LLVM  7.0.0svn
AArch64LoadStoreOptimizer.cpp
Go to the documentation of this file.
1 //===- AArch64LoadStoreOptimizer.cpp - AArch64 load/store opt. pass -------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains a pass that performs load / store related peephole
11 // optimizations. This pass should be run after register allocation.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "AArch64InstrInfo.h"
16 #include "AArch64Subtarget.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"
31 #include "llvm/IR/DebugLoc.h"
32 #include "llvm/MC/MCRegisterInfo.h"
33 #include "llvm/Pass.h"
35 #include "llvm/Support/Debug.h"
38 #include <cassert>
39 #include <cstdint>
40 #include <iterator>
41 #include <limits>
42 
43 using namespace llvm;
44 
45 #define DEBUG_TYPE "aarch64-ldst-opt"
46 
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");
54 
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);
58 
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),
62  cl::Hidden);
63 
64 #define AARCH64_LOAD_STORE_OPT_NAME "AArch64 load / store optimization pass"
65 
66 namespace {
67 
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;
73 
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.
78  int SExtIdx = -1;
79 
80  LdStPairFlags() = default;
81 
82  void setMergeForward(bool V = true) { MergeForward = V; }
83  bool getMergeForward() const { return MergeForward; }
84 
85  void setSExtIdx(int V) { SExtIdx = V; }
86  int getSExtIdx() const { return SExtIdx; }
87 };
88 
89 struct AArch64LoadStoreOpt : public MachineFunctionPass {
90  static char ID;
91 
92  AArch64LoadStoreOpt() : MachineFunctionPass(ID) {
94  }
95 
96  AliasAnalysis *AA;
97  const AArch64InstrInfo *TII;
98  const TargetRegisterInfo *TRI;
99  const AArch64Subtarget *Subtarget;
100 
101  // Track which registers have been modified and used.
102  BitVector ModifiedRegs, UsedRegs;
103 
104  void getAnalysisUsage(AnalysisUsage &AU) const override {
107  }
108 
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().
113  LdStPairFlags &Flags,
114  unsigned Limit,
115  bool FindNarrowMerge);
116 
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,
121 
122  // Merge the two instructions indicated into a wider narrow store instruction.
124  mergeNarrowZeroStores(MachineBasicBlock::iterator I,
126  const LdStPairFlags &Flags);
127 
128  // Merge the two instructions indicated into a single pair-wise instruction.
130  mergePairedInsns(MachineBasicBlock::iterator I,
132  const LdStPairFlags &Flags);
133 
134  // Promote the load that reads directly from the address stored to.
136  promoteLoadFromStore(MachineBasicBlock::iterator LoadI,
138 
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.
143  findMatchingUpdateInsnForward(MachineBasicBlock::iterator I,
144  int UnscaledOffset, unsigned Limit);
145 
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.
150  findMatchingUpdateInsnBackward(MachineBasicBlock::iterator I, unsigned Limit);
151 
152  // Find an instruction that updates the base register of the ld/st
153  // instruction.
154  bool isMatchingUpdateInsn(MachineInstr &MemMI, MachineInstr &MI,
155  unsigned BaseReg, int Offset);
156 
157  // Merge a pre- or post-index base register update into a ld/st instruction.
159  mergeUpdateInsn(MachineBasicBlock::iterator I,
160  MachineBasicBlock::iterator Update, bool IsPreIdx);
161 
162  // Find and merge zero store instructions.
163  bool tryToMergeZeroStInst(MachineBasicBlock::iterator &MBBI);
164 
165  // Find and pair ldr/str instructions.
166  bool tryToPairLdStInst(MachineBasicBlock::iterator &MBBI);
167 
168  // Find and promote load instructions which read directly from store.
169  bool tryToPromoteLoadFromStore(MachineBasicBlock::iterator &MBBI);
170 
171  // Find and merge a base register updates before or after a ld/st instruction.
172  bool tryToMergeLdStUpdate(MachineBasicBlock::iterator &MBBI);
173 
174  bool optimizeBlock(MachineBasicBlock &MBB, bool EnableNarrowZeroStOpt);
175 
176  bool runOnMachineFunction(MachineFunction &Fn) override;
177 
178  MachineFunctionProperties getRequiredProperties() const override {
181  }
182 
183  StringRef getPassName() const override { return AARCH64_LOAD_STORE_OPT_NAME; }
184 };
185 
186 char AArch64LoadStoreOpt::ID = 0;
187 
188 } // end anonymous namespace
189 
190 INITIALIZE_PASS(AArch64LoadStoreOpt, "aarch64-ldst-opt",
191  AARCH64_LOAD_STORE_OPT_NAME, false, false)
192 
193 static bool isNarrowStore(unsigned Opc) {
194  switch (Opc) {
195  default:
196  return false;
197  case AArch64::STRBBui:
198  case AArch64::STURBBi:
199  case AArch64::STRHHui:
200  case AArch64::STURHHi:
201  return true;
202  }
203 }
204 
205 // Scaling factor for unscaled load or store.
207  switch (MI.getOpcode()) {
208  default:
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:
216  return 1;
217  case AArch64::LDRHHui:
218  case AArch64::LDURHHi:
219  case AArch64::LDRSHWui:
220  case AArch64::LDURSHWi:
221  case AArch64::STRHHui:
222  case AArch64::STURHHi:
223  return 2;
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:
234  case AArch64::LDPSi:
235  case AArch64::LDPSWi:
236  case AArch64::LDPWi:
237  case AArch64::STPSi:
238  case AArch64::STPWi:
239  return 4;
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:
248  case AArch64::LDPDi:
249  case AArch64::LDPXi:
250  case AArch64::STPDi:
251  case AArch64::STPXi:
252  return 8;
253  case AArch64::LDRQui:
254  case AArch64::LDURQi:
255  case AArch64::STRQui:
256  case AArch64::STURQi:
257  case AArch64::LDPQi:
258  case AArch64::STPQi:
259  return 16;
260  }
261 }
262 
263 static unsigned getMatchingNonSExtOpcode(unsigned Opc,
264  bool *IsValidLdStrOpc = nullptr) {
265  if (IsValidLdStrOpc)
266  *IsValidLdStrOpc = true;
267  switch (Opc) {
268  default:
269  if (IsValidLdStrOpc)
270  *IsValidLdStrOpc = false;
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:
296  return Opc;
297  case AArch64::LDRSWui:
298  return AArch64::LDRWui;
299  case AArch64::LDURSWi:
300  return AArch64::LDURWi;
301  }
302 }
303 
304 static unsigned getMatchingWideOpcode(unsigned Opc) {
305  switch (Opc) {
306  default:
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;
320  }
321 }
322 
323 static unsigned getMatchingPairOpcode(unsigned Opc) {
324  switch (Opc) {
325  default:
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;
360  }
361 }
362 
365  unsigned LdOpc = LoadInst.getOpcode();
366  unsigned StOpc = StoreInst.getOpcode();
367  switch (LdOpc) {
368  default:
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;
390  }
391 }
392 
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.
