1 //===- EarlyCSE.cpp - Simple and fast CSE 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 // This pass performs a simple dominator tree walk that eliminates trivially
10 // redundant instructions.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Transforms/Scalar/EarlyCSE.h"
15 #include "llvm/ADT/DenseMapInfo.h"
16 #include "llvm/ADT/Hashing.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/ScopedHashTable.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AssumptionCache.h"
22 #include "llvm/Analysis/GlobalsModRef.h"
23 #include "llvm/Analysis/GuardUtils.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/MemorySSA.h"
26 #include "llvm/Analysis/MemorySSAUpdater.h"
27 #include "llvm/Analysis/TargetLibraryInfo.h"
28 #include "llvm/Analysis/TargetTransformInfo.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/InstrTypes.h"
35 #include "llvm/IR/Instruction.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/LLVMContext.h"
39 #include "llvm/IR/PassManager.h"
40 #include "llvm/IR/PatternMatch.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/IR/Value.h"
43 #include "llvm/InitializePasses.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/Allocator.h"
46 #include "llvm/Support/AtomicOrdering.h"
47 #include "llvm/Support/Casting.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/DebugCounter.h"
50 #include "llvm/Support/RecyclingAllocator.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include "llvm/Transforms/Scalar.h"
53 #include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
54 #include "llvm/Transforms/Utils/Local.h"
55 #include <cassert>
56 #include <deque>
57 #include <memory>
58 #include <utility>
59 
60 using namespace llvm;
61 using namespace llvm::PatternMatch;
62 
63 #define DEBUG_TYPE "early-cse"
64 
65 STATISTIC(NumSimplify, "Number of instructions simplified or DCE'd");
66 STATISTIC(NumCSE,      "Number of instructions CSE'd");
67 STATISTIC(NumCSECVP,   "Number of compare instructions CVP'd");
68 STATISTIC(NumCSELoad,  "Number of load instructions CSE'd");
69 STATISTIC(NumCSECall,  "Number of call instructions CSE'd");
70 STATISTIC(NumDSE,      "Number of trivial dead stores removed");
71 
72 DEBUG_COUNTER(CSECounter, "early-cse",
73               "Controls which instructions are removed");
74 
75 static cl::opt<unsigned> EarlyCSEMssaOptCap(
76     "earlycse-mssa-optimization-cap", cl::init(500), cl::Hidden,
77     cl::desc("Enable imprecision in EarlyCSE in pathological cases, in exchange "
78              "for faster compile. Caps the MemorySSA clobbering calls."));
79 
80 static cl::opt<bool> EarlyCSEDebugHash(
81     "earlycse-debug-hash", cl::init(false), cl::Hidden,
82     cl::desc("Perform extra assertion checking to verify that SimpleValue's hash "
83              "function is well-behaved w.r.t. its isEqual predicate"));
84 
85 //===----------------------------------------------------------------------===//
86 // SimpleValue
87 //===----------------------------------------------------------------------===//
88 
89 namespace {
90 
91 /// Struct representing the available values in the scoped hash table.
92 struct SimpleValue {
93   Instruction *Inst;
94 
95   SimpleValue(Instruction *I) : Inst(I) {
96     assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
97   }
98 
99   bool isSentinel() const {
100     return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
101            Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
102   }
103 
104   static bool canHandle(Instruction *Inst) {
105     // This can only handle non-void readnone functions.
106     // Also handled are constrained intrinsic that look like the types
107     // of instruction handled below (UnaryOperator, etc.).
108     if (CallInst *CI = dyn_cast<CallInst>(Inst)) {
109       if (Function *F = CI->getCalledFunction()) {
110         switch ((Intrinsic::ID)F->getIntrinsicID()) {
111         case Intrinsic::experimental_constrained_fadd:
112         case Intrinsic::experimental_constrained_fsub:
113         case Intrinsic::experimental_constrained_fmul:
114         case Intrinsic::experimental_constrained_fdiv:
115         case Intrinsic::experimental_constrained_frem:
116         case Intrinsic::experimental_constrained_fptosi:
117         case Intrinsic::experimental_constrained_sitofp:
118         case Intrinsic::experimental_constrained_fptoui:
119         case Intrinsic::experimental_constrained_uitofp:
120         case Intrinsic::experimental_constrained_fcmp:
121         case Intrinsic::experimental_constrained_fcmps: {
122           auto *CFP = cast<ConstrainedFPIntrinsic>(CI);
123           return CFP->isDefaultFPEnvironment();
124         }
125         }
126       }
127       return CI->doesNotAccessMemory() && !CI->getType()->isVoidTy();
128     }
129     return isa<CastInst>(Inst) || isa<UnaryOperator>(Inst) ||
130            isa<BinaryOperator>(Inst) || isa<GetElementPtrInst>(Inst) ||
131            isa<CmpInst>(Inst) || isa<SelectInst>(Inst) ||
132            isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) ||
133            isa<ShuffleVectorInst>(Inst) || isa<ExtractValueInst>(Inst) ||
134            isa<InsertValueInst>(Inst) || isa<FreezeInst>(Inst);
135   }
136 };
137 
138 } // end anonymous namespace
139 
140 namespace llvm {
141 
142 template <> struct DenseMapInfo<SimpleValue> {
143   static inline SimpleValue getEmptyKey() {
144     return DenseMapInfo<Instruction *>::getEmptyKey();
145   }
146 
147   static inline SimpleValue getTombstoneKey() {
148     return DenseMapInfo<Instruction *>::getTombstoneKey();
149   }
150 
151   static unsigned getHashValue(SimpleValue Val);
152   static bool isEqual(SimpleValue LHS, SimpleValue RHS);
153 };
154 
155 } // end namespace llvm
156 
157 /// Match a 'select' including an optional 'not's of the condition.
158 static bool matchSelectWithOptionalNotCond(Value *V, Value *&Cond, Value *&A,
159                                            Value *&B,
160                                            SelectPatternFlavor &Flavor) {
161   // Return false if V is not even a select.
162   if (!match(V, m_Select(m_Value(Cond), m_Value(A), m_Value(B))))
163     return false;
164 
165   // Look through a 'not' of the condition operand by swapping A/B.
166   Value *CondNot;
167   if (match(Cond, m_Not(m_Value(CondNot)))) {
168     Cond = CondNot;
169     std::swap(A, B);
170   }
171 
172   // Match canonical forms of min/max. We are not using ValueTracking's
173   // more powerful matchSelectPattern() because it may rely on instruction flags
174   // such as "nsw". That would be incompatible with the current hashing
175   // mechanism that may remove flags to increase the likelihood of CSE.
176 
177   Flavor = SPF_UNKNOWN;
178   CmpInst::Predicate Pred;
179 
180   if (!match(Cond, m_ICmp(Pred, m_Specific(A), m_Specific(B)))) {
181     // Check for commuted variants of min/max by swapping predicate.
182     // If we do not match the standard or commuted patterns, this is not a
183     // recognized form of min/max, but it is still a select, so return true.
184     if (!match(Cond, m_ICmp(Pred, m_Specific(B), m_Specific(A))))
185       return true;
186     Pred = ICmpInst::getSwappedPredicate(Pred);
187   }
188 
189   switch (Pred) {
190   case CmpInst::ICMP_UGT: Flavor = SPF_UMAX; break;
191   case CmpInst::ICMP_ULT: Flavor = SPF_UMIN; break;
192   case CmpInst::ICMP_SGT: Flavor = SPF_SMAX; break;
193   case CmpInst::ICMP_SLT: Flavor = SPF_SMIN; break;
194   // Non-strict inequalities.
195   case CmpInst::ICMP_ULE: Flavor = SPF_UMIN; break;
196   case CmpInst::ICMP_UGE: Flavor = SPF_UMAX; break;
197   case CmpInst::ICMP_SLE: Flavor = SPF_SMIN; break;
198   case CmpInst::ICMP_SGE: Flavor = SPF_SMAX; break;
199   default: break;
200   }
201 
202   return true;
203 }
204 
205 static unsigned getHashValueImpl(SimpleValue Val) {
206   Instruction *Inst = Val.Inst;
207   // Hash in all of the operands as pointers.
208   if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst)) {
209     Value *LHS = BinOp->getOperand(0);
210     Value *RHS = BinOp->getOperand(1);
211     if (BinOp->isCommutative() && BinOp->getOperand(0) > BinOp->getOperand(1))
212       std::swap(LHS, RHS);
213 
214     return hash_combine(BinOp->getOpcode(), LHS, RHS);
215   }
216 
217   if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) {
218     // Compares can be commuted by swapping the comparands and
219     // updating the predicate.  Choose the form that has the
220     // comparands in sorted order, or in the case of a tie, the
221     // one with the lower predicate.
222     Value *LHS = CI->getOperand(0);
223     Value *RHS = CI->getOperand(1);
224     CmpInst::Predicate Pred = CI->getPredicate();
225     CmpInst::Predicate SwappedPred = CI->getSwappedPredicate();
226     if (std::tie(LHS, Pred) > std::tie(RHS, SwappedPred)) {
227       std::swap(LHS, RHS);
228       Pred = SwappedPred;
229     }
230     return hash_combine(Inst->getOpcode(), Pred, LHS, RHS);
231   }
232 
233   // Hash general selects to allow matching commuted true/false operands.
