xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/Reassociate.cpp (revision 734e82fe33aa764367791a7d603b383996c6b40b)
1 //===- Reassociate.cpp - Reassociate binary expressions -------------------===//
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 reassociates commutative expressions in an order that is designed
10 // to promote better constant propagation, GCSE, LICM, PRE, etc.
11 //
12 // For example: 4 + (x + 5) -> x + (4 + 5)
13 //
14 // In the implementation of this algorithm, constants are assigned rank = 0,
15 // function arguments are rank = 1, and other values are assigned ranks
16 // corresponding to the reverse post order traversal of current function
17 // (starting at 2), which effectively gives values in deep loops higher rank
18 // than values not in loops.
19 //
20 //===----------------------------------------------------------------------===//
21 
22 #include "llvm/Transforms/Scalar/Reassociate.h"
23 #include "llvm/ADT/APFloat.h"
24 #include "llvm/ADT/APInt.h"
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/PostOrderIterator.h"
27 #include "llvm/ADT/SmallPtrSet.h"
28 #include "llvm/ADT/SmallSet.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/BasicAliasAnalysis.h"
32 #include "llvm/Analysis/ConstantFolding.h"
33 #include "llvm/Analysis/GlobalsModRef.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/Argument.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/CFG.h"
38 #include "llvm/IR/Constant.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/InstrTypes.h"
43 #include "llvm/IR/Instruction.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/Operator.h"
46 #include "llvm/IR/PassManager.h"
47 #include "llvm/IR/PatternMatch.h"
48 #include "llvm/IR/Type.h"
49 #include "llvm/IR/User.h"
50 #include "llvm/IR/Value.h"
51 #include "llvm/IR/ValueHandle.h"
52 #include "llvm/InitializePasses.h"
53 #include "llvm/Pass.h"
54 #include "llvm/Support/Casting.h"
55 #include "llvm/Support/Debug.h"
56 #include "llvm/Support/raw_ostream.h"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Transforms/Utils/Local.h"
59 #include <algorithm>
60 #include <cassert>
61 #include <utility>
62 
63 using namespace llvm;
64 using namespace reassociate;
65 using namespace PatternMatch;
66 
67 #define DEBUG_TYPE "reassociate"
68 
69 STATISTIC(NumChanged, "Number of insts reassociated");
70 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
71 STATISTIC(NumFactor , "Number of multiplies factored");
72 
73 #ifndef NDEBUG
74 /// Print out the expression identified in the Ops list.
75 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
76   Module *M = I->getModule();
77   dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
78        << *Ops[0].Op->getType() << '\t';
79   for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
80     dbgs() << "[ ";
81     Ops[i].Op->printAsOperand(dbgs(), false, M);
82     dbgs() << ", #" << Ops[i].Rank << "] ";
83   }
84 }
85 #endif
86 
87 /// Utility class representing a non-constant Xor-operand. We classify
88 /// non-constant Xor-Operands into two categories:
89 ///  C1) The operand is in the form "X & C", where C is a constant and C != ~0
90 ///  C2)
91 ///    C2.1) The operand is in the form of "X | C", where C is a non-zero
92 ///          constant.
93 ///    C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
94 ///          operand as "E | 0"
95 class llvm::reassociate::XorOpnd {
96 public:
97   XorOpnd(Value *V);
98 
99   bool isInvalid() const { return SymbolicPart == nullptr; }
100   bool isOrExpr() const { return isOr; }
101   Value *getValue() const { return OrigVal; }
102   Value *getSymbolicPart() const { return SymbolicPart; }
103   unsigned getSymbolicRank() const { return SymbolicRank; }
104   const APInt &getConstPart() const { return ConstPart; }
105 
106   void Invalidate() { SymbolicPart = OrigVal = nullptr; }
107   void setSymbolicRank(unsigned R) { SymbolicRank = R; }
108 
109 private:
110   Value *OrigVal;
111   Value *SymbolicPart;
112   APInt ConstPart;
113   unsigned SymbolicRank;
114   bool isOr;
115 };
116 
117 XorOpnd::XorOpnd(Value *V) {
118   assert(!isa<ConstantInt>(V) && "No ConstantInt");
119   OrigVal = V;
120   Instruction *I = dyn_cast<Instruction>(V);
121   SymbolicRank = 0;
122 
123   if (I && (I->getOpcode() == Instruction::Or ||
124             I->getOpcode() == Instruction::And)) {
125     Value *V0 = I->getOperand(0);
126     Value *V1 = I->getOperand(1);
127     const APInt *C;
128     if (match(V0, m_APInt(C)))
129       std::swap(V0, V1);
130 
131     if (match(V1, m_APInt(C))) {
132       ConstPart = *C;
133       SymbolicPart = V0;
134       isOr = (I->getOpcode() == Instruction::Or);
135       return;
136     }
137   }
138 
139   // view the operand as "V | 0"
140   SymbolicPart = V;
141   ConstPart = APInt::getZero(V->getType()->getScalarSizeInBits());
142   isOr = true;
143 }
144 
145 /// Return true if I is an instruction with the FastMathFlags that are needed
146 /// for general reassociation set.  This is not the same as testing
147 /// Instruction::isAssociative() because it includes operations like fsub.
148 /// (This routine is only intended to be called for floating-point operations.)
149 static bool hasFPAssociativeFlags(Instruction *I) {
150   assert(I && isa<FPMathOperator>(I) && "Should only check FP ops");
151   return I->hasAllowReassoc() && I->hasNoSignedZeros();
152 }
153 
154 /// Return true if V is an instruction of the specified opcode and if it
155 /// only has one use.
156 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
157   auto *BO = dyn_cast<BinaryOperator>(V);
158   if (BO && BO->hasOneUse() && BO->getOpcode() == Opcode)
159     if (!isa<FPMathOperator>(BO) || hasFPAssociativeFlags(BO))
160       return BO;
161   return nullptr;
162 }
163 
164 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
165                                         unsigned Opcode2) {
166   auto *BO = dyn_cast<BinaryOperator>(V);
167   if (BO && BO->hasOneUse() &&
168       (BO->getOpcode() == Opcode1 || BO->getOpcode() == Opcode2))
169     if (!isa<FPMathOperator>(BO) || hasFPAssociativeFlags(BO))
170       return BO;
171   return nullptr;
172 }
173 
174 void ReassociatePass::BuildRankMap(Function &F,
175                                    ReversePostOrderTraversal<Function*> &RPOT) {
176   unsigned Rank = 2;
177 
178   // Assign distinct ranks to function arguments.
179   for (auto &Arg : F.args()) {
180     ValueRankMap[&Arg] = ++Rank;
181     LLVM_DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank
182                       << "\n");
183   }
184 
185   // Traverse basic blocks in ReversePostOrder.
186   for (BasicBlock *BB : RPOT) {
187     unsigned BBRank = RankMap[BB] = ++Rank << 16;
188 
189     // Walk the basic block, adding precomputed ranks for any instructions that
190     // we cannot move.  This ensures that the ranks for these instructions are
191     // all different in the block.
192     for (Instruction &I : *BB)
193       if (mayHaveNonDefUseDependency(I))
194         ValueRankMap[&I] = ++BBRank;
195   }
196 }
197 
198 unsigned ReassociatePass::getRank(Value *V) {
199   Instruction *I = dyn_cast<Instruction>(V);
200   if (!I) {
201     if (isa<Argument>(V)) return ValueRankMap[V];   // Function argument.
202     return 0;  // Otherwise it's a global or constant, rank 0.
203   }
204 
205   if (unsigned Rank = ValueRankMap[I])
206     return Rank;    // Rank already known?
207 
208   // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
209   // we can reassociate expressions for code motion!  Since we do not recurse
210   // for PHI nodes, we cannot have infinite recursion here, because there
211   // cannot be loops in the value graph that do not go through PHI nodes.
212   unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
213   for (unsigned i = 0, e = I->getNumOperands(); i != e && Rank != MaxRank; ++i)
214     Rank = std::max(Rank, getRank(I->getOperand(i)));
215 
216   // If this is a 'not' or 'neg' instruction, do not count it for rank. This
217   // assures us that X and ~X will have the same rank.
218   if (!match(I, m_Not(m_Value())) && !match(I, m_Neg(m_Value())) &&
219       !match(I, m_FNeg(m_Value())))
220     ++Rank;
221 
222   LLVM_DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank
223                     << "\n");
224 
225   return ValueRankMap[I] = Rank;
226 }
227 
228 // Canonicalize constants to RHS.  Otherwise, sort the operands by rank.
229 void ReassociatePass::canonicalizeOperands(Instruction *I) {
230   assert(isa<BinaryOperator>(I) && "Expected binary operator.");
231   assert(I->isCommutative() && "Expected commutative operator.");
232 
233   Value *LHS = I->getOperand(0);
234   Value *RHS = I->getOperand(1);
235   if (LHS == RHS || isa<Constant>(RHS))
236     return;
237   if (isa<Constant>(LHS) || getRank(RHS) < getRank(LHS))
238     cast<BinaryOperator>(I)->swapOperands();
239 }
240 
241 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
242                                  Instruction *InsertBefore, Value *FlagsOp) {
243   if (S1->getType()->isIntOrIntVectorTy())
244     return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
245   else {
246     BinaryOperator *Res =
247         BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
248     Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
249     return Res;
250   }
251 }
252 
253 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
254                                  Instruction *InsertBefore, Value *FlagsOp) {
255   if (S1->getType()->isIntOrIntVectorTy())
256     return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
257   else {
258     BinaryOperator *Res =
259       BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
260     Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
261     return Res;
262   }
263 }
264 
265 static Instruction *CreateNeg(Value *S1, const Twine &Name,
266                               Instruction *InsertBefore, Value *FlagsOp) {
267   if (S1->getType()->isIntOrIntVectorTy())
268     return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
269 
270   if (auto *FMFSource = dyn_cast<Instruction>(FlagsOp))
271     return UnaryOperator::CreateFNegFMF(S1, FMFSource, Name, InsertBefore);
272 
273   return UnaryOperator::CreateFNeg(S1, Name, InsertBefore);
274 }
275 
276 /// Replace 0-X with X*-1.
277 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
278   assert((isa<UnaryOperator>(Neg) || isa<BinaryOperator>(Neg)) &&
279          "Expected a Negate!");
280   // FIXME: It's not safe to lower a unary FNeg into a FMul by -1.0.
281   unsigned OpNo = isa<BinaryOperator>(Neg) ? 1 : 0;
282   Type *Ty = Neg->getType();
283   Constant *NegOne = Ty->isIntOrIntVectorTy() ?
284     ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
285 
286   BinaryOperator *Res = CreateMul(Neg->getOperand(OpNo), NegOne, "", Neg, Neg);
287   Neg->setOperand(OpNo, Constant::getNullValue(Ty)); // Drop use of op.
288   Res->takeName(Neg);
289   Neg->replaceAllUsesWith(Res);
290   Res->setDebugLoc(Neg->getDebugLoc());
291   return Res;
292 }
293 
294 /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
295 /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for
296 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
297 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
298 /// even x in Bitwidth-bit arithmetic.
299 static unsigned CarmichaelShift(unsigned Bitwidth) {
300   if (Bitwidth < 3)
301     return Bitwidth - 1;
302   return Bitwidth - 2;
303 }
304 
305 /// Add the extra weight 'RHS' to the existing weight 'LHS',
306 /// reducing the combined weight using any special properties of the operation.
307 /// The existing weight LHS represents the computation X op X op ... op X where
308 /// X occurs LHS times.  The combined weight represents  X op X op ... op X with
309 /// X occurring LHS + RHS times.  If op is "Xor" for example then the combined
310 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
311 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
312 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
313   // If we were working with infinite precision arithmetic then the combined
314   // weight would be LHS + RHS.  But we are using finite precision arithmetic,
315   // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
316   // for nilpotent operations and addition, but not for idempotent operations
317   // and multiplication), so it is important to correctly reduce the combined
318   // weight back into range if wrapping would be wrong.
319 
320   // If RHS is zero then the weight didn't change.
321   if (RHS.isMinValue())
322     return;
323   // If LHS is zero then the combined weight is RHS.
324   if (LHS.isMinValue()) {
325     LHS = RHS;
326     return;
327   }
328   // From this point on we know that neither LHS nor RHS is zero.
329 
330   if (Instruction::isIdempotent(Opcode)) {
331     // Idempotent means X op X === X, so any non-zero weight is equivalent to a
332     // weight of 1.  Keeping weights at zero or one also means that wrapping is
333     // not a problem.
334     assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
335     return; // Return a weight of 1.
336   }
337   if (Instruction::isNilpotent(Opcode)) {
338     // Nilpotent means X op X === 0, so reduce weights modulo 2.
