1 //===- X86OptimizeLEAs.cpp - optimize usage of LEA instructions -----------===//
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 // This file defines the pass that performs some optimizations with LEA
10 // instructions in order to improve performance and code size.
11 // Currently, it does two things:
12 // 1) If there are two LEA instructions calculating addresses which only differ
13 //    by displacement inside a basic block, one of them is removed.
14 // 2) Address calculations in load and store instructions are replaced by
15 //    existing LEA def registers where possible.
16 //
17 //===----------------------------------------------------------------------===//
18 
19 #include "MCTargetDesc/X86BaseInfo.h"
20 #include "X86.h"
21 #include "X86InstrInfo.h"
22 #include "X86Subtarget.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/DenseMapInfo.h"
25 #include "llvm/ADT/Hashing.h"
26 #include "llvm/ADT/SmallVector.h"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/Analysis/ProfileSummaryInfo.h"
29 #include "llvm/CodeGen/LazyMachineBlockFrequencyInfo.h"
30 #include "llvm/CodeGen/MachineBasicBlock.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineFunctionPass.h"
33 #include "llvm/CodeGen/MachineInstr.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineOperand.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/MachineSizeOpts.h"
38 #include "llvm/CodeGen/TargetOpcodes.h"
39 #include "llvm/CodeGen/TargetRegisterInfo.h"
40 #include "llvm/IR/DebugInfoMetadata.h"
41 #include "llvm/IR/DebugLoc.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/MC/MCInstrDesc.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include <cassert>
50 #include <cstdint>
51 #include <iterator>
52 
53 using namespace llvm;
54 
55 #define DEBUG_TYPE "x86-optimize-LEAs"
56 
57 static cl::opt<bool>
58     DisableX86LEAOpt("disable-x86-lea-opt", cl::Hidden,
59                      cl::desc("X86: Disable LEA optimizations."),
60                      cl::init(false));
61 
62 STATISTIC(NumSubstLEAs, "Number of LEA instruction substitutions");
63 STATISTIC(NumRedundantLEAs, "Number of redundant LEA instructions removed");
64 
65 /// Returns true if two machine operands are identical and they are not
66 /// physical registers.
67 static inline bool isIdenticalOp(const MachineOperand &MO1,
68                                  const MachineOperand &MO2);
69 
70 /// Returns true if two address displacement operands are of the same
71 /// type and use the same symbol/index/address regardless of the offset.
72 static bool isSimilarDispOp(const MachineOperand &MO1,
73                             const MachineOperand &MO2);
74 
75 /// Returns true if the instruction is LEA.
76 static inline bool isLEA(const MachineInstr &MI);
77 
78 namespace {
79 
80 /// A key based on instruction's memory operands.
81 class MemOpKey {
82 public:
MemOpKey(const MachineOperand * Base,const MachineOperand * Scale,const MachineOperand * Index,const MachineOperand * Segment,const MachineOperand * Disp)83   MemOpKey(const MachineOperand *Base, const MachineOperand *Scale,
84            const MachineOperand *Index, const MachineOperand *Segment,
85            const MachineOperand *Disp)
86       : Disp(Disp) {
87     Operands[0] = Base;
88     Operands[1] = Scale;
89     Operands[2] = Index;
90     Operands[3] = Segment;
91   }
92 
operator ==(const MemOpKey & Other) const93   bool operator==(const MemOpKey &Other) const {
94     // Addresses' bases, scales, indices and segments must be identical.
95     for (int i = 0; i < 4; ++i)
96       if (!isIdenticalOp(*Operands[i], *Other.Operands[i]))
97         return false;
98 
99     // Addresses' displacements don't have to be exactly the same. It only
100     // matters that they use the same symbol/index/address. Immediates' or
101     // offsets' differences will be taken care of during instruction
102     // substitution.
103     return isSimilarDispOp(*Disp, *Other.Disp);
104   }
105 
106   // Address' base, scale, index and segment operands.
107   const MachineOperand *Operands[4];
108 
109   // Address' displacement operand.
110   const MachineOperand *Disp;
111 };
112 
113 } // end anonymous namespace
114 
115 /// Provide DenseMapInfo for MemOpKey.
