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(NumCSEGEP, "Number of GEP instructions CSE'd");
71 STATISTIC(NumDSE,      "Number of trivial dead stores removed");
72 
73 DEBUG_COUNTER(CSECounter, "early-cse",
74               "Controls which instructions are removed");
75 
76 static cl::opt<unsigned> EarlyCSEMssaOptCap(
77     "earlycse-mssa-optimization-cap", cl::init(500), cl::Hidden,
78     cl::desc("Enable imprecision in EarlyCSE in pathological cases, in exchange "
79              "for faster compile. Caps the MemorySSA clobbering calls."));
80 
81 static cl::opt<bool> EarlyCSEDebugHash(
82     "earlycse-debug-hash", cl::init(false), cl::Hidden,
83     cl::desc("Perform extra assertion checking to verify that SimpleValue's hash "
84              "function is well-behaved w.r.t. its isEqual predicate"));
85 
86 //===----------------------------------------------------------------------===//
87 // SimpleValue
88 //===----------------------------------------------------------------------===//
89 
90 namespace {
91 
92 /// Struct representing the available values in the scoped hash table.
93 struct SimpleValue {
94   Instruction *Inst;
95 
96   SimpleValue(Instruction *I) : Inst(I) {
97     assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
98   }
99 
100   bool isSentinel() const {
101     return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
102            Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
103   }
104 
105   static bool canHandle(Instruction *Inst) {
106     // This can only handle non-void readnone functions.
107     // Also handled are constrained intrinsic that look like the types
108     // of instruction handled below (UnaryOperator, etc.).
109     if (CallInst *CI = dyn_cast<CallInst>(Inst)) {
110       if (Function *F = CI->getCalledFunction()) {
111         switch ((Intrinsic::ID)F->getIntrinsicID()) {
112         case Intrinsic::experimental_constrained_fadd:
113         case Intrinsic::experimental_constrained_fsub:
114         case Intrinsic::experimental_constrained_fmul:
115         case Intrinsic::experimental_constrained_fdiv:
116         case Intrinsic::experimental_constrained_frem:
117         case Intrinsic::experimental_constrained_fptosi:
118         case Intrinsic::experimental_constrained_sitofp:
119         case Intrinsic::experimental_constrained_fptoui:
120         case Intrinsic::experimental_constrained_uitofp:
121         case Intrinsic::experimental_constrained_fcmp:
122         case Intrinsic::experimental_constrained_fcmps: {
123           auto *CFP = cast<ConstrainedFPIntrinsic>(CI);
124           if (CFP->getExceptionBehavior() &&
125               CFP->getExceptionBehavior() == fp::ebStrict)
126             return false;
127           // Since we CSE across function calls we must not allow
128           // the rounding mode to change.
129           if (CFP->getRoundingMode() &&
130               CFP->getRoundingMode() == RoundingMode::Dynamic)
131             return false;
132           return true;
133         }
134         }
135       }
136       return CI->doesNotAccessMemory() && !CI->getType()->isVoidTy() &&
137              // FIXME: Currently the calls which may access the thread id may
138              // be considered as not accessing the memory. But this is
139              // problematic for coroutines, since coroutines may resume in a
140              // different thread. So we disable the optimization here for the
141              // correctness. However, it may block many other correct
142              // optimizations. Revert this one when we detect the memory
143              // accessing kind more precisely.
144              !CI->getFunction()->isPresplitCoroutine();
145     }
146     return isa<CastInst>(Inst) || isa<UnaryOperator>(Inst) ||
147            isa<BinaryOperator>(Inst) || isa<CmpInst>(Inst) ||
148            isa<SelectInst>(Inst) || isa<ExtractElementInst>(Inst) ||
149            isa<InsertElementInst>(Inst) || isa<ShuffleVectorInst>(Inst) ||
150            isa<ExtractValueInst>(Inst) || isa<InsertValueInst>(Inst) ||
151            isa<FreezeInst>(Inst);
152   }
153 };
154 
155 } // end anonymous namespace
156 
157 namespace llvm {
158 
159 template <> struct DenseMapInfo<SimpleValue> {
160   static inline SimpleValue getEmptyKey() {
161     return DenseMapInfo<Instruction *>::getEmptyKey();
162   }
163 
164   static inline SimpleValue getTombstoneKey() {
165     return DenseMapInfo<Instruction *>::getTombstoneKey();
166   }
167 
168   static unsigned getHashValue(SimpleValue Val);
169   static bool isEqual(SimpleValue LHS, SimpleValue RHS);
170 };
171 
172 } // end namespace llvm
173 
174 /// Match a 'select' including an optional 'not's of the condition.
175 static bool matchSelectWithOptionalNotCond(Value *V, Value *&Cond, Value *&A,
176                                            Value *&B,
177                                            SelectPatternFlavor &Flavor) {
178   // Return false if V is not even a select.
179   if (!match(V, m_Select(m_Value(Cond), m_Value(A), m_Value(B))))
180     return false;
181 
182   // Look through a 'not' of the condition operand by swapping A/B.
183   Value *CondNot;
184   if (match(Cond, m_Not(m_Value(CondNot)))) {
185     Cond = CondNot;
186     std::swap(A, B);
187   }
188 
189   // Match canonical forms of min/max. We are not using ValueTracking's
190   // more powerful matchSelectPattern() because it may rely on instruction flags
191   // such as "nsw". That would be incompatible with the current hashing
192   // mechanism that may remove flags to increase the likelihood of CSE.
193 
194   Flavor = SPF_UNKNOWN;
195   CmpInst::Predicate Pred;
196 
197   if (!match(Cond, m_ICmp(Pred, m_Specific(A), m_Specific(B)))) {
198     // Check for commuted variants of min/max by swapping predicate.
199     // If we do not match the standard or commuted patterns, this is not a
200     // recognized form of min/max, but it is still a select, so return true.
201     if (!match(Cond, m_ICmp(Pred, m_Specific(B), m_Specific(A))))
202       return true;
203     Pred = ICmpInst::getSwappedPredicate(Pred);
204   }
205 
206   switch (Pred) {
207   case CmpInst::ICMP_UGT: Flavor = SPF_UMAX; break;
208   case CmpInst::ICMP_ULT: Flavor = SPF_UMIN; break;
209   case CmpInst::ICMP_SGT: Flavor = SPF_SMAX; break;
210   case CmpInst::ICMP_SLT: Flavor = SPF_SMIN; break;
211   // Non-strict inequalities.
212   case CmpInst::ICMP_ULE: Flavor = SPF_UMIN; break;
213   case CmpInst::ICMP_UGE: Flavor = SPF_UMAX; break;
214   case CmpInst::ICMP_SLE: Flavor = SPF_SMIN; break;
215   case CmpInst::ICMP_SGE: Flavor = SPF_SMAX; break;
216   default: break;
217   }
218 
219   return true;
220 }
221 
222 static unsigned hashCallInst(CallInst *CI) {
223   // Don't CSE convergent calls in different basic blocks, because they
224   // implicitly depend on the set of threads that is currently executing.
225   if (CI->isConvergent()) {
226     return hash_combine(
227         CI->getOpcode(), CI->getParent(),
228         hash_combine_range(CI->value_op_begin(), CI->value_op_end()));
229   }
230   return hash_combine(
231       CI->getOpcode(),
232       hash_combine_range(CI->value_op_begin(), CI->value_op_end()));
233 }
234 
235 static unsigned getHashValueImpl(SimpleValue Val) {
236   Instruction *Inst = Val.Inst;
237   // Hash in all of the operands as pointers.
238   if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst)) {
239     Value *LHS = BinOp->getOperand(0);
240     Value *RHS = BinOp->getOperand(1);
241     if (BinOp->isCommutative() && BinOp->getOperand(0) > BinOp->getOperand(1))
242       std::swap(LHS, RHS);
243 
244     return hash_combine(BinOp->getOpcode(), LHS, RHS);
245   }
246 
247   if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) {
248     // Compares can be commuted by swapping the comparands and
249     // updating the predicate.  Choose the form that has the
250     // comparands in sorted order, or in the case of a tie, the
251     // one with the lower predicate.
252     Value *LHS = CI->getOperand(0);
253     Value *RHS = CI->getOperand(1);
254     CmpInst::Predicate Pred = CI->getPredicate();
255     CmpInst::Predicate SwappedPred = CI->getSwappedPredicate();
256     if (std::tie(LHS, Pred) > std::tie(RHS, SwappedPred)) {
257       std::swap(LHS, RHS);
258       Pred = SwappedPred;
259     }
260     return hash_combine(Inst->getOpcode(), Pred, LHS, RHS);
261   }
262 
263   // Hash general selects to allow matching commuted true/false operands.
264   SelectPatternFlavor SPF;
265   Value *Cond, *A, *B;
266   if (matchSelectWithOptionalNotCond(Inst, Cond, A, B, SPF)) {
267     // Hash min/max (cmp + select) to allow for commuted operands.
268     // Min/max may also have non-canonical compare predicate (eg, the compare for
269     // smin may use 'sgt' rather than 'slt'), and non-canonical operands in the
270     // compare.
271     // TODO: We should also detect FP min/max.
272     if (SPF == SPF_SMIN || SPF == SPF_SMAX ||
273         SPF == SPF_UMIN || SPF == SPF_UMAX) {
274       if (A > B)
275         std::swap(A, B);
276       return hash_combine(Inst->getOpcode(), SPF, A, B);
277     }
278 
279     // Hash general selects to allow matching commuted true/false operands.
280 
281     // If we do not have a compare as the condition, just hash in the condition.
