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/SetVector.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/AssumptionCache.h"
23 #include "llvm/Analysis/GlobalsModRef.h"
24 #include "llvm/Analysis/GuardUtils.h"
25 #include "llvm/Analysis/InstructionSimplify.h"
26 #include "llvm/Analysis/MemorySSA.h"
27 #include "llvm/Analysis/MemorySSAUpdater.h"
28 #include "llvm/Analysis/TargetLibraryInfo.h"
29 #include "llvm/Analysis/TargetTransformInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/Constants.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/InstrTypes.h"
37 #include "llvm/IR/Instruction.h"
38 #include "llvm/IR/Instructions.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/IR/PassManager.h"
43 #include "llvm/IR/PatternMatch.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/IR/Use.h"
46 #include "llvm/IR/Value.h"
47 #include "llvm/InitializePasses.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/Allocator.h"
50 #include "llvm/Support/AtomicOrdering.h"
51 #include "llvm/Support/Casting.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/DebugCounter.h"
54 #include "llvm/Support/RecyclingAllocator.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Scalar.h"
57 #include "llvm/Transforms/Utils/GuardUtils.h"
58 #include "llvm/Transforms/Utils/Local.h"
59 #include <cassert>
60 #include <deque>
61 #include <memory>
62 #include <utility>
63 
64 using namespace llvm;
65 using namespace llvm::PatternMatch;
66 
67 #define DEBUG_TYPE "early-cse"
68 
69 STATISTIC(NumSimplify, "Number of instructions simplified or DCE'd");
70 STATISTIC(NumCSE,      "Number of instructions CSE'd");
71 STATISTIC(NumCSECVP,   "Number of compare instructions CVP'd");
72 STATISTIC(NumCSELoad,  "Number of load instructions CSE'd");
73 STATISTIC(NumCSECall,  "Number of call instructions CSE'd");
74 STATISTIC(NumDSE,      "Number of trivial dead stores removed");
75 
76 DEBUG_COUNTER(CSECounter, "early-cse",
77               "Controls which instructions are removed");
78 
79 static cl::opt<unsigned> EarlyCSEMssaOptCap(
80     "earlycse-mssa-optimization-cap", cl::init(500), cl::Hidden,
81     cl::desc("Enable imprecision in EarlyCSE in pathological cases, in exchange "
82              "for faster compile. Caps the MemorySSA clobbering calls."));
83 
84 static cl::opt<bool> EarlyCSEDebugHash(
85     "earlycse-debug-hash", cl::init(false), cl::Hidden,
86     cl::desc("Perform extra assertion checking to verify that SimpleValue's hash "
87              "function is well-behaved w.r.t. its isEqual predicate"));
88 
89 //===----------------------------------------------------------------------===//
90 // SimpleValue
91 //===----------------------------------------------------------------------===//
92 
93 namespace {
94 
95 /// Struct representing the available values in the scoped hash table.
96 struct SimpleValue {
97   Instruction *Inst;
98 
99   SimpleValue(Instruction *I) : Inst(I) {
100     assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
101   }
102 
103   bool isSentinel() const {
104     return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
105            Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
106   }
107 
108   static bool canHandle(Instruction *Inst) {
109     // This can only handle non-void readnone functions.
110     if (CallInst *CI = dyn_cast<CallInst>(Inst))
111       return CI->doesNotAccessMemory() && !CI->getType()->isVoidTy();
112     return isa<CastInst>(Inst) || isa<UnaryOperator>(Inst) ||
113            isa<BinaryOperator>(Inst) || isa<GetElementPtrInst>(Inst) ||
114            isa<CmpInst>(Inst) || isa<SelectInst>(Inst) ||
115            isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) ||
116            isa<ShuffleVectorInst>(Inst) || isa<ExtractValueInst>(Inst) ||
117            isa<InsertValueInst>(Inst);
118   }
119 };
120 
121 } // end anonymous namespace
122 
123 namespace llvm {
124 
125 template <> struct DenseMapInfo<SimpleValue> {
126   static inline SimpleValue getEmptyKey() {
127     return DenseMapInfo<Instruction *>::getEmptyKey();
128   }
129 
130   static inline SimpleValue getTombstoneKey() {
131     return DenseMapInfo<Instruction *>::getTombstoneKey();
132   }
133 
134   static unsigned getHashValue(SimpleValue Val);
135   static bool isEqual(SimpleValue LHS, SimpleValue RHS);
136 };
137 
138 } // end namespace llvm
139 
140 /// Match a 'select' including an optional 'not's of the condition.
141 static bool matchSelectWithOptionalNotCond(Value *V, Value *&Cond, Value *&A,
142                                            Value *&B,
143                                            SelectPatternFlavor &Flavor) {
144   // Return false if V is not even a select.
145   if (!match(V, m_Select(m_Value(Cond), m_Value(A), m_Value(B))))
146     return false;
147 
148   // Look through a 'not' of the condition operand by swapping A/B.
149   Value *CondNot;
150   if (match(Cond, m_Not(m_Value(CondNot)))) {
151     Cond = CondNot;
152     std::swap(A, B);
153   }
154 
155   // Match canonical forms of abs/nabs/min/max. We are not using ValueTracking's
156   // more powerful matchSelectPattern() because it may rely on instruction flags
157   // such as "nsw". That would be incompatible with the current hashing
158   // mechanism that may remove flags to increase the likelihood of CSE.
159 
160   // These are the canonical forms of abs(X) and nabs(X) created by instcombine:
161   // %N = sub i32 0, %X
162   // %C = icmp slt i32 %X, 0
163   // %ABS = select i1 %C, i32 %N, i32 %X
164   //
165   // %N = sub i32 0, %X
166   // %C = icmp slt i32 %X, 0
167   // %NABS = select i1 %C, i32 %X, i32 %N
168   Flavor = SPF_UNKNOWN;
169   CmpInst::Predicate Pred;
170   if (match(Cond, m_ICmp(Pred, m_Specific(B), m_ZeroInt())) &&
171       Pred == ICmpInst::ICMP_SLT && match(A, m_Neg(m_Specific(B)))) {
172     // ABS: B < 0 ? -B : B
173     Flavor = SPF_ABS;
174     return true;
175   }
176   if (match(Cond, m_ICmp(Pred, m_Specific(A), m_ZeroInt())) &&
177       Pred == ICmpInst::ICMP_SLT && match(B, m_Neg(m_Specific(A)))) {
178     // NABS: A < 0 ? A : -A
179     Flavor = SPF_NABS;
180     return true;
181   }
182 
183   if (!match(Cond, m_ICmp(Pred, m_Specific(A), m_Specific(B)))) {
184     // Check for commuted variants of min/max by swapping predicate.
185     // If we do not match the standard or commuted patterns, this is not a
186     // recognized form of min/max, but it is still a select, so return true.
187     if (!match(Cond, m_ICmp(Pred, m_Specific(B), m_Specific(A))))
188       return true;
189     Pred = ICmpInst::getSwappedPredicate(Pred);
190   }
191 
192   switch (Pred) {
193   case CmpInst::ICMP_UGT: Flavor = SPF_UMAX; break;
194   case CmpInst::ICMP_ULT: Flavor = SPF_UMIN; break;
195   case CmpInst::ICMP_SGT: Flavor = SPF_SMAX; break;
196   case CmpInst::ICMP_SLT: Flavor = SPF_SMIN; break;
197   default: break;
198   }
199 
200   return true;
201 }
202 
203 static unsigned getHashValueImpl(SimpleValue Val) {
204   Instruction *Inst = Val.Inst;
205   // Hash in all of the operands as pointers.