339     assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
340     LHS = 0; // 1 + 1 === 0 modulo 2.
341     return;
342   }
343   if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
344     // TODO: Reduce the weight by exploiting nsw/nuw?
345     LHS += RHS;
346     return;
347   }
348 
349   assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
350          "Unknown associative operation!");
351   unsigned Bitwidth = LHS.getBitWidth();
352   // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
353   // can be replaced with W-CM.  That's because x^W=x^(W-CM) for every Bitwidth
354   // bit number x, since either x is odd in which case x^CM = 1, or x is even in
355   // which case both x^W and x^(W - CM) are zero.  By subtracting off multiples
356   // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
357   // which by a happy accident means that they can always be represented using
358   // Bitwidth bits.
359   // TODO: Reduce the weight by exploiting nsw/nuw?  (Could do much better than
360   // the Carmichael number).
361   if (Bitwidth > 3) {
362     /// CM - The value of Carmichael's lambda function.
363     APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
364     // Any weight W >= Threshold can be replaced with W - CM.
365     APInt Threshold = CM + Bitwidth;
366     assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
367     // For Bitwidth 4 or more the following sum does not overflow.
368     LHS += RHS;
369     while (LHS.uge(Threshold))
370       LHS -= CM;
371   } else {
372     // To avoid problems with overflow do everything the same as above but using
373     // a larger type.
374     unsigned CM = 1U << CarmichaelShift(Bitwidth);
375     unsigned Threshold = CM + Bitwidth;
376     assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
377            "Weights not reduced!");
378     unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
379     while (Total >= Threshold)
380       Total -= CM;
381     LHS = Total;
382   }
383 }
384 
385 using RepeatedValue = std::pair<Value*, APInt>;
386 
387 /// Given an associative binary expression, return the leaf
388 /// nodes in Ops along with their weights (how many times the leaf occurs).  The
389 /// original expression is the same as
390 ///   (Ops[0].first op Ops[0].first op ... Ops[0].first)  <- Ops[0].second times
391 /// op
392 ///   (Ops[1].first op Ops[1].first op ... Ops[1].first)  <- Ops[1].second times
393 /// op
394 ///   ...
395 /// op
396 ///   (Ops[N].first op Ops[N].first op ... Ops[N].first)  <- Ops[N].second times
397 ///
398 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
399 ///
400 /// This routine may modify the function, in which case it returns 'true'.  The
401 /// changes it makes may well be destructive, changing the value computed by 'I'
402 /// to something completely different.  Thus if the routine returns 'true' then
403 /// you MUST either replace I with a new expression computed from the Ops array,
404 /// or use RewriteExprTree to put the values back in.
405 ///
406 /// A leaf node is either not a binary operation of the same kind as the root
407 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
408 /// opcode), or is the same kind of binary operator but has a use which either
409 /// does not belong to the expression, or does belong to the expression but is
410 /// a leaf node.  Every leaf node has at least one use that is a non-leaf node
411 /// of the expression, while for non-leaf nodes (except for the root 'I') every
412 /// use is a non-leaf node of the expression.
413 ///
414 /// For example:
415 ///           expression graph        node names
416 ///
417 ///                     +        |        I
418 ///                    / \       |
419 ///                   +   +      |      A,  B
420 ///                  / \ / \     |
421 ///                 *   +   *    |    C,  D,  E
422 ///                / \ / \ / \   |
423 ///                   +   *      |      F,  G
424 ///
425 /// The leaf nodes are C, E, F and G.  The Ops array will contain (maybe not in
426 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
427 ///
428 /// The expression is maximal: if some instruction is a binary operator of the
429 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
430 /// then the instruction also belongs to the expression, is not a leaf node of
431 /// it, and its operands also belong to the expression (but may be leaf nodes).
432 ///
433 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
434 /// order to ensure that every non-root node in the expression has *exactly one*
435 /// use by a non-leaf node of the expression.  This destruction means that the
436 /// caller MUST either replace 'I' with a new expression or use something like
437 /// RewriteExprTree to put the values back in if the routine indicates that it
438 /// made a change by returning 'true'.
439 ///
440 /// In the above example either the right operand of A or the left operand of B
441 /// will be replaced by undef.  If it is B's operand then this gives:
442 ///
443 ///                     +        |        I
444 ///                    / \       |
445 ///                   +   +      |      A,  B - operand of B replaced with undef
446 ///                  / \   \     |
447 ///                 *   +   *    |    C,  D,  E
448 ///                / \ / \ / \   |
449 ///                   +   *      |      F,  G
450 ///
451 /// Note that such undef operands can only be reached by passing through 'I'.
452 /// For example, if you visit operands recursively starting from a leaf node
453 /// then you will never see such an undef operand unless you get back to 'I',
454 /// which requires passing through a phi node.
455 ///
456 /// Note that this routine may also mutate binary operators of the wrong type
457 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
458 /// of the expression) if it can turn them into binary operators of the right
459 /// type and thus make the expression bigger.
460 static bool LinearizeExprTree(Instruction *I,
461                               SmallVectorImpl<RepeatedValue> &Ops,
462                               ReassociatePass::OrderedSet &ToRedo) {
463   assert((isa<UnaryOperator>(I) || isa<BinaryOperator>(I)) &&
464          "Expected a UnaryOperator or BinaryOperator!");
465   LLVM_DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
466   unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
467   unsigned Opcode = I->getOpcode();
468   assert(I->isAssociative() && I->isCommutative() &&
469          "Expected an associative and commutative operation!");
470 
471   // Visit all operands of the expression, keeping track of their weight (the
472   // number of paths from the expression root to the operand, or if you like
473   // the number of times that operand occurs in the linearized expression).
474   // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
475   // while A has weight two.
476 
477   // Worklist of non-leaf nodes (their operands are in the expression too) along
478   // with their weights, representing a certain number of paths to the operator.
479   // If an operator occurs in the worklist multiple times then we found multiple
480   // ways to get to it.
481   SmallVector<std::pair<Instruction*, APInt>, 8> Worklist; // (Op, Weight)
482   Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
483   bool Changed = false;
484 
485   // Leaves of the expression are values that either aren't the right kind of
486   // operation (eg: a constant, or a multiply in an add tree), or are, but have
487   // some uses that are not inside the expression.  For example, in I = X + X,
488   // X = A + B, the value X has two uses (by I) that are in the expression.  If
489   // X has any other uses, for example in a return instruction, then we consider
490   // X to be a leaf, and won't analyze it further.  When we first visit a value,
491   // if it has more than one use then at first we conservatively consider it to
492   // be a leaf.  Later, as the expression is explored, we may discover some more
493   // uses of the value from inside the expression.  If all uses turn out to be
494   // from within the expression (and the value is a binary operator of the right
495   // kind) then the value is no longer considered to be a leaf, and its operands
496   // are explored.
497 
498   // Leaves - Keeps track of the set of putative leaves as well as the number of
499   // paths to each leaf seen so far.
500   using LeafMap = DenseMap<Value *, APInt>;
501   LeafMap Leaves; // Leaf -> Total weight so far.
502   SmallVector<Value *, 8> LeafOrder; // Ensure deterministic leaf output order.
503 
504 #ifndef NDEBUG
505   SmallPtrSet<Value *, 8> Visited; // For checking the iteration scheme.
506 #endif
507   while (!Worklist.empty()) {
508     std::pair<Instruction*, APInt> P = Worklist.pop_back_val();
509     I = P.first; // We examine the operands of this binary operator.
510 
511     for (unsigned OpIdx = 0; OpIdx < I->getNumOperands(); ++OpIdx) { // Visit operands.
512       Value *Op = I->getOperand(OpIdx);
513       APInt Weight = P.second; // Number of paths to this operand.
514       LLVM_DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
515       assert(!Op->use_empty() && "No uses, so how did we get to it?!");
516 
517       // If this is a binary operation of the right kind with only one use then
518       // add its operands to the expression.
519       if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
520         assert(Visited.insert(Op).second && "Not first visit!");
521         LLVM_DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
522         Worklist.push_back(std::make_pair(BO, Weight));
523         continue;
524       }
525 
526       // Appears to be a leaf.  Is the operand already in the set of leaves?
527       LeafMap::iterator It = Leaves.find(Op);
528       if (It == Leaves.end()) {
529         // Not in the leaf map.  Must be the first time we saw this operand.
530         assert(Visited.insert(Op).second && "Not first visit!");
531         if (!Op->hasOneUse()) {
532           // This value has uses not accounted for by the expression, so it is
533           // not safe to modify.  Mark it as being a leaf.
534           LLVM_DEBUG(dbgs()
535                      << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
536           LeafOrder.push_back(Op);
537           Leaves[Op] = Weight;
538           continue;
539         }
540         // No uses outside the expression, try morphing it.
541       } else {
542         // Already in the leaf map.
543         assert(It != Leaves.end() && Visited.count(Op) &&
544                "In leaf map but not visited!");
545 
546         // Update the number of paths to the leaf.
547         IncorporateWeight(It->second, Weight, Opcode);
548 
549 #if 0   // TODO: Re-enable once PR13021 is fixed.
550         // The leaf already has one use from inside the expression.  As we want
551         // exactly one such use, drop this new use of the leaf.
552         assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
553         I->setOperand(OpIdx, UndefValue::get(I->getType()));
554         Changed = true;
555 
556         // If the leaf is a binary operation of the right kind and we now see
557         // that its multiple original uses were in fact all by nodes belonging
558         // to the expression, then no longer consider it to be a leaf and add
559         // its operands to the expression.
560         if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
561           LLVM_DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
562           Worklist.push_back(std::make_pair(BO, It->second));
563           Leaves.erase(It);
564           continue;
565         }
566 #endif
567 
568         // If we still have uses that are not accounted for by the expression
569         // then it is not safe to modify the value.
570         if (!Op->hasOneUse())
571           continue;
572 
573         // No uses outside the expression, try morphing it.
574         Weight = It->second;
575         Leaves.erase(It); // Since the value may be morphed below.
576       }
577 
578       // At this point we have a value which, first of all, is not a binary
579       // expression of the right kind, and secondly, is only used inside the
580       // expression.  This means that it can safely be modified.  See if we
581       // can usefully morph it into an expression of the right kind.
582       assert((!isa<Instruction>(Op) ||
583               cast<Instruction>(Op)->getOpcode() != Opcode
584               || (isa<FPMathOperator>(Op) &&
585                   !hasFPAssociativeFlags(cast<Instruction>(Op)))) &&
586              "Should have been handled above!");
587       assert(Op->hasOneUse() && "Has uses outside the expression tree!");
588 
589       // If this is a multiply expression, turn any internal negations into
590       // multiplies by -1 so they can be reassociated.  Add any users of the
591       // newly created multiplication by -1 to the redo list, so any
592       // reassociation opportunities that are exposed will be reassociated
593       // further.
594       Instruction *Neg;
595       if (((Opcode == Instruction::Mul && match(Op, m_Neg(m_Value()))) ||
596            (Opcode == Instruction::FMul && match(Op, m_FNeg(m_Value())))) &&
597            match(Op, m_Instruction(Neg))) {
598         LLVM_DEBUG(dbgs()
599                    << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
600         Instruction *Mul = LowerNegateToMultiply(Neg);
601         LLVM_DEBUG(dbgs() << *Mul << '\n');
602         Worklist.push_back(std::make_pair(Mul, Weight));
603         for (User *U : Mul->users()) {
604           if (BinaryOperator *UserBO = dyn_cast<BinaryOperator>(U))
605             ToRedo.insert(UserBO);
606         }
607         ToRedo.insert(Neg);
608         Changed = true;
609         continue;
610       }
611 
612       // Failed to morph into an expression of the right type.  This really is
613       // a leaf.
614       LLVM_DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
615       assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
616       LeafOrder.push_back(Op);
617       Leaves[Op] = Weight;
618     }
619   }
620 
621   // The leaves, repeated according to their weights, represent the linearized
622   // form of the expression.
623   for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
624     Value *V = LeafOrder[i];
625     LeafMap::iterator It = Leaves.find(V);
626     if (It == Leaves.end())
627       // Node initially thought to be a leaf wasn't.
628       continue;
629     assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
630     APInt Weight = It->second;
631     if (Weight.isMinValue())
632       // Leaf already output or weight reduction eliminated it.
633       continue;
634     // Ensure the leaf is only output once.
635     It->second = 0;
636     Ops.push_back(std::make_pair(V, Weight));
637   }
638 
639   // For nilpotent operations or addition there may be no operands, for example
640   // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
641   // in both cases the weight reduces to 0 causing the value to be skipped.
642   if (Ops.empty()) {
643     Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
644     assert(Identity && "Associative operation without identity!");
645     Ops.emplace_back(Identity, APInt(Bitwidth, 1));
646   }
647 
648   return Changed;
649 }
650 
651 /// Now that the operands for this expression tree are
652 /// linearized and optimized, emit them in-order.
