xref: /llvm-project/llvm/lib/Transforms/Scalar/ConstraintElimination.cpp (revision 9aa004a04c0ef5f6f317fa3e28ee96713bd4a2cd)
1 //===-- ConstraintElimination.cpp - Eliminate conds using constraints. ----===//
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 // Eliminate conditions based on constraints collected from dominating
10 // conditions.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Transforms/Scalar/ConstraintElimination.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/ScopeExit.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/ConstraintSystem.h"
20 #include "llvm/Analysis/GlobalsModRef.h"
21 #include "llvm/Analysis/ValueTracking.h"
22 #include "llvm/IR/Dominators.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/PatternMatch.h"
27 #include "llvm/InitializePasses.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/DebugCounter.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/Transforms/Scalar.h"
33 
34 #include <cmath>
35 #include <string>
36 
37 using namespace llvm;
38 using namespace PatternMatch;
39 
40 #define DEBUG_TYPE "constraint-elimination"
41 
42 STATISTIC(NumCondsRemoved, "Number of instructions removed");
43 DEBUG_COUNTER(EliminatedCounter, "conds-eliminated",
44               "Controls which conditions are eliminated");
45 
46 static int64_t MaxConstraintValue = std::numeric_limits<int64_t>::max();
47 static int64_t MinSignedConstraintValue = std::numeric_limits<int64_t>::min();
48 
49 namespace {
50 
51 class ConstraintInfo;
52 
53 struct StackEntry {
54   unsigned NumIn;
55   unsigned NumOut;
56   bool IsSigned = false;
57   /// Variables that can be removed from the system once the stack entry gets
58   /// removed.
59   SmallVector<Value *, 2> ValuesToRelease;
60 
61   StackEntry(unsigned NumIn, unsigned NumOut, bool IsSigned,
62              SmallVector<Value *, 2> ValuesToRelease)
63       : NumIn(NumIn), NumOut(NumOut), IsSigned(IsSigned),
64         ValuesToRelease(ValuesToRelease) {}
65 };
66 
67 /// Struct to express a pre-condition of the form %Op0 Pred %Op1.
68 struct PreconditionTy {
69   CmpInst::Predicate Pred;
70   Value *Op0;
71   Value *Op1;
72 
73   PreconditionTy(CmpInst::Predicate Pred, Value *Op0, Value *Op1)
74       : Pred(Pred), Op0(Op0), Op1(Op1) {}
75 };
76 
77 struct ConstraintTy {
78   SmallVector<int64_t, 8> Coefficients;
79   SmallVector<PreconditionTy, 2> Preconditions;
80 
81   SmallVector<SmallVector<int64_t, 8>> ExtraInfo;
82 
83   bool IsSigned = false;
84   bool IsEq = false;
85 
86   ConstraintTy() = default;
87 
88   ConstraintTy(SmallVector<int64_t, 8> Coefficients, bool IsSigned)
89       : Coefficients(Coefficients), IsSigned(IsSigned) {}
90 
91   unsigned size() const { return Coefficients.size(); }
92 
93   unsigned empty() const { return Coefficients.empty(); }
94 
95   /// Returns true if all preconditions for this list of constraints are
96   /// satisfied given \p CS and the corresponding \p Value2Index mapping.
97   bool isValid(const ConstraintInfo &Info) const;
98 };
99 
100 /// Wrapper encapsulating separate constraint systems and corresponding value
101 /// mappings for both unsigned and signed information. Facts are added to and
102 /// conditions are checked against the corresponding system depending on the
103 /// signed-ness of their predicates. While the information is kept separate
104 /// based on signed-ness, certain conditions can be transferred between the two
105 /// systems.
106 class ConstraintInfo {
107   DenseMap<Value *, unsigned> UnsignedValue2Index;
108   DenseMap<Value *, unsigned> SignedValue2Index;
109 
110   ConstraintSystem UnsignedCS;
111   ConstraintSystem SignedCS;
112 
113 public:
114   DenseMap<Value *, unsigned> &getValue2Index(bool Signed) {
115     return Signed ? SignedValue2Index : UnsignedValue2Index;
116   }
117   const DenseMap<Value *, unsigned> &getValue2Index(bool Signed) const {
118     return Signed ? SignedValue2Index : UnsignedValue2Index;
119   }
120 
121   ConstraintSystem &getCS(bool Signed) {
122     return Signed ? SignedCS : UnsignedCS;
123   }
124   const ConstraintSystem &getCS(bool Signed) const {
125     return Signed ? SignedCS : UnsignedCS;
126   }
127 
128   void popLastConstraint(bool Signed) { getCS(Signed).popLastConstraint(); }
129   void popLastNVariables(bool Signed, unsigned N) {
130     getCS(Signed).popLastNVariables(N);
131   }
132 
133   bool doesHold(CmpInst::Predicate Pred, Value *A, Value *B) const;
134 
135   void addFact(CmpInst::Predicate Pred, Value *A, Value *B, unsigned NumIn,
136                unsigned NumOut, SmallVectorImpl<StackEntry> &DFSInStack);
137 
138   /// Turn a comparison of the form \p Op0 \p Pred \p Op1 into a vector of
139   /// constraints, using indices from the corresponding constraint system.