206   if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst)) {
207     Value *LHS = BinOp->getOperand(0);
208     Value *RHS = BinOp->getOperand(1);
209     if (BinOp->isCommutative() && BinOp->getOperand(0) > BinOp->getOperand(1))
210       std::swap(LHS, RHS);
211 
212     return hash_combine(BinOp->getOpcode(), LHS, RHS);
213   }
214 
215   if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) {
216     // Compares can be commuted by swapping the comparands and
217     // updating the predicate.  Choose the form that has the
218     // comparands in sorted order, or in the case of a tie, the
219     // one with the lower predicate.
220     Value *LHS = CI->getOperand(0);
221     Value *RHS = CI->getOperand(1);
222     CmpInst::Predicate Pred = CI->getPredicate();
223     CmpInst::Predicate SwappedPred = CI->getSwappedPredicate();
224     if (std::tie(LHS, Pred) > std::tie(RHS, SwappedPred)) {
225       std::swap(LHS, RHS);
226       Pred = SwappedPred;
227     }
228     return hash_combine(Inst->getOpcode(), Pred, LHS, RHS);
229   }
230 
231   // Hash general selects to allow matching commuted true/false operands.
232   SelectPatternFlavor SPF;
233   Value *Cond, *A, *B;
234   if (matchSelectWithOptionalNotCond(Inst, Cond, A, B, SPF)) {
235     // Hash min/max/abs (cmp + select) to allow for commuted operands.
236     // Min/max may also have non-canonical compare predicate (eg, the compare for
237     // smin may use 'sgt' rather than 'slt'), and non-canonical operands in the
238     // compare.
239     // TODO: We should also detect FP min/max.
240     if (SPF == SPF_SMIN || SPF == SPF_SMAX ||
241         SPF == SPF_UMIN || SPF == SPF_UMAX) {
242       if (A > B)
243         std::swap(A, B);
244       return hash_combine(Inst->getOpcode(), SPF, A, B);
245     }
246     if (SPF == SPF_ABS || SPF == SPF_NABS) {
247       // ABS/NABS always puts the input in A and its negation in B.
248       return hash_combine(Inst->getOpcode(), SPF, A, B);
249     }
250 
251     // Hash general selects to allow matching commuted true/false operands.
252 
253     // If we do not have a compare as the condition, just hash in the condition.
254     CmpInst::Predicate Pred;
255     Value *X, *Y;
256     if (!match(Cond, m_Cmp(Pred, m_Value(X), m_Value(Y))))
257       return hash_combine(Inst->getOpcode(), Cond, A, B);
258 
259     // Similar to cmp normalization (above) - canonicalize the predicate value:
260     // select (icmp Pred, X, Y), A, B --> select (icmp InvPred, X, Y), B, A
261     if (CmpInst::getInversePredicate(Pred) < Pred) {
262       Pred = CmpInst::getInversePredicate(Pred);
263       std::swap(A, B);
264     }
265     return hash_combine(Inst->getOpcode(), Pred, X, Y, A, B);
266   }
267 
268   if (CastInst *CI = dyn_cast<CastInst>(Inst))
269     return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0));
270 
271   if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst))
272     return hash_combine(EVI->getOpcode(), EVI->getOperand(0),
273                         hash_combine_range(EVI->idx_begin(), EVI->idx_end()));
274 
275   if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst))
276     return hash_combine(IVI->getOpcode(), IVI->getOperand(0),
277                         IVI->getOperand(1),
278                         hash_combine_range(IVI->idx_begin(), IVI->idx_end()));
279 
280   assert((isa<CallInst>(Inst) || isa<GetElementPtrInst>(Inst) ||
281           isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) ||
282           isa<ShuffleVectorInst>(Inst) || isa<UnaryOperator>(Inst)) &&
283          "Invalid/unknown instruction");
284 
285   // Mix in the opcode.
286   return hash_combine(
287       Inst->getOpcode(),
288       hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
289 }
290 
291 unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) {
292 #ifndef NDEBUG
293   // If -earlycse-debug-hash was specified, return a constant -- this
294   // will force all hashing to collide, so we'll exhaustively search
295   // the table for a match, and the assertion in isEqual will fire if
296   // there's a bug causing equal keys to hash differently.
297   if (EarlyCSEDebugHash)
298     return 0;
299 #endif
300   return getHashValueImpl(Val);
301 }
302 
303 static bool isEqualImpl(SimpleValue LHS, SimpleValue RHS) {
304   Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
305 
306   if (LHS.isSentinel() || RHS.isSentinel())
307     return LHSI == RHSI;
308 
309   if (LHSI->getOpcode() != RHSI->getOpcode())
310     return false;
311   if (LHSI->isIdenticalToWhenDefined(RHSI))
312     return true;
313 
314   // If we're not strictly identical, we still might be a commutable instruction
315   if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) {
316     if (!LHSBinOp->isCommutative())
317       return false;
318 
319     assert(isa<BinaryOperator>(RHSI) &&
320            "same opcode, but different instruction type?");
321     BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI);
322 
323     // Commuted equality
324     return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) &&
325            LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0);
326   }
327   if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) {
328     assert(isa<CmpInst>(RHSI) &&
329            "same opcode, but different instruction type?");
330     CmpInst *RHSCmp = cast<CmpInst>(RHSI);
331     // Commuted equality
332     return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) &&
333            LHSCmp->getOperand(1) == RHSCmp->getOperand(0) &&
334            LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate();
335   }
336 
337   // Min/max/abs can occur with commuted operands, non-canonical predicates,
338   // and/or non-canonical operands.
339   // Selects can be non-trivially equivalent via inverted conditions and swaps.
340   SelectPatternFlavor LSPF, RSPF;
341   Value *CondL, *CondR, *LHSA, *RHSA, *LHSB, *RHSB;
342   if (matchSelectWithOptionalNotCond(LHSI, CondL, LHSA, LHSB, LSPF) &&
343       matchSelectWithOptionalNotCond(RHSI, CondR, RHSA, RHSB, RSPF)) {
344     if (LSPF == RSPF) {
345       // TODO: We should also detect FP min/max.
346       if (LSPF == SPF_SMIN || LSPF == SPF_SMAX ||
347           LSPF == SPF_UMIN || LSPF == SPF_UMAX)
348         return ((LHSA == RHSA && LHSB == RHSB) ||
349                 (LHSA == RHSB && LHSB == RHSA));
350 
351       if (LSPF == SPF_ABS || LSPF == SPF_NABS) {
352         // Abs results are placed in a defined order by matchSelectPattern.
353         return LHSA == RHSA && LHSB == RHSB;
354       }
355 
356       // select Cond, A, B <--> select not(Cond), B, A
357       if (CondL == CondR && LHSA == RHSA && LHSB == RHSB)
358         return true;
359     }
360 
361     // If the true/false operands are swapped and the conditions are compares
362     // with inverted predicates, the selects are equal:
363     // select (icmp Pred, X, Y), A, B <--> select (icmp InvPred, X, Y), B, A
364     //
365     // This also handles patterns with a double-negation in the sense of not +
366     // inverse, because we looked through a 'not' in the matching function and
367     // swapped A/B:
368     // select (cmp Pred, X, Y), A, B <--> select (not (cmp InvPred, X, Y)), B, A
369     //
370     // This intentionally does NOT handle patterns with a double-negation in
371     // the sense of not + not, because doing so could result in values
372     // comparing
373     // as equal that hash differently in the min/max/abs cases like:
374     // select (cmp slt, X, Y), X, Y <--> select (not (not (cmp slt, X, Y))), X, Y
375     //   ^ hashes as min                  ^ would not hash as min
376     // In the context of the EarlyCSE pass, however, such cases never reach
377     // this code, as we simplify the double-negation before hashing the second
378     // select (and so still succeed at CSEing them).