116 namespace llvm {
117 
118 template <> struct DenseMapInfo<MemOpKey> {
119   using PtrInfo = DenseMapInfo<const MachineOperand *>;
120 
getEmptyKeyllvm::DenseMapInfo121   static inline MemOpKey getEmptyKey() {
122     return MemOpKey(PtrInfo::getEmptyKey(), PtrInfo::getEmptyKey(),
123                     PtrInfo::getEmptyKey(), PtrInfo::getEmptyKey(),
124                     PtrInfo::getEmptyKey());
125   }
126 
getTombstoneKeyllvm::DenseMapInfo127   static inline MemOpKey getTombstoneKey() {
128     return MemOpKey(PtrInfo::getTombstoneKey(), PtrInfo::getTombstoneKey(),
129                     PtrInfo::getTombstoneKey(), PtrInfo::getTombstoneKey(),
130                     PtrInfo::getTombstoneKey());
131   }
132 
getHashValuellvm::DenseMapInfo133   static unsigned getHashValue(const MemOpKey &Val) {
134     // Checking any field of MemOpKey is enough to determine if the key is
135     // empty or tombstone.
136     assert(Val.Disp != PtrInfo::getEmptyKey() && "Cannot hash the empty key");
137     assert(Val.Disp != PtrInfo::getTombstoneKey() &&
138            "Cannot hash the tombstone key");
139 
140     hash_code Hash = hash_combine(*Val.Operands[0], *Val.Operands[1],
141                                   *Val.Operands[2], *Val.Operands[3]);
142 
143     // If the address displacement is an immediate, it should not affect the
144     // hash so that memory operands which differ only be immediate displacement
145     // would have the same hash. If the address displacement is something else,
146     // we should reflect symbol/index/address in the hash.
147     switch (Val.Disp->getType()) {
148     case MachineOperand::MO_Immediate:
149       break;
150     case MachineOperand::MO_ConstantPoolIndex:
151     case MachineOperand::MO_JumpTableIndex:
152       Hash = hash_combine(Hash, Val.Disp->getIndex());
153       break;
154     case MachineOperand::MO_ExternalSymbol:
155       Hash = hash_combine(Hash, Val.Disp->getSymbolName());
156       break;
157     case MachineOperand::MO_GlobalAddress:
158       Hash = hash_combine(Hash, Val.Disp->getGlobal());
159       break;
160     case MachineOperand::MO_BlockAddress:
161       Hash = hash_combine(Hash, Val.Disp->getBlockAddress());
162       break;
163     case MachineOperand::MO_MCSymbol:
164       Hash = hash_combine(Hash, Val.Disp->getMCSymbol());
165       break;
166     case MachineOperand::MO_MachineBasicBlock:
167       Hash = hash_combine(Hash, Val.Disp->getMBB());
168       break;
169     default:
170       llvm_unreachable("Invalid address displacement operand");
171     }
172 
173     return (unsigned)Hash;
174   }
175 
isEqualllvm::DenseMapInfo176   static bool isEqual(const MemOpKey &LHS, const MemOpKey &RHS) {
177     // Checking any field of MemOpKey is enough to determine if the key is
178     // empty or tombstone.
179     if (RHS.Disp == PtrInfo::getEmptyKey())
180       return LHS.Disp == PtrInfo::getEmptyKey();
181     if (RHS.Disp == PtrInfo::getTombstoneKey())
182       return LHS.Disp == PtrInfo::getTombstoneKey();
183     return LHS == RHS;
184   }
185 };
186 
187 } // end namespace llvm
188 
189 /// Returns a hash table key based on memory operands of \p MI. The
190 /// number of the first memory operand of \p MI is specified through \p N.