282     CmpInst::Predicate Pred;
283     Value *X, *Y;
284     if (!match(Cond, m_Cmp(Pred, m_Value(X), m_Value(Y))))
285       return hash_combine(Inst->getOpcode(), Cond, A, B);
286 
287     // Similar to cmp normalization (above) - canonicalize the predicate value:
288     // select (icmp Pred, X, Y), A, B --> select (icmp InvPred, X, Y), B, A
289     if (CmpInst::getInversePredicate(Pred) < Pred) {
290       Pred = CmpInst::getInversePredicate(Pred);
291       std::swap(A, B);
292     }
293     return hash_combine(Inst->getOpcode(), Pred, X, Y, A, B);
294   }
295 
296   if (CastInst *CI = dyn_cast<CastInst>(Inst))
297     return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0));
298 
299   if (FreezeInst *FI = dyn_cast<FreezeInst>(Inst))
300     return hash_combine(FI->getOpcode(), FI->getOperand(0));
301 
302   if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst))
303     return hash_combine(EVI->getOpcode(), EVI->getOperand(0),
304                         hash_combine_range(EVI->idx_begin(), EVI->idx_end()));
305 
306   if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst))
307     return hash_combine(IVI->getOpcode(), IVI->getOperand(0),
308                         IVI->getOperand(1),
309                         hash_combine_range(IVI->idx_begin(), IVI->idx_end()));
310 
311   assert((isa<CallInst>(Inst) || isa<ExtractElementInst>(Inst) ||
312           isa<InsertElementInst>(Inst) || isa<ShuffleVectorInst>(Inst) ||
313           isa<UnaryOperator>(Inst) || isa<FreezeInst>(Inst)) &&
314          "Invalid/unknown instruction");
315 
316   // Handle intrinsics with commutative operands.
317   auto *II = dyn_cast<IntrinsicInst>(Inst);
318   if (II && II->isCommutative() && II->arg_size() >= 2) {
319     Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
320     if (LHS > RHS)
321       std::swap(LHS, RHS);
322     return hash_combine(
323         II->getOpcode(), LHS, RHS,
324         hash_combine_range(II->value_op_begin() + 2, II->value_op_end()));
325   }
326 
327   // gc.relocate is 'special' call: its second and third operands are
328   // not real values, but indices into statepoint's argument list.
329   // Get values they point to.
330   if (const GCRelocateInst *GCR = dyn_cast<GCRelocateInst>(Inst))
331     return hash_combine(GCR->getOpcode(), GCR->getOperand(0),
332                         GCR->getBasePtr(), GCR->getDerivedPtr());
333 
334   // Don't CSE convergent calls in different basic blocks, because they
335   // implicitly depend on the set of threads that is currently executing.
336   if (CallInst *CI = dyn_cast<CallInst>(Inst))
337     return hashCallInst(CI);
338 
339   // Mix in the opcode.
340   return hash_combine(
341       Inst->getOpcode(),
342       hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
343 }
344 
345 unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) {
346 #ifndef NDEBUG
347   // If -earlycse-debug-hash was specified, return a constant -- this
348   // will force all hashing to collide, so we'll exhaustively search
349   // the table for a match, and the assertion in isEqual will fire if
350   // there's a bug causing equal keys to hash differently.
351   if (EarlyCSEDebugHash)
352     return 0;
353 #endif
354   return getHashValueImpl(Val);
355 }
356 
357 static bool isEqualImpl(SimpleValue LHS, SimpleValue RHS) {
358   Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
359 
360   if (LHS.isSentinel() || RHS.isSentinel())
361     return LHSI == RHSI;
362 
363   if (LHSI->getOpcode() != RHSI->getOpcode())
364     return false;
365   if (LHSI->isIdenticalToWhenDefined(RHSI)) {
366     // Convergent calls implicitly depend on the set of threads that is
367     // currently executing, so conservatively return false if they are in
368     // different basic blocks.
369     if (CallInst *CI = dyn_cast<CallInst>(LHSI);
370         CI && CI->isConvergent() && LHSI->getParent() != RHSI->getParent())
371       return false;
372 
373     return true;
374   }
375 
376   // If we're not strictly identical, we still might be a commutable instruction
377   if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) {
378     if (!LHSBinOp->isCommutative())
379       return false;
380 
381     assert(isa<BinaryOperator>(RHSI) &&
382            "same opcode, but different instruction type?");
383     BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI);
384 
385     // Commuted equality
386     return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) &&
387            LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0);
388   }
389   if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) {
390     assert(isa<CmpInst>(RHSI) &&
391            "same opcode, but different instruction type?");
392     CmpInst *RHSCmp = cast<CmpInst>(RHSI);
393     // Commuted equality
394     return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) &&
395            LHSCmp->getOperand(1) == RHSCmp->getOperand(0) &&
396            LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate();
397   }
398 
399   auto *LII = dyn_cast<IntrinsicInst>(LHSI);
400   auto *RII = dyn_cast<IntrinsicInst>(RHSI);
401   if (LII && RII && LII->getIntrinsicID() == RII->getIntrinsicID() &&
402       LII->isCommutative() && LII->arg_size() >= 2) {
403     return LII->getArgOperand(0) == RII->getArgOperand(1) &&
404            LII->getArgOperand(1) == RII->getArgOperand(0) &&
405            std::equal(LII->arg_begin() + 2, LII->arg_end(),
406                       RII->arg_begin() + 2, RII->arg_end());
407   }
408 
409   // See comment above in `getHashValue()`.
410   if (const GCRelocateInst *GCR1 = dyn_cast<GCRelocateInst>(LHSI))
411     if (const GCRelocateInst *GCR2 = dyn_cast<GCRelocateInst>(RHSI))
412       return GCR1->getOperand(0) == GCR2->getOperand(0) &&
413              GCR1->getBasePtr() == GCR2->getBasePtr() &&
414              GCR1->getDerivedPtr() == GCR2->getDerivedPtr();
415 
416   // Min/max can occur with commuted operands, non-canonical predicates,
417   // and/or non-canonical operands.
418   // Selects can be non-trivially equivalent via inverted conditions and swaps.
419   SelectPatternFlavor LSPF, RSPF;
420   Value *CondL, *CondR, *LHSA, *RHSA, *LHSB, *RHSB;
421   if (matchSelectWithOptionalNotCond(LHSI, CondL, LHSA, LHSB, LSPF) &&
422       matchSelectWithOptionalNotCond(RHSI, CondR, RHSA, RHSB, RSPF)) {
423     if (LSPF == RSPF) {
424       // TODO: We should also detect FP min/max.
425       if (LSPF == SPF_SMIN || LSPF == SPF_SMAX ||
426           LSPF == SPF_UMIN || LSPF == SPF_UMAX)
427         return ((LHSA == RHSA && LHSB == RHSB) ||
428                 (LHSA == RHSB && LHSB == RHSA));
429 
430       // select Cond, A, B <--> select not(Cond), B, A
431       if (CondL == CondR && LHSA == RHSA && LHSB == RHSB)
432         return true;
433     }
434 
435     // If the true/false operands are swapped and the conditions are compares
436     // with inverted predicates, the selects are equal:
437     // select (icmp Pred, X, Y), A, B <--> select (icmp InvPred, X, Y), B, A
438     //
439     // This also handles patterns with a double-negation in the sense of not +
440     // inverse, because we looked through a 'not' in the matching function and
441     // swapped A/B:
442     // select (cmp Pred, X, Y), A, B <--> select (not (cmp InvPred, X, Y)), B, A
443     //
444     // This intentionally does NOT handle patterns with a double-negation in
445     // the sense of not + not, because doing so could result in values
446     // comparing
447     // as equal that hash differently in the min/max cases like:
448     // select (cmp slt, X, Y), X, Y <--> select (not (not (cmp slt, X, Y))), X, Y
449     //   ^ hashes as min                  ^ would not hash as min
450     // In the context of the EarlyCSE pass, however, such cases never reach
451     // this code, as we simplify the double-negation before hashing the second
452     // select (and so still succeed at CSEing them).
453     if (LHSA == RHSB && LHSB == RHSA) {
454       CmpInst::Predicate PredL, PredR;
455       Value *X, *Y;
456       if (match(CondL, m_Cmp(PredL, m_Value(X), m_Value(Y))) &&
457           match(CondR, m_Cmp(PredR, m_Specific(X), m_Specific(Y))) &&
458           CmpInst::getInversePredicate(PredL) == PredR)
459         return true;
460     }
461   }
462 
463   return false;
464 }
465 
466 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) {
467   // These comparisons are nontrivial, so assert that equality implies
468   // hash equality (DenseMap demands this as an invariant).
469   bool Result = isEqualImpl(LHS, RHS);
470   assert(!Result || (LHS.isSentinel() && LHS.Inst == RHS.Inst) ||
471          getHashValueImpl(LHS) == getHashValueImpl(RHS));
472   return Result;
473 }
474 
475 //===----------------------------------------------------------------------===//
476 // CallValue
477 //===----------------------------------------------------------------------===//
478 
479 namespace {
480 
481 /// Struct representing the available call values in the scoped hash
482 /// table.
483 struct CallValue {
484   Instruction *Inst;
485 
486   CallValue(Instruction *I) : Inst(I) {
487     assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
488   }
489 
490   bool isSentinel() const {
491     return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
492            Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
493   }
494 
495   static bool canHandle(Instruction *Inst) {
496     // Don't value number anything that returns void.
497     if (Inst->getType()->isVoidTy())
498       return false;
499 
500     CallInst *CI = dyn_cast<CallInst>(Inst);
501     if (!CI || !CI->onlyReadsMemory() ||
502         // FIXME: Currently the calls which may access the thread id may
503         // be considered as not accessing the memory. But this is
504         // problematic for coroutines, since coroutines may resume in a
505         // different thread. So we disable the optimization here for the
506         // correctness. However, it may block many other correct
507         // optimizations. Revert this one when we detect the memory
508         // accessing kind more precisely.
509         CI->getFunction()->isPresplitCoroutine())
510       return false;
511     return true;
512   }
513 };
514 
515 } // end anonymous namespace
516 
517 namespace llvm {
518 
519 template <> struct DenseMapInfo<CallValue> {
520   static inline CallValue getEmptyKey() {
521     return DenseMapInfo<Instruction *>::getEmptyKey();
522   }
523 
524   static inline CallValue getTombstoneKey() {
525     return DenseMapInfo<Instruction *>::getTombstoneKey();
526   }
527 
528   static unsigned getHashValue(CallValue Val);
529   static bool isEqual(CallValue LHS, CallValue RHS);
530 };
531 
532 } // end namespace llvm
533 
534 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) {
535   Instruction *Inst = Val.Inst;
536 
537   // Hash all of the operands as pointers and mix in the opcode.