234   SelectPatternFlavor SPF;
235   Value *Cond, *A, *B;
236   if (matchSelectWithOptionalNotCond(Inst, Cond, A, B, SPF)) {
237     // Hash min/max (cmp + select) to allow for commuted operands.
238     // Min/max may also have non-canonical compare predicate (eg, the compare for
239     // smin may use 'sgt' rather than 'slt'), and non-canonical operands in the
240     // compare.
241     // TODO: We should also detect FP min/max.
242     if (SPF == SPF_SMIN || SPF == SPF_SMAX ||
243         SPF == SPF_UMIN || SPF == SPF_UMAX) {
244       if (A > B)
245         std::swap(A, B);
246       return hash_combine(Inst->getOpcode(), SPF, A, B);
247     }
248 
249     // Hash general selects to allow matching commuted true/false operands.
250 
251     // If we do not have a compare as the condition, just hash in the condition.
252     CmpInst::Predicate Pred;
253     Value *X, *Y;
254     if (!match(Cond, m_Cmp(Pred, m_Value(X), m_Value(Y))))
255       return hash_combine(Inst->getOpcode(), Cond, A, B);
256 
257     // Similar to cmp normalization (above) - canonicalize the predicate value:
258     // select (icmp Pred, X, Y), A, B --> select (icmp InvPred, X, Y), B, A
259     if (CmpInst::getInversePredicate(Pred) < Pred) {
260       Pred = CmpInst::getInversePredicate(Pred);
261       std::swap(A, B);
262     }
263     return hash_combine(Inst->getOpcode(), Pred, X, Y, A, B);
264   }
265 
266   if (CastInst *CI = dyn_cast<CastInst>(Inst))
267     return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0));
268 
269   if (FreezeInst *FI = dyn_cast<FreezeInst>(Inst))
270     return hash_combine(FI->getOpcode(), FI->getOperand(0));
271 
272   if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst))
273     return hash_combine(EVI->getOpcode(), EVI->getOperand(0),
274                         hash_combine_range(EVI->idx_begin(), EVI->idx_end()));
275 
276   if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst))
277     return hash_combine(IVI->getOpcode(), IVI->getOperand(0),
278                         IVI->getOperand(1),
279                         hash_combine_range(IVI->idx_begin(), IVI->idx_end()));
280 
281   assert((isa<CallInst>(Inst) || isa<GetElementPtrInst>(Inst) ||
282           isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) ||
283           isa<ShuffleVectorInst>(Inst) || isa<UnaryOperator>(Inst) ||
284           isa<FreezeInst>(Inst)) &&
285          "Invalid/unknown instruction");
286 
287   // Handle intrinsics with commutative operands.
288   // TODO: Extend this to handle intrinsics with >2 operands where the 1st
289   //       2 operands are commutative.
290   auto *II = dyn_cast<IntrinsicInst>(Inst);
291   if (II && II->isCommutative() && II->arg_size() == 2) {
292     Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
293     if (LHS > RHS)
294       std::swap(LHS, RHS);
295     return hash_combine(II->getOpcode(), LHS, RHS);
296   }
297 
298   // gc.relocate is 'special' call: its second and third operands are
299   // not real values, but indices into statepoint's argument list.
300   // Get values they point to.
301   if (const GCRelocateInst *GCR = dyn_cast<GCRelocateInst>(Inst))
302     return hash_combine(GCR->getOpcode(), GCR->getOperand(0),
303                         GCR->getBasePtr(), GCR->getDerivedPtr());
304 
305   // Mix in the opcode.
306   return hash_combine(
307       Inst->getOpcode(),
308       hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
309 }
310 
311 unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) {
312 #ifndef NDEBUG
313   // If -earlycse-debug-hash was specified, return a constant -- this
314   // will force all hashing to collide, so we'll exhaustively search
315   // the table for a match, and the assertion in isEqual will fire if
316   // there's a bug causing equal keys to hash differently.
317   if (EarlyCSEDebugHash)
318     return 0;
319 #endif
320   return getHashValueImpl(Val);
321 }
322 
323 static bool isEqualImpl(SimpleValue LHS, SimpleValue RHS) {
324   Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
325 
326   if (LHS.isSentinel() || RHS.isSentinel())
327     return LHSI == RHSI;
328 
329   if (LHSI->getOpcode() != RHSI->getOpcode())
330     return false;
331   if (LHSI->isIdenticalToWhenDefined(RHSI))
332     return true;
333 
334   // If we're not strictly identical, we still might be a commutable instruction
335   if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) {
336     if (!LHSBinOp->isCommutative())
337       return false;
338 
339     assert(isa<BinaryOperator>(RHSI) &&
340            "same opcode, but different instruction type?");
341     BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI);
342 
343     // Commuted equality
344     return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) &&
345            LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0);
346   }
347   if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) {
348     assert(isa<CmpInst>(RHSI) &&
349            "same opcode, but different instruction type?");
350     CmpInst *RHSCmp = cast<CmpInst>(RHSI);
351     // Commuted equality
352     return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) &&
353            LHSCmp->getOperand(1) == RHSCmp->getOperand(0) &&
354            LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate();
355   }
356 
357   // TODO: Extend this for >2 args by matching the trailing N-2 args.
358   auto *LII = dyn_cast<IntrinsicInst>(LHSI);
359   auto *RII = dyn_cast<IntrinsicInst>(RHSI);
360   if (LII && RII && LII->getIntrinsicID() == RII->getIntrinsicID() &&
361       LII->isCommutative() && LII->arg_size() == 2) {
362     return LII->getArgOperand(0) == RII->getArgOperand(1) &&
363            LII->getArgOperand(1) == RII->getArgOperand(0);
364   }
365 
366   // See comment above in `getHashValue()`.
367   if (const GCRelocateInst *GCR1 = dyn_cast<GCRelocateInst>(LHSI))
368     if (const GCRelocateInst *GCR2 = dyn_cast<GCRelocateInst>(RHSI))
369       return GCR1->getOperand(0) == GCR2->getOperand(0) &&
370              GCR1->getBasePtr() == GCR2->getBasePtr() &&
371              GCR1->getDerivedPtr() == GCR2->getDerivedPtr();
372 
373   // Min/max can occur with commuted operands, non-canonical predicates,
374   // and/or non-canonical operands.
375   // Selects can be non-trivially equivalent via inverted conditions and swaps.
376   SelectPatternFlavor LSPF, RSPF;
377   Value *CondL, *CondR, *LHSA, *RHSA, *LHSB, *RHSB;
378   if (matchSelectWithOptionalNotCond(LHSI, CondL, LHSA, LHSB, LSPF) &&
379       matchSelectWithOptionalNotCond(RHSI, CondR, RHSA, RHSB, RSPF)) {
380     if (LSPF == RSPF) {
381       // TODO: We should also detect FP min/max.
382       if (LSPF == SPF_SMIN || LSPF == SPF_SMAX ||
383           LSPF == SPF_UMIN || LSPF == SPF_UMAX)
384         return ((LHSA == RHSA && LHSB == RHSB) ||
385                 (LHSA == RHSB && LHSB == RHSA));
386 
387       // select Cond, A, B <--> select not(Cond), B, A
388       if (CondL == CondR && LHSA == RHSA && LHSB == RHSB)
389         return true;
390     }
391 
392     // If the true/false operands are swapped and the conditions are compares
393     // with inverted predicates, the selects are equal:
394     // select (icmp Pred, X, Y), A, B <--> select (icmp InvPred, X, Y), B, A
395     //
396     // This also handles patterns with a double-negation in the sense of not +
397     // inverse, because we looked through a 'not' in the matching function and
398     // swapped A/B:
399     // select (cmp Pred, X, Y), A, B <--> select (not (cmp InvPred, X, Y)), B, A
400     //
401     // This intentionally does NOT handle patterns with a double-negation in
402     // the sense of not + not, because doing so could result in values
403     // comparing
404     // as equal that hash differently in the min/max cases like:
405     // select (cmp slt, X, Y), X, Y <--> select (not (not (cmp slt, X, Y))), X, Y
406     //   ^ hashes as min                  ^ would not hash as min
407     // In the context of the EarlyCSE pass, however, such cases never reach
408     // this code, as we simplify the double-negation before hashing the second
409     // select (and so still succeed at CSEing them).
410     if (LHSA == RHSB && LHSB == RHSA) {
411       CmpInst::Predicate PredL, PredR;
412       Value *X, *Y;
413       if (match(CondL, m_Cmp(PredL, m_Value(X), m_Value(Y))) &&
414           match(CondR, m_Cmp(PredR, m_Specific(X), m_Specific(Y))) &&
415           CmpInst::getInversePredicate(PredL) == PredR)
416         return true;
417     }
418   }
419 
420   return false;
421 }
422 
423 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) {
424   // These comparisons are nontrivial, so assert that equality implies
425   // hash equality (DenseMap demands this as an invariant).
426   bool Result = isEqualImpl(LHS, RHS);
427   assert(!Result || (LHS.isSentinel() && LHS.Inst == RHS.Inst) ||
428          getHashValueImpl(LHS) == getHashValueImpl(RHS));
429   return Result;
430 }
431 
432 //===----------------------------------------------------------------------===//
433 // CallValue
434 //===----------------------------------------------------------------------===//
435 
436 namespace {
437 
438 /// Struct representing the available call values in the scoped hash
439 /// table.