379     if (LHSA == RHSB && LHSB == RHSA) {
380       CmpInst::Predicate PredL, PredR;
381       Value *X, *Y;
382       if (match(CondL, m_Cmp(PredL, m_Value(X), m_Value(Y))) &&
383           match(CondR, m_Cmp(PredR, m_Specific(X), m_Specific(Y))) &&
384           CmpInst::getInversePredicate(PredL) == PredR)
385         return true;
386     }
387   }
388 
389   return false;
390 }
391 
392 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) {
393   // These comparisons are nontrivial, so assert that equality implies
394   // hash equality (DenseMap demands this as an invariant).
395   bool Result = isEqualImpl(LHS, RHS);
396   assert(!Result || (LHS.isSentinel() && LHS.Inst == RHS.Inst) ||
397          getHashValueImpl(LHS) == getHashValueImpl(RHS));
398   return Result;
399 }
400 
401 //===----------------------------------------------------------------------===//
402 // CallValue
403 //===----------------------------------------------------------------------===//
404 
405 namespace {
406 
407 /// Struct representing the available call values in the scoped hash
408 /// table.
409 struct CallValue {
410   Instruction *Inst;
411 
412   CallValue(Instruction *I) : Inst(I) {
413     assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
414   }
415 
416   bool isSentinel() const {
417     return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
418            Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
419   }
420 
421   static bool canHandle(Instruction *Inst) {
422     // Don't value number anything that returns void.
423     if (Inst->getType()->isVoidTy())
424       return false;
425 
426     CallInst *CI = dyn_cast<CallInst>(Inst);
427     if (!CI || !CI->onlyReadsMemory())
428       return false;
429     return true;
430   }
431 };
432 
433 } // end anonymous namespace
434 
435 namespace llvm {
436 
437 template <> struct DenseMapInfo<CallValue> {
438   static inline CallValue getEmptyKey() {
439     return DenseMapInfo<Instruction *>::getEmptyKey();
440   }
441 
442   static inline CallValue getTombstoneKey() {
443     return DenseMapInfo<Instruction *>::getTombstoneKey();
444   }
445 
446   static unsigned getHashValue(CallValue Val);
447   static bool isEqual(CallValue LHS, CallValue RHS);
448 };
449 
450 } // end namespace llvm
451 
452 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) {
453   Instruction *Inst = Val.Inst;
454   // Hash all of the operands as pointers and mix in the opcode.
455   return hash_combine(
456       Inst->getOpcode(),
457       hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
458 }
459 
460 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) {
461   Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
462   if (LHS.isSentinel() || RHS.isSentinel())
463     return LHSI == RHSI;
464   return LHSI->isIdenticalTo(RHSI);
465 }
466 
467 //===----------------------------------------------------------------------===//
468 // EarlyCSE implementation
469 //===----------------------------------------------------------------------===//
470 
471 namespace {
472 
473 /// A simple and fast domtree-based CSE pass.
474 ///
475 /// This pass does a simple depth-first walk over the dominator tree,
476 /// eliminating trivially redundant instructions and using instsimplify to
477 /// canonicalize things as it goes. It is intended to be fast and catch obvious
478 /// cases so that instcombine and other passes are more effective. It is
479 /// expected that a later pass of GVN will catch the interesting/hard cases.
480 class EarlyCSE {
481 public:
482   const TargetLibraryInfo &TLI;
483   const TargetTransformInfo &TTI;
484   DominatorTree &DT;
485   AssumptionCache &AC;
486   const SimplifyQuery SQ;
487   MemorySSA *MSSA;
488   std::unique_ptr<MemorySSAUpdater> MSSAUpdater;
489 
490   using AllocatorTy =
491       RecyclingAllocator<BumpPtrAllocator,
492                          ScopedHashTableVal<SimpleValue, Value *>>;
493   using ScopedHTType =
494       ScopedHashTable<SimpleValue, Value *, DenseMapInfo<SimpleValue>,
495                       AllocatorTy>;
496 
497   /// A scoped hash table of the current values of all of our simple
498   /// scalar expressions.
499   ///
500   /// As we walk down the domtree, we look to see if instructions are in this:
501   /// if so, we replace them with what we find, otherwise we insert them so
502   /// that dominated values can succeed in their lookup.
503   ScopedHTType AvailableValues;
504 
505   /// A scoped hash table of the current values of previously encountered
506   /// memory locations.
507   ///
508   /// This allows us to get efficient access to dominating loads or stores when
509   /// we have a fully redundant load.  In addition to the most recent load, we
510   /// keep track of a generation count of the read, which is compared against
511   /// the current generation count.  The current generation count is incremented
512   /// after every possibly writing memory operation, which ensures that we only
513   /// CSE loads with other loads that have no intervening store.  Ordering
514   /// events (such as fences or atomic instructions) increment the generation
515   /// count as well; essentially, we model these as writes to all possible
516   /// locations.  Note that atomic and/or volatile loads and stores can be
517   /// present the table; it is the responsibility of the consumer to inspect
518   /// the atomicity/volatility if needed.
519   struct LoadValue {
520     Instruction *DefInst = nullptr;
521     unsigned Generation = 0;
522     int MatchingId = -1;
523     bool IsAtomic = false;
524 
525     LoadValue() = default;
526     LoadValue(Instruction *Inst, unsigned Generation, unsigned MatchingId,
527               bool IsAtomic)
528         : DefInst(Inst), Generation(Generation), MatchingId(MatchingId),
529           IsAtomic(IsAtomic) {}
530   };
531 
532   using LoadMapAllocator =
533       RecyclingAllocator<BumpPtrAllocator,
534                          ScopedHashTableVal<Value *, LoadValue>>;
535   using LoadHTType =
536       ScopedHashTable<Value *, LoadValue, DenseMapInfo<Value *>,
537                       LoadMapAllocator>;
538 
539   LoadHTType AvailableLoads;
540 
541   // A scoped hash table mapping memory locations (represented as typed
542   // addresses) to generation numbers at which that memory location became
543   // (henceforth indefinitely) invariant.
544   using InvariantMapAllocator =
545       RecyclingAllocator<BumpPtrAllocator,
546                          ScopedHashTableVal<MemoryLocation, unsigned>>;
547   using InvariantHTType =
548       ScopedHashTable<MemoryLocation, unsigned, DenseMapInfo<MemoryLocation>,
549                       InvariantMapAllocator>;
550   InvariantHTType AvailableInvariants;
551 
552   /// A scoped hash table of the current values of read-only call
553   /// values.
554   ///
555   /// It uses the same generation count as loads.
556   using CallHTType =
557       ScopedHashTable<CallValue, std::pair<Instruction *, unsigned>>;
558   CallHTType AvailableCalls;
559 
560   /// This is the current generation of the memory value.
561   unsigned CurrentGeneration = 0;
562 
563   /// Set up the EarlyCSE runner for a particular function.