140   /// New variables that need to be added to the system are collected in
141   /// \p NewVariables.
142   ConstraintTy getConstraint(CmpInst::Predicate Pred, Value *Op0, Value *Op1,
143                              SmallVectorImpl<Value *> &NewVariables) const;
144 
145   /// Turn a condition \p CmpI into a vector of constraints, using indices from
146   /// the corresponding constraint system. New variables that need to be added
147   /// to the system are collected in \p NewVariables.
148   ConstraintTy getConstraint(CmpInst *Cmp,
149                              SmallVectorImpl<Value *> &NewVariables) {
150     return getConstraint(Cmp->getPredicate(), Cmp->getOperand(0),
151                          Cmp->getOperand(1), NewVariables);
152   }
153 
154   /// Try to add information from \p A \p Pred \p B to the unsigned/signed
155   /// system if \p Pred is signed/unsigned.
156   void transferToOtherSystem(CmpInst::Predicate Pred, Value *A, Value *B,
157                              unsigned NumIn, unsigned NumOut,
158                              SmallVectorImpl<StackEntry> &DFSInStack);
159 };
160 
161 /// Represents a (Coefficient * Variable) entry after IR decomposition.
162 struct DecompEntry {
163   int64_t Coefficient;
164   Value *Variable;
165   /// True if the variable is known positive in the current constraint.
166   bool IsKnownPositive;
167 
168   DecompEntry(int64_t Coefficient, Value *Variable,
169               bool IsKnownPositive = false)
170       : Coefficient(Coefficient), Variable(Variable),
171         IsKnownPositive(IsKnownPositive) {}
172 };
173 
174 } // namespace
175 
176 // Decomposes \p V into a vector of entries of the form { Coefficient, Variable
177 // } where Coefficient * Variable. The sum of the pairs equals \p V.  The first
178 // pair is the constant-factor and X must be nullptr. If the expression cannot
179 // be decomposed, returns an empty vector.
180 static SmallVector<DecompEntry, 4>
181 decompose(Value *V, SmallVector<PreconditionTy, 4> &Preconditions,
182           bool IsSigned) {
183 
184   auto CanUseSExt = [](ConstantInt *CI) {
185     const APInt &Val = CI->getValue();
186     return Val.sgt(MinSignedConstraintValue) && Val.slt(MaxConstraintValue);
187   };
188   // Decompose \p V used with a signed predicate.
189   if (IsSigned) {
190     if (auto *CI = dyn_cast<ConstantInt>(V)) {
191       if (CanUseSExt(CI))
192         return {{CI->getSExtValue(), nullptr}};
193     }
194 
195     return {{0, nullptr}, {1, V}};
196   }
197 
198   if (auto *CI = dyn_cast<ConstantInt>(V)) {
199     if (CI->uge(MaxConstraintValue))
200       return {};
201     return {{int64_t(CI->getZExtValue()), nullptr}};
202   }
203   auto *GEP = dyn_cast<GetElementPtrInst>(V);
204   if (GEP && GEP->getNumOperands() == 2 && GEP->isInBounds()) {
205     Value *Op0, *Op1;
206     ConstantInt *CI;
207 
208     // If the index is zero-extended, it is guaranteed to be positive.
209     if (match(GEP->getOperand(GEP->getNumOperands() - 1),
210               m_ZExt(m_Value(Op0)))) {
211       if (match(Op0, m_NUWShl(m_Value(Op1), m_ConstantInt(CI))) &&
212           CanUseSExt(CI))
213         return {{0, nullptr},
214                 {1, GEP->getPointerOperand()},
215                 {int64_t(std::pow(int64_t(2), CI->getSExtValue())), Op1}};
216       if (match(Op0, m_NSWAdd(m_Value(Op1), m_ConstantInt(CI))) &&
217           CanUseSExt(CI))
218         return {{CI->getSExtValue(), nullptr},
219                 {1, GEP->getPointerOperand()},
220                 {1, Op1}};
221       return {{0, nullptr}, {1, GEP->getPointerOperand()}, {1, Op0, true}};
222     }
223 
224     if (match(GEP->getOperand(GEP->getNumOperands() - 1), m_ConstantInt(CI)) &&
225         !CI->isNegative() && CanUseSExt(CI))
226       return {{CI->getSExtValue(), nullptr}, {1, GEP->getPointerOperand()}};
227 
228     SmallVector<DecompEntry, 4> Result;
229     if (match(GEP->getOperand(GEP->getNumOperands() - 1),
230               m_NUWShl(m_Value(Op0), m_ConstantInt(CI))) &&
231         CanUseSExt(CI))
232       Result = {{0, nullptr},
233                 {1, GEP->getPointerOperand()},
234                 {int(std::pow(int64_t(2), CI->getSExtValue())), Op0}};
235     else if (match(GEP->getOperand(GEP->getNumOperands() - 1),
236                    m_NSWAdd(m_Value(Op0), m_ConstantInt(CI))) &&
237              CanUseSExt(CI))
238       Result = {{CI->getSExtValue(), nullptr},
239                 {1, GEP->getPointerOperand()},
240                 {1, Op0}};
241     else {
242       Op0 = GEP->getOperand(GEP->getNumOperands() - 1);
243       Result = {{0, nullptr}, {1, GEP->getPointerOperand()}, {1, Op0}};
244     }
245     // If Op0 is signed non-negative, the GEP is increasing monotonically and
246     // can be de-composed.