440 struct CallValue {
441   Instruction *Inst;
442 
443   CallValue(Instruction *I) : Inst(I) {
444     assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
445   }
446 
447   bool isSentinel() const {
448     return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
449            Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
450   }
451 
452   static bool canHandle(Instruction *Inst) {
453     // Don't value number anything that returns void.
454     if (Inst->getType()->isVoidTy())
455       return false;
456 
457     CallInst *CI = dyn_cast<CallInst>(Inst);
458     if (!CI || !CI->onlyReadsMemory())
459       return false;
460     return true;
461   }
462 };
463 
464 } // end anonymous namespace
465 
466 namespace llvm {
467 
468 template <> struct DenseMapInfo<CallValue> {
469   static inline CallValue getEmptyKey() {
470     return DenseMapInfo<Instruction *>::getEmptyKey();
471   }
472 
473   static inline CallValue getTombstoneKey() {
474     return DenseMapInfo<Instruction *>::getTombstoneKey();
475   }
476 
477   static unsigned getHashValue(CallValue Val);
478   static bool isEqual(CallValue LHS, CallValue RHS);
479 };
480 
481 } // end namespace llvm
482 
483 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) {
484   Instruction *Inst = Val.Inst;
485 
486   // Hash all of the operands as pointers and mix in the opcode.
487   return hash_combine(
488       Inst->getOpcode(),
489       hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
490 }
491 
492 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) {
493   Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
494   if (LHS.isSentinel() || RHS.isSentinel())
495     return LHSI == RHSI;
496 
497   return LHSI->isIdenticalTo(RHSI);
498 }
499 
500 //===----------------------------------------------------------------------===//
501 // EarlyCSE implementation
502 //===----------------------------------------------------------------------===//
503 
504 namespace {
505 
506 /// A simple and fast domtree-based CSE pass.
507 ///
508 /// This pass does a simple depth-first walk over the dominator tree,
509 /// eliminating trivially redundant instructions and using instsimplify to
510 /// canonicalize things as it goes. It is intended to be fast and catch obvious
511 /// cases so that instcombine and other passes are more effective. It is
512 /// expected that a later pass of GVN will catch the interesting/hard cases.
513 class EarlyCSE {
514 public:
515   const TargetLibraryInfo &TLI;
516   const TargetTransformInfo &TTI;
517   DominatorTree &DT;
518   AssumptionCache &AC;
519   const SimplifyQuery SQ;
520   MemorySSA *MSSA;
521   std::unique_ptr<MemorySSAUpdater> MSSAUpdater;
522 
523   using AllocatorTy =
524       RecyclingAllocator<BumpPtrAllocator,
525                          ScopedHashTableVal<SimpleValue, Value *>>;
526   using ScopedHTType =
527       ScopedHashTable<SimpleValue, Value *, DenseMapInfo<SimpleValue>,
528                       AllocatorTy>;
529 
530   /// A scoped hash table of the current values of all of our simple
531   /// scalar expressions.
532   ///
533   /// As we walk down the domtree, we look to see if instructions are in this:
534   /// if so, we replace them with what we find, otherwise we insert them so
535   /// that dominated values can succeed in their lookup.
536   ScopedHTType AvailableValues;
537 
538   /// A scoped hash table of the current values of previously encountered
539   /// memory locations.
540   ///
541   /// This allows us to get efficient access to dominating loads or stores when
542   /// we have a fully redundant load.  In addition to the most recent load, we
543   /// keep track of a generation count of the read, which is compared against
544   /// the current generation count.  The current generation count is incremented
545   /// after every possibly writing memory operation, which ensures that we only
546   /// CSE loads with other loads that have no intervening store.  Ordering
547   /// events (such as fences or atomic instructions) increment the generation
548   /// count as well; essentially, we model these as writes to all possible
549   /// locations.  Note that atomic and/or volatile loads and stores can be
550   /// present the table; it is the responsibility of the consumer to inspect
551   /// the atomicity/volatility if needed.
552   struct LoadValue {
553     Instruction *DefInst = nullptr;
554     unsigned Generation = 0;
555     int MatchingId = -1;
556     bool IsAtomic = false;
557 
558     LoadValue() = default;
559     LoadValue(Instruction *Inst, unsigned Generation, unsigned MatchingId,
560               bool IsAtomic)
561         : DefInst(Inst), Generation(Generation), MatchingId(MatchingId),
562           IsAtomic(IsAtomic) {}
563   };
564 
565   using LoadMapAllocator =
566       RecyclingAllocator<BumpPtrAllocator,
567                          ScopedHashTableVal<Value *, LoadValue>>;
568   using LoadHTType =
569       ScopedHashTable<Value *, LoadValue, DenseMapInfo<Value *>,
570                       LoadMapAllocator>;
571 
572   LoadHTType AvailableLoads;
573 
574   // A scoped hash table mapping memory locations (represented as typed
575   // addresses) to generation numbers at which that memory location became
576   // (henceforth indefinitely) invariant.
577   using InvariantMapAllocator =
578       RecyclingAllocator<BumpPtrAllocator,
579                          ScopedHashTableVal<MemoryLocation, unsigned>>;
580   using InvariantHTType =
581       ScopedHashTable<MemoryLocation, unsigned, DenseMapInfo<MemoryLocation>,
582                       InvariantMapAllocator>;
583   InvariantHTType AvailableInvariants;
584 
585   /// A scoped hash table of the current values of read-only call
586   /// values.
587   ///
588   /// It uses the same generation count as loads.
589   using CallHTType =
590       ScopedHashTable<CallValue, std::pair<Instruction *, unsigned>>;
591   CallHTType AvailableCalls;
592 
593   /// This is the current generation of the memory value.
594   unsigned CurrentGeneration = 0;
595 
596   /// Set up the EarlyCSE runner for a particular function.
597   EarlyCSE(const DataLayout &DL, const TargetLibraryInfo &TLI,
598            const TargetTransformInfo &TTI, DominatorTree &DT,
599            AssumptionCache &AC, MemorySSA *MSSA)
600       : TLI(TLI), TTI(TTI), DT(DT), AC(AC), SQ(DL, &TLI, &DT, &AC), MSSA(MSSA),
601         MSSAUpdater(std::make_unique<MemorySSAUpdater>(MSSA)) {}
602 
603   bool run();
604 
605 private:
606   unsigned ClobberCounter = 0;
607   // Almost a POD, but needs to call the constructors for the scoped hash
608   // tables so that a new scope gets pushed on. These are RAII so that the
609   // scope gets popped when the NodeScope is destroyed.
610   class NodeScope {
611   public:
612     NodeScope(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
613               InvariantHTType &AvailableInvariants, CallHTType &AvailableCalls)
614       : Scope(AvailableValues), LoadScope(AvailableLoads),
615         InvariantScope(AvailableInvariants), CallScope(AvailableCalls) {}
616     NodeScope(const NodeScope &) = delete;
617     NodeScope &operator=(const NodeScope &) = delete;
618 
619   private:
620     ScopedHTType::ScopeTy Scope;
621     LoadHTType::ScopeTy LoadScope;
622     InvariantHTType::ScopeTy InvariantScope;
623     CallHTType::ScopeTy CallScope;
624   };
625 
626   // Contains all the needed information to create a stack for doing a depth
627   // first traversal of the tree. This includes scopes for values, loads, and
628   // calls as well as the generation. There is a child iterator so that the
629   // children do not need to be store separately.
630   class StackNode {
631   public:
632     StackNode(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
633               InvariantHTType &AvailableInvariants, CallHTType &AvailableCalls,
634               unsigned cg, DomTreeNode *n, DomTreeNode::const_iterator child,
635               DomTreeNode::const_iterator end)
636         : CurrentGeneration(cg), ChildGeneration(cg), Node(n), ChildIter(child),
637           EndIter(end),
638           Scopes(AvailableValues, AvailableLoads, AvailableInvariants,
639                  AvailableCalls)
640           {}
641     StackNode(const StackNode &) = delete;
642     StackNode &operator=(const StackNode &) = delete;
643 
644     // Accessors.
645     unsigned currentGeneration() const { return CurrentGeneration; }
646     unsigned childGeneration() const { return ChildGeneration; }
647     void childGeneration(unsigned generation) { ChildGeneration = generation; }
648     DomTreeNode *node() { return Node; }
649     DomTreeNode::const_iterator childIter() const { return ChildIter; }
650 
651     DomTreeNode *nextChild() {
652       DomTreeNode *child = *ChildIter;
653       ++ChildIter;
654       return child;
655     }
656 
657     DomTreeNode::const_iterator end() const { return EndIter; }
658     bool isProcessed() const { return Processed; }
659     void process() { Processed = true; }
660 
661   private:
662     unsigned CurrentGeneration;
663     unsigned ChildGeneration;
664     DomTreeNode *Node;
665     DomTreeNode::const_iterator ChildIter;
666     DomTreeNode::const_iterator EndIter;
667     NodeScope Scopes;
668     bool Processed = false;
669   };
670 
671   /// Wrapper class to handle memory instructions, including loads,
672   /// stores and intrinsic loads and stores defined by the target.