564   EarlyCSE(const DataLayout &DL, const TargetLibraryInfo &TLI,
565            const TargetTransformInfo &TTI, DominatorTree &DT,
566            AssumptionCache &AC, MemorySSA *MSSA)
567       : TLI(TLI), TTI(TTI), DT(DT), AC(AC), SQ(DL, &TLI, &DT, &AC), MSSA(MSSA),
568         MSSAUpdater(std::make_unique<MemorySSAUpdater>(MSSA)) {}
569 
570   bool run();
571 
572 private:
573   unsigned ClobberCounter = 0;
574   // Almost a POD, but needs to call the constructors for the scoped hash
575   // tables so that a new scope gets pushed on. These are RAII so that the
576   // scope gets popped when the NodeScope is destroyed.
577   class NodeScope {
578   public:
579     NodeScope(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
580               InvariantHTType &AvailableInvariants, CallHTType &AvailableCalls)
581       : Scope(AvailableValues), LoadScope(AvailableLoads),
582         InvariantScope(AvailableInvariants), CallScope(AvailableCalls) {}
583     NodeScope(const NodeScope &) = delete;
584     NodeScope &operator=(const NodeScope &) = delete;
585 
586   private:
587     ScopedHTType::ScopeTy Scope;
588     LoadHTType::ScopeTy LoadScope;
589     InvariantHTType::ScopeTy InvariantScope;
590     CallHTType::ScopeTy CallScope;
591   };
592 
593   // Contains all the needed information to create a stack for doing a depth
594   // first traversal of the tree. This includes scopes for values, loads, and
595   // calls as well as the generation. There is a child iterator so that the
596   // children do not need to be store separately.
597   class StackNode {
598   public:
599     StackNode(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
600               InvariantHTType &AvailableInvariants, CallHTType &AvailableCalls,
601               unsigned cg, DomTreeNode *n, DomTreeNode::iterator child,
602               DomTreeNode::iterator end)
603         : CurrentGeneration(cg), ChildGeneration(cg), Node(n), ChildIter(child),
604           EndIter(end),
605           Scopes(AvailableValues, AvailableLoads, AvailableInvariants,
606                  AvailableCalls)
607           {}
608     StackNode(const StackNode &) = delete;
609     StackNode &operator=(const StackNode &) = delete;
610 
611     // Accessors.
612     unsigned currentGeneration() { return CurrentGeneration; }
613     unsigned childGeneration() { return ChildGeneration; }
614     void childGeneration(unsigned generation) { ChildGeneration = generation; }
615     DomTreeNode *node() { return Node; }
616     DomTreeNode::iterator childIter() { return ChildIter; }
617 
618     DomTreeNode *nextChild() {
619       DomTreeNode *child = *ChildIter;
620       ++ChildIter;
621       return child;
622     }
623 
624     DomTreeNode::iterator end() { return EndIter; }
625     bool isProcessed() { return Processed; }
626     void process() { Processed = true; }
627 
628   private:
629     unsigned CurrentGeneration;
630     unsigned ChildGeneration;
631     DomTreeNode *Node;
632     DomTreeNode::iterator ChildIter;
633     DomTreeNode::iterator EndIter;
634     NodeScope Scopes;
635     bool Processed = false;
636   };
637 
638   /// Wrapper class to handle memory instructions, including loads,
639   /// stores and intrinsic loads and stores defined by the target.
640   class ParseMemoryInst {
641   public:
642     ParseMemoryInst(Instruction *Inst, const TargetTransformInfo &TTI)
643       : Inst(Inst) {
644       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
645         if (TTI.getTgtMemIntrinsic(II, Info))
646           IsTargetMemInst = true;
647     }
648 
649     bool isLoad() const {
650       if (IsTargetMemInst) return Info.ReadMem;
651       return isa<LoadInst>(Inst);
652     }
653 
654     bool isStore() const {
655       if (IsTargetMemInst) return Info.WriteMem;
656       return isa<StoreInst>(Inst);
657     }
658 
659     bool isAtomic() const {
660       if (IsTargetMemInst)
661         return Info.Ordering != AtomicOrdering::NotAtomic;
662       return Inst->isAtomic();
663     }
664 
665     bool isUnordered() const {
666       if (IsTargetMemInst)
667         return Info.isUnordered();
668 
669       if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
670         return LI->isUnordered();
671       } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
672         return SI->isUnordered();
673       }
674       // Conservative answer
675       return !Inst->isAtomic();
676     }
677 
678     bool isVolatile() const {
679       if (IsTargetMemInst)
680         return Info.IsVolatile;
681 
682       if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
683         return LI->isVolatile();
684       } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
685         return SI->isVolatile();
686       }
687       // Conservative answer
688       return true;
689     }
690 
691     bool isInvariantLoad() const {
692       if (auto *LI = dyn_cast<LoadInst>(Inst))
693         return LI->hasMetadata(LLVMContext::MD_invariant_load);
694       return false;
695     }
696 
697     bool isMatchingMemLoc(const ParseMemoryInst &Inst) const {
698       return (getPointerOperand() == Inst.getPointerOperand() &&
699               getMatchingId() == Inst.getMatchingId());
700     }
701 
702     bool isValid() const { return getPointerOperand() != nullptr; }
703 
704     // For regular (non-intrinsic) loads/stores, this is set to -1. For
705     // intrinsic loads/stores, the id is retrieved from the corresponding
706     // field in the MemIntrinsicInfo structure.  That field contains
707     // non-negative values only.
708     int getMatchingId() const {
709       if (IsTargetMemInst) return Info.MatchingId;
710       return -1;
711     }
712 
713     Value *getPointerOperand() const {
714       if (IsTargetMemInst) return Info.PtrVal;
715       return getLoadStorePointerOperand(Inst);
716     }
717 
718     bool mayReadFromMemory() const {
719       if (IsTargetMemInst) return Info.ReadMem;
720       return Inst->mayReadFromMemory();
721     }
722 
723     bool mayWriteToMemory() const {
724       if (IsTargetMemInst) return Info.WriteMem;
725       return Inst->mayWriteToMemory();
726     }
727 
728   private:
729     bool IsTargetMemInst = false;
730     MemIntrinsicInfo Info;
731     Instruction *Inst;
732   };
733 
734   bool processNode(DomTreeNode *Node);
735 
736   bool handleBranchCondition(Instruction *CondInst, const BranchInst *BI,
737                              const BasicBlock *BB, const BasicBlock *Pred);
738 
739   Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const {
740     if (auto *LI = dyn_cast<LoadInst>(Inst))
741       return LI;
742     if (auto *SI = dyn_cast<StoreInst>(Inst))
743       return SI->getValueOperand();
744     assert(isa<IntrinsicInst>(Inst) && "Instruction not supported");
745     return TTI.getOrCreateResultFromMemIntrinsic(cast<IntrinsicInst>(Inst),
746                                                  ExpectedType);
747   }
748 
749   /// Return true if the instruction is known to only operate on memory
750   /// provably invariant in the given "generation".
751   bool isOperatingOnInvariantMemAt(Instruction *I, unsigned GenAt);
752 
753   bool isSameMemGeneration(unsigned EarlierGeneration, unsigned LaterGeneration,
754                            Instruction *EarlierInst, Instruction *LaterInst);
755 
756   void removeMSSA(Instruction *Inst) {
757     if (!MSSA)
758       return;
759     if (VerifyMemorySSA)
760       MSSA->verifyMemorySSA();
761     // Removing a store here can leave MemorySSA in an unoptimized state by
762     // creating MemoryPhis that have identical arguments and by creating
763     // MemoryUses whose defining access is not an actual clobber. The phi case
764     // is handled by MemorySSA when passing OptimizePhis = true to
765     // removeMemoryAccess.  The non-optimized MemoryUse case is lazily updated
766     // by MemorySSA's getClobberingMemoryAccess.
