xref: /freebsd/contrib/llvm-project/llvm/lib/Target/ARM/ARMLoadStoreOptimizer.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- ARMLoadStoreOptimizer.cpp - ARM load / store opt. pass -------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file This file contains a pass that performs load / store related peephole
10 /// optimizations. This pass should be run after register allocation.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "ARM.h"
15 #include "ARMBaseInstrInfo.h"
16 #include "ARMBaseRegisterInfo.h"
17 #include "ARMISelLowering.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMSubtarget.h"
20 #include "MCTargetDesc/ARMAddressingModes.h"
21 #include "MCTargetDesc/ARMBaseInfo.h"
22 #include "Utils/ARMBaseInfo.h"
23 #include "llvm/ADT/ArrayRef.h"
24 #include "llvm/ADT/DenseMap.h"
25 #include "llvm/ADT/DenseSet.h"
26 #include "llvm/ADT/STLExtras.h"
27 #include "llvm/ADT/SetVector.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/SmallSet.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/ADT/iterator_range.h"
33 #include "llvm/Analysis/AliasAnalysis.h"
34 #include "llvm/CodeGen/LiveRegUnits.h"
35 #include "llvm/CodeGen/MachineBasicBlock.h"
36 #include "llvm/CodeGen/MachineDominators.h"
37 #include "llvm/CodeGen/MachineFrameInfo.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineFunctionPass.h"
40 #include "llvm/CodeGen/MachineInstr.h"
41 #include "llvm/CodeGen/MachineInstrBuilder.h"
42 #include "llvm/CodeGen/MachineMemOperand.h"
43 #include "llvm/CodeGen/MachineOperand.h"
44 #include "llvm/CodeGen/MachineRegisterInfo.h"
45 #include "llvm/CodeGen/RegisterClassInfo.h"
46 #include "llvm/CodeGen/TargetFrameLowering.h"
47 #include "llvm/CodeGen/TargetInstrInfo.h"
48 #include "llvm/CodeGen/TargetLowering.h"
49 #include "llvm/CodeGen/TargetRegisterInfo.h"
50 #include "llvm/CodeGen/TargetSubtargetInfo.h"
51 #include "llvm/IR/DataLayout.h"
52 #include "llvm/IR/DebugLoc.h"
53 #include "llvm/IR/DerivedTypes.h"
54 #include "llvm/IR/Function.h"
55 #include "llvm/IR/Type.h"
56 #include "llvm/InitializePasses.h"
57 #include "llvm/MC/MCInstrDesc.h"
58 #include "llvm/Pass.h"
59 #include "llvm/Support/Allocator.h"
60 #include "llvm/Support/CommandLine.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/ErrorHandling.h"
63 #include "llvm/Support/raw_ostream.h"
64 #include <algorithm>
65 #include <cassert>
66 #include <cstddef>
67 #include <cstdlib>
68 #include <iterator>
69 #include <limits>
70 #include <utility>
71 
72 using namespace llvm;
73 
74 #define DEBUG_TYPE "arm-ldst-opt"
75 
76 STATISTIC(NumLDMGened , "Number of ldm instructions generated");
77 STATISTIC(NumSTMGened , "Number of stm instructions generated");
78 STATISTIC(NumVLDMGened, "Number of vldm instructions generated");
79 STATISTIC(NumVSTMGened, "Number of vstm instructions generated");
80 STATISTIC(NumLdStMoved, "Number of load / store instructions moved");
81 STATISTIC(NumLDRDFormed,"Number of ldrd created before allocation");
82 STATISTIC(NumSTRDFormed,"Number of strd created before allocation");
83 STATISTIC(NumLDRD2LDM,  "Number of ldrd instructions turned back into ldm");
84 STATISTIC(NumSTRD2STM,  "Number of strd instructions turned back into stm");
85 STATISTIC(NumLDRD2LDR,  "Number of ldrd instructions turned back into ldr's");
86 STATISTIC(NumSTRD2STR,  "Number of strd instructions turned back into str's");
87 
88 /// This switch disables formation of double/multi instructions that could
89 /// potentially lead to (new) alignment traps even with CCR.UNALIGN_TRP
90 /// disabled. This can be used to create libraries that are robust even when
91 /// users provoke undefined behaviour by supplying misaligned pointers.
92 /// \see mayCombineMisaligned()
93 static cl::opt<bool>
94 AssumeMisalignedLoadStores("arm-assume-misaligned-load-store", cl::Hidden,
95     cl::init(false), cl::desc("Be more conservative in ARM load/store opt"));
96 
97 #define ARM_LOAD_STORE_OPT_NAME "ARM load / store optimization pass"
98 
99 namespace {
100 
101   /// Post- register allocation pass the combine load / store instructions to
102   /// form ldm / stm instructions.
103   struct ARMLoadStoreOpt : public MachineFunctionPass {
104     static char ID;
105 
106     const MachineFunction *MF;
107     const TargetInstrInfo *TII;
108     const TargetRegisterInfo *TRI;
109     const ARMSubtarget *STI;
110     const TargetLowering *TL;
111     ARMFunctionInfo *AFI;
112     LiveRegUnits LiveRegs;
113     RegisterClassInfo RegClassInfo;
114     MachineBasicBlock::const_iterator LiveRegPos;
115     bool LiveRegsValid;
116     bool RegClassInfoValid;
117     bool isThumb1, isThumb2;
118 
ARMLoadStoreOpt__anonf2a56f050111::ARMLoadStoreOpt119     ARMLoadStoreOpt() : MachineFunctionPass(ID) {}
120 
121     bool runOnMachineFunction(MachineFunction &Fn) override;
122 
getRequiredProperties__anonf2a56f050111::ARMLoadStoreOpt123     MachineFunctionProperties getRequiredProperties() const override {
124       return MachineFunctionProperties().set(
125           MachineFunctionProperties::Property::NoVRegs);
126     }
127 
getPassName__anonf2a56f050111::ARMLoadStoreOpt128     StringRef getPassName() const override { return ARM_LOAD_STORE_OPT_NAME; }
129 
130   private:
131     /// A set of load/store MachineInstrs with same base register sorted by
132     /// offset.
133     struct MemOpQueueEntry {
134       MachineInstr *MI;
135       int Offset;        ///< Load/Store offset.
136       unsigned Position; ///< Position as counted from end of basic block.
137 
MemOpQueueEntry__anonf2a56f050111::ARMLoadStoreOpt::MemOpQueueEntry138       MemOpQueueEntry(MachineInstr &MI, int Offset, unsigned Position)
139           : MI(&MI), Offset(Offset), Position(Position) {}
140     };
141     using MemOpQueue = SmallVector<MemOpQueueEntry, 8>;
142 
143     /// A set of MachineInstrs that fulfill (nearly all) conditions to get
144     /// merged into a LDM/STM.
145     struct MergeCandidate {
146       /// List of instructions ordered by load/store offset.
147       SmallVector<MachineInstr*, 4> Instrs;
148 
149       /// Index in Instrs of the instruction being latest in the schedule.
150       unsigned LatestMIIdx;
151 
152       /// Index in Instrs of the instruction being earliest in the schedule.
153       unsigned EarliestMIIdx;
154 
155       /// Index into the basic block where the merged instruction will be
156       /// inserted. (See MemOpQueueEntry.Position)
157       unsigned InsertPos;
158 
159       /// Whether the instructions can be merged into a ldm/stm instruction.
160       bool CanMergeToLSMulti;
161 
162       /// Whether the instructions can be merged into a ldrd/strd instruction.
163       bool CanMergeToLSDouble;
164     };
165     SpecificBumpPtrAllocator<MergeCandidate> Allocator;
166     SmallVector<const MergeCandidate*,4> Candidates;
167     SmallVector<MachineInstr*,4> MergeBaseCandidates;
168 
169     void moveLiveRegsBefore(const MachineBasicBlock &MBB,
170                             MachineBasicBlock::const_iterator Before);
171     unsigned findFreeReg(const TargetRegisterClass &RegClass);
172     void UpdateBaseRegUses(MachineBasicBlock &MBB,
173                            MachineBasicBlock::iterator MBBI, const DebugLoc &DL,
174                            unsigned Base, unsigned WordOffset,
175                            ARMCC::CondCodes Pred, unsigned PredReg);
176     MachineInstr *CreateLoadStoreMulti(
177         MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore,
178         int Offset, unsigned Base, bool BaseKill, unsigned Opcode,
179         ARMCC::CondCodes Pred, unsigned PredReg, const DebugLoc &DL,
180         ArrayRef<std::pair<unsigned, bool>> Regs,
181         ArrayRef<MachineInstr*> Instrs);
182     MachineInstr *CreateLoadStoreDouble(
183         MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore,
184         int Offset, unsigned Base, bool BaseKill, unsigned Opcode,
185         ARMCC::CondCodes Pred, unsigned PredReg, const DebugLoc &DL,
186         ArrayRef<std::pair<unsigned, bool>> Regs,
187         ArrayRef<MachineInstr*> Instrs) const;
188     void FormCandidates(const MemOpQueue &MemOps);
189     MachineInstr *MergeOpsUpdate(const MergeCandidate &Cand);
190     bool FixInvalidRegPairOp(MachineBasicBlock &MBB,
191                              MachineBasicBlock::iterator &MBBI);
192     bool MergeBaseUpdateLoadStore(MachineInstr *MI);
193     bool MergeBaseUpdateLSMultiple(MachineInstr *MI);
194     bool MergeBaseUpdateLSDouble(MachineInstr &MI) const;
195     bool LoadStoreMultipleOpti(MachineBasicBlock &MBB);
196     bool MergeReturnIntoLDM(MachineBasicBlock &MBB);
197     bool CombineMovBx(MachineBasicBlock &MBB);
198   };
199 
200 } // end anonymous namespace
201 
202 char ARMLoadStoreOpt::ID = 0;
203 
204 INITIALIZE_PASS(ARMLoadStoreOpt, "arm-ldst-opt", ARM_LOAD_STORE_OPT_NAME, false,
205                 false)
206 
definesCPSR(const MachineInstr & MI)207 static bool definesCPSR(const MachineInstr &MI) {
208   for (const auto &MO : MI.operands()) {
209     if (!MO.isReg())
210       continue;
211     if (MO.isDef() && MO.getReg() == ARM::CPSR && !MO.isDead())
212       // If the instruction has live CPSR def, then it's not safe to fold it
213       // into load / store.
214       return true;
215   }
216 
217   return false;
218 }
219 
getMemoryOpOffset(const MachineInstr & MI)220 static int getMemoryOpOffset(const MachineInstr &MI) {
221   unsigned Opcode = MI.getOpcode();
222   bool isAM3 = Opcode == ARM::LDRD || Opcode == ARM::STRD;
223   unsigned NumOperands = MI.getDesc().getNumOperands();
224   unsigned OffField = MI.getOperand(NumOperands - 3).getImm();
225 
226   if (Opcode == ARM::t2LDRi12 || Opcode == ARM::t2LDRi8 ||
227       Opcode == ARM::t2STRi12 || Opcode == ARM::t2STRi8 ||
228       Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8 ||
229       Opcode == ARM::LDRi12   || Opcode == ARM::STRi12)
230     return OffField;
231 
232   // Thumb1 immediate offsets are scaled by 4
233   if (Opcode == ARM::tLDRi || Opcode == ARM::tSTRi ||
234       Opcode == ARM::tLDRspi || Opcode == ARM::tSTRspi)
235     return OffField * 4;
236 
237   int Offset = isAM3 ? ARM_AM::getAM3Offset(OffField)
238     : ARM_AM::getAM5Offset(OffField) * 4;
239   ARM_AM::AddrOpc Op = isAM3 ? ARM_AM::getAM3Op(OffField)
240     : ARM_AM::getAM5Op(OffField);
241 
242   if (Op == ARM_AM::sub)
243     return -Offset;
244 
245   return Offset;
246 }
247 
getLoadStoreBaseOp(const MachineInstr & MI)248 static const MachineOperand &getLoadStoreBaseOp(const MachineInstr &MI) {
249   return MI.getOperand(1);
250 }
251 
getLoadStoreRegOp(const MachineInstr & MI)252 static const MachineOperand &getLoadStoreRegOp(const MachineInstr &MI) {
253   return MI.getOperand(0);
254 }
255 
getLoadStoreMultipleOpcode(unsigned Opcode,ARM_AM::AMSubMode Mode)256 static int getLoadStoreMultipleOpcode(unsigned Opcode, ARM_AM::AMSubMode Mode) {
257   switch (Opcode) {
258   default: llvm_unreachable("Unhandled opcode!");
259   case ARM::LDRi12:
260     ++NumLDMGened;
261     switch (Mode) {
262     default: llvm_unreachable("Unhandled submode!");
263     case ARM_AM::ia: return ARM::LDMIA;
264     case ARM_AM::da: return ARM::LDMDA;
265     case ARM_AM::db: return ARM::LDMDB;
266     case ARM_AM::ib: return ARM::LDMIB;
267     }
268   case ARM::STRi12:
269     ++NumSTMGened;
270     switch (Mode) {
271     default: llvm_unreachable("Unhandled submode!");
272     case ARM_AM::ia: return ARM::STMIA;
273     case ARM_AM::da: return ARM::STMDA;
274     case ARM_AM::db: return ARM::STMDB;
275     case ARM_AM::ib: return ARM::STMIB;
276     }
277   case ARM::tLDRi:
278   case ARM::tLDRspi:
279     // tLDMIA is writeback-only - unless the base register is in the input
280     // reglist.
281     ++NumLDMGened;
282     switch (Mode) {
283     default: llvm_unreachable("Unhandled submode!");
284     case ARM_AM::ia: return ARM::tLDMIA;
285     }
286   case ARM::tSTRi:
287   case ARM::tSTRspi:
288     // There is no non-writeback tSTMIA either.
289     ++NumSTMGened;
290     switch (Mode) {
291     default: llvm_unreachable("Unhandled submode!");
292     case ARM_AM::ia: return ARM::tSTMIA_UPD;
293     }
294   case ARM::t2LDRi8:
295   case ARM::t2LDRi12:
296     ++NumLDMGened;
297     switch (Mode) {
298     default: llvm_unreachable("Unhandled submode!");
299     case ARM_AM::ia: return ARM::t2LDMIA;
300     case ARM_AM::db: return ARM::t2LDMDB;
301     }
302   case ARM::t2STRi8:
303   case ARM::t2STRi12:
304     ++NumSTMGened;
305     switch (Mode) {
306     default: llvm_unreachable("Unhandled submode!");
307     case ARM_AM::ia: return ARM::t2STMIA;
308     case ARM_AM::db: return ARM::t2STMDB;
309     }
310   case ARM::VLDRS:
311     ++NumVLDMGened;
312     switch (Mode) {
313     default: llvm_unreachable("Unhandled submode!");
314     case ARM_AM::ia: return ARM::VLDMSIA;
315     case ARM_AM::db: return 0; // Only VLDMSDB_UPD exists.
316     }
317   case ARM::VSTRS:
318     ++NumVSTMGened;
319     switch (Mode) {
320     default: llvm_unreachable("Unhandled submode!");
321     case ARM_AM::ia: return ARM::VSTMSIA;
322     case ARM_AM::db: return 0; // Only VSTMSDB_UPD exists.
323     }
324   case ARM::VLDRD:
325     ++NumVLDMGened;
326     switch (Mode) {
327     default: llvm_unreachable("Unhandled submode!");
328     case ARM_AM::ia: return ARM::VLDMDIA;
329     case ARM_AM::db: return 0; // Only VLDMDDB_UPD exists.
330     }
331   case ARM::VSTRD:
332     ++NumVSTMGened;
333     switch (Mode) {
334     default: llvm_unreachable("Unhandled submode!");
335     case ARM_AM::ia: return ARM::VSTMDIA;
336     case ARM_AM::db: return 0; // Only VSTMDDB_UPD exists.
337     }
338   }
339 }
340 
getLoadStoreMultipleSubMode(unsigned Opcode)341 static ARM_AM::AMSubMode getLoadStoreMultipleSubMode(unsigned Opcode) {
342   switch (Opcode) {
343   default: llvm_unreachable("Unhandled opcode!");
344   case ARM::LDMIA_RET:
345   case ARM::LDMIA:
346   case ARM::LDMIA_UPD:
347   case ARM::STMIA:
348   case ARM::STMIA_UPD:
349   case ARM::tLDMIA:
350   case ARM::tLDMIA_UPD:
351   case ARM::tSTMIA_UPD:
352   case ARM::t2LDMIA_RET:
353   case ARM::t2LDMIA:
354   case ARM::t2LDMIA_UPD:
355   case ARM::t2STMIA:
356   case ARM::t2STMIA_UPD:
357   case ARM::VLDMSIA:
358   case ARM::VLDMSIA_UPD:
359   case ARM::VSTMSIA:
360   case ARM::VSTMSIA_UPD:
361   case ARM::VLDMDIA:
362   case ARM::VLDMDIA_UPD:
363   case ARM::VSTMDIA:
364   case ARM::VSTMDIA_UPD:
365     return ARM_AM::ia;
366 
367   case ARM::LDMDA:
368   case ARM::LDMDA_UPD:
369   case ARM::STMDA:
370   case ARM::STMDA_UPD:
371     return ARM_AM::da;
372 
373   case ARM::LDMDB:
374   case ARM::LDMDB_UPD:
375   case ARM::STMDB:
376   case ARM::STMDB_UPD:
377   case ARM::t2LDMDB:
378   case ARM::t2LDMDB_UPD:
379   case ARM::t2STMDB:
380   case ARM::t2STMDB_UPD:
381   case ARM::VLDMSDB_UPD:
382   case ARM::VSTMSDB_UPD:
383   case ARM::VLDMDDB_UPD:
384   case ARM::VSTMDDB_UPD:
385     return ARM_AM::db;
386 
387   case ARM::LDMIB:
388   case ARM::LDMIB_UPD:
389   case ARM::STMIB:
390   case ARM::STMIB_UPD:
391     return ARM_AM::ib;
392   }
393 }
394 
isT1i32Load(unsigned Opc)395 static bool isT1i32Load(unsigned Opc) {
396   return Opc == ARM::tLDRi || Opc == ARM::tLDRspi;
397 }
398 
isT2i32Load(unsigned Opc)399 static bool isT2i32Load(unsigned Opc) {
400   return Opc == ARM::t2LDRi12 || Opc == ARM::t2LDRi8;
401 }
402 
isi32Load(unsigned Opc)403 static bool isi32Load(unsigned Opc) {
404   return Opc == ARM::LDRi12 || isT1i32Load(Opc) || isT2i32Load(Opc) ;
405 }
406 
isT1i32Store(unsigned Opc)407 static bool isT1i32Store(unsigned Opc) {
408   return Opc == ARM::tSTRi || Opc == ARM::tSTRspi;
409 }
410 
isT2i32Store(unsigned Opc)411 static bool isT2i32Store(unsigned Opc) {
412   return Opc == ARM::t2STRi12 || Opc == ARM::t2STRi8;
413 }
414 
isi32Store(unsigned Opc)415 static bool isi32Store(unsigned Opc) {
416   return Opc == ARM::STRi12 || isT1i32Store(Opc) || isT2i32Store(Opc);
417 }
418 
isLoadSingle(unsigned Opc)419 static bool isLoadSingle(unsigned Opc) {
420   return isi32Load(Opc) || Opc == ARM::VLDRS || Opc == ARM::VLDRD;
421 }
422 
getImmScale(unsigned Opc)423 static unsigned getImmScale(unsigned Opc) {
424   switch (Opc) {
425   default: llvm_unreachable("Unhandled opcode!");
426   case ARM::tLDRi:
427   case ARM::tSTRi:
428   case ARM::tLDRspi:
429   case ARM::tSTRspi:
430     return 1;
431   case ARM::tLDRHi:
432   case ARM::tSTRHi:
433     return 2;
434   case ARM::tLDRBi:
435   case ARM::tSTRBi:
436     return 4;
437   }
438 }
439 
getLSMultipleTransferSize(const MachineInstr * MI)440 static unsigned getLSMultipleTransferSize(const MachineInstr *MI) {
441   switch (MI->getOpcode()) {
442   default: return 0;
443   case ARM::LDRi12:
444   case ARM::STRi12:
445   case ARM::tLDRi:
446   case ARM::tSTRi:
447   case ARM::tLDRspi:
448   case ARM::tSTRspi:
449   case ARM::t2LDRi8:
450   case ARM::t2LDRi12:
451   case ARM::t2STRi8:
452   case ARM::t2STRi12:
453   case ARM::VLDRS:
454   case ARM::VSTRS:
455     return 4;
456   case ARM::VLDRD:
457   case ARM::VSTRD:
458     return 8;
459   case ARM::LDMIA:
460   case ARM::LDMDA:
461   case ARM::LDMDB:
462   case ARM::LDMIB:
463   case ARM::STMIA:
464   case ARM::STMDA:
465   case ARM::STMDB:
466   case ARM::STMIB:
467   case ARM::tLDMIA:
468   case ARM::tLDMIA_UPD:
469   case ARM::tSTMIA_UPD:
470   case ARM::t2LDMIA:
471   case ARM::t2LDMDB:
472   case ARM::t2STMIA:
473   case ARM::t2STMDB:
474   case ARM::VLDMSIA:
475   case ARM::VSTMSIA:
476     return (MI->getNumOperands() - MI->getDesc().getNumOperands() + 1) * 4;
477   case ARM::VLDMDIA:
478   case ARM::VSTMDIA:
479     return (MI->getNumOperands() - MI->getDesc().getNumOperands() + 1) * 8;
480   }
481 }
482 
483 /// Update future uses of the base register with the offset introduced
484 /// due to writeback. This function only works on Thumb1.