673   class ParseMemoryInst {
674   public:
675     ParseMemoryInst(Instruction *Inst, const TargetTransformInfo &TTI)
676       : Inst(Inst) {
677       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
678         IntrID = II->getIntrinsicID();
679         if (TTI.getTgtMemIntrinsic(II, Info))
680           return;
681         if (isHandledNonTargetIntrinsic(IntrID)) {
682           switch (IntrID) {
683           case Intrinsic::masked_load:
684             Info.PtrVal = Inst->getOperand(0);
685             Info.MatchingId = Intrinsic::masked_load;
686             Info.ReadMem = true;
687             Info.WriteMem = false;
688             Info.IsVolatile = false;
689             break;
690           case Intrinsic::masked_store:
691             Info.PtrVal = Inst->getOperand(1);
692             // Use the ID of masked load as the "matching id". This will
693             // prevent matching non-masked loads/stores with masked ones
694             // (which could be done), but at the moment, the code here
695             // does not support matching intrinsics with non-intrinsics,
696             // so keep the MatchingIds specific to masked instructions
697             // for now (TODO).
698             Info.MatchingId = Intrinsic::masked_load;
699             Info.ReadMem = false;
700             Info.WriteMem = true;
701             Info.IsVolatile = false;
702             break;
703           }
704         }
705       }
706     }
707 
708     Instruction *get() { return Inst; }
709     const Instruction *get() const { return Inst; }
710 
711     bool isLoad() const {
712       if (IntrID != 0)
713         return Info.ReadMem;
714       return isa<LoadInst>(Inst);
715     }
716 
717     bool isStore() const {
718       if (IntrID != 0)
719         return Info.WriteMem;
720       return isa<StoreInst>(Inst);
721     }
722 
723     bool isAtomic() const {
724       if (IntrID != 0)
725         return Info.Ordering != AtomicOrdering::NotAtomic;
726       return Inst->isAtomic();
727     }
728 
729     bool isUnordered() const {
730       if (IntrID != 0)
731         return Info.isUnordered();
732 
733       if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
734         return LI->isUnordered();
735       } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
736         return SI->isUnordered();
737       }
738       // Conservative answer
739       return !Inst->isAtomic();
740     }
741 
742     bool isVolatile() const {
743       if (IntrID != 0)
744         return Info.IsVolatile;
745 
746       if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
747         return LI->isVolatile();
748       } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
749         return SI->isVolatile();
750       }
751       // Conservative answer
752       return true;
753     }
754 
755     bool isInvariantLoad() const {
756       if (auto *LI = dyn_cast<LoadInst>(Inst))
757         return LI->hasMetadata(LLVMContext::MD_invariant_load);
758       return false;
759     }
760 
761     bool isValid() const { return getPointerOperand() != nullptr; }
762 
763     // For regular (non-intrinsic) loads/stores, this is set to -1. For
764     // intrinsic loads/stores, the id is retrieved from the corresponding
765     // field in the MemIntrinsicInfo structure.  That field contains
766     // non-negative values only.
767     int getMatchingId() const {
768       if (IntrID != 0)
769         return Info.MatchingId;
770       return -1;
771     }
772 
773     Value *getPointerOperand() const {
774       if (IntrID != 0)
775         return Info.PtrVal;
776       return getLoadStorePointerOperand(Inst);
777     }
778 
779     Type *getValueType() const {
780       // TODO: handle target-specific intrinsics.
781       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
782         switch (II->getIntrinsicID()) {
783         case Intrinsic::masked_load:
784           return II->getType();
785         case Intrinsic::masked_store:
786           return II->getArgOperand(0)->getType();
787         default:
788           return nullptr;
789         }
790       }
791       return getLoadStoreType(Inst);
792     }
793 
794     bool mayReadFromMemory() const {
795       if (IntrID != 0)
796         return Info.ReadMem;
797       return Inst->mayReadFromMemory();
798     }
799 
800     bool mayWriteToMemory() const {
801       if (IntrID != 0)
802         return Info.WriteMem;
803       return Inst->mayWriteToMemory();
804     }
805 
806   private:
807     Intrinsic::ID IntrID = 0;
808     MemIntrinsicInfo Info;
809     Instruction *Inst;
810   };
811 
812   // This function is to prevent accidentally passing a non-target
813   // intrinsic ID to TargetTransformInfo.
814   static bool isHandledNonTargetIntrinsic(Intrinsic::ID ID) {
815     switch (ID) {
816     case Intrinsic::masked_load:
817     case Intrinsic::masked_store:
818       return true;
819     }
820     return false;
821   }
822   static bool isHandledNonTargetIntrinsic(const Value *V) {
823     if (auto *II = dyn_cast<IntrinsicInst>(V))
824       return isHandledNonTargetIntrinsic(II->getIntrinsicID());
825     return false;
826   }
827 
828   bool processNode(DomTreeNode *Node);
829 
830   bool handleBranchCondition(Instruction *CondInst, const BranchInst *BI,
831                              const BasicBlock *BB, const BasicBlock *Pred);
832 
833   Value *getMatchingValue(LoadValue &InVal, ParseMemoryInst &MemInst,
834                           unsigned CurrentGeneration);
835 
836   bool overridingStores(const ParseMemoryInst &Earlier,
837                         const ParseMemoryInst &Later);
838 
839   Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const {
840     // TODO: We could insert relevant casts on type mismatch here.
841     if (auto *LI = dyn_cast<LoadInst>(Inst))
842       return LI->getType() == ExpectedType ? LI : nullptr;
843     else if (auto *SI = dyn_cast<StoreInst>(Inst)) {
844       Value *V = SI->getValueOperand();
845       return V->getType() == ExpectedType ? V : nullptr;
846     }
847     assert(isa<IntrinsicInst>(Inst) && "Instruction not supported");
848     auto *II = cast<IntrinsicInst>(Inst);
849     if (isHandledNonTargetIntrinsic(II->getIntrinsicID()))
850       return getOrCreateResultNonTargetMemIntrinsic(II, ExpectedType);
851     return TTI.getOrCreateResultFromMemIntrinsic(II, ExpectedType);
852   }
853 
854   Value *getOrCreateResultNonTargetMemIntrinsic(IntrinsicInst *II,
855                                                 Type *ExpectedType) const {
856     switch (II->getIntrinsicID()) {
857     case Intrinsic::masked_load:
858       return II;
859     case Intrinsic::masked_store:
860       return II->getOperand(0);
861     }
862     return nullptr;
863   }
864 
865   /// Return true if the instruction is known to only operate on memory
866   /// provably invariant in the given "generation".
867   bool isOperatingOnInvariantMemAt(Instruction *I, unsigned GenAt);
868 
869   bool isSameMemGeneration(unsigned EarlierGeneration, unsigned LaterGeneration,
870                            Instruction *EarlierInst, Instruction *LaterInst);
871 
872   bool isNonTargetIntrinsicMatch(const IntrinsicInst *Earlier,
873                                  const IntrinsicInst *Later) {
874     auto IsSubmask = [](const Value *Mask0, const Value *Mask1) {
875       // Is Mask0 a submask of Mask1?
876       if (Mask0 == Mask1)
877         return true;
878       if (isa<UndefValue>(Mask0) || isa<UndefValue>(Mask1))
879         return false;
880       auto *Vec0 = dyn_cast<ConstantVector>(Mask0);
881       auto *Vec1 = dyn_cast<ConstantVector>(Mask1);
882       if (!Vec0 || !Vec1)
883         return false;
884       assert(Vec0->getType() == Vec1->getType() &&
885              "Masks should have the same type");
886       for (int i = 0, e = Vec0->getNumOperands(); i != e; ++i) {
887         Constant *Elem0 = Vec0->getOperand(i);
888         Constant *Elem1 = Vec1->getOperand(i);
889         auto *Int0 = dyn_cast<ConstantInt>(Elem0);
890         if (Int0 && Int0->isZero())
891           continue;
892         auto *Int1 = dyn_cast<ConstantInt>(Elem1);
893         if (Int1 && !Int1->isZero())
894           continue;
895         if (isa<UndefValue>(Elem0) || isa<UndefValue>(Elem1))
896           return false;
897         if (Elem0 == Elem1)
898           continue;
899         return false;
900       }
901       return true;
902     };
903     auto PtrOp = [](const IntrinsicInst *II) {
904       if (II->getIntrinsicID() == Intrinsic::masked_load)
905         return II->getOperand(0);
906       if (II->getIntrinsicID() == Intrinsic::masked_store)
907         return II->getOperand(1);
908       llvm_unreachable("Unexpected IntrinsicInst");
909     };
910     auto MaskOp = [](const IntrinsicInst *II) {
911       if (II->getIntrinsicID() == Intrinsic::masked_load)
912         return II->getOperand(2);
913       if (II->getIntrinsicID() == Intrinsic::masked_store)
914         return II->getOperand(3);
915       llvm_unreachable("Unexpected IntrinsicInst");
916     };
917     auto ThruOp = [](const IntrinsicInst *II) {
918       if (II->getIntrinsicID() == Intrinsic::masked_load)
919         return II->getOperand(3);
920       llvm_unreachable("Unexpected IntrinsicInst");
921     };
922 
923     if (PtrOp(Earlier) != PtrOp(Later))
924       return false;
925 
926     Intrinsic::ID IDE = Earlier->getIntrinsicID();
927     Intrinsic::ID IDL = Later->getIntrinsicID();
928     // We could really use specific intrinsic classes for masked loads
929     // and stores in IntrinsicInst.h.
930     if (IDE == Intrinsic::masked_load && IDL == Intrinsic::masked_load) {
931       // Trying to replace later masked load with the earlier one.
932       // Check that the pointers are the same, and
933       // - masks and pass-throughs are the same, or
934       // - replacee's pass-through is "undef" and replacer's mask is a
935       //   super-set of the replacee's mask.