getMemOpKey(const MachineInstr & MI,unsigned N)191 static inline MemOpKey getMemOpKey(const MachineInstr &MI, unsigned N) {
192   assert((isLEA(MI) || MI.mayLoadOrStore()) &&
193          "The instruction must be a LEA, a load or a store");
194   return MemOpKey(&MI.getOperand(N + X86::AddrBaseReg),
195                   &MI.getOperand(N + X86::AddrScaleAmt),
196                   &MI.getOperand(N + X86::AddrIndexReg),
197                   &MI.getOperand(N + X86::AddrSegmentReg),
198                   &MI.getOperand(N + X86::AddrDisp));
199 }
200 
isIdenticalOp(const MachineOperand & MO1,const MachineOperand & MO2)201 static inline bool isIdenticalOp(const MachineOperand &MO1,
202                                  const MachineOperand &MO2) {
203   return MO1.isIdenticalTo(MO2) &&
204          (!MO1.isReg() || !Register::isPhysicalRegister(MO1.getReg()));
205 }
206 
207 #ifndef NDEBUG
isValidDispOp(const MachineOperand & MO)208 static bool isValidDispOp(const MachineOperand &MO) {
209   return MO.isImm() || MO.isCPI() || MO.isJTI() || MO.isSymbol() ||
210          MO.isGlobal() || MO.isBlockAddress() || MO.isMCSymbol() || MO.isMBB();
211 }
212 #endif
213 
isSimilarDispOp(const MachineOperand & MO1,const MachineOperand & MO2)214 static bool isSimilarDispOp(const MachineOperand &MO1,
215                             const MachineOperand &MO2) {
216   assert(isValidDispOp(MO1) && isValidDispOp(MO2) &&
217          "Address displacement operand is not valid");
218   return (MO1.isImm() && MO2.isImm()) ||
219          (MO1.isCPI() && MO2.isCPI() && MO1.getIndex() == MO2.getIndex()) ||
220          (MO1.isJTI() && MO2.isJTI() && MO1.getIndex() == MO2.getIndex()) ||
221          (MO1.isSymbol() && MO2.isSymbol() &&
222           MO1.getSymbolName() == MO2.getSymbolName()) ||
223          (MO1.isGlobal() && MO2.isGlobal() &&
224           MO1.getGlobal() == MO2.getGlobal()) ||
225          (MO1.isBlockAddress() && MO2.isBlockAddress() &&
226           MO1.getBlockAddress() == MO2.getBlockAddress()) ||
227          (MO1.isMCSymbol() && MO2.isMCSymbol() &&
228           MO1.getMCSymbol() == MO2.getMCSymbol()) ||
229          (MO1.isMBB() && MO2.isMBB() && MO1.getMBB() == MO2.getMBB());
230 }
231 
isLEA(const MachineInstr & MI)232 static inline bool isLEA(const MachineInstr &MI) {
233   unsigned Opcode = MI.getOpcode();
234   return Opcode == X86::LEA16r || Opcode == X86::LEA32r ||
235          Opcode == X86::LEA64r || Opcode == X86::LEA64_32r;
236 }
237 
238 namespace {
239 
240 class X86OptimizeLEAPass : public MachineFunctionPass {
241 public:
X86OptimizeLEAPass()242   X86OptimizeLEAPass() : MachineFunctionPass(ID) {}
243 
getPassName() const244   StringRef getPassName() const override { return "X86 LEA Optimize"; }
245 
246   /// Loop over all of the basic blocks, replacing address
247   /// calculations in load and store instructions, if it's already
248   /// been calculated by LEA. Also, remove redundant LEAs.
249   bool runOnMachineFunction(MachineFunction &MF) override;
250 
251   static char ID;
252 
getAnalysisUsage(AnalysisUsage & AU) const253   void getAnalysisUsage(AnalysisUsage &AU) const override {
254     AU.addRequired<ProfileSummaryInfoWrapperPass>();
255     AU.addRequired<LazyMachineBlockFrequencyInfoPass>();
256     MachineFunctionPass::getAnalysisUsage(AU);
257   }
258 
259 private:
260   using MemOpMap = DenseMap<MemOpKey, SmallVector<MachineInstr *, 16>>;
261 
262   /// Returns a distance between two instructions inside one basic block.
263   /// Negative result means, that instructions occur in reverse order.
264   int calcInstrDist(const MachineInstr &First, const MachineInstr &Last);
265 
266   /// Choose the best \p LEA instruction from the \p List to replace
267   /// address calculation in \p MI instruction. Return the address displacement
268   /// and the distance between \p MI and the chosen \p BestLEA in
269   /// \p AddrDispShift and \p Dist.