UpdateBaseRegUses(MachineBasicBlock & MBB,MachineBasicBlock::iterator MBBI,const DebugLoc & DL,unsigned Base,unsigned WordOffset,ARMCC::CondCodes Pred,unsigned PredReg)485 void ARMLoadStoreOpt::UpdateBaseRegUses(MachineBasicBlock &MBB,
486                                         MachineBasicBlock::iterator MBBI,
487                                         const DebugLoc &DL, unsigned Base,
488                                         unsigned WordOffset,
489                                         ARMCC::CondCodes Pred,
490                                         unsigned PredReg) {
491   assert(isThumb1 && "Can only update base register uses for Thumb1!");
492   // Start updating any instructions with immediate offsets. Insert a SUB before
493   // the first non-updateable instruction (if any).
494   for (; MBBI != MBB.end(); ++MBBI) {
495     bool InsertSub = false;
496     unsigned Opc = MBBI->getOpcode();
497 
498     if (MBBI->readsRegister(Base, /*TRI=*/nullptr)) {
499       int Offset;
500       bool IsLoad =
501         Opc == ARM::tLDRi || Opc == ARM::tLDRHi || Opc == ARM::tLDRBi;
502       bool IsStore =
503         Opc == ARM::tSTRi || Opc == ARM::tSTRHi || Opc == ARM::tSTRBi;
504 
505       if (IsLoad || IsStore) {
506         // Loads and stores with immediate offsets can be updated, but only if
507         // the new offset isn't negative.
508         // The MachineOperand containing the offset immediate is the last one
509         // before predicates.
510         MachineOperand &MO =
511           MBBI->getOperand(MBBI->getDesc().getNumOperands() - 3);
512         // The offsets are scaled by 1, 2 or 4 depending on the Opcode.
513         Offset = MO.getImm() - WordOffset * getImmScale(Opc);
514 
515         // If storing the base register, it needs to be reset first.
516         Register InstrSrcReg = getLoadStoreRegOp(*MBBI).getReg();
517 
518         if (Offset >= 0 && !(IsStore && InstrSrcReg == Base))
519           MO.setImm(Offset);
520         else
521           InsertSub = true;
522       } else if ((Opc == ARM::tSUBi8 || Opc == ARM::tADDi8) &&
523                  !definesCPSR(*MBBI)) {
524         // SUBS/ADDS using this register, with a dead def of the CPSR.
525         // Merge it with the update; if the merged offset is too large,
526         // insert a new sub instead.
527         MachineOperand &MO =
528           MBBI->getOperand(MBBI->getDesc().getNumOperands() - 3);
529         Offset = (Opc == ARM::tSUBi8) ?
530           MO.getImm() + WordOffset * 4 :
531           MO.getImm() - WordOffset * 4 ;
532         if (Offset >= 0 && TL->isLegalAddImmediate(Offset)) {
533           // FIXME: Swap ADDS<->SUBS if Offset < 0, erase instruction if
534           // Offset == 0.
535           MO.setImm(Offset);
536           // The base register has now been reset, so exit early.
537           return;
538         } else {
539           InsertSub = true;
540         }
541       } else {
542         // Can't update the instruction.
543         InsertSub = true;
544       }
545     } else if (definesCPSR(*MBBI) || MBBI->isCall() || MBBI->isBranch()) {
546       // Since SUBS sets the condition flags, we can't place the base reset
547       // after an instruction that has a live CPSR def.
548       // The base register might also contain an argument for a function call.
549       InsertSub = true;
550     }
551 
552     if (InsertSub) {
553       // An instruction above couldn't be updated, so insert a sub.
554       BuildMI(MBB, MBBI, DL, TII->get(ARM::tSUBi8), Base)
555           .add(t1CondCodeOp(true))
556           .addReg(Base)
557           .addImm(WordOffset * 4)
558           .addImm(Pred)
559           .addReg(PredReg);
560       return;
561     }
562 
563     if (MBBI->killsRegister(Base, /*TRI=*/nullptr) ||
564         MBBI->definesRegister(Base, /*TRI=*/nullptr))
565       // Register got killed. Stop updating.
566       return;
567   }
568 
569   // End of block was reached.
570   if (!MBB.succ_empty()) {
571     // FIXME: Because of a bug, live registers are sometimes missing from
572     // the successor blocks' live-in sets. This means we can't trust that
573     // information and *always* have to reset at the end of a block.
574     // See PR21029.
575     if (MBBI != MBB.end()) --MBBI;
576     BuildMI(MBB, MBBI, DL, TII->get(ARM::tSUBi8), Base)
577         .add(t1CondCodeOp(true))
578         .addReg(Base)
579         .addImm(WordOffset * 4)
580         .addImm(Pred)
581         .addReg(PredReg);
582   }
583 }
584 
585 /// Return the first register of class \p RegClass that is not in \p Regs.
findFreeReg(const TargetRegisterClass & RegClass)586 unsigned ARMLoadStoreOpt::findFreeReg(const TargetRegisterClass &RegClass) {
587   if (!RegClassInfoValid) {
588     RegClassInfo.runOnMachineFunction(*MF);
589     RegClassInfoValid = true;
590   }
591 
592   for (unsigned Reg : RegClassInfo.getOrder(&RegClass))
593     if (LiveRegs.available(Reg) && !MF->getRegInfo().isReserved(Reg))
594       return Reg;
595   return 0;
596 }
597 
598 /// Compute live registers just before instruction \p Before (in normal schedule
599 /// direction). Computes backwards so multiple queries in the same block must
600 /// come in reverse order.
moveLiveRegsBefore(const MachineBasicBlock & MBB,MachineBasicBlock::const_iterator Before)601 void ARMLoadStoreOpt::moveLiveRegsBefore(const MachineBasicBlock &MBB,
602     MachineBasicBlock::const_iterator Before) {
603   // Initialize if we never queried in this block.
604   if (!LiveRegsValid) {
605     LiveRegs.init(*TRI);
606     LiveRegs.addLiveOuts(MBB);
607     LiveRegPos = MBB.end();
608     LiveRegsValid = true;
609   }
610   // Move backward just before the "Before" position.
611   while (LiveRegPos != Before) {
612     --LiveRegPos;
613     LiveRegs.stepBackward(*LiveRegPos);
614   }
615 }
616 
ContainsReg(const ArrayRef<std::pair<unsigned,bool>> & Regs,unsigned Reg)617 static bool ContainsReg(const ArrayRef<std::pair<unsigned, bool>> &Regs,
618                         unsigned Reg) {
619   for (const std::pair<unsigned, bool> &R : Regs)
620     if (R.first == Reg)
621       return true;
622   return false;
623 }
624 
625 /// Create and insert a LDM or STM with Base as base register and registers in
626 /// Regs as the register operands that would be loaded / stored.  It returns
627 /// true if the transformation is done.
CreateLoadStoreMulti(MachineBasicBlock & MBB,MachineBasicBlock::iterator InsertBefore,int Offset,unsigned Base,bool BaseKill,unsigned Opcode,ARMCC::CondCodes Pred,unsigned PredReg,const DebugLoc & DL,ArrayRef<std::pair<unsigned,bool>> Regs,ArrayRef<MachineInstr * > Instrs)628 MachineInstr *ARMLoadStoreOpt::CreateLoadStoreMulti(
629     MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore,
630     int Offset, unsigned Base, bool BaseKill, unsigned Opcode,
631     ARMCC::CondCodes Pred, unsigned PredReg, const DebugLoc &DL,
632     ArrayRef<std::pair<unsigned, bool>> Regs,
633     ArrayRef<MachineInstr*> Instrs) {
634   unsigned NumRegs = Regs.size();
635   assert(NumRegs > 1);
636 
637   // For Thumb1 targets, it might be necessary to clobber the CPSR to merge.
638   // Compute liveness information for that register to make the decision.
639   bool SafeToClobberCPSR = !isThumb1 ||
640     (MBB.computeRegisterLiveness(TRI, ARM::CPSR, InsertBefore, 20) ==
641      MachineBasicBlock::LQR_Dead);
642 
643   bool Writeback = isThumb1; // Thumb1 LDM/STM have base reg writeback.
644 
645   // Exception: If the base register is in the input reglist, Thumb1 LDM is
646   // non-writeback.
647   // It's also not possible to merge an STR of the base register in Thumb1.
648   if (isThumb1 && ContainsReg(Regs, Base)) {
649     assert(Base != ARM::SP && "Thumb1 does not allow SP in register list");
650     if (Opcode == ARM::tLDRi)
651       Writeback = false;
652     else if (Opcode == ARM::tSTRi)
653       return nullptr;
654   }
655 
656   ARM_AM::AMSubMode Mode = ARM_AM::ia;
657   // VFP and Thumb2 do not support IB or DA modes. Thumb1 only supports IA.
658   bool isNotVFP = isi32Load(Opcode) || isi32Store(Opcode);
659   bool haveIBAndDA = isNotVFP && !isThumb2 && !isThumb1;
660 
661   if (Offset == 4 && haveIBAndDA) {
662     Mode = ARM_AM::ib;
663   } else if (Offset == -4 * (int)NumRegs + 4 && haveIBAndDA) {
664     Mode = ARM_AM::da;
665   } else if (Offset == -4 * (int)NumRegs && isNotVFP && !isThumb1) {
666     // VLDM/VSTM do not support DB mode without also updating the base reg.
667     Mode = ARM_AM::db;
668   } else if (Offset != 0 || Opcode == ARM::tLDRspi || Opcode == ARM::tSTRspi) {
669     // Check if this is a supported opcode before inserting instructions to
670     // calculate a new base register.
671     if (!getLoadStoreMultipleOpcode(Opcode, Mode)) return nullptr;
672 
673     // If starting offset isn't zero, insert a MI to materialize a new base.
674     // But only do so if it is cost effective, i.e. merging more than two
675     // loads / stores.
676     if (NumRegs <= 2)
677       return nullptr;
678 
679     // On Thumb1, it's not worth materializing a new base register without
680     // clobbering the CPSR (i.e. not using ADDS/SUBS).
681     if (!SafeToClobberCPSR)
682       return nullptr;
683 
684     unsigned NewBase;
685     if (isi32Load(Opcode)) {
686       // If it is a load, then just use one of the destination registers
687       // as the new base. Will no longer be writeback in Thumb1.
688       NewBase = Regs[NumRegs-1].first;
689       Writeback = false;
690     } else {
691       // Find a free register that we can use as scratch register.
692       moveLiveRegsBefore(MBB, InsertBefore);
693       // The merged instruction does not exist yet but will use several Regs if
694       // it is a Store.
695       if (!isLoadSingle(Opcode))
696         for (const std::pair<unsigned, bool> &R : Regs)
697           LiveRegs.addReg(R.first);
698 
699       NewBase = findFreeReg(isThumb1 ? ARM::tGPRRegClass : ARM::GPRRegClass);
700       if (NewBase == 0)
701         return nullptr;
702     }
703 
704     int BaseOpc = isThumb2 ? (BaseKill && Base == ARM::SP ? ARM::t2ADDspImm
705                                                           : ARM::t2ADDri)
706                            : (isThumb1 && Base == ARM::SP)
707                                  ? ARM::tADDrSPi
708                                  : (isThumb1 && Offset < 8)
709                                        ? ARM::tADDi3
710                                        : isThumb1 ? ARM::tADDi8 : ARM::ADDri;
711 
712     if (Offset < 0) {
713       // FIXME: There are no Thumb1 load/store instructions with negative
714       // offsets. So the Base != ARM::SP might be unnecessary.
715       Offset = -Offset;
716       BaseOpc = isThumb2 ? (BaseKill && Base == ARM::SP ? ARM::t2SUBspImm
717                                                         : ARM::t2SUBri)
718                          : (isThumb1 && Offset < 8 && Base != ARM::SP)
719                                ? ARM::tSUBi3
720                                : isThumb1 ? ARM::tSUBi8 : ARM::SUBri;
721     }
722 
723     if (!TL->isLegalAddImmediate(Offset))
724       // FIXME: Try add with register operand?
725       return nullptr; // Probably not worth it then.
726 
727     // We can only append a kill flag to the add/sub input if the value is not
728     // used in the register list of the stm as well.
729     bool KillOldBase = BaseKill &&
730       (!isi32Store(Opcode) || !ContainsReg(Regs, Base));
731 
732     if (isThumb1) {
733       // Thumb1: depending on immediate size, use either
734       //   ADDS NewBase, Base, #imm3
735       // or
736       //   MOV  NewBase, Base
737       //   ADDS NewBase, #imm8.
738       if (Base != NewBase &&
739           (BaseOpc == ARM::tADDi8 || BaseOpc == ARM::tSUBi8)) {
740         // Need to insert a MOV to the new base first.
741         if (isARMLowRegister(NewBase) && isARMLowRegister(Base) &&
742             !STI->hasV6Ops()) {
743           // thumbv4t doesn't have lo->lo copies, and we can't predicate tMOVSr
744           if (Pred != ARMCC::AL)
745             return nullptr;
746           BuildMI(MBB, InsertBefore, DL, TII->get(ARM::tMOVSr), NewBase)
747             .addReg(Base, getKillRegState(KillOldBase));
748         } else
749           BuildMI(MBB, InsertBefore, DL, TII->get(ARM::tMOVr), NewBase)
750               .addReg(Base, getKillRegState(KillOldBase))
751               .add(predOps(Pred, PredReg));
752 
753         // The following ADDS/SUBS becomes an update.
754         Base = NewBase;
755         KillOldBase = true;
756       }
757       if (BaseOpc == ARM::tADDrSPi) {
758         assert(Offset % 4 == 0 && "tADDrSPi offset is scaled by 4");
759         BuildMI(MBB, InsertBefore, DL, TII->get(BaseOpc), NewBase)
760             .addReg(Base, getKillRegState(KillOldBase))
761             .addImm(Offset / 4)
762             .add(predOps(Pred, PredReg));
763       } else
764         BuildMI(MBB, InsertBefore, DL, TII->get(BaseOpc), NewBase)
765             .add(t1CondCodeOp(true))
766             .addReg(Base, getKillRegState(KillOldBase))
767             .addImm(Offset)
768             .add(predOps(Pred, PredReg));
769     } else {
770       BuildMI(MBB, InsertBefore, DL, TII->get(BaseOpc), NewBase)
771           .addReg(Base, getKillRegState(KillOldBase))
772           .addImm(Offset)
773           .add(predOps(Pred, PredReg))
774           .add(condCodeOp());
775     }
776     Base = NewBase;
777     BaseKill = true; // New base is always killed straight away.
778   }
779 
780   bool isDef = isLoadSingle(Opcode);
781 
782   // Get LS multiple opcode. Note that for Thumb1 this might be an opcode with
783   // base register writeback.
784   Opcode = getLoadStoreMultipleOpcode(Opcode, Mode);
785   if (!Opcode)
786     return nullptr;
787 
788   // Check if a Thumb1 LDM/STM merge is safe. This is the case if:
789   // - There is no writeback (LDM of base register),
790   // - the base register is killed by the merged instruction,
791   // - or it's safe to overwrite the condition flags, i.e. to insert a SUBS
792   //   to reset the base register.
793   // Otherwise, don't merge.
794   // It's safe to return here since the code to materialize a new base register
795   // above is also conditional on SafeToClobberCPSR.
796   if (isThumb1 && !SafeToClobberCPSR && Writeback && !BaseKill)
797     return nullptr;
798 
799   MachineInstrBuilder MIB;
800 
801   if (Writeback) {
802     assert(isThumb1 && "expected Writeback only inThumb1");
803     if (Opcode == ARM::tLDMIA) {
804       assert(!(ContainsReg(Regs, Base)) && "Thumb1 can't LDM ! with Base in Regs");
805       // Update tLDMIA with writeback if necessary.
806       Opcode = ARM::tLDMIA_UPD;
807     }
808 
809     MIB = BuildMI(MBB, InsertBefore, DL, TII->get(Opcode));
810 
811     // Thumb1: we might need to set base writeback when building the MI.
812     MIB.addReg(Base, getDefRegState(true))
813        .addReg(Base, getKillRegState(BaseKill));
814 
815     // The base isn't dead after a merged instruction with writeback.
816     // Insert a sub instruction after the newly formed instruction to reset.
817     if (!BaseKill)
818       UpdateBaseRegUses(MBB, InsertBefore, DL, Base, NumRegs, Pred, PredReg);
819   } else {
820     // No writeback, simply build the MachineInstr.
821     MIB = BuildMI(MBB, InsertBefore, DL, TII->get(Opcode));
822     MIB.addReg(Base, getKillRegState(BaseKill));
823   }
824 
825   MIB.addImm(Pred).addReg(PredReg);
826 
827   for (const std::pair<unsigned, bool> &R : Regs)
828     MIB.addReg(R.first, getDefRegState(isDef) | getKillRegState(R.second));
829 
830   MIB.cloneMergedMemRefs(Instrs);
831 
832   return MIB.getInstr();
833 }
834 
CreateLoadStoreDouble(MachineBasicBlock & MBB,MachineBasicBlock::iterator InsertBefore,int Offset,unsigned Base,bool BaseKill,unsigned Opcode,ARMCC::CondCodes Pred,unsigned PredReg,const DebugLoc & DL,ArrayRef<std::pair<unsigned,bool>> Regs,ArrayRef<MachineInstr * > Instrs) const835 MachineInstr *ARMLoadStoreOpt::CreateLoadStoreDouble(
836     MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore,
837     int Offset, unsigned Base, bool BaseKill, unsigned Opcode,
838     ARMCC::CondCodes Pred, unsigned PredReg, const DebugLoc &DL,
839     ArrayRef<std::pair<unsigned, bool>> Regs,
840     ArrayRef<MachineInstr*> Instrs) const {
841   bool IsLoad = isi32Load(Opcode);
842   assert((IsLoad || isi32Store(Opcode)) && "Must have integer load or store");
843   unsigned LoadStoreOpcode = IsLoad ? ARM::t2LDRDi8 : ARM::t2STRDi8;
844 
845   assert(Regs.size() == 2);
846   MachineInstrBuilder MIB = BuildMI(MBB, InsertBefore, DL,
847                                     TII->get(LoadStoreOpcode));
848   if (IsLoad) {
849     MIB.addReg(Regs[0].first, RegState::Define)
850        .addReg(Regs[1].first, RegState::Define);
851   } else {
852     MIB.addReg(Regs[0].first, getKillRegState(Regs[0].second))
853        .addReg(Regs[1].first, getKillRegState(Regs[1].second));
854   }
855   MIB.addReg(Base).addImm(Offset).addImm(Pred).addReg(PredReg);
856   MIB.cloneMergedMemRefs(Instrs);
857   return MIB.getInstr();
858 }
859 
860 /// Call MergeOps and update MemOps and merges accordingly on success.
