags.setMergeForward(false);
1290 // Likewise, if the Rt of the first instruction is not modified or used
1291 // between the two instructions and none of the instructions between the
1292 // first and the second alias with the first, we can combine the first
1294 if (!ModifiedRegs[getLdStRegOp(FirstMI).getReg()] &&
1295 !(MayLoad && UsedRegs[getLdStRegOp(FirstMI).getReg()]) &&
1296 !mayAlias(FirstMI, MemInsns, AA)) {
1297 Flags.setMergeForward(true);
1300 // Unable to combine these instructions due to interference in between.
1305 // If the instruction wasn't a matching load or store. Stop searching if we
1306 // encounter a call instruction that might modify memory.
1310 // Update modified / uses register lists.
1311 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1313 // Otherwise, if the base register is modified, we have no match, so
1315 if (ModifiedRegs[BaseReg])
1318 // Update list of instructions that read/write memory.
1319 if (MI.mayLoadOrStore())
1320 MemInsns.push_back(&MI);
1325 MachineBasicBlock::iterator
1326 AArch64LoadStoreOpt::mergeUpdateInsn(MachineBasicBlock::iterator I,
1327 MachineBasicBlock::iterator Update,
1329 assert((Update->getOpcode() == AArch64::ADDXri ||
1330 Update->getOpcode() == AArch64::SUBXri) &&
1331 "Unexpected base register update instruction to merge!");
1332 MachineBasicBlock::iterator NextI = I;
1333 // Return the instruction following the merged instruction, which is
1334 // the instruction following our unmerged load. Unless that's the add/sub
1335 // instruction we're merging, in which case it's the one after that.
1336 if (++NextI == Update)
1339 int Value = Update->getOperand(2).getImm();
1340 assert(AArch64_AM::getShiftValue(Update->getOperand(3).getImm()) == 0 &&
1341 "Can't merge 1 << 12 offset into pre-/post-indexed load / store");
1342 if (Update->getOpcode() == AArch64::SUBXri)
1345 unsigned NewOpc = IsPreIdx ? getPreIndexedOpcode(I->getOpcode())
1346 : getPostIndexedOpcode(I->getOpcode());
1347 MachineInstrBuilder MIB;
1348 if (!isPairedLdSt(*I)) {
1349 // Non-paired instruction.
1350 MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc))
1351 .add(getLdStRegOp(*Update))
1352 .add(getLdStRegOp(*I))
1353 .add(getLdStBaseOp(*I))
1355 .setMemRefs(I->memoperands_begin(), I->memoperands_end());
1357 // Paired instruction.
1358 int Scale = getMemScale(*I);
1359 MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc))
1360 .add(getLdStRegOp(*Update))
1361 .add(getLdStRegOp(*I, 0))
1362 .add(getLdStRegOp(*I, 1))
1363 .add(getLdStBaseOp(*I))
1364 .addImm(Value / Scale)
1365 .setMemRefs(I->memoperands_begin(), I->memoperands_end());
1371 DEBUG(dbgs() << "Creating pre-indexed load/store.");
1374 DEBUG(dbgs() << "Creating post-indexed load/store.");
1376 DEBUG(dbgs() << " Replacing instructions:\n ");
1377 DEBUG(I->print(dbgs()));
1378 DEBUG(dbgs() << " ");
1379 DEBUG(Update->print(dbgs()));
1380 DEBUG(dbgs() << " with instruction:\n ");
1381 DEBUG(((MachineInstr *)MIB)->print(dbgs()));
1382 DEBUG(dbgs() << "\n");
1384 // Erase the old instructions for the block.
1385 I->eraseFromParent();
1386 Update->eraseFromParent();
1391 bool AArch64LoadStoreOpt::isMatchingUpdateInsn(MachineInstr &MemMI,
1393 unsigned BaseReg, int Offset) {
1394 switch (MI.getOpcode()) {
1397 case AArch64::SUBXri:
1398 case AArch64::ADDXri:
1399 // Make sure it's a vanilla immediate operand, not a relocation or
1400 // anything else we can't handle.
1401 if (!MI.getOperand(2).isImm())
1403 // Watch out for 1 << 12 shifted value.
1404 if (AArch64_AM::getShiftValue(MI.getOperand(3).getImm()))
1407 // The update instruction source and destination register must be the
1408 // same as the load/store base register.