270   bool chooseBestLEA(const SmallVectorImpl<MachineInstr *> &List,
271                      const MachineInstr &MI, MachineInstr *&BestLEA,
272                      int64_t &AddrDispShift, int &Dist);
273 
274   /// Returns the difference between addresses' displacements of \p MI1
275   /// and \p MI2. The numbers of the first memory operands for the instructions
276   /// are specified through \p N1 and \p N2.
277   int64_t getAddrDispShift(const MachineInstr &MI1, unsigned N1,
278                            const MachineInstr &MI2, unsigned N2) const;
279 
280   /// Returns true if the \p Last LEA instruction can be replaced by the
281   /// \p First. The difference between displacements of the addresses calculated
282   /// by these LEAs is returned in \p AddrDispShift. It'll be used for proper
283   /// replacement of the \p Last LEA's uses with the \p First's def register.
284   bool isReplaceable(const MachineInstr &First, const MachineInstr &Last,
285                      int64_t &AddrDispShift) const;
286 
287   /// Find all LEA instructions in the basic block. Also, assign position
288   /// numbers to all instructions in the basic block to speed up calculation of
289   /// distance between them.
290   void findLEAs(const MachineBasicBlock &MBB, MemOpMap &LEAs);
291 
292   /// Removes redundant address calculations.
293   bool removeRedundantAddrCalc(MemOpMap &LEAs);
294 
295   /// Replace debug value MI with a new debug value instruction using register
296   /// VReg with an appropriate offset and DIExpression to incorporate the
297   /// address displacement AddrDispShift. Return new debug value instruction.
298   MachineInstr *replaceDebugValue(MachineInstr &MI, unsigned VReg,
299                                   int64_t AddrDispShift);
300 
301   /// Removes LEAs which calculate similar addresses.
302   bool removeRedundantLEAs(MemOpMap &LEAs);
303 
304   DenseMap<const MachineInstr *, unsigned> InstrPos;
305 
306   MachineRegisterInfo *MRI = nullptr;
307   const X86InstrInfo *TII = nullptr;
308   const X86RegisterInfo *TRI = nullptr;
309 };
310 
311 } // end anonymous namespace
312 
313 char X86OptimizeLEAPass::ID = 0;
314 
createX86OptimizeLEAs()315 FunctionPass *llvm::createX86OptimizeLEAs() { return new X86OptimizeLEAPass(); }
316 INITIALIZE_PASS(X86OptimizeLEAPass, DEBUG_TYPE, "X86 optimize LEA pass", false,
317                 false)
318 
calcInstrDist(const MachineInstr & First,const MachineInstr & Last)319 int X86OptimizeLEAPass::calcInstrDist(const MachineInstr &First,
320                                       const MachineInstr &Last) {
321   // Both instructions must be in the same basic block and they must be
322   // presented in InstrPos.
323   assert(Last.getParent() == First.getParent() &&
324          "Instructions are in different basic blocks");
325   assert(InstrPos.find(&First) != InstrPos.end() &&
326          InstrPos.find(&Last) != InstrPos.end() &&
327          "Instructions' positions are undefined");
328 
329   return InstrPos[&Last] - InstrPos[&First];
330 }
331 
332 // Find the best LEA instruction in the List to replace address recalculation in
333 // MI. Such LEA must meet these requirements:
334 // 1) The address calculated by the LEA differs only by the displacement from
335 //    the address used in MI.
336 // 2) The register class of the definition of the LEA is compatible with the
337 //    register class of the address base register of MI.
338 // 3) Displacement of the new memory operand should fit in 1 byte if possible.
339 // 4) The LEA should be as close to MI as possible, and prior to it if
340 //    possible.
chooseBestLEA(const SmallVectorImpl<MachineInstr * > & List,const MachineInstr & MI,MachineInstr * & BestLEA,int64_t & AddrDispShift,int & Dist)341 bool X86OptimizeLEAPass::chooseBestLEA(
342     const SmallVectorImpl<MachineInstr *> &List, const MachineInstr &MI,
343     MachineInstr *&BestLEA, int64_t &AddrDispShift, int &Dist) {
344   const MachineFunction *MF = MI.getParent()->getParent();
345   const MCInstrDesc &Desc = MI.getDesc();
346   int MemOpNo = X86II::getMemoryOperandNo(Desc.TSFlags) +
347                 X86II::getOperandBias(Desc);
348 
349   BestLEA = nullptr;
350 
351   // Loop over all LEA instructions.
