1 //===- Reassociate.cpp - Reassociate binary expressions -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass reassociates commutative expressions in an order that is designed
11 // to promote better constant propagation, GCSE, LICM, PRE, etc.
13 // For example: 4 + (x + 5) -> x + (4 + 5)
15 // In the implementation of this algorithm, constants are assigned rank = 0,
16 // function arguments are rank = 1, and other values are assigned ranks
17 // corresponding to the reverse post order traversal of current function
18 // (starting at 2), which effectively gives values in deep loops higher rank
19 // than values not in loops.
21 //===----------------------------------------------------------------------===//
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/ADT/DenseMap.h"
25 #include "llvm/ADT/PostOrderIterator.h"
26 #include "llvm/ADT/STLExtras.h"
27 #include "llvm/ADT/SetVector.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/Analysis/GlobalsModRef.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/CFG.h"
32 #include "llvm/IR/Constants.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/IR/IRBuilder.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/ValueHandle.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Transforms/Utils/Local.h"
46 #define DEBUG_TYPE "reassociate"
48 STATISTIC(NumChanged, "Number of insts reassociated");
49 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
50 STATISTIC(NumFactor , "Number of multiplies factored");
56 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {}
58 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) {
59 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start.
64 /// Print out the expression identified in the Ops list.
66 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
67 Module *M = I->getParent()->getParent()->getParent();
68 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
69 << *Ops[0].Op->getType() << '\t';
70 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
72 Ops[i].Op->printAsOperand(dbgs(), false, M);
73 dbgs() << ", #" << Ops[i].Rank << "] ";
79 /// \brief Utility class representing a base and exponent pair which form one
80 /// factor of some product.
85 Factor(Value *Base, unsigned Power) : Base(Base), Power(Power) {}
87 /// \brief Sort factors by their Base.
89 bool operator()(const Factor &LHS, const Factor &RHS) {
90 return LHS.Base < RHS.Base;
94 /// \brief Compare factors for equal bases.
96 bool operator()(const Factor &LHS, const Factor &RHS) {
97 return LHS.Base == RHS.Base;
101 /// \brief Sort factors in descending order by their power.
102 struct PowerDescendingSorter {
103 bool operator()(const Factor &LHS, const Factor &RHS) {
104 return LHS.Power > RHS.Power;
108 /// \brief Compare factors for equal powers.
110 bool operator()(const Factor &LHS, const Factor &RHS) {
111 return LHS.Power == RHS.Power;
116 /// Utility class representing a non-constant Xor-operand. We classify
117 /// non-constant Xor-Operands into two categories:
118 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0
120 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
122 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
123 /// operand as "E | 0"
128 bool isInvalid() const { return SymbolicPart == nullptr; }
129 bool isOrExpr() const { return isOr; }
130 Value *getValue() const { return OrigVal; }
131 Value *getSymbolicPart() const { return SymbolicPart; }
132 unsigned getSymbolicRank() const { return SymbolicRank; }
133 const APInt &getConstPart() const { return ConstPart; }
135 void Invalidate() { SymbolicPart = OrigVal = nullptr; }
136 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
138 // Sort the XorOpnd-Pointer in ascending order of symbolic-value-rank.
139 // The purpose is twofold:
140 // 1) Cluster together the operands sharing the same symbolic-value.
141 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
142 // could potentially shorten crital path, and expose more loop-invariants.
143 // Note that values' rank are basically defined in RPO order (FIXME).
144 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
145 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
146 // "z" in the order of X-Y-Z is better than any other orders.
147 struct PtrSortFunctor {
148 bool operator()(XorOpnd * const &LHS, XorOpnd * const &RHS) {
149 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
156 unsigned SymbolicRank;
162 class Reassociate : public FunctionPass {
163 DenseMap<BasicBlock*, unsigned> RankMap;
164 DenseMap<AssertingVH<Value>, unsigned> ValueRankMap;
165 SetVector<AssertingVH<Instruction> > RedoInsts;
168 static char ID; // Pass identification, replacement for typeid
169 Reassociate() : FunctionPass(ID) {
170 initializeReassociatePass(*PassRegistry::getPassRegistry());
173 bool runOnFunction(Function &F) override;
175 void getAnalysisUsage(AnalysisUsage &AU) const override {
176 AU.setPreservesCFG();
177 AU.addPreserved<GlobalsAAWrapperPass>();
180 void BuildRankMap(Function &F);
181 unsigned getRank(Value *V);
182 void canonicalizeOperands(Instruction *I);
183 void ReassociateExpression(BinaryOperator *I);
184 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
185 Value *OptimizeExpression(BinaryOperator *I,
186 SmallVectorImpl<ValueEntry> &Ops);
187 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
188 Value *OptimizeXor(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
189 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, APInt &ConstOpnd,
191 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
192 APInt &ConstOpnd, Value *&Res);
193 bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
194 SmallVectorImpl<Factor> &Factors);
195 Value *buildMinimalMultiplyDAG(IRBuilder<> &Builder,
196 SmallVectorImpl<Factor> &Factors);
197 Value *OptimizeMul(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
198 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
199 void EraseInst(Instruction *I);
200 void OptimizeInst(Instruction *I);
201 Instruction *canonicalizeNegConstExpr(Instruction *I);
205 XorOpnd::XorOpnd(Value *V) {
206 assert(!isa<ConstantInt>(V) && "No ConstantInt");
208 Instruction *I = dyn_cast<Instruction>(V);
211 if (I && (I->getOpcode() == Instruction::Or ||
212 I->getOpcode() == Instruction::And)) {
213 Value *V0 = I->getOperand(0);
214 Value *V1 = I->getOperand(1);
215 if (isa<ConstantInt>(V0))
218 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) {
219 ConstPart = C->getValue();
221 isOr = (I->getOpcode() == Instruction::Or);
226 // view the operand as "V | 0"
228 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth());
232 char Reassociate::ID = 0;
233 INITIALIZE_PASS(Reassociate, "reassociate",
234 "Reassociate expressions", false, false)
236 // Public interface to the Reassociate pass
237 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
239 /// Return true if V is an instruction of the specified opcode and if it
240 /// only has one use.