398  switch (Opc) {
399  default:
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;
431  case AArch64::LDPSi:
432  return AArch64::LDPSpre;
433  case AArch64::LDPSWi:
434  return AArch64::LDPSWpre;
435  case AArch64::LDPDi:
436  return AArch64::LDPDpre;
437  case AArch64::LDPQi:
438  return AArch64::LDPQpre;
439  case AArch64::LDPWi:
440  return AArch64::LDPWpre;
441  case AArch64::LDPXi:
442  return AArch64::LDPXpre;
443  case AArch64::STPSi:
444  return AArch64::STPSpre;
445  case AArch64::STPDi:
446  return AArch64::STPDpre;
447  case AArch64::STPQi:
448  return AArch64::STPQpre;
449  case AArch64::STPWi:
450  return AArch64::STPWpre;
451  case AArch64::STPXi:
452  return AArch64::STPXpre;
453  }
454 }
455 
456 static unsigned getPostIndexedOpcode(unsigned Opc) {
457  switch (Opc) {
458  default:
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;
500  case AArch64::LDPSi:
501  return AArch64::LDPSpost;
502  case AArch64::LDPSWi:
503  return AArch64::LDPSWpost;
504  case AArch64::LDPDi:
505  return AArch64::LDPDpost;
506  case AArch64::LDPQi:
507  return AArch64::LDPQpost;
508  case AArch64::LDPWi:
509  return AArch64::LDPWpost;
510  case AArch64::LDPXi:
511  return AArch64::LDPXpost;
512  case AArch64::STPSi:
513  return AArch64::STPSpost;
514  case AArch64::STPDi:
515  return AArch64::STPDpost;
516  case AArch64::STPQi:
517  return AArch64::STPQpost;
518  case AArch64::STPWi:
519  return AArch64::STPWpost;
520  case AArch64::STPXi:
521  return AArch64::STPXpost;
522  }
523 }
524 
525 static bool isPairedLdSt(const MachineInstr &MI) {
526  switch (MI.getOpcode()) {
527  default:
528  return false;
529  case AArch64::LDPSi:
530  case AArch64::LDPSWi:
531  case AArch64::LDPDi:
532  case AArch64::LDPQi:
533  case AArch64::LDPWi:
534  case AArch64::LDPXi:
535  case AArch64::STPSi:
536  case AArch64::STPDi:
537  case AArch64::STPQi:
538  case AArch64::STPWi:
539  case AArch64::STPXi:
540  return true;
541  }
542 }
543 
545  unsigned PairedRegOp = 0) {
546  assert(PairedRegOp < 2 && "Unexpected register operand idx.");
547  unsigned Idx = isPairedLdSt(MI) ? PairedRegOp : 0;
548  return MI.getOperand(Idx);
549 }
550 
552  unsigned Idx = isPairedLdSt(MI) ? 2 : 1;
553  return MI.getOperand(Idx);
554 }
555 
557  unsigned Idx = isPairedLdSt(MI) ? 3 : 2;
558  return MI.getOperand(Idx);
559 }
560 
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));
575 }
576 
578  unsigned Opc = MI.getOpcode();
579  return (Opc == AArch64::STRWui || Opc == AArch64::STURWi ||
580  isNarrowStore(Opc)) &&
581  getLdStRegOp(MI).getReg() == AArch64::WZR;
582 }
583 
585  switch (MI.getOpcode()) {
586  default:
587  return false;
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:
598  return true;
599  }
600 }
601 
603  unsigned Opc = MI.getOpcode();
604  switch (Opc) {
605  default:
606  return false;
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.
634  case AArch64::LDPSi:
635  case AArch64::LDPSWi:
636  case AArch64::LDPDi:
637  case AArch64::LDPQi:
638  case AArch64::LDPWi:
639  case AArch64::LDPXi:
640  case AArch64::STPSi:
641  case AArch64::STPDi:
642  case AArch64::STPQi:
643  case AArch64::STPWi:
644  case AArch64::STPXi:
645  // Make sure this is a reg+imm (as opposed to an address reloc).
646  if (!getLdStOffsetOp(MI).isImm())
647  return false;
648 
649  return true;
650  }
651 }
652 
654 AArch64LoadStoreOpt::mergeNarrowZeroStores(MachineBasicBlock::iterator I,
656  const LdStPairFlags &Flags) {
658  "Expected promotable zero stores.");
659 
661  ++NextI;
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)
667  ++NextI;
668 
669  unsigned Opc = I->getOpcode();
670  bool IsScaled = !TII->isUnscaledLdSt(Opc);
671  int OffsetStride = IsScaled ? 1 : getMemScale(*I);
672 
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);
681 
682  // Which register is Rt and which is Rt2 depends on the offset order.
683  MachineInstr *RtMI;
684  if (getLdStOffsetOp(*I).getImm() ==
685  getLdStOffsetOp(*MergeMI).getImm() + OffsetStride)
686  RtMI = &*MergeMI;
687  else
688  RtMI = &*I;
689 
690  int OffsetImm = getLdStOffsetOp(*RtMI).getImm();
691  // Change the scaled offset from small to large type.
692  if (IsScaled) {
693  assert(((OffsetImm & 1) == 0) && "Unexpected offset to merge");
694  OffsetImm /= 2;
695  }
696 
697  // Construct the new instruction.
698  DebugLoc DL = I->getDebugLoc();
699  MachineBasicBlock *MBB = I->getParent();
701  MIB = BuildMI(*MBB, InsertionPoint, DL, TII->get(getMatchingWideOpcode(Opc)))
702  .addReg(isNarrowStore(Opc) ? AArch64::WZR : AArch64::XZR)
703  .add(BaseRegOp)
704  .addImm(OffsetImm)
705  .setMemRefs(I->mergeMemRefsWith(*MergeMI))
706  .setMIFlags(I->mergeFlagsWith(*MergeMI));
707  (void)MIB;
708 
709  DEBUG(dbgs() << "Creating wider store. Replacing instructions:\n ");
710  DEBUG(I->print(dbgs()));
711  DEBUG(dbgs() << " ");
712  DEBUG(MergeMI->print(dbgs()));
713  DEBUG(dbgs() << " with instruction:\n ");
714  DEBUG(((MachineInstr *)MIB)->print(dbgs()));
715  DEBUG(dbgs() << "\n");
716 
717  // Erase the old instructions.
718  I->eraseFromParent();
719  MergeMI->eraseFromParent();
720  return NextI;
721 }
722 
724 AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I,
726  const LdStPairFlags &Flags) {
728  ++NextI;
729  // If NextI is the second of the two instructions to be merged, we need
730  // to skip one further. Either way we merge will invalidate the iterator,
731  // and we don't need to scan the new instruction, as it's a pairwise
732  // instruction, which we're not considering for further action anyway.
733  if (NextI == Paired)
734  ++NextI;
735 
736  int SExtIdx = Flags.getSExtIdx();
737  unsigned Opc =
738  SExtIdx == -1 ? I->getOpcode() : getMatchingNonSExtOpcode(I->getOpcode());
739  bool IsUnscaled = TII->isUnscaledLdSt(Opc);
740  int OffsetStride = IsUnscaled ? getMemScale(*I) : 1;
741 
742  bool MergeForward = Flags.getMergeForward();
743  // Insert our new paired instruction after whichever of the paired
744  // instructions MergeForward indicates.
745  MachineBasicBlock::iterator InsertionPoint = MergeForward ? Paired : I;
746  // Also based on MergeForward is from where we copy the base register operand
747  // so we get the flags compatible with the input code.
748  const MachineOperand &BaseRegOp =
749  MergeForward ? getLdStBaseOp(*Paired) : getLdStBaseOp(*I);
750 
751  int Offset = getLdStOffsetOp(*I).getImm();
752  int PairedOffset = getLdStOffsetOp(*Paired).getImm();
753  bool PairedIsUnscaled = TII->isUnscaledLdSt(Paired->getOpcode());
754  if (IsUnscaled != PairedIsUnscaled) {
755  // We're trying to pair instructions that differ in how they are scaled. If
756  // I is scaled then scale the offset of Paired accordingly. Otherwise, do
757  // the opposite (i.e., make Paired's offset unscaled).
758  int MemSize = getMemScale(*Paired);
759  if (PairedIsUnscaled) {
760  // If the unscaled offset isn't a multiple of the MemSize, we can't
761  // pair the operations together.
762  assert(!(PairedOffset % getMemScale(*Paired)) &&
763  "Offset should be a multiple of the stride!");
764  PairedOffset /= MemSize;
765  } else {
766  PairedOffset *= MemSize;
767  }
768  }
769 
770  // Which register is Rt and which is Rt2 depends on the offset order.
771  MachineInstr *RtMI, *Rt2MI;
772  if (Offset == PairedOffset + OffsetStride) {
773  RtMI = &*Paired;
774  Rt2MI = &*I;
775  // Here we swapped the assumption made for SExtIdx.
776  // I.e., we turn ldp I, Paired into ldp Paired, I.
777  // Update the index accordingly.
778  if (SExtIdx != -1)
779  SExtIdx = (SExtIdx + 1) % 2;
780  } else {
781  RtMI = &*I;
782  Rt2MI = &*Paired;
783  }
784  int OffsetImm = getLdStOffsetOp(*RtMI).getImm();
785  // Scale the immediate offset, if necessary.