352   for (auto DefMI : List) {
353     // Get new address displacement.
354     int64_t AddrDispShiftTemp = getAddrDispShift(MI, MemOpNo, *DefMI, 1);
355 
356     // Make sure address displacement fits 4 bytes.
357     if (!isInt<32>(AddrDispShiftTemp))
358       continue;
359 
360     // Check that LEA def register can be used as MI address base. Some
361     // instructions can use a limited set of registers as address base, for
362     // example MOV8mr_NOREX. We could constrain the register class of the LEA
363     // def to suit MI, however since this case is very rare and hard to
364     // reproduce in a test it's just more reliable to skip the LEA.
365     if (TII->getRegClass(Desc, MemOpNo + X86::AddrBaseReg, TRI, *MF) !=
366         MRI->getRegClass(DefMI->getOperand(0).getReg()))
367       continue;
368 
369     // Choose the closest LEA instruction from the list, prior to MI if
370     // possible. Note that we took into account resulting address displacement
371     // as well. Also note that the list is sorted by the order in which the LEAs
372     // occur, so the break condition is pretty simple.
373     int DistTemp = calcInstrDist(*DefMI, MI);
374     assert(DistTemp != 0 &&
375            "The distance between two different instructions cannot be zero");
376     if (DistTemp > 0 || BestLEA == nullptr) {
377       // Do not update return LEA, if the current one provides a displacement
378       // which fits in 1 byte, while the new candidate does not.
379       if (BestLEA != nullptr && !isInt<8>(AddrDispShiftTemp) &&
380           isInt<8>(AddrDispShift))
381         continue;
382 
383       BestLEA = DefMI;
384       AddrDispShift = AddrDispShiftTemp;
385       Dist = DistTemp;
386     }
387 
388     // FIXME: Maybe we should not always stop at the first LEA after MI.
389     if (DistTemp < 0)
390       break;
391   }
392 
393   return BestLEA != nullptr;
394 }
395 
396 // Get the difference between the addresses' displacements of the two
397 // instructions \p MI1 and \p MI2. The numbers of the first memory operands are
398 // passed through \p N1 and \p N2.
getAddrDispShift(const MachineInstr & MI1,unsigned N1,const MachineInstr & MI2,unsigned N2) const399 int64_t X86OptimizeLEAPass::getAddrDispShift(const MachineInstr &MI1,
400                                              unsigned N1,
401                                              const MachineInstr &MI2,
402                                              unsigned N2) const {
403   const MachineOperand &Op1 = MI1.getOperand(N1 + X86::AddrDisp);
404   const MachineOperand &Op2 = MI2.getOperand(N2 + X86::AddrDisp);
405 
406   assert(isSimilarDispOp(Op1, Op2) &&
407          "Address displacement operands are not compatible");
408 
409   // After the assert above we can be sure that both operands are of the same
410   // valid type and use the same symbol/index/address, thus displacement shift
411   // calculation is rather simple.
412   if (Op1.isJTI())
413     return 0;
414   return Op1.isImm() ? Op1.getImm() - Op2.getImm()
415                      : Op1.getOffset() - Op2.getOffset();
416 }
417 
418 // Check that the Last LEA can be replaced by the First LEA. To be so,
419 // these requirements must be met:
420 // 1) Addresses calculated by LEAs differ only by displacement.
421 // 2) Def registers of LEAs belong to the same class.
422 // 3) All uses of the Last LEA def register are replaceable, thus the
423 //    register is used only as address base.
isReplaceable(const MachineInstr & First,const MachineInstr & Last,int64_t & AddrDispShift) const424 bool X86OptimizeLEAPass::isReplaceable(const MachineInstr &First,
425                                        const MachineInstr &Last,
426                                        int64_t &AddrDispShift) const {
427   assert(isLEA(First) && isLEA(Last) &&
428          "The function works only with LEA instructions");
429 
430   // Make sure that LEA def registers belong to the same class. There may be
431   // instructions (like MOV8mr_NOREX) which allow a limited set of registers to
432   // be used as their operands, so we must be sure that replacing one LEA
433   // with another won't lead to putting a wrong register in the instruction.