767     MSSAUpdater->removeMemoryAccess(Inst, true);
768   }
769 };
770 
771 } // end anonymous namespace
772 
773 /// Determine if the memory referenced by LaterInst is from the same heap
774 /// version as EarlierInst.
775 /// This is currently called in two scenarios:
776 ///
777 ///   load p
778 ///   ...
779 ///   load p
780 ///
781 /// and
782 ///
783 ///   x = load p
784 ///   ...
785 ///   store x, p
786 ///
787 /// in both cases we want to verify that there are no possible writes to the
788 /// memory referenced by p between the earlier and later instruction.
789 bool EarlyCSE::isSameMemGeneration(unsigned EarlierGeneration,
790                                    unsigned LaterGeneration,
791                                    Instruction *EarlierInst,
792                                    Instruction *LaterInst) {
793   // Check the simple memory generation tracking first.
794   if (EarlierGeneration == LaterGeneration)
795     return true;
796 
797   if (!MSSA)
798     return false;
799 
800   // If MemorySSA has determined that one of EarlierInst or LaterInst does not
801   // read/write memory, then we can safely return true here.
802   // FIXME: We could be more aggressive when checking doesNotAccessMemory(),
803   // onlyReadsMemory(), mayReadFromMemory(), and mayWriteToMemory() in this pass
804   // by also checking the MemorySSA MemoryAccess on the instruction.  Initial
805   // experiments suggest this isn't worthwhile, at least for C/C++ code compiled
806   // with the default optimization pipeline.
807   auto *EarlierMA = MSSA->getMemoryAccess(EarlierInst);
808   if (!EarlierMA)
809     return true;
810   auto *LaterMA = MSSA->getMemoryAccess(LaterInst);
811   if (!LaterMA)
812     return true;
813 
814   // Since we know LaterDef dominates LaterInst and EarlierInst dominates
815   // LaterInst, if LaterDef dominates EarlierInst then it can't occur between
816   // EarlierInst and LaterInst and neither can any other write that potentially
817   // clobbers LaterInst.
818   MemoryAccess *LaterDef;
819   if (ClobberCounter < EarlyCSEMssaOptCap) {
820     LaterDef = MSSA->getWalker()->getClobberingMemoryAccess(LaterInst);
821     ClobberCounter++;
822   } else
823     LaterDef = LaterMA->getDefiningAccess();
824 
825   return MSSA->dominates(LaterDef, EarlierMA);
826 }
827 
828 bool EarlyCSE::isOperatingOnInvariantMemAt(Instruction *I, unsigned GenAt) {
829   // A location loaded from with an invariant_load is assumed to *never* change
830   // within the visible scope of the compilation.
831   if (auto *LI = dyn_cast<LoadInst>(I))
832     if (LI->hasMetadata(LLVMContext::MD_invariant_load))
833       return true;
834 
835   auto MemLocOpt = MemoryLocation::getOrNone(I);
836   if (!MemLocOpt)
837     // "target" intrinsic forms of loads aren't currently known to
838     // MemoryLocation::get.  TODO
839     return false;
840   MemoryLocation MemLoc = *MemLocOpt;
841   if (!AvailableInvariants.count(MemLoc))
842     return false;
843 
844   // Is the generation at which this became invariant older than the
845   // current one?
846   return AvailableInvariants.lookup(MemLoc) <= GenAt;
847 }
848 
849 bool EarlyCSE::handleBranchCondition(Instruction *CondInst,
850                                      const BranchInst *BI, const BasicBlock *BB,
851                                      const BasicBlock *Pred) {
852   assert(BI->isConditional() && "Should be a conditional branch!");
853   assert(BI->getCondition() == CondInst && "Wrong condition?");
854   assert(BI->getSuccessor(0) == BB || BI->getSuccessor(1) == BB);
855   auto *TorF = (BI->getSuccessor(0) == BB)
856                    ? ConstantInt::getTrue(BB->getContext())
857                    : ConstantInt::getFalse(BB->getContext());
858   auto MatchBinOp = [](Instruction *I, unsigned Opcode) {
859     if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(I))
860       return BOp->getOpcode() == Opcode;
861     return false;
862   };
863   // If the condition is AND operation, we can propagate its operands into the
864   // true branch. If it is OR operation, we can propagate them into the false
865   // branch.
866   unsigned PropagateOpcode =
867       (BI->getSuccessor(0) == BB) ? Instruction::And : Instruction::Or;
868 
869   bool MadeChanges = false;
870   SmallVector<Instruction *, 4> WorkList;
871   SmallPtrSet<Instruction *, 4> Visited;
872   WorkList.push_back(CondInst);
873   while (!WorkList.empty()) {
874     Instruction *Curr = WorkList.pop_back_val();
875 
876     AvailableValues.insert(Curr, TorF);
877     LLVM_DEBUG(dbgs() << "EarlyCSE CVP: Add conditional value for '"
878                       << Curr->getName() << "' as " << *TorF << " in "
879                       << BB->getName() << "\n");
880     if (!DebugCounter::shouldExecute(CSECounter)) {
881       LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
882     } else {
883       // Replace all dominated uses with the known value.
884       if (unsigned Count = replaceDominatedUsesWith(Curr, TorF, DT,
885                                                     BasicBlockEdge(Pred, BB))) {
886         NumCSECVP += Count;
887         MadeChanges = true;
888       }
889     }
890 
891     if (MatchBinOp(Curr, PropagateOpcode))
892       for (auto &Op : cast<BinaryOperator>(Curr)->operands())
893         if (Instruction *OPI = dyn_cast<Instruction>(Op))
894           if (SimpleValue::canHandle(OPI) && Visited.insert(OPI).second)
895             WorkList.push_back(OPI);
896   }
897 
898   return MadeChanges;
899 }
900 
901 bool EarlyCSE::processNode(DomTreeNode *Node) {
902   bool Changed = false;
903   BasicBlock *BB = Node->getBlock();
904 
905   // If this block has a single predecessor, then the predecessor is the parent
906   // of the domtree node and all of the live out memory values are still current
907   // in this block.  If this block has multiple predecessors, then they could
908   // have invalidated the live-out memory values of our parent value.  For now,
909   // just be conservative and invalidate memory if this block has multiple
910   // predecessors.
911   if (!BB->getSinglePredecessor())
912     ++CurrentGeneration;
913 
914   // If this node has a single predecessor which ends in a conditional branch,
915   // we can infer the value of the branch condition given that we took this
916   // path.  We need the single predecessor to ensure there's not another path
917   // which reaches this block where the condition might hold a different
918   // value.  Since we're adding this to the scoped hash table (like any other
919   // def), it will have been popped if we encounter a future merge block.
920   if (BasicBlock *Pred = BB->getSinglePredecessor()) {
921     auto *BI = dyn_cast<BranchInst>(Pred->getTerminator());
922     if (BI && BI->isConditional()) {
923       auto *CondInst = dyn_cast<Instruction>(BI->getCondition());
924       if (CondInst && SimpleValue::canHandle(CondInst))
925         Changed |= handleBranchCondition(CondInst, BI, BB, Pred);
926     }
927   }
928 
929   /// LastStore - Keep track of the last non-volatile store that we saw... for
930   /// as long as there in no instruction that reads memory.  If we see a store
931   /// to the same location, we delete the dead store.  This zaps trivial dead
932   /// stores which can occur in bitfield code among other things.
933   Instruction *LastStore = nullptr;
934 
935   // See if any instructions in the block can be eliminated.  If so, do it.  If
936   // not, add them to AvailableValues.