653 void ReassociatePass::RewriteExprTree(BinaryOperator *I,
654                                       SmallVectorImpl<ValueEntry> &Ops) {
655   assert(Ops.size() > 1 && "Single values should be used directly!");
656 
657   // Since our optimizations should never increase the number of operations, the
658   // new expression can usually be written reusing the existing binary operators
659   // from the original expression tree, without creating any new instructions,
660   // though the rewritten expression may have a completely different topology.
661   // We take care to not change anything if the new expression will be the same
662   // as the original.  If more than trivial changes (like commuting operands)
663   // were made then we are obliged to clear out any optional subclass data like
664   // nsw flags.
665 
666   /// NodesToRewrite - Nodes from the original expression available for writing
667   /// the new expression into.
668   SmallVector<BinaryOperator*, 8> NodesToRewrite;
669   unsigned Opcode = I->getOpcode();
670   BinaryOperator *Op = I;
671 
672   /// NotRewritable - The operands being written will be the leaves of the new
673   /// expression and must not be used as inner nodes (via NodesToRewrite) by
674   /// mistake.  Inner nodes are always reassociable, and usually leaves are not
675   /// (if they were they would have been incorporated into the expression and so
676   /// would not be leaves), so most of the time there is no danger of this.  But
677   /// in rare cases a leaf may become reassociable if an optimization kills uses
678   /// of it, or it may momentarily become reassociable during rewriting (below)
679   /// due it being removed as an operand of one of its uses.  Ensure that misuse
680   /// of leaf nodes as inner nodes cannot occur by remembering all of the future
681   /// leaves and refusing to reuse any of them as inner nodes.
682   SmallPtrSet<Value*, 8> NotRewritable;
683   for (unsigned i = 0, e = Ops.size(); i != e; ++i)
684     NotRewritable.insert(Ops[i].Op);
685 
686   // ExpressionChanged - Non-null if the rewritten expression differs from the
687   // original in some non-trivial way, requiring the clearing of optional flags.
688   // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
689   BinaryOperator *ExpressionChanged = nullptr;
690   for (unsigned i = 0; ; ++i) {
691     // The last operation (which comes earliest in the IR) is special as both
692     // operands will come from Ops, rather than just one with the other being
693     // a subexpression.
694     if (i+2 == Ops.size()) {
695       Value *NewLHS = Ops[i].Op;
696       Value *NewRHS = Ops[i+1].Op;
697       Value *OldLHS = Op->getOperand(0);
698       Value *OldRHS = Op->getOperand(1);
699 
700       if (NewLHS == OldLHS && NewRHS == OldRHS)
701         // Nothing changed, leave it alone.
702         break;
703 
704       if (NewLHS == OldRHS && NewRHS == OldLHS) {
705         // The order of the operands was reversed.  Swap them.
706         LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
707         Op->swapOperands();
708         LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
709         MadeChange = true;
710         ++NumChanged;
711         break;
712       }
713 
714       // The new operation differs non-trivially from the original. Overwrite
715       // the old operands with the new ones.
716       LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
717       if (NewLHS != OldLHS) {
718         BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
719         if (BO && !NotRewritable.count(BO))
720           NodesToRewrite.push_back(BO);
721         Op->setOperand(0, NewLHS);
722       }
723       if (NewRHS != OldRHS) {
724         BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
725         if (BO && !NotRewritable.count(BO))
726           NodesToRewrite.push_back(BO);
727         Op->setOperand(1, NewRHS);
728       }
729       LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
730 
731       ExpressionChanged = Op;
732       MadeChange = true;
733       ++NumChanged;
734 
735       break;
736     }
737 
738     // Not the last operation.  The left-hand side will be a sub-expression
739     // while the right-hand side will be the current element of Ops.
740     Value *NewRHS = Ops[i].Op;
741     if (NewRHS != Op->getOperand(1)) {
742       LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
743       if (NewRHS == Op->getOperand(0)) {
744         // The new right-hand side was already present as the left operand.  If
745         // we are lucky then swapping the operands will sort out both of them.
746         Op->swapOperands();
747       } else {
748         // Overwrite with the new right-hand side.
749         BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
750         if (BO && !NotRewritable.count(BO))
751           NodesToRewrite.push_back(BO);
752         Op->setOperand(1, NewRHS);
753         ExpressionChanged = Op;
754       }
755       LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
756       MadeChange = true;
757       ++NumChanged;
758     }
759 
760     // Now deal with the left-hand side.  If this is already an operation node
761     // from the original expression then just rewrite the rest of the expression
762     // into it.
763     BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
764     if (BO && !NotRewritable.count(BO)) {
765       Op = BO;
766       continue;
767     }
768 
769     // Otherwise, grab a spare node from the original expression and use that as
770     // the left-hand side.  If there are no nodes left then the optimizers made
771     // an expression with more nodes than the original!  This usually means that
772     // they did something stupid but it might mean that the problem was just too
773     // hard (finding the mimimal number of multiplications needed to realize a
774     // multiplication expression is NP-complete).  Whatever the reason, smart or
775     // stupid, create a new node if there are none left.
776     BinaryOperator *NewOp;
777     if (NodesToRewrite.empty()) {
778       Constant *Undef = UndefValue::get(I->getType());
779       NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
780                                      Undef, Undef, "", I);
781       if (isa<FPMathOperator>(NewOp))
782         NewOp->setFastMathFlags(I->getFastMathFlags());
783     } else {
784       NewOp = NodesToRewrite.pop_back_val();
785     }
786 
787     LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
788     Op->setOperand(0, NewOp);
789     LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
790     ExpressionChanged = Op;
791     MadeChange = true;
792     ++NumChanged;
793     Op = NewOp;
794   }
795 
796   // If the expression changed non-trivially then clear out all subclass data
797   // starting from the operator specified in ExpressionChanged, and compactify
798   // the operators to just before the expression root to guarantee that the
799   // expression tree is dominated by all of Ops.
800   if (ExpressionChanged)
801     do {
802       // Preserve FastMathFlags.
803       if (isa<FPMathOperator>(I)) {
804         FastMathFlags Flags = I->getFastMathFlags();
805         ExpressionChanged->clearSubclassOptionalData();
806         ExpressionChanged->setFastMathFlags(Flags);
807       } else
808         ExpressionChanged->clearSubclassOptionalData();
809 
810       if (ExpressionChanged == I)
811         break;
812 
813       // Discard any debug info related to the expressions that has changed (we
814       // can leave debug infor related to the root, since the result of the
815       // expression tree should be the same even after reassociation).
816       replaceDbgUsesWithUndef(ExpressionChanged);
817 
818       ExpressionChanged->moveBefore(I);
819       ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
820     } while (true);
821 
822   // Throw away any left over nodes from the original expression.
823   for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
824     RedoInsts.insert(NodesToRewrite[i]);
825 }
826 
827 /// Insert instructions before the instruction pointed to by BI,
828 /// that computes the negative version of the value specified.  The negative
829 /// version of the value is returned, and BI is left pointing at the instruction
830 /// that should be processed next by the reassociation pass.
831 /// Also add intermediate instructions to the redo list that are modified while
832 /// pushing the negates through adds.  These will be revisited to see if
833 /// additional opportunities have been exposed.
834 static Value *NegateValue(Value *V, Instruction *BI,
835                           ReassociatePass::OrderedSet &ToRedo) {
836   if (auto *C = dyn_cast<Constant>(V)) {
837     const DataLayout &DL = BI->getModule()->getDataLayout();
838     Constant *Res = C->getType()->isFPOrFPVectorTy()
839                         ? ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)
840                         : ConstantExpr::getNeg(C);
841     if (Res)
842       return Res;
843   }
844 
845   // We are trying to expose opportunity for reassociation.  One of the things
846   // that we want to do to achieve this is to push a negation as deep into an
847   // expression chain as possible, to expose the add instructions.  In practice,
848   // this means that we turn this:
849   //   X = -(A+12+C+D)   into    X = -A + -12 + -C + -D = -12 + -A + -C + -D
850   // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
851   // the constants.  We assume that instcombine will clean up the mess later if
852   // we introduce tons of unnecessary negation instructions.
853   //
854   if (BinaryOperator *I =
855           isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
856     // Push the negates through the add.
857     I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo));
858     I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo));
859     if (I->getOpcode() == Instruction::Add) {
860       I->setHasNoUnsignedWrap(false);
861       I->setHasNoSignedWrap(false);
862     }
863 
864     // We must move the add instruction here, because the neg instructions do
865     // not dominate the old add instruction in general.  By moving it, we are
866     // assured that the neg instructions we just inserted dominate the
867     // instruction we are about to insert after them.
868     //
869     I->moveBefore(BI);
870     I->setName(I->getName()+".neg");
871 
872     // Add the intermediate negates to the redo list as processing them later
873     // could expose more reassociating opportunities.
874     ToRedo.insert(I);
875     return I;
876   }
877 
878   // Okay, we need to materialize a negated version of V with an instruction.
879   // Scan the use lists of V to see if we have one already.
880   for (User *U : V->users()) {
881     if (!match(U, m_Neg(m_Value())) && !match(U, m_FNeg(m_Value())))
882       continue;
883 
884     // We found one!  Now we have to make sure that the definition dominates
885     // this use.  We do this by moving it to the entry block (if it is a
886     // non-instruction value) or right after the definition.  These negates will
887     // be zapped by reassociate later, so we don't need much finesse here.
888     Instruction *TheNeg = dyn_cast<Instruction>(U);
889 
890     // We can't safely propagate a vector zero constant with poison/undef lanes.
891     Constant *C;
892     if (match(TheNeg, m_BinOp(m_Constant(C), m_Value())) &&
893         C->containsUndefOrPoisonElement())
894       continue;
895 
896     // Verify that the negate is in this function, V might be a constant expr.
897     if (!TheNeg ||
898         TheNeg->getParent()->getParent() != BI->getParent()->getParent())
899       continue;
900 
901     Instruction *InsertPt;
902     if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
903       InsertPt = InstInput->getInsertionPointAfterDef();
904       if (!InsertPt)
905         continue;
906     } else {
907       InsertPt = &*TheNeg->getFunction()->getEntryBlock().begin();
908     }
909 
910     TheNeg->moveBefore(InsertPt);
911     if (TheNeg->getOpcode() == Instruction::Sub) {
912       TheNeg->setHasNoUnsignedWrap(false);
913       TheNeg->setHasNoSignedWrap(false);
914     } else {
915       TheNeg->andIRFlags(BI);
916     }
917     ToRedo.insert(TheNeg);
918     return TheNeg;
919   }
920 
921   // Insert a 'neg' instruction that subtracts the value from zero to get the
922   // negation.
923   Instruction *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI);
924   ToRedo.insert(NewNeg);
925   return NewNeg;
926 }
927 
928 // See if this `or` looks like an load widening reduction, i.e. that it
929 // consists of an `or`/`shl`/`zext`/`load` nodes only. Note that we don't
930 // ensure that the pattern is *really* a load widening reduction,
931 // we do not ensure that it can really be replaced with a widened load,
932 // only that it mostly looks like one.
933 static bool isLoadCombineCandidate(Instruction *Or) {
934   SmallVector<Instruction *, 8> Worklist;
935   SmallSet<Instruction *, 8> Visited;
936 
937   auto Enqueue = [&](Value *V) {
938     auto *I = dyn_cast<Instruction>(V);
939     // Each node of an `or` reduction must be an instruction,
940     if (!I)
941       return false; // Node is certainly not part of an `or` load reduction.
942     // Only process instructions we have never processed before.
943     if (Visited.insert(I).second)
944       Worklist.emplace_back(I);
945     return true; // Will need to look at parent nodes.
946   };
947 
948   if (!Enqueue(Or))
949     return false; // Not an `or` reduction pattern.
950 
951   while (!Worklist.empty()) {
952     auto *I = Worklist.pop_back_val();
953 
954     // Okay, which instruction is this node?
955     switch (I->getOpcode()) {
956     case Instruction::Or:
957       // Got an `or` node. That's fine, just recurse into it's operands.
958       for (Value *Op : I->operands())
959         if (!Enqueue(Op))
960           return false; // Not an `or` reduction pattern.
961       continue;
962 
963     case Instruction::Shl:
964     case Instruction::ZExt:
965       // `shl`/`zext` nodes are fine, just recurse into their base operand.
966       if (!Enqueue(I->getOperand(0)))
967         return false; // Not an `or` reduction pattern.
968       continue;
969 
970     case Instruction::Load:
971       // Perfect, `load` node means we've reached an edge of the graph.
972       continue;
973 
974     default:        // Unknown node.
975       return false; // Not an `or` reduction pattern.
976     }
977   }
978 
979   return true;
980 }
981 
982 /// Return true if it may be profitable to convert this (X|Y) into (X+Y).