538   return hashCallInst(cast<CallInst>(Inst));
539 }
540 
541 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) {
542   if (LHS.isSentinel() || RHS.isSentinel())
543     return LHS.Inst == RHS.Inst;
544 
545   CallInst *LHSI = cast<CallInst>(LHS.Inst);
546   CallInst *RHSI = cast<CallInst>(RHS.Inst);
547 
548   // Convergent calls implicitly depend on the set of threads that is
549   // currently executing, so conservatively return false if they are in
550   // different basic blocks.
551   if (LHSI->isConvergent() && LHSI->getParent() != RHSI->getParent())
552     return false;
553 
554   return LHSI->isIdenticalTo(RHSI);
555 }
556 
557 //===----------------------------------------------------------------------===//
558 // GEPValue
559 //===----------------------------------------------------------------------===//
560 
561 namespace {
562 
563 struct GEPValue {
564   Instruction *Inst;
565   std::optional<int64_t> ConstantOffset;
566 
567   GEPValue(Instruction *I) : Inst(I) {
568     assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
569   }
570 
571   GEPValue(Instruction *I, std::optional<int64_t> ConstantOffset)
572       : Inst(I), ConstantOffset(ConstantOffset) {
573     assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
574   }
575 
576   bool isSentinel() const {
577     return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
578            Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
579   }
580 
581   static bool canHandle(Instruction *Inst) {
582     return isa<GetElementPtrInst>(Inst);
583   }
584 };
585 
586 } // namespace
587 
588 namespace llvm {
589 
590 template <> struct DenseMapInfo<GEPValue> {
591   static inline GEPValue getEmptyKey() {
592     return DenseMapInfo<Instruction *>::getEmptyKey();
593   }
594 
595   static inline GEPValue getTombstoneKey() {
596     return DenseMapInfo<Instruction *>::getTombstoneKey();
597   }
598 
599   static unsigned getHashValue(const GEPValue &Val);
600   static bool isEqual(const GEPValue &LHS, const GEPValue &RHS);
601 };
602 
603 } // end namespace llvm
604 
605 unsigned DenseMapInfo<GEPValue>::getHashValue(const GEPValue &Val) {
606   auto *GEP = cast<GetElementPtrInst>(Val.Inst);
607   if (Val.ConstantOffset.has_value())
608     return hash_combine(GEP->getOpcode(), GEP->getPointerOperand(),
609                         Val.ConstantOffset.value());
610   return hash_combine(
611       GEP->getOpcode(),
612       hash_combine_range(GEP->value_op_begin(), GEP->value_op_end()));
613 }
614 
615 bool DenseMapInfo<GEPValue>::isEqual(const GEPValue &LHS, const GEPValue &RHS) {
616   if (LHS.isSentinel() || RHS.isSentinel())
617     return LHS.Inst == RHS.Inst;
618   auto *LGEP = cast<GetElementPtrInst>(LHS.Inst);
619   auto *RGEP = cast<GetElementPtrInst>(RHS.Inst);
620   if (LGEP->getPointerOperand() != RGEP->getPointerOperand())
621     return false;
622   if (LHS.ConstantOffset.has_value() && RHS.ConstantOffset.has_value())
623     return LHS.ConstantOffset.value() == RHS.ConstantOffset.value();
624   return LGEP->isIdenticalToWhenDefined(RGEP);
625 }
626 
627 //===----------------------------------------------------------------------===//
628 // EarlyCSE implementation
629 //===----------------------------------------------------------------------===//
630 
631 namespace {
632 
633 /// A simple and fast domtree-based CSE pass.
634 ///
635 /// This pass does a simple depth-first walk over the dominator tree,
636 /// eliminating trivially redundant instructions and using instsimplify to
637 /// canonicalize things as it goes. It is intended to be fast and catch obvious
638 /// cases so that instcombine and other passes are more effective. It is
639 /// expected that a later pass of GVN will catch the interesting/hard cases.
640 class EarlyCSE {
641 public:
642   const TargetLibraryInfo &TLI;
643   const TargetTransformInfo &TTI;
644   DominatorTree &DT;
645   AssumptionCache &AC;
646   const SimplifyQuery SQ;
647   MemorySSA *MSSA;
648   std::unique_ptr<MemorySSAUpdater> MSSAUpdater;
649 
650   using AllocatorTy =
651       RecyclingAllocator<BumpPtrAllocator,
652                          ScopedHashTableVal<SimpleValue, Value *>>;
653   using ScopedHTType =
654       ScopedHashTable<SimpleValue, Value *, DenseMapInfo<SimpleValue>,
655                       AllocatorTy>;
656 
657   /// A scoped hash table of the current values of all of our simple
658   /// scalar expressions.
659   ///
660   /// As we walk down the domtree, we look to see if instructions are in this:
661   /// if so, we replace them with what we find, otherwise we insert them so
662   /// that dominated values can succeed in their lookup.
663   ScopedHTType AvailableValues;
664 
665   /// A scoped hash table of the current values of previously encountered
666   /// memory locations.
667   ///
668   /// This allows us to get efficient access to dominating loads or stores when
669   /// we have a fully redundant load.  In addition to the most recent load, we
670   /// keep track of a generation count of the read, which is compared against
671   /// the current generation count.  The current generation count is incremented
672   /// after every possibly writing memory operation, which ensures that we only
673   /// CSE loads with other loads that have no intervening store.  Ordering
674   /// events (such as fences or atomic instructions) increment the generation
675   /// count as well; essentially, we model these as writes to all possible
676   /// locations.  Note that atomic and/or volatile loads and stores can be
677   /// present the table; it is the responsibility of the consumer to inspect
678   /// the atomicity/volatility if needed.
679   struct LoadValue {
680     Instruction *DefInst = nullptr;
681     unsigned Generation = 0;
682     int MatchingId = -1;
683     bool IsAtomic = false;
684     bool IsLoad = false;
685 
686     LoadValue() = default;
687     LoadValue(Instruction *Inst, unsigned Generation, unsigned MatchingId,
688               bool IsAtomic, bool IsLoad)
689         : DefInst(Inst), Generation(Generation), MatchingId(MatchingId),
690           IsAtomic(IsAtomic), IsLoad(IsLoad) {}
691   };
692 
693   using LoadMapAllocator =
694       RecyclingAllocator<BumpPtrAllocator,
695                          ScopedHashTableVal<Value *, LoadValue>>;
696   using LoadHTType =
697       ScopedHashTable<Value *, LoadValue, DenseMapInfo<Value *>,
698                       LoadMapAllocator>;
699 
700   LoadHTType AvailableLoads;
701 
702   // A scoped hash table mapping memory locations (represented as typed
703   // addresses) to generation numbers at which that memory location became
704   // (henceforth indefinitely) invariant.
705   using InvariantMapAllocator =
706       RecyclingAllocator<BumpPtrAllocator,
707                          ScopedHashTableVal<MemoryLocation, unsigned>>;
708   using InvariantHTType =
709       ScopedHashTable<MemoryLocation, unsigned, DenseMapInfo<MemoryLocation>,
710                       InvariantMapAllocator>;
711   InvariantHTType AvailableInvariants;
712 
713   /// A scoped hash table of the current values of read-only call
714   /// values.
715   ///
716   /// It uses the same generation count as loads.
717   using CallHTType =
718       ScopedHashTable<CallValue, std::pair<Instruction *, unsigned>>;
719   CallHTType AvailableCalls;
720 
721   using GEPMapAllocatorTy =
722       RecyclingAllocator<BumpPtrAllocator,
723                          ScopedHashTableVal<GEPValue, Value *>>;
724   using GEPHTType = ScopedHashTable<GEPValue, Value *, DenseMapInfo<GEPValue>,
725                                     GEPMapAllocatorTy>;
726   GEPHTType AvailableGEPs;
727 
728   /// This is the current generation of the memory value.
729   unsigned CurrentGeneration = 0;
730 
731   /// Set up the EarlyCSE runner for a particular function.
732   EarlyCSE(const DataLayout &DL, const TargetLibraryInfo &TLI,
733            const TargetTransformInfo &TTI, DominatorTree &DT,
734            AssumptionCache &AC, MemorySSA *MSSA)
735       : TLI(TLI), TTI(TTI), DT(DT), AC(AC), SQ(DL, &TLI, &DT, &AC), MSSA(MSSA),
736         MSSAUpdater(std::make_unique<MemorySSAUpdater>(MSSA)) {}
737 
738   bool run();
739 
740 private:
741   unsigned ClobberCounter = 0;
742   // Almost a POD, but needs to call the constructors for the scoped hash
743   // tables so that a new scope gets pushed on. These are RAII so that the
744   // scope gets popped when the NodeScope is destroyed.
745   class NodeScope {
746   public:
747     NodeScope(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
748               InvariantHTType &AvailableInvariants, CallHTType &AvailableCalls,
749               GEPHTType &AvailableGEPs)
750         : Scope(AvailableValues), LoadScope(AvailableLoads),
751           InvariantScope(AvailableInvariants), CallScope(AvailableCalls),
752           GEPScope(AvailableGEPs) {}
753     NodeScope(const NodeScope &) = delete;
754     NodeScope &operator=(const NodeScope &) = delete;
755 
756   private:
757     ScopedHTType::ScopeTy Scope;
758     LoadHTType::ScopeTy LoadScope;
759     InvariantHTType::ScopeTy InvariantScope;
760     CallHTType::ScopeTy CallScope;
761     GEPHTType::ScopeTy GEPScope;
762   };
763 
764   // Contains all the needed information to create a stack for doing a depth
765   // first traversal of the tree. This includes scopes for values, loads, and
766   // calls as well as the generation. There is a child iterator so that the
767   // children do not need to be store separately.
768   class StackNode {
769   public:
770     StackNode(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
771               InvariantHTType &AvailableInvariants, CallHTType &AvailableCalls,
772               GEPHTType &AvailableGEPs, unsigned cg, DomTreeNode *n,
773               DomTreeNode::const_iterator child,
774               DomTreeNode::const_iterator end)
775         : CurrentGeneration(cg), ChildGeneration(cg), Node(n), ChildIter(child),
776           EndIter(end),
777           Scopes(AvailableValues, AvailableLoads, AvailableInvariants,
778                  AvailableCalls, AvailableGEPs) {}
779     StackNode(const StackNode &) = delete;
780     StackNode &operator=(const StackNode &) = delete;
781 
782     // Accessors.