937   for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
938     Instruction *Inst = &*I++;
939 
940     // Dead instructions should just be removed.
941     if (isInstructionTriviallyDead(Inst, &TLI)) {
942       LLVM_DEBUG(dbgs() << "EarlyCSE DCE: " << *Inst << '\n');
943       if (!DebugCounter::shouldExecute(CSECounter)) {
944         LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
945         continue;
946       }
947 
948       salvageDebugInfoOrMarkUndef(*Inst);
949       removeMSSA(Inst);
950       Inst->eraseFromParent();
951       Changed = true;
952       ++NumSimplify;
953       continue;
954     }
955 
956     // Skip assume intrinsics, they don't really have side effects (although
957     // they're marked as such to ensure preservation of control dependencies),
958     // and this pass will not bother with its removal. However, we should mark
959     // its condition as true for all dominated blocks.
960     if (match(Inst, m_Intrinsic<Intrinsic::assume>())) {
961       auto *CondI =
962           dyn_cast<Instruction>(cast<CallInst>(Inst)->getArgOperand(0));
963       if (CondI && SimpleValue::canHandle(CondI)) {
964         LLVM_DEBUG(dbgs() << "EarlyCSE considering assumption: " << *Inst
965                           << '\n');
966         AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext()));
967       } else
968         LLVM_DEBUG(dbgs() << "EarlyCSE skipping assumption: " << *Inst << '\n');
969       continue;
970     }
971 
972     // Skip sideeffect intrinsics, for the same reason as assume intrinsics.
973     if (match(Inst, m_Intrinsic<Intrinsic::sideeffect>())) {
974       LLVM_DEBUG(dbgs() << "EarlyCSE skipping sideeffect: " << *Inst << '\n');
975       continue;
976     }
977 
978     // We can skip all invariant.start intrinsics since they only read memory,
979     // and we can forward values across it. For invariant starts without
980     // invariant ends, we can use the fact that the invariantness never ends to
981     // start a scope in the current generaton which is true for all future
982     // generations.  Also, we dont need to consume the last store since the
983     // semantics of invariant.start allow us to perform   DSE of the last
984     // store, if there was a store following invariant.start. Consider:
985     //
986     // store 30, i8* p
987     // invariant.start(p)
988     // store 40, i8* p
989     // We can DSE the store to 30, since the store 40 to invariant location p
990     // causes undefined behaviour.
991     if (match(Inst, m_Intrinsic<Intrinsic::invariant_start>())) {
992       // If there are any uses, the scope might end.
993       if (!Inst->use_empty())
994         continue;
995       auto *CI = cast<CallInst>(Inst);
996       MemoryLocation MemLoc = MemoryLocation::getForArgument(CI, 1, TLI);
997       // Don't start a scope if we already have a better one pushed
998       if (!AvailableInvariants.count(MemLoc))
999         AvailableInvariants.insert(MemLoc, CurrentGeneration);
1000       continue;
1001     }
1002 
1003     if (isGuard(Inst)) {
1004       if (auto *CondI =
1005               dyn_cast<Instruction>(cast<CallInst>(Inst)->getArgOperand(0))) {
1006         if (SimpleValue::canHandle(CondI)) {
1007           // Do we already know the actual value of this condition?
1008           if (auto *KnownCond = AvailableValues.lookup(CondI)) {
1009             // Is the condition known to be true?
1010             if (isa<ConstantInt>(KnownCond) &&
1011                 cast<ConstantInt>(KnownCond)->isOne()) {
1012               LLVM_DEBUG(dbgs()
1013                          << "EarlyCSE removing guard: " << *Inst << '\n');
1014               removeMSSA(Inst);
1015               Inst->eraseFromParent();
1016               Changed = true;
1017               continue;
1018             } else
1019               // Use the known value if it wasn't true.
1020               cast<CallInst>(Inst)->setArgOperand(0, KnownCond);
1021           }
1022           // The condition we're on guarding here is true for all dominated
1023           // locations.
1024           AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext()));
1025         }
1026       }
1027 
1028       // Guard intrinsics read all memory, but don't write any memory.
1029       // Accordingly, don't update the generation but consume the last store (to
1030       // avoid an incorrect DSE).
1031       LastStore = nullptr;
1032       continue;
1033     }
1034 
1035     // If the instruction can be simplified (e.g. X+0 = X) then replace it with
1036     // its simpler value.
1037     if (Value *V = SimplifyInstruction(Inst, SQ)) {
1038       LLVM_DEBUG(dbgs() << "EarlyCSE Simplify: " << *Inst << "  to: " << *V
1039                         << '\n');
1040       if (!DebugCounter::shouldExecute(CSECounter)) {
1041         LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1042       } else {
1043         bool Killed = false;
1044         if (!Inst->use_empty()) {
1045           Inst->replaceAllUsesWith(V);
1046           Changed = true;
1047         }
1048         if (isInstructionTriviallyDead(Inst, &TLI)) {
1049           removeMSSA(Inst);
1050           Inst->eraseFromParent();
1051           Changed = true;
1052           Killed = true;
1053         }
1054         if (Changed)
1055           ++NumSimplify;
1056         if (Killed)
1057           continue;
1058       }
1059     }
1060 
1061     // If this is a simple instruction that we can value number, process it.
1062     if (SimpleValue::canHandle(Inst)) {
1063       // See if the instruction has an available value.  If so, use it.
1064       if (Value *V = AvailableValues.lookup(Inst)) {
1065         LLVM_DEBUG(dbgs() << "EarlyCSE CSE: " << *Inst << "  to: " << *V
1066                           << '\n');
1067         if (!DebugCounter::shouldExecute(CSECounter)) {
1068           LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1069           continue;
1070         }
1071         if (auto *I = dyn_cast<Instruction>(V))
1072           I->andIRFlags(Inst);
1073         Inst->replaceAllUsesWith(V);
1074         removeMSSA(Inst);
1075         Inst->eraseFromParent();
1076         Changed = true;
1077         ++NumCSE;
1078         continue;
1079       }
1080 
1081       // Otherwise, just remember that this value is available.
1082       AvailableValues.insert(Inst, Inst);
1083       continue;
1084     }
1085 
1086     ParseMemoryInst MemInst(Inst, TTI);
1087     // If this is a non-volatile load, process it.
1088     if (MemInst.isValid() && MemInst.isLoad()) {
1089       // (conservatively) we can't peak past the ordering implied by this
1090       // operation, but we can add this load to our set of available values
1091       if (MemInst.isVolatile() || !MemInst.isUnordered()) {
1092         LastStore = nullptr;
1093         ++CurrentGeneration;
1094       }
1095 
1096       if (MemInst.isInvariantLoad()) {
1097         // If we pass an invariant load, we know that memory location is
1098         // indefinitely constant from the moment of first dereferenceability.
1099         // We conservatively treat the invariant_load as that moment.  If we
1100         // pass a invariant load after already establishing a scope, don't
1101         // restart it since we want to preserve the earliest point seen.
1102         auto MemLoc = MemoryLocation::get(Inst);
1103         if (!AvailableInvariants.count(MemLoc))
1104           AvailableInvariants.insert(MemLoc, CurrentGeneration);
1105       }
1106 
1107       // If we have an available version of this load, and if it is the right
1108       // generation or the load is known to be from an invariant location,
1109       // replace this instruction.
1110       //
1111       // If either the dominating load or the current load are invariant, then
1112       // we can assume the current load loads the same value as the dominating
1113       // load.