MergeOpsUpdate(const MergeCandidate & Cand)861 MachineInstr *ARMLoadStoreOpt::MergeOpsUpdate(const MergeCandidate &Cand) {
862   const MachineInstr *First = Cand.Instrs.front();
863   unsigned Opcode = First->getOpcode();
864   bool IsLoad = isLoadSingle(Opcode);
865   SmallVector<std::pair<unsigned, bool>, 8> Regs;
866   SmallVector<unsigned, 4> ImpDefs;
867   DenseSet<unsigned> KilledRegs;
868   DenseSet<unsigned> UsedRegs;
869   // Determine list of registers and list of implicit super-register defs.
870   for (const MachineInstr *MI : Cand.Instrs) {
871     const MachineOperand &MO = getLoadStoreRegOp(*MI);
872     Register Reg = MO.getReg();
873     bool IsKill = MO.isKill();
874     if (IsKill)
875       KilledRegs.insert(Reg);
876     Regs.push_back(std::make_pair(Reg, IsKill));
877     UsedRegs.insert(Reg);
878 
879     if (IsLoad) {
880       // Collect any implicit defs of super-registers, after merging we can't
881       // be sure anymore that we properly preserved these live ranges and must
882       // removed these implicit operands.
883       for (const MachineOperand &MO : MI->implicit_operands()) {
884         if (!MO.isReg() || !MO.isDef() || MO.isDead())
885           continue;
886         assert(MO.isImplicit());
887         Register DefReg = MO.getReg();
888 
889         if (is_contained(ImpDefs, DefReg))
890           continue;
891         // We can ignore cases where the super-reg is read and written.
892         if (MI->readsRegister(DefReg, /*TRI=*/nullptr))
893           continue;
894         ImpDefs.push_back(DefReg);
895       }
896     }
897   }
898 
899   // Attempt the merge.
900   using iterator = MachineBasicBlock::iterator;
901 
902   MachineInstr *LatestMI = Cand.Instrs[Cand.LatestMIIdx];
903   iterator InsertBefore = std::next(iterator(LatestMI));
904   MachineBasicBlock &MBB = *LatestMI->getParent();
905   unsigned Offset = getMemoryOpOffset(*First);
906   Register Base = getLoadStoreBaseOp(*First).getReg();
907   bool BaseKill = LatestMI->killsRegister(Base, /*TRI=*/nullptr);
908   Register PredReg;
909   ARMCC::CondCodes Pred = getInstrPredicate(*First, PredReg);
910   DebugLoc DL = First->getDebugLoc();
911   MachineInstr *Merged = nullptr;
912   if (Cand.CanMergeToLSDouble)
913     Merged = CreateLoadStoreDouble(MBB, InsertBefore, Offset, Base, BaseKill,
914                                    Opcode, Pred, PredReg, DL, Regs,
915                                    Cand.Instrs);
916   if (!Merged && Cand.CanMergeToLSMulti)
917     Merged = CreateLoadStoreMulti(MBB, InsertBefore, Offset, Base, BaseKill,
918                                   Opcode, Pred, PredReg, DL, Regs, Cand.Instrs);
919   if (!Merged)
920     return nullptr;
921 
922   // Determine earliest instruction that will get removed. We then keep an
923   // iterator just above it so the following erases don't invalidated it.
924   iterator EarliestI(Cand.Instrs[Cand.EarliestMIIdx]);
925   bool EarliestAtBegin = false;
926   if (EarliestI == MBB.begin()) {
927     EarliestAtBegin = true;
928   } else {
929     EarliestI = std::prev(EarliestI);
930   }
931 
932   // Remove instructions which have been merged.
933   for (MachineInstr *MI : Cand.Instrs)
934     MBB.erase(MI);
935 
936   // Determine range between the earliest removed instruction and the new one.
937   if (EarliestAtBegin)
938     EarliestI = MBB.begin();
939   else
940     EarliestI = std::next(EarliestI);
941   auto FixupRange = make_range(EarliestI, iterator(Merged));
942 
943   if (isLoadSingle(Opcode)) {
944     // If the previous loads defined a super-reg, then we have to mark earlier
945     // operands undef; Replicate the super-reg def on the merged instruction.
946     for (MachineInstr &MI : FixupRange) {
947       for (unsigned &ImpDefReg : ImpDefs) {
948         for (MachineOperand &MO : MI.implicit_operands()) {
949           if (!MO.isReg() || MO.getReg() != ImpDefReg)
950             continue;
951           if (MO.readsReg())
952             MO.setIsUndef();
953           else if (MO.isDef())
954             ImpDefReg = 0;
955         }
956       }
957     }
958 
959     MachineInstrBuilder MIB(*Merged->getParent()->getParent(), Merged);
960     for (unsigned ImpDef : ImpDefs)
961       MIB.addReg(ImpDef, RegState::ImplicitDefine);
962   } else {
963     // Remove kill flags: We are possibly storing the values later now.
964     assert(isi32Store(Opcode) || Opcode == ARM::VSTRS || Opcode == ARM::VSTRD);
965     for (MachineInstr &MI : FixupRange) {
966       for (MachineOperand &MO : MI.uses()) {
967         if (!MO.isReg() || !MO.isKill())
968           continue;
969         if (UsedRegs.count(MO.getReg()))
970           MO.setIsKill(false);
971       }
972     }
973     assert(ImpDefs.empty());
974   }
975 
976   return Merged;
977 }
978 
isValidLSDoubleOffset(int Offset)979 static bool isValidLSDoubleOffset(int Offset) {
980   unsigned Value = abs(Offset);
981   // t2LDRDi8/t2STRDi8 supports an 8 bit immediate which is internally
982   // multiplied by 4.
983   return (Value % 4) == 0 && Value < 1024;
984 }
985 
986 /// Return true for loads/stores that can be combined to a double/multi
987 /// operation without increasing the requirements for alignment.
mayCombineMisaligned(const TargetSubtargetInfo & STI,const MachineInstr & MI)988 static bool mayCombineMisaligned(const TargetSubtargetInfo &STI,
989                                  const MachineInstr &MI) {
990   // vldr/vstr trap on misaligned pointers anyway, forming vldm makes no
991   // difference.
992   unsigned Opcode = MI.getOpcode();
993   if (!isi32Load(Opcode) && !isi32Store(Opcode))
994     return true;
995 
996   // Stack pointer alignment is out of the programmers control so we can trust
997   // SP-relative loads/stores.
998   if (getLoadStoreBaseOp(MI).getReg() == ARM::SP &&
999       STI.getFrameLowering()->getTransientStackAlign() >= Align(4))
1000     return true;
1001   return false;
1002 }
1003 
1004 /// Find candidates for load/store multiple merge in list of MemOpQueueEntries.
FormCandidates(const MemOpQueue & MemOps)1005 void ARMLoadStoreOpt::FormCandidates(const MemOpQueue &MemOps) {
1006   const MachineInstr *FirstMI = MemOps[0].MI;
1007   unsigned Opcode = FirstMI->getOpcode();
1008   bool isNotVFP = isi32Load(Opcode) || isi32Store(Opcode);
1009   unsigned Size = getLSMultipleTransferSize(FirstMI);
1010 
1011   unsigned SIndex = 0;
1012   unsigned EIndex = MemOps.size();
1013   do {
1014     // Look at the first instruction.
1015     const MachineInstr *MI = MemOps[SIndex].MI;
1016     int Offset = MemOps[SIndex].Offset;
1017     const MachineOperand &PMO = getLoadStoreRegOp(*MI);
1018     Register PReg = PMO.getReg();
1019     unsigned PRegNum = PMO.isUndef() ? std::numeric_limits<unsigned>::max()
1020                                      : TRI->getEncodingValue(PReg);
1021     unsigned Latest = SIndex;
1022     unsigned Earliest = SIndex;
1023     unsigned Count = 1;
1024     bool CanMergeToLSDouble =
1025       STI->isThumb2() && isNotVFP && isValidLSDoubleOffset(Offset);
1026     // ARM errata 602117: LDRD with base in list may result in incorrect base
1027     // register when interrupted or faulted.
1028     if (STI->isCortexM3() && isi32Load(Opcode) &&
1029         PReg == getLoadStoreBaseOp(*MI).getReg())
1030       CanMergeToLSDouble = false;
1031 
1032     bool CanMergeToLSMulti = true;
1033     // On swift vldm/vstm starting with an odd register number as that needs
1034     // more uops than single vldrs.
1035     if (STI->hasSlowOddRegister() && !isNotVFP && (PRegNum % 2) == 1)
1036       CanMergeToLSMulti = false;
1037 
1038     // LDRD/STRD do not allow SP/PC. LDM/STM do not support it or have it
1039     // deprecated; LDM to PC is fine but cannot happen here.
1040     if (PReg == ARM::SP || PReg == ARM::PC)
1041       CanMergeToLSMulti = CanMergeToLSDouble = false;
1042 
1043     // Should we be conservative?
1044     if (AssumeMisalignedLoadStores && !mayCombineMisaligned(*STI, *MI))
1045       CanMergeToLSMulti = CanMergeToLSDouble = false;
1046 
1047     // vldm / vstm limit are 32 for S variants, 16 for D variants.
1048     unsigned Limit;
1049     switch (Opcode) {
1050     default:
1051       Limit = UINT_MAX;
1052       break;
1053     case ARM::VLDRD:
1054     case ARM::VSTRD:
1055       Limit = 16;
1056       break;
1057     }
1058 
1059     // Merge following instructions where possible.
1060     for (unsigned I = SIndex+1; I < EIndex; ++I, ++Count) {
1061       int NewOffset = MemOps[I].Offset;
1062       if (NewOffset != Offset + (int)Size)
1063         break;
1064       const MachineOperand &MO = getLoadStoreRegOp(*MemOps[I].MI);
1065       Register Reg = MO.getReg();
1066       if (Reg == ARM::SP || Reg == ARM::PC)
1067         break;
1068       if (Count == Limit)
1069         break;
1070 
1071       // See if the current load/store may be part of a multi load/store.
1072       unsigned RegNum = MO.isUndef() ? std::numeric_limits<unsigned>::max()
1073                                      : TRI->getEncodingValue(Reg);
1074       bool PartOfLSMulti = CanMergeToLSMulti;
1075       if (PartOfLSMulti) {
1076         // Register numbers must be in ascending order.
1077         if (RegNum <= PRegNum)
1078           PartOfLSMulti = false;
1079         // For VFP / NEON load/store multiples, the registers must be
1080         // consecutive and within the limit on the number of registers per
1081         // instruction.
1082         else if (!isNotVFP && RegNum != PRegNum+1)
1083           PartOfLSMulti = false;
1084       }
1085       // See if the current load/store may be part of a double load/store.
1086       bool PartOfLSDouble = CanMergeToLSDouble && Count <= 1;
1087 
1088       if (!PartOfLSMulti && !PartOfLSDouble)
1089         break;
1090       CanMergeToLSMulti &= PartOfLSMulti;
1091       CanMergeToLSDouble &= PartOfLSDouble;
1092       // Track MemOp with latest and earliest position (Positions are
1093       // counted in reverse).
1094       unsigned Position = MemOps[I].Position;
1095       if (Position < MemOps[Latest].Position)
1096         Latest = I;
1097       else if (Position > MemOps[Earliest].Position)
1098         Earliest = I;
1099       // Prepare for next MemOp.
1100       Offset += Size;
1101       PRegNum = RegNum;
1102     }
1103 
1104     // Form a candidate from the Ops collected so far.
1105     MergeCandidate *Candidate = new(Allocator.Allocate()) MergeCandidate;
1106     for (unsigned C = SIndex, CE = SIndex + Count; C < CE; ++C)
1107       Candidate->Instrs.push_back(MemOps[C].MI);
1108     Candidate->LatestMIIdx = Latest - SIndex;
1109     Candidate->EarliestMIIdx = Earliest - SIndex;
1110     Candidate->InsertPos = MemOps[Latest].Position;
1111     if (Count == 1)
1112       CanMergeToLSMulti = CanMergeToLSDouble = false;
1113     Candidate->CanMergeToLSMulti = CanMergeToLSMulti;
1114     Candidate->CanMergeToLSDouble = CanMergeToLSDouble;
1115     Candidates.push_back(Candidate);
1116     // Continue after the chain.
1117     SIndex += Count;
1118   } while (SIndex < EIndex);
1119 }
1120 
getUpdatingLSMultipleOpcode(unsigned Opc,ARM_AM::AMSubMode Mode)1121 static unsigned getUpdatingLSMultipleOpcode(unsigned Opc,
1122                                             ARM_AM::AMSubMode Mode) {
1123   switch (Opc) {
1124   default: llvm_unreachable("Unhandled opcode!");
1125   case ARM::LDMIA:
1126   case ARM::LDMDA:
1127   case ARM::LDMDB:
1128   case ARM::LDMIB:
1129     switch (Mode) {
1130     default: llvm_unreachable("Unhandled submode!");
1131     case ARM_AM::ia: return ARM::LDMIA_UPD;
1132     case ARM_AM::ib: return ARM::LDMIB_UPD;
1133     case ARM_AM::da: return ARM::LDMDA_UPD;
1134     case ARM_AM::db: return ARM::LDMDB_UPD;
1135     }
1136   case ARM::STMIA:
1137   case ARM::STMDA:
1138   case ARM::STMDB:
1139   case ARM::STMIB:
1140     switch (Mode) {
1141     default: llvm_unreachable("Unhandled submode!");
1142     case ARM_AM::ia: return ARM::STMIA_UPD;
1143     case ARM_AM::ib: return ARM::STMIB_UPD;
1144     case ARM_AM::da: return ARM::STMDA_UPD;
1145     case ARM_AM::db: return ARM::STMDB_UPD;
1146     }
1147   case ARM::t2LDMIA:
1148   case ARM::t2LDMDB:
1149     switch (Mode) {
1150     default: llvm_unreachable("Unhandled submode!");
1151     case ARM_AM::ia: return ARM::t2LDMIA_UPD;
1152     case ARM_AM::db: return ARM::t2LDMDB_UPD;
1153     }
1154   case ARM::t2STMIA:
1155   case ARM::t2STMDB:
1156     switch (Mode) {
1157     default: llvm_unreachable("Unhandled submode!");
1158     case ARM_AM::ia: return ARM::t2STMIA_UPD;
1159     case ARM_AM::db: return ARM::t2STMDB_UPD;
1160     }
1161   case ARM::VLDMSIA:
1162     switch (Mode) {
1163     default: llvm_unreachable("Unhandled submode!");
1164     case ARM_AM::ia: return ARM::VLDMSIA_UPD;
1165     case ARM_AM::db: return ARM::VLDMSDB_UPD;
1166     }
1167   case ARM::VLDMDIA:
1168     switch (Mode) {
1169     default: llvm_unreachable("Unhandled submode!");
1170     case ARM_AM::ia: return ARM::VLDMDIA_UPD;
1171     case ARM_AM::db: return ARM::VLDMDDB_UPD;
1172     }
1173   case ARM::VSTMSIA:
1174     switch (Mode) {
1175     default: llvm_unreachable("Unhandled submode!");
1176     case ARM_AM::ia: return ARM::VSTMSIA_UPD;
1177     case ARM_AM::db: return ARM::VSTMSDB_UPD;
1178     }
1179   case ARM::VSTMDIA:
1180     switch (Mode) {
1181     default: llvm_unreachable("Unhandled submode!");
1182     case ARM_AM::ia: return ARM::VSTMDIA_UPD;
1183     case ARM_AM::db: return ARM::VSTMDDB_UPD;
1184     }
1185   }
1186 }
1187 
1188 /// Check if the given instruction increments or decrements a register and
1189 /// return the amount it is incremented/decremented. Returns 0 if the CPSR flags
1190 /// generated by the instruction are possibly read as well.
isIncrementOrDecrement(const MachineInstr & MI,Register Reg,ARMCC::CondCodes Pred,Register PredReg)1191 static int isIncrementOrDecrement(const MachineInstr &MI, Register Reg,
1192                                   ARMCC::CondCodes Pred, Register PredReg) {
1193   bool CheckCPSRDef;
1194   int Scale;
1195   switch (MI.getOpcode()) {
1196   case ARM::tADDi8:  Scale =  4; CheckCPSRDef = true; break;
1197   case ARM::tSUBi8:  Scale = -4; CheckCPSRDef = true; break;
1198   case ARM::t2SUBri:
1199   case ARM::t2SUBspImm:
1200   case ARM::SUBri:   Scale = -1; CheckCPSRDef = true; break;
1201   case ARM::t2ADDri:
1202   case ARM::t2ADDspImm:
1203   case ARM::ADDri:   Scale =  1; CheckCPSRDef = true; break;
1204   case ARM::tADDspi: Scale =  4; CheckCPSRDef = false; break;
1205   case ARM::tSUBspi: Scale = -4; CheckCPSRDef = false; break;
1206   default: return 0;
1207   }
1208 
1209   Register MIPredReg;
1210   if (MI.getOperand(0).getReg() != Reg ||
1211       MI.getOperand(1).getReg() != Reg ||
1212       getInstrPredicate(MI, MIPredReg) != Pred ||
1213       MIPredReg != PredReg)
1214     return 0;
1215 
1216   if (CheckCPSRDef && definesCPSR(MI))
1217     return 0;
1218   return MI.getOperand(2).getImm() * Scale;
1219 }
1220 
1221 /// Searches for an increment or decrement of \p Reg before \p MBBI.
1222 static MachineBasicBlock::iterator
findIncDecBefore(MachineBasicBlock::iterator MBBI,Register Reg,ARMCC::CondCodes Pred,Register PredReg,int & Offset)1223 findIncDecBefore(MachineBasicBlock::iterator MBBI, Register Reg,
1224                  ARMCC::CondCodes Pred, Register PredReg, int &Offset) {
1225   Offset = 0;
1226   MachineBasicBlock &MBB = *MBBI->getParent();
1227   MachineBasicBlock::iterator BeginMBBI = MBB.begin();
1228   MachineBasicBlock::iterator EndMBBI = MBB.end();
1229   if (MBBI == BeginMBBI)
1230     return EndMBBI;
1231 
1232   // Skip debug values.
1233   MachineBasicBlock::iterator PrevMBBI = std::prev(MBBI);
1234   while (PrevMBBI->isDebugInstr() && PrevMBBI != BeginMBBI)
1235     --PrevMBBI;
1236 
1237   Offset = isIncrementOrDecrement(*PrevMBBI, Reg, Pred, PredReg);
1238   return Offset == 0 ? EndMBBI : PrevMBBI;
1239 }
1240 
1241 /// Searches for a increment or decrement of \p Reg after \p MBBI.
1242 static MachineBasicBlock::iterator
findIncDecAfter(MachineBasicBlock::iterator MBBI,Register Reg,ARMCC::CondCodes Pred,Register PredReg,int & Offset,const TargetRegisterInfo * TRI)1243 findIncDecAfter(MachineBasicBlock::iterator MBBI, Register Reg,
1244                 ARMCC::CondCodes Pred, Register PredReg, int &Offset,
1245                 const TargetRegisterInfo *TRI) {
1246   Offset = 0;
1247   MachineBasicBlock &MBB = *MBBI->getParent();
1248   MachineBasicBlock::iterator EndMBBI = MBB.end();
1249   MachineBasicBlock::iterator NextMBBI = std::next(MBBI);
1250   while (NextMBBI != EndMBBI) {
1251     // Skip debug values.
1252     while (NextMBBI != EndMBBI && NextMBBI->isDebugInstr())
1253       ++NextMBBI;
1254     if (NextMBBI == EndMBBI)
1255       return EndMBBI;
1256 
1257     unsigned Off = isIncrementOrDecrement(*NextMBBI, Reg, Pred, PredReg);
1258     if (Off) {
1259       Offset = Off;
1260       return NextMBBI;
1261     }
1262 
1263     // SP can only be combined if it is the next instruction after the original
1264     // MBBI, otherwise we may be incrementing the stack pointer (invalidating
1265     // anything below the new pointer) when its frame elements are still in
1266     // use. Other registers can attempt to look further, until a different use
1267     // or def of the register is found.