783     unsigned currentGeneration() const { return CurrentGeneration; }
784     unsigned childGeneration() const { return ChildGeneration; }
785     void childGeneration(unsigned generation) { ChildGeneration = generation; }
786     DomTreeNode *node() { return Node; }
787     DomTreeNode::const_iterator childIter() const { return ChildIter; }
788 
789     DomTreeNode *nextChild() {
790       DomTreeNode *child = *ChildIter;
791       ++ChildIter;
792       return child;
793     }
794 
795     DomTreeNode::const_iterator end() const { return EndIter; }
796     bool isProcessed() const { return Processed; }
797     void process() { Processed = true; }
798 
799   private:
800     unsigned CurrentGeneration;
801     unsigned ChildGeneration;
802     DomTreeNode *Node;
803     DomTreeNode::const_iterator ChildIter;
804     DomTreeNode::const_iterator EndIter;
805     NodeScope Scopes;
806     bool Processed = false;
807   };
808 
809   /// Wrapper class to handle memory instructions, including loads,
810   /// stores and intrinsic loads and stores defined by the target.
811   class ParseMemoryInst {
812   public:
813     ParseMemoryInst(Instruction *Inst, const TargetTransformInfo &TTI)
814       : Inst(Inst) {
815       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
816         IntrID = II->getIntrinsicID();
817         if (TTI.getTgtMemIntrinsic(II, Info))
818           return;
819         if (isHandledNonTargetIntrinsic(IntrID)) {
820           switch (IntrID) {
821           case Intrinsic::masked_load:
822             Info.PtrVal = Inst->getOperand(0);
823             Info.MatchingId = Intrinsic::masked_load;
824             Info.ReadMem = true;
825             Info.WriteMem = false;
826             Info.IsVolatile = false;
827             break;
828           case Intrinsic::masked_store:
829             Info.PtrVal = Inst->getOperand(1);
830             // Use the ID of masked load as the "matching id". This will
831             // prevent matching non-masked loads/stores with masked ones
832             // (which could be done), but at the moment, the code here
833             // does not support matching intrinsics with non-intrinsics,
834             // so keep the MatchingIds specific to masked instructions
835             // for now (TODO).
836             Info.MatchingId = Intrinsic::masked_load;
837             Info.ReadMem = false;
838             Info.WriteMem = true;
839             Info.IsVolatile = false;
840             break;
841           }
842         }
843       }
844     }
845 
846     Instruction *get() { return Inst; }
847     const Instruction *get() const { return Inst; }
848 
849     bool isLoad() const {
850       if (IntrID != 0)
851         return Info.ReadMem;
852       return isa<LoadInst>(Inst);
853     }
854 
855     bool isStore() const {
856       if (IntrID != 0)
857         return Info.WriteMem;
858       return isa<StoreInst>(Inst);
859     }
860 
861     bool isAtomic() const {
862       if (IntrID != 0)
863         return Info.Ordering != AtomicOrdering::NotAtomic;
864       return Inst->isAtomic();
865     }
866 
867     bool isUnordered() const {
868       if (IntrID != 0)
869         return Info.isUnordered();
870 
871       if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
872         return LI->isUnordered();
873       } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
874         return SI->isUnordered();
875       }
876       // Conservative answer
877       return !Inst->isAtomic();
878     }
879 
880     bool isVolatile() const {
881       if (IntrID != 0)
882         return Info.IsVolatile;
883 
884       if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
885         return LI->isVolatile();
886       } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
887         return SI->isVolatile();
888       }
889       // Conservative answer
890       return true;
891     }
892 
893     bool isInvariantLoad() const {
894       if (auto *LI = dyn_cast<LoadInst>(Inst))
895         return LI->hasMetadata(LLVMContext::MD_invariant_load);
896       return false;
897     }
898 
899     bool isValid() const { return getPointerOperand() != nullptr; }
900 
901     // For regular (non-intrinsic) loads/stores, this is set to -1. For
902     // intrinsic loads/stores, the id is retrieved from the corresponding
903     // field in the MemIntrinsicInfo structure.  That field contains
904     // non-negative values only.
905     int getMatchingId() const {
906       if (IntrID != 0)
907         return Info.MatchingId;
908       return -1;
909     }
910 
911     Value *getPointerOperand() const {
912       if (IntrID != 0)
913         return Info.PtrVal;
914       return getLoadStorePointerOperand(Inst);
915     }
916 
917     Type *getValueType() const {
918       // TODO: handle target-specific intrinsics.
919       return Inst->getAccessType();
920     }
921 
922     bool mayReadFromMemory() const {
923       if (IntrID != 0)
924         return Info.ReadMem;
925       return Inst->mayReadFromMemory();
926     }
927 
928     bool mayWriteToMemory() const {
929       if (IntrID != 0)
930         return Info.WriteMem;
931       return Inst->mayWriteToMemory();
932     }
933 
934   private:
935     Intrinsic::ID IntrID = 0;
936     MemIntrinsicInfo Info;
937     Instruction *Inst;
938   };
939 
940   // This function is to prevent accidentally passing a non-target
941   // intrinsic ID to TargetTransformInfo.
942   static bool isHandledNonTargetIntrinsic(Intrinsic::ID ID) {
943     switch (ID) {
944     case Intrinsic::masked_load:
945     case Intrinsic::masked_store:
946       return true;
947     }
948     return false;
949   }
950   static bool isHandledNonTargetIntrinsic(const Value *V) {
951     if (auto *II = dyn_cast<IntrinsicInst>(V))
952       return isHandledNonTargetIntrinsic(II->getIntrinsicID());
953     return false;
954   }
955 
956   bool processNode(DomTreeNode *Node);
957 
958   bool handleBranchCondition(Instruction *CondInst, const BranchInst *BI,
959                              const BasicBlock *BB, const BasicBlock *Pred);
960 
961   Value *getMatchingValue(LoadValue &InVal, ParseMemoryInst &MemInst,
962                           unsigned CurrentGeneration);
963 
964   bool overridingStores(const ParseMemoryInst &Earlier,
965                         const ParseMemoryInst &Later);
966 
967   Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const {
968     // TODO: We could insert relevant casts on type mismatch here.
969     if (auto *LI = dyn_cast<LoadInst>(Inst))
970       return LI->getType() == ExpectedType ? LI : nullptr;
971     if (auto *SI = dyn_cast<StoreInst>(Inst)) {
972       Value *V = SI->getValueOperand();
973       return V->getType() == ExpectedType ? V : nullptr;
974     }
975     assert(isa<IntrinsicInst>(Inst) && "Instruction not supported");
976     auto *II = cast<IntrinsicInst>(Inst);
977     if (isHandledNonTargetIntrinsic(II->getIntrinsicID()))
978       return getOrCreateResultNonTargetMemIntrinsic(II, ExpectedType);
979     return TTI.getOrCreateResultFromMemIntrinsic(II, ExpectedType);
980   }
981 
982   Value *getOrCreateResultNonTargetMemIntrinsic(IntrinsicInst *II,
983                                                 Type *ExpectedType) const {
984     // TODO: We could insert relevant casts on type mismatch here.
985     switch (II->getIntrinsicID()) {
986     case Intrinsic::masked_load:
987       return II->getType() == ExpectedType ? II : nullptr;
988     case Intrinsic::masked_store: {
989       Value *V = II->getOperand(0);
990       return V->getType() == ExpectedType ? V : nullptr;
991     }
992     }
993     return nullptr;
994   }
995 
996   /// Return true if the instruction is known to only operate on memory
997   /// provably invariant in the given "generation".
998   bool isOperatingOnInvariantMemAt(Instruction *I, unsigned GenAt);
999 
1000   bool isSameMemGeneration(unsigned EarlierGeneration, unsigned LaterGeneration,
1001                            Instruction *EarlierInst, Instruction *LaterInst);
1002 
1003   bool isNonTargetIntrinsicMatch(const IntrinsicInst *Earlier,
1004                                  const IntrinsicInst *Later) {
1005     auto IsSubmask = [](const Value *Mask0, const Value *Mask1) {
1006       // Is Mask0 a submask of Mask1?
1007       if (Mask0 == Mask1)
1008         return true;
1009       if (isa<UndefValue>(Mask0) || isa<UndefValue>(Mask1))
1010         return false;
1011       auto *Vec0 = dyn_cast<ConstantVector>(Mask0);
1012       auto *Vec1 = dyn_cast<ConstantVector>(Mask1);
1013       if (!Vec0 || !Vec1)
1014         return false;
1015       if (Vec0->getType() != Vec1->getType())
1016         return false;
1017       for (int i = 0, e = Vec0->getNumOperands(); i != e; ++i) {
1018         Constant *Elem0 = Vec0->getOperand(i);
1019         Constant *Elem1 = Vec1->getOperand(i);
1020         auto *Int0 = dyn_cast<ConstantInt>(Elem0);
1021         if (Int0 && Int0->isZero())
1022           continue;
1023         auto *Int1 = dyn_cast<ConstantInt>(Elem1);
1024         if (Int1 && !Int1->isZero())
1025           continue;
1026         if (isa<UndefValue>(Elem0) || isa<UndefValue>(Elem1))
1027           return false;
1028         if (Elem0 == Elem1)
1029           continue;
1030         return false;
1031       }
1032       return true;
1033     };
1034     auto PtrOp = [](const IntrinsicInst *II) {
1035       if (II->getIntrinsicID() == Intrinsic::masked_load)
1036         return II->getOperand(0);
1037       if (II->getIntrinsicID() == Intrinsic::masked_store)
1038         return II->getOperand(1);
1039       llvm_unreachable("Unexpected IntrinsicInst");
1040     };
1041     auto MaskOp = [](const IntrinsicInst *II) {
1042       if (II->getIntrinsicID() == Intrinsic::masked_load)
1043         return II->getOperand(2);
1044       if (II->getIntrinsicID() == Intrinsic::masked_store)
1045         return II->getOperand(3);
1046       llvm_unreachable("Unexpected IntrinsicInst");
1047     };
1048     auto ThruOp = [](const IntrinsicInst *II) {
1049       if (II->getIntrinsicID() == Intrinsic::masked_load)
1050         return II->getOperand(3);
1051       llvm_unreachable("Unexpected IntrinsicInst");
1052     };
1053 
1054     if (PtrOp(Earlier) != PtrOp(Later))
1055       return false;
1056 
1057     Intrinsic::ID IDE = Earlier->getIntrinsicID();
1058     Intrinsic::ID IDL = Later->getIntrinsicID();
1059     // We could really use specific intrinsic classes for masked loads
1060     // and stores in IntrinsicInst.h.