1409 if (MI.getOperand(0).getReg() != BaseReg ||
1410 MI.getOperand(1).getReg() != BaseReg)
1413 bool IsPairedInsn = isPairedLdSt(MemMI);
1414 int UpdateOffset = MI.getOperand(2).getImm();
1415 if (MI.getOpcode() == AArch64::SUBXri)
1416 UpdateOffset = -UpdateOffset;
1418 // For non-paired load/store instructions, the immediate must fit in a
1419 // signed 9-bit integer.
1420 if (!IsPairedInsn && (UpdateOffset > 255 || UpdateOffset < -256))
1423 // For paired load/store instructions, the immediate must be a multiple of
1424 // the scaling factor. The scaled offset must also fit into a signed 7-bit
1427 int Scale = getMemScale(MemMI);
1428 if (UpdateOffset % Scale != 0)
1431 int ScaledOffset = UpdateOffset / Scale;
1432 if (ScaledOffset > 63 || ScaledOffset < -64)
1436 // If we have a non-zero Offset, we check that it matches the amount
1437 // we're adding to the register.
1438 if (!Offset || Offset == UpdateOffset)
1445 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnForward(
1446 MachineBasicBlock::iterator I, int UnscaledOffset, unsigned Limit) {
1447 MachineBasicBlock::iterator E = I->getParent()->end();
1448 MachineInstr &MemMI = *I;
1449 MachineBasicBlock::iterator MBBI = I;
1451 unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
1452 int MIUnscaledOffset = getLdStOffsetOp(MemMI).getImm() * getMemScale(MemMI);
1454 // Scan forward looking for post-index opportunities. Updating instructions
1455 // can't be formed if the memory instruction doesn't have the offset we're
1457 if (MIUnscaledOffset != UnscaledOffset)
1460 // If the base register overlaps a destination register, we can't
1461 // merge the update.
1462 bool IsPairedInsn = isPairedLdSt(MemMI);
1463 for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) {
1464 unsigned DestReg = getLdStRegOp(MemMI, i).getReg();
1465 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg))
1469 // Track which registers have been modified and used between the first insn
1470 // (inclusive) and the second insn.
1471 ModifiedRegs.reset();
1474 for (unsigned Count = 0; MBBI != E && Count < Limit; ++MBBI) {
1475 MachineInstr &MI = *MBBI;
1477 // Don't count transient instructions towards the search limit since there
1478 // may be different numbers of them if e.g. debug information is present.
1479 if (!MI.isTransient())
1482 // If we found a match, return it.
1483 if (isMatchingUpdateInsn(*I, MI, BaseReg, UnscaledOffset))
1486 // Update the status of what the instruction clobbered and used.
1487 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1489 // Otherwise, if the base register is used or modified, we have no match, so
1491 if (ModifiedRegs[BaseReg] || UsedRegs[BaseReg])
1497 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnBackward(
1498 MachineBasicBlock::iterator I, unsigned Limit) {
1499 MachineBasicBlock::iterator B = I->getParent()->begin();
1500 MachineBasicBlock::iterator E = I->getParent()->end();
1501 MachineInstr &MemMI = *I;
1502 MachineBasicBlock::iterator MBBI = I;
1504 unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
1505 int Offset = getLdStOffsetOp(MemMI).getImm();
1507 // If the load/store is the first instruction in the block, there's obviously
1508 // not any matching update. Ditto if the memory offset isn't zero.
1509 if (MBBI == B || Offset != 0)
1511 // If the base register overlaps a destination register, we can't
1512 // merge the update.
1513 bool IsPairedInsn = isPairedLdSt(MemMI);
1514 for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) {
1515 unsigned DestReg = getLdStRegOp(MemMI, i).getReg();
1516 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg))
1520 // Track which registers have been modified and used between the first insn
1521 // (inclusive) and the second insn.
1522 ModifiedRegs.reset();
1527 MachineInstr &MI = *MBBI;
1529 // Don't count transient instructions towards the search limit since there
1530 // may be different numbers of them if e.g. debug information is present.
1531 if (!MI.isTransient())
1534 // If we found a match, return it.