247     Preconditions.emplace_back(CmpInst::ICMP_SGE, Op0,
248                                ConstantInt::get(Op0->getType(), 0));
249     return Result;
250   }
251 
252   Value *Op0;
253   bool IsKnownPositive = false;
254   if (match(V, m_ZExt(m_Value(Op0)))) {
255     IsKnownPositive = true;
256     V = Op0;
257   }
258 
259   auto MergeResults = [&Preconditions, IsSigned](
260                           Value *A, Value *B,
261                           bool IsSignedB) -> SmallVector<DecompEntry, 4> {
262     auto ResA = decompose(A, Preconditions, IsSigned);
263     auto ResB = decompose(B, Preconditions, IsSignedB);
264     if (ResA.empty() || ResB.empty())
265       return {};
266     ResA[0].Coefficient += ResB[0].Coefficient;
267     append_range(ResA, drop_begin(ResB));
268     return ResA;
269   };
270   Value *Op1;
271   ConstantInt *CI;
272   if (match(V, m_NUWAdd(m_Value(Op0), m_Value(Op1)))) {
273     return MergeResults(Op0, Op1, IsSigned);
274   }
275   if (match(V, m_Add(m_Value(Op0), m_ConstantInt(CI))) && CI->isNegative() &&
276       CanUseSExt(CI)) {
277     Preconditions.emplace_back(
278         CmpInst::ICMP_UGE, Op0,
279         ConstantInt::get(Op0->getType(), CI->getSExtValue() * -1));
280     return MergeResults(Op0, CI, true);
281   }
282 
283   if (match(V, m_NUWSub(m_Value(Op0), m_ConstantInt(CI))) && CanUseSExt(CI))
284     return {{-1 * CI->getSExtValue(), nullptr}, {1, Op0}};
285   if (match(V, m_NUWSub(m_Value(Op0), m_Value(Op1))))
286     return {{0, nullptr}, {1, Op0}, {-1, Op1}};
287 
288   return {{0, nullptr}, {1, V, IsKnownPositive}};
289 }
290 
291 ConstraintTy
292 ConstraintInfo::getConstraint(CmpInst::Predicate Pred, Value *Op0, Value *Op1,
293                               SmallVectorImpl<Value *> &NewVariables) const {
294   assert(NewVariables.empty() && "NewVariables must be empty when passed in");
295   bool IsEq = false;
296   // Try to convert Pred to one of ULE/SLT/SLE/SLT.
297   switch (Pred) {
298   case CmpInst::ICMP_UGT:
299   case CmpInst::ICMP_UGE:
300   case CmpInst::ICMP_SGT:
301   case CmpInst::ICMP_SGE: {
302     Pred = CmpInst::getSwappedPredicate(Pred);
303     std::swap(Op0, Op1);
304     break;
305   }
306   case CmpInst::ICMP_EQ:
307     if (match(Op1, m_Zero())) {
308       Pred = CmpInst::ICMP_ULE;
309     } else {
310       IsEq = true;
311       Pred = CmpInst::ICMP_ULE;
312     }
313     break;
314   case CmpInst::ICMP_NE:
315     if (!match(Op1, m_Zero()))
316       return {};
317     Pred = CmpInst::getSwappedPredicate(CmpInst::ICMP_UGT);
318     std::swap(Op0, Op1);
319     break;
320   default:
321     break;
322   }
323 
324   // Only ULE and ULT predicates are supported at the moment.
325   if (Pred != CmpInst::ICMP_ULE && Pred != CmpInst::ICMP_ULT &&
326       Pred != CmpInst::ICMP_SLE && Pred != CmpInst::ICMP_SLT)
327     return {};
328 
329   SmallVector<PreconditionTy, 4> Preconditions;
330   bool IsSigned = CmpInst::isSigned(Pred);
331   auto &Value2Index = getValue2Index(IsSigned);
332   auto ADec = decompose(Op0->stripPointerCastsSameRepresentation(),
333                         Preconditions, IsSigned);
334   auto BDec = decompose(Op1->stripPointerCastsSameRepresentation(),
335                         Preconditions, IsSigned);
336   // Skip if decomposing either of the values failed.
337   if (ADec.empty() || BDec.empty())
338     return {};
339 
340   int64_t Offset1 = ADec[0].Coefficient;
341   int64_t Offset2 = BDec[0].Coefficient;
342   Offset1 *= -1;
343 
344   // Create iterator ranges that skip the constant-factor.
345   auto VariablesA = llvm::drop_begin(ADec);
346   auto VariablesB = llvm::drop_begin(BDec);
347 
348   // First try to look up \p V in Value2Index and NewVariables. Otherwise add a
349   // new entry to NewVariables.