1061     if (IDE == Intrinsic::masked_load && IDL == Intrinsic::masked_load) {
1062       // Trying to replace later masked load with the earlier one.
1063       // Check that the pointers are the same, and
1064       // - masks and pass-throughs are the same, or
1065       // - replacee's pass-through is "undef" and replacer's mask is a
1066       //   super-set of the replacee's mask.
1067       if (MaskOp(Earlier) == MaskOp(Later) && ThruOp(Earlier) == ThruOp(Later))
1068         return true;
1069       if (!isa<UndefValue>(ThruOp(Later)))
1070         return false;
1071       return IsSubmask(MaskOp(Later), MaskOp(Earlier));
1072     }
1073     if (IDE == Intrinsic::masked_store && IDL == Intrinsic::masked_load) {
1074       // Trying to replace a load of a stored value with the store's value.
1075       // Check that the pointers are the same, and
1076       // - load's mask is a subset of store's mask, and
1077       // - load's pass-through is "undef".
1078       if (!IsSubmask(MaskOp(Later), MaskOp(Earlier)))
1079         return false;
1080       return isa<UndefValue>(ThruOp(Later));
1081     }
1082     if (IDE == Intrinsic::masked_load && IDL == Intrinsic::masked_store) {
1083       // Trying to remove a store of the loaded value.
1084       // Check that the pointers are the same, and
1085       // - store's mask is a subset of the load's mask.
1086       return IsSubmask(MaskOp(Later), MaskOp(Earlier));
1087     }
1088     if (IDE == Intrinsic::masked_store && IDL == Intrinsic::masked_store) {
1089       // Trying to remove a dead store (earlier).
1090       // Check that the pointers are the same,
1091       // - the to-be-removed store's mask is a subset of the other store's
1092       //   mask.
1093       return IsSubmask(MaskOp(Earlier), MaskOp(Later));
1094     }
1095     return false;
1096   }
1097 
1098   void removeMSSA(Instruction &Inst) {
1099     if (!MSSA)
1100       return;
1101     if (VerifyMemorySSA)
1102       MSSA->verifyMemorySSA();
1103     // Removing a store here can leave MemorySSA in an unoptimized state by
1104     // creating MemoryPhis that have identical arguments and by creating
1105     // MemoryUses whose defining access is not an actual clobber. The phi case
1106     // is handled by MemorySSA when passing OptimizePhis = true to
1107     // removeMemoryAccess.  The non-optimized MemoryUse case is lazily updated
1108     // by MemorySSA's getClobberingMemoryAccess.
1109     MSSAUpdater->removeMemoryAccess(&Inst, true);
1110   }
1111 };
1112 
1113 } // end anonymous namespace
1114 
1115 /// Determine if the memory referenced by LaterInst is from the same heap
1116 /// version as EarlierInst.
1117 /// This is currently called in two scenarios:
1118 ///
1119 ///   load p
1120 ///   ...
1121 ///   load p
1122 ///
1123 /// and
1124 ///
1125 ///   x = load p
1126 ///   ...
1127 ///   store x, p
1128 ///
1129 /// in both cases we want to verify that there are no possible writes to the
1130 /// memory referenced by p between the earlier and later instruction.
1131 bool EarlyCSE::isSameMemGeneration(unsigned EarlierGeneration,
1132                                    unsigned LaterGeneration,
1133                                    Instruction *EarlierInst,
1134                                    Instruction *LaterInst) {
1135   // Check the simple memory generation tracking first.
1136   if (EarlierGeneration == LaterGeneration)
1137     return true;
1138 
1139   if (!MSSA)
1140     return false;
1141 
1142   // If MemorySSA has determined that one of EarlierInst or LaterInst does not
1143   // read/write memory, then we can safely return true here.
1144   // FIXME: We could be more aggressive when checking doesNotAccessMemory(),
1145   // onlyReadsMemory(), mayReadFromMemory(), and mayWriteToMemory() in this pass
1146   // by also checking the MemorySSA MemoryAccess on the instruction.  Initial
1147   // experiments suggest this isn't worthwhile, at least for C/C++ code compiled
1148   // with the default optimization pipeline.
1149   auto *EarlierMA = MSSA->getMemoryAccess(EarlierInst);
1150   if (!EarlierMA)
1151     return true;
1152   auto *LaterMA = MSSA->getMemoryAccess(LaterInst);
1153   if (!LaterMA)
1154     return true;
1155 
1156   // Since we know LaterDef dominates LaterInst and EarlierInst dominates
1157   // LaterInst, if LaterDef dominates EarlierInst then it can't occur between
1158   // EarlierInst and LaterInst and neither can any other write that potentially
1159   // clobbers LaterInst.
1160   MemoryAccess *LaterDef;
1161   if (ClobberCounter < EarlyCSEMssaOptCap) {
1162     LaterDef = MSSA->getWalker()->getClobberingMemoryAccess(LaterInst);
1163     ClobberCounter++;
1164   } else
1165     LaterDef = LaterMA->getDefiningAccess();
1166 
1167   return MSSA->dominates(LaterDef, EarlierMA);
1168 }
1169 
1170 bool EarlyCSE::isOperatingOnInvariantMemAt(Instruction *I, unsigned GenAt) {
1171   // A location loaded from with an invariant_load is assumed to *never* change
1172   // within the visible scope of the compilation.
1173   if (auto *LI = dyn_cast<LoadInst>(I))
1174     if (LI->hasMetadata(LLVMContext::MD_invariant_load))
1175       return true;
1176 
1177   auto MemLocOpt = MemoryLocation::getOrNone(I);
1178   if (!MemLocOpt)
1179     // "target" intrinsic forms of loads aren't currently known to
1180     // MemoryLocation::get.  TODO
1181     return false;
1182   MemoryLocation MemLoc = *MemLocOpt;
1183   if (!AvailableInvariants.count(MemLoc))
1184     return false;
1185 
1186   // Is the generation at which this became invariant older than the
1187   // current one?
1188   return AvailableInvariants.lookup(MemLoc) <= GenAt;
1189 }
1190 
1191 bool EarlyCSE::handleBranchCondition(Instruction *CondInst,
1192                                      const BranchInst *BI, const BasicBlock *BB,
1193                                      const BasicBlock *Pred) {
1194   assert(BI->isConditional() && "Should be a conditional branch!");
1195   assert(BI->getCondition() == CondInst && "Wrong condition?");
1196   assert(BI->getSuccessor(0) == BB || BI->getSuccessor(1) == BB);
1197   auto *TorF = (BI->getSuccessor(0) == BB)
1198                    ? ConstantInt::getTrue(BB->getContext())
1199                    : ConstantInt::getFalse(BB->getContext());
1200   auto MatchBinOp = [](Instruction *I, unsigned Opcode, Value *&LHS,
1201                        Value *&RHS) {
1202     if (Opcode == Instruction::And &&
1203         match(I, m_LogicalAnd(m_Value(LHS), m_Value(RHS))))
1204       return true;
1205     else if (Opcode == Instruction::Or &&
1206              match(I, m_LogicalOr(m_Value(LHS), m_Value(RHS))))
1207       return true;
1208     return false;
1209   };
1210   // If the condition is AND operation, we can propagate its operands into the
1211   // true branch. If it is OR operation, we can propagate them into the false
1212   // branch.
1213   unsigned PropagateOpcode =
1214       (BI->getSuccessor(0) == BB) ? Instruction::And : Instruction::Or;
1215 
1216   bool MadeChanges = false;
1217   SmallVector<Instruction *, 4> WorkList;
1218   SmallPtrSet<Instruction *, 4> Visited;
1219   WorkList.push_back(CondInst);
1220   while (!WorkList.empty()) {
1221     Instruction *Curr = WorkList.pop_back_val();
1222 
1223     AvailableValues.insert(Curr, TorF);
1224     LLVM_DEBUG(dbgs() << "EarlyCSE CVP: Add conditional value for '"
1225                       << Curr->getName() << "' as " << *TorF << " in "
1226                       << BB->getName() << "\n");
1227     if (!DebugCounter::shouldExecute(CSECounter)) {
1228       LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1229     } else {
1230       // Replace all dominated uses with the known value.
1231       if (unsigned Count = replaceDominatedUsesWith(Curr, TorF, DT,
1232                                                     BasicBlockEdge(Pred, BB))) {
1233         NumCSECVP += Count;
1234         MadeChanges = true;
1235       }
1236     }
1237 
1238     Value *LHS, *RHS;
1239     if (MatchBinOp(Curr, PropagateOpcode, LHS, RHS))
1240       for (auto *Op : { LHS, RHS })
1241         if (Instruction *OPI = dyn_cast<Instruction>(Op))
1242           if (SimpleValue::canHandle(OPI) && Visited.insert(OPI).second)
1243             WorkList.push_back(OPI);
1244   }
1245 
1246   return MadeChanges;
1247 }
1248 
1249 Value *EarlyCSE::getMatchingValue(LoadValue &InVal, ParseMemoryInst &MemInst,
1250                                   unsigned CurrentGeneration) {
1251   if (InVal.DefInst == nullptr)
1252     return nullptr;
1253   if (InVal.MatchingId != MemInst.getMatchingId())
1254     return nullptr;
1255   // We don't yet handle removing loads with ordering of any kind.
1256   if (MemInst.isVolatile() || !MemInst.isUnordered())
1257     return nullptr;
1258   // We can't replace an atomic load with one which isn't also atomic.