1268     if (Reg == ARM::SP || NextMBBI->readsRegister(Reg, TRI) ||
1269         NextMBBI->definesRegister(Reg, TRI))
1270       return EndMBBI;
1271 
1272     ++NextMBBI;
1273   }
1274   return EndMBBI;
1275 }
1276 
1277 /// Fold proceeding/trailing inc/dec of base register into the
1278 /// LDM/STM/VLDM{D|S}/VSTM{D|S} op when possible:
1279 ///
1280 /// stmia rn, <ra, rb, rc>
1281 /// rn := rn + 4 * 3;
1282 /// =>
1283 /// stmia rn!, <ra, rb, rc>
1284 ///
1285 /// rn := rn - 4 * 3;
1286 /// ldmia rn, <ra, rb, rc>
1287 /// =>
1288 /// ldmdb rn!, <ra, rb, rc>
MergeBaseUpdateLSMultiple(MachineInstr * MI)1289 bool ARMLoadStoreOpt::MergeBaseUpdateLSMultiple(MachineInstr *MI) {
1290   // Thumb1 is already using updating loads/stores.
1291   if (isThumb1) return false;
1292   LLVM_DEBUG(dbgs() << "Attempting to merge update of: " << *MI);
1293 
1294   const MachineOperand &BaseOP = MI->getOperand(0);
1295   Register Base = BaseOP.getReg();
1296   bool BaseKill = BaseOP.isKill();
1297   Register PredReg;
1298   ARMCC::CondCodes Pred = getInstrPredicate(*MI, PredReg);
1299   unsigned Opcode = MI->getOpcode();
1300   DebugLoc DL = MI->getDebugLoc();
1301 
1302   // Can't use an updating ld/st if the base register is also a dest
1303   // register. e.g. ldmdb r0!, {r0, r1, r2}. The behavior is undefined.
1304   for (const MachineOperand &MO : llvm::drop_begin(MI->operands(), 2))
1305     if (MO.getReg() == Base)
1306       return false;
1307 
1308   int Bytes = getLSMultipleTransferSize(MI);
1309   MachineBasicBlock &MBB = *MI->getParent();
1310   MachineBasicBlock::iterator MBBI(MI);
1311   int Offset;
1312   MachineBasicBlock::iterator MergeInstr
1313     = findIncDecBefore(MBBI, Base, Pred, PredReg, Offset);
1314   ARM_AM::AMSubMode Mode = getLoadStoreMultipleSubMode(Opcode);
1315   if (Mode == ARM_AM::ia && Offset == -Bytes) {
1316     Mode = ARM_AM::db;
1317   } else if (Mode == ARM_AM::ib && Offset == -Bytes) {
1318     Mode = ARM_AM::da;
1319   } else {
1320     MergeInstr = findIncDecAfter(MBBI, Base, Pred, PredReg, Offset, TRI);
1321     if (((Mode != ARM_AM::ia && Mode != ARM_AM::ib) || Offset != Bytes) &&
1322         ((Mode != ARM_AM::da && Mode != ARM_AM::db) || Offset != -Bytes)) {
1323 
1324       // We couldn't find an inc/dec to merge. But if the base is dead, we
1325       // can still change to a writeback form as that will save us 2 bytes
1326       // of code size. It can create WAW hazards though, so only do it if
1327       // we're minimizing code size.
1328       if (!STI->hasMinSize() || !BaseKill)
1329         return false;
1330 
1331       bool HighRegsUsed = false;
1332       for (const MachineOperand &MO : llvm::drop_begin(MI->operands(), 2))
1333         if (MO.getReg() >= ARM::R8) {
1334           HighRegsUsed = true;
1335           break;
1336         }
1337 
1338       if (!HighRegsUsed)
1339         MergeInstr = MBB.end();
1340       else
1341         return false;
1342     }
1343   }
1344   if (MergeInstr != MBB.end()) {
1345     LLVM_DEBUG(dbgs() << "  Erasing old increment: " << *MergeInstr);
1346     MBB.erase(MergeInstr);
1347   }
1348 
1349   unsigned NewOpc = getUpdatingLSMultipleOpcode(Opcode, Mode);
1350   MachineInstrBuilder MIB = BuildMI(MBB, MBBI, DL, TII->get(NewOpc))
1351     .addReg(Base, getDefRegState(true)) // WB base register
1352     .addReg(Base, getKillRegState(BaseKill))
1353     .addImm(Pred).addReg(PredReg);
1354 
1355   // Transfer the rest of operands.
1356   for (const MachineOperand &MO : llvm::drop_begin(MI->operands(), 3))
1357     MIB.add(MO);
1358 
1359   // Transfer memoperands.
1360   MIB.setMemRefs(MI->memoperands());
1361 
1362   LLVM_DEBUG(dbgs() << "  Added new load/store: " << *MIB);
1363   MBB.erase(MBBI);
1364   return true;
1365 }
1366 
getPreIndexedLoadStoreOpcode(unsigned Opc,ARM_AM::AddrOpc Mode)1367 static unsigned getPreIndexedLoadStoreOpcode(unsigned Opc,
1368                                              ARM_AM::AddrOpc Mode) {
1369   switch (Opc) {
1370   case ARM::LDRi12:
1371     return ARM::LDR_PRE_IMM;
1372   case ARM::STRi12:
1373     return ARM::STR_PRE_IMM;
1374   case ARM::VLDRS:
1375     return Mode == ARM_AM::add ? ARM::VLDMSIA_UPD : ARM::VLDMSDB_UPD;
1376   case ARM::VLDRD:
1377     return Mode == ARM_AM::add ? ARM::VLDMDIA_UPD : ARM::VLDMDDB_UPD;
1378   case ARM::VSTRS:
1379     return Mode == ARM_AM::add ? ARM::VSTMSIA_UPD : ARM::VSTMSDB_UPD;
1380   case ARM::VSTRD:
1381     return Mode == ARM_AM::add ? ARM::VSTMDIA_UPD : ARM::VSTMDDB_UPD;
1382   case ARM::t2LDRi8:
1383   case ARM::t2LDRi12:
1384     return ARM::t2LDR_PRE;
1385   case ARM::t2STRi8:
1386   case ARM::t2STRi12:
1387     return ARM::t2STR_PRE;
1388   default: llvm_unreachable("Unhandled opcode!");
1389   }
1390 }
1391 
getPostIndexedLoadStoreOpcode(unsigned Opc,ARM_AM::AddrOpc Mode)1392 static unsigned getPostIndexedLoadStoreOpcode(unsigned Opc,
1393                                               ARM_AM::AddrOpc Mode) {
1394   switch (Opc) {
1395   case ARM::LDRi12:
1396     return ARM::LDR_POST_IMM;
1397   case ARM::STRi12:
1398     return ARM::STR_POST_IMM;
1399   case ARM::VLDRS:
1400     return Mode == ARM_AM::add ? ARM::VLDMSIA_UPD : ARM::VLDMSDB_UPD;
1401   case ARM::VLDRD:
1402     return Mode == ARM_AM::add ? ARM::VLDMDIA_UPD : ARM::VLDMDDB_UPD;
1403   case ARM::VSTRS:
1404     return Mode == ARM_AM::add ? ARM::VSTMSIA_UPD : ARM::VSTMSDB_UPD;
1405   case ARM::VSTRD:
1406     return Mode == ARM_AM::add ? ARM::VSTMDIA_UPD : ARM::VSTMDDB_UPD;
1407   case ARM::t2LDRi8:
1408   case ARM::t2LDRi12:
1409     return ARM::t2LDR_POST;
1410   case ARM::t2LDRBi8:
1411   case ARM::t2LDRBi12:
1412     return ARM::t2LDRB_POST;
1413   case ARM::t2LDRSBi8:
1414   case ARM::t2LDRSBi12:
1415     return ARM::t2LDRSB_POST;
1416   case ARM::t2LDRHi8:
1417   case ARM::t2LDRHi12:
1418     return ARM::t2LDRH_POST;
1419   case ARM::t2LDRSHi8:
1420   case ARM::t2LDRSHi12:
1421     return ARM::t2LDRSH_POST;
1422   case ARM::t2STRi8:
1423   case ARM::t2STRi12:
1424     return ARM::t2STR_POST;
1425   case ARM::t2STRBi8:
1426   case ARM::t2STRBi12:
1427     return ARM::t2STRB_POST;
1428   case ARM::t2STRHi8:
1429   case ARM::t2STRHi12:
1430     return ARM::t2STRH_POST;
1431 
1432   case ARM::MVE_VLDRBS16:
1433     return ARM::MVE_VLDRBS16_post;
1434   case ARM::MVE_VLDRBS32:
1435     return ARM::MVE_VLDRBS32_post;
1436   case ARM::MVE_VLDRBU16:
1437     return ARM::MVE_VLDRBU16_post;
1438   case ARM::MVE_VLDRBU32:
1439     return ARM::MVE_VLDRBU32_post;
1440   case ARM::MVE_VLDRHS32:
1441     return ARM::MVE_VLDRHS32_post;
1442   case ARM::MVE_VLDRHU32:
1443     return ARM::MVE_VLDRHU32_post;
1444   case ARM::MVE_VLDRBU8:
1445     return ARM::MVE_VLDRBU8_post;
1446   case ARM::MVE_VLDRHU16:
1447     return ARM::MVE_VLDRHU16_post;
1448   case ARM::MVE_VLDRWU32:
1449     return ARM::MVE_VLDRWU32_post;
1450   case ARM::MVE_VSTRB16:
1451     return ARM::MVE_VSTRB16_post;
1452   case ARM::MVE_VSTRB32:
1453     return ARM::MVE_VSTRB32_post;
1454   case ARM::MVE_VSTRH32:
1455     return ARM::MVE_VSTRH32_post;
1456   case ARM::MVE_VSTRBU8:
1457     return ARM::MVE_VSTRBU8_post;
1458   case ARM::MVE_VSTRHU16:
1459     return ARM::MVE_VSTRHU16_post;
1460   case ARM::MVE_VSTRWU32:
1461     return ARM::MVE_VSTRWU32_post;
1462 
1463   default: llvm_unreachable("Unhandled opcode!");
1464   }
1465 }
1466 
1467 /// Fold proceeding/trailing inc/dec of base register into the
1468 /// LDR/STR/FLD{D|S}/FST{D|S} op when possible:
MergeBaseUpdateLoadStore(MachineInstr * MI)1469 bool ARMLoadStoreOpt::MergeBaseUpdateLoadStore(MachineInstr *MI) {
1470   // Thumb1 doesn't have updating LDR/STR.
1471   // FIXME: Use LDM/STM with single register instead.
1472   if (isThumb1) return false;
1473   LLVM_DEBUG(dbgs() << "Attempting to merge update of: " << *MI);
1474 
1475   Register Base = getLoadStoreBaseOp(*MI).getReg();
1476   bool BaseKill = getLoadStoreBaseOp(*MI).isKill();
1477   unsigned Opcode = MI->getOpcode();
1478   DebugLoc DL = MI->getDebugLoc();
1479   bool isAM5 = (Opcode == ARM::VLDRD || Opcode == ARM::VLDRS ||
1480                 Opcode == ARM::VSTRD || Opcode == ARM::VSTRS);
1481   bool isAM2 = (Opcode == ARM::LDRi12 || Opcode == ARM::STRi12);
1482   if (isi32Load(Opcode) || isi32Store(Opcode))
1483     if (MI->getOperand(2).getImm() != 0)
1484       return false;
1485   if (isAM5 && ARM_AM::getAM5Offset(MI->getOperand(2).getImm()) != 0)
1486     return false;
1487 
1488   // Can't do the merge if the destination register is the same as the would-be
1489   // writeback register.
1490   if (MI->getOperand(0).getReg() == Base)
1491     return false;
1492 
1493   Register PredReg;
1494   ARMCC::CondCodes Pred = getInstrPredicate(*MI, PredReg);
1495   int Bytes = getLSMultipleTransferSize(MI);
1496   MachineBasicBlock &MBB = *MI->getParent();
1497   MachineBasicBlock::iterator MBBI(MI);
1498   int Offset;
1499   MachineBasicBlock::iterator MergeInstr
1500     = findIncDecBefore(MBBI, Base, Pred, PredReg, Offset);
1501   unsigned NewOpc;
1502   if (!isAM5 && Offset == Bytes) {
1503     NewOpc = getPreIndexedLoadStoreOpcode(Opcode, ARM_AM::add);
1504   } else if (Offset == -Bytes) {
1505     NewOpc = getPreIndexedLoadStoreOpcode(Opcode, ARM_AM::sub);
1506   } else {
1507     MergeInstr = findIncDecAfter(MBBI, Base, Pred, PredReg, Offset, TRI);
1508     if (MergeInstr == MBB.end())
1509       return false;
1510 
1511     NewOpc = getPostIndexedLoadStoreOpcode(Opcode, ARM_AM::add);
1512     if ((isAM5 && Offset != Bytes) ||
1513         (!isAM5 && !isLegalAddressImm(NewOpc, Offset, TII))) {
1514       NewOpc = getPostIndexedLoadStoreOpcode(Opcode, ARM_AM::sub);
1515       if (isAM5 || !isLegalAddressImm(NewOpc, Offset, TII))
1516         return false;
1517     }
1518   }
1519   LLVM_DEBUG(dbgs() << "  Erasing old increment: " << *MergeInstr);
1520   MBB.erase(MergeInstr);
1521 
1522   ARM_AM::AddrOpc AddSub = Offset < 0 ? ARM_AM::sub : ARM_AM::add;
1523 
1524   bool isLd = isLoadSingle(Opcode);
1525   if (isAM5) {
1526     // VLDM[SD]_UPD, VSTM[SD]_UPD
1527     // (There are no base-updating versions of VLDR/VSTR instructions, but the
1528     // updating load/store-multiple instructions can be used with only one
1529     // register.)
1530     MachineOperand &MO = MI->getOperand(0);
1531     auto MIB = BuildMI(MBB, MBBI, DL, TII->get(NewOpc))
1532                    .addReg(Base, getDefRegState(true)) // WB base register
1533                    .addReg(Base, getKillRegState(isLd ? BaseKill : false))
1534                    .addImm(Pred)
1535                    .addReg(PredReg)
1536                    .addReg(MO.getReg(), (isLd ? getDefRegState(true)
1537                                               : getKillRegState(MO.isKill())))
1538                    .cloneMemRefs(*MI);
1539     (void)MIB;
1540     LLVM_DEBUG(dbgs() << "  Added new instruction: " << *MIB);
1541   } else if (isLd) {
1542     if (isAM2) {
1543       // LDR_PRE, LDR_POST
1544       if (NewOpc == ARM::LDR_PRE_IMM || NewOpc == ARM::LDRB_PRE_IMM) {
1545         auto MIB =
1546             BuildMI(MBB, MBBI, DL, TII->get(NewOpc), MI->getOperand(0).getReg())
1547                 .addReg(Base, RegState::Define)
1548                 .addReg(Base)
1549                 .addImm(Offset)
1550                 .addImm(Pred)
1551                 .addReg(PredReg)
1552                 .cloneMemRefs(*MI);
1553         (void)MIB;
1554         LLVM_DEBUG(dbgs() << "  Added new instruction: " << *MIB);
1555       } else {
1556         int Imm = ARM_AM::getAM2Opc(AddSub, abs(Offset), ARM_AM::no_shift);
1557         auto MIB =
1558             BuildMI(MBB, MBBI, DL, TII->get(NewOpc), MI->getOperand(0).getReg())
1559                 .addReg(Base, RegState::Define)
1560                 .addReg(Base)
1561                 .addReg(0)
1562                 .addImm(Imm)
1563                 .add(predOps(Pred, PredReg))
1564                 .cloneMemRefs(*MI);
1565         (void)MIB;
1566         LLVM_DEBUG(dbgs() << "  Added new instruction: " << *MIB);
1567       }
1568     } else {
1569       // t2LDR_PRE, t2LDR_POST
1570       auto MIB =
1571           BuildMI(MBB, MBBI, DL, TII->get(NewOpc), MI->getOperand(0).getReg())
1572               .addReg(Base, RegState::Define)
1573               .addReg(Base)
1574               .addImm(Offset)
1575               .add(predOps(Pred, PredReg))
1576               .cloneMemRefs(*MI);
1577       (void)MIB;
1578       LLVM_DEBUG(dbgs() << "  Added new instruction: " << *MIB);
1579     }
1580   } else {
1581     MachineOperand &MO = MI->getOperand(0);
1582     // FIXME: post-indexed stores use am2offset_imm, which still encodes
1583     // the vestigal zero-reg offset register. When that's fixed, this clause
1584     // can be removed entirely.
1585     if (isAM2 && NewOpc == ARM::STR_POST_IMM) {
1586       int Imm = ARM_AM::getAM2Opc(AddSub, abs(Offset), ARM_AM::no_shift);
1587       // STR_PRE, STR_POST
1588       auto MIB = BuildMI(MBB, MBBI, DL, TII->get(NewOpc), Base)
1589                      .addReg(MO.getReg(), getKillRegState(MO.isKill()))
1590                      .addReg(Base)
1591                      .addReg(0)
1592                      .addImm(Imm)
1593                      .add(predOps(Pred, PredReg))
1594                      .cloneMemRefs(*MI);
1595       (void)MIB;
1596       LLVM_DEBUG(dbgs() << "  Added new instruction: " << *MIB);
1597     } else {
1598       // t2STR_PRE, t2STR_POST
1599       auto MIB = BuildMI(MBB, MBBI, DL, TII->get(NewOpc), Base)
1600                      .addReg(MO.getReg(), getKillRegState(MO.isKill()))
1601                      .addReg(Base)
1602                      .addImm(Offset)
1603                      .add(predOps(Pred, PredReg))
1604                      .cloneMemRefs(*MI);
1605       (void)MIB;
1606       LLVM_DEBUG(dbgs() << "  Added new instruction: " << *MIB);
1607     }
1608   }
1609   MBB.erase(MBBI);
1610 
1611   return true;
1612 }
1613 
MergeBaseUpdateLSDouble(MachineInstr & MI) const1614 bool ARMLoadStoreOpt::MergeBaseUpdateLSDouble(MachineInstr &MI) const {
1615   unsigned Opcode = MI.getOpcode();
1616   assert((Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8) &&
1617          "Must have t2STRDi8 or t2LDRDi8");
1618   if (MI.getOperand(3).getImm() != 0)
1619     return false;
1620   LLVM_DEBUG(dbgs() << "Attempting to merge update of: " << MI);
1621 
1622   // Behaviour for writeback is undefined if base register is the same as one
1623   // of the others.
1624   const MachineOperand &BaseOp = MI.getOperand(2);
1625   Register Base = BaseOp.getReg();
1626   const MachineOperand &Reg0Op = MI.getOperand(0);
1627   const MachineOperand &Reg1Op = MI.getOperand(1);
1628   if (Reg0Op.getReg() == Base || Reg1Op.getReg() == Base)
1629     return false;
1630 
1631   Register PredReg;
1632   ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
1633   MachineBasicBlock::iterator MBBI(MI);
1634   MachineBasicBlock &MBB = *MI.getParent();
1635   int Offset;
1636   MachineBasicBlock::iterator MergeInstr = findIncDecBefore(MBBI, Base, Pred,
1637                                                             PredReg, Offset);
1638   unsigned NewOpc;
1639   if (Offset == 8 || Offset == -8) {
1640     NewOpc = Opcode == ARM::t2LDRDi8 ? ARM::t2LDRD_PRE : ARM::t2STRD_PRE;
1641   } else {
1642     MergeInstr = findIncDecAfter(MBBI, Base, Pred, PredReg, Offset, TRI);
1643     if (MergeInstr == MBB.end())
1644       return false;
1645     NewOpc = Opcode == ARM::t2LDRDi8 ? ARM::t2LDRD_POST : ARM::t2STRD_POST;
1646     if (!isLegalAddressImm(NewOpc, Offset, TII))
1647       return false;
1648   }
1649   LLVM_DEBUG(dbgs() << "  Erasing old increment: " << *MergeInstr);
1650   MBB.erase(MergeInstr);
1651 
1652   DebugLoc DL = MI.getDebugLoc();
1653   MachineInstrBuilder MIB = BuildMI(MBB, MBBI, DL, TII->get(NewOpc));
1654   if (NewOpc == ARM::t2LDRD_PRE || NewOpc == ARM::t2LDRD_POST) {
1655     MIB.add(Reg0Op).add(Reg1Op).addReg(BaseOp.getReg(), RegState::Define);
1656   } else {
1657     assert(NewOpc == ARM::t2STRD_PRE || NewOpc == ARM::t2STRD_POST);
1658     MIB.addReg(BaseOp.getReg(), RegState::Define).add(Reg0Op).add(Reg1Op);
1659   }
1660   MIB.addReg(BaseOp.getReg(), RegState::Kill)
1661      .addImm(Offset).addImm(Pred).addReg(PredReg);
1662   assert(TII->get(Opcode).getNumOperands() == 6 &&
1663          TII->get(NewOpc).getNumOperands() == 7 &&
1664          "Unexpected number of operands in Opcode specification.");
1665 
1666   // Transfer implicit operands.