350   DenseMap<Value *, unsigned> NewIndexMap;
351   auto GetOrAddIndex = [&Value2Index, &NewVariables,
352                         &NewIndexMap](Value *V) -> unsigned {
353     auto V2I = Value2Index.find(V);
354     if (V2I != Value2Index.end())
355       return V2I->second;
356     auto Insert =
357         NewIndexMap.insert({V, Value2Index.size() + NewVariables.size() + 1});
358     if (Insert.second)
359       NewVariables.push_back(V);
360     return Insert.first->second;
361   };
362 
363   // Make sure all variables have entries in Value2Index or NewVariables.
364   for (const auto &KV : concat<DecompEntry>(VariablesA, VariablesB))
365     GetOrAddIndex(KV.Variable);
366 
367   // Build result constraint, by first adding all coefficients from A and then
368   // subtracting all coefficients from B.
369   ConstraintTy Res(
370       SmallVector<int64_t, 8>(Value2Index.size() + NewVariables.size() + 1, 0),
371       IsSigned);
372   // Collect variables that are known to be positive in all uses in the
373   // constraint.
374   DenseMap<Value *, bool> KnownPositiveVariables;
375   Res.IsEq = IsEq;
376   auto &R = Res.Coefficients;
377   for (const auto &KV : VariablesA) {
378     R[GetOrAddIndex(KV.Variable)] += KV.Coefficient;
379     auto I = KnownPositiveVariables.insert({KV.Variable, KV.IsKnownPositive});
380     I.first->second &= KV.IsKnownPositive;
381   }
382 
383   for (const auto &KV : VariablesB) {
384     R[GetOrAddIndex(KV.Variable)] -= KV.Coefficient;
385     auto I = KnownPositiveVariables.insert({KV.Variable, KV.IsKnownPositive});
386     I.first->second &= KV.IsKnownPositive;
387   }
388 
389   int64_t OffsetSum;
390   if (AddOverflow(Offset1, Offset2, OffsetSum))
391     return {};
392   if (Pred == (IsSigned ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT))
393     if (AddOverflow(OffsetSum, int64_t(-1), OffsetSum))
394       return {};
395   R[0] = OffsetSum;
396   Res.Preconditions = std::move(Preconditions);
397 
398   // Remove any (Coefficient, Variable) entry where the Coefficient is 0 for new
399   // variables.
400   while (!NewVariables.empty()) {
401     int64_t Last = R.back();
402     if (Last != 0)
403       break;
404     R.pop_back();
405     Value *RemovedV = NewVariables.pop_back_val();
406     NewIndexMap.erase(RemovedV);
407   }
408 
409   // Add extra constraints for variables that are known positive.
410   for (auto &KV : KnownPositiveVariables) {
411     if (!KV.second || (Value2Index.find(KV.first) == Value2Index.end() &&
412                        NewIndexMap.find(KV.first) == NewIndexMap.end()))
413       continue;
414     SmallVector<int64_t, 8> C(Value2Index.size() + NewVariables.size() + 1, 0);
415     C[GetOrAddIndex(KV.first)] = -1;
416     Res.ExtraInfo.push_back(C);
417   }
418   return Res;
419 }
420 
421 bool ConstraintTy::isValid(const ConstraintInfo &Info) const {
422   return Coefficients.size() > 0 &&
423          all_of(Preconditions, [&Info](const PreconditionTy &C) {
424            return Info.doesHold(C.Pred, C.Op0, C.Op1);
425          });
426 }
427 
428 bool ConstraintInfo::doesHold(CmpInst::Predicate Pred, Value *A,
429                               Value *B) const {
430   SmallVector<Value *> NewVariables;
431   auto R = getConstraint(Pred, A, B, NewVariables);
432 
433   if (!NewVariables.empty())
434     return false;
435 
436   return NewVariables.empty() && R.Preconditions.empty() && !R.IsEq &&
437          !R.empty() &&
438          getCS(CmpInst::isSigned(Pred)).isConditionImplied(R.Coefficients);
439 }
440 
441 void ConstraintInfo::transferToOtherSystem(
442     CmpInst::Predicate Pred, Value *A, Value *B, unsigned NumIn,
443     unsigned NumOut, SmallVectorImpl<StackEntry> &DFSInStack) {
444   // Check if we can combine facts from the signed and unsigned systems to
445   // derive additional facts.
446   if (!A->getType()->isIntegerTy())
447     return;
448   // FIXME: This currently depends on the order we add facts. Ideally we
449   // would first add all known facts and only then try to add additional
450   // facts.
451   switch (Pred) {
452   default:
453     break;
454   case CmpInst::ICMP_ULT:
455     //  If B is a signed positive constant, A >=s 0 and A <s B.