786  if (TII->isUnscaledLdSt(RtMI->getOpcode())) {
787  assert(!(OffsetImm % getMemScale(*RtMI)) &&
788  "Unscaled offset cannot be scaled.");
789  OffsetImm /= getMemScale(*RtMI);
790  }
791 
792  // Construct the new instruction.
794  DebugLoc DL = I->getDebugLoc();
795  MachineBasicBlock *MBB = I->getParent();
796  MachineOperand RegOp0 = getLdStRegOp(*RtMI);
797  MachineOperand RegOp1 = getLdStRegOp(*Rt2MI);
798  // Kill flags may become invalid when moving stores for pairing.
799  if (RegOp0.isUse()) {
800  if (!MergeForward) {
801  // Clear kill flags on store if moving upwards. Example:
802  // STRWui %w0, ...
803  // USE %w1
804  // STRWui kill %w1 ; need to clear kill flag when moving STRWui upwards
805  RegOp0.setIsKill(false);
806  RegOp1.setIsKill(false);
807  } else {
808  // Clear kill flags of the first stores register. Example:
809  // STRWui %w1, ...
810  // USE kill %w1 ; need to clear kill flag when moving STRWui downwards
811  // STRW %w0
812  unsigned Reg = getLdStRegOp(*I).getReg();
813  for (MachineInstr &MI : make_range(std::next(I), Paired))
814  MI.clearRegisterKills(Reg, TRI);
815  }
816  }
817  MIB = BuildMI(*MBB, InsertionPoint, DL, TII->get(getMatchingPairOpcode(Opc)))
818  .add(RegOp0)
819  .add(RegOp1)
820  .add(BaseRegOp)
821  .addImm(OffsetImm)
822  .setMemRefs(I->mergeMemRefsWith(*Paired))
823  .setMIFlags(I->mergeFlagsWith(*Paired));
824 
825  (void)MIB;
826 
827  DEBUG(dbgs() << "Creating pair load/store. Replacing instructions:\n ");
828  DEBUG(I->print(dbgs()));
829  DEBUG(dbgs() << " ");
830  DEBUG(Paired->print(dbgs()));
831  DEBUG(dbgs() << " with instruction:\n ");
832  if (SExtIdx != -1) {
833  // Generate the sign extension for the proper result of the ldp.
834  // I.e., with X1, that would be:
835  // %w1 = KILL %w1, implicit-def %x1
836  // %x1 = SBFMXri killed %x1, 0, 31
837  MachineOperand &DstMO = MIB->getOperand(SExtIdx);
838  // Right now, DstMO has the extended register, since it comes from an
839  // extended opcode.
840  unsigned DstRegX = DstMO.getReg();
841  // Get the W variant of that register.
842  unsigned DstRegW = TRI->getSubReg(DstRegX, AArch64::sub_32);
843  // Update the result of LDP to use the W instead of the X variant.
844  DstMO.setReg(DstRegW);
845  DEBUG(((MachineInstr *)MIB)->print(dbgs()));
846  DEBUG(dbgs() << "\n");
847  // Make the machine verifier happy by providing a definition for
848  // the X register.
849  // Insert this definition right after the generated LDP, i.e., before
850  // InsertionPoint.
851  MachineInstrBuilder MIBKill =
852  BuildMI(*MBB, InsertionPoint, DL, TII->get(TargetOpcode::KILL), DstRegW)
853  .addReg(DstRegW)
854  .addReg(DstRegX, RegState::Define);
855  MIBKill->getOperand(2).setImplicit();
856  // Create the sign extension.
857  MachineInstrBuilder MIBSXTW =
858  BuildMI(*MBB, InsertionPoint, DL, TII->get(AArch64::SBFMXri), DstRegX)
859  .addReg(DstRegX)
860  .addImm(0)
861  .addImm(31);
862  (void)MIBSXTW;
863  DEBUG(dbgs() << " Extend operand:\n ");
864  DEBUG(((MachineInstr *)MIBSXTW)->print(dbgs()));
865  } else {
866  DEBUG(((MachineInstr *)MIB)->print(dbgs()));
867  }
868  DEBUG(dbgs() << "\n");
869 
870  // Erase the old instructions.
871  I->eraseFromParent();
872  Paired->eraseFromParent();
873 
874  return NextI;
875 }
876 
878 AArch64LoadStoreOpt::promoteLoadFromStore(MachineBasicBlock::iterator LoadI,
880  MachineBasicBlock::iterator NextI = LoadI;
881  ++NextI;
882 
883  int LoadSize = getMemScale(*LoadI);
884  int StoreSize = getMemScale(*StoreI);
885  unsigned LdRt = getLdStRegOp(*LoadI).getReg();
886  const MachineOperand &StMO = getLdStRegOp(*StoreI);
887  unsigned StRt = getLdStRegOp(*StoreI).getReg();
888  bool IsStoreXReg = TRI->getRegClass(AArch64::GPR64RegClassID)->contains(StRt);
889 
890  assert((IsStoreXReg ||
891  TRI->getRegClass(AArch64::GPR32RegClassID)->contains(StRt)) &&
892  "Unexpected RegClass");
893 
894  MachineInstr *BitExtMI;
895  if (LoadSize == StoreSize && (LoadSize == 4 || LoadSize == 8)) {
896  // Remove the load, if the destination register of the loads is the same
897  // register for stored value.
898  if (StRt == LdRt && LoadSize == 8) {
899  for (MachineInstr &MI : make_range(StoreI->getIterator(),
900  LoadI->getIterator())) {
901  if (MI.killsRegister(StRt, TRI)) {
902  MI.clearRegisterKills(StRt, TRI);
903  break;
904  }
905  }
906  DEBUG(dbgs() << "Remove load instruction:\n ");
907  DEBUG(LoadI->print(dbgs()));
908  DEBUG(dbgs() << "\n");
909  LoadI->eraseFromParent();
910  return NextI;
911  }
912  // Replace the load with a mov if the load and store are in the same size.
913  BitExtMI =
914  BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(),
915  TII->get(IsStoreXReg ? AArch64::ORRXrs : AArch64::ORRWrs), LdRt)
916  .addReg(IsStoreXReg ? AArch64::XZR : AArch64::WZR)
917  .add(StMO)
919  .setMIFlags(LoadI->getFlags());
920  } else {
921  // FIXME: Currently we disable this transformation in big-endian targets as
922  // performance and correctness are verified only in little-endian.
923  if (!Subtarget->isLittleEndian())
924  return NextI;
925  bool IsUnscaled = TII->isUnscaledLdSt(*LoadI);
926  assert(IsUnscaled == TII->isUnscaledLdSt(*StoreI) &&
927  "Unsupported ld/st match");
928  assert(LoadSize <= StoreSize && "Invalid load size");
929  int UnscaledLdOffset = IsUnscaled
930  ? getLdStOffsetOp(*LoadI).getImm()
931  : getLdStOffsetOp(*LoadI).getImm() * LoadSize;
932  int UnscaledStOffset = IsUnscaled
933  ? getLdStOffsetOp(*StoreI).getImm()
934  : getLdStOffsetOp(*StoreI).getImm() * StoreSize;
935  int Width = LoadSize * 8;
936  int Immr = 8 * (UnscaledLdOffset - UnscaledStOffset);
937  int Imms = Immr + Width - 1;
938  unsigned DestReg = IsStoreXReg
939  ? TRI->getMatchingSuperReg(LdRt, AArch64::sub_32,
940  &AArch64::GPR64RegClass)
941  : LdRt;
942 
943  assert((UnscaledLdOffset >= UnscaledStOffset &&
944  (UnscaledLdOffset + LoadSize) <= UnscaledStOffset + StoreSize) &&
945  "Invalid offset");
946 
947  Immr = 8 * (UnscaledLdOffset - UnscaledStOffset);
948  Imms = Immr + Width - 1;
949  if (UnscaledLdOffset == UnscaledStOffset) {
950  uint32_t AndMaskEncoded = ((IsStoreXReg ? 1 : 0) << 12) // N
951  | ((Immr) << 6) // immr
952  | ((Imms) << 0) // imms
953  ;
954 
955  BitExtMI =
956  BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(),
957  TII->get(IsStoreXReg ? AArch64::ANDXri : AArch64::ANDWri),
958  DestReg)
959  .add(StMO)
960  .addImm(AndMaskEncoded)
961  .setMIFlags(LoadI->getFlags());
962  } else {
963  BitExtMI =
964  BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(),
965  TII->get(IsStoreXReg ? AArch64::UBFMXri : AArch64::UBFMWri),
966  DestReg)
967  .add(StMO)
968  .addImm(Immr)
969  .addImm(Imms)
970  .setMIFlags(LoadI->getFlags());
971  }
972  }
973 
974  // Clear kill flags between store and load.