1667   for (const MachineOperand &MO : MI.implicit_operands())
1668     MIB.add(MO);
1669   MIB.cloneMemRefs(MI);
1670 
1671   LLVM_DEBUG(dbgs() << "  Added new load/store: " << *MIB);
1672   MBB.erase(MBBI);
1673   return true;
1674 }
1675 
1676 /// Returns true if instruction is a memory operation that this pass is capable
1677 /// of operating on.
isMemoryOp(const MachineInstr & MI)1678 static bool isMemoryOp(const MachineInstr &MI) {
1679   unsigned Opcode = MI.getOpcode();
1680   switch (Opcode) {
1681   case ARM::VLDRS:
1682   case ARM::VSTRS:
1683   case ARM::VLDRD:
1684   case ARM::VSTRD:
1685   case ARM::LDRi12:
1686   case ARM::STRi12:
1687   case ARM::tLDRi:
1688   case ARM::tSTRi:
1689   case ARM::tLDRspi:
1690   case ARM::tSTRspi:
1691   case ARM::t2LDRi8:
1692   case ARM::t2LDRi12:
1693   case ARM::t2STRi8:
1694   case ARM::t2STRi12:
1695     break;
1696   default:
1697     return false;
1698   }
1699   if (!MI.getOperand(1).isReg())
1700     return false;
1701 
1702   // When no memory operands are present, conservatively assume unaligned,
1703   // volatile, unfoldable.
1704   if (!MI.hasOneMemOperand())
1705     return false;
1706 
1707   const MachineMemOperand &MMO = **MI.memoperands_begin();
1708 
1709   // Don't touch volatile memory accesses - we may be changing their order.
1710   // TODO: We could allow unordered and monotonic atomics here, but we need to
1711   // make sure the resulting ldm/stm is correctly marked as atomic.
1712   if (MMO.isVolatile() || MMO.isAtomic())
1713     return false;
1714 
1715   // Unaligned ldr/str is emulated by some kernels, but unaligned ldm/stm is
1716   // not.
1717   if (MMO.getAlign() < Align(4))
1718     return false;
1719 
1720   // str <undef> could probably be eliminated entirely, but for now we just want
1721   // to avoid making a mess of it.
1722   // FIXME: Use str <undef> as a wildcard to enable better stm folding.
1723   if (MI.getOperand(0).isReg() && MI.getOperand(0).isUndef())
1724     return false;
1725 
1726   // Likewise don't mess with references to undefined addresses.
1727   if (MI.getOperand(1).isUndef())
1728     return false;
1729 
1730   return true;
1731 }
1732 
InsertLDR_STR(MachineBasicBlock & MBB,MachineBasicBlock::iterator & MBBI,int Offset,bool isDef,unsigned NewOpc,unsigned Reg,bool RegDeadKill,bool RegUndef,unsigned BaseReg,bool BaseKill,bool BaseUndef,ARMCC::CondCodes Pred,unsigned PredReg,const TargetInstrInfo * TII,MachineInstr * MI)1733 static void InsertLDR_STR(MachineBasicBlock &MBB,
1734                           MachineBasicBlock::iterator &MBBI, int Offset,
1735                           bool isDef, unsigned NewOpc, unsigned Reg,
1736                           bool RegDeadKill, bool RegUndef, unsigned BaseReg,
1737                           bool BaseKill, bool BaseUndef, ARMCC::CondCodes Pred,
1738                           unsigned PredReg, const TargetInstrInfo *TII,
1739                           MachineInstr *MI) {
1740   if (isDef) {
1741     MachineInstrBuilder MIB = BuildMI(MBB, MBBI, MBBI->getDebugLoc(),
1742                                       TII->get(NewOpc))
1743       .addReg(Reg, getDefRegState(true) | getDeadRegState(RegDeadKill))
1744       .addReg(BaseReg, getKillRegState(BaseKill)|getUndefRegState(BaseUndef));
1745     MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
1746     // FIXME: This is overly conservative; the new instruction accesses 4
1747     // bytes, not 8.
1748     MIB.cloneMemRefs(*MI);
1749   } else {
1750     MachineInstrBuilder MIB = BuildMI(MBB, MBBI, MBBI->getDebugLoc(),
1751                                       TII->get(NewOpc))
1752       .addReg(Reg, getKillRegState(RegDeadKill) | getUndefRegState(RegUndef))
1753       .addReg(BaseReg, getKillRegState(BaseKill)|getUndefRegState(BaseUndef));
1754     MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
1755     // FIXME: This is overly conservative; the new instruction accesses 4
1756     // bytes, not 8.
1757     MIB.cloneMemRefs(*MI);
1758   }
1759 }
1760 
FixInvalidRegPairOp(MachineBasicBlock & MBB,MachineBasicBlock::iterator & MBBI)1761 bool ARMLoadStoreOpt::FixInvalidRegPairOp(MachineBasicBlock &MBB,
1762                                           MachineBasicBlock::iterator &MBBI) {
1763   MachineInstr *MI = &*MBBI;
1764   unsigned Opcode = MI->getOpcode();
1765   // FIXME: Code/comments below check Opcode == t2STRDi8, but this check returns
1766   // if we see this opcode.
1767   if (Opcode != ARM::LDRD && Opcode != ARM::STRD && Opcode != ARM::t2LDRDi8)
1768     return false;
1769 
1770   const MachineOperand &BaseOp = MI->getOperand(2);
1771   Register BaseReg = BaseOp.getReg();
1772   Register EvenReg = MI->getOperand(0).getReg();
1773   Register OddReg = MI->getOperand(1).getReg();
1774   unsigned EvenRegNum = TRI->getDwarfRegNum(EvenReg, false);
1775   unsigned OddRegNum  = TRI->getDwarfRegNum(OddReg, false);
1776 
1777   // ARM errata 602117: LDRD with base in list may result in incorrect base
1778   // register when interrupted or faulted.
1779   bool Errata602117 = EvenReg == BaseReg &&
1780     (Opcode == ARM::LDRD || Opcode == ARM::t2LDRDi8) && STI->isCortexM3();
1781   // ARM LDRD/STRD needs consecutive registers.
1782   bool NonConsecutiveRegs = (Opcode == ARM::LDRD || Opcode == ARM::STRD) &&
1783     (EvenRegNum % 2 != 0 || EvenRegNum + 1 != OddRegNum);
1784 
1785   if (!Errata602117 && !NonConsecutiveRegs)
1786     return false;
1787 
1788   bool isT2 = Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8;
1789   bool isLd = Opcode == ARM::LDRD || Opcode == ARM::t2LDRDi8;
1790   bool EvenDeadKill = isLd ?
1791     MI->getOperand(0).isDead() : MI->getOperand(0).isKill();
1792   bool EvenUndef = MI->getOperand(0).isUndef();
1793   bool OddDeadKill  = isLd ?
1794     MI->getOperand(1).isDead() : MI->getOperand(1).isKill();
1795   bool OddUndef = MI->getOperand(1).isUndef();
1796   bool BaseKill = BaseOp.isKill();
1797   bool BaseUndef = BaseOp.isUndef();
1798   assert((isT2 || MI->getOperand(3).getReg() == ARM::NoRegister) &&
1799          "register offset not handled below");
1800   int OffImm = getMemoryOpOffset(*MI);
1801   Register PredReg;
1802   ARMCC::CondCodes Pred = getInstrPredicate(*MI, PredReg);
1803 
1804   if (OddRegNum > EvenRegNum && OffImm == 0) {
1805     // Ascending register numbers and no offset. It's safe to change it to a
1806     // ldm or stm.
1807     unsigned NewOpc = (isLd)
1808       ? (isT2 ? ARM::t2LDMIA : ARM::LDMIA)
1809       : (isT2 ? ARM::t2STMIA : ARM::STMIA);
1810     if (isLd) {
1811       BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
1812         .addReg(BaseReg, getKillRegState(BaseKill))
1813         .addImm(Pred).addReg(PredReg)
1814         .addReg(EvenReg, getDefRegState(isLd) | getDeadRegState(EvenDeadKill))
1815         .addReg(OddReg,  getDefRegState(isLd) | getDeadRegState(OddDeadKill))
1816         .cloneMemRefs(*MI);
1817       ++NumLDRD2LDM;
1818     } else {
1819       BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
1820         .addReg(BaseReg, getKillRegState(BaseKill))
1821         .addImm(Pred).addReg(PredReg)
1822         .addReg(EvenReg,
1823                 getKillRegState(EvenDeadKill) | getUndefRegState(EvenUndef))
1824         .addReg(OddReg,
1825                 getKillRegState(OddDeadKill)  | getUndefRegState(OddUndef))
1826         .cloneMemRefs(*MI);
1827       ++NumSTRD2STM;
1828     }
1829   } else {
1830     // Split into two instructions.
1831     unsigned NewOpc = (isLd)
1832       ? (isT2 ? (OffImm < 0 ? ARM::t2LDRi8 : ARM::t2LDRi12) : ARM::LDRi12)
1833       : (isT2 ? (OffImm < 0 ? ARM::t2STRi8 : ARM::t2STRi12) : ARM::STRi12);
1834     // Be extra careful for thumb2. t2LDRi8 can't reference a zero offset,
1835     // so adjust and use t2LDRi12 here for that.
1836     unsigned NewOpc2 = (isLd)
1837       ? (isT2 ? (OffImm+4 < 0 ? ARM::t2LDRi8 : ARM::t2LDRi12) : ARM::LDRi12)
1838       : (isT2 ? (OffImm+4 < 0 ? ARM::t2STRi8 : ARM::t2STRi12) : ARM::STRi12);
1839     // If this is a load, make sure the first load does not clobber the base
1840     // register before the second load reads it.
1841     if (isLd && TRI->regsOverlap(EvenReg, BaseReg)) {
1842       assert(!TRI->regsOverlap(OddReg, BaseReg));
1843       InsertLDR_STR(MBB, MBBI, OffImm + 4, isLd, NewOpc2, OddReg, OddDeadKill,
1844                     false, BaseReg, false, BaseUndef, Pred, PredReg, TII, MI);
1845       InsertLDR_STR(MBB, MBBI, OffImm, isLd, NewOpc, EvenReg, EvenDeadKill,
1846                     false, BaseReg, BaseKill, BaseUndef, Pred, PredReg, TII,
1847                     MI);
1848     } else {
1849       if (OddReg == EvenReg && EvenDeadKill) {
1850         // If the two source operands are the same, the kill marker is
1851         // probably on the first one. e.g.
1852         // t2STRDi8 killed %r5, %r5, killed %r9, 0, 14, %reg0
1853         EvenDeadKill = false;
1854         OddDeadKill = true;
1855       }
1856       // Never kill the base register in the first instruction.
1857       if (EvenReg == BaseReg)
1858         EvenDeadKill = false;
1859       InsertLDR_STR(MBB, MBBI, OffImm, isLd, NewOpc, EvenReg, EvenDeadKill,
1860                     EvenUndef, BaseReg, false, BaseUndef, Pred, PredReg, TII,
1861                     MI);
1862       InsertLDR_STR(MBB, MBBI, OffImm + 4, isLd, NewOpc2, OddReg, OddDeadKill,
1863                     OddUndef, BaseReg, BaseKill, BaseUndef, Pred, PredReg, TII,
1864                     MI);
1865     }
1866     if (isLd)
1867       ++NumLDRD2LDR;
1868     else
1869       ++NumSTRD2STR;
1870   }
1871 
1872   MBBI = MBB.erase(MBBI);
1873   return true;
1874 }
1875 
1876 /// An optimization pass to turn multiple LDR / STR ops of the same base and
1877 /// incrementing offset into LDM / STM ops.
LoadStoreMultipleOpti(MachineBasicBlock & MBB)1878 bool ARMLoadStoreOpt::LoadStoreMultipleOpti(MachineBasicBlock &MBB) {
1879   MemOpQueue MemOps;
1880   unsigned CurrBase = 0;
1881   unsigned CurrOpc = ~0u;
1882   ARMCC::CondCodes CurrPred = ARMCC::AL;
1883   unsigned Position = 0;
1884   assert(Candidates.size() == 0);
1885   assert(MergeBaseCandidates.size() == 0);
1886   LiveRegsValid = false;
1887 
1888   for (MachineBasicBlock::iterator I = MBB.end(), MBBI; I != MBB.begin();
1889        I = MBBI) {
1890     // The instruction in front of the iterator is the one we look at.
1891     MBBI = std::prev(I);
1892     if (FixInvalidRegPairOp(MBB, MBBI))
1893       continue;
1894     ++Position;
1895 
1896     if (isMemoryOp(*MBBI)) {
1897       unsigned Opcode = MBBI->getOpcode();
1898       const MachineOperand &MO = MBBI->getOperand(0);
1899       Register Reg = MO.getReg();
1900       Register Base = getLoadStoreBaseOp(*MBBI).getReg();
1901       Register PredReg;
1902       ARMCC::CondCodes Pred = getInstrPredicate(*MBBI, PredReg);
1903       int Offset = getMemoryOpOffset(*MBBI);
1904       if (CurrBase == 0) {
1905         // Start of a new chain.
1906         CurrBase = Base;
1907         CurrOpc  = Opcode;
1908         CurrPred = Pred;
1909         MemOps.push_back(MemOpQueueEntry(*MBBI, Offset, Position));
1910         continue;
1911       }
1912       // Note: No need to match PredReg in the next if.
1913       if (CurrOpc == Opcode && CurrBase == Base && CurrPred == Pred) {
1914         // Watch out for:
1915         //   r4 := ldr [r0, #8]
1916         //   r4 := ldr [r0, #4]
1917         // or
1918         //   r0 := ldr [r0]
1919         // If a load overrides the base register or a register loaded by
1920         // another load in our chain, we cannot take this instruction.
1921         bool Overlap = false;
1922         if (isLoadSingle(Opcode)) {
1923           Overlap = (Base == Reg);
1924           if (!Overlap) {
1925             for (const MemOpQueueEntry &E : MemOps) {
1926               if (TRI->regsOverlap(Reg, E.MI->getOperand(0).getReg())) {
1927                 Overlap = true;
1928                 break;
1929               }
1930             }
1931           }
1932         }
1933 
1934         if (!Overlap) {
1935           // Check offset and sort memory operation into the current chain.
1936           if (Offset > MemOps.back().Offset) {
1937             MemOps.push_back(MemOpQueueEntry(*MBBI, Offset, Position));
1938             continue;
1939           } else {
1940             MemOpQueue::iterator MI, ME;
1941             for (MI = MemOps.begin(), ME = MemOps.end(); MI != ME; ++MI) {
1942               if (Offset < MI->Offset) {
1943                 // Found a place to insert.
1944                 break;
1945               }
1946               if (Offset == MI->Offset) {
1947                 // Collision, abort.
1948                 MI = ME;
1949                 break;
1950               }
1951             }
1952             if (MI != MemOps.end()) {
1953               MemOps.insert(MI, MemOpQueueEntry(*MBBI, Offset, Position));
1954               continue;
1955             }
1956           }
1957         }
1958       }
1959 
1960       // Don't advance the iterator; The op will start a new chain next.
1961       MBBI = I;
1962       --Position;
1963       // Fallthrough to look into existing chain.
1964     } else if (MBBI->isDebugInstr()) {
1965       continue;
1966     } else if (MBBI->getOpcode() == ARM::t2LDRDi8 ||
1967                MBBI->getOpcode() == ARM::t2STRDi8) {
1968       // ARMPreAllocLoadStoreOpt has already formed some LDRD/STRD instructions
1969       // remember them because we may still be able to merge add/sub into them.
1970       MergeBaseCandidates.push_back(&*MBBI);
1971     }
1972 
1973     // If we are here then the chain is broken; Extract candidates for a merge.
1974     if (MemOps.size() > 0) {
1975       FormCandidates(MemOps);
1976       // Reset for the next chain.
1977       CurrBase = 0;
1978       CurrOpc = ~0u;
1979       CurrPred = ARMCC::AL;
1980       MemOps.clear();
1981     }
1982   }
1983   if (MemOps.size() > 0)
1984     FormCandidates(MemOps);
1985 
1986   // Sort candidates so they get processed from end to begin of the basic
1987   // block later; This is necessary for liveness calculation.
1988   auto LessThan = [](const MergeCandidate* M0, const MergeCandidate *M1) {
1989     return M0->InsertPos < M1->InsertPos;
1990   };
1991   llvm::sort(Candidates, LessThan);
1992 
1993   // Go through list of candidates and merge.
1994   bool Changed = false;
1995   for (const MergeCandidate *Candidate : Candidates) {
1996     if (Candidate->CanMergeToLSMulti || Candidate->CanMergeToLSDouble) {
1997       MachineInstr *Merged = MergeOpsUpdate(*Candidate);
1998       // Merge preceding/trailing base inc/dec into the merged op.
1999       if (Merged) {
2000         Changed = true;
2001         unsigned Opcode = Merged->getOpcode();
2002         if (Opcode == ARM::t2STRDi8 || Opcode == ARM::t2LDRDi8)
2003           MergeBaseUpdateLSDouble(*Merged);
2004         else
2005           MergeBaseUpdateLSMultiple(Merged);
2006       } else {
2007         for (MachineInstr *MI : Candidate->Instrs) {
2008           if (MergeBaseUpdateLoadStore(MI))
2009             Changed = true;
2010         }
2011       }
2012     } else {
2013       assert(Candidate->Instrs.size() == 1);
2014       if (MergeBaseUpdateLoadStore(Candidate->Instrs.front()))
2015         Changed = true;
2016     }
2017   }
2018   Candidates.clear();
2019   // Try to fold add/sub into the LDRD/STRD formed by ARMPreAllocLoadStoreOpt.
2020   for (MachineInstr *MI : MergeBaseCandidates)
2021     MergeBaseUpdateLSDouble(*MI);
2022   MergeBaseCandidates.clear();
2023 
2024   return Changed;
2025 }
2026 
2027 /// If this is a exit BB, try merging the return ops ("bx lr" and "mov pc, lr")
2028 /// into the preceding stack restore so it directly restore the value of LR
2029 /// into pc.
2030 ///   ldmfd sp!, {..., lr}
2031 ///   bx lr
2032 /// or
2033 ///   ldmfd sp!, {..., lr}
2034 ///   mov pc, lr
2035 /// =>
2036 ///   ldmfd sp!, {..., pc}
MergeReturnIntoLDM(MachineBasicBlock & MBB)2037 bool ARMLoadStoreOpt::MergeReturnIntoLDM(MachineBasicBlock &MBB) {
2038   // Thumb1 LDM doesn't allow high registers.
2039   if (isThumb1) return false;
2040   if (MBB.empty()) return false;
2041 
2042   MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr();
2043   if (MBBI != MBB.begin() && MBBI != MBB.end() &&
2044       (MBBI->getOpcode() == ARM::BX_RET ||
2045        MBBI->getOpcode() == ARM::tBX_RET ||
2046        MBBI->getOpcode() == ARM::MOVPCLR)) {
2047     MachineBasicBlock::iterator PrevI = std::prev(MBBI);
2048     // Ignore any debug instructions.