1259   if (MemInst.isLoad() && !InVal.IsAtomic && MemInst.isAtomic())
1260     return nullptr;
1261   // The value V returned from this function is used differently depending
1262   // on whether MemInst is a load or a store. If it's a load, we will replace
1263   // MemInst with V, if it's a store, we will check if V is the same as the
1264   // available value.
1265   bool MemInstMatching = !MemInst.isLoad();
1266   Instruction *Matching = MemInstMatching ? MemInst.get() : InVal.DefInst;
1267   Instruction *Other = MemInstMatching ? InVal.DefInst : MemInst.get();
1268 
1269   // For stores check the result values before checking memory generation
1270   // (otherwise isSameMemGeneration may crash).
1271   Value *Result = MemInst.isStore()
1272                       ? getOrCreateResult(Matching, Other->getType())
1273                       : nullptr;
1274   if (MemInst.isStore() && InVal.DefInst != Result)
1275     return nullptr;
1276 
1277   // Deal with non-target memory intrinsics.
1278   bool MatchingNTI = isHandledNonTargetIntrinsic(Matching);
1279   bool OtherNTI = isHandledNonTargetIntrinsic(Other);
1280   if (OtherNTI != MatchingNTI)
1281     return nullptr;
1282   if (OtherNTI && MatchingNTI) {
1283     if (!isNonTargetIntrinsicMatch(cast<IntrinsicInst>(InVal.DefInst),
1284                                    cast<IntrinsicInst>(MemInst.get())))
1285       return nullptr;
1286   }
1287 
1288   if (!isOperatingOnInvariantMemAt(MemInst.get(), InVal.Generation) &&
1289       !isSameMemGeneration(InVal.Generation, CurrentGeneration, InVal.DefInst,
1290                            MemInst.get()))
1291     return nullptr;
1292 
1293   if (!Result)
1294     Result = getOrCreateResult(Matching, Other->getType());
1295   return Result;
1296 }
1297 
1298 static void combineIRFlags(Instruction &From, Value *To) {
1299   if (auto *I = dyn_cast<Instruction>(To)) {
1300     // If I being poison triggers UB, there is no need to drop those
1301     // flags. Otherwise, only retain flags present on both I and Inst.
1302     // TODO: Currently some fast-math flags are not treated as
1303     // poison-generating even though they should. Until this is fixed,
1304     // always retain flags present on both I and Inst for floating point
1305     // instructions.
1306     if (isa<FPMathOperator>(I) ||
1307         (I->hasPoisonGeneratingFlags() && !programUndefinedIfPoison(I)))
1308       I->andIRFlags(&From);
1309   }
1310 }
1311 
1312 bool EarlyCSE::overridingStores(const ParseMemoryInst &Earlier,
1313                                 const ParseMemoryInst &Later) {
1314   // Can we remove Earlier store because of Later store?
1315 
1316   assert(Earlier.isUnordered() && !Earlier.isVolatile() &&
1317          "Violated invariant");
1318   if (Earlier.getPointerOperand() != Later.getPointerOperand())
1319     return false;
1320   if (!Earlier.getValueType() || !Later.getValueType() ||
1321       Earlier.getValueType() != Later.getValueType())
1322     return false;
1323   if (Earlier.getMatchingId() != Later.getMatchingId())
1324     return false;
1325   // At the moment, we don't remove ordered stores, but do remove
1326   // unordered atomic stores.  There's no special requirement (for
1327   // unordered atomics) about removing atomic stores only in favor of
1328   // other atomic stores since we were going to execute the non-atomic
1329   // one anyway and the atomic one might never have become visible.
1330   if (!Earlier.isUnordered() || !Later.isUnordered())
1331     return false;
1332 
1333   // Deal with non-target memory intrinsics.
1334   bool ENTI = isHandledNonTargetIntrinsic(Earlier.get());
1335   bool LNTI = isHandledNonTargetIntrinsic(Later.get());
1336   if (ENTI && LNTI)
1337     return isNonTargetIntrinsicMatch(cast<IntrinsicInst>(Earlier.get()),
1338                                      cast<IntrinsicInst>(Later.get()));
1339 
1340   // Because of the check above, at least one of them is false.
1341   // For now disallow matching intrinsics with non-intrinsics,
1342   // so assume that the stores match if neither is an intrinsic.
1343   return ENTI == LNTI;
1344 }
1345 
1346 bool EarlyCSE::processNode(DomTreeNode *Node) {
1347   bool Changed = false;
1348   BasicBlock *BB = Node->getBlock();
1349 
1350   // If this block has a single predecessor, then the predecessor is the parent
1351   // of the domtree node and all of the live out memory values are still current
1352   // in this block.  If this block has multiple predecessors, then they could
1353   // have invalidated the live-out memory values of our parent value.  For now,
1354   // just be conservative and invalidate memory if this block has multiple
1355   // predecessors.
1356   if (!BB->getSinglePredecessor())
1357     ++CurrentGeneration;
1358 
1359   // If this node has a single predecessor which ends in a conditional branch,
1360   // we can infer the value of the branch condition given that we took this
1361   // path.  We need the single predecessor to ensure there's not another path
1362   // which reaches this block where the condition might hold a different
1363   // value.  Since we're adding this to the scoped hash table (like any other
1364   // def), it will have been popped if we encounter a future merge block.
1365   if (BasicBlock *Pred = BB->getSinglePredecessor()) {
1366     auto *BI = dyn_cast<BranchInst>(Pred->getTerminator());
1367     if (BI && BI->isConditional()) {
1368       auto *CondInst = dyn_cast<Instruction>(BI->getCondition());
1369       if (CondInst && SimpleValue::canHandle(CondInst))
1370         Changed |= handleBranchCondition(CondInst, BI, BB, Pred);
1371     }
1372   }
1373 
1374   /// LastStore - Keep track of the last non-volatile store that we saw... for
1375   /// as long as there in no instruction that reads memory.  If we see a store
1376   /// to the same location, we delete the dead store.  This zaps trivial dead
1377   /// stores which can occur in bitfield code among other things.
1378   Instruction *LastStore = nullptr;
1379 
1380   // See if any instructions in the block can be eliminated.  If so, do it.  If
1381   // not, add them to AvailableValues.
1382   for (Instruction &Inst : make_early_inc_range(*BB)) {
1383     // Dead instructions should just be removed.
1384     if (isInstructionTriviallyDead(&Inst, &TLI)) {
1385       LLVM_DEBUG(dbgs() << "EarlyCSE DCE: " << Inst << '\n');
1386       if (!DebugCounter::shouldExecute(CSECounter)) {
1387         LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1388         continue;
1389       }
1390 
1391       salvageKnowledge(&Inst, &AC);
1392       salvageDebugInfo(Inst);
1393       removeMSSA(Inst);
1394       Inst.eraseFromParent();
1395       Changed = true;
1396       ++NumSimplify;
1397       continue;
1398     }
1399 
1400     // Skip assume intrinsics, they don't really have side effects (although
1401     // they're marked as such to ensure preservation of control dependencies),
1402     // and this pass will not bother with its removal. However, we should mark
1403     // its condition as true for all dominated blocks.
1404     if (auto *Assume = dyn_cast<AssumeInst>(&Inst)) {
1405       auto *CondI = dyn_cast<Instruction>(Assume->getArgOperand(0));
1406       if (CondI && SimpleValue::canHandle(CondI)) {
1407         LLVM_DEBUG(dbgs() << "EarlyCSE considering assumption: " << Inst
1408                           << '\n');
1409         AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext()));
1410       } else
1411         LLVM_DEBUG(dbgs() << "EarlyCSE skipping assumption: " << Inst << '\n');
1412       continue;
1413     }
1414 
1415     // Likewise, noalias intrinsics don't actually write.
1416     if (match(&Inst,
1417               m_Intrinsic<Intrinsic::experimental_noalias_scope_decl>())) {
1418       LLVM_DEBUG(dbgs() << "EarlyCSE skipping noalias intrinsic: " << Inst
1419                         << '\n');
1420       continue;
1421     }
1422 
1423     // Skip sideeffect intrinsics, for the same reason as assume intrinsics.
1424     if (match(&Inst, m_Intrinsic<Intrinsic::sideeffect>())) {
1425       LLVM_DEBUG(dbgs() << "EarlyCSE skipping sideeffect: " << Inst << '\n');
1426       continue;
1427     }
1428 
1429     // Skip pseudoprobe intrinsics, for the same reason as assume intrinsics.
1430     if (match(&Inst, m_Intrinsic<Intrinsic::pseudoprobe>())) {
1431       LLVM_DEBUG(dbgs() << "EarlyCSE skipping pseudoprobe: " << Inst << '\n');
1432       continue;
1433     }
1434 
1435     // We can skip all invariant.start intrinsics since they only read memory,
1436     // and we can forward values across it. For invariant starts without
1437     // invariant ends, we can use the fact that the invariantness never ends to
1438     // start a scope in the current generaton which is true for all future
1439     // generations.  Also, we dont need to consume the last store since the
1440     // semantics of invariant.start allow us to perform   DSE of the last
1441     // store, if there was a store following invariant.start. Consider:
1442     //
1443     // store 30, i8* p
1444     // invariant.start(p)
1445     // store 40, i8* p
1446     // We can DSE the store to 30, since the store 40 to invariant location p
1447     // causes undefined behaviour.
1448     if (match(&Inst, m_Intrinsic<Intrinsic::invariant_start>())) {
1449       // If there are any uses, the scope might end.
1450       if (!Inst.use_empty())
1451         continue;
1452       MemoryLocation MemLoc =
1453           MemoryLocation::getForArgument(&cast<CallInst>(Inst), 1, TLI);
1454       // Don't start a scope if we already have a better one pushed
1455       if (!AvailableInvariants.count(MemLoc))
1456         AvailableInvariants.insert(MemLoc, CurrentGeneration);
1457       continue;
1458     }
1459 
1460     if (isGuard(&Inst)) {
1461       if (auto *CondI =
1462               dyn_cast<Instruction>(cast<CallInst>(Inst).getArgOperand(0))) {
1463         if (SimpleValue::canHandle(CondI)) {
1464           // Do we already know the actual value of this condition?
1465           if (auto *KnownCond = AvailableValues.lookup(CondI)) {
1466             // Is the condition known to be true?