1535 if (isMatchingUpdateInsn(*I, MI, BaseReg, Offset))
1538 // Update the status of what the instruction clobbered and used.
1539 trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1541 // Otherwise, if the base register is used or modified, we have no match, so
1543 if (ModifiedRegs[BaseReg] || UsedRegs[BaseReg])
1545 } while (MBBI != B && Count < Limit);
1549 bool AArch64LoadStoreOpt::tryToPromoteLoadFromStore(
1550 MachineBasicBlock::iterator &MBBI) {
1551 MachineInstr &MI = *MBBI;
1552 // If this is a volatile load, don't mess with it.
1553 if (MI.hasOrderedMemoryRef())
1556 // Make sure this is a reg+imm.
1557 // FIXME: It is possible to extend it to handle reg+reg cases.
1558 if (!getLdStOffsetOp(MI).isImm())
1561 // Look backward up to LdStLimit instructions.
1562 MachineBasicBlock::iterator StoreI;
1563 if (findMatchingStore(MBBI, LdStLimit, StoreI)) {
1564 ++NumLoadsFromStoresPromoted;
1565 // Promote the load. Keeping the iterator straight is a
1566 // pain, so we let the merge routine tell us what the next instruction
1567 // is after it's done mucking about.
1568 MBBI = promoteLoadFromStore(MBBI, StoreI);
1574 // Merge adjacent zero stores into a wider store.
1575 bool AArch64LoadStoreOpt::tryToMergeZeroStInst(
1576 MachineBasicBlock::iterator &MBBI) {
1577 assert(isPromotableZeroStoreInst(*MBBI) && "Expected narrow store.");
1578 MachineInstr &MI = *MBBI;
1579 MachineBasicBlock::iterator E = MI.getParent()->end();
1581 if (!TII->isCandidateToMergeOrPair(MI))
1584 // Look ahead up to LdStLimit instructions for a mergable instruction.
1585 LdStPairFlags Flags;
1586 MachineBasicBlock::iterator MergeMI =
1587 findMatchingInsn(MBBI, Flags, LdStLimit, /* FindNarrowMerge = */ true);
1589 ++NumZeroStoresPromoted;
1591 // Keeping the iterator straight is a pain, so we let the merge routine tell
1592 // us what the next instruction is after it's done mucking about.
1593 MBBI = mergeNarrowZeroStores(MBBI, MergeMI, Flags);
1599 // Find loads and stores that can be merged into a single load or store pair
1601 bool AArch64LoadStoreOpt::tryToPairLdStInst(MachineBasicBlock::iterator &MBBI) {
1602 MachineInstr &MI = *MBBI;
1603 MachineBasicBlock::iterator E = MI.getParent()->end();
1605 if (!TII->isCandidateToMergeOrPair(MI))
1608 // Early exit if the offset is not possible to match. (6 bits of positive
1609 // range, plus allow an extra one in case we find a later insn that matches
1611 bool IsUnscaled = TII->isUnscaledLdSt(MI);
1612 int Offset = getLdStOffsetOp(MI).getImm();
1613 int OffsetStride = IsUnscaled ? getMemScale(MI) : 1;
1614 // Allow one more for offset.
1616 Offset -= OffsetStride;
1617 if (!inBoundsForPair(IsUnscaled, Offset, OffsetStride))
1620 // Look ahead up to LdStLimit instructions for a pairable instruction.
1621 LdStPairFlags Flags;
1622 MachineBasicBlock::iterator Paired =
1623 findMatchingInsn(MBBI, Flags, LdStLimit, /* FindNarrowMerge = */ false);
1626 if (TII->isUnscaledLdSt(MI))
1627 ++NumUnscaledPairCreated;
1628 // Keeping the iterator straight is a pain, so we let the merge routine tell
1629 // us what the next instruction is after it's done mucking about.
1630 MBBI = mergePairedInsns(MBBI, Paired, Flags);
1636 bool AArch64LoadStoreOpt::tryToMergeLdStUpdate
1637 (MachineBasicBlock::iterator &MBBI) {
1638 MachineInstr &MI = *MBBI;
1639 MachineBasicBlock::iterator E = MI.getParent()->end();
1640 MachineBasicBlock::iterator Update;
1642 // Look forward to try to form a post-index instruction. For example,
1644 // add x20, x20, #32
1646 // ldr x0, [x20], #32
1647 Update = findMatchingUpdateInsnForward(MBBI, 0, UpdateLimit);
1649 // Merge the update into the ld/st.