983 static bool shouldConvertOrWithNoCommonBitsToAdd(Instruction *Or) {
984   // Don't bother to convert this up unless either the LHS is an associable add
985   // or subtract or mul or if this is only used by one of the above.
986   // This is only a compile-time improvement, it is not needed for correctness!
987   auto isInteresting = [](Value *V) {
988     for (auto Op : {Instruction::Add, Instruction::Sub, Instruction::Mul,
989                     Instruction::Shl})
990       if (isReassociableOp(V, Op))
991         return true;
992     return false;
993   };
994 
995   if (any_of(Or->operands(), isInteresting))
996     return true;
997 
998   Value *VB = Or->user_back();
999   if (Or->hasOneUse() && isInteresting(VB))
1000     return true;
1001 
1002   return false;
1003 }
1004 
1005 /// If we have (X|Y), and iff X and Y have no common bits set,
1006 /// transform this into (X+Y) to allow arithmetics reassociation.
1007 static BinaryOperator *convertOrWithNoCommonBitsToAdd(Instruction *Or) {
1008   // Convert an or into an add.
1009   BinaryOperator *New =
1010       CreateAdd(Or->getOperand(0), Or->getOperand(1), "", Or, Or);
1011   New->setHasNoSignedWrap();
1012   New->setHasNoUnsignedWrap();
1013   New->takeName(Or);
1014 
1015   // Everyone now refers to the add instruction.
1016   Or->replaceAllUsesWith(New);
1017   New->setDebugLoc(Or->getDebugLoc());
1018 
1019   LLVM_DEBUG(dbgs() << "Converted or into an add: " << *New << '\n');
1020   return New;
1021 }
1022 
1023 /// Return true if we should break up this subtract of X-Y into (X + -Y).
1024 static bool ShouldBreakUpSubtract(Instruction *Sub) {
1025   // If this is a negation, we can't split it up!
1026   if (match(Sub, m_Neg(m_Value())) || match(Sub, m_FNeg(m_Value())))
1027     return false;
1028 
1029   // Don't breakup X - undef.
1030   if (isa<UndefValue>(Sub->getOperand(1)))
1031     return false;
1032 
1033   // Don't bother to break this up unless either the LHS is an associable add or
1034   // subtract or if this is only used by one.
1035   Value *V0 = Sub->getOperand(0);
1036   if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
1037       isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
1038     return true;
1039   Value *V1 = Sub->getOperand(1);
1040   if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
1041       isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
1042     return true;
1043   Value *VB = Sub->user_back();
1044   if (Sub->hasOneUse() &&
1045       (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
1046        isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
1047     return true;
1048 
1049   return false;
1050 }
1051 
1052 /// If we have (X-Y), and if either X is an add, or if this is only used by an
1053 /// add, transform this into (X+(0-Y)) to promote better reassociation.
1054 static BinaryOperator *BreakUpSubtract(Instruction *Sub,
1055                                        ReassociatePass::OrderedSet &ToRedo) {
1056   // Convert a subtract into an add and a neg instruction. This allows sub
1057   // instructions to be commuted with other add instructions.
1058   //
1059   // Calculate the negative value of Operand 1 of the sub instruction,
1060   // and set it as the RHS of the add instruction we just made.
1061   Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo);
1062   BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
1063   Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
1064   Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
1065   New->takeName(Sub);
1066 
1067   // Everyone now refers to the add instruction.
1068   Sub->replaceAllUsesWith(New);
1069   New->setDebugLoc(Sub->getDebugLoc());
1070 
1071   LLVM_DEBUG(dbgs() << "Negated: " << *New << '\n');
1072   return New;
1073 }
1074 
1075 /// If this is a shift of a reassociable multiply or is used by one, change
1076 /// this into a multiply by a constant to assist with further reassociation.
1077 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1078   Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
1079   auto *SA = cast<ConstantInt>(Shl->getOperand(1));
1080   MulCst = ConstantExpr::getShl(MulCst, SA);
1081 
1082   BinaryOperator *Mul =
1083     BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
1084   Shl->setOperand(0, PoisonValue::get(Shl->getType())); // Drop use of op.
1085   Mul->takeName(Shl);
1086 
1087   // Everyone now refers to the mul instruction.
1088   Shl->replaceAllUsesWith(Mul);
1089   Mul->setDebugLoc(Shl->getDebugLoc());
1090 
1091   // We can safely preserve the nuw flag in all cases.  It's also safe to turn a
1092   // nuw nsw shl into a nuw nsw mul.  However, nsw in isolation requires special
1093   // handling.  It can be preserved as long as we're not left shifting by
1094   // bitwidth - 1.
1095   bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
1096   bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
1097   unsigned BitWidth = Shl->getType()->getIntegerBitWidth();
1098   if (NSW && (NUW || SA->getValue().ult(BitWidth - 1)))
1099     Mul->setHasNoSignedWrap(true);
1100   Mul->setHasNoUnsignedWrap(NUW);
1101   return Mul;
1102 }
1103 
1104 /// Scan backwards and forwards among values with the same rank as element i
1105 /// to see if X exists.  If X does not exist, return i.  This is useful when
1106 /// scanning for 'x' when we see '-x' because they both get the same rank.
1107 static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops,
1108                                   unsigned i, Value *X) {
1109   unsigned XRank = Ops[i].Rank;
1110   unsigned e = Ops.size();
1111   for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
1112     if (Ops[j].Op == X)
1113       return j;
1114     if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1115       if (Instruction *I2 = dyn_cast<Instruction>(X))
1116         if (I1->isIdenticalTo(I2))
1117           return j;
1118   }
1119   // Scan backwards.
1120   for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
1121     if (Ops[j].Op == X)
1122       return j;
1123     if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1124       if (Instruction *I2 = dyn_cast<Instruction>(X))
1125         if (I1->isIdenticalTo(I2))
1126           return j;
1127   }
1128   return i;
1129 }
1130 
1131 /// Emit a tree of add instructions, summing Ops together
1132 /// and returning the result.  Insert the tree before I.
1133 static Value *EmitAddTreeOfValues(Instruction *I,
1134                                   SmallVectorImpl<WeakTrackingVH> &Ops) {
1135   if (Ops.size() == 1) return Ops.back();
1136 
1137   Value *V1 = Ops.pop_back_val();
1138   Value *V2 = EmitAddTreeOfValues(I, Ops);
1139   return CreateAdd(V2, V1, "reass.add", I, I);
1140 }
1141 
1142 /// If V is an expression tree that is a multiplication sequence,
1143 /// and if this sequence contains a multiply by Factor,
1144 /// remove Factor from the tree and return the new tree.
1145 Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) {
1146   BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1147   if (!BO)
1148     return nullptr;
1149 
1150   SmallVector<RepeatedValue, 8> Tree;
1151   MadeChange |= LinearizeExprTree(BO, Tree, RedoInsts);
1152   SmallVector<ValueEntry, 8> Factors;
1153   Factors.reserve(Tree.size());
1154   for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1155     RepeatedValue E = Tree[i];
1156     Factors.append(E.second.getZExtValue(),
1157                    ValueEntry(getRank(E.first), E.first));
1158   }
1159 
1160   bool FoundFactor = false;
1161   bool NeedsNegate = false;
1162   for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1163     if (Factors[i].Op == Factor) {
1164       FoundFactor = true;
1165       Factors.erase(Factors.begin()+i);
1166       break;
1167     }
1168 
1169     // If this is a negative version of this factor, remove it.
1170     if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1171       if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1172         if (FC1->getValue() == -FC2->getValue()) {
1173           FoundFactor = NeedsNegate = true;
1174           Factors.erase(Factors.begin()+i);
1175           break;
1176         }
1177     } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1178       if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1179         const APFloat &F1 = FC1->getValueAPF();
1180         APFloat F2(FC2->getValueAPF());
1181         F2.changeSign();
1182         if (F1 == F2) {
1183           FoundFactor = NeedsNegate = true;
1184           Factors.erase(Factors.begin() + i);
1185           break;
1186         }
1187       }
1188     }
1189   }
1190 
1191   if (!FoundFactor) {
1192     // Make sure to restore the operands to the expression tree.
1193     RewriteExprTree(BO, Factors);
1194     return nullptr;
1195   }
1196 
1197   BasicBlock::iterator InsertPt = ++BO->getIterator();
1198 
1199   // If this was just a single multiply, remove the multiply and return the only
1200   // remaining operand.
1201   if (Factors.size() == 1) {
1202     RedoInsts.insert(BO);
1203     V = Factors[0].Op;
1204   } else {
1205     RewriteExprTree(BO, Factors);
1206     V = BO;
1207   }
1208 
1209   if (NeedsNegate)
1210     V = CreateNeg(V, "neg", &*InsertPt, BO);
1211 
1212   return V;
1213 }
1214 
1215 /// If V is a single-use multiply, recursively add its operands as factors,
1216 /// otherwise add V to the list of factors.
1217 ///
1218 /// Ops is the top-level list of add operands we're trying to factor.
1219 static void FindSingleUseMultiplyFactors(Value *V,
1220                                          SmallVectorImpl<Value*> &Factors) {
1221   BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1222   if (!BO) {
1223     Factors.push_back(V);
1224     return;
1225   }
1226 
1227   // Otherwise, add the LHS and RHS to the list of factors.
1228   FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
1229   FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
1230 }
1231 
1232 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1233 /// This optimizes based on identities.  If it can be reduced to a single Value,
1234 /// it is returned, otherwise the Ops list is mutated as necessary.
1235 static Value *OptimizeAndOrXor(unsigned Opcode,
1236                                SmallVectorImpl<ValueEntry> &Ops) {
1237   // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1238   // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1239   for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1240     // First, check for X and ~X in the operand list.
1241     assert(i < Ops.size());
1242     Value *X;
1243     if (match(Ops[i].Op, m_Not(m_Value(X)))) {    // Cannot occur for ^.
1244       unsigned FoundX = FindInOperandList(Ops, i, X);
1245       if (FoundX != i) {
1246         if (Opcode == Instruction::And)   // ...&X&~X = 0
1247           return Constant::getNullValue(X->getType());
1248 
1249         if (Opcode == Instruction::Or)    // ...|X|~X = -1
1250           return Constant::getAllOnesValue(X->getType());
1251       }
1252     }
1253 
1254     // Next, check for duplicate pairs of values, which we assume are next to
1255     // each other, due to our sorting criteria.
1256     assert(i < Ops.size());
1257     if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1258       if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1259         // Drop duplicate values for And and Or.
1260         Ops.erase(Ops.begin()+i);
1261         --i; --e;
1262         ++NumAnnihil;
1263         continue;
1264       }
1265 
1266       // Drop pairs of values for Xor.
1267       assert(Opcode == Instruction::Xor);
1268       if (e == 2)
1269         return Constant::getNullValue(Ops[0].Op->getType());
1270 
1271       // Y ^ X^X -> Y
1272       Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1273       i -= 1; e -= 2;
1274       ++NumAnnihil;
1275     }
1276   }
1277   return nullptr;
1278 }
1279 
1280 /// Helper function of CombineXorOpnd(). It creates a bitwise-and
1281 /// instruction with the given two operands, and return the resulting
1282 /// instruction. There are two special cases: 1) if the constant operand is 0,
1283 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1284 /// be returned.
1285 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1286                              const APInt &ConstOpnd) {
1287   if (ConstOpnd.isZero())
1288     return nullptr;
1289 
1290   if (ConstOpnd.isAllOnes())
1291     return Opnd;
1292 
1293   Instruction *I = BinaryOperator::CreateAnd(
1294       Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra",
1295       InsertBefore);
1296   I->setDebugLoc(InsertBefore->getDebugLoc());
1297   return I;
1298 }
1299 
1300 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1301 // into "R ^ C", where C would be 0, and R is a symbolic value.
1302 //
1303 // If it was successful, true is returned, and the "R" and "C" is returned
1304 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1305 // and both "Res" and "ConstOpnd" remain unchanged.
1306 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1307                                      APInt &ConstOpnd, Value *&Res) {
1308   // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1309   //                       = ((x | c1) ^ c1) ^ (c1 ^ c2)
1310   //                       = (x & ~c1) ^ (c1 ^ c2)
1311   // It is useful only when c1 == c2.
1312   if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isZero())
1313     return false;
1314 
1315   if (!Opnd1->getValue()->hasOneUse())
1316     return false;
1317 
1318   const APInt &C1 = Opnd1->getConstPart();
1319   if (C1 != ConstOpnd)
1320     return false;
1321 
1322   Value *X = Opnd1->getSymbolicPart();
1323   Res = createAndInstr(I, X, ~C1);
1324   // ConstOpnd was C2, now C1 ^ C2.