1467             if (isa<ConstantInt>(KnownCond) &&
1468                 cast<ConstantInt>(KnownCond)->isOne()) {
1469               LLVM_DEBUG(dbgs()
1470                          << "EarlyCSE removing guard: " << Inst << '\n');
1471               salvageKnowledge(&Inst, &AC);
1472               removeMSSA(Inst);
1473               Inst.eraseFromParent();
1474               Changed = true;
1475               continue;
1476             } else
1477               // Use the known value if it wasn't true.
1478               cast<CallInst>(Inst).setArgOperand(0, KnownCond);
1479           }
1480           // The condition we're on guarding here is true for all dominated
1481           // locations.
1482           AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext()));
1483         }
1484       }
1485 
1486       // Guard intrinsics read all memory, but don't write any memory.
1487       // Accordingly, don't update the generation but consume the last store (to
1488       // avoid an incorrect DSE).
1489       LastStore = nullptr;
1490       continue;
1491     }
1492 
1493     // If the instruction can be simplified (e.g. X+0 = X) then replace it with
1494     // its simpler value.
1495     if (Value *V = simplifyInstruction(&Inst, SQ)) {
1496       LLVM_DEBUG(dbgs() << "EarlyCSE Simplify: " << Inst << "  to: " << *V
1497                         << '\n');
1498       if (!DebugCounter::shouldExecute(CSECounter)) {
1499         LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1500       } else {
1501         bool Killed = false;
1502         if (!Inst.use_empty()) {
1503           Inst.replaceAllUsesWith(V);
1504           Changed = true;
1505         }
1506         if (isInstructionTriviallyDead(&Inst, &TLI)) {
1507           salvageKnowledge(&Inst, &AC);
1508           removeMSSA(Inst);
1509           Inst.eraseFromParent();
1510           Changed = true;
1511           Killed = true;
1512         }
1513         if (Changed)
1514           ++NumSimplify;
1515         if (Killed)
1516           continue;
1517       }
1518     }
1519 
1520     // If this is a simple instruction that we can value number, process it.
1521     if (SimpleValue::canHandle(&Inst)) {
1522       if ([[maybe_unused]] auto *CI = dyn_cast<ConstrainedFPIntrinsic>(&Inst)) {
1523         assert(CI->getExceptionBehavior() != fp::ebStrict &&
1524                "Unexpected ebStrict from SimpleValue::canHandle()");
1525         assert((!CI->getRoundingMode() ||
1526                 CI->getRoundingMode() != RoundingMode::Dynamic) &&
1527                "Unexpected dynamic rounding from SimpleValue::canHandle()");
1528       }
1529       // See if the instruction has an available value.  If so, use it.
1530       if (Value *V = AvailableValues.lookup(&Inst)) {
1531         LLVM_DEBUG(dbgs() << "EarlyCSE CSE: " << Inst << "  to: " << *V
1532                           << '\n');
1533         if (!DebugCounter::shouldExecute(CSECounter)) {
1534           LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1535           continue;
1536         }
1537         combineIRFlags(Inst, V);
1538         Inst.replaceAllUsesWith(V);
1539         salvageKnowledge(&Inst, &AC);
1540         removeMSSA(Inst);
1541         Inst.eraseFromParent();
1542         Changed = true;
1543         ++NumCSE;
1544         continue;
1545       }
1546 
1547       // Otherwise, just remember that this value is available.
1548       AvailableValues.insert(&Inst, &Inst);
1549       continue;
1550     }
1551 
1552     ParseMemoryInst MemInst(&Inst, TTI);
1553     // If this is a non-volatile load, process it.
1554     if (MemInst.isValid() && MemInst.isLoad()) {
1555       // (conservatively) we can't peak past the ordering implied by this
1556       // operation, but we can add this load to our set of available values
1557       if (MemInst.isVolatile() || !MemInst.isUnordered()) {
1558         LastStore = nullptr;
1559         ++CurrentGeneration;
1560       }
1561 
1562       if (MemInst.isInvariantLoad()) {
1563         // If we pass an invariant load, we know that memory location is
1564         // indefinitely constant from the moment of first dereferenceability.
1565         // We conservatively treat the invariant_load as that moment.  If we
1566         // pass a invariant load after already establishing a scope, don't
1567         // restart it since we want to preserve the earliest point seen.
1568         auto MemLoc = MemoryLocation::get(&Inst);
1569         if (!AvailableInvariants.count(MemLoc))
1570           AvailableInvariants.insert(MemLoc, CurrentGeneration);
1571       }
1572 
1573       // If we have an available version of this load, and if it is the right
1574       // generation or the load is known to be from an invariant location,
1575       // replace this instruction.
1576       //
1577       // If either the dominating load or the current load are invariant, then
1578       // we can assume the current load loads the same value as the dominating
1579       // load.
1580       LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand());
1581       if (Value *Op = getMatchingValue(InVal, MemInst, CurrentGeneration)) {
1582         LLVM_DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << Inst
1583                           << "  to: " << *InVal.DefInst << '\n');
1584         if (!DebugCounter::shouldExecute(CSECounter)) {
1585           LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1586           continue;
1587         }
1588         if (InVal.IsLoad)
1589           if (auto *I = dyn_cast<Instruction>(Op))
1590             combineMetadataForCSE(I, &Inst, false);
1591         if (!Inst.use_empty())
1592           Inst.replaceAllUsesWith(Op);
1593         salvageKnowledge(&Inst, &AC);
1594         removeMSSA(Inst);
1595         Inst.eraseFromParent();
1596         Changed = true;
1597         ++NumCSELoad;
1598         continue;
1599       }
1600 
1601       // Otherwise, remember that we have this instruction.
1602       AvailableLoads.insert(MemInst.getPointerOperand(),
1603                             LoadValue(&Inst, CurrentGeneration,
1604                                       MemInst.getMatchingId(),
1605                                       MemInst.isAtomic(),
1606                                       MemInst.isLoad()));
1607       LastStore = nullptr;
1608       continue;
1609     }
1610 
1611     // If this instruction may read from memory or throw (and potentially read
1612     // from memory in the exception handler), forget LastStore.  Load/store
1613     // intrinsics will indicate both a read and a write to memory.  The target
1614     // may override this (e.g. so that a store intrinsic does not read from
1615     // memory, and thus will be treated the same as a regular store for
1616     // commoning purposes).
1617     if ((Inst.mayReadFromMemory() || Inst.mayThrow()) &&
1618         !(MemInst.isValid() && !MemInst.mayReadFromMemory()))
1619       LastStore = nullptr;
1620 
1621     // If this is a read-only call, process it.
1622     if (CallValue::canHandle(&Inst)) {
1623       // If we have an available version of this call, and if it is the right
1624       // generation, replace this instruction.
1625       std::pair<Instruction *, unsigned> InVal = AvailableCalls.lookup(&Inst);
1626       if (InVal.first != nullptr &&
1627           isSameMemGeneration(InVal.second, CurrentGeneration, InVal.first,
1628                               &Inst)) {
1629         LLVM_DEBUG(dbgs() << "EarlyCSE CSE CALL: " << Inst
1630                           << "  to: " << *InVal.first << '\n');
1631         if (!DebugCounter::shouldExecute(CSECounter)) {
1632           LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1633           continue;
1634         }
1635         if (!Inst.use_empty())
1636           Inst.replaceAllUsesWith(InVal.first);
1637         salvageKnowledge(&Inst, &AC);
1638         removeMSSA(Inst);
1639         Inst.eraseFromParent();
1640         Changed = true;
1641         ++NumCSECall;
1642         continue;
1643       }
1644 
1645       // Otherwise, remember that we have this instruction.
1646       AvailableCalls.insert(&Inst, std::make_pair(&Inst, CurrentGeneration));
1647       continue;
1648     }
1649 
1650     // Compare GEP instructions based on offset.
1651     if (GEPValue::canHandle(&Inst)) {
1652       auto *GEP = cast<GetElementPtrInst>(&Inst);
1653       APInt Offset = APInt(SQ.DL.getIndexTypeSizeInBits(GEP->getType()), 0);
1654       GEPValue GEPVal(GEP, GEP->accumulateConstantOffset(SQ.DL, Offset)
1655                                ? Offset.trySExtValue()
1656                                : std::nullopt);
1657       if (Value *V = AvailableGEPs.lookup(GEPVal)) {
1658         LLVM_DEBUG(dbgs() << "EarlyCSE CSE GEP: " << Inst << "  to: " << *V
1659                           << '\n');
1660         combineIRFlags(Inst, V);
1661         Inst.replaceAllUsesWith(V);
1662         salvageKnowledge(&Inst, &AC);
1663         removeMSSA(Inst);
1664         Inst.eraseFromParent();
1665         Changed = true;
1666         ++NumCSEGEP;
1667         continue;
1668       }
1669 
1670       // Otherwise, just remember that we have this GEP.
1671       AvailableGEPs.insert(GEPVal, &Inst);
1672       continue;
1673     }
1674 
1675     // A release fence requires that all stores complete before it, but does
1676     // not prevent the reordering of following loads 'before' the fence.  As a
1677     // result, we don't need to consider it as writing to memory and don't need
1678     // to advance the generation.  We do need to prevent DSE across the fence,
1679     // but that's handled above.
1680     if (auto *FI = dyn_cast<FenceInst>(&Inst))
1681       if (FI->getOrdering() == AtomicOrdering::Release) {
1682         assert(Inst.mayReadFromMemory() && "relied on to prevent DSE above");
1683         continue;
1684       }
1685 
1686     // write back DSE - If we write back the same value we just loaded from
1687     // the same location and haven't passed any intervening writes or ordering
1688     // operations, we can remove the write.  The primary benefit is in allowing
1689     // the available load table to remain valid and value forward past where
1690     // the store originally was.
1691     if (MemInst.isValid() && MemInst.isStore()) {
1692       LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand());
1693       if (InVal.DefInst &&
1694           InVal.DefInst == getMatchingValue(InVal, MemInst, CurrentGeneration)) {
1695         // It is okay to have a LastStore to a different pointer here if MemorySSA
1696         // tells us that the load and store are from the same memory generation.