456     if (doesHold(CmpInst::ICMP_SGE, B, ConstantInt::get(B->getType(), 0))) {
457       addFact(CmpInst::ICMP_SGE, A, ConstantInt::get(B->getType(), 0), NumIn,
458               NumOut, DFSInStack);
459       addFact(CmpInst::ICMP_SLT, A, B, NumIn, NumOut, DFSInStack);
460     }
461     break;
462   case CmpInst::ICMP_SLT:
463     if (doesHold(CmpInst::ICMP_SGE, A, ConstantInt::get(B->getType(), 0)))
464       addFact(CmpInst::ICMP_ULT, A, B, NumIn, NumOut, DFSInStack);
465     break;
466   case CmpInst::ICMP_SGT:
467     if (doesHold(CmpInst::ICMP_SGE, B, ConstantInt::get(B->getType(), -1)))
468       addFact(CmpInst::ICMP_UGE, A, ConstantInt::get(B->getType(), 0), NumIn,
469               NumOut, DFSInStack);
470     break;
471   case CmpInst::ICMP_SGE:
472     if (doesHold(CmpInst::ICMP_SGE, B, ConstantInt::get(B->getType(), 0))) {
473       addFact(CmpInst::ICMP_UGE, A, B, NumIn, NumOut, DFSInStack);
474     }
475     break;
476   }
477 }
478 
479 namespace {
480 /// Represents either a condition that holds on entry to a block or a basic
481 /// block, with their respective Dominator DFS in and out numbers.
482 struct ConstraintOrBlock {
483   unsigned NumIn;
484   unsigned NumOut;
485   bool IsBlock;
486   bool Not;
487   union {
488     BasicBlock *BB;
489     CmpInst *Condition;
490   };
491 
492   ConstraintOrBlock(DomTreeNode *DTN)
493       : NumIn(DTN->getDFSNumIn()), NumOut(DTN->getDFSNumOut()), IsBlock(true),
494         BB(DTN->getBlock()) {}
495   ConstraintOrBlock(DomTreeNode *DTN, CmpInst *Condition, bool Not)
496       : NumIn(DTN->getDFSNumIn()), NumOut(DTN->getDFSNumOut()), IsBlock(false),
497         Not(Not), Condition(Condition) {}
498 };
499 
500 /// Keep state required to build worklist.
501 struct State {
502   DominatorTree &DT;
503   SmallVector<ConstraintOrBlock, 64> WorkList;
504 
505   State(DominatorTree &DT) : DT(DT) {}
506 
507   /// Process block \p BB and add known facts to work-list.
508   void addInfoFor(BasicBlock &BB);
509 
510   /// Returns true if we can add a known condition from BB to its successor
511   /// block Succ. Each predecessor of Succ can either be BB or be dominated
512   /// by Succ (e.g. the case when adding a condition from a pre-header to a
513   /// loop header).
514   bool canAddSuccessor(BasicBlock &BB, BasicBlock *Succ) const {
515     if (BB.getSingleSuccessor()) {
516       assert(BB.getSingleSuccessor() == Succ);
517       return DT.properlyDominates(&BB, Succ);
518     }
519     return any_of(successors(&BB),
520                   [Succ](const BasicBlock *S) { return S != Succ; }) &&
521            all_of(predecessors(Succ), [&BB, Succ, this](BasicBlock *Pred) {
522              return Pred == &BB || DT.dominates(Succ, Pred);
523            });
524   }
525 };
526 
527 } // namespace
528 
529 #ifndef NDEBUG
530 static void dumpWithNames(const ConstraintSystem &CS,
531                           DenseMap<Value *, unsigned> &Value2Index) {
532   SmallVector<std::string> Names(Value2Index.size(), "");
533   for (auto &KV : Value2Index) {
534     Names[KV.second - 1] = std::string("%") + KV.first->getName().str();
535   }
536   CS.dump(Names);
537 }
538 
539 static void dumpWithNames(ArrayRef<int64_t> C,
540                           DenseMap<Value *, unsigned> &Value2Index) {
541   ConstraintSystem CS;
542   CS.addVariableRowFill(C);
543   dumpWithNames(CS, Value2Index);
544 }
545 #endif
546 
547 void State::addInfoFor(BasicBlock &BB) {
548   WorkList.emplace_back(DT.getNode(&BB));
549 
550   // True as long as long as the current instruction is guaranteed to execute.
551   bool GuaranteedToExecute = true;
552   // Scan BB for assume calls.
553   // TODO: also use this scan to queue conditions to simplify, so we can
554   // interleave facts from assumes and conditions to simplify in a single
555   // basic block. And to skip another traversal of each basic block when
556   // simplifying.
557   for (Instruction &I : BB) {
558     Value *Cond;
559     // For now, just handle assumes with a single compare as condition.
560     if (match(&I, m_Intrinsic<Intrinsic::assume>(m_Value(Cond))) &&
561         isa<ICmpInst>(Cond)) {
562       if (GuaranteedToExecute) {
563         // The assume is guaranteed to execute when BB is entered, hence Cond
564         // holds on entry to BB.
565         WorkList.emplace_back(DT.getNode(&BB), cast<ICmpInst>(Cond), false);
566       } else {
567         // Otherwise the condition only holds in the successors.