975  for (MachineInstr &MI : make_range(StoreI->getIterator(),
976  BitExtMI->getIterator()))
977  if (MI.killsRegister(StRt, TRI)) {
978  MI.clearRegisterKills(StRt, TRI);
979  break;
980  }
981 
982  DEBUG(dbgs() << "Promoting load by replacing :\n ");
983  DEBUG(StoreI->print(dbgs()));
984  DEBUG(dbgs() << " ");
985  DEBUG(LoadI->print(dbgs()));
986  DEBUG(dbgs() << " with instructions:\n ");
987  DEBUG(StoreI->print(dbgs()));
988  DEBUG(dbgs() << " ");
989  DEBUG((BitExtMI)->print(dbgs()));
990  DEBUG(dbgs() << "\n");
991 
992  // Erase the old instructions.
993  LoadI->eraseFromParent();
994  return NextI;
995 }
996 
997 static bool inBoundsForPair(bool IsUnscaled, int Offset, int OffsetStride) {
998  // Convert the byte-offset used by unscaled into an "element" offset used
999  // by the scaled pair load/store instructions.
1000  if (IsUnscaled) {
1001  // If the byte-offset isn't a multiple of the stride, there's no point
1002  // trying to match it.
1003  if (Offset % OffsetStride)
1004  return false;
1005  Offset /= OffsetStride;
1006  }
1007  return Offset <= 63 && Offset >= -64;
1008 }
1009 
1010 // Do alignment, specialized to power of 2 and for signed ints,
1011 // avoiding having to do a C-style cast from uint_64t to int when
1012 // using alignTo from include/llvm/Support/MathExtras.h.
1013 // FIXME: Move this function to include/MathExtras.h?
1014 static int alignTo(int Num, int PowOf2) {
1015  return (Num + PowOf2 - 1) & ~(PowOf2 - 1);
1016 }
1017 
1018 static bool mayAlias(MachineInstr &MIa, MachineInstr &MIb,
1019  AliasAnalysis *AA) {
1020  // One of the instructions must modify memory.
1021  if (!MIa.mayStore() && !MIb.mayStore())
1022  return false;
1023 
1024  // Both instructions must be memory operations.
1025  if (!MIa.mayLoadOrStore() && !MIb.mayLoadOrStore())
1026  return false;
1027 
1028  return MIa.mayAlias(AA, MIb, /*UseTBAA*/false);
1029 }
1030 
1031 static bool mayAlias(MachineInstr &MIa,
1033  AliasAnalysis *AA) {
1034  for (MachineInstr *MIb : MemInsns)
1035  if (mayAlias(MIa, *MIb, AA))
1036  return true;
1037 
1038  return false;
1039 }
1040 
1041 bool AArch64LoadStoreOpt::findMatchingStore(
1042  MachineBasicBlock::iterator I, unsigned Limit,
1043  MachineBasicBlock::iterator &StoreI) {
1044  MachineBasicBlock::iterator B = I->getParent()->begin();
1046  MachineInstr &LoadMI = *I;
1047  unsigned BaseReg = getLdStBaseOp(LoadMI).getReg();
1048 
1049  // If the load is the first instruction in the block, there's obviously
1050  // not any matching store.
1051  if (MBBI == B)
1052  return false;
1053 
1054  // Track which registers have been modified and used between the first insn
1055  // and the second insn.
1056  ModifiedRegs.reset();
1057  UsedRegs.reset();
1058 
1059  unsigned Count = 0;
1060  do {
1061  --MBBI;
1062  MachineInstr &MI = *MBBI;
1063 
1064  // Don't count transient instructions towards the search limit since there
1065  // may be different numbers of them if e.g. debug information is present.
1066  if (!MI.isTransient())
1067  ++Count;
1068 
1069  // If the load instruction reads directly from the address to which the
1070  // store instruction writes and the stored value is not modified, we can
1071  // promote the load. Since we do not handle stores with pre-/post-index,
1072  // it's unnecessary to check if BaseReg is modified by the store itself.
1073  if (MI.mayStore() && isMatchingStore(LoadMI, MI) &&
1074  BaseReg == getLdStBaseOp(MI).getReg() &&
1075  isLdOffsetInRangeOfSt(LoadMI, MI, TII) &&
1076  !ModifiedRegs[getLdStRegOp(MI).getReg()]) {
1077  StoreI = MBBI;
1078  return true;
1079  }
1080 
1081  if (MI.isCall())
1082  return false;
1083 
1084  // Update modified / uses register lists.
1085  TII->trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1086 
1087  // Otherwise, if the base register is modified, we have no match, so
1088  // return early.
1089  if (ModifiedRegs[BaseReg])
1090  return false;
1091 
1092  // If we encounter a store aliased with the load, return early.
1093  if (MI.mayStore() && mayAlias(LoadMI, MI, AA))
1094  return false;
1095  } while (MBBI != B && Count < Limit);
1096  return false;
1097 }
1098 
1099 // Returns true if FirstMI and MI are candidates for merging or pairing.
1100 // Otherwise, returns false.
1102  LdStPairFlags &Flags,
1103  const AArch64InstrInfo *TII) {
1104  // If this is volatile or if pairing is suppressed, not a candidate.
1105  if (MI.hasOrderedMemoryRef() || TII->isLdStPairSuppressed(MI))
1106  return false;
1107 
1108  // We should have already checked FirstMI for pair suppression and volatility.
1109  assert(!FirstMI.hasOrderedMemoryRef() &&
1110  !TII->isLdStPairSuppressed(FirstMI) &&
1111  "FirstMI shouldn't get here if either of these checks are true.");
1112 
1113  unsigned OpcA = FirstMI.getOpcode();
1114  unsigned OpcB = MI.getOpcode();
1115 
1116  // Opcodes match: nothing more to check.
1117  if (OpcA == OpcB)
1118  return true;
1119 
1120  // Try to match a sign-extended load/store with a zero-extended load/store.
1121  bool IsValidLdStrOpc, PairIsValidLdStrOpc;
1122  unsigned NonSExtOpc = getMatchingNonSExtOpcode(OpcA, &IsValidLdStrOpc);
1123  assert(IsValidLdStrOpc &&
1124  "Given Opc should be a Load or Store with an immediate");
1125  // OpcA will be the first instruction in the pair.
1126  if (NonSExtOpc == getMatchingNonSExtOpcode(OpcB, &PairIsValidLdStrOpc)) {
1127  Flags.setSExtIdx(NonSExtOpc == (unsigned)OpcA ? 1 : 0);
1128  return true;
1129  }
1130 
1131  // If the second instruction isn't even a mergable/pairable load/store, bail
1132  // out.
1133  if (!PairIsValidLdStrOpc)
1134  return false;
1135 
1136  // FIXME: We don't support merging narrow stores with mixed scaled/unscaled
1137  // offsets.
1138  if (isNarrowStore(OpcA) || isNarrowStore(OpcB))
1139  return false;
1140 
1141  // Try to match an unscaled load/store with a scaled load/store.
1142  return TII->isUnscaledLdSt(OpcA) != TII->isUnscaledLdSt(OpcB) &&
1144 
1145  // FIXME: Can we also match a mixed sext/zext unscaled/scaled pair?
1146 }
1147 
1148 /// Scan the instructions looking for a load/store that can be combined with the
1149 /// current instruction into a wider equivalent or a load/store pair.