434   if (MRI->getRegClass(First.getOperand(0).getReg()) !=
435       MRI->getRegClass(Last.getOperand(0).getReg()))
436     return false;
437 
438   // Get new address displacement.
439   AddrDispShift = getAddrDispShift(Last, 1, First, 1);
440 
441   // Loop over all uses of the Last LEA to check that its def register is
442   // used only as address base for memory accesses. If so, it can be
443   // replaced, otherwise - no.
444   for (auto &MO : MRI->use_nodbg_operands(Last.getOperand(0).getReg())) {
445     MachineInstr &MI = *MO.getParent();
446 
447     // Get the number of the first memory operand.
448     const MCInstrDesc &Desc = MI.getDesc();
449     int MemOpNo = X86II::getMemoryOperandNo(Desc.TSFlags);
450 
451     // If the use instruction has no memory operand - the LEA is not
452     // replaceable.
453     if (MemOpNo < 0)
454       return false;
455 
456     MemOpNo += X86II::getOperandBias(Desc);
457 
458     // If the address base of the use instruction is not the LEA def register -
459     // the LEA is not replaceable.
460     if (!isIdenticalOp(MI.getOperand(MemOpNo + X86::AddrBaseReg), MO))
461       return false;
462 
463     // If the LEA def register is used as any other operand of the use
464     // instruction - the LEA is not replaceable.
465     for (unsigned i = 0; i < MI.getNumOperands(); i++)
466       if (i != (unsigned)(MemOpNo + X86::AddrBaseReg) &&
467           isIdenticalOp(MI.getOperand(i), MO))
468         return false;
469 
470     // Check that the new address displacement will fit 4 bytes.
471     if (MI.getOperand(MemOpNo + X86::AddrDisp).isImm() &&
472         !isInt<32>(MI.getOperand(MemOpNo + X86::AddrDisp).getImm() +
473                    AddrDispShift))
474       return false;
475   }
476 
477   return true;
478 }
479 
findLEAs(const MachineBasicBlock & MBB,MemOpMap & LEAs)480 void X86OptimizeLEAPass::findLEAs(const MachineBasicBlock &MBB,
481                                   MemOpMap &LEAs) {
482   unsigned Pos = 0;
483   for (auto &MI : MBB) {
484     // Assign the position number to the instruction. Note that we are going to
485     // move some instructions during the optimization however there will never
486     // be a need to move two instructions before any selected instruction. So to
487     // avoid multiple positions' updates during moves we just increase position
488     // counter by two leaving a free space for instructions which will be moved.
489     InstrPos[&MI] = Pos += 2;
490 
491     if (isLEA(MI))
492       LEAs[getMemOpKey(MI, 1)].push_back(const_cast<MachineInstr *>(&MI));
493   }
494 }
495 
496 // Try to find load and store instructions which recalculate addresses already
497 // calculated by some LEA and replace their memory operands with its def
498 // register.
removeRedundantAddrCalc(MemOpMap & LEAs)499 bool X86OptimizeLEAPass::removeRedundantAddrCalc(MemOpMap &LEAs) {
500   bool Changed = false;
501 
502   assert(!LEAs.empty());
503   MachineBasicBlock *MBB = (*LEAs.begin()->second.begin())->getParent();
504 
505   // Process all instructions in basic block.
506   for (auto I = MBB->begin(), E = MBB->end(); I != E;) {
507     MachineInstr &MI = *I++;
508 
509     // Instruction must be load or store.
510     if (!MI.mayLoadOrStore())
511       continue;
512 
513     // Get the number of the first memory operand.
514     const MCInstrDesc &Desc = MI.getDesc();
515     int MemOpNo = X86II::getMemoryOperandNo(Desc.TSFlags);
516 
517     // If instruction has no memory operand - skip it.
518     if (MemOpNo < 0)
519       continue;
520 
521     MemOpNo += X86II::getOperandBias(Desc);
522 
523     // Do not call chooseBestLEA if there was no matching LEA
524     auto Insns = LEAs.find(getMemOpKey(MI, MemOpNo));
525     if (Insns == LEAs.end())
526       continue;
527 
528     // Get the best LEA instruction to replace address calculation.