568         for (BasicBlock *Succ : successors(&BB)) {
569           if (!canAddSuccessor(BB, Succ))
570             continue;
571           WorkList.emplace_back(DT.getNode(Succ), cast<ICmpInst>(Cond), false);
572         }
573       }
574     }
575     GuaranteedToExecute &= isGuaranteedToTransferExecutionToSuccessor(&I);
576   }
577 
578   auto *Br = dyn_cast<BranchInst>(BB.getTerminator());
579   if (!Br || !Br->isConditional())
580     return;
581 
582   // If the condition is an OR of 2 compares and the false successor only has
583   // the current block as predecessor, queue both negated conditions for the
584   // false successor.
585   Value *Op0, *Op1;
586   if (match(Br->getCondition(), m_LogicalOr(m_Value(Op0), m_Value(Op1))) &&
587       isa<ICmpInst>(Op0) && isa<ICmpInst>(Op1)) {
588     BasicBlock *FalseSuccessor = Br->getSuccessor(1);
589     if (canAddSuccessor(BB, FalseSuccessor)) {
590       WorkList.emplace_back(DT.getNode(FalseSuccessor), cast<ICmpInst>(Op0),
591                             true);
592       WorkList.emplace_back(DT.getNode(FalseSuccessor), cast<ICmpInst>(Op1),
593                             true);
594     }
595     return;
596   }
597 
598   // If the condition is an AND of 2 compares and the true successor only has
599   // the current block as predecessor, queue both conditions for the true
600   // successor.
601   if (match(Br->getCondition(), m_LogicalAnd(m_Value(Op0), m_Value(Op1))) &&
602       isa<ICmpInst>(Op0) && isa<ICmpInst>(Op1)) {
603     BasicBlock *TrueSuccessor = Br->getSuccessor(0);
604     if (canAddSuccessor(BB, TrueSuccessor)) {
605       WorkList.emplace_back(DT.getNode(TrueSuccessor), cast<ICmpInst>(Op0),
606                             false);
607       WorkList.emplace_back(DT.getNode(TrueSuccessor), cast<ICmpInst>(Op1),
608                             false);
609     }
610     return;
611   }
612 
613   auto *CmpI = dyn_cast<ICmpInst>(Br->getCondition());
614   if (!CmpI)
615     return;
616   if (canAddSuccessor(BB, Br->getSuccessor(0)))
617     WorkList.emplace_back(DT.getNode(Br->getSuccessor(0)), CmpI, false);
618   if (canAddSuccessor(BB, Br->getSuccessor(1)))
619     WorkList.emplace_back(DT.getNode(Br->getSuccessor(1)), CmpI, true);
620 }
621 
622 void ConstraintInfo::addFact(CmpInst::Predicate Pred, Value *A, Value *B,
623                              unsigned NumIn, unsigned NumOut,
624                              SmallVectorImpl<StackEntry> &DFSInStack) {
625   // If the constraint has a pre-condition, skip the constraint if it does not
626   // hold.
627   SmallVector<Value *> NewVariables;
628   auto R = getConstraint(Pred, A, B, NewVariables);
629   if (!R.isValid(*this))
630     return;
631 
632   LLVM_DEBUG(dbgs() << "Adding '" << CmpInst::getPredicateName(Pred) << " ";
633              A->printAsOperand(dbgs(), false); dbgs() << ", ";
634              B->printAsOperand(dbgs(), false); dbgs() << "'\n");
635   bool Added = false;
636   assert(CmpInst::isSigned(Pred) == R.IsSigned &&
637          "condition and constraint signs must match");
638   auto &CSToUse = getCS(R.IsSigned);
639   if (R.Coefficients.empty())
640     return;
641 
642   Added |= CSToUse.addVariableRowFill(R.Coefficients);
643 
644   // If R has been added to the system, add the new variables and queue it for
645   // removal once it goes out-of-scope.
646   if (Added) {
647     SmallVector<Value *, 2> ValuesToRelease;
648     auto &Value2Index = getValue2Index(R.IsSigned);
649     for (Value *V : NewVariables) {
650       Value2Index.insert({V, Value2Index.size() + 1});
651       ValuesToRelease.push_back(V);
652     }
653 
654     LLVM_DEBUG({
655       dbgs() << "  constraint: ";
656       dumpWithNames(R.Coefficients, getValue2Index(R.IsSigned));
657       dbgs() << "\n";
658     });
659 
660     DFSInStack.emplace_back(NumIn, NumOut, R.IsSigned, ValuesToRelease);
661 
662     if (R.IsEq) {
663       // Also add the inverted constraint for equality constraints.
664       for (auto &Coeff : R.Coefficients)
665         Coeff *= -1;
666       CSToUse.addVariableRowFill(R.Coefficients);
667 
668       DFSInStack.emplace_back(NumIn, NumOut, R.IsSigned,
669                               SmallVector<Value *, 2>());
670     }
671   }
672 }
673 
674 static bool
675 tryToSimplifyOverflowMath(IntrinsicInst *II, ConstraintInfo &Info,
676                           SmallVectorImpl<Instruction *> &ToRemove) {
677   auto DoesConditionHold = [](CmpInst::Predicate Pred, Value *A, Value *B,
678                               ConstraintInfo &Info) {
679     SmallVector<Value *> NewVariables;
680     auto R = Info.getConstraint(Pred, A, B, NewVariables);
681     if (R.size() < 2 || !NewVariables.empty() || !R.isValid(Info))
682       return false;
683 
684     auto &CSToUse = Info.getCS(CmpInst::isSigned(Pred));
685     return CSToUse.isConditionImplied(R.Coefficients);
686   };
687 
688   bool Changed = false;
689   if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow) {
690     // If A s>= B && B s>= 0, ssub.with.overflow(a, b) should not overflow and
691     // can be simplified to a regular sub.