1151 AArch64LoadStoreOpt::findMatchingInsn(MachineBasicBlock::iterator I,
1152  LdStPairFlags &Flags, unsigned Limit,
1153  bool FindNarrowMerge) {
1154  MachineBasicBlock::iterator E = I->getParent()->end();
1156  MachineInstr &FirstMI = *I;
1157  ++MBBI;
1158 
1159  bool MayLoad = FirstMI.mayLoad();
1160  bool IsUnscaled = TII->isUnscaledLdSt(FirstMI);
1161  unsigned Reg = getLdStRegOp(FirstMI).getReg();
1162  unsigned BaseReg = getLdStBaseOp(FirstMI).getReg();
1163  int Offset = getLdStOffsetOp(FirstMI).getImm();
1164  int OffsetStride = IsUnscaled ? getMemScale(FirstMI) : 1;
1165  bool IsPromotableZeroStore = isPromotableZeroStoreInst(FirstMI);
1166 
1167  // Track which registers have been modified and used between the first insn
1168  // (inclusive) and the second insn.
1169  ModifiedRegs.reset();
1170  UsedRegs.reset();
1171 
1172  // Remember any instructions that read/write memory between FirstMI and MI.
1174 
1175  for (unsigned Count = 0; MBBI != E && Count < Limit; ++MBBI) {
1176  MachineInstr &MI = *MBBI;
1177 
1178  // Don't count transient instructions towards the search limit since there
1179  // may be different numbers of them if e.g. debug information is present.
1180  if (!MI.isTransient())
1181  ++Count;
1182 
1183  Flags.setSExtIdx(-1);
1184  if (areCandidatesToMergeOrPair(FirstMI, MI, Flags, TII) &&
1185  getLdStOffsetOp(MI).isImm()) {
1186  assert(MI.mayLoadOrStore() && "Expected memory operation.");
1187  // If we've found another instruction with the same opcode, check to see
1188  // if the base and offset are compatible with our starting instruction.
1189  // These instructions all have scaled immediate operands, so we just
1190  // check for +1/-1. Make sure to check the new instruction offset is
1191  // actually an immediate and not a symbolic reference destined for
1192  // a relocation.
1193  unsigned MIBaseReg = getLdStBaseOp(MI).getReg();
1194  int MIOffset = getLdStOffsetOp(MI).getImm();
1195  bool MIIsUnscaled = TII->isUnscaledLdSt(MI);
1196  if (IsUnscaled != MIIsUnscaled) {
1197  // We're trying to pair instructions that differ in how they are scaled.
1198  // If FirstMI is scaled then scale the offset of MI accordingly.
1199  // Otherwise, do the opposite (i.e., make MI's offset unscaled).
1200  int MemSize = getMemScale(MI);
1201  if (MIIsUnscaled) {
1202  // If the unscaled offset isn't a multiple of the MemSize, we can't
1203  // pair the operations together: bail and keep looking.
1204  if (MIOffset % MemSize) {
1205  TII->trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1206  MemInsns.push_back(&MI);
1207  continue;
1208  }
1209  MIOffset /= MemSize;
1210  } else {
1211  MIOffset *= MemSize;
1212  }
1213  }
1214 
1215  if (BaseReg == MIBaseReg && ((Offset == MIOffset + OffsetStride) ||
1216  (Offset + OffsetStride == MIOffset))) {
1217  int MinOffset = Offset < MIOffset ? Offset : MIOffset;
1218  if (FindNarrowMerge) {
1219  // If the alignment requirements of the scaled wide load/store
1220  // instruction can't express the offset of the scaled narrow input,
1221  // bail and keep looking. For promotable zero stores, allow only when
1222  // the stored value is the same (i.e., WZR).
1223  if ((!IsUnscaled && alignTo(MinOffset, 2) != MinOffset) ||
1224  (IsPromotableZeroStore && Reg != getLdStRegOp(MI).getReg())) {
1225  TII->trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1226  MemInsns.push_back(&MI);
1227  continue;
1228  }
1229  } else {
1230  // Pairwise instructions have a 7-bit signed offset field. Single
1231  // insns have a 12-bit unsigned offset field. If the resultant
1232  // immediate offset of merging these instructions is out of range for
1233  // a pairwise instruction, bail and keep looking.
1234  if (!inBoundsForPair(IsUnscaled, MinOffset, OffsetStride)) {
1235  TII->trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1236  MemInsns.push_back(&MI);
1237  continue;
1238  }
1239  // If the alignment requirements of the paired (scaled) instruction
1240  // can't express the offset of the unscaled input, bail and keep
1241  // looking.
1242  if (IsUnscaled && (alignTo(MinOffset, OffsetStride) != MinOffset)) {
1243  TII->trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1244  MemInsns.push_back(&MI);
1245  continue;
1246  }
1247  }
1248  // If the destination register of the loads is the same register, bail
1249  // and keep looking. A load-pair instruction with both destination
1250  // registers the same is UNPREDICTABLE and will result in an exception.
1251  if (MayLoad && Reg == getLdStRegOp(MI).getReg()) {
1252  TII->trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1253  MemInsns.push_back(&MI);
1254  continue;
1255  }
1256 
1257  // If the Rt of the second instruction was not modified or used between
1258  // the two instructions and none of the instructions between the second
1259  // and first alias with the second, we can combine the second into the
1260  // first.
1261  if (!ModifiedRegs[getLdStRegOp(MI).getReg()] &&
1262  !(MI.mayLoad() && UsedRegs[getLdStRegOp(MI).getReg()]) &&
1263  !mayAlias(MI, MemInsns, AA)) {
1264  Flags.setMergeForward(false);
1265  return MBBI;
1266  }
1267 
1268  // Likewise, if the Rt of the first instruction is not modified or used
1269  // between the two instructions and none of the instructions between the
1270  // first and the second alias with the first, we can combine the first
1271  // into the second.
1272  if (!ModifiedRegs[getLdStRegOp(FirstMI).getReg()] &&
1273  !(MayLoad && UsedRegs[getLdStRegOp(FirstMI).getReg()]) &&
1274  !mayAlias(FirstMI, MemInsns, AA)) {
1275  Flags.setMergeForward(true);
1276  return MBBI;
1277  }
1278  // Unable to combine these instructions due to interference in between.
1279  // Keep looking.
1280  }
1281  }
1282 
1283  // If the instruction wasn't a matching load or store. Stop searching if we
1284  // encounter a call instruction that might modify memory.
1285  if (MI.isCall())
1286  return E;
1287 
1288  // Update modified / uses register lists.
1289  TII->trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1290 
1291  // Otherwise, if the base register is modified, we have no match, so
1292  // return early.
1293  if (ModifiedRegs[BaseReg])
1294  return E;
1295 
1296  // Update list of instructions that read/write memory.
1297  if (MI.mayLoadOrStore())
1298  MemInsns.push_back(&MI);
1299  }
1300  return E;
1301 }
1302 
1304 AArch64LoadStoreOpt::mergeUpdateInsn(MachineBasicBlock::iterator I,
1306  bool IsPreIdx) {
1307  assert((Update->getOpcode() == AArch64::ADDXri ||
1308  Update->getOpcode() == AArch64::SUBXri) &&
1309  "Unexpected base register update instruction to merge!");
1311  // Return the instruction following the merged instruction, which is
1312  // the instruction following our unmerged load. Unless that's the add/sub
1313  // instruction we're merging, in which case it's the one after that.
1314  if (++NextI == Update)
1315  ++NextI;
1316 
1317  int Value = Update->getOperand(2).getImm();
1318  assert(AArch64_AM::getShiftValue(Update->getOperand(3).getImm()) == 0 &&
1319  "Can't merge 1 << 12 offset into pre-/post-indexed load / store");
1320  if (Update->getOpcode() == AArch64::SUBXri)
1321  Value = -Value;
1322 
1323  unsigned NewOpc = IsPreIdx ? getPreIndexedOpcode(I->getOpcode())
1324  : getPostIndexedOpcode(I->getOpcode());
1325  MachineInstrBuilder MIB;
1326  if (!isPairedLdSt(*I)) {
1327  // Non-paired instruction.
1328  MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc))
1329  .add(getLdStRegOp(*Update))
1330  .add(getLdStRegOp(*I))
1331  .add(getLdStBaseOp(*I))
1332  .addImm(Value)
1333  .setMemRefs(I->memoperands_begin(), I->memoperands_end())
1334  .setMIFlags(I->mergeFlagsWith(*Update));
1335  } else {
1336  // Paired instruction.