2049     while (PrevI->isDebugInstr() && PrevI != MBB.begin())
2050       --PrevI;
2051     MachineInstr &PrevMI = *PrevI;
2052     unsigned Opcode = PrevMI.getOpcode();
2053     if (Opcode == ARM::LDMIA_UPD || Opcode == ARM::LDMDA_UPD ||
2054         Opcode == ARM::LDMDB_UPD || Opcode == ARM::LDMIB_UPD ||
2055         Opcode == ARM::t2LDMIA_UPD || Opcode == ARM::t2LDMDB_UPD) {
2056       MachineOperand &MO = PrevMI.getOperand(PrevMI.getNumOperands() - 1);
2057       if (MO.getReg() != ARM::LR)
2058         return false;
2059       unsigned NewOpc = (isThumb2 ? ARM::t2LDMIA_RET : ARM::LDMIA_RET);
2060       assert(((isThumb2 && Opcode == ARM::t2LDMIA_UPD) ||
2061               Opcode == ARM::LDMIA_UPD) && "Unsupported multiple load-return!");
2062       PrevMI.setDesc(TII->get(NewOpc));
2063       MO.setReg(ARM::PC);
2064       PrevMI.copyImplicitOps(*MBB.getParent(), *MBBI);
2065       MBB.erase(MBBI);
2066       return true;
2067     }
2068   }
2069   return false;
2070 }
2071 
CombineMovBx(MachineBasicBlock & MBB)2072 bool ARMLoadStoreOpt::CombineMovBx(MachineBasicBlock &MBB) {
2073   MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator();
2074   if (MBBI == MBB.begin() || MBBI == MBB.end() ||
2075       MBBI->getOpcode() != ARM::tBX_RET)
2076     return false;
2077 
2078   MachineBasicBlock::iterator Prev = MBBI;
2079   --Prev;
2080   if (Prev->getOpcode() != ARM::tMOVr ||
2081       !Prev->definesRegister(ARM::LR, /*TRI=*/nullptr))
2082     return false;
2083 
2084   for (auto Use : Prev->uses())
2085     if (Use.isKill()) {
2086       assert(STI->hasV4TOps());
2087       BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(ARM::tBX))
2088           .addReg(Use.getReg(), RegState::Kill)
2089           .add(predOps(ARMCC::AL))
2090           .copyImplicitOps(*MBBI);
2091       MBB.erase(MBBI);
2092       MBB.erase(Prev);
2093       return true;
2094     }
2095 
2096   llvm_unreachable("tMOVr doesn't kill a reg before tBX_RET?");
2097 }
2098 
runOnMachineFunction(MachineFunction & Fn)2099 bool ARMLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
2100   if (skipFunction(Fn.getFunction()))
2101     return false;
2102 
2103   MF = &Fn;
2104   STI = &Fn.getSubtarget<ARMSubtarget>();
2105   TL = STI->getTargetLowering();
2106   AFI = Fn.getInfo<ARMFunctionInfo>();
2107   TII = STI->getInstrInfo();
2108   TRI = STI->getRegisterInfo();
2109 
2110   RegClassInfoValid = false;
2111   isThumb2 = AFI->isThumb2Function();
2112   isThumb1 = AFI->isThumbFunction() && !isThumb2;
2113 
2114   bool Modified = false, ModifiedLDMReturn = false;
2115   for (MachineBasicBlock &MBB : Fn) {
2116     Modified |= LoadStoreMultipleOpti(MBB);
2117     if (STI->hasV5TOps() && !AFI->shouldSignReturnAddress())
2118       ModifiedLDMReturn |= MergeReturnIntoLDM(MBB);
2119     if (isThumb1)
2120       Modified |= CombineMovBx(MBB);
2121   }
2122   Modified |= ModifiedLDMReturn;
2123 
2124   // If we merged a BX instruction into an LDM, we need to re-calculate whether
2125   // LR is restored. This check needs to consider the whole function, not just
2126   // the instruction(s) we changed, because there may be other BX returns which
2127   // still need LR to be restored.
2128   if (ModifiedLDMReturn)
2129     ARMFrameLowering::updateLRRestored(Fn);
2130 
2131   Allocator.DestroyAll();
2132   return Modified;
2133 }
2134 
2135 #define ARM_PREALLOC_LOAD_STORE_OPT_NAME                                       \
2136   "ARM pre- register allocation load / store optimization pass"
2137 
2138 namespace {
2139 
2140   /// Pre- register allocation pass that move load / stores from consecutive
2141   /// locations close to make it more likely they will be combined later.
2142   struct ARMPreAllocLoadStoreOpt : public MachineFunctionPass{
2143     static char ID;
2144 
2145     AliasAnalysis *AA;
2146     const DataLayout *TD;
2147     const TargetInstrInfo *TII;
2148     const TargetRegisterInfo *TRI;
2149     const ARMSubtarget *STI;
2150     MachineRegisterInfo *MRI;
2151     MachineDominatorTree *DT;
2152     MachineFunction *MF;
2153 
ARMPreAllocLoadStoreOpt__anonf2a56f050311::ARMPreAllocLoadStoreOpt2154     ARMPreAllocLoadStoreOpt() : MachineFunctionPass(ID) {}
2155 
2156     bool runOnMachineFunction(MachineFunction &Fn) override;
2157 
getPassName__anonf2a56f050311::ARMPreAllocLoadStoreOpt2158     StringRef getPassName() const override {
2159       return ARM_PREALLOC_LOAD_STORE_OPT_NAME;
2160     }
2161 
getAnalysisUsage__anonf2a56f050311::ARMPreAllocLoadStoreOpt2162     void getAnalysisUsage(AnalysisUsage &AU) const override {
2163       AU.addRequired<AAResultsWrapperPass>();
2164       AU.addRequired<MachineDominatorTreeWrapperPass>();
2165       AU.addPreserved<MachineDominatorTreeWrapperPass>();
2166       MachineFunctionPass::getAnalysisUsage(AU);
2167     }
2168 
2169   private:
2170     bool CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1, DebugLoc &dl,
2171                           unsigned &NewOpc, Register &EvenReg, Register &OddReg,
2172                           Register &BaseReg, int &Offset, Register &PredReg,
2173                           ARMCC::CondCodes &Pred, bool &isT2);
2174     bool RescheduleOps(
2175         MachineBasicBlock *MBB, SmallVectorImpl<MachineInstr *> &Ops,
2176         unsigned Base, bool isLd, DenseMap<MachineInstr *, unsigned> &MI2LocMap,
2177         SmallDenseMap<Register, SmallVector<MachineInstr *>, 8> &RegisterMap);
2178     bool RescheduleLoadStoreInstrs(MachineBasicBlock *MBB);
2179     bool DistributeIncrements();
2180     bool DistributeIncrements(Register Base);
2181   };
2182 
2183 } // end anonymous namespace
2184 
2185 char ARMPreAllocLoadStoreOpt::ID = 0;
2186 
2187 INITIALIZE_PASS_BEGIN(ARMPreAllocLoadStoreOpt, "arm-prera-ldst-opt",
2188                       ARM_PREALLOC_LOAD_STORE_OPT_NAME, false, false)
2189 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
2190 INITIALIZE_PASS_END(ARMPreAllocLoadStoreOpt, "arm-prera-ldst-opt",
2191                     ARM_PREALLOC_LOAD_STORE_OPT_NAME, false, false)
2192 
2193 // Limit the number of instructions to be rescheduled.
2194 // FIXME: tune this limit, and/or come up with some better heuristics.
2195 static cl::opt<unsigned> InstReorderLimit("arm-prera-ldst-opt-reorder-limit",
2196                                           cl::init(8), cl::Hidden);
2197 
runOnMachineFunction(MachineFunction & Fn)2198 bool ARMPreAllocLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
2199   if (AssumeMisalignedLoadStores || skipFunction(Fn.getFunction()))
2200     return false;
2201 
2202   TD = &Fn.getDataLayout();
2203   STI = &Fn.getSubtarget<ARMSubtarget>();
2204   TII = STI->getInstrInfo();
2205   TRI = STI->getRegisterInfo();
2206   MRI = &Fn.getRegInfo();
2207   DT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree();
2208   MF  = &Fn;
2209   AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2210 
2211   bool Modified = DistributeIncrements();
2212   for (MachineBasicBlock &MFI : Fn)
2213     Modified |= RescheduleLoadStoreInstrs(&MFI);
2214 
2215   return Modified;
2216 }
2217 
IsSafeAndProfitableToMove(bool isLd,unsigned Base,MachineBasicBlock::iterator I,MachineBasicBlock::iterator E,SmallPtrSetImpl<MachineInstr * > & MemOps,SmallSet<unsigned,4> & MemRegs,const TargetRegisterInfo * TRI,AliasAnalysis * AA)2218 static bool IsSafeAndProfitableToMove(bool isLd, unsigned Base,
2219                                       MachineBasicBlock::iterator I,
2220                                       MachineBasicBlock::iterator E,
2221                                       SmallPtrSetImpl<MachineInstr*> &MemOps,
2222                                       SmallSet<unsigned, 4> &MemRegs,
2223                                       const TargetRegisterInfo *TRI,
2224                                       AliasAnalysis *AA) {
2225   // Are there stores / loads / calls between them?
2226   SmallSet<unsigned, 4> AddedRegPressure;
2227   while (++I != E) {
2228     if (I->isDebugInstr() || MemOps.count(&*I))
2229       continue;
2230     if (I->isCall() || I->isTerminator() || I->hasUnmodeledSideEffects())
2231       return false;
2232     if (I->mayStore() || (!isLd && I->mayLoad()))
2233       for (MachineInstr *MemOp : MemOps)
2234         if (I->mayAlias(AA, *MemOp, /*UseTBAA*/ false))
2235           return false;
2236     for (unsigned j = 0, NumOps = I->getNumOperands(); j != NumOps; ++j) {
2237       MachineOperand &MO = I->getOperand(j);
2238       if (!MO.isReg())
2239         continue;
2240       Register Reg = MO.getReg();
2241       if (MO.isDef() && TRI->regsOverlap(Reg, Base))
2242         return false;
2243       if (Reg != Base && !MemRegs.count(Reg))
2244         AddedRegPressure.insert(Reg);
2245     }
2246   }
2247 
2248   // Estimate register pressure increase due to the transformation.
2249   if (MemRegs.size() <= 4)
2250     // Ok if we are moving small number of instructions.
2251     return true;
2252   return AddedRegPressure.size() <= MemRegs.size() * 2;
2253 }
2254 
CanFormLdStDWord(MachineInstr * Op0,MachineInstr * Op1,DebugLoc & dl,unsigned & NewOpc,Register & FirstReg,Register & SecondReg,Register & BaseReg,int & Offset,Register & PredReg,ARMCC::CondCodes & Pred,bool & isT2)2255 bool ARMPreAllocLoadStoreOpt::CanFormLdStDWord(
2256     MachineInstr *Op0, MachineInstr *Op1, DebugLoc &dl, unsigned &NewOpc,
2257     Register &FirstReg, Register &SecondReg, Register &BaseReg, int &Offset,
2258     Register &PredReg, ARMCC::CondCodes &Pred, bool &isT2) {
2259   // Make sure we're allowed to generate LDRD/STRD.
2260   if (!STI->hasV5TEOps())
2261     return false;
2262 
2263   // FIXME: VLDRS / VSTRS -> VLDRD / VSTRD
2264   unsigned Scale = 1;
2265   unsigned Opcode = Op0->getOpcode();
2266   if (Opcode == ARM::LDRi12) {
2267     NewOpc = ARM::LDRD;
2268   } else if (Opcode == ARM::STRi12) {
2269     NewOpc = ARM::STRD;
2270   } else if (Opcode == ARM::t2LDRi8 || Opcode == ARM::t2LDRi12) {
2271     NewOpc = ARM::t2LDRDi8;
2272     Scale = 4;
2273     isT2 = true;
2274   } else if (Opcode == ARM::t2STRi8 || Opcode == ARM::t2STRi12) {
2275     NewOpc = ARM::t2STRDi8;
2276     Scale = 4;
2277     isT2 = true;
2278   } else {
2279     return false;
2280   }
2281 
2282   // Make sure the base address satisfies i64 ld / st alignment requirement.
2283   // At the moment, we ignore the memoryoperand's value.
2284   // If we want to use AliasAnalysis, we should check it accordingly.
2285   if (!Op0->hasOneMemOperand() ||
2286       (*Op0->memoperands_begin())->isVolatile() ||
2287       (*Op0->memoperands_begin())->isAtomic())
2288     return false;
2289 
2290   Align Alignment = (*Op0->memoperands_begin())->getAlign();
2291   Align ReqAlign = STI->getDualLoadStoreAlignment();
2292   if (Alignment < ReqAlign)
2293     return false;
2294 
2295   // Then make sure the immediate offset fits.
2296   int OffImm = getMemoryOpOffset(*Op0);
2297   if (isT2) {
2298     int Limit = (1 << 8) * Scale;
2299     if (OffImm >= Limit || (OffImm <= -Limit) || (OffImm & (Scale-1)))
2300       return false;
2301     Offset = OffImm;
2302   } else {
2303     ARM_AM::AddrOpc AddSub = ARM_AM::add;
2304     if (OffImm < 0) {
2305       AddSub = ARM_AM::sub;
2306       OffImm = - OffImm;
2307     }
2308     int Limit = (1 << 8) * Scale;
2309     if (OffImm >= Limit || (OffImm & (Scale-1)))
2310       return false;
2311     Offset = ARM_AM::getAM3Opc(AddSub, OffImm);
2312   }
2313   FirstReg = Op0->getOperand(0).getReg();
2314   SecondReg = Op1->getOperand(0).getReg();
2315   if (FirstReg == SecondReg)
2316     return false;
2317   BaseReg = Op0->getOperand(1).getReg();
2318   Pred = getInstrPredicate(*Op0, PredReg);
2319   dl = Op0->getDebugLoc();
2320   return true;
2321 }
2322 
RescheduleOps(MachineBasicBlock * MBB,SmallVectorImpl<MachineInstr * > & Ops,unsigned Base,bool isLd,DenseMap<MachineInstr *,unsigned> & MI2LocMap,SmallDenseMap<Register,SmallVector<MachineInstr * >,8> & RegisterMap)2323 bool ARMPreAllocLoadStoreOpt::RescheduleOps(
2324     MachineBasicBlock *MBB, SmallVectorImpl<MachineInstr *> &Ops, unsigned Base,
2325     bool isLd, DenseMap<MachineInstr *, unsigned> &MI2LocMap,
2326     SmallDenseMap<Register, SmallVector<MachineInstr *>, 8> &RegisterMap) {
2327   bool RetVal = false;
2328 
2329   // Sort by offset (in reverse order).
2330   llvm::sort(Ops, [](const MachineInstr *LHS, const MachineInstr *RHS) {
2331     int LOffset = getMemoryOpOffset(*LHS);
2332     int ROffset = getMemoryOpOffset(*RHS);
2333     assert(LHS == RHS || LOffset != ROffset);
2334     return LOffset > ROffset;
2335   });
2336 
2337   // The loads / stores of the same base are in order. Scan them from first to
2338   // last and check for the following:
2339   // 1. Any def of base.
2340   // 2. Any gaps.
2341   while (Ops.size() > 1) {
2342     unsigned FirstLoc = ~0U;
2343     unsigned LastLoc = 0;
2344     MachineInstr *FirstOp = nullptr;
2345     MachineInstr *LastOp = nullptr;
2346     int LastOffset = 0;
2347     unsigned LastOpcode = 0;
2348     unsigned LastBytes = 0;
2349     unsigned NumMove = 0;
2350     for (MachineInstr *Op : llvm::reverse(Ops)) {
2351       // Make sure each operation has the same kind.
2352       unsigned LSMOpcode
2353         = getLoadStoreMultipleOpcode(Op->getOpcode(), ARM_AM::ia);
2354       if (LastOpcode && LSMOpcode != LastOpcode)
2355         break;
2356 
2357       // Check that we have a continuous set of offsets.
2358       int Offset = getMemoryOpOffset(*Op);
2359       unsigned Bytes = getLSMultipleTransferSize(Op);
2360       if (LastBytes) {
2361         if (Bytes != LastBytes || Offset != (LastOffset + (int)Bytes))
2362           break;
2363       }
2364 
2365       // Don't try to reschedule too many instructions.
2366       if (NumMove == InstReorderLimit)
2367         break;
2368 
2369       // Found a mergable instruction; save information about it.
2370       ++NumMove;
2371       LastOffset = Offset;
2372       LastBytes = Bytes;
2373       LastOpcode = LSMOpcode;
2374 
2375       unsigned Loc = MI2LocMap[Op];
2376       if (Loc <= FirstLoc) {
2377         FirstLoc = Loc;
2378         FirstOp = Op;
2379       }
2380       if (Loc >= LastLoc) {
2381         LastLoc = Loc;
2382         LastOp = Op;
2383       }
2384     }
2385 
2386     if (NumMove <= 1)
2387       Ops.pop_back();
2388     else {
2389       SmallPtrSet<MachineInstr*, 4> MemOps;
2390       SmallSet<unsigned, 4> MemRegs;
2391       for (size_t i = Ops.size() - NumMove, e = Ops.size(); i != e; ++i) {
2392         MemOps.insert(Ops[i]);
2393         MemRegs.insert(Ops[i]->getOperand(0).getReg());
2394       }
2395 
2396       // Be conservative, if the instructions are too far apart, don't
2397       // move them. We want to limit the increase of register pressure.
2398       bool DoMove = (LastLoc - FirstLoc) <= NumMove*4; // FIXME: Tune this.
2399       if (DoMove)
2400         DoMove = IsSafeAndProfitableToMove(isLd, Base, FirstOp, LastOp,
2401                                            MemOps, MemRegs, TRI, AA);
2402       if (!DoMove) {
2403         for (unsigned i = 0; i != NumMove; ++i)
2404           Ops.pop_back();
2405       } else {
2406         // This is the new location for the loads / stores.
2407         MachineBasicBlock::iterator InsertPos = isLd ? FirstOp : LastOp;
2408         while (InsertPos != MBB->end() &&
2409                (MemOps.count(&*InsertPos) || InsertPos->isDebugInstr()))
2410           ++InsertPos;
2411 
2412         // If we are moving a pair of loads / stores, see if it makes sense
2413         // to try to allocate a pair of registers that can form register pairs.
2414         MachineInstr *Op0 = Ops.back();
2415         MachineInstr *Op1 = Ops[Ops.size()-2];
2416         Register FirstReg, SecondReg;
2417         Register BaseReg, PredReg;
2418         ARMCC::CondCodes Pred = ARMCC::AL;
2419         bool isT2 = false;
2420         unsigned NewOpc = 0;
2421         int Offset = 0;
2422         DebugLoc dl;
2423         if (NumMove == 2 && CanFormLdStDWord(Op0, Op1, dl, NewOpc,
2424                                              FirstReg, SecondReg, BaseReg,
2425                                              Offset, PredReg, Pred, isT2)) {
2426           Ops.pop_back();
2427           Ops.pop_back();
2428 
2429           const MCInstrDesc &MCID = TII->get(NewOpc);
2430           const TargetRegisterClass *TRC = TII->getRegClass(MCID, 0, TRI, *MF);
2431           MRI->constrainRegClass(FirstReg, TRC);
2432           MRI->constrainRegClass(SecondReg, TRC);
2433 
2434           // Form the pair instruction.
2435           if (isLd) {
2436             MachineInstrBuilder MIB = BuildMI(*MBB, InsertPos, dl, MCID)
2437               .addReg(FirstReg, RegState::Define)
2438               .addReg(SecondReg, RegState::Define)
2439               .addReg(BaseReg);
2440             // FIXME: We're converting from LDRi12 to an insn that still
2441             // uses addrmode2, so we need an explicit offset reg. It should
2442             // always by reg0 since we're transforming LDRi12s.
2443             if (!isT2)
2444               MIB.addReg(0);
2445             MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
2446             MIB.cloneMergedMemRefs({Op0, Op1});
2447             LLVM_DEBUG(dbgs() << "Formed " << *MIB << "\n");
2448             ++NumLDRDFormed;
2449           } else {
2450             MachineInstrBuilder MIB = BuildMI(*MBB, InsertPos, dl, MCID)
2451               .addReg(FirstReg)
2452               .addReg(SecondReg)
2453               .addReg(BaseReg);
2454             // FIXME: We're converting from LDRi12 to an insn that still
2455             // uses addrmode2, so we need an explicit offset reg. It should
2456             // always by reg0 since we're transforming STRi12s.
2457             if (!isT2)
2458               MIB.addReg(0);
2459             MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
2460             MIB.cloneMergedMemRefs({Op0, Op1});
2461             LLVM_DEBUG(dbgs() << "Formed " << *MIB << "\n");
2462             ++NumSTRDFormed;
2463           }
2464           MBB->erase(Op0);
2465           MBB->erase(Op1);
2466 
2467           if (!isT2) {
2468             // Add register allocation hints to form register pairs.