936       if (MaskOp(Earlier) == MaskOp(Later) && ThruOp(Earlier) == ThruOp(Later))
937         return true;
938       if (!isa<UndefValue>(ThruOp(Later)))
939         return false;
940       return IsSubmask(MaskOp(Later), MaskOp(Earlier));
941     }
942     if (IDE == Intrinsic::masked_store && IDL == Intrinsic::masked_load) {
943       // Trying to replace a load of a stored value with the store's value.
944       // Check that the pointers are the same, and
945       // - load's mask is a subset of store's mask, and
946       // - load's pass-through is "undef".
947       if (!IsSubmask(MaskOp(Later), MaskOp(Earlier)))
948         return false;
949       return isa<UndefValue>(ThruOp(Later));
950     }
951     if (IDE == Intrinsic::masked_load && IDL == Intrinsic::masked_store) {
952       // Trying to remove a store of the loaded value.
953       // Check that the pointers are the same, and
954       // - store's mask is a subset of the load's mask.
955       return IsSubmask(MaskOp(Later), MaskOp(Earlier));
956     }
957     if (IDE == Intrinsic::masked_store && IDL == Intrinsic::masked_store) {
958       // Trying to remove a dead store (earlier).
959       // Check that the pointers are the same,
960       // - the to-be-removed store's mask is a subset of the other store's
961       //   mask.
962       return IsSubmask(MaskOp(Earlier), MaskOp(Later));
963     }
964     return false;
965   }
966 
967   void removeMSSA(Instruction &Inst) {
968     if (!MSSA)
969       return;
970     if (VerifyMemorySSA)
971       MSSA->verifyMemorySSA();
972     // Removing a store here can leave MemorySSA in an unoptimized state by
973     // creating MemoryPhis that have identical arguments and by creating
974     // MemoryUses whose defining access is not an actual clobber. The phi case
975     // is handled by MemorySSA when passing OptimizePhis = true to
976     // removeMemoryAccess.  The non-optimized MemoryUse case is lazily updated
977     // by MemorySSA's getClobberingMemoryAccess.
978     MSSAUpdater->removeMemoryAccess(&Inst, true);
979   }
980 };
981 
982 } // end anonymous namespace
983 
984 /// Determine if the memory referenced by LaterInst is from the same heap
985 /// version as EarlierInst.
986 /// This is currently called in two scenarios:
987 ///
988 ///   load p
989 ///   ...
990 ///   load p
991 ///
992 /// and
993 ///
994 ///   x = load p
995 ///   ...
996 ///   store x, p
997 ///
998 /// in both cases we want to verify that there are no possible writes to the
999 /// memory referenced by p between the earlier and later instruction.
1000 bool EarlyCSE::isSameMemGeneration(unsigned EarlierGeneration,
1001                                    unsigned LaterGeneration,
1002                                    Instruction *EarlierInst,
1003                                    Instruction *LaterInst) {
1004   // Check the simple memory generation tracking first.
1005   if (EarlierGeneration == LaterGeneration)
1006     return true;
1007 
1008   if (!MSSA)
1009     return false;
1010 
1011   // If MemorySSA has determined that one of EarlierInst or LaterInst does not
1012   // read/write memory, then we can safely return true here.
1013   // FIXME: We could be more aggressive when checking doesNotAccessMemory(),
1014   // onlyReadsMemory(), mayReadFromMemory(), and mayWriteToMemory() in this pass
1015   // by also checking the MemorySSA MemoryAccess on the instruction.  Initial
1016   // experiments suggest this isn't worthwhile, at least for C/C++ code compiled
1017   // with the default optimization pipeline.
1018   auto *EarlierMA = MSSA->getMemoryAccess(EarlierInst);
1019   if (!EarlierMA)
1020     return true;
1021   auto *LaterMA = MSSA->getMemoryAccess(LaterInst);
1022   if (!LaterMA)
1023     return true;
1024 
1025   // Since we know LaterDef dominates LaterInst and EarlierInst dominates
1026   // LaterInst, if LaterDef dominates EarlierInst then it can't occur between
1027   // EarlierInst and LaterInst and neither can any other write that potentially
1028   // clobbers LaterInst.
1029   MemoryAccess *LaterDef;
1030   if (ClobberCounter < EarlyCSEMssaOptCap) {
1031     LaterDef = MSSA->getWalker()->getClobberingMemoryAccess(LaterInst);
1032     ClobberCounter++;
1033   } else
1034     LaterDef = LaterMA->getDefiningAccess();
1035 
1036   return MSSA->dominates(LaterDef, EarlierMA);
1037 }
1038 
1039 bool EarlyCSE::isOperatingOnInvariantMemAt(Instruction *I, unsigned GenAt) {
1040   // A location loaded from with an invariant_load is assumed to *never* change
1041   // within the visible scope of the compilation.
1042   if (auto *LI = dyn_cast<LoadInst>(I))
1043     if (LI->hasMetadata(LLVMContext::MD_invariant_load))
1044       return true;
1045 
1046   auto MemLocOpt = MemoryLocation::getOrNone(I);
1047   if (!MemLocOpt)
1048     // "target" intrinsic forms of loads aren't currently known to
1049     // MemoryLocation::get.  TODO
1050     return false;
1051   MemoryLocation MemLoc = *MemLocOpt;
1052   if (!AvailableInvariants.count(MemLoc))
1053     return false;
1054 
1055   // Is the generation at which this became invariant older than the
1056   // current one?
1057   return AvailableInvariants.lookup(MemLoc) <= GenAt;
1058 }
1059 
1060 bool EarlyCSE::handleBranchCondition(Instruction *CondInst,
1061                                      const BranchInst *BI, const BasicBlock *BB,
1062                                      const BasicBlock *Pred) {
1063   assert(BI->isConditional() && "Should be a conditional branch!");
1064   assert(BI->getCondition() == CondInst && "Wrong condition?");
1065   assert(BI->getSuccessor(0) == BB || BI->getSuccessor(1) == BB);
1066   auto *TorF = (BI->getSuccessor(0) == BB)
1067                    ? ConstantInt::getTrue(BB->getContext())
1068                    : ConstantInt::getFalse(BB->getContext());
1069   auto MatchBinOp = [](Instruction *I, unsigned Opcode, Value *&LHS,
1070                        Value *&RHS) {
1071     if (Opcode == Instruction::And &&
1072         match(I, m_LogicalAnd(m_Value(LHS), m_Value(RHS))))
1073       return true;
1074     else if (Opcode == Instruction::Or &&
1075              match(I, m_LogicalOr(m_Value(LHS), m_Value(RHS))))
1076       return true;
1077     return false;
1078   };
1079   // If the condition is AND operation, we can propagate its operands into the
1080   // true branch. If it is OR operation, we can propagate them into the false
1081   // branch.
1082   unsigned PropagateOpcode =
1083       (BI->getSuccessor(0) == BB) ? Instruction::And : Instruction::Or;
1084 
1085   bool MadeChanges = false;
1086   SmallVector<Instruction *, 4> WorkList;
1087   SmallPtrSet<Instruction *, 4> Visited;
1088   WorkList.push_back(CondInst);
1089   while (!WorkList.empty()) {
1090     Instruction *Curr = WorkList.pop_back_val();
1091 
1092     AvailableValues.insert(Curr, TorF);
1093     LLVM_DEBUG(dbgs() << "EarlyCSE CVP: Add conditional value for '"
1094                       << Curr->getName() << "' as " << *TorF << " in "
1095                       << BB->getName() << "\n");
1096     if (!DebugCounter::shouldExecute(CSECounter)) {
1097       LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1098     } else {
1099       // Replace all dominated uses with the known value.
1100       if (unsigned Count = replaceDominatedUsesWith(Curr, TorF, DT,
1101                                                     BasicBlockEdge(Pred, BB))) {
1102         NumCSECVP += Count;
1103         MadeChanges = true;
1104       }
1105     }
1106 
1107     Value *LHS, *RHS;
1108     if (MatchBinOp(Curr, PropagateOpcode, LHS, RHS))
1109       for (auto &Op : { LHS, RHS })
1110         if (Instruction *OPI = dyn_cast<Instruction>(Op))
1111           if (SimpleValue::canHandle(OPI) && Visited.insert(OPI).second)
1112             WorkList.push_back(OPI);
1113   }
1114 
1115   return MadeChanges;
1116 }
1117 
1118 Value *EarlyCSE::getMatchingValue(LoadValue &InVal, ParseMemoryInst &MemInst,
1119                                   unsigned CurrentGeneration) {
1120   if (InVal.DefInst == nullptr)
1121     return nullptr;
1122   if (InVal.MatchingId != MemInst.getMatchingId())
1123     return nullptr;
1124   // We don't yet handle removing loads with ordering of any kind.
1125   if (MemInst.isVolatile() || !MemInst.isUnordered())
1126     return nullptr;
1127   // We can't replace an atomic load with one which isn't also atomic.
1128   if (MemInst.isLoad() && !InVal.IsAtomic && MemInst.isAtomic())
1129     return nullptr;
1130   // The value V returned from this function is used differently depending
1131   // on whether MemInst is a load or a store. If it's a load, we will replace
1132   // MemInst with V, if it's a store, we will check if V is the same as the
1133   // available value.