1337  int Scale = getMemScale(*I);
1338  MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc))
1339  .add(getLdStRegOp(*Update))
1340  .add(getLdStRegOp(*I, 0))
1341  .add(getLdStRegOp(*I, 1))
1342  .add(getLdStBaseOp(*I))
1343  .addImm(Value / Scale)
1344  .setMemRefs(I->memoperands_begin(), I->memoperands_end())
1345  .setMIFlags(I->mergeFlagsWith(*Update));
1346  }
1347  (void)MIB;
1348 
1349  if (IsPreIdx) {
1350  ++NumPreFolded;
1351  DEBUG(dbgs() << "Creating pre-indexed load/store.");
1352  } else {
1353  ++NumPostFolded;
1354  DEBUG(dbgs() << "Creating post-indexed load/store.");
1355  }
1356  DEBUG(dbgs() << " Replacing instructions:\n ");
1357  DEBUG(I->print(dbgs()));
1358  DEBUG(dbgs() << " ");
1359  DEBUG(Update->print(dbgs()));
1360  DEBUG(dbgs() << " with instruction:\n ");
1361  DEBUG(((MachineInstr *)MIB)->print(dbgs()));
1362  DEBUG(dbgs() << "\n");
1363 
1364  // Erase the old instructions for the block.
1365  I->eraseFromParent();
1366  Update->eraseFromParent();
1367 
1368  return NextI;
1369 }
1370 
1371 bool AArch64LoadStoreOpt::isMatchingUpdateInsn(MachineInstr &MemMI,
1372  MachineInstr &MI,
1373  unsigned BaseReg, int Offset) {
1374  switch (MI.getOpcode()) {
1375  default:
1376  break;
1377  case AArch64::SUBXri:
1378  case AArch64::ADDXri:
1379  // Make sure it's a vanilla immediate operand, not a relocation or
1380  // anything else we can't handle.
1381  if (!MI.getOperand(2).isImm())
1382  break;
1383  // Watch out for 1 << 12 shifted value.
1385  break;
1386 
1387  // The update instruction source and destination register must be the
1388  // same as the load/store base register.
1389  if (MI.getOperand(0).getReg() != BaseReg ||
1390  MI.getOperand(1).getReg() != BaseReg)
1391  break;
1392 
1393  bool IsPairedInsn = isPairedLdSt(MemMI);
1394  int UpdateOffset = MI.getOperand(2).getImm();
1395  if (MI.getOpcode() == AArch64::SUBXri)
1396  UpdateOffset = -UpdateOffset;
1397 
1398  // For non-paired load/store instructions, the immediate must fit in a
1399  // signed 9-bit integer.
1400  if (!IsPairedInsn && (UpdateOffset > 255 || UpdateOffset < -256))
1401  break;
1402 
1403  // For paired load/store instructions, the immediate must be a multiple of
1404  // the scaling factor. The scaled offset must also fit into a signed 7-bit
1405  // integer.
1406  if (IsPairedInsn) {
1407  int Scale = getMemScale(MemMI);
1408  if (UpdateOffset % Scale != 0)
1409  break;
1410 
1411  int ScaledOffset = UpdateOffset / Scale;
1412  if (ScaledOffset > 63 || ScaledOffset < -64)
1413  break;
1414  }
1415 
1416  // If we have a non-zero Offset, we check that it matches the amount
1417  // we're adding to the register.
1418  if (!Offset || Offset == UpdateOffset)
1419  return true;
1420  break;
1421  }
1422  return false;
1423 }
1424 
1425 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnForward(
1426  MachineBasicBlock::iterator I, int UnscaledOffset, unsigned Limit) {
1427  MachineBasicBlock::iterator E = I->getParent()->end();
1428  MachineInstr &MemMI = *I;
1430 
1431  unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
1432  int MIUnscaledOffset = getLdStOffsetOp(MemMI).getImm() * getMemScale(MemMI);
1433 
1434  // Scan forward looking for post-index opportunities. Updating instructions
1435  // can't be formed if the memory instruction doesn't have the offset we're
1436  // looking for.
1437  if (MIUnscaledOffset != UnscaledOffset)
1438  return E;
1439 
1440  // If the base register overlaps a destination register, we can't
1441  // merge the update.
1442  bool IsPairedInsn = isPairedLdSt(MemMI);
1443  for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) {
1444  unsigned DestReg = getLdStRegOp(MemMI, i).getReg();
1445  if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg))
1446  return E;
1447  }
1448 
1449  // Track which registers have been modified and used between the first insn
1450  // (inclusive) and the second insn.
1451  ModifiedRegs.reset();
1452  UsedRegs.reset();
1453  ++MBBI;
1454  for (unsigned Count = 0; MBBI != E && Count < Limit; ++MBBI) {
1455  MachineInstr &MI = *MBBI;
1456 
1457  // Don't count transient instructions towards the search limit since there
1458  // may be different numbers of them if e.g. debug information is present.
1459  if (!MI.isTransient())
1460  ++Count;
1461 
1462  // If we found a match, return it.
1463  if (isMatchingUpdateInsn(*I, MI, BaseReg, UnscaledOffset))
1464  return MBBI;
1465 
1466  // Update the status of what the instruction clobbered and used.
1467  TII->trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1468 
1469  // Otherwise, if the base register is used or modified, we have no match, so
1470  // return early.
1471  if (ModifiedRegs[BaseReg] || UsedRegs[BaseReg])
1472  return E;
1473  }
1474  return E;
1475 }
1476 
1477 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnBackward(
1478  MachineBasicBlock::iterator I, unsigned Limit) {
1479  MachineBasicBlock::iterator B = I->getParent()->begin();
1480  MachineBasicBlock::iterator E = I->getParent()->end();
1481  MachineInstr &MemMI = *I;
1483 
1484  unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
1485  int Offset = getLdStOffsetOp(MemMI).getImm();
1486 
1487  // If the load/store is the first instruction in the block, there's obviously
1488  // not any matching update. Ditto if the memory offset isn't zero.
1489  if (MBBI == B || Offset != 0)
1490  return E;
1491  // If the base register overlaps a destination register, we can't
1492  // merge the update.
1493  bool IsPairedInsn = isPairedLdSt(MemMI);
1494  for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) {
1495  unsigned DestReg = getLdStRegOp(MemMI, i).getReg();
1496  if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg))
1497  return E;
1498  }
1499 
1500  // Track which registers have been modified and used between the first insn
1501  // (inclusive) and the second insn.
1502  ModifiedRegs.reset();
1503  UsedRegs.reset();
1504  unsigned Count = 0;
1505  do {
1506  --MBBI;
1507  MachineInstr &MI = *MBBI;
1508 
1509  // Don't count transient instructions towards the search limit since there
1510  // may be different numbers of them if e.g. debug information is present.
1511  if (!MI.isTransient())
1512  ++Count;
1513 
1514  // If we found a match, return it.
1515  if (isMatchingUpdateInsn(*I, MI, BaseReg, Offset))
1516  return MBBI;
1517 
1518  // Update the status of what the instruction clobbered and used.
1519  TII->trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
1520 
1521  // Otherwise, if the base register is used or modified, we have no match, so
1522  // return early.
1523  if (ModifiedRegs[BaseReg] || UsedRegs[BaseReg])
1524  return E;
1525  } while (MBBI != B && Count < Limit);
1526  return E;
1527 }
1528 
1529 bool AArch64LoadStoreOpt::tryToPromoteLoadFromStore(
1531  MachineInstr &MI = *MBBI;
1532  // If this is a volatile load, don't mess with it.
1533  if (MI.hasOrderedMemoryRef())
1534  return false;
1535 
1536  // Make sure this is a reg+imm.
1537  // FIXME: It is possible to extend it to handle reg+reg cases.
1538  if (!getLdStOffsetOp(MI).isImm())
1539  return false;
1540 
1541  // Look backward up to LdStLimit instructions.