2469             MRI->setRegAllocationHint(FirstReg, ARMRI::RegPairEven, SecondReg);
2470             MRI->setRegAllocationHint(SecondReg,  ARMRI::RegPairOdd, FirstReg);
2471           }
2472         } else {
2473           for (unsigned i = 0; i != NumMove; ++i) {
2474             MachineInstr *Op = Ops.pop_back_val();
2475             if (isLd) {
2476               // Populate RegisterMap with all Registers defined by loads.
2477               Register Reg = Op->getOperand(0).getReg();
2478               RegisterMap[Reg];
2479             }
2480 
2481             MBB->splice(InsertPos, MBB, Op);
2482           }
2483         }
2484 
2485         NumLdStMoved += NumMove;
2486         RetVal = true;
2487       }
2488     }
2489   }
2490 
2491   return RetVal;
2492 }
2493 
forEachDbgRegOperand(MachineInstr * MI,std::function<void (MachineOperand &)> Fn)2494 static void forEachDbgRegOperand(MachineInstr *MI,
2495                                  std::function<void(MachineOperand &)> Fn) {
2496   if (MI->isNonListDebugValue()) {
2497     auto &Op = MI->getOperand(0);
2498     if (Op.isReg())
2499       Fn(Op);
2500   } else {
2501     for (unsigned I = 2; I < MI->getNumOperands(); I++) {
2502       auto &Op = MI->getOperand(I);
2503       if (Op.isReg())
2504         Fn(Op);
2505     }
2506   }
2507 }
2508 
2509 // Update the RegisterMap with the instruction that was moved because a
2510 // DBG_VALUE_LIST may need to be moved again.
updateRegisterMapForDbgValueListAfterMove(SmallDenseMap<Register,SmallVector<MachineInstr * >,8> & RegisterMap,MachineInstr * DbgValueListInstr,MachineInstr * InstrToReplace)2511 static void updateRegisterMapForDbgValueListAfterMove(
2512     SmallDenseMap<Register, SmallVector<MachineInstr *>, 8> &RegisterMap,
2513     MachineInstr *DbgValueListInstr, MachineInstr *InstrToReplace) {
2514 
2515   forEachDbgRegOperand(DbgValueListInstr, [&](MachineOperand &Op) {
2516     auto RegIt = RegisterMap.find(Op.getReg());
2517     if (RegIt == RegisterMap.end())
2518       return;
2519     auto &InstrVec = RegIt->getSecond();
2520     for (unsigned I = 0; I < InstrVec.size(); I++)
2521       if (InstrVec[I] == InstrToReplace)
2522         InstrVec[I] = DbgValueListInstr;
2523   });
2524 }
2525 
createDebugVariableFromMachineInstr(MachineInstr * MI)2526 static DebugVariable createDebugVariableFromMachineInstr(MachineInstr *MI) {
2527   auto DbgVar = DebugVariable(MI->getDebugVariable(), MI->getDebugExpression(),
2528                               MI->getDebugLoc()->getInlinedAt());
2529   return DbgVar;
2530 }
2531 
2532 bool
RescheduleLoadStoreInstrs(MachineBasicBlock * MBB)2533 ARMPreAllocLoadStoreOpt::RescheduleLoadStoreInstrs(MachineBasicBlock *MBB) {
2534   bool RetVal = false;
2535 
2536   DenseMap<MachineInstr*, unsigned> MI2LocMap;
2537   using MapIt = DenseMap<unsigned, SmallVector<MachineInstr *, 4>>::iterator;
2538   using Base2InstMap = DenseMap<unsigned, SmallVector<MachineInstr *, 4>>;
2539   using BaseVec = SmallVector<unsigned, 4>;
2540   Base2InstMap Base2LdsMap;
2541   Base2InstMap Base2StsMap;
2542   BaseVec LdBases;
2543   BaseVec StBases;
2544   // This map is used to track the relationship between the virtual
2545   // register that is the result of a load that is moved and the DBG_VALUE
2546   // MachineInstr pointer that uses that virtual register.
2547   SmallDenseMap<Register, SmallVector<MachineInstr *>, 8> RegisterMap;
2548 
2549   unsigned Loc = 0;
2550   MachineBasicBlock::iterator MBBI = MBB->begin();
2551   MachineBasicBlock::iterator E = MBB->end();
2552   while (MBBI != E) {
2553     for (; MBBI != E; ++MBBI) {
2554       MachineInstr &MI = *MBBI;
2555       if (MI.isCall() || MI.isTerminator()) {
2556         // Stop at barriers.
2557         ++MBBI;
2558         break;
2559       }
2560 
2561       if (!MI.isDebugInstr())
2562         MI2LocMap[&MI] = ++Loc;
2563 
2564       if (!isMemoryOp(MI))
2565         continue;
2566       Register PredReg;
2567       if (getInstrPredicate(MI, PredReg) != ARMCC::AL)
2568         continue;
2569 
2570       int Opc = MI.getOpcode();
2571       bool isLd = isLoadSingle(Opc);
2572       Register Base = MI.getOperand(1).getReg();
2573       int Offset = getMemoryOpOffset(MI);
2574       bool StopHere = false;
2575       auto FindBases = [&] (Base2InstMap &Base2Ops, BaseVec &Bases) {
2576         MapIt BI = Base2Ops.find(Base);
2577         if (BI == Base2Ops.end()) {
2578           Base2Ops[Base].push_back(&MI);
2579           Bases.push_back(Base);
2580           return;
2581         }
2582         for (const MachineInstr *MI : BI->second) {
2583           if (Offset == getMemoryOpOffset(*MI)) {
2584             StopHere = true;
2585             break;
2586           }
2587         }
2588         if (!StopHere)
2589           BI->second.push_back(&MI);
2590       };
2591 
2592       if (isLd)
2593         FindBases(Base2LdsMap, LdBases);
2594       else
2595         FindBases(Base2StsMap, StBases);
2596 
2597       if (StopHere) {
2598         // Found a duplicate (a base+offset combination that's seen earlier).
2599         // Backtrack.
2600         --Loc;
2601         break;
2602       }
2603     }
2604 
2605     // Re-schedule loads.
2606     for (unsigned Base : LdBases) {
2607       SmallVectorImpl<MachineInstr *> &Lds = Base2LdsMap[Base];
2608       if (Lds.size() > 1)
2609         RetVal |= RescheduleOps(MBB, Lds, Base, true, MI2LocMap, RegisterMap);
2610     }
2611 
2612     // Re-schedule stores.
2613     for (unsigned Base : StBases) {
2614       SmallVectorImpl<MachineInstr *> &Sts = Base2StsMap[Base];
2615       if (Sts.size() > 1)
2616         RetVal |= RescheduleOps(MBB, Sts, Base, false, MI2LocMap, RegisterMap);
2617     }
2618 
2619     if (MBBI != E) {
2620       Base2LdsMap.clear();
2621       Base2StsMap.clear();
2622       LdBases.clear();
2623       StBases.clear();
2624     }
2625   }
2626 
2627   // Reschedule DBG_VALUEs to match any loads that were moved. When a load is
2628   // sunk beyond a DBG_VALUE that is referring to it, the DBG_VALUE becomes a
2629   // use-before-def, resulting in a loss of debug info.
2630 
2631   // Example:
2632   // Before the Pre Register Allocation Load Store Pass
2633   // inst_a
2634   // %2 = ld ...
2635   // inst_b
2636   // DBG_VALUE %2, "x", ...
2637   // %3 = ld ...
2638 
2639   // After the Pass:
2640   // inst_a
2641   // inst_b
2642   // DBG_VALUE %2, "x", ...
2643   // %2 = ld ...
2644   // %3 = ld ...
2645 
2646   // The code below addresses this by moving the DBG_VALUE to the position
2647   // immediately after the load.
2648 
2649   // Example:
2650   // After the code below:
2651   // inst_a
2652   // inst_b
2653   // %2 = ld ...
2654   // DBG_VALUE %2, "x", ...
2655   // %3 = ld ...
2656 
2657   // The algorithm works in two phases: First RescheduleOps() populates the
2658   // RegisterMap with registers that were moved as keys, there is no value
2659   // inserted. In the next phase, every MachineInstr in a basic block is
2660   // iterated over. If it is a valid DBG_VALUE or DBG_VALUE_LIST and it uses one
2661   // or more registers in the RegisterMap, the RegisterMap and InstrMap are
2662   // populated with the MachineInstr. If the DBG_VALUE or DBG_VALUE_LIST
2663   // describes debug information for a variable that already exists in the
2664   // DbgValueSinkCandidates, the MachineInstr in the DbgValueSinkCandidates must
2665   // be set to undef. If the current MachineInstr is a load that was moved,
2666   // undef the corresponding DBG_VALUE or DBG_VALUE_LIST and clone it to below
2667   // the load.
2668 
2669   // To illustrate the above algorithm visually let's take this example.
2670 
2671   // Before the Pre Register Allocation Load Store Pass:
2672   // %2 = ld ...
2673   // DBG_VALUE %2, A, .... # X
2674   // DBG_VALUE 0, A, ... # Y
2675   // %3 = ld ...
2676   // DBG_VALUE %3, A, ..., # Z
2677   // %4 = ld ...
2678 
2679   // After Pre Register Allocation Load Store Pass:
2680   // DBG_VALUE %2, A, .... # X
2681   // DBG_VALUE 0, A, ... # Y
2682   // DBG_VALUE %3, A, ..., # Z
2683   // %2 = ld ...
2684   // %3 = ld ...
2685   // %4 = ld ...
2686 
2687   // The algorithm below does the following:
2688 
2689   // In the beginning, the RegisterMap will have been populated with the virtual
2690   // registers %2, and %3, the DbgValueSinkCandidates and the InstrMap will be
2691   // empty. DbgValueSinkCandidates = {}, RegisterMap = {2 -> {}, 3 -> {}},
2692   // InstrMap {}
2693   // -> DBG_VALUE %2, A, .... # X
2694   // DBG_VALUE 0, A, ... # Y
2695   // DBG_VALUE %3, A, ..., # Z
2696   // %2 = ld ...
2697   // %3 = ld ...
2698   // %4 = ld ...
2699 
2700   // After the first DBG_VALUE (denoted with an X) is processed, the
2701   // DbgValueSinkCandidates and InstrMap will be populated and the RegisterMap
2702   // entry for %2 will be populated as well. DbgValueSinkCandidates = {A -> X},
2703   // RegisterMap = {2 -> {X}, 3 -> {}}, InstrMap {X -> 2}
2704   // DBG_VALUE %2, A, .... # X
2705   // -> DBG_VALUE 0, A, ... # Y
2706   // DBG_VALUE %3, A, ..., # Z
2707   // %2 = ld ...
2708   // %3 = ld ...
2709   // %4 = ld ...
2710 
2711   // After the DBG_VALUE Y is processed, the DbgValueSinkCandidates is updated
2712   // to now hold Y for A and the RegisterMap is also updated to remove X from
2713   // %2, this is because both X and Y describe the same debug variable A. X is
2714   // also updated to have a $noreg as the first operand.
2715   // DbgValueSinkCandidates = {A -> {Y}}, RegisterMap = {2 -> {}, 3 -> {}},
2716   // InstrMap = {X-> 2}
2717   // DBG_VALUE $noreg, A, .... # X
2718   // DBG_VALUE 0, A, ... # Y
2719   // -> DBG_VALUE %3, A, ..., # Z
2720   // %2 = ld ...
2721   // %3 = ld ...
2722   // %4 = ld ...
2723 
2724   // After DBG_VALUE Z is processed, the DbgValueSinkCandidates is updated to
2725   // hold Z fr A, the RegisterMap is updated to hold Z for %3, and the InstrMap
2726   // is updated to have Z mapped to %3. This is again because Z describes the
2727   // debug variable A, Y is not updated to have $noreg as first operand because
2728   // its first operand is an immediate, not a register.
2729   // DbgValueSinkCandidates = {A -> {Z}}, RegisterMap = {2 -> {}, 3 -> {Z}},
2730   // InstrMap = {X -> 2, Z -> 3}
2731   // DBG_VALUE $noreg, A, .... # X
2732   // DBG_VALUE 0, A, ... # Y
2733   // DBG_VALUE %3, A, ..., # Z
2734   // -> %2 = ld ...
2735   // %3 = ld ...
2736   // %4 = ld ...
2737 
2738   // Nothing happens here since the RegisterMap for %2 contains no value.
2739   // DbgValueSinkCandidates = {A -> {Z}}, RegisterMap = {2 -> {}, 3 -> {Z}},
2740   // InstrMap = {X -> 2, Z -> 3}
2741   // DBG_VALUE $noreg, A, .... # X
2742   // DBG_VALUE 0, A, ... # Y
2743   // DBG_VALUE %3, A, ..., # Z
2744   // %2 = ld ...
2745   // -> %3 = ld ...
2746   // %4 = ld ...
2747 
2748   // Since the RegisterMap contains Z as a value for %3, the MachineInstr
2749   // pointer Z is copied to come after the load for %3 and the old Z's first
2750   // operand is changed to $noreg the Basic Block iterator is moved to after the
2751   // DBG_VALUE Z's new position.
2752   // DbgValueSinkCandidates = {A -> {Z}}, RegisterMap = {2 -> {}, 3 -> {Z}},
2753   // InstrMap = {X -> 2, Z -> 3}
2754   // DBG_VALUE $noreg, A, .... # X
2755   // DBG_VALUE 0, A, ... # Y
2756   // DBG_VALUE $noreg, A, ..., # Old Z
2757   // %2 = ld ...
2758   // %3 = ld ...
2759   // DBG_VALUE %3, A, ..., # Z
2760   // -> %4 = ld ...
2761 
2762   // Nothing happens for %4 and the algorithm exits having processed the entire
2763   // Basic Block.
2764   // DbgValueSinkCandidates = {A -> {Z}}, RegisterMap = {2 -> {}, 3 -> {Z}},
2765   // InstrMap = {X -> 2, Z -> 3}
2766   // DBG_VALUE $noreg, A, .... # X
2767   // DBG_VALUE 0, A, ... # Y
2768   // DBG_VALUE $noreg, A, ..., # Old Z
2769   // %2 = ld ...
2770   // %3 = ld ...
2771   // DBG_VALUE %3, A, ..., # Z
2772   // %4 = ld ...
2773 
2774   // This map is used to track the relationship between
2775   // a Debug Variable and the DBG_VALUE MachineInstr pointer that describes the
2776   // debug information for that Debug Variable.
2777   SmallDenseMap<DebugVariable, MachineInstr *, 8> DbgValueSinkCandidates;
2778   // This map is used to track the relationship between a DBG_VALUE or
2779   // DBG_VALUE_LIST MachineInstr pointer and Registers that it uses.
2780   SmallDenseMap<MachineInstr *, SmallVector<Register>, 8> InstrMap;
2781   for (MBBI = MBB->begin(), E = MBB->end(); MBBI != E; ++MBBI) {
2782     MachineInstr &MI = *MBBI;
2783 
2784     auto PopulateRegisterAndInstrMapForDebugInstr = [&](Register Reg) {
2785       auto RegIt = RegisterMap.find(Reg);
2786       if (RegIt == RegisterMap.end())
2787         return;
2788       auto &InstrVec = RegIt->getSecond();
2789       InstrVec.push_back(&MI);
2790       InstrMap[&MI].push_back(Reg);
2791     };
2792 
2793     if (MI.isDebugValue()) {
2794       assert(MI.getDebugVariable() &&
2795              "DBG_VALUE or DBG_VALUE_LIST must contain a DILocalVariable");
2796 
2797       auto DbgVar = createDebugVariableFromMachineInstr(&MI);
2798       // If the first operand is a register and it exists in the RegisterMap, we
2799       // know this is a DBG_VALUE that uses the result of a load that was moved,
2800       // and is therefore a candidate to also be moved, add it to the
2801       // RegisterMap and InstrMap.
2802       forEachDbgRegOperand(&MI, [&](MachineOperand &Op) {
2803         PopulateRegisterAndInstrMapForDebugInstr(Op.getReg());
2804       });
2805 
2806       // If the current DBG_VALUE describes the same variable as one of the
2807       // in-flight DBG_VALUEs, remove the candidate from the list and set it to
2808       // undef. Moving one DBG_VALUE past another would result in the variable's
2809       // value going back in time when stepping through the block in the
2810       // debugger.
2811       auto InstrIt = DbgValueSinkCandidates.find(DbgVar);
2812       if (InstrIt != DbgValueSinkCandidates.end()) {
2813         auto *Instr = InstrIt->getSecond();
2814         auto RegIt = InstrMap.find(Instr);
2815         if (RegIt != InstrMap.end()) {
2816           const auto &RegVec = RegIt->getSecond();
2817           // For every Register in the RegVec, remove the MachineInstr in the
2818           // RegisterMap that describes the DbgVar.
2819           for (auto &Reg : RegVec) {
2820             auto RegIt = RegisterMap.find(Reg);
2821             if (RegIt == RegisterMap.end())
2822               continue;
2823             auto &InstrVec = RegIt->getSecond();
2824             auto IsDbgVar = [&](MachineInstr *I) -> bool {
2825               auto Var = createDebugVariableFromMachineInstr(I);
2826               return Var == DbgVar;
2827             };
2828 
2829             llvm::erase_if(InstrVec, IsDbgVar);
2830           }
2831           forEachDbgRegOperand(Instr,
2832                                [&](MachineOperand &Op) { Op.setReg(0); });
2833         }
2834       }
2835       DbgValueSinkCandidates[DbgVar] = &MI;
2836     } else {
2837       // If the first operand of a load matches with a DBG_VALUE in RegisterMap,
2838       // then move that DBG_VALUE to below the load.
2839       auto Opc = MI.getOpcode();
2840       if (!isLoadSingle(Opc))
2841         continue;
2842       auto Reg = MI.getOperand(0).getReg();
2843       auto RegIt = RegisterMap.find(Reg);
2844       if (RegIt == RegisterMap.end())
2845         continue;
2846       auto &DbgInstrVec = RegIt->getSecond();
2847       if (!DbgInstrVec.size())
2848         continue;
2849       for (auto *DbgInstr : DbgInstrVec) {
2850         MachineBasicBlock::iterator InsertPos = std::next(MBBI);
2851         auto *ClonedMI = MI.getMF()->CloneMachineInstr(DbgInstr);
2852         MBB->insert(InsertPos, ClonedMI);
2853         MBBI++;
2854         //  Erase the entry into the DbgValueSinkCandidates for the DBG_VALUE
2855         //  that was moved.
2856         auto DbgVar = createDebugVariableFromMachineInstr(DbgInstr);
2857         auto DbgIt = DbgValueSinkCandidates.find(DbgVar);
2858         // If the instruction is a DBG_VALUE_LIST, it may have already been
2859         // erased from the DbgValueSinkCandidates. Only erase if it exists in
2860         // the DbgValueSinkCandidates.
2861         if (DbgIt != DbgValueSinkCandidates.end())
2862           DbgValueSinkCandidates.erase(DbgIt);
2863         // Zero out original dbg instr
2864         forEachDbgRegOperand(DbgInstr,
2865                              [&](MachineOperand &Op) { Op.setReg(0); });
2866         // Update RegisterMap with ClonedMI because it might have to be moved
2867         // again.