1697         // In that case, LastStore should keep its present value since we're
1698         // removing the current store.
1699         assert((!LastStore ||
1700                 ParseMemoryInst(LastStore, TTI).getPointerOperand() ==
1701                     MemInst.getPointerOperand() ||
1702                 MSSA) &&
1703                "can't have an intervening store if not using MemorySSA!");
1704         LLVM_DEBUG(dbgs() << "EarlyCSE DSE (writeback): " << Inst << '\n');
1705         if (!DebugCounter::shouldExecute(CSECounter)) {
1706           LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1707           continue;
1708         }
1709         salvageKnowledge(&Inst, &AC);
1710         removeMSSA(Inst);
1711         Inst.eraseFromParent();
1712         Changed = true;
1713         ++NumDSE;
1714         // We can avoid incrementing the generation count since we were able
1715         // to eliminate this store.
1716         continue;
1717       }
1718     }
1719 
1720     // Okay, this isn't something we can CSE at all.  Check to see if it is
1721     // something that could modify memory.  If so, our available memory values
1722     // cannot be used so bump the generation count.
1723     if (Inst.mayWriteToMemory()) {
1724       ++CurrentGeneration;
1725 
1726       if (MemInst.isValid() && MemInst.isStore()) {
1727         // We do a trivial form of DSE if there are two stores to the same
1728         // location with no intervening loads.  Delete the earlier store.
1729         if (LastStore) {
1730           if (overridingStores(ParseMemoryInst(LastStore, TTI), MemInst)) {
1731             LLVM_DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore
1732                               << "  due to: " << Inst << '\n');
1733             if (!DebugCounter::shouldExecute(CSECounter)) {
1734               LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n");
1735             } else {
1736               salvageKnowledge(&Inst, &AC);
1737               removeMSSA(*LastStore);
1738               LastStore->eraseFromParent();
1739               Changed = true;
1740               ++NumDSE;
1741               LastStore = nullptr;
1742             }
1743           }
1744           // fallthrough - we can exploit information about this store
1745         }
1746 
1747         // Okay, we just invalidated anything we knew about loaded values.  Try
1748         // to salvage *something* by remembering that the stored value is a live
1749         // version of the pointer.  It is safe to forward from volatile stores
1750         // to non-volatile loads, so we don't have to check for volatility of
1751         // the store.
1752         AvailableLoads.insert(MemInst.getPointerOperand(),
1753                               LoadValue(&Inst, CurrentGeneration,
1754                                         MemInst.getMatchingId(),
1755                                         MemInst.isAtomic(),
1756                                         MemInst.isLoad()));
1757 
1758         // Remember that this was the last unordered store we saw for DSE. We
1759         // don't yet handle DSE on ordered or volatile stores since we don't
1760         // have a good way to model the ordering requirement for following
1761         // passes  once the store is removed.  We could insert a fence, but
1762         // since fences are slightly stronger than stores in their ordering,
1763         // it's not clear this is a profitable transform. Another option would
1764         // be to merge the ordering with that of the post dominating store.
1765         if (MemInst.isUnordered() && !MemInst.isVolatile())
1766           LastStore = &Inst;
1767         else
1768           LastStore = nullptr;
1769       }
1770     }
1771   }
1772 
1773   return Changed;
1774 }
1775 
1776 bool EarlyCSE::run() {
1777   // Note, deque is being used here because there is significant performance
1778   // gains over vector when the container becomes very large due to the
1779   // specific access patterns. For more information see the mailing list
1780   // discussion on this:
1781   // http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html
1782   std::deque<StackNode *> nodesToProcess;
1783 
1784   bool Changed = false;
1785 
1786   // Process the root node.
1787   nodesToProcess.push_back(new StackNode(
1788       AvailableValues, AvailableLoads, AvailableInvariants, AvailableCalls,
1789       AvailableGEPs, CurrentGeneration, DT.getRootNode(),
1790       DT.getRootNode()->begin(), DT.getRootNode()->end()));
1791 
1792   assert(!CurrentGeneration && "Create a new EarlyCSE instance to rerun it.");
1793 
1794   // Process the stack.
1795   while (!nodesToProcess.empty()) {
1796     // Grab the first item off the stack. Set the current generation, remove
1797     // the node from the stack, and process it.
1798     StackNode *NodeToProcess = nodesToProcess.back();
1799 
1800     // Initialize class members.
1801     CurrentGeneration = NodeToProcess->currentGeneration();
1802 
1803     // Check if the node needs to be processed.
1804     if (!NodeToProcess->isProcessed()) {
1805       // Process the node.
1806       Changed |= processNode(NodeToProcess->node());
1807       NodeToProcess->childGeneration(CurrentGeneration);
1808       NodeToProcess->process();
1809     } else if (NodeToProcess->childIter() != NodeToProcess->end()) {
1810       // Push the next child onto the stack.
1811       DomTreeNode *child = NodeToProcess->nextChild();
1812       nodesToProcess.push_back(new StackNode(
1813           AvailableValues, AvailableLoads, AvailableInvariants, AvailableCalls,
1814           AvailableGEPs, NodeToProcess->childGeneration(), child,
1815           child->begin(), child->end()));
1816     } else {
1817       // It has been processed, and there are no more children to process,
1818       // so delete it and pop it off the stack.
1819       delete NodeToProcess;
1820       nodesToProcess.pop_back();
1821     }
1822   } // while (!nodes...)
1823 
1824   return Changed;
1825 }
1826 
1827 PreservedAnalyses EarlyCSEPass::run(Function &F,
1828                                     FunctionAnalysisManager &AM) {
1829   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1830   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
1831   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1832   auto &AC = AM.getResult<AssumptionAnalysis>(F);
1833   auto *MSSA =
1834       UseMemorySSA ? &AM.getResult<MemorySSAAnalysis>(F).getMSSA() : nullptr;
1835 
1836   EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA);
1837 
1838   if (!CSE.run())
1839     return PreservedAnalyses::all();
1840 
1841   PreservedAnalyses PA;
1842   PA.preserveSet<CFGAnalyses>();
1843   if (UseMemorySSA)
1844     PA.preserve<MemorySSAAnalysis>();
1845   return PA;
1846 }
1847 
1848 void EarlyCSEPass::printPipeline(
1849     raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
1850   static_cast<PassInfoMixin<EarlyCSEPass> *>(this)->printPipeline(
1851       OS, MapClassName2PassName);
1852   OS << '<';
1853   if (UseMemorySSA)
1854     OS << "memssa";
1855   OS << '>';
1856 }
1857 
1858 namespace {
1859 
1860 /// A simple and fast domtree-based CSE pass.
1861 ///
1862 /// This pass does a simple depth-first walk over the dominator tree,
1863 /// eliminating trivially redundant instructions and using instsimplify to
1864 /// canonicalize things as it goes. It is intended to be fast and catch obvious
1865 /// cases so that instcombine and other passes are more effective. It is
1866 /// expected that a later pass of GVN will catch the interesting/hard cases.
1867 template<bool UseMemorySSA>
1868 class EarlyCSELegacyCommonPass : public FunctionPass {
1869 public:
1870   static char ID;
1871 
1872   EarlyCSELegacyCommonPass() : FunctionPass(ID) {
1873     if (UseMemorySSA)
1874       initializeEarlyCSEMemSSALegacyPassPass(*PassRegistry::getPassRegistry());
1875     else
1876       initializeEarlyCSELegacyPassPass(*PassRegistry::getPassRegistry());
1877   }
1878 
1879   bool runOnFunction(Function &F) override {
1880     if (skipFunction(F))
1881       return false;
1882 
1883     auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1884     auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1885     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1886     auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1887     auto *MSSA =
1888         UseMemorySSA ? &getAnalysis<MemorySSAWrapperPass>().getMSSA() : nullptr;
1889 
1890     EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA);
1891 
1892     return CSE.run();
1893   }
1894 
1895   void getAnalysisUsage(AnalysisUsage &AU) const override {
1896     AU.addRequired<AssumptionCacheTracker>();
1897     AU.addRequired<DominatorTreeWrapperPass>();
1898     AU.addRequired<TargetLibraryInfoWrapperPass>();
1899     AU.addRequired<TargetTransformInfoWrapperPass>();
1900     if (UseMemorySSA) {
1901       AU.addRequired<AAResultsWrapperPass>();
1902       AU.addRequired<MemorySSAWrapperPass>();
1903       AU.addPreserved<MemorySSAWrapperPass>();
1904     }
1905     AU.addPreserved<GlobalsAAWrapperPass>();
1906     AU.addPreserved<AAResultsWrapperPass>();
1907     AU.setPreservesCFG();
1908   }
1909 };
1910 
1911 } // end anonymous namespace
1912 
1913 using EarlyCSELegacyPass = EarlyCSELegacyCommonPass</*UseMemorySSA=*/false>;
1914 
1915 template<>
1916 char EarlyCSELegacyPass::ID = 0;
1917 
1918 INITIALIZE_PASS_BEGIN(EarlyCSELegacyPass, "early-cse", "Early CSE", false,
1919                       false)
1920 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1921 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1922 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1923 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1924 INITIALIZE_PASS_END(EarlyCSELegacyPass, "early-cse", "Early CSE", false, false)
1925 
1926 using EarlyCSEMemSSALegacyPass =
1927     EarlyCSELegacyCommonPass</*UseMemorySSA=*/true>;
1928 
1929 template<>
1930 char EarlyCSEMemSSALegacyPass::ID = 0;
1931 
1932 FunctionPass *llvm::createEarlyCSEPass(bool UseMemorySSA) {
1933   if (UseMemorySSA)
1934     return new EarlyCSEMemSSALegacyPass();
1935   else
1936     return new EarlyCSELegacyPass();
1937 }
1938 
1939 INITIALIZE_PASS_BEGIN(EarlyCSEMemSSALegacyPass, "early-cse-memssa",
1940                       "Early CSE w/ MemorySSA", false, false)
1941 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1942 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1943 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1944 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1945 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1946 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
1947 INITIALIZE_PASS_END(EarlyCSEMemSSALegacyPass, "early-cse-memssa",
1948                     "Early CSE w/ MemorySSA", false, false)
1949