1543  if (findMatchingStore(MBBI, LdStLimit, StoreI)) {
1544  ++NumLoadsFromStoresPromoted;
1545  // Promote the load. Keeping the iterator straight is a
1546  // pain, so we let the merge routine tell us what the next instruction
1547  // is after it's done mucking about.
1548  MBBI = promoteLoadFromStore(MBBI, StoreI);
1549  return true;
1550  }
1551  return false;
1552 }
1553 
1554 // Merge adjacent zero stores into a wider store.
1555 bool AArch64LoadStoreOpt::tryToMergeZeroStInst(
1557  assert(isPromotableZeroStoreInst(*MBBI) && "Expected narrow store.");
1558  MachineInstr &MI = *MBBI;
1560 
1561  if (!TII->isCandidateToMergeOrPair(MI))
1562  return false;
1563 
1564  // Look ahead up to LdStLimit instructions for a mergable instruction.
1565  LdStPairFlags Flags;
1566  MachineBasicBlock::iterator MergeMI =
1567  findMatchingInsn(MBBI, Flags, LdStLimit, /* FindNarrowMerge = */ true);
1568  if (MergeMI != E) {
1569  ++NumZeroStoresPromoted;
1570 
1571  // Keeping the iterator straight is a pain, so we let the merge routine tell
1572  // us what the next instruction is after it's done mucking about.
1573  MBBI = mergeNarrowZeroStores(MBBI, MergeMI, Flags);
1574  return true;
1575  }
1576  return false;
1577 }
1578 
1579 // Find loads and stores that can be merged into a single load or store pair
1580 // instruction.
1581 bool AArch64LoadStoreOpt::tryToPairLdStInst(MachineBasicBlock::iterator &MBBI) {
1582  MachineInstr &MI = *MBBI;
1584 
1585  if (!TII->isCandidateToMergeOrPair(MI))
1586  return false;
1587 
1588  // Early exit if the offset is not possible to match. (6 bits of positive
1589  // range, plus allow an extra one in case we find a later insn that matches
1590  // with Offset-1)
1591  bool IsUnscaled = TII->isUnscaledLdSt(MI);
1592  int Offset = getLdStOffsetOp(MI).getImm();
1593  int OffsetStride = IsUnscaled ? getMemScale(MI) : 1;
1594  // Allow one more for offset.
1595  if (Offset > 0)
1596  Offset -= OffsetStride;
1597  if (!inBoundsForPair(IsUnscaled, Offset, OffsetStride))
1598  return false;
1599 
1600  // Look ahead up to LdStLimit instructions for a pairable instruction.
1601  LdStPairFlags Flags;
1603  findMatchingInsn(MBBI, Flags, LdStLimit, /* FindNarrowMerge = */ false);
1604  if (Paired != E) {
1605  ++NumPairCreated;
1606  if (TII->isUnscaledLdSt(MI))
1607  ++NumUnscaledPairCreated;
1608  // Keeping the iterator straight is a pain, so we let the merge routine tell
1609  // us what the next instruction is after it's done mucking about.
1610  MBBI = mergePairedInsns(MBBI, Paired, Flags);
1611  return true;
1612  }
1613  return false;
1614 }
1615 
1616 bool AArch64LoadStoreOpt::tryToMergeLdStUpdate
1617  (MachineBasicBlock::iterator &MBBI) {
1618  MachineInstr &MI = *MBBI;
1621 
1622  // Look forward to try to form a post-index instruction. For example,
1623  // ldr x0, [x20]
1624  // add x20, x20, #32
1625  // merged into:
1626  // ldr x0, [x20], #32
1627  Update = findMatchingUpdateInsnForward(MBBI, 0, UpdateLimit);
1628  if (Update != E) {
1629  // Merge the update into the ld/st.
1630  MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/false);
1631  return true;
1632  }
1633 
1634  // Don't know how to handle unscaled pre/post-index versions below, so bail.
1635  if (TII->isUnscaledLdSt(MI.getOpcode()))
1636  return false;
1637 
1638  // Look back to try to find a pre-index instruction. For example,
1639  // add x0, x0, #8
1640  // ldr x1, [x0]
1641  // merged into:
1642  // ldr x1, [x0, #8]!
1643  Update = findMatchingUpdateInsnBackward(MBBI, UpdateLimit);
1644  if (Update != E) {
1645  // Merge the update into the ld/st.
1646  MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true);
1647  return true;
1648  }
1649 
1650  // The immediate in the load/store is scaled by the size of the memory
1651  // operation. The immediate in the add we're looking for,
1652  // however, is not, so adjust here.
1653  int UnscaledOffset = getLdStOffsetOp(MI).getImm() * getMemScale(MI);
1654 
1655  // Look forward to try to find a post-index instruction. For example,
1656  // ldr x1, [x0, #64]
1657  // add x0, x0, #64
1658  // merged into:
1659  // ldr x1, [x0, #64]!
1660  Update = findMatchingUpdateInsnForward(MBBI, UnscaledOffset, UpdateLimit);
1661  if (Update != E) {
1662  // Merge the update into the ld/st.
1663  MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true);
1664  return true;
1665  }
1666 
1667  return false;
1668 }
1669 
1670 bool AArch64LoadStoreOpt::optimizeBlock(MachineBasicBlock &MBB,
1671  bool EnableNarrowZeroStOpt) {
1672  bool Modified = false;
1673  // Four tranformations to do here:
1674  // 1) Find loads that directly read from stores and promote them by
1675  // replacing with mov instructions. If the store is wider than the load,
1676  // the load will be replaced with a bitfield extract.
1677  // e.g.,
1678  // str w1, [x0, #4]
1679  // ldrh w2, [x0, #6]
1680  // ; becomes
1681  // str w1, [x0, #4]
1682  // lsr w2, w1, #16
1683  for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1684  MBBI != E;) {
1685  if (isPromotableLoadFromStore(*MBBI) && tryToPromoteLoadFromStore(MBBI))
1686  Modified = true;
1687  else
1688  ++MBBI;
1689  }
1690  // 2) Merge adjacent zero stores into a wider store.
1691  // e.g.,
1692  // strh wzr, [x0]
1693  // strh wzr, [x0, #2]
1694  // ; becomes
1695  // str wzr, [x0]
1696  // e.g.,
1697  // str wzr, [x0]
1698  // str wzr, [x0, #4]
1699  // ; becomes
1700  // str xzr, [x0]
1701  if (EnableNarrowZeroStOpt)
1702  for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1703  MBBI != E;) {
1704  if (isPromotableZeroStoreInst(*MBBI) && tryToMergeZeroStInst(MBBI))
1705  Modified = true;
1706  else
1707  ++MBBI;
1708  }
1709  // 3) Find loads and stores that can be merged into a single load or store
1710  // pair instruction.
1711  // e.g.,
1712  // ldr x0, [x2]
1713  // ldr x1, [x2, #8]
1714  // ; becomes
1715  // ldp x0, x1, [x2]
1716  for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1717  MBBI != E;) {
1718  if (TII->isPairableLdStInst(*MBBI) && tryToPairLdStInst(MBBI))
1719  Modified = true;
1720  else
1721  ++MBBI;
1722  }
1723  // 4) Find base register updates that can be merged into the load or store
1724  // as a base-reg writeback.
1725  // e.g.,
1726  // ldr x0, [x2]
1727  // add x2, x2, #4
1728  // ; becomes
1729  // ldr x0, [x2], #4
1730  for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
1731  MBBI != E;) {
1732  if (isMergeableLdStUpdate(*MBBI) && tryToMergeLdStUpdate(MBBI))
1733  Modified = true;
1734  else
1735  ++MBBI;
1736  }
1737 
1738  return Modified;
1739 }
1740 
1741 bool AArch64LoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
1742  if (skipFunction(Fn.getFunction()))
1743  return false;
1744 
1745  Subtarget = &static_cast<const AArch64Subtarget &>(Fn.getSubtarget());
1746  TII = static_cast<const AArch64InstrInfo *>(Subtarget->getInstrInfo());
1747  TRI = Subtarget->getRegisterInfo();
1748  AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1749 
1750  // Resize the modified and used register bitfield trackers. We do this once
1751  // per function and then clear the bitfield each time we optimize a load or
1752  // store.