1134   bool MemInstMatching = !MemInst.isLoad();
1135   Instruction *Matching = MemInstMatching ? MemInst.get() : InVal.DefInst;
1136   Instruction *Other = MemInstMatching ? InVal.DefInst : MemInst.get();
1137 
1138   // For stores check the result values before checking memory generation
1139   // (otherwise isSameMemGeneration may crash).
1140   Value *Result = MemInst.isStore()
1141                       ? getOrCreateResult(Matching, Other->getType())
1142                       : nullptr;
1143   if (MemInst.isStore() && InVal.DefInst != Result)
1144     return nullptr;
1145 
1146   // Deal with non-target memory intrinsics.
1147   bool MatchingNTI = isHandledNonTargetIntrinsic(Matching);
1148   bool OtherNTI = isHandledNonTargetIntrinsic(Other);
1149   if (OtherNTI != MatchingNTI)
1150     return nullptr;
1151   if (OtherNTI && MatchingNTI) {
1152     if (!isNonTargetIntrinsicMatch(cast<IntrinsicInst>(InVal.DefInst),
1153                                    cast<IntrinsicInst>(MemInst.get())))
1154       return nullptr;
1155   }
1156 
1157   if (!isOperatingOnInvariantMemAt(MemInst.get(), InVal.Generation) &&
1158       !isSameMemGeneration(InVal.Generation, CurrentGeneration, InVal.DefInst,
1159                            MemInst.get()))
1160     return nullptr;
1161 
1162   if (!Result)
1163     Result = getOrCreateResult(Matching, Other->getType());
1164   return Result;
1165 }
1166 
1167 bool EarlyCSE::overridingStores(const ParseMemoryInst &Earlier,
1168                                 const ParseMemoryInst &Later) {
1169   // Can we remove Earlier store because of Later store?
1170 
1171   assert(Earlier.isUnordered() && !Earlier.isVolatile() &&
1172          "Violated invariant");
1173   if (Earlier.getPointerOperand() != Later.getPointerOperand())
1174     return false;
1175   if (!Earlier.getValueType() || !Later.getValueType() ||
1176       Earlier.getValueType() != Later.getValueType())
1177     return false;
1178   if (Earlier.getMatchingId() != Later.getMatchingId())
1179     return false;
1180   // At the moment, we don't remove ordered stores, but do remove
1181   // unordered atomic stores.  There's no special requirement (for
1182   // unordered atomics) about removing atomic stores only in favor of
1183   // other atomic stores since we were going to execute the non-atomic
1184   // one anyway and the atomic one might never have become visible.
1185   if (!Earlier.isUnordered() || !Later.isUnordered())
1186     return false;
1187 
1188   // Deal with non-target memory intrinsics.
1189   bool ENTI = isHandledNonTargetIntrinsic(Earlier.get());
1190   bool LNTI = isHandledNonTargetIntrinsic(Later.get());
1191   if (ENTI && LNTI)
1192     return isNonTargetIntrinsicMatch(cast<IntrinsicInst>(Earlier.get()),
1193                                      cast<IntrinsicInst>(Later.get()));
1194 
1195   // Because of the check above, at least one of them is false.
1196   // For now disallow matching intrinsics with non-intrinsics,
1197   // so assume that the stores match if neither is an intrinsic.
1198   return ENTI == LNTI;
1199 }
1200 
1201 bool EarlyCSE::processNode(DomTreeNode *Node) {
1202   bool Changed = false;
1203   BasicBlock *BB = Node->getBlock();
1204 
1205   // If this block has a single predecessor, then the predecessor is the parent
1206   // of the domtree node and all of the live out memory values are still current
1207   // in this block.  If this block has multiple predecessors, then they could
1208   // have invalidated the live-out memory values of our parent value.  For now,
1209   // just be conservative and invalidate memory if this block has multiple
1210   // predecessors.
1211   if (!BB->getSinglePredecessor())
1212     ++CurrentGeneration;
1213 
1214   // If this node has a single predecessor which ends in a conditional branch,
1215   // we can infer the value of the branch condition given that we took this
1216   // path.  We need the single predecessor to ensure there's not another path
1217   // which reaches this block where the condition might hold a different
1218   // value.  Since we're adding this to the scoped hash table (like any other
1219   // def), it will have been popped if we encounter a future merge block.
1220   if (BasicBlock *Pred = BB->getSinglePredecessor()) {
1221     auto *BI = dyn_cast<BranchInst>(Pred->getTerminator());
1222     if (BI && BI->isConditional()) {
1223       auto *CondInst = dyn_cast<Instruction>(BI->getCondition());
1224       if (CondInst && SimpleValue::canHandle(CondInst))
1225         Changed |= handleBranchCondition(CondInst, BI, BB, Pred);
1226     }
1227   }
1228 
1229   /// LastStore - Keep track of the last non-volatile store that we saw... for
1230   /// as long as there in no instruction that reads memory.  If we see a store
1231   /// to the same location, we delete the dead store.  This zaps trivial dead
1232   /// stores which can occur in bitfield code among other things.
1233   Instruction *LastStore = nullptr;
1234 
1235   // See if any instructions in the block can be eliminated.  If so, do it.  If
1236   // not, add them to AvailableValues.
1237   for (Instruction &Inst : make_early_inc_range(BB->getInstList())) {
1238     // Dead instructions should just be removed.
1239     if (isInstructionTriviallyDead(&Inst, &TLI)) {
1240       LLVM_DEBUG(dbgs() << "EarlyCSE DCE: " << Inst << '\n');
1241       if (!DebugCounter::shouldExecute(CSECounter)) {
1242         LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1243         continue;
1244       }
1245 
1246       salvageKnowledge(&Inst, &AC);
1247       salvageDebugInfo(Inst);
1248       removeMSSA(Inst);
1249       Inst.eraseFromParent();
1250       Changed = true;
1251       ++NumSimplify;
1252       continue;
1253     }
1254 
1255     // Skip assume intrinsics, they don't really have side effects (although
1256     // they're marked as such to ensure preservation of control dependencies),
1257     // and this pass will not bother with its removal. However, we should mark
1258     // its condition as true for all dominated blocks.
1259     if (auto *Assume = dyn_cast<AssumeInst>(&Inst)) {
1260       auto *CondI = dyn_cast<Instruction>(Assume->getArgOperand(0));
1261       if (CondI && SimpleValue::canHandle(CondI)) {
1262         LLVM_DEBUG(dbgs() << "EarlyCSE considering assumption: " << Inst
1263                           << '\n');
1264         AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext()));
1265       } else
1266         LLVM_DEBUG(dbgs() << "EarlyCSE skipping assumption: " << Inst << '\n');
1267       continue;
1268     }
1269 
1270     // Likewise, noalias intrinsics don't actually write.
1271     if (match(&Inst,
1272               m_Intrinsic<Intrinsic::experimental_noalias_scope_decl>())) {
1273       LLVM_DEBUG(dbgs() << "EarlyCSE skipping noalias intrinsic: " << Inst
1274                         << '\n');
1275       continue;
1276     }
1277 
1278     // Skip sideeffect intrinsics, for the same reason as assume intrinsics.
1279     if (match(&Inst, m_Intrinsic<Intrinsic::sideeffect>())) {
1280       LLVM_DEBUG(dbgs() << "EarlyCSE skipping sideeffect: " << Inst << '\n');
1281       continue;
1282     }
1283 
1284     // Skip pseudoprobe intrinsics, for the same reason as assume intrinsics.
1285     if (match(&Inst, m_Intrinsic<Intrinsic::pseudoprobe>())) {
1286       LLVM_DEBUG(dbgs() << "EarlyCSE skipping pseudoprobe: " << Inst << '\n');
1287       continue;
1288     }
1289 
1290     // We can skip all invariant.start intrinsics since they only read memory,
1291     // and we can forward values across it. For invariant starts without
1292     // invariant ends, we can use the fact that the invariantness never ends to
1293     // start a scope in the current generaton which is true for all future
1294     // generations.  Also, we dont need to consume the last store since the
1295     // semantics of invariant.start allow us to perform   DSE of the last
1296     // store, if there was a store following invariant.start. Consider:
1297     //
1298     // store 30, i8* p
1299     // invariant.start(p)
1300     // store 40, i8* p
1301     // We can DSE the store to 30, since the store 40 to invariant location p
1302     // causes undefined behaviour.
1303     if (match(&Inst, m_Intrinsic<Intrinsic::invariant_start>())) {
1304       // If there are any uses, the scope might end.
1305       if (!Inst.use_empty())
1306         continue;
1307       MemoryLocation MemLoc =
1308           MemoryLocation::getForArgument(&cast<CallInst>(Inst), 1, TLI);
1309       // Don't start a scope if we already have a better one pushed
1310       if (!AvailableInvariants.count(MemLoc))
1311         AvailableInvariants.insert(MemLoc, CurrentGeneration);
1312       continue;
1313     }
1314 
1315     if (isGuard(&Inst)) {
1316       if (auto *CondI =
1317               dyn_cast<Instruction>(cast<CallInst>(Inst).getArgOperand(0))) {
1318         if (SimpleValue::canHandle(CondI)) {
1319           // Do we already know the actual value of this condition?
1320           if (auto *KnownCond = AvailableValues.lookup(CondI)) {
1321             // Is the condition known to be true?