1325   ConstOpnd ^= C1;
1326 
1327   if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1328     RedoInsts.insert(T);
1329   return true;
1330 }
1331 
1332 // Helper function of OptimizeXor(). It tries to simplify
1333 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1334 // symbolic value.
1335 //
1336 // If it was successful, true is returned, and the "R" and "C" is returned
1337 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1338 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1339 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1340 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1341                                      XorOpnd *Opnd2, APInt &ConstOpnd,
1342                                      Value *&Res) {
1343   Value *X = Opnd1->getSymbolicPart();
1344   if (X != Opnd2->getSymbolicPart())
1345     return false;
1346 
1347   // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1348   int DeadInstNum = 1;
1349   if (Opnd1->getValue()->hasOneUse())
1350     DeadInstNum++;
1351   if (Opnd2->getValue()->hasOneUse())
1352     DeadInstNum++;
1353 
1354   // Xor-Rule 2:
1355   //  (x | c1) ^ (x & c2)
1356   //   = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1357   //   = (x & ~c1) ^ (x & c2) ^ c1               // Xor-Rule 1
1358   //   = (x & c3) ^ c1, where c3 = ~c1 ^ c2      // Xor-rule 3
1359   //
1360   if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1361     if (Opnd2->isOrExpr())
1362       std::swap(Opnd1, Opnd2);
1363 
1364     const APInt &C1 = Opnd1->getConstPart();
1365     const APInt &C2 = Opnd2->getConstPart();
1366     APInt C3((~C1) ^ C2);
1367 
1368     // Do not increase code size!
1369     if (!C3.isZero() && !C3.isAllOnes()) {
1370       int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1371       if (NewInstNum > DeadInstNum)
1372         return false;
1373     }
1374 
1375     Res = createAndInstr(I, X, C3);
1376     ConstOpnd ^= C1;
1377   } else if (Opnd1->isOrExpr()) {
1378     // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1379     //
1380     const APInt &C1 = Opnd1->getConstPart();
1381     const APInt &C2 = Opnd2->getConstPart();
1382     APInt C3 = C1 ^ C2;
1383 
1384     // Do not increase code size
1385     if (!C3.isZero() && !C3.isAllOnes()) {
1386       int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1387       if (NewInstNum > DeadInstNum)
1388         return false;
1389     }
1390 
1391     Res = createAndInstr(I, X, C3);
1392     ConstOpnd ^= C3;
1393   } else {
1394     // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1395     //
1396     const APInt &C1 = Opnd1->getConstPart();
1397     const APInt &C2 = Opnd2->getConstPart();
1398     APInt C3 = C1 ^ C2;
1399     Res = createAndInstr(I, X, C3);
1400   }
1401 
1402   // Put the original operands in the Redo list; hope they will be deleted
1403   // as dead code.
1404   if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1405     RedoInsts.insert(T);
1406   if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1407     RedoInsts.insert(T);
1408 
1409   return true;
1410 }
1411 
1412 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1413 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1414 /// necessary.
1415 Value *ReassociatePass::OptimizeXor(Instruction *I,
1416                                     SmallVectorImpl<ValueEntry> &Ops) {
1417   if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1418     return V;
1419 
1420   if (Ops.size() == 1)
1421     return nullptr;
1422 
1423   SmallVector<XorOpnd, 8> Opnds;
1424   SmallVector<XorOpnd*, 8> OpndPtrs;
1425   Type *Ty = Ops[0].Op->getType();
1426   APInt ConstOpnd(Ty->getScalarSizeInBits(), 0);
1427 
1428   // Step 1: Convert ValueEntry to XorOpnd
1429   for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1430     Value *V = Ops[i].Op;
1431     const APInt *C;
1432     // TODO: Support non-splat vectors.
1433     if (match(V, m_APInt(C))) {
1434       ConstOpnd ^= *C;
1435     } else {
1436       XorOpnd O(V);
1437       O.setSymbolicRank(getRank(O.getSymbolicPart()));
1438       Opnds.push_back(O);
1439     }
1440   }
1441 
1442   // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1443   //  It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1444   //  the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1445   //  with the previous loop --- the iterator of the "Opnds" may be invalidated
1446   //  when new elements are added to the vector.
1447   for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1448     OpndPtrs.push_back(&Opnds[i]);
1449 
1450   // Step 2: Sort the Xor-Operands in a way such that the operands containing
1451   //  the same symbolic value cluster together. For instance, the input operand
1452   //  sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1453   //  ("x | 123", "x & 789", "y & 456").
1454   //
1455   //  The purpose is twofold:
1456   //  1) Cluster together the operands sharing the same symbolic-value.
1457   //  2) Operand having smaller symbolic-value-rank is permuted earlier, which
1458   //     could potentially shorten crital path, and expose more loop-invariants.
1459   //     Note that values' rank are basically defined in RPO order (FIXME).
1460   //     So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1461   //     than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1462   //     "z" in the order of X-Y-Z is better than any other orders.
1463   llvm::stable_sort(OpndPtrs, [](XorOpnd *LHS, XorOpnd *RHS) {
1464     return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1465   });
1466 
1467   // Step 3: Combine adjacent operands
1468   XorOpnd *PrevOpnd = nullptr;
1469   bool Changed = false;
1470   for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1471     XorOpnd *CurrOpnd = OpndPtrs[i];
1472     // The combined value
1473     Value *CV;
1474 
1475     // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1476     if (!ConstOpnd.isZero() && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1477       Changed = true;
1478       if (CV)
1479         *CurrOpnd = XorOpnd(CV);
1480       else {
1481         CurrOpnd->Invalidate();
1482         continue;
1483       }
1484     }
1485 
1486     if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1487       PrevOpnd = CurrOpnd;
1488       continue;
1489     }
1490 
1491     // step 3.2: When previous and current operands share the same symbolic
1492     //  value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1493     if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1494       // Remove previous operand
1495       PrevOpnd->Invalidate();
1496       if (CV) {
1497         *CurrOpnd = XorOpnd(CV);
1498         PrevOpnd = CurrOpnd;
1499       } else {
1500         CurrOpnd->Invalidate();
1501         PrevOpnd = nullptr;
1502       }
1503       Changed = true;
1504     }
1505   }
1506 
1507   // Step 4: Reassemble the Ops
1508   if (Changed) {
1509     Ops.clear();
1510     for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1511       XorOpnd &O = Opnds[i];
1512       if (O.isInvalid())
1513         continue;
1514       ValueEntry VE(getRank(O.getValue()), O.getValue());
1515       Ops.push_back(VE);
1516     }
1517     if (!ConstOpnd.isZero()) {
1518       Value *C = ConstantInt::get(Ty, ConstOpnd);
1519       ValueEntry VE(getRank(C), C);
1520       Ops.push_back(VE);
1521     }
1522     unsigned Sz = Ops.size();
1523     if (Sz == 1)
1524       return Ops.back().Op;
1525     if (Sz == 0) {
1526       assert(ConstOpnd.isZero());
1527       return ConstantInt::get(Ty, ConstOpnd);
1528     }
1529   }
1530 
1531   return nullptr;
1532 }
1533 
1534 /// Optimize a series of operands to an 'add' instruction.  This
1535 /// optimizes based on identities.  If it can be reduced to a single Value, it
1536 /// is returned, otherwise the Ops list is mutated as necessary.
1537 Value *ReassociatePass::OptimizeAdd(Instruction *I,
1538                                     SmallVectorImpl<ValueEntry> &Ops) {
1539   // Scan the operand lists looking for X and -X pairs.  If we find any, we
1540   // can simplify expressions like X+-X == 0 and X+~X ==-1.  While we're at it,
1541   // scan for any
1542   // duplicates.  We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1543 
1544   for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1545     Value *TheOp = Ops[i].Op;
1546     // Check to see if we've seen this operand before.  If so, we factor all
1547     // instances of the operand together.  Due to our sorting criteria, we know
1548     // that these need to be next to each other in the vector.
1549     if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1550       // Rescan the list, remove all instances of this operand from the expr.
1551       unsigned NumFound = 0;
1552       do {
1553         Ops.erase(Ops.begin()+i);
1554         ++NumFound;
1555       } while (i != Ops.size() && Ops[i].Op == TheOp);
1556 
1557       LLVM_DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp
1558                         << '\n');
1559       ++NumFactor;
1560 
1561       // Insert a new multiply.
1562       Type *Ty = TheOp->getType();
1563       Constant *C = Ty->isIntOrIntVectorTy() ?
1564         ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1565       Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1566 
1567       // Now that we have inserted a multiply, optimize it. This allows us to
1568       // handle cases that require multiple factoring steps, such as this:
1569       // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1570       RedoInsts.insert(Mul);
1571 
1572       // If every add operand was a duplicate, return the multiply.
1573       if (Ops.empty())
1574         return Mul;
1575 
1576       // Otherwise, we had some input that didn't have the dupe, such as
1577       // "A + A + B" -> "A*2 + B".  Add the new multiply to the list of
1578       // things being added by this operation.
1579       Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1580 
1581       --i;
1582       e = Ops.size();
1583       continue;
1584     }
1585 
1586     // Check for X and -X or X and ~X in the operand list.
1587     Value *X;
1588     if (!match(TheOp, m_Neg(m_Value(X))) && !match(TheOp, m_Not(m_Value(X))) &&
1589         !match(TheOp, m_FNeg(m_Value(X))))
1590       continue;
1591 
1592     unsigned FoundX = FindInOperandList(Ops, i, X);
1593     if (FoundX == i)
1594       continue;
1595 
1596     // Remove X and -X from the operand list.
1597     if (Ops.size() == 2 &&
1598         (match(TheOp, m_Neg(m_Value())) || match(TheOp, m_FNeg(m_Value()))))
1599       return Constant::getNullValue(X->getType());
1600 
1601     // Remove X and ~X from the operand list.
1602     if (Ops.size() == 2 && match(TheOp, m_Not(m_Value())))
1603       return Constant::getAllOnesValue(X->getType());
1604 
1605     Ops.erase(Ops.begin()+i);
1606     if (i < FoundX)
1607       --FoundX;
1608     else
1609       --i;   // Need to back up an extra one.
1610     Ops.erase(Ops.begin()+FoundX);
1611     ++NumAnnihil;
1612     --i;     // Revisit element.
1613     e -= 2;  // Removed two elements.
1614 
1615     // if X and ~X we append -1 to the operand list.
1616     if (match(TheOp, m_Not(m_Value()))) {
1617       Value *V = Constant::getAllOnesValue(X->getType());
1618       Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1619       e += 1;
1620     }
1621   }
1622 
1623   // Scan the operand list, checking to see if there are any common factors
1624   // between operands.  Consider something like A*A+A*B*C+D.  We would like to
1625   // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1626   // To efficiently find this, we count the number of times a factor occurs
1627   // for any ADD operands that are MULs.
1628   DenseMap<Value*, unsigned> FactorOccurrences;
1629 
1630   // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1631   // where they are actually the same multiply.
1632   unsigned MaxOcc = 0;
1633   Value *MaxOccVal = nullptr;
1634   for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1635     BinaryOperator *BOp =
1636         isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1637     if (!BOp)
1638       continue;
1639 
1640     // Compute all of the factors of this added value.
1641     SmallVector<Value*, 8> Factors;
1642     FindSingleUseMultiplyFactors(BOp, Factors);
1643     assert(Factors.size() > 1 && "Bad linearize!");
1644 
1645     // Add one to FactorOccurrences for each unique factor in this op.
1646     SmallPtrSet<Value*, 8> Duplicates;
1647     for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1648       Value *Factor = Factors[i];
1649       if (!Duplicates.insert(Factor).second)
1650         continue;
1651 
1652       unsigned Occ = ++FactorOccurrences[Factor];
1653       if (Occ > MaxOcc) {
1654         MaxOcc = Occ;
1655         MaxOccVal = Factor;
1656       }
1657 
1658       // If Factor is a negative constant, add the negated value as a factor
1659       // because we can percolate the negate out.  Watch for minint, which
1660       // cannot be positivified.
1661       if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1662         if (CI->isNegative() && !CI->isMinValue(true)) {
1663           Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1664           if (!Duplicates.insert(Factor).second)
1665             continue;
1666           unsigned Occ = ++FactorOccurrences[Factor];
1667           if (Occ > MaxOcc) {
1668             MaxOcc = Occ;
1669             MaxOccVal = Factor;
1670           }
1671         }
1672       } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1673         if (CF->isNegative()) {
1674           APFloat F(CF->getValueAPF());
1675           F.changeSign();
1676           Factor = ConstantFP::get(CF->getContext(), F);
1677           if (!Duplicates.insert(Factor).second)
1678             continue;
1679           unsigned Occ = ++FactorOccurrences[Factor];
1680           if (Occ > MaxOcc) {
1681             MaxOcc = Occ;
1682             MaxOccVal = Factor;
1683           }
1684         }
1685       }
1686     }
1687   }
1688 
1689   // If any factor occurred more than one time, we can pull it out.