1753  ModifiedRegs.resize(TRI->getNumRegs());
1754  UsedRegs.resize(TRI->getNumRegs());
1755 
1756  bool Modified = false;
1757  bool enableNarrowZeroStOpt = !Subtarget->requiresStrictAlign();
1758  for (auto &MBB : Fn)
1759  Modified |= optimizeBlock(MBB, enableNarrowZeroStOpt);
1760 
1761  return Modified;
1762 }
1763 
1764 // FIXME: Do we need/want a pre-alloc pass like ARM has to try to keep loads and
1765 // stores near one another? Note: The pre-RA instruction scheduler already has
1766 // hooks to try and schedule pairable loads/stores together to improve pairing
1767 // opportunities. Thus, pre-RA pairing pass may not be worth the effort.
1768 
1769 // FIXME: When pairing store instructions it's very possible for this pass to
1770 // hoist a store with a KILL marker above another use (without a KILL marker).
1771 // The resulting IR is invalid, but nothing uses the KILL markers after this
1772 // pass, so it's never caused a problem in practice.
1773 
1774 /// createAArch64LoadStoreOptimizationPass - returns an instance of the
1775 /// load / store optimization pass.
1777  return new AArch64LoadStoreOpt();
1778 }
void push_back(const T &Elt)
Definition: SmallVector.h:212
const MachineInstrBuilder & add(const MachineOperand &MO) const
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
bool isCall(QueryType Type=AnyInBundle) const
Definition: MachineInstr.h:461
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
static bool mayAlias(MachineInstr &MIa, MachineInstr &MIb, AliasAnalysis *AA)
void initializeAArch64LoadStoreOptPass(PassRegistry &)
This provides a very simple, boring adaptor for a begin and end iterator into a range type...
unsigned getReg() const
getReg - Returns the register number.
bool mayLoadOrStore(QueryType Type=AnyInBundle) const
Return true if this instruction could possibly read or modify memory.
Definition: MachineInstr.h:652
static bool inBoundsForPair(bool IsUnscaled, int Offset, int OffsetStride)
bool isTransient() const
Return true if this is a transient instruction that is either very likely to be eliminated during reg...
Definition: MachineInstr.h:905
STATISTIC(NumFunctions, "Total number of functions")
A debug info location.
Definition: DebugLoc.h:34
uint64_t alignTo(uint64_t Value, uint64_t Align, uint64_t Skew=0)
Returns the next integer (mod 2**64) that is greater than or equal to Value and is a multiple of Alig...
Definition: MathExtras.h:677
An instruction for reading from memory.
Definition: Instructions.h:164
static bool isLdOffsetInRangeOfSt(MachineInstr &LoadInst, MachineInstr &StoreInst, const AArch64InstrInfo *TII)
bool isImm() const
isImm - Tests if this is a MO_Immediate operand.
static bool isLdStPairSuppressed(const MachineInstr &MI)
Return true if pairing the given load or store is hinted to be unprofitable.
static bool isUnscaledLdSt(unsigned Opc)
Return true if this is an unscaled load/store.
static bool isPromotableLoadFromStore(MachineInstr &MI)
AnalysisUsage & addRequired()
MachineFunctionPass - This class adapts the FunctionPass interface to allow convenient creation of pa...
static unsigned getMatchingPairOpcode(unsigned Opc)
const HexagonInstrInfo * TII
void setImplicit(bool Val=true)
Reg
All possible values of the reg field in the ModR/M byte.
unsigned getOpcode() const
Returns the opcode of this MachineInstr.
Definition: MachineInstr.h:293
static bool isPairedLdSt(const MachineInstr &MI)
static unsigned isMatchingStore(MachineInstr &LoadInst, MachineInstr &StoreInst)
An instruction for storing to memory.
Definition: Instructions.h:306
FunctionPass * createAArch64LoadStoreOptimizationPass()
createAArch64LoadStoreOptimizationPass - returns an instance of the load / store optimization pass...
bool mayAlias(AliasAnalysis *AA, MachineInstr &Other, bool UseTBAA)
Returns true if this instruction&#39;s memory access aliases the memory access of Other.
static unsigned getShifterImm(AArch64_AM::ShiftExtendType ST, unsigned Imm)
getShifterImm - Encode the shift type and amount: imm: 6-bit shift amount shifter: 000 ==> lsl 001 ==...
static unsigned getShiftValue(unsigned Imm)
getShiftValue - Extract the shift value.
static int getMemScale(MachineInstr &MI)
MachineInstrBuilder BuildMI(MachineFunction &MF, const DebugLoc &DL, const MCInstrDesc &MCID)
Builder interface. Specify how to create the initial instruction itself.
bool mayStore(QueryType Type=AnyInBundle) const
Return true if this instruction could possibly modify memory.
Definition: MachineInstr.h:642
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:406
#define AARCH64_LOAD_STORE_OPT_NAME
const AArch64RegisterInfo & getRegisterInfo() const
getRegisterInfo - TargetInstrInfo is a superset of MRegister info.
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static unsigned getMatchingNonSExtOpcode(unsigned Opc, bool *IsValidLdStrOpc=nullptr)
const TargetSubtargetInfo & getSubtarget() const
getSubtarget - Return the subtarget for which this machine code is being compiled.
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - Subclasses that override getAnalysisUsage must call this.
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
Represent the analysis usage information of a pass.
static unsigned getMatchingWideOpcode(unsigned Opc)
static cl::opt< unsigned > LdStLimit("aarch64-load-store-scan-limit", cl::init(20), cl::Hidden)
static bool isMergeableLdStUpdate(MachineInstr &MI)
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
static void print(raw_ostream &Out, object::Archive::Kind Kind, T Val)
self_iterator getIterator()
Definition: ilist_node.h:82
TargetRegisterInfo base class - We assume that the target defines a static array of TargetRegisterDes...
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
void setIsKill(bool Val=true)
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
static const MachineOperand & getLdStBaseOp(const MachineInstr &MI)
MachineOperand class - Representation of each machine instruction operand.
static bool isPromotableZeroStoreInst(MachineInstr &MI)
bool hasOrderedMemoryRef() const
Return true if this instruction may have an ordered or volatile memory reference, or if the informati...
int64_t getImm() const
const Function & getFunction() const
Return the LLVM function that this machine code represents.
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
static unsigned getReg(const void *D, unsigned RC, unsigned RegNo)
const MachineBasicBlock * getParent() const
Definition: MachineInstr.h:142
MachineFunctionProperties & set(Property P)
Representation of each machine instruction.
Definition: MachineInstr.h:60
const MachineInstrBuilder & addImm(int64_t Val) const
Add a new immediate operand.
const MachineInstrBuilder & setMIFlags(unsigned Flags) const
void setReg(unsigned Reg)
Change the register this operand corresponds to.
#define I(x, y, z)
Definition: MD5.cpp:58
static const MachineOperand & getLdStRegOp(const MachineInstr &MI, unsigned PairedRegOp=0)
const MachineInstrBuilder & addReg(unsigned RegNo, unsigned flags=0, unsigned SubReg=0) const
Add a new virtual register operand.
static unsigned getPreIndexedOpcode(unsigned Opc)
static const MachineOperand & getLdStOffsetOp(const MachineInstr &MI)
bool mayLoad(QueryType Type=AnyInBundle) const
Return true if this instruction could possibly read memory.
Definition: MachineInstr.h:629
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
static unsigned getPostIndexedOpcode(unsigned Opc)
LLVM Value Representation.
Definition: Value.h:73
static bool areCandidatesToMergeOrPair(MachineInstr &FirstMI, MachineInstr &MI, LdStPairFlags &Flags, const AArch64InstrInfo *TII)
#define DEBUG(X)
Definition: Debug.h:118
IRTranslator LLVM IR MI
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
INITIALIZE_PASS(AArch64LoadStoreOpt, "aarch64-ldst-opt", AARCH64_LOAD_STORE_OPT_NAME, false, false) static bool isNarrowStore(unsigned Opc)
const MachineOperand & getOperand(unsigned i) const
Definition: MachineInstr.h:298
static cl::opt< unsigned > UpdateLimit("aarch64-update-scan-limit", cl::init(100), cl::Hidden)
Properties which a MachineFunction may have at a given point in time.