692     Value *A = II->getArgOperand(0);
693     Value *B = II->getArgOperand(1);
694     if (!DoesConditionHold(CmpInst::ICMP_SGE, A, B, Info) ||
695         !DoesConditionHold(CmpInst::ICMP_SGE, B,
696                            ConstantInt::get(A->getType(), 0), Info))
697       return false;
698 
699     IRBuilder<> Builder(II->getParent(), II->getIterator());
700     Value *Sub = nullptr;
701     for (User *U : make_early_inc_range(II->users())) {
702       if (match(U, m_ExtractValue<0>(m_Value()))) {
703         if (!Sub)
704           Sub = Builder.CreateSub(A, B);
705         U->replaceAllUsesWith(Sub);
706         Changed = true;
707       } else if (match(U, m_ExtractValue<1>(m_Value()))) {
708         U->replaceAllUsesWith(Builder.getFalse());
709         Changed = true;
710       } else
711         continue;
712 
713       if (U->use_empty()) {
714         auto *I = cast<Instruction>(U);
715         ToRemove.push_back(I);
716         I->setOperand(0, PoisonValue::get(II->getType()));
717         Changed = true;
718       }
719     }
720 
721     if (II->use_empty()) {
722       II->eraseFromParent();
723       Changed = true;
724     }
725   }
726   return Changed;
727 }
728 
729 static bool eliminateConstraints(Function &F, DominatorTree &DT) {
730   bool Changed = false;
731   DT.updateDFSNumbers();
732 
733   ConstraintInfo Info;
734   State S(DT);
735 
736   // First, collect conditions implied by branches and blocks with their
737   // Dominator DFS in and out numbers.
738   for (BasicBlock &BB : F) {
739     if (!DT.getNode(&BB))
740       continue;
741     S.addInfoFor(BB);
742   }
743 
744   // Next, sort worklist by dominance, so that dominating blocks and conditions
745   // come before blocks and conditions dominated by them. If a block and a
746   // condition have the same numbers, the condition comes before the block, as
747   // it holds on entry to the block.
748   stable_sort(S.WorkList, [](const ConstraintOrBlock &A, const ConstraintOrBlock &B) {
749     return std::tie(A.NumIn, A.IsBlock) < std::tie(B.NumIn, B.IsBlock);
750   });
751 
752   SmallVector<Instruction *> ToRemove;
753 
754   // Finally, process ordered worklist and eliminate implied conditions.
755   SmallVector<StackEntry, 16> DFSInStack;
756   for (ConstraintOrBlock &CB : S.WorkList) {
757     // First, pop entries from the stack that are out-of-scope for CB. Remove
758     // the corresponding entry from the constraint system.
759     while (!DFSInStack.empty()) {
760       auto &E = DFSInStack.back();
761       LLVM_DEBUG(dbgs() << "Top of stack : " << E.NumIn << " " << E.NumOut
762                         << "\n");
763       LLVM_DEBUG(dbgs() << "CB: " << CB.NumIn << " " << CB.NumOut << "\n");
764       assert(E.NumIn <= CB.NumIn);
765       if (CB.NumOut <= E.NumOut)
766         break;
767       LLVM_DEBUG({
768         dbgs() << "Removing ";
769         dumpWithNames(Info.getCS(E.IsSigned).getLastConstraint(),
770                       Info.getValue2Index(E.IsSigned));
771         dbgs() << "\n";
772       });
773 
774       Info.popLastConstraint(E.IsSigned);
775       // Remove variables in the system that went out of scope.
776       auto &Mapping = Info.getValue2Index(E.IsSigned);
777       for (Value *V : E.ValuesToRelease)
778         Mapping.erase(V);
779       Info.popLastNVariables(E.IsSigned, E.ValuesToRelease.size());
780       DFSInStack.pop_back();
781     }
782 
783     LLVM_DEBUG({
784       dbgs() << "Processing ";
785       if (CB.IsBlock)
786         dbgs() << *CB.BB;
787       else
788         dbgs() << *CB.Condition;
789       dbgs() << "\n";
790     });
791 
792     // For a block, check if any CmpInsts become known based on the current set
793     // of constraints.
794     if (CB.IsBlock) {
795       for (Instruction &I : make_early_inc_range(*CB.BB)) {
796         if (auto *II = dyn_cast<WithOverflowInst>(&I)) {
797           Changed |= tryToSimplifyOverflowMath(II, Info, ToRemove);
798           continue;
799         }
800         auto *Cmp = dyn_cast<ICmpInst>(&I);
801         if (!Cmp)
802           continue;
803 
804         LLVM_DEBUG(dbgs() << "Checking " << *Cmp << "\n");
805         SmallVector<Value *> NewVariables;
806         auto R = Info.getConstraint(Cmp, NewVariables);
807         if (R.IsEq || R.empty() || !NewVariables.empty() || !R.isValid(Info))
808           continue;
809 
810         auto &CSToUse = Info.getCS(R.IsSigned);
811 
812         // If there was extra information collected during decomposition, apply
813         // it now and remove it immediately once we are done with reasoning
814         // about the constraint.