1690   if (MaxOcc > 1) {
1691     LLVM_DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal
1692                       << '\n');
1693     ++NumFactor;
1694 
1695     // Create a new instruction that uses the MaxOccVal twice.  If we don't do
1696     // this, we could otherwise run into situations where removing a factor
1697     // from an expression will drop a use of maxocc, and this can cause
1698     // RemoveFactorFromExpression on successive values to behave differently.
1699     Instruction *DummyInst =
1700         I->getType()->isIntOrIntVectorTy()
1701             ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1702             : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1703 
1704     SmallVector<WeakTrackingVH, 4> NewMulOps;
1705     for (unsigned i = 0; i != Ops.size(); ++i) {
1706       // Only try to remove factors from expressions we're allowed to.
1707       BinaryOperator *BOp =
1708           isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1709       if (!BOp)
1710         continue;
1711 
1712       if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1713         // The factorized operand may occur several times.  Convert them all in
1714         // one fell swoop.
1715         for (unsigned j = Ops.size(); j != i;) {
1716           --j;
1717           if (Ops[j].Op == Ops[i].Op) {
1718             NewMulOps.push_back(V);
1719             Ops.erase(Ops.begin()+j);
1720           }
1721         }
1722         --i;
1723       }
1724     }
1725 
1726     // No need for extra uses anymore.
1727     DummyInst->deleteValue();
1728 
1729     unsigned NumAddedValues = NewMulOps.size();
1730     Value *V = EmitAddTreeOfValues(I, NewMulOps);
1731 
1732     // Now that we have inserted the add tree, optimize it. This allows us to
1733     // handle cases that require multiple factoring steps, such as this:
1734     // A*A*B + A*A*C   -->   A*(A*B+A*C)   -->   A*(A*(B+C))
1735     assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1736     (void)NumAddedValues;
1737     if (Instruction *VI = dyn_cast<Instruction>(V))
1738       RedoInsts.insert(VI);
1739 
1740     // Create the multiply.
1741     Instruction *V2 = CreateMul(V, MaxOccVal, "reass.mul", I, I);
1742 
1743     // Rerun associate on the multiply in case the inner expression turned into
1744     // a multiply.  We want to make sure that we keep things in canonical form.
1745     RedoInsts.insert(V2);
1746 
1747     // If every add operand included the factor (e.g. "A*B + A*C"), then the
1748     // entire result expression is just the multiply "A*(B+C)".
1749     if (Ops.empty())
1750       return V2;
1751 
1752     // Otherwise, we had some input that didn't have the factor, such as
1753     // "A*B + A*C + D" -> "A*(B+C) + D".  Add the new multiply to the list of
1754     // things being added by this operation.
1755     Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1756   }
1757 
1758   return nullptr;
1759 }
1760 
1761 /// Build up a vector of value/power pairs factoring a product.
1762 ///
1763 /// Given a series of multiplication operands, build a vector of factors and
1764 /// the powers each is raised to when forming the final product. Sort them in
1765 /// the order of descending power.
1766 ///
1767 ///      (x*x)          -> [(x, 2)]
1768 ///     ((x*x)*x)       -> [(x, 3)]
1769 ///   ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1770 ///
1771 /// \returns Whether any factors have a power greater than one.
1772 static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1773                                    SmallVectorImpl<Factor> &Factors) {
1774   // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1775   // Compute the sum of powers of simplifiable factors.
1776   unsigned FactorPowerSum = 0;
1777   for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1778     Value *Op = Ops[Idx-1].Op;
1779 
1780     // Count the number of occurrences of this value.
1781     unsigned Count = 1;
1782     for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1783       ++Count;
1784     // Track for simplification all factors which occur 2 or more times.
1785     if (Count > 1)
1786       FactorPowerSum += Count;
1787   }
1788 
1789   // We can only simplify factors if the sum of the powers of our simplifiable
1790   // factors is 4 or higher. When that is the case, we will *always* have
1791   // a simplification. This is an important invariant to prevent cyclicly
1792   // trying to simplify already minimal formations.
1793   if (FactorPowerSum < 4)
1794     return false;
1795 
1796   // Now gather the simplifiable factors, removing them from Ops.
1797   FactorPowerSum = 0;
1798   for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1799     Value *Op = Ops[Idx-1].Op;
1800 
1801     // Count the number of occurrences of this value.
1802     unsigned Count = 1;
1803     for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1804       ++Count;
1805     if (Count == 1)
1806       continue;
1807     // Move an even number of occurrences to Factors.
1808     Count &= ~1U;
1809     Idx -= Count;
1810     FactorPowerSum += Count;
1811     Factors.push_back(Factor(Op, Count));
1812     Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1813   }
1814 
1815   // None of the adjustments above should have reduced the sum of factor powers
1816   // below our mininum of '4'.
1817   assert(FactorPowerSum >= 4);
1818 
1819   llvm::stable_sort(Factors, [](const Factor &LHS, const Factor &RHS) {
1820     return LHS.Power > RHS.Power;
1821   });
1822   return true;
1823 }
1824 
1825 /// Build a tree of multiplies, computing the product of Ops.
1826 static Value *buildMultiplyTree(IRBuilderBase &Builder,
1827                                 SmallVectorImpl<Value*> &Ops) {
1828   if (Ops.size() == 1)
1829     return Ops.back();
1830 
1831   Value *LHS = Ops.pop_back_val();
1832   do {
1833     if (LHS->getType()->isIntOrIntVectorTy())
1834       LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1835     else
1836       LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1837   } while (!Ops.empty());
1838 
1839   return LHS;
1840 }
1841 
1842 /// Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1843 ///
1844 /// Given a vector of values raised to various powers, where no two values are
1845 /// equal and the powers are sorted in decreasing order, compute the minimal
1846 /// DAG of multiplies to compute the final product, and return that product
1847 /// value.
1848 Value *
1849 ReassociatePass::buildMinimalMultiplyDAG(IRBuilderBase &Builder,
1850                                          SmallVectorImpl<Factor> &Factors) {
1851   assert(Factors[0].Power);
1852   SmallVector<Value *, 4> OuterProduct;
1853   for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1854        Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1855     if (Factors[Idx].Power != Factors[LastIdx].Power) {
1856       LastIdx = Idx;
1857       continue;
1858     }
1859 
1860     // We want to multiply across all the factors with the same power so that
1861     // we can raise them to that power as a single entity. Build a mini tree
1862     // for that.
1863     SmallVector<Value *, 4> InnerProduct;
1864     InnerProduct.push_back(Factors[LastIdx].Base);
1865     do {
1866       InnerProduct.push_back(Factors[Idx].Base);
1867       ++Idx;
1868     } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1869 
1870     // Reset the base value of the first factor to the new expression tree.
1871     // We'll remove all the factors with the same power in a second pass.
1872     Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1873     if (Instruction *MI = dyn_cast<Instruction>(M))
1874       RedoInsts.insert(MI);
1875 
1876     LastIdx = Idx;
1877   }
1878   // Unique factors with equal powers -- we've folded them into the first one's
1879   // base.
1880   Factors.erase(std::unique(Factors.begin(), Factors.end(),
1881                             [](const Factor &LHS, const Factor &RHS) {
1882                               return LHS.Power == RHS.Power;
1883                             }),
1884                 Factors.end());
1885 
1886   // Iteratively collect the base of each factor with an add power into the
1887   // outer product, and halve each power in preparation for squaring the
1888   // expression.
1889   for (Factor &F : Factors) {
1890     if (F.Power & 1)
1891       OuterProduct.push_back(F.Base);
1892     F.Power >>= 1;
1893   }
1894   if (Factors[0].Power) {
1895     Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1896     OuterProduct.push_back(SquareRoot);
1897     OuterProduct.push_back(SquareRoot);
1898   }
1899   if (OuterProduct.size() == 1)
1900     return OuterProduct.front();
1901 
1902   Value *V = buildMultiplyTree(Builder, OuterProduct);
1903   return V;
1904 }
1905 
1906 Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
1907                                     SmallVectorImpl<ValueEntry> &Ops) {
1908   // We can only optimize the multiplies when there is a chain of more than
1909   // three, such that a balanced tree might require fewer total multiplies.
1910   if (Ops.size() < 4)
1911     return nullptr;
1912 
1913   // Try to turn linear trees of multiplies without other uses of the
1914   // intermediate stages into minimal multiply DAGs with perfect sub-expression
1915   // re-use.
1916   SmallVector<Factor, 4> Factors;
1917   if (!collectMultiplyFactors(Ops, Factors))
1918     return nullptr; // All distinct factors, so nothing left for us to do.
1919 
1920   IRBuilder<> Builder(I);
1921   // The reassociate transformation for FP operations is performed only
1922   // if unsafe algebra is permitted by FastMathFlags. Propagate those flags
1923   // to the newly generated operations.
1924   if (auto FPI = dyn_cast<FPMathOperator>(I))
1925     Builder.setFastMathFlags(FPI->getFastMathFlags());
1926 
1927   Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1928   if (Ops.empty())
1929     return V;
1930 
1931   ValueEntry NewEntry = ValueEntry(getRank(V), V);
1932   Ops.insert(llvm::lower_bound(Ops, NewEntry), NewEntry);
1933   return nullptr;
1934 }
1935 
1936 Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
1937                                            SmallVectorImpl<ValueEntry> &Ops) {
1938   // Now that we have the linearized expression tree, try to optimize it.
1939   // Start by folding any constants that we found.
1940   const DataLayout &DL = I->getModule()->getDataLayout();
1941   Constant *Cst = nullptr;
1942   unsigned Opcode = I->getOpcode();
1943   while (!Ops.empty()) {
1944     if (auto *C = dyn_cast<Constant>(Ops.back().Op)) {
1945       if (!Cst) {
1946         Ops.pop_back();
1947         Cst = C;
1948         continue;
1949       }
1950       if (Constant *Res = ConstantFoldBinaryOpOperands(Opcode, C, Cst, DL)) {
1951         Ops.pop_back();
1952         Cst = Res;
1953         continue;
1954       }
1955     }
1956     break;
1957   }
1958   // If there was nothing but constants then we are done.
1959   if (Ops.empty())
1960     return Cst;
1961 
1962   // Put the combined constant back at the end of the operand list, except if
1963   // there is no point.  For example, an add of 0 gets dropped here, while a
1964   // multiplication by zero turns the whole expression into zero.
1965   if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1966     if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1967       return Cst;
1968     Ops.push_back(ValueEntry(0, Cst));
1969   }
1970 
1971   if (Ops.size() == 1) return Ops[0].Op;
1972 
1973   // Handle destructive annihilation due to identities between elements in the
1974   // argument list here.
1975   unsigned NumOps = Ops.size();
1976   switch (Opcode) {
1977   default: break;
1978   case Instruction::And:
1979   case Instruction::Or:
1980     if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1981       return Result;
1982     break;
1983 
1984   case Instruction::Xor:
1985     if (Value *Result = OptimizeXor(I, Ops))
1986       return Result;
1987     break;
1988 
1989   case Instruction::Add:
1990   case Instruction::FAdd:
1991     if (Value *Result = OptimizeAdd(I, Ops))
1992       return Result;
1993     break;
1994 
1995   case Instruction::Mul:
1996   case Instruction::FMul:
1997     if (Value *Result = OptimizeMul(I, Ops))
1998       return Result;
1999     break;
2000   }
2001 
2002   if (Ops.size() != NumOps)
2003     return OptimizeExpression(I, Ops);
2004   return nullptr;
2005 }
2006 
2007 // Remove dead instructions and if any operands are trivially dead add them to
2008 // Insts so they will be removed as well.
2009 void ReassociatePass::RecursivelyEraseDeadInsts(Instruction *I,
2010                                                 OrderedSet &Insts) {
2011   assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
2012   SmallVector<Value *, 4> Ops(I->operands());
2013   ValueRankMap.erase(I);
2014   Insts.remove(I);
2015   RedoInsts.remove(I);
2016   llvm::salvageDebugInfo(*I);
2017   I->eraseFromParent();
2018   for (auto *Op : Ops)
2019     if (Instruction *OpInst = dyn_cast<Instruction>(Op))
2020       if (OpInst->use_empty())
2021         Insts.insert(OpInst);
2022 }
2023 
2024 /// Zap the given instruction, adding interesting operands to the work list.
2025 void ReassociatePass::EraseInst(Instruction *I) {
2026   assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
2027   LLVM_DEBUG(dbgs() << "Erasing dead inst: "; I->dump());
2028 
2029   SmallVector<Value *, 8> Ops(I->operands());
2030   // Erase the dead instruction.
2031   ValueRankMap.erase(I);
2032   RedoInsts.remove(I);
2033   llvm::salvageDebugInfo(*I);
2034   I->eraseFromParent();
2035   // Optimize its operands.