2868         if (DbgInstr->isDebugValueList())
2869           updateRegisterMapForDbgValueListAfterMove(RegisterMap, ClonedMI,
2870                                                     DbgInstr);
2871       }
2872     }
2873   }
2874   return RetVal;
2875 }
2876 
2877 // Get the Base register operand index from the memory access MachineInst if we
2878 // should attempt to distribute postinc on it. Return -1 if not of a valid
2879 // instruction type. If it returns an index, it is assumed that instruction is a
2880 // r+i indexing mode, and getBaseOperandIndex() + 1 is the Offset index.
getBaseOperandIndex(MachineInstr & MI)2881 static int getBaseOperandIndex(MachineInstr &MI) {
2882   switch (MI.getOpcode()) {
2883   case ARM::MVE_VLDRBS16:
2884   case ARM::MVE_VLDRBS32:
2885   case ARM::MVE_VLDRBU16:
2886   case ARM::MVE_VLDRBU32:
2887   case ARM::MVE_VLDRHS32:
2888   case ARM::MVE_VLDRHU32:
2889   case ARM::MVE_VLDRBU8:
2890   case ARM::MVE_VLDRHU16:
2891   case ARM::MVE_VLDRWU32:
2892   case ARM::MVE_VSTRB16:
2893   case ARM::MVE_VSTRB32:
2894   case ARM::MVE_VSTRH32:
2895   case ARM::MVE_VSTRBU8:
2896   case ARM::MVE_VSTRHU16:
2897   case ARM::MVE_VSTRWU32:
2898   case ARM::t2LDRHi8:
2899   case ARM::t2LDRHi12:
2900   case ARM::t2LDRSHi8:
2901   case ARM::t2LDRSHi12:
2902   case ARM::t2LDRBi8:
2903   case ARM::t2LDRBi12:
2904   case ARM::t2LDRSBi8:
2905   case ARM::t2LDRSBi12:
2906   case ARM::t2STRBi8:
2907   case ARM::t2STRBi12:
2908   case ARM::t2STRHi8:
2909   case ARM::t2STRHi12:
2910     return 1;
2911   case ARM::MVE_VLDRBS16_post:
2912   case ARM::MVE_VLDRBS32_post:
2913   case ARM::MVE_VLDRBU16_post:
2914   case ARM::MVE_VLDRBU32_post:
2915   case ARM::MVE_VLDRHS32_post:
2916   case ARM::MVE_VLDRHU32_post:
2917   case ARM::MVE_VLDRBU8_post:
2918   case ARM::MVE_VLDRHU16_post:
2919   case ARM::MVE_VLDRWU32_post:
2920   case ARM::MVE_VSTRB16_post:
2921   case ARM::MVE_VSTRB32_post:
2922   case ARM::MVE_VSTRH32_post:
2923   case ARM::MVE_VSTRBU8_post:
2924   case ARM::MVE_VSTRHU16_post:
2925   case ARM::MVE_VSTRWU32_post:
2926   case ARM::MVE_VLDRBS16_pre:
2927   case ARM::MVE_VLDRBS32_pre:
2928   case ARM::MVE_VLDRBU16_pre:
2929   case ARM::MVE_VLDRBU32_pre:
2930   case ARM::MVE_VLDRHS32_pre:
2931   case ARM::MVE_VLDRHU32_pre:
2932   case ARM::MVE_VLDRBU8_pre:
2933   case ARM::MVE_VLDRHU16_pre:
2934   case ARM::MVE_VLDRWU32_pre:
2935   case ARM::MVE_VSTRB16_pre:
2936   case ARM::MVE_VSTRB32_pre:
2937   case ARM::MVE_VSTRH32_pre:
2938   case ARM::MVE_VSTRBU8_pre:
2939   case ARM::MVE_VSTRHU16_pre:
2940   case ARM::MVE_VSTRWU32_pre:
2941     return 2;
2942   }
2943   return -1;
2944 }
2945 
isPostIndex(MachineInstr & MI)2946 static bool isPostIndex(MachineInstr &MI) {
2947   switch (MI.getOpcode()) {
2948   case ARM::MVE_VLDRBS16_post:
2949   case ARM::MVE_VLDRBS32_post:
2950   case ARM::MVE_VLDRBU16_post:
2951   case ARM::MVE_VLDRBU32_post:
2952   case ARM::MVE_VLDRHS32_post:
2953   case ARM::MVE_VLDRHU32_post:
2954   case ARM::MVE_VLDRBU8_post:
2955   case ARM::MVE_VLDRHU16_post:
2956   case ARM::MVE_VLDRWU32_post:
2957   case ARM::MVE_VSTRB16_post:
2958   case ARM::MVE_VSTRB32_post:
2959   case ARM::MVE_VSTRH32_post:
2960   case ARM::MVE_VSTRBU8_post:
2961   case ARM::MVE_VSTRHU16_post:
2962   case ARM::MVE_VSTRWU32_post:
2963     return true;
2964   }
2965   return false;
2966 }
2967 
isPreIndex(MachineInstr & MI)2968 static bool isPreIndex(MachineInstr &MI) {
2969   switch (MI.getOpcode()) {
2970   case ARM::MVE_VLDRBS16_pre:
2971   case ARM::MVE_VLDRBS32_pre:
2972   case ARM::MVE_VLDRBU16_pre:
2973   case ARM::MVE_VLDRBU32_pre:
2974   case ARM::MVE_VLDRHS32_pre:
2975   case ARM::MVE_VLDRHU32_pre:
2976   case ARM::MVE_VLDRBU8_pre:
2977   case ARM::MVE_VLDRHU16_pre:
2978   case ARM::MVE_VLDRWU32_pre:
2979   case ARM::MVE_VSTRB16_pre:
2980   case ARM::MVE_VSTRB32_pre:
2981   case ARM::MVE_VSTRH32_pre:
2982   case ARM::MVE_VSTRBU8_pre:
2983   case ARM::MVE_VSTRHU16_pre:
2984   case ARM::MVE_VSTRWU32_pre:
2985     return true;
2986   }
2987   return false;
2988 }
2989 
2990 // Given a memory access Opcode, check that the give Imm would be a valid Offset
2991 // for this instruction (same as isLegalAddressImm), Or if the instruction
2992 // could be easily converted to one where that was valid. For example converting
2993 // t2LDRi12 to t2LDRi8 for negative offsets. Works in conjunction with
2994 // AdjustBaseAndOffset below.
isLegalOrConvertableAddressImm(unsigned Opcode,int Imm,const TargetInstrInfo * TII,int & CodesizeEstimate)2995 static bool isLegalOrConvertableAddressImm(unsigned Opcode, int Imm,
2996                                            const TargetInstrInfo *TII,
2997                                            int &CodesizeEstimate) {
2998   if (isLegalAddressImm(Opcode, Imm, TII))
2999     return true;
3000 
3001   // We can convert AddrModeT2_i12 to AddrModeT2_i8neg.
3002   const MCInstrDesc &Desc = TII->get(Opcode);
3003   unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
3004   switch (AddrMode) {
3005   case ARMII::AddrModeT2_i12:
3006     CodesizeEstimate += 1;
3007     return Imm < 0 && -Imm < ((1 << 8) * 1);
3008   }
3009   return false;
3010 }
3011 
3012 // Given an MI adjust its address BaseReg to use NewBaseReg and address offset
3013 // by -Offset. This can either happen in-place or be a replacement as MI is
3014 // converted to another instruction type.
AdjustBaseAndOffset(MachineInstr * MI,Register NewBaseReg,int Offset,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI)3015 static void AdjustBaseAndOffset(MachineInstr *MI, Register NewBaseReg,
3016                                 int Offset, const TargetInstrInfo *TII,
3017                                 const TargetRegisterInfo *TRI) {
3018   // Set the Base reg
3019   unsigned BaseOp = getBaseOperandIndex(*MI);
3020   MI->getOperand(BaseOp).setReg(NewBaseReg);
3021   // and constrain the reg class to that required by the instruction.
3022   MachineFunction *MF = MI->getMF();
3023   MachineRegisterInfo &MRI = MF->getRegInfo();
3024   const MCInstrDesc &MCID = TII->get(MI->getOpcode());
3025   const TargetRegisterClass *TRC = TII->getRegClass(MCID, BaseOp, TRI, *MF);
3026   MRI.constrainRegClass(NewBaseReg, TRC);
3027 
3028   int OldOffset = MI->getOperand(BaseOp + 1).getImm();
3029   if (isLegalAddressImm(MI->getOpcode(), OldOffset - Offset, TII))
3030     MI->getOperand(BaseOp + 1).setImm(OldOffset - Offset);
3031   else {
3032     unsigned ConvOpcode;
3033     switch (MI->getOpcode()) {
3034     case ARM::t2LDRHi12:
3035       ConvOpcode = ARM::t2LDRHi8;
3036       break;
3037     case ARM::t2LDRSHi12:
3038       ConvOpcode = ARM::t2LDRSHi8;
3039       break;
3040     case ARM::t2LDRBi12:
3041       ConvOpcode = ARM::t2LDRBi8;
3042       break;
3043     case ARM::t2LDRSBi12:
3044       ConvOpcode = ARM::t2LDRSBi8;
3045       break;
3046     case ARM::t2STRHi12:
3047       ConvOpcode = ARM::t2STRHi8;
3048       break;
3049     case ARM::t2STRBi12:
3050       ConvOpcode = ARM::t2STRBi8;
3051       break;
3052     default:
3053       llvm_unreachable("Unhandled convertable opcode");
3054     }
3055     assert(isLegalAddressImm(ConvOpcode, OldOffset - Offset, TII) &&
3056            "Illegal Address Immediate after convert!");
3057 
3058     const MCInstrDesc &MCID = TII->get(ConvOpcode);
3059     BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), MCID)
3060         .add(MI->getOperand(0))
3061         .add(MI->getOperand(1))
3062         .addImm(OldOffset - Offset)
3063         .add(MI->getOperand(3))
3064         .add(MI->getOperand(4))
3065         .cloneMemRefs(*MI);
3066     MI->eraseFromParent();
3067   }
3068 }
3069 
createPostIncLoadStore(MachineInstr * MI,int Offset,Register NewReg,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI)3070 static MachineInstr *createPostIncLoadStore(MachineInstr *MI, int Offset,
3071                                             Register NewReg,
3072                                             const TargetInstrInfo *TII,
3073                                             const TargetRegisterInfo *TRI) {
3074   MachineFunction *MF = MI->getMF();
3075   MachineRegisterInfo &MRI = MF->getRegInfo();
3076 
3077   unsigned NewOpcode = getPostIndexedLoadStoreOpcode(
3078       MI->getOpcode(), Offset > 0 ? ARM_AM::add : ARM_AM::sub);
3079 
3080   const MCInstrDesc &MCID = TII->get(NewOpcode);
3081   // Constrain the def register class
3082   const TargetRegisterClass *TRC = TII->getRegClass(MCID, 0, TRI, *MF);
3083   MRI.constrainRegClass(NewReg, TRC);
3084   // And do the same for the base operand
3085   TRC = TII->getRegClass(MCID, 2, TRI, *MF);
3086   MRI.constrainRegClass(MI->getOperand(1).getReg(), TRC);
3087 
3088   unsigned AddrMode = (MCID.TSFlags & ARMII::AddrModeMask);
3089   switch (AddrMode) {
3090   case ARMII::AddrModeT2_i7:
3091   case ARMII::AddrModeT2_i7s2:
3092   case ARMII::AddrModeT2_i7s4:
3093     // Any MVE load/store
3094     return BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), MCID)
3095         .addReg(NewReg, RegState::Define)
3096         .add(MI->getOperand(0))
3097         .add(MI->getOperand(1))
3098         .addImm(Offset)
3099         .add(MI->getOperand(3))
3100         .add(MI->getOperand(4))
3101         .add(MI->getOperand(5))
3102         .cloneMemRefs(*MI);
3103   case ARMII::AddrModeT2_i8:
3104     if (MI->mayLoad()) {
3105       return BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), MCID)
3106           .add(MI->getOperand(0))
3107           .addReg(NewReg, RegState::Define)
3108           .add(MI->getOperand(1))
3109           .addImm(Offset)
3110           .add(MI->getOperand(3))
3111           .add(MI->getOperand(4))
3112           .cloneMemRefs(*MI);
3113     } else {
3114       return BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), MCID)
3115           .addReg(NewReg, RegState::Define)
3116           .add(MI->getOperand(0))
3117           .add(MI->getOperand(1))
3118           .addImm(Offset)
3119           .add(MI->getOperand(3))
3120           .add(MI->getOperand(4))
3121           .cloneMemRefs(*MI);
3122     }
3123   default:
3124     llvm_unreachable("Unhandled createPostIncLoadStore");
3125   }
3126 }
3127 
3128 // Given a Base Register, optimise the load/store uses to attempt to create more
3129 // post-inc accesses and less register moves. We do this by taking zero offset
3130 // loads/stores with an add, and convert them to a postinc load/store of the
3131 // same type. Any subsequent accesses will be adjusted to use and account for
3132 // the post-inc value.
3133 // For example:
3134 // LDR #0            LDR_POSTINC #16
3135 // LDR #4            LDR #-12
3136 // LDR #8            LDR #-8
3137 // LDR #12           LDR #-4
3138 // ADD #16
3139 //
3140 // At the same time if we do not find an increment but do find an existing
3141 // pre/post inc instruction, we can still adjust the offsets of subsequent
3142 // instructions to save the register move that would otherwise be needed for the
3143 // in-place increment.
DistributeIncrements(Register Base)3144 bool ARMPreAllocLoadStoreOpt::DistributeIncrements(Register Base) {
3145   // We are looking for:
3146   // One zero offset load/store that can become postinc
3147   MachineInstr *BaseAccess = nullptr;
3148   MachineInstr *PrePostInc = nullptr;
3149   // An increment that can be folded in
3150   MachineInstr *Increment = nullptr;
3151   // Other accesses after BaseAccess that will need to be updated to use the
3152   // postinc value.
3153   SmallPtrSet<MachineInstr *, 8> OtherAccesses;
3154   for (auto &Use : MRI->use_nodbg_instructions(Base)) {
3155     if (!Increment && getAddSubImmediate(Use) != 0) {
3156       Increment = &Use;
3157       continue;
3158     }
3159 
3160     int BaseOp = getBaseOperandIndex(Use);
3161     if (BaseOp == -1)
3162       return false;
3163 
3164     if (!Use.getOperand(BaseOp).isReg() ||
3165         Use.getOperand(BaseOp).getReg() != Base)
3166       return false;
3167     if (isPreIndex(Use) || isPostIndex(Use))
3168       PrePostInc = &Use;
3169     else if (Use.getOperand(BaseOp + 1).getImm() == 0)
3170       BaseAccess = &Use;
3171     else
3172       OtherAccesses.insert(&Use);
3173   }
3174 
3175   int IncrementOffset;
3176   Register NewBaseReg;
3177   if (BaseAccess && Increment) {
3178     if (PrePostInc || BaseAccess->getParent() != Increment->getParent())
3179       return false;
3180     Register PredReg;
3181     if (Increment->definesRegister(ARM::CPSR, /*TRI=*/nullptr) ||
3182         getInstrPredicate(*Increment, PredReg) != ARMCC::AL)
3183       return false;
3184 
3185     LLVM_DEBUG(dbgs() << "\nAttempting to distribute increments on VirtualReg "
3186                       << Base.virtRegIndex() << "\n");
3187 
3188     // Make sure that Increment has no uses before BaseAccess that are not PHI
3189     // uses.
3190     for (MachineInstr &Use :
3191         MRI->use_nodbg_instructions(Increment->getOperand(0).getReg())) {
3192       if (&Use == BaseAccess || (Use.getOpcode() != TargetOpcode::PHI &&
3193                                  !DT->dominates(BaseAccess, &Use))) {
3194         LLVM_DEBUG(dbgs() << "  BaseAccess doesn't dominate use of increment\n");
3195         return false;
3196       }
3197     }
3198 
3199     // Make sure that Increment can be folded into Base
3200     IncrementOffset = getAddSubImmediate(*Increment);
3201     unsigned NewPostIncOpcode = getPostIndexedLoadStoreOpcode(
3202         BaseAccess->getOpcode(), IncrementOffset > 0 ? ARM_AM::add : ARM_AM::sub);
3203     if (!isLegalAddressImm(NewPostIncOpcode, IncrementOffset, TII)) {
3204       LLVM_DEBUG(dbgs() << "  Illegal addressing mode immediate on postinc\n");
3205       return false;
3206     }
3207   }
3208   else if (PrePostInc) {
3209     // If we already have a pre/post index load/store then set BaseAccess,
3210     // IncrementOffset and NewBaseReg to the values it already produces,
3211     // allowing us to update and subsequent uses of BaseOp reg with the
3212     // incremented value.
3213     if (Increment)
3214       return false;
3215 
3216     LLVM_DEBUG(dbgs() << "\nAttempting to distribute increments on already "
3217                       << "indexed VirtualReg " << Base.virtRegIndex() << "\n");
3218     int BaseOp = getBaseOperandIndex(*PrePostInc);
3219     IncrementOffset = PrePostInc->getOperand(BaseOp+1).getImm();
3220     BaseAccess = PrePostInc;
3221     NewBaseReg = PrePostInc->getOperand(0).getReg();
3222   }
3223   else
3224     return false;
3225 
3226   // And make sure that the negative value of increment can be added to all
3227   // other offsets after the BaseAccess. We rely on either
3228   // dominates(BaseAccess, OtherAccess) or dominates(OtherAccess, BaseAccess)
3229   // to keep things simple.
3230   // This also adds a simple codesize metric, to detect if an instruction (like
3231   // t2LDRBi12) which can often be shrunk to a thumb1 instruction (tLDRBi)
3232   // cannot because it is converted to something else (t2LDRBi8). We start this
3233   // at -1 for the gain from removing the increment.
3234   SmallPtrSet<MachineInstr *, 4> SuccessorAccesses;
3235   int CodesizeEstimate = -1;
3236   for (auto *Use : OtherAccesses) {
3237     if (DT->dominates(BaseAccess, Use)) {
3238       SuccessorAccesses.insert(Use);
3239       unsigned BaseOp = getBaseOperandIndex(*Use);
3240       if (!isLegalOrConvertableAddressImm(Use->getOpcode(),
3241                                           Use->getOperand(BaseOp + 1).getImm() -
3242                                               IncrementOffset,
3243                                           TII, CodesizeEstimate)) {
3244         LLVM_DEBUG(dbgs() << "  Illegal addressing mode immediate on use\n");
3245         return false;
3246       }
3247     } else if (!DT->dominates(Use, BaseAccess)) {
3248       LLVM_DEBUG(
3249           dbgs() << "  Unknown dominance relation between Base and Use\n");
3250       return false;
3251     }
3252   }
3253   if (STI->hasMinSize() && CodesizeEstimate > 0) {
3254     LLVM_DEBUG(dbgs() << "  Expected to grow instructions under minsize\n");
3255     return false;
3256   }
3257 
3258   if (!PrePostInc) {
3259     // Replace BaseAccess with a post inc
3260     LLVM_DEBUG(dbgs() << "Changing: "; BaseAccess->dump());
3261     LLVM_DEBUG(dbgs() << "  And   : "; Increment->dump());
3262     NewBaseReg = Increment->getOperand(0).getReg();
3263     MachineInstr *BaseAccessPost =
3264         createPostIncLoadStore(BaseAccess, IncrementOffset, NewBaseReg, TII, TRI);
3265     BaseAccess->eraseFromParent();
3266     Increment->eraseFromParent();
3267     (void)BaseAccessPost;
3268     LLVM_DEBUG(dbgs() << "  To    : "; BaseAccessPost->dump());
3269   }
3270 
3271   for (auto *Use : SuccessorAccesses) {
3272     LLVM_DEBUG(dbgs() << "Changing: "; Use->dump());
3273     AdjustBaseAndOffset(Use, NewBaseReg, IncrementOffset, TII, TRI);
3274     LLVM_DEBUG(dbgs() << "  To    : "; Use->dump());
3275   }
3276 
3277   // Remove the kill flag from all uses of NewBaseReg, in case any old uses
3278   // remain.
3279   for (MachineOperand &Op : MRI->use_nodbg_operands(NewBaseReg))
3280     Op.setIsKill(false);
3281   return true;
3282 }
3283 
DistributeIncrements()3284 bool ARMPreAllocLoadStoreOpt::DistributeIncrements() {
3285   bool Changed = false;
3286   SmallSetVector<Register, 4> Visited;
3287   for (auto &MBB : *MF) {
3288     for (auto &MI : MBB) {
3289       int BaseOp = getBaseOperandIndex(MI);
3290       if (BaseOp == -1 || !MI.getOperand(BaseOp).isReg())
3291         continue;
3292 
3293       Register Base = MI.getOperand(BaseOp).getReg();
3294       if (!Base.isVirtual() || Visited.count(Base))
3295         continue;
3296 
3297       Visited.insert(Base);
3298     }
3299   }
3300 
3301   for (auto Base : Visited)
3302     Changed |= DistributeIncrements(Base);
3303 
3304   return Changed;
3305 }
3306 
3307 /// Returns an instance of the load / store optimization pass.
createARMLoadStoreOptimizationPass(bool PreAlloc)3308 FunctionPass *llvm::createARMLoadStoreOptimizationPass(bool PreAlloc) {
3309   if (PreAlloc)
3310     return new ARMPreAllocLoadStoreOpt();
3311   return new ARMLoadStoreOpt();
3312 }
3313