1322             if (isa<ConstantInt>(KnownCond) &&
1323                 cast<ConstantInt>(KnownCond)->isOne()) {
1324               LLVM_DEBUG(dbgs()
1325                          << "EarlyCSE removing guard: " << Inst << '\n');
1326               salvageKnowledge(&Inst, &AC);
1327               removeMSSA(Inst);
1328               Inst.eraseFromParent();
1329               Changed = true;
1330               continue;
1331             } else
1332               // Use the known value if it wasn't true.
1333               cast<CallInst>(Inst).setArgOperand(0, KnownCond);
1334           }
1335           // The condition we're on guarding here is true for all dominated
1336           // locations.
1337           AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext()));
1338         }
1339       }
1340 
1341       // Guard intrinsics read all memory, but don't write any memory.
1342       // Accordingly, don't update the generation but consume the last store (to
1343       // avoid an incorrect DSE).
1344       LastStore = nullptr;
1345       continue;
1346     }
1347 
1348     // If the instruction can be simplified (e.g. X+0 = X) then replace it with
1349     // its simpler value.
1350     if (Value *V = simplifyInstruction(&Inst, SQ)) {
1351       LLVM_DEBUG(dbgs() << "EarlyCSE Simplify: " << Inst << "  to: " << *V
1352                         << '\n');
1353       if (!DebugCounter::shouldExecute(CSECounter)) {
1354         LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1355       } else {
1356         bool Killed = false;
1357         if (!Inst.use_empty()) {
1358           Inst.replaceAllUsesWith(V);
1359           Changed = true;
1360         }
1361         if (isInstructionTriviallyDead(&Inst, &TLI)) {
1362           salvageKnowledge(&Inst, &AC);
1363           removeMSSA(Inst);
1364           Inst.eraseFromParent();
1365           Changed = true;
1366           Killed = true;
1367         }
1368         if (Changed)
1369           ++NumSimplify;
1370         if (Killed)
1371           continue;
1372       }
1373     }
1374 
1375     // If this is a simple instruction that we can value number, process it.
1376     if (SimpleValue::canHandle(&Inst)) {
1377       // See if the instruction has an available value.  If so, use it.
1378       if (Value *V = AvailableValues.lookup(&Inst)) {
1379         LLVM_DEBUG(dbgs() << "EarlyCSE CSE: " << Inst << "  to: " << *V
1380                           << '\n');
1381         if (!DebugCounter::shouldExecute(CSECounter)) {
1382           LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1383           continue;
1384         }
1385         if (auto *I = dyn_cast<Instruction>(V)) {
1386           // If I being poison triggers UB, there is no need to drop those
1387           // flags. Otherwise, only retain flags present on both I and Inst.
1388           // TODO: Currently some fast-math flags are not treated as
1389           // poison-generating even though they should. Until this is fixed,
1390           // always retain flags present on both I and Inst for floating point
1391           // instructions.
1392           if (isa<FPMathOperator>(I) || (I->hasPoisonGeneratingFlags() && !programUndefinedIfPoison(I)))
1393             I->andIRFlags(&Inst);
1394         }
1395         Inst.replaceAllUsesWith(V);
1396         salvageKnowledge(&Inst, &AC);
1397         removeMSSA(Inst);
1398         Inst.eraseFromParent();
1399         Changed = true;
1400         ++NumCSE;
1401         continue;
1402       }
1403 
1404       // Otherwise, just remember that this value is available.
1405       AvailableValues.insert(&Inst, &Inst);
1406       continue;
1407     }
1408 
1409     ParseMemoryInst MemInst(&Inst, TTI);
1410     // If this is a non-volatile load, process it.
1411     if (MemInst.isValid() && MemInst.isLoad()) {
1412       // (conservatively) we can't peak past the ordering implied by this
1413       // operation, but we can add this load to our set of available values
1414       if (MemInst.isVolatile() || !MemInst.isUnordered()) {
1415         LastStore = nullptr;
1416         ++CurrentGeneration;
1417       }
1418 
1419       if (MemInst.isInvariantLoad()) {
1420         // If we pass an invariant load, we know that memory location is
1421         // indefinitely constant from the moment of first dereferenceability.
1422         // We conservatively treat the invariant_load as that moment.  If we
1423         // pass a invariant load after already establishing a scope, don't
1424         // restart it since we want to preserve the earliest point seen.
1425         auto MemLoc = MemoryLocation::get(&Inst);
1426         if (!AvailableInvariants.count(MemLoc))
1427           AvailableInvariants.insert(MemLoc, CurrentGeneration);
1428       }
1429 
1430       // If we have an available version of this load, and if it is the right
1431       // generation or the load is known to be from an invariant location,
1432       // replace this instruction.
1433       //
1434       // If either the dominating load or the current load are invariant, then
1435       // we can assume the current load loads the same value as the dominating
1436       // load.
1437       LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand());
1438       if (Value *Op = getMatchingValue(InVal, MemInst, CurrentGeneration)) {
1439         LLVM_DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << Inst
1440                           << "  to: " << *InVal.DefInst << '\n');
1441         if (!DebugCounter::shouldExecute(CSECounter)) {
1442           LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1443           continue;
1444         }
1445         if (!Inst.use_empty())
1446           Inst.replaceAllUsesWith(Op);
1447         salvageKnowledge(&Inst, &AC);
1448         removeMSSA(Inst);
1449         Inst.eraseFromParent();
1450         Changed = true;
1451         ++NumCSELoad;
1452         continue;
1453       }
1454 
1455       // Otherwise, remember that we have this instruction.
1456       AvailableLoads.insert(MemInst.getPointerOperand(),
1457                             LoadValue(&Inst, CurrentGeneration,
1458                                       MemInst.getMatchingId(),
1459                                       MemInst.isAtomic()));
1460       LastStore = nullptr;
1461       continue;
1462     }
1463 
1464     // If this instruction may read from memory or throw (and potentially read
1465     // from memory in the exception handler), forget LastStore.  Load/store
1466     // intrinsics will indicate both a read and a write to memory.  The target
1467     // may override this (e.g. so that a store intrinsic does not read from
1468     // memory, and thus will be treated the same as a regular store for
1469     // commoning purposes).
1470     if ((Inst.mayReadFromMemory() || Inst.mayThrow()) &&
1471         !(MemInst.isValid() && !MemInst.mayReadFromMemory()))
1472       LastStore = nullptr;
1473 
1474     // If this is a read-only call, process it.
1475     if (CallValue::canHandle(&Inst)) {
1476       // If we have an available version of this call, and if it is the right
1477       // generation, replace this instruction.
1478       std::pair<Instruction *, unsigned> InVal = AvailableCalls.lookup(&Inst);
1479       if (InVal.first != nullptr &&
1480           isSameMemGeneration(InVal.second, CurrentGeneration, InVal.first,
1481                               &Inst)) {
1482         LLVM_DEBUG(dbgs() << "EarlyCSE CSE CALL: " << Inst
1483                           << "  to: " << *InVal.first << '\n');
1484         if (!DebugCounter::shouldExecute(CSECounter)) {
1485           LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1486           continue;
1487         }
1488         if (!Inst.use_empty())
1489           Inst.replaceAllUsesWith(InVal.first);
1490         salvageKnowledge(&Inst, &AC);
1491         removeMSSA(Inst);
1492         Inst.eraseFromParent();
1493         Changed = true;
1494         ++NumCSECall;
1495         continue;
1496       }
1497 
1498       // Otherwise, remember that we have this instruction.
1499       AvailableCalls.insert(&Inst, std::make_pair(&Inst, CurrentGeneration));
1500       continue;
1501     }
1502 
1503     // A release fence requires that all stores complete before it, but does
1504     // not prevent the reordering of following loads 'before' the fence.  As a
1505     // result, we don't need to consider it as writing to memory and don't need
1506     // to advance the generation.  We do need to prevent DSE across the fence,
1507     // but that's handled above.
1508     if (auto *FI = dyn_cast<FenceInst>(&Inst))
1509       if (FI->getOrdering() == AtomicOrdering::Release) {
1510         assert(Inst.mayReadFromMemory() && "relied on to prevent DSE above");
1511         continue;
1512       }
1513 
1514     // write back DSE - If we write back the same value we just loaded from
1515     // the same location and haven't passed any intervening writes or ordering
1516     // operations, we can remove the write.  The primary benefit is in allowing
1517     // the available load table to remain valid and value forward past where
1518     // the store originally was.
1519     if (MemInst.isValid() && MemInst.isStore()) {
1520       LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand());
1521       if (InVal.DefInst &&
1522           InVal.DefInst == getMatchingValue(InVal, MemInst, CurrentGeneration)) {
1523         // It is okay to have a LastStore to a different pointer here if MemorySSA
1524         // tells us that the load and store are from the same memory generation.
1525         // In that case, LastStore should keep its present value since we're
1526         // removing the current store.
1527         assert((!LastStore ||
1528                 ParseMemoryInst(LastStore, TTI).getPointerOperand() ==
1529                     MemInst.getPointerOperand() ||
1530                 MSSA) &&
1531                "can't have an intervening store if not using MemorySSA!");
1532         LLVM_DEBUG(dbgs() << "EarlyCSE DSE (writeback): " << Inst << '\n');
1533         if (!DebugCounter::shouldExecute(CSECounter)) {
1534           LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1535           continue;
1536         }
1537         salvageKnowledge(&Inst, &AC);
1538         removeMSSA(Inst);
1539         Inst.eraseFromParent();
1540         Changed = true;
1541         ++NumDSE;
1542         // We can avoid incrementing the generation count since we were able
1543         // to eliminate this store.