529     MachineInstr *DefMI;
530     int64_t AddrDispShift;
531     int Dist;
532     if (!chooseBestLEA(Insns->second, MI, DefMI, AddrDispShift, Dist))
533       continue;
534 
535     // If LEA occurs before current instruction, we can freely replace
536     // the instruction. If LEA occurs after, we can lift LEA above the
537     // instruction and this way to be able to replace it. Since LEA and the
538     // instruction have similar memory operands (thus, the same def
539     // instructions for these operands), we can always do that, without
540     // worries of using registers before their defs.
541     if (Dist < 0) {
542       DefMI->removeFromParent();
543       MBB->insert(MachineBasicBlock::iterator(&MI), DefMI);
544       InstrPos[DefMI] = InstrPos[&MI] - 1;
545 
546       // Make sure the instructions' position numbers are sane.
547       assert(((InstrPos[DefMI] == 1 &&
548                MachineBasicBlock::iterator(DefMI) == MBB->begin()) ||
549               InstrPos[DefMI] >
550                   InstrPos[&*std::prev(MachineBasicBlock::iterator(DefMI))]) &&
551              "Instruction positioning is broken");
552     }
553 
554     // Since we can possibly extend register lifetime, clear kill flags.
555     MRI->clearKillFlags(DefMI->getOperand(0).getReg());
556 
557     ++NumSubstLEAs;
558     LLVM_DEBUG(dbgs() << "OptimizeLEAs: Candidate to replace: "; MI.dump(););
559 
560     // Change instruction operands.
561     MI.getOperand(MemOpNo + X86::AddrBaseReg)
562         .ChangeToRegister(DefMI->getOperand(0).getReg(), false);
563     MI.getOperand(MemOpNo + X86::AddrScaleAmt).ChangeToImmediate(1);
564     MI.getOperand(MemOpNo + X86::AddrIndexReg)
565         .ChangeToRegister(X86::NoRegister, false);
566     MI.getOperand(MemOpNo + X86::AddrDisp).ChangeToImmediate(AddrDispShift);
567     MI.getOperand(MemOpNo + X86::AddrSegmentReg)
568         .ChangeToRegister(X86::NoRegister, false);
569 
570     LLVM_DEBUG(dbgs() << "OptimizeLEAs: Replaced by: "; MI.dump(););
571 
572     Changed = true;
573   }
574 
575   return Changed;
576 }
577 
replaceDebugValue(MachineInstr & MI,unsigned VReg,int64_t AddrDispShift)578 MachineInstr *X86OptimizeLEAPass::replaceDebugValue(MachineInstr &MI,
579                                                     unsigned VReg,
580                                                     int64_t AddrDispShift) {
581   const DIExpression *Expr = MI.getDebugExpression();
582   if (AddrDispShift != 0)
583     Expr = DIExpression::prepend(Expr, DIExpression::StackValue, AddrDispShift);
584 
585   // Replace DBG_VALUE instruction with modified version.
586   MachineBasicBlock *MBB = MI.getParent();
587   DebugLoc DL = MI.getDebugLoc();
588   bool IsIndirect = MI.isIndirectDebugValue();
589   const MDNode *Var = MI.getDebugVariable();
590   if (IsIndirect)
591     assert(MI.getOperand(1).getImm() == 0 && "DBG_VALUE with nonzero offset");
592   return BuildMI(*MBB, MBB->erase(&MI), DL, TII->get(TargetOpcode::DBG_VALUE),
593                  IsIndirect, VReg, Var, Expr);
594 }
595 
596 // Try to find similar LEAs in the list and replace one with another.
removeRedundantLEAs(MemOpMap & LEAs)597 bool X86OptimizeLEAPass::removeRedundantLEAs(MemOpMap &LEAs) {
598   bool Changed = false;
599 
600   // Loop over all entries in the table.
601   for (auto &E : LEAs) {
602     auto &List = E.second;
603 
604     // Loop over all LEA pairs.
605     auto I1 = List.begin();
606     while (I1 != List.end()) {
607       MachineInstr &First = **I1;
608       auto I2 = std::next(I1);
609       while (I2 != List.end()) {
610         MachineInstr &Last = **I2;
611         int64_t AddrDispShift;
612 
613         // LEAs should be in occurrence order in the list, so we can freely
614         // replace later LEAs with earlier ones.