2036   SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
2037   for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2038     if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
2039       // If this is a node in an expression tree, climb to the expression root
2040       // and add that since that's where optimization actually happens.
2041       unsigned Opcode = Op->getOpcode();
2042       while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
2043              Visited.insert(Op).second)
2044         Op = Op->user_back();
2045 
2046       // The instruction we're going to push may be coming from a
2047       // dead block, and Reassociate skips the processing of unreachable
2048       // blocks because it's a waste of time and also because it can
2049       // lead to infinite loop due to LLVM's non-standard definition
2050       // of dominance.
2051       if (ValueRankMap.find(Op) != ValueRankMap.end())
2052         RedoInsts.insert(Op);
2053     }
2054 
2055   MadeChange = true;
2056 }
2057 
2058 /// Recursively analyze an expression to build a list of instructions that have
2059 /// negative floating-point constant operands. The caller can then transform
2060 /// the list to create positive constants for better reassociation and CSE.
2061 static void getNegatibleInsts(Value *V,
2062                               SmallVectorImpl<Instruction *> &Candidates) {
2063   // Handle only one-use instructions. Combining negations does not justify
2064   // replicating instructions.
2065   Instruction *I;
2066   if (!match(V, m_OneUse(m_Instruction(I))))
2067     return;
2068 
2069   // Handle expressions of multiplications and divisions.
2070   // TODO: This could look through floating-point casts.
2071   const APFloat *C;
2072   switch (I->getOpcode()) {
2073     case Instruction::FMul:
2074       // Not expecting non-canonical code here. Bail out and wait.
2075       if (match(I->getOperand(0), m_Constant()))
2076         break;
2077 
2078       if (match(I->getOperand(1), m_APFloat(C)) && C->isNegative()) {
2079         Candidates.push_back(I);
2080         LLVM_DEBUG(dbgs() << "FMul with negative constant: " << *I << '\n');
2081       }
2082       getNegatibleInsts(I->getOperand(0), Candidates);
2083       getNegatibleInsts(I->getOperand(1), Candidates);
2084       break;
2085     case Instruction::FDiv:
2086       // Not expecting non-canonical code here. Bail out and wait.
2087       if (match(I->getOperand(0), m_Constant()) &&
2088           match(I->getOperand(1), m_Constant()))
2089         break;
2090 
2091       if ((match(I->getOperand(0), m_APFloat(C)) && C->isNegative()) ||
2092           (match(I->getOperand(1), m_APFloat(C)) && C->isNegative())) {
2093         Candidates.push_back(I);
2094         LLVM_DEBUG(dbgs() << "FDiv with negative constant: " << *I << '\n');
2095       }
2096       getNegatibleInsts(I->getOperand(0), Candidates);
2097       getNegatibleInsts(I->getOperand(1), Candidates);
2098       break;
2099     default:
2100       break;
2101   }
2102 }
2103 
2104 /// Given an fadd/fsub with an operand that is a one-use instruction
2105 /// (the fadd/fsub), try to change negative floating-point constants into
2106 /// positive constants to increase potential for reassociation and CSE.
2107 Instruction *ReassociatePass::canonicalizeNegFPConstantsForOp(Instruction *I,
2108                                                               Instruction *Op,
2109                                                               Value *OtherOp) {
2110   assert((I->getOpcode() == Instruction::FAdd ||
2111           I->getOpcode() == Instruction::FSub) && "Expected fadd/fsub");
2112 
2113   // Collect instructions with negative FP constants from the subtree that ends
2114   // in Op.
2115   SmallVector<Instruction *, 4> Candidates;
2116   getNegatibleInsts(Op, Candidates);
2117   if (Candidates.empty())
2118     return nullptr;
2119 
2120   // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the
2121   // resulting subtract will be broken up later.  This can get us into an
2122   // infinite loop during reassociation.
2123   bool IsFSub = I->getOpcode() == Instruction::FSub;
2124   bool NeedsSubtract = !IsFSub && Candidates.size() % 2 == 1;
2125   if (NeedsSubtract && ShouldBreakUpSubtract(I))
2126     return nullptr;
2127 
2128   for (Instruction *Negatible : Candidates) {
2129     const APFloat *C;
2130     if (match(Negatible->getOperand(0), m_APFloat(C))) {
2131       assert(!match(Negatible->getOperand(1), m_Constant()) &&
2132              "Expecting only 1 constant operand");
2133       assert(C->isNegative() && "Expected negative FP constant");
2134       Negatible->setOperand(0, ConstantFP::get(Negatible->getType(), abs(*C)));
2135       MadeChange = true;
2136     }
2137     if (match(Negatible->getOperand(1), m_APFloat(C))) {
2138       assert(!match(Negatible->getOperand(0), m_Constant()) &&
2139              "Expecting only 1 constant operand");
2140       assert(C->isNegative() && "Expected negative FP constant");
2141       Negatible->setOperand(1, ConstantFP::get(Negatible->getType(), abs(*C)));
2142       MadeChange = true;
2143     }
2144   }
2145   assert(MadeChange == true && "Negative constant candidate was not changed");
2146 
2147   // Negations cancelled out.
2148   if (Candidates.size() % 2 == 0)
2149     return I;
2150 
2151   // Negate the final operand in the expression by flipping the opcode of this
2152   // fadd/fsub.
2153   assert(Candidates.size() % 2 == 1 && "Expected odd number");
2154   IRBuilder<> Builder(I);
2155   Value *NewInst = IsFSub ? Builder.CreateFAddFMF(OtherOp, Op, I)
2156                           : Builder.CreateFSubFMF(OtherOp, Op, I);
2157   I->replaceAllUsesWith(NewInst);
2158   RedoInsts.insert(I);
2159   return dyn_cast<Instruction>(NewInst);
2160 }
2161 
2162 /// Canonicalize expressions that contain a negative floating-point constant
2163 /// of the following form:
2164 ///   OtherOp + (subtree) -> OtherOp {+/-} (canonical subtree)
2165 ///   (subtree) + OtherOp -> OtherOp {+/-} (canonical subtree)
2166 ///   OtherOp - (subtree) -> OtherOp {+/-} (canonical subtree)
2167 ///
2168 /// The fadd/fsub opcode may be switched to allow folding a negation into the
2169 /// input instruction.
2170 Instruction *ReassociatePass::canonicalizeNegFPConstants(Instruction *I) {
2171   LLVM_DEBUG(dbgs() << "Combine negations for: " << *I << '\n');
2172   Value *X;
2173   Instruction *Op;
2174   if (match(I, m_FAdd(m_Value(X), m_OneUse(m_Instruction(Op)))))
2175     if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X))
2176       I = R;
2177   if (match(I, m_FAdd(m_OneUse(m_Instruction(Op)), m_Value(X))))
2178     if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X))
2179       I = R;
2180   if (match(I, m_FSub(m_Value(X), m_OneUse(m_Instruction(Op)))))
2181     if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X))
2182       I = R;
2183   return I;
2184 }
2185 
2186 /// Inspect and optimize the given instruction. Note that erasing
2187 /// instructions is not allowed.
2188 void ReassociatePass::OptimizeInst(Instruction *I) {
2189   // Only consider operations that we understand.
2190   if (!isa<UnaryOperator>(I) && !isa<BinaryOperator>(I))
2191     return;
2192 
2193   if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
2194     // If an operand of this shift is a reassociable multiply, or if the shift
2195     // is used by a reassociable multiply or add, turn into a multiply.
2196     if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
2197         (I->hasOneUse() &&
2198          (isReassociableOp(I->user_back(), Instruction::Mul) ||
2199           isReassociableOp(I->user_back(), Instruction::Add)))) {
2200       Instruction *NI = ConvertShiftToMul(I);
2201       RedoInsts.insert(I);
2202       MadeChange = true;
2203       I = NI;
2204     }
2205 
2206   // Commute binary operators, to canonicalize the order of their operands.
2207   // This can potentially expose more CSE opportunities, and makes writing other
2208   // transformations simpler.
2209   if (I->isCommutative())
2210     canonicalizeOperands(I);
2211 
2212   // Canonicalize negative constants out of expressions.
2213   if (Instruction *Res = canonicalizeNegFPConstants(I))
2214     I = Res;
2215 
2216   // Don't optimize floating-point instructions unless they have the
2217   // appropriate FastMathFlags for reassociation enabled.
2218   if (isa<FPMathOperator>(I) && !hasFPAssociativeFlags(I))
2219     return;
2220 
2221   // Do not reassociate boolean (i1) expressions.  We want to preserve the
2222   // original order of evaluation for short-circuited comparisons that
2223   // SimplifyCFG has folded to AND/OR expressions.  If the expression
2224   // is not further optimized, it is likely to be transformed back to a
2225   // short-circuited form for code gen, and the source order may have been
2226   // optimized for the most likely conditions.
2227   if (I->getType()->isIntegerTy(1))
2228     return;
2229 
2230   // If this is a bitwise or instruction of operands
2231   // with no common bits set, convert it to X+Y.
2232   if (I->getOpcode() == Instruction::Or &&
2233       shouldConvertOrWithNoCommonBitsToAdd(I) && !isLoadCombineCandidate(I) &&
2234       haveNoCommonBitsSet(I->getOperand(0), I->getOperand(1),
2235                           I->getModule()->getDataLayout(), /*AC=*/nullptr, I,
2236                           /*DT=*/nullptr)) {
2237     Instruction *NI = convertOrWithNoCommonBitsToAdd(I);
2238     RedoInsts.insert(I);
2239     MadeChange = true;
2240     I = NI;
2241   }
2242 
2243   // If this is a subtract instruction which is not already in negate form,
2244   // see if we can convert it to X+-Y.
2245   if (I->getOpcode() == Instruction::Sub) {
2246     if (ShouldBreakUpSubtract(I)) {
2247       Instruction *NI = BreakUpSubtract(I, RedoInsts);
2248       RedoInsts.insert(I);
2249       MadeChange = true;
2250       I = NI;
2251     } else if (match(I, m_Neg(m_Value()))) {
2252       // Otherwise, this is a negation.  See if the operand is a multiply tree
2253       // and if this is not an inner node of a multiply tree.
2254       if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2255           (!I->hasOneUse() ||
2256            !isReassociableOp(I->user_back(), Instruction::Mul))) {
2257         Instruction *NI = LowerNegateToMultiply(I);
2258         // If the negate was simplified, revisit the users to see if we can
2259         // reassociate further.
2260         for (User *U : NI->users()) {
2261           if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2262             RedoInsts.insert(Tmp);
2263         }
2264         RedoInsts.insert(I);
2265         MadeChange = true;
2266         I = NI;
2267       }
2268     }
2269   } else if (I->getOpcode() == Instruction::FNeg ||
2270              I->getOpcode() == Instruction::FSub) {
2271     if (ShouldBreakUpSubtract(I)) {
2272       Instruction *NI = BreakUpSubtract(I, RedoInsts);
2273       RedoInsts.insert(I);
2274       MadeChange = true;
2275       I = NI;
2276     } else if (match(I, m_FNeg(m_Value()))) {
2277       // Otherwise, this is a negation.  See if the operand is a multiply tree
2278       // and if this is not an inner node of a multiply tree.
2279       Value *Op = isa<BinaryOperator>(I) ? I->getOperand(1) :
2280                                            I->getOperand(0);
2281       if (isReassociableOp(Op, Instruction::FMul) &&
2282           (!I->hasOneUse() ||
2283            !isReassociableOp(I->user_back(), Instruction::FMul))) {
2284         // If the negate was simplified, revisit the users to see if we can
2285         // reassociate further.
2286         Instruction *NI = LowerNegateToMultiply(I);
2287         for (User *U : NI->users()) {
2288           if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2289             RedoInsts.insert(Tmp);
2290         }
2291         RedoInsts.insert(I);
2292         MadeChange = true;
2293         I = NI;
2294       }
2295     }
2296   }
2297 
2298   // If this instruction is an associative binary operator, process it.
2299   if (!I->isAssociative()) return;
2300   BinaryOperator *BO = cast<BinaryOperator>(I);
2301 
2302   // If this is an interior node of a reassociable tree, ignore it until we
2303   // get to the root of the tree, to avoid N^2 analysis.
2304   unsigned Opcode = BO->getOpcode();
2305   if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2306     // During the initial run we will get to the root of the tree.
2307     // But if we get here while we are redoing instructions, there is no
2308     // guarantee that the root will be visited. So Redo later
2309     if (BO->user_back() != BO &&
2310         BO->getParent() == BO->user_back()->getParent())
2311       RedoInsts.insert(BO->user_back());
2312     return;
2313   }
2314 
2315   // If this is an add tree that is used by a sub instruction, ignore it
2316   // until we process the subtract.