1544         continue;
1545       }
1546     }
1547 
1548     // Okay, this isn't something we can CSE at all.  Check to see if it is
1549     // something that could modify memory.  If so, our available memory values
1550     // cannot be used so bump the generation count.
1551     if (Inst.mayWriteToMemory()) {
1552       ++CurrentGeneration;
1553 
1554       if (MemInst.isValid() && MemInst.isStore()) {
1555         // We do a trivial form of DSE if there are two stores to the same
1556         // location with no intervening loads.  Delete the earlier store.
1557         if (LastStore) {
1558           if (overridingStores(ParseMemoryInst(LastStore, TTI), MemInst)) {
1559             LLVM_DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore
1560                               << "  due to: " << Inst << '\n');
1561             if (!DebugCounter::shouldExecute(CSECounter)) {
1562               LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1563             } else {
1564               salvageKnowledge(&Inst, &AC);
1565               removeMSSA(*LastStore);
1566               LastStore->eraseFromParent();
1567               Changed = true;
1568               ++NumDSE;
1569               LastStore = nullptr;
1570             }
1571           }
1572           // fallthrough - we can exploit information about this store
1573         }
1574 
1575         // Okay, we just invalidated anything we knew about loaded values.  Try
1576         // to salvage *something* by remembering that the stored value is a live
1577         // version of the pointer.  It is safe to forward from volatile stores
1578         // to non-volatile loads, so we don't have to check for volatility of
1579         // the store.
1580         AvailableLoads.insert(MemInst.getPointerOperand(),
1581                               LoadValue(&Inst, CurrentGeneration,
1582                                         MemInst.getMatchingId(),
1583                                         MemInst.isAtomic()));
1584 
1585         // Remember that this was the last unordered store we saw for DSE. We
1586         // don't yet handle DSE on ordered or volatile stores since we don't
1587         // have a good way to model the ordering requirement for following
1588         // passes  once the store is removed.  We could insert a fence, but
1589         // since fences are slightly stronger than stores in their ordering,
1590         // it's not clear this is a profitable transform. Another option would
1591         // be to merge the ordering with that of the post dominating store.
1592         if (MemInst.isUnordered() && !MemInst.isVolatile())
1593           LastStore = &Inst;
1594         else
1595           LastStore = nullptr;
1596       }
1597     }
1598   }
1599 
1600   return Changed;
1601 }
1602 
1603 bool EarlyCSE::run() {
1604   // Note, deque is being used here because there is significant performance
1605   // gains over vector when the container becomes very large due to the
1606   // specific access patterns. For more information see the mailing list
1607   // discussion on this:
1608   // http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html
1609   std::deque<StackNode *> nodesToProcess;
1610 
1611   bool Changed = false;
1612 
1613   // Process the root node.
1614   nodesToProcess.push_back(new StackNode(
1615       AvailableValues, AvailableLoads, AvailableInvariants, AvailableCalls,
1616       CurrentGeneration, DT.getRootNode(),
1617       DT.getRootNode()->begin(), DT.getRootNode()->end()));
1618 
1619   assert(!CurrentGeneration && "Create a new EarlyCSE instance to rerun it.");
1620 
1621   // Process the stack.
1622   while (!nodesToProcess.empty()) {
1623     // Grab the first item off the stack. Set the current generation, remove
1624     // the node from the stack, and process it.
1625     StackNode *NodeToProcess = nodesToProcess.back();
1626 
1627     // Initialize class members.
1628     CurrentGeneration = NodeToProcess->currentGeneration();
1629 
1630     // Check if the node needs to be processed.
1631     if (!NodeToProcess->isProcessed()) {
1632       // Process the node.
1633       Changed |= processNode(NodeToProcess->node());
1634       NodeToProcess->childGeneration(CurrentGeneration);
1635       NodeToProcess->process();
1636     } else if (NodeToProcess->childIter() != NodeToProcess->end()) {
1637       // Push the next child onto the stack.
1638       DomTreeNode *child = NodeToProcess->nextChild();
1639       nodesToProcess.push_back(
1640           new StackNode(AvailableValues, AvailableLoads, AvailableInvariants,
1641                         AvailableCalls, NodeToProcess->childGeneration(),
1642                         child, child->begin(), child->end()));
1643     } else {
1644       // It has been processed, and there are no more children to process,
1645       // so delete it and pop it off the stack.
1646       delete NodeToProcess;
1647       nodesToProcess.pop_back();
1648     }
1649   } // while (!nodes...)
1650 
1651   return Changed;
1652 }
1653 
1654 PreservedAnalyses EarlyCSEPass::run(Function &F,
1655                                     FunctionAnalysisManager &AM) {
1656   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1657   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
1658   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1659   auto &AC = AM.getResult<AssumptionAnalysis>(F);
1660   auto *MSSA =
1661       UseMemorySSA ? &AM.getResult<MemorySSAAnalysis>(F).getMSSA() : nullptr;
1662 
1663   EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA);
1664 
1665   if (!CSE.run())
1666     return PreservedAnalyses::all();
1667 
1668   PreservedAnalyses PA;
1669   PA.preserveSet<CFGAnalyses>();
1670   if (UseMemorySSA)
1671     PA.preserve<MemorySSAAnalysis>();
1672   return PA;
1673 }
1674 
1675 void EarlyCSEPass::printPipeline(
1676     raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
1677   static_cast<PassInfoMixin<EarlyCSEPass> *>(this)->printPipeline(
1678       OS, MapClassName2PassName);
1679   OS << "<";
1680   if (UseMemorySSA)
1681     OS << "memssa";
1682   OS << ">";
1683 }
1684 
1685 namespace {
1686 
1687 /// A simple and fast domtree-based CSE pass.
1688 ///
1689 /// This pass does a simple depth-first walk over the dominator tree,
1690 /// eliminating trivially redundant instructions and using instsimplify to
1691 /// canonicalize things as it goes. It is intended to be fast and catch obvious
1692 /// cases so that instcombine and other passes are more effective. It is
1693 /// expected that a later pass of GVN will catch the interesting/hard cases.
1694 template<bool UseMemorySSA>
1695 class EarlyCSELegacyCommonPass : public FunctionPass {
1696 public:
1697   static char ID;
1698 
1699   EarlyCSELegacyCommonPass() : FunctionPass(ID) {
1700     if (UseMemorySSA)
1701       initializeEarlyCSEMemSSALegacyPassPass(*PassRegistry::getPassRegistry());
1702     else
1703       initializeEarlyCSELegacyPassPass(*PassRegistry::getPassRegistry());
1704   }
1705 
1706   bool runOnFunction(Function &F) override {
1707     if (skipFunction(F))
1708       return false;
1709 
1710     auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1711     auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1712     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1713     auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1714     auto *MSSA =
1715         UseMemorySSA ? &getAnalysis<MemorySSAWrapperPass>().getMSSA() : nullptr;
1716 
1717     EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA);
1718 
1719     return CSE.run();
1720   }
1721 
1722   void getAnalysisUsage(AnalysisUsage &AU) const override {
1723     AU.addRequired<AssumptionCacheTracker>();
1724     AU.addRequired<DominatorTreeWrapperPass>();
1725     AU.addRequired<TargetLibraryInfoWrapperPass>();
1726     AU.addRequired<TargetTransformInfoWrapperPass>();
1727     if (UseMemorySSA) {
1728       AU.addRequired<AAResultsWrapperPass>();
1729       AU.addRequired<MemorySSAWrapperPass>();
1730       AU.addPreserved<MemorySSAWrapperPass>();
1731     }
1732     AU.addPreserved<GlobalsAAWrapperPass>();
1733     AU.addPreserved<AAResultsWrapperPass>();
1734     AU.setPreservesCFG();
1735   }
1736 };
1737 
1738 } // end anonymous namespace
1739 
1740 using EarlyCSELegacyPass = EarlyCSELegacyCommonPass</*UseMemorySSA=*/false>;
1741 
1742 template<>
1743 char EarlyCSELegacyPass::ID = 0;
1744 
1745 INITIALIZE_PASS_BEGIN(EarlyCSELegacyPass, "early-cse", "Early CSE", false,
1746                       false)
1747 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1748 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1749 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1750 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1751 INITIALIZE_PASS_END(EarlyCSELegacyPass, "early-cse", "Early CSE", false, false)
1752 
1753 using EarlyCSEMemSSALegacyPass =
1754     EarlyCSELegacyCommonPass</*UseMemorySSA=*/true>;
1755 
1756 template<>
1757 char EarlyCSEMemSSALegacyPass::ID = 0;
1758 
1759 FunctionPass *llvm::createEarlyCSEPass(bool UseMemorySSA) {
1760   if (UseMemorySSA)
1761     return new EarlyCSEMemSSALegacyPass();
1762   else
1763     return new EarlyCSELegacyPass();
1764 }
1765 
1766 INITIALIZE_PASS_BEGIN(EarlyCSEMemSSALegacyPass, "early-cse-memssa",
1767                       "Early CSE w/ MemorySSA", false, false)
1768 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1769 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1770 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1771 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1772 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1773 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
1774 INITIALIZE_PASS_END(EarlyCSEMemSSALegacyPass, "early-cse-memssa",
1775                     "Early CSE w/ MemorySSA", false, false)
1776