615         assert(calcInstrDist(First, Last) > 0 &&
616                "LEAs must be in occurrence order in the list");
617 
618         // Check that the Last LEA instruction can be replaced by the First.
619         if (!isReplaceable(First, Last, AddrDispShift)) {
620           ++I2;
621           continue;
622         }
623 
624         // Loop over all uses of the Last LEA and update their operands. Note
625         // that the correctness of this has already been checked in the
626         // isReplaceable function.
627         Register FirstVReg = First.getOperand(0).getReg();
628         Register LastVReg = Last.getOperand(0).getReg();
629         for (auto UI = MRI->use_begin(LastVReg), UE = MRI->use_end();
630              UI != UE;) {
631           MachineOperand &MO = *UI++;
632           MachineInstr &MI = *MO.getParent();
633 
634           if (MI.isDebugValue()) {
635             // Replace DBG_VALUE instruction with modified version using the
636             // register from the replacing LEA and the address displacement
637             // between the LEA instructions.
638             replaceDebugValue(MI, FirstVReg, AddrDispShift);
639             continue;
640           }
641 
642           // Get the number of the first memory operand.
643           const MCInstrDesc &Desc = MI.getDesc();
644           int MemOpNo =
645               X86II::getMemoryOperandNo(Desc.TSFlags) +
646               X86II::getOperandBias(Desc);
647 
648           // Update address base.
649           MO.setReg(FirstVReg);
650 
651           // Update address disp.
652           MachineOperand &Op = MI.getOperand(MemOpNo + X86::AddrDisp);
653           if (Op.isImm())
654             Op.setImm(Op.getImm() + AddrDispShift);
655           else if (!Op.isJTI())
656             Op.setOffset(Op.getOffset() + AddrDispShift);
657         }
658 
659         // Since we can possibly extend register lifetime, clear kill flags.
660         MRI->clearKillFlags(FirstVReg);
661 
662         ++NumRedundantLEAs;
663         LLVM_DEBUG(dbgs() << "OptimizeLEAs: Remove redundant LEA: ";
664                    Last.dump(););
665 
666         // By this moment, all of the Last LEA's uses must be replaced. So we
667         // can freely remove it.
668         assert(MRI->use_empty(LastVReg) &&
669                "The LEA's def register must have no uses");
670         Last.eraseFromParent();
671 
672         // Erase removed LEA from the list.
673         I2 = List.erase(I2);
674 
675         Changed = true;
676       }
677       ++I1;
678     }
679   }
680 
681   return Changed;
682 }
683 
runOnMachineFunction(MachineFunction & MF)684 bool X86OptimizeLEAPass::runOnMachineFunction(MachineFunction &MF) {
685   bool Changed = false;
686 
687   if (DisableX86LEAOpt || skipFunction(MF.getFunction()))
688     return false;
689 
690   MRI = &MF.getRegInfo();
691   TII = MF.getSubtarget<X86Subtarget>().getInstrInfo();
692   TRI = MF.getSubtarget<X86Subtarget>().getRegisterInfo();
693   auto *PSI =
694       &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
695   auto *MBFI = (PSI && PSI->hasProfileSummary()) ?
696                &getAnalysis<LazyMachineBlockFrequencyInfoPass>().getBFI() :
697                nullptr;
698 
699   // Process all basic blocks.
700   for (auto &MBB : MF) {
701     MemOpMap LEAs;
702     InstrPos.clear();
703 
704     // Find all LEA instructions in basic block.
705     findLEAs(MBB, LEAs);
706 
707     // If current basic block has no LEAs, move on to the next one.
708     if (LEAs.empty())
709       continue;
710 
711     // Remove redundant LEA instructions.
712     Changed |= removeRedundantLEAs(LEAs);
713 
714     // Remove redundant address calculations. Do it only for -Os/-Oz since only
715     // a code size gain is expected from this part of the pass.
716     bool OptForSize = MF.getFunction().hasOptSize() ||
717                       llvm::shouldOptimizeForSize(&MBB, PSI, MBFI);
718     if (OptForSize)
719       Changed |= removeRedundantAddrCalc(LEAs);
720   }
721 
722   return Changed;
723 }
724