1114       LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand());
1115       if (InVal.DefInst != nullptr &&
1116           InVal.MatchingId == MemInst.getMatchingId() &&
1117           // We don't yet handle removing loads with ordering of any kind.
1118           !MemInst.isVolatile() && MemInst.isUnordered() &&
1119           // We can't replace an atomic load with one which isn't also atomic.
1120           InVal.IsAtomic >= MemInst.isAtomic() &&
1121           (isOperatingOnInvariantMemAt(Inst, InVal.Generation) ||
1122            isSameMemGeneration(InVal.Generation, CurrentGeneration,
1123                                InVal.DefInst, Inst))) {
1124         Value *Op = getOrCreateResult(InVal.DefInst, Inst->getType());
1125         if (Op != nullptr) {
1126           LLVM_DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << *Inst
1127                             << "  to: " << *InVal.DefInst << '\n');
1128           if (!DebugCounter::shouldExecute(CSECounter)) {
1129             LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1130             continue;
1131           }
1132           if (!Inst->use_empty())
1133             Inst->replaceAllUsesWith(Op);
1134           removeMSSA(Inst);
1135           Inst->eraseFromParent();
1136           Changed = true;
1137           ++NumCSELoad;
1138           continue;
1139         }
1140       }
1141 
1142       // Otherwise, remember that we have this instruction.
1143       AvailableLoads.insert(
1144           MemInst.getPointerOperand(),
1145           LoadValue(Inst, CurrentGeneration, MemInst.getMatchingId(),
1146                     MemInst.isAtomic()));
1147       LastStore = nullptr;
1148       continue;
1149     }
1150 
1151     // If this instruction may read from memory or throw (and potentially read
1152     // from memory in the exception handler), forget LastStore.  Load/store
1153     // intrinsics will indicate both a read and a write to memory.  The target
1154     // may override this (e.g. so that a store intrinsic does not read from
1155     // memory, and thus will be treated the same as a regular store for
1156     // commoning purposes).
1157     if ((Inst->mayReadFromMemory() || Inst->mayThrow()) &&
1158         !(MemInst.isValid() && !MemInst.mayReadFromMemory()))
1159       LastStore = nullptr;
1160 
1161     // If this is a read-only call, process it.
1162     if (CallValue::canHandle(Inst)) {
1163       // If we have an available version of this call, and if it is the right
1164       // generation, replace this instruction.
1165       std::pair<Instruction *, unsigned> InVal = AvailableCalls.lookup(Inst);
1166       if (InVal.first != nullptr &&
1167           isSameMemGeneration(InVal.second, CurrentGeneration, InVal.first,
1168                               Inst)) {
1169         LLVM_DEBUG(dbgs() << "EarlyCSE CSE CALL: " << *Inst
1170                           << "  to: " << *InVal.first << '\n');
1171         if (!DebugCounter::shouldExecute(CSECounter)) {
1172           LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1173           continue;
1174         }
1175         if (!Inst->use_empty())
1176           Inst->replaceAllUsesWith(InVal.first);
1177         removeMSSA(Inst);
1178         Inst->eraseFromParent();
1179         Changed = true;
1180         ++NumCSECall;
1181         continue;
1182       }
1183 
1184       // Otherwise, remember that we have this instruction.
1185       AvailableCalls.insert(
1186           Inst, std::pair<Instruction *, unsigned>(Inst, CurrentGeneration));
1187       continue;
1188     }
1189 
1190     // A release fence requires that all stores complete before it, but does
1191     // not prevent the reordering of following loads 'before' the fence.  As a
1192     // result, we don't need to consider it as writing to memory and don't need
1193     // to advance the generation.  We do need to prevent DSE across the fence,
1194     // but that's handled above.
1195     if (FenceInst *FI = dyn_cast<FenceInst>(Inst))
1196       if (FI->getOrdering() == AtomicOrdering::Release) {
1197         assert(Inst->mayReadFromMemory() && "relied on to prevent DSE above");
1198         continue;
1199       }
1200 
1201     // write back DSE - If we write back the same value we just loaded from
1202     // the same location and haven't passed any intervening writes or ordering
1203     // operations, we can remove the write.  The primary benefit is in allowing
1204     // the available load table to remain valid and value forward past where
1205     // the store originally was.
1206     if (MemInst.isValid() && MemInst.isStore()) {
1207       LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand());
1208       if (InVal.DefInst &&
1209           InVal.DefInst == getOrCreateResult(Inst, InVal.DefInst->getType()) &&
1210           InVal.MatchingId == MemInst.getMatchingId() &&
1211           // We don't yet handle removing stores with ordering of any kind.
1212           !MemInst.isVolatile() && MemInst.isUnordered() &&
1213           (isOperatingOnInvariantMemAt(Inst, InVal.Generation) ||
1214            isSameMemGeneration(InVal.Generation, CurrentGeneration,
1215                                InVal.DefInst, Inst))) {
1216         // It is okay to have a LastStore to a different pointer here if MemorySSA
1217         // tells us that the load and store are from the same memory generation.
1218         // In that case, LastStore should keep its present value since we're
1219         // removing the current store.
1220         assert((!LastStore ||
1221                 ParseMemoryInst(LastStore, TTI).getPointerOperand() ==
1222                     MemInst.getPointerOperand() ||
1223                 MSSA) &&
1224                "can't have an intervening store if not using MemorySSA!");
1225         LLVM_DEBUG(dbgs() << "EarlyCSE DSE (writeback): " << *Inst << '\n');
1226         if (!DebugCounter::shouldExecute(CSECounter)) {
1227           LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1228           continue;
1229         }
1230         removeMSSA(Inst);
1231         Inst->eraseFromParent();
1232         Changed = true;
1233         ++NumDSE;
1234         // We can avoid incrementing the generation count since we were able
1235         // to eliminate this store.
1236         continue;
1237       }
1238     }
1239 
1240     // Okay, this isn't something we can CSE at all.  Check to see if it is
1241     // something that could modify memory.  If so, our available memory values
1242     // cannot be used so bump the generation count.
1243     if (Inst->mayWriteToMemory()) {
1244       ++CurrentGeneration;
1245 
1246       if (MemInst.isValid() && MemInst.isStore()) {
1247         // We do a trivial form of DSE if there are two stores to the same
1248         // location with no intervening loads.  Delete the earlier store.
1249         // At the moment, we don't remove ordered stores, but do remove
1250         // unordered atomic stores.  There's no special requirement (for
1251         // unordered atomics) about removing atomic stores only in favor of
1252         // other atomic stores since we were going to execute the non-atomic
1253         // one anyway and the atomic one might never have become visible.
1254         if (LastStore) {
1255           ParseMemoryInst LastStoreMemInst(LastStore, TTI);
1256           assert(LastStoreMemInst.isUnordered() &&
1257                  !LastStoreMemInst.isVolatile() &&
1258                  "Violated invariant");
1259           if (LastStoreMemInst.isMatchingMemLoc(MemInst)) {
1260             LLVM_DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore
1261                               << "  due to: " << *Inst << '\n');
1262             if (!DebugCounter::shouldExecute(CSECounter)) {
1263               LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1264             } else {
1265               removeMSSA(LastStore);
1266               LastStore->eraseFromParent();
1267               Changed = true;
1268               ++NumDSE;
1269               LastStore = nullptr;
1270             }
1271           }
1272           // fallthrough - we can exploit information about this store
1273         }
1274 
1275         // Okay, we just invalidated anything we knew about loaded values.  Try
1276         // to salvage *something* by remembering that the stored value is a live
1277         // version of the pointer.  It is safe to forward from volatile stores
1278         // to non-volatile loads, so we don't have to check for volatility of
1279         // the store.