1650 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/false);
1654 // Don't know how to handle unscaled pre/post-index versions below, so bail.
1655 if (TII->isUnscaledLdSt(MI.getOpcode()))
1658 // Look back to try to find a pre-index instruction. For example,
1662 // ldr x1, [x0, #8]!
1663 Update = findMatchingUpdateInsnBackward(MBBI, UpdateLimit);
1665 // Merge the update into the ld/st.
1666 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true);
1670 // The immediate in the load/store is scaled by the size of the memory
1671 // operation. The immediate in the add we're looking for,
1672 // however, is not, so adjust here.
1673 int UnscaledOffset = getLdStOffsetOp(MI).getImm() * getMemScale(MI);
1675 // Look forward to try to find a post-index instruction. For example,
1676 // ldr x1, [x0, #64]
1679 // ldr x1, [x0, #64]!
1680 Update = findMatchingUpdateInsnForward(MBBI, UnscaledOffset, UpdateLimit);
1682 // Merge the update into the ld/st.
1683 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true);
1690 bool AArch64LoadStoreOpt::optimizeBlock(MachineBasicBlock &MBB,
1691 bool EnableNarrowZeroStOpt) {
1692 bool Modified = false;
1693 // Four tranformations to do here:
1694 // 1) Find loads that directly read from stores and promote them by
1695 // replacing with mov instructions. If the store is wider than the load,
1696 // the load will be replaced with a bitfield extract.
1699 // ldrh w2, [x0, #6]
1703 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1705 if (isPromotableLoadFromStore(*MBBI) && tryToPromoteLoadFromStore(MBBI))
1710 // 2) Merge adjacent zero stores into a wider store.
1713 // strh wzr, [x0, #2]
1718 // str wzr, [x0, #4]
1721 if (EnableNarrowZeroStOpt)
1722 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1724 if (isPromotableZeroStoreInst(*MBBI) && tryToMergeZeroStInst(MBBI))
1729 // 3) Find loads and stores that can be merged into a single load or store
1730 // pair instruction.
1736 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1738 if (TII->isPairableLdStInst(*MBBI) && tryToPairLdStInst(MBBI))
1743 // 4) Find base register updates that can be merged into the load or store
1744 // as a base-reg writeback.
1750 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1752 if (isMergeableLdStUpdate(*MBBI) && tryToMergeLdStUpdate(MBBI))
1761 bool AArch64LoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
1762 if (skipFunction(Fn.getFunction()))
1765 Subtarget = &static_cast<const AArch64Subtarget &>(Fn.getSubtarget());
1766 TII = static_cast<const AArch64InstrInfo *>(Subtarget->getInstrInfo());
1767 TRI = Subtarget->getRegisterInfo();
1768 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1770 // Resize the modified and used register bitfield trackers. We do this once
1771 // per function and then clear the bitfield each time we optimize a load or
1773 ModifiedRegs.resize(TRI->getNumRegs());
1774 UsedRegs.resize(TRI->getNumRegs());
1776 bool Modified = false;
1777 bool enableNarrowZeroStOpt = !Subtarget->requiresStrictAlign();
1778 for (auto &MBB : Fn)
1779 Modified |= optimizeBlock(MBB, enableNarrowZeroStOpt);
1784 // FIXME: Do we need/want a pre-alloc pass like ARM has to try to keep loads and
1785 // stores near one another? Note: The pre-RA instruction scheduler already has
1786 // hooks to try and schedule pairable loads/stores together to improve pairing
1787 // opportunities. Thus, pre-RA pairing pass may not be worth the effort.
1789 // FIXME: When pairing store instructions it's very possible for this pass to
1790 // hoist a store with a KILL marker above another use (without a KILL marker).
1791 // The resulting IR is invalid, but nothing uses the KILL markers after this
1792 // pass, so it's never caused a problem in practice.
1794 /// createAArch64LoadStoreOptimizationPass - returns an instance of the
1795 /// load / store optimization pass.
1796 FunctionPass *llvm::createAArch64LoadStoreOptimizationPass() {
1797 return new AArch64LoadStoreOpt();