815         for (auto &Row : R.ExtraInfo)
816           CSToUse.addVariableRow(Row);
817         auto InfoRestorer = make_scope_exit([&]() {
818           for (unsigned I = 0; I < R.ExtraInfo.size(); ++I)
819             CSToUse.popLastConstraint();
820         });
821 
822         if (CSToUse.isConditionImplied(R.Coefficients)) {
823           if (!DebugCounter::shouldExecute(EliminatedCounter))
824             continue;
825 
826           LLVM_DEBUG({
827             dbgs() << "Condition " << *Cmp
828                    << " implied by dominating constraints\n";
829             dumpWithNames(CSToUse, Info.getValue2Index(R.IsSigned));
830           });
831           Cmp->replaceUsesWithIf(
832               ConstantInt::getTrue(F.getParent()->getContext()), [](Use &U) {
833                 // Conditions in an assume trivially simplify to true. Skip uses
834                 // in assume calls to not destroy the available information.
835                 auto *II = dyn_cast<IntrinsicInst>(U.getUser());
836                 return !II || II->getIntrinsicID() != Intrinsic::assume;
837               });
838           NumCondsRemoved++;
839           Changed = true;
840         }
841         if (CSToUse.isConditionImplied(
842                 ConstraintSystem::negate(R.Coefficients))) {
843           if (!DebugCounter::shouldExecute(EliminatedCounter))
844             continue;
845 
846           LLVM_DEBUG({
847             dbgs() << "Condition !" << *Cmp
848                    << " implied by dominating constraints\n";
849             dumpWithNames(CSToUse, Info.getValue2Index(R.IsSigned));
850           });
851           Cmp->replaceAllUsesWith(
852               ConstantInt::getFalse(F.getParent()->getContext()));
853           NumCondsRemoved++;
854           Changed = true;
855         }
856       }
857       continue;
858     }
859 
860     ICmpInst::Predicate Pred;
861     Value *A, *B;
862     if (match(CB.Condition, m_ICmp(Pred, m_Value(A), m_Value(B)))) {
863       // Use the inverse predicate if required.
864       if (CB.Not)
865         Pred = CmpInst::getInversePredicate(Pred);
866 
867       Info.addFact(Pred, A, B, CB.NumIn, CB.NumOut, DFSInStack);
868       Info.transferToOtherSystem(Pred, A, B, CB.NumIn, CB.NumOut, DFSInStack);
869     }
870   }
871 
872 #ifndef NDEBUG
873   unsigned SignedEntries =
874       count_if(DFSInStack, [](const StackEntry &E) { return E.IsSigned; });
875   assert(Info.getCS(false).size() == DFSInStack.size() - SignedEntries &&
876          "updates to CS and DFSInStack are out of sync");
877   assert(Info.getCS(true).size() == SignedEntries &&
878          "updates to CS and DFSInStack are out of sync");
879 #endif
880 
881   for (Instruction *I : ToRemove)
882     I->eraseFromParent();
883   return Changed;
884 }
885 
886 PreservedAnalyses ConstraintEliminationPass::run(Function &F,
887                                                  FunctionAnalysisManager &AM) {
888   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
889   if (!eliminateConstraints(F, DT))
890     return PreservedAnalyses::all();
891 
892   PreservedAnalyses PA;
893   PA.preserve<DominatorTreeAnalysis>();
894   PA.preserveSet<CFGAnalyses>();
895   return PA;
896 }
897 
898 namespace {
899 
900 class ConstraintElimination : public FunctionPass {
901 public:
902   static char ID;
903 
904   ConstraintElimination() : FunctionPass(ID) {
905     initializeConstraintEliminationPass(*PassRegistry::getPassRegistry());
906   }
907 
908   bool runOnFunction(Function &F) override {
909     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
910     return eliminateConstraints(F, DT);
911   }
912 
913   void getAnalysisUsage(AnalysisUsage &AU) const override {
914     AU.setPreservesCFG();
915     AU.addRequired<DominatorTreeWrapperPass>();
916     AU.addPreserved<GlobalsAAWrapperPass>();
917     AU.addPreserved<DominatorTreeWrapperPass>();
918   }
919 };
920 
921 } // end anonymous namespace
922 
923 char ConstraintElimination::ID = 0;
924 
925 INITIALIZE_PASS_BEGIN(ConstraintElimination, "constraint-elimination",
926                       "Constraint Elimination", false, false)
927 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
928 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
929 INITIALIZE_PASS_END(ConstraintElimination, "constraint-elimination",
930                     "Constraint Elimination", false, false)
931 
932 FunctionPass *llvm::createConstraintEliminationPass() {
933   return new ConstraintElimination();
934 }
935