1280         AvailableLoads.insert(
1281             MemInst.getPointerOperand(),
1282             LoadValue(Inst, CurrentGeneration, MemInst.getMatchingId(),
1283                       MemInst.isAtomic()));
1284 
1285         // Remember that this was the last unordered store we saw for DSE. We
1286         // don't yet handle DSE on ordered or volatile stores since we don't
1287         // have a good way to model the ordering requirement for following
1288         // passes  once the store is removed.  We could insert a fence, but
1289         // since fences are slightly stronger than stores in their ordering,
1290         // it's not clear this is a profitable transform. Another option would
1291         // be to merge the ordering with that of the post dominating store.
1292         if (MemInst.isUnordered() && !MemInst.isVolatile())
1293           LastStore = Inst;
1294         else
1295           LastStore = nullptr;
1296       }
1297     }
1298   }
1299 
1300   return Changed;
1301 }
1302 
1303 bool EarlyCSE::run() {
1304   // Note, deque is being used here because there is significant performance
1305   // gains over vector when the container becomes very large due to the
1306   // specific access patterns. For more information see the mailing list
1307   // discussion on this:
1308   // http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html
1309   std::deque<StackNode *> nodesToProcess;
1310 
1311   bool Changed = false;
1312 
1313   // Process the root node.
1314   nodesToProcess.push_back(new StackNode(
1315       AvailableValues, AvailableLoads, AvailableInvariants, AvailableCalls,
1316       CurrentGeneration, DT.getRootNode(),
1317       DT.getRootNode()->begin(), DT.getRootNode()->end()));
1318 
1319   assert(!CurrentGeneration && "Create a new EarlyCSE instance to rerun it.");
1320 
1321   // Process the stack.
1322   while (!nodesToProcess.empty()) {
1323     // Grab the first item off the stack. Set the current generation, remove
1324     // the node from the stack, and process it.
1325     StackNode *NodeToProcess = nodesToProcess.back();
1326 
1327     // Initialize class members.
1328     CurrentGeneration = NodeToProcess->currentGeneration();
1329 
1330     // Check if the node needs to be processed.
1331     if (!NodeToProcess->isProcessed()) {
1332       // Process the node.
1333       Changed |= processNode(NodeToProcess->node());
1334       NodeToProcess->childGeneration(CurrentGeneration);
1335       NodeToProcess->process();
1336     } else if (NodeToProcess->childIter() != NodeToProcess->end()) {
1337       // Push the next child onto the stack.
1338       DomTreeNode *child = NodeToProcess->nextChild();
1339       nodesToProcess.push_back(
1340           new StackNode(AvailableValues, AvailableLoads, AvailableInvariants,
1341                         AvailableCalls, NodeToProcess->childGeneration(),
1342                         child, child->begin(), child->end()));
1343     } else {
1344       // It has been processed, and there are no more children to process,
1345       // so delete it and pop it off the stack.
1346       delete NodeToProcess;
1347       nodesToProcess.pop_back();
1348     }
1349   } // while (!nodes...)
1350 
1351   return Changed;
1352 }
1353 
1354 PreservedAnalyses EarlyCSEPass::run(Function &F,
1355                                     FunctionAnalysisManager &AM) {
1356   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1357   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
1358   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1359   auto &AC = AM.getResult<AssumptionAnalysis>(F);
1360   auto *MSSA =
1361       UseMemorySSA ? &AM.getResult<MemorySSAAnalysis>(F).getMSSA() : nullptr;
1362 
1363   EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA);
1364 
1365   if (!CSE.run())
1366     return PreservedAnalyses::all();
1367 
1368   PreservedAnalyses PA;
1369   PA.preserveSet<CFGAnalyses>();
1370   PA.preserve<GlobalsAA>();
1371   if (UseMemorySSA)
1372     PA.preserve<MemorySSAAnalysis>();
1373   return PA;
1374 }
1375 
1376 namespace {
1377 
1378 /// A simple and fast domtree-based CSE pass.
1379 ///
1380 /// This pass does a simple depth-first walk over the dominator tree,
1381 /// eliminating trivially redundant instructions and using instsimplify to
1382 /// canonicalize things as it goes. It is intended to be fast and catch obvious
1383 /// cases so that instcombine and other passes are more effective. It is
1384 /// expected that a later pass of GVN will catch the interesting/hard cases.
1385 template<bool UseMemorySSA>
1386 class EarlyCSELegacyCommonPass : public FunctionPass {
1387 public:
1388   static char ID;
1389 
1390   EarlyCSELegacyCommonPass() : FunctionPass(ID) {
1391     if (UseMemorySSA)
1392       initializeEarlyCSEMemSSALegacyPassPass(*PassRegistry::getPassRegistry());
1393     else
1394       initializeEarlyCSELegacyPassPass(*PassRegistry::getPassRegistry());
1395   }
1396 
1397   bool runOnFunction(Function &F) override {
1398     if (skipFunction(F))
1399       return false;
1400 
1401     auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1402     auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1403     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1404     auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1405     auto *MSSA =
1406         UseMemorySSA ? &getAnalysis<MemorySSAWrapperPass>().getMSSA() : nullptr;
1407 
1408     EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA);
1409 
1410     return CSE.run();
1411   }
1412 
1413   void getAnalysisUsage(AnalysisUsage &AU) const override {
1414     AU.addRequired<AssumptionCacheTracker>();
1415     AU.addRequired<DominatorTreeWrapperPass>();
1416     AU.addRequired<TargetLibraryInfoWrapperPass>();
1417     AU.addRequired<TargetTransformInfoWrapperPass>();
1418     if (UseMemorySSA) {
1419       AU.addRequired<MemorySSAWrapperPass>();
1420       AU.addPreserved<MemorySSAWrapperPass>();
1421     }
1422     AU.addPreserved<GlobalsAAWrapperPass>();
1423     AU.addPreserved<AAResultsWrapperPass>();
1424     AU.setPreservesCFG();
1425   }
1426 };
1427 
1428 } // end anonymous namespace
1429 
1430 using EarlyCSELegacyPass = EarlyCSELegacyCommonPass</*UseMemorySSA=*/false>;
1431 
1432 template<>
1433 char EarlyCSELegacyPass::ID = 0;
1434 
1435 INITIALIZE_PASS_BEGIN(EarlyCSELegacyPass, "early-cse", "Early CSE", false,
1436                       false)
1437 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1438 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1439 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1440 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1441 INITIALIZE_PASS_END(EarlyCSELegacyPass, "early-cse", "Early CSE", false, false)
1442 
1443 using EarlyCSEMemSSALegacyPass =
1444     EarlyCSELegacyCommonPass</*UseMemorySSA=*/true>;
1445 
1446 template<>
1447 char EarlyCSEMemSSALegacyPass::ID = 0;
1448 
1449 FunctionPass *llvm::createEarlyCSEPass(bool UseMemorySSA) {
1450   if (UseMemorySSA)
1451     return new EarlyCSEMemSSALegacyPass();
1452   else
1453     return new EarlyCSELegacyPass();
1454 }
1455 
1456 INITIALIZE_PASS_BEGIN(EarlyCSEMemSSALegacyPass, "early-cse-memssa",
1457                       "Early CSE w/ MemorySSA", false, false)
1458 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1459 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1460 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1461 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1462 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
1463 INITIALIZE_PASS_END(EarlyCSEMemSSALegacyPass, "early-cse-memssa",
1464                     "Early CSE w/ MemorySSA", false, false)
1465