2317   if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2318       cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2319     return;
2320   if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2321       cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2322     return;
2323 
2324   ReassociateExpression(BO);
2325 }
2326 
2327 void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2328   // First, walk the expression tree, linearizing the tree, collecting the
2329   // operand information.
2330   SmallVector<RepeatedValue, 8> Tree;
2331   MadeChange |= LinearizeExprTree(I, Tree, RedoInsts);
2332   SmallVector<ValueEntry, 8> Ops;
2333   Ops.reserve(Tree.size());
2334   for (const RepeatedValue &E : Tree)
2335     Ops.append(E.second.getZExtValue(), ValueEntry(getRank(E.first), E.first));
2336 
2337   LLVM_DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2338 
2339   // Now that we have linearized the tree to a list and have gathered all of
2340   // the operands and their ranks, sort the operands by their rank.  Use a
2341   // stable_sort so that values with equal ranks will have their relative
2342   // positions maintained (and so the compiler is deterministic).  Note that
2343   // this sorts so that the highest ranking values end up at the beginning of
2344   // the vector.
2345   llvm::stable_sort(Ops);
2346 
2347   // Now that we have the expression tree in a convenient
2348   // sorted form, optimize it globally if possible.
2349   if (Value *V = OptimizeExpression(I, Ops)) {
2350     if (V == I)
2351       // Self-referential expression in unreachable code.
2352       return;
2353     // This expression tree simplified to something that isn't a tree,
2354     // eliminate it.
2355     LLVM_DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2356     I->replaceAllUsesWith(V);
2357     if (Instruction *VI = dyn_cast<Instruction>(V))
2358       if (I->getDebugLoc())
2359         VI->setDebugLoc(I->getDebugLoc());
2360     RedoInsts.insert(I);
2361     ++NumAnnihil;
2362     return;
2363   }
2364 
2365   // We want to sink immediates as deeply as possible except in the case where
2366   // this is a multiply tree used only by an add, and the immediate is a -1.
2367   // In this case we reassociate to put the negation on the outside so that we
2368   // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2369   if (I->hasOneUse()) {
2370     if (I->getOpcode() == Instruction::Mul &&
2371         cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2372         isa<ConstantInt>(Ops.back().Op) &&
2373         cast<ConstantInt>(Ops.back().Op)->isMinusOne()) {
2374       ValueEntry Tmp = Ops.pop_back_val();
2375       Ops.insert(Ops.begin(), Tmp);
2376     } else if (I->getOpcode() == Instruction::FMul &&
2377                cast<Instruction>(I->user_back())->getOpcode() ==
2378                    Instruction::FAdd &&
2379                isa<ConstantFP>(Ops.back().Op) &&
2380                cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2381       ValueEntry Tmp = Ops.pop_back_val();
2382       Ops.insert(Ops.begin(), Tmp);
2383     }
2384   }
2385 
2386   LLVM_DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2387 
2388   if (Ops.size() == 1) {
2389     if (Ops[0].Op == I)
2390       // Self-referential expression in unreachable code.
2391       return;
2392 
2393     // This expression tree simplified to something that isn't a tree,
2394     // eliminate it.
2395     I->replaceAllUsesWith(Ops[0].Op);
2396     if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2397       OI->setDebugLoc(I->getDebugLoc());
2398     RedoInsts.insert(I);
2399     return;
2400   }
2401 
2402   if (Ops.size() > 2 && Ops.size() <= GlobalReassociateLimit) {
2403     // Find the pair with the highest count in the pairmap and move it to the
2404     // back of the list so that it can later be CSE'd.
2405     // example:
2406     //   a*b*c*d*e
2407     // if c*e is the most "popular" pair, we can express this as
2408     //   (((c*e)*d)*b)*a
2409     unsigned Max = 1;
2410     unsigned BestRank = 0;
2411     std::pair<unsigned, unsigned> BestPair;
2412     unsigned Idx = I->getOpcode() - Instruction::BinaryOpsBegin;
2413     for (unsigned i = 0; i < Ops.size() - 1; ++i)
2414       for (unsigned j = i + 1; j < Ops.size(); ++j) {
2415         unsigned Score = 0;
2416         Value *Op0 = Ops[i].Op;
2417         Value *Op1 = Ops[j].Op;
2418         if (std::less<Value *>()(Op1, Op0))
2419           std::swap(Op0, Op1);
2420         auto it = PairMap[Idx].find({Op0, Op1});
2421         if (it != PairMap[Idx].end()) {
2422           // Functions like BreakUpSubtract() can erase the Values we're using
2423           // as keys and create new Values after we built the PairMap. There's a
2424           // small chance that the new nodes can have the same address as
2425           // something already in the table. We shouldn't accumulate the stored
2426           // score in that case as it refers to the wrong Value.
2427           if (it->second.isValid())
2428             Score += it->second.Score;
2429         }
2430 
2431         unsigned MaxRank = std::max(Ops[i].Rank, Ops[j].Rank);
2432         if (Score > Max || (Score == Max && MaxRank < BestRank)) {
2433           BestPair = {i, j};
2434           Max = Score;
2435           BestRank = MaxRank;
2436         }
2437       }
2438     if (Max > 1) {
2439       auto Op0 = Ops[BestPair.first];
2440       auto Op1 = Ops[BestPair.second];
2441       Ops.erase(&Ops[BestPair.second]);
2442       Ops.erase(&Ops[BestPair.first]);
2443       Ops.push_back(Op0);
2444       Ops.push_back(Op1);
2445     }
2446   }
2447   // Now that we ordered and optimized the expressions, splat them back into
2448   // the expression tree, removing any unneeded nodes.
2449   RewriteExprTree(I, Ops);
2450 }
2451 
2452 void
2453 ReassociatePass::BuildPairMap(ReversePostOrderTraversal<Function *> &RPOT) {
2454   // Make a "pairmap" of how often each operand pair occurs.
2455   for (BasicBlock *BI : RPOT) {
2456     for (Instruction &I : *BI) {
2457       if (!I.isAssociative())
2458         continue;
2459 
2460       // Ignore nodes that aren't at the root of trees.
2461       if (I.hasOneUse() && I.user_back()->getOpcode() == I.getOpcode())
2462         continue;
2463 
2464       // Collect all operands in a single reassociable expression.
2465       // Since Reassociate has already been run once, we can assume things
2466       // are already canonical according to Reassociation's regime.
2467       SmallVector<Value *, 8> Worklist = { I.getOperand(0), I.getOperand(1) };
2468       SmallVector<Value *, 8> Ops;
2469       while (!Worklist.empty() && Ops.size() <= GlobalReassociateLimit) {
2470         Value *Op = Worklist.pop_back_val();
2471         Instruction *OpI = dyn_cast<Instruction>(Op);
2472         if (!OpI || OpI->getOpcode() != I.getOpcode() || !OpI->hasOneUse()) {
2473           Ops.push_back(Op);
2474           continue;
2475         }
2476         // Be paranoid about self-referencing expressions in unreachable code.
2477         if (OpI->getOperand(0) != OpI)
2478           Worklist.push_back(OpI->getOperand(0));
2479         if (OpI->getOperand(1) != OpI)
2480           Worklist.push_back(OpI->getOperand(1));
2481       }
2482       // Skip extremely long expressions.
2483       if (Ops.size() > GlobalReassociateLimit)
2484         continue;
2485 
2486       // Add all pairwise combinations of operands to the pair map.
2487       unsigned BinaryIdx = I.getOpcode() - Instruction::BinaryOpsBegin;
2488       SmallSet<std::pair<Value *, Value*>, 32> Visited;
2489       for (unsigned i = 0; i < Ops.size() - 1; ++i) {
2490         for (unsigned j = i + 1; j < Ops.size(); ++j) {
2491           // Canonicalize operand orderings.
2492           Value *Op0 = Ops[i];
2493           Value *Op1 = Ops[j];
2494           if (std::less<Value *>()(Op1, Op0))
2495             std::swap(Op0, Op1);
2496           if (!Visited.insert({Op0, Op1}).second)
2497             continue;
2498           auto res = PairMap[BinaryIdx].insert({{Op0, Op1}, {Op0, Op1, 1}});
2499           if (!res.second) {
2500             // If either key value has been erased then we've got the same
2501             // address by coincidence. That can't happen here because nothing is
2502             // erasing values but it can happen by the time we're querying the
2503             // map.
2504             assert(res.first->second.isValid() && "WeakVH invalidated");
2505             ++res.first->second.Score;
2506           }
2507         }
2508       }
2509     }
2510   }
2511 }
2512 
2513 PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) {
2514   // Get the functions basic blocks in Reverse Post Order. This order is used by
2515   // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic
2516   // blocks (it has been seen that the analysis in this pass could hang when
2517   // analysing dead basic blocks).
2518   ReversePostOrderTraversal<Function *> RPOT(&F);
2519 
2520   // Calculate the rank map for F.
2521   BuildRankMap(F, RPOT);
2522 
2523   // Build the pair map before running reassociate.
2524   // Technically this would be more accurate if we did it after one round
2525   // of reassociation, but in practice it doesn't seem to help much on
2526   // real-world code, so don't waste the compile time running reassociate
2527   // twice.
2528   // If a user wants, they could expicitly run reassociate twice in their
2529   // pass pipeline for further potential gains.
2530   // It might also be possible to update the pair map during runtime, but the
2531   // overhead of that may be large if there's many reassociable chains.
2532   BuildPairMap(RPOT);
2533 
2534   MadeChange = false;
2535 
2536   // Traverse the same blocks that were analysed by BuildRankMap.
2537   for (BasicBlock *BI : RPOT) {
2538     assert(RankMap.count(&*BI) && "BB should be ranked.");
2539     // Optimize every instruction in the basic block.
2540     for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;)
2541       if (isInstructionTriviallyDead(&*II)) {
2542         EraseInst(&*II++);
2543       } else {
2544         OptimizeInst(&*II);
2545         assert(II->getParent() == &*BI && "Moved to a different block!");
2546         ++II;
2547       }
2548 
2549     // Make a copy of all the instructions to be redone so we can remove dead
2550     // instructions.
2551     OrderedSet ToRedo(RedoInsts);
2552     // Iterate over all instructions to be reevaluated and remove trivially dead
2553     // instructions. If any operand of the trivially dead instruction becomes
2554     // dead mark it for deletion as well. Continue this process until all
2555     // trivially dead instructions have been removed.
2556     while (!ToRedo.empty()) {
2557       Instruction *I = ToRedo.pop_back_val();
2558       if (isInstructionTriviallyDead(I)) {
2559         RecursivelyEraseDeadInsts(I, ToRedo);
2560         MadeChange = true;
2561       }
2562     }
2563 
2564     // Now that we have removed dead instructions, we can reoptimize the
2565     // remaining instructions.
2566     while (!RedoInsts.empty()) {
2567       Instruction *I = RedoInsts.front();
2568       RedoInsts.erase(RedoInsts.begin());
2569       if (isInstructionTriviallyDead(I))
2570         EraseInst(I);
2571       else
2572         OptimizeInst(I);
2573     }
2574   }
2575 
2576   // We are done with the rank map and pair map.
2577   RankMap.clear();
2578   ValueRankMap.clear();
2579   for (auto &Entry : PairMap)
2580     Entry.clear();
2581 
2582   if (MadeChange) {
2583     PreservedAnalyses PA;
2584     PA.preserveSet<CFGAnalyses>();
2585     return PA;
2586   }
2587 
2588   return PreservedAnalyses::all();
2589 }
2590 
2591 namespace {
2592 
2593   class ReassociateLegacyPass : public FunctionPass {
2594     ReassociatePass Impl;
2595 
2596   public:
2597     static char ID; // Pass identification, replacement for typeid
2598 
2599     ReassociateLegacyPass() : FunctionPass(ID) {
2600       initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry());
2601     }
2602 
2603     bool runOnFunction(Function &F) override {
2604       if (skipFunction(F))
2605         return false;
2606 
2607       FunctionAnalysisManager DummyFAM;
2608       auto PA = Impl.run(F, DummyFAM);
2609       return !PA.areAllPreserved();
2610     }
2611 
2612     void getAnalysisUsage(AnalysisUsage &AU) const override {
2613       AU.setPreservesCFG();
2614       AU.addPreserved<AAResultsWrapperPass>();
2615       AU.addPreserved<BasicAAWrapperPass>();
2616       AU.addPreserved<GlobalsAAWrapperPass>();
2617     }
2618   };
2619 
2620 } // end anonymous namespace
2621 
2622 char ReassociateLegacyPass::ID = 0;
2623 
2624 INITIALIZE_PASS(ReassociateLegacyPass, "reassociate",
2625                 "Reassociate expressions", false, false)
2626 
2627 // Public interface to the Reassociate pass
2628 FunctionPass *llvm::createReassociatePass() {
2629   return new ReassociateLegacyPass();
2630 }
2631