241 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
242 if (V->hasOneUse() && isa<Instruction>(V) &&
243 cast<Instruction>(V)->getOpcode() == Opcode &&
244 (!isa<FPMathOperator>(V) ||
245 cast<Instruction>(V)->hasUnsafeAlgebra()))
246 return cast<BinaryOperator>(V);
250 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
252 if (V->hasOneUse() && isa<Instruction>(V) &&
253 (cast<Instruction>(V)->getOpcode() == Opcode1 ||
254 cast<Instruction>(V)->getOpcode() == Opcode2) &&
255 (!isa<FPMathOperator>(V) ||
256 cast<Instruction>(V)->hasUnsafeAlgebra()))
257 return cast<BinaryOperator>(V);
261 void Reassociate::BuildRankMap(Function &F) {
264 // Assign distinct ranks to function arguments.
265 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
266 ValueRankMap[&*I] = ++i;
267 DEBUG(dbgs() << "Calculated Rank[" << I->getName() << "] = " << i << "\n");
270 ReversePostOrderTraversal<Function*> RPOT(&F);
271 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
272 E = RPOT.end(); I != E; ++I) {
274 unsigned BBRank = RankMap[BB] = ++i << 16;
276 // Walk the basic block, adding precomputed ranks for any instructions that
277 // we cannot move. This ensures that the ranks for these instructions are
278 // all different in the block.
279 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
280 if (mayBeMemoryDependent(*I))
281 ValueRankMap[&*I] = ++BBRank;
285 unsigned Reassociate::getRank(Value *V) {
286 Instruction *I = dyn_cast<Instruction>(V);
288 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
289 return 0; // Otherwise it's a global or constant, rank 0.
292 if (unsigned Rank = ValueRankMap[I])
293 return Rank; // Rank already known?
295 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
296 // we can reassociate expressions for code motion! Since we do not recurse
297 // for PHI nodes, we cannot have infinite recursion here, because there
298 // cannot be loops in the value graph that do not go through PHI nodes.
299 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
300 for (unsigned i = 0, e = I->getNumOperands();
301 i != e && Rank != MaxRank; ++i)
302 Rank = std::max(Rank, getRank(I->getOperand(i)));
304 // If this is a not or neg instruction, do not count it for rank. This
305 // assures us that X and ~X will have the same rank.
306 if (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
307 !BinaryOperator::isFNeg(I))
310 DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
312 return ValueRankMap[I] = Rank;
315 // Canonicalize constants to RHS. Otherwise, sort the operands by rank.
316 void Reassociate::canonicalizeOperands(Instruction *I) {
317 assert(isa<BinaryOperator>(I) && "Expected binary operator.");
318 assert(I->isCommutative() && "Expected commutative operator.");
320 Value *LHS = I->getOperand(0);
321 Value *RHS = I->getOperand(1);
322 unsigned LHSRank = getRank(LHS);
323 unsigned RHSRank = getRank(RHS);
325 if (isa<Constant>(RHS))
328 if (isa<Constant>(LHS) || RHSRank < LHSRank)
329 cast<BinaryOperator>(I)->swapOperands();
332 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
333 Instruction *InsertBefore, Value *FlagsOp) {
334 if (S1->getType()->isIntOrIntVectorTy())
335 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
337 BinaryOperator *Res =
338 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
339 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
344 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
345 Instruction *InsertBefore, Value *FlagsOp) {
346 if (S1->getType()->isIntOrIntVectorTy())
347 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
349 BinaryOperator *Res =
350 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
351 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
356 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
357 Instruction *InsertBefore, Value *FlagsOp) {
358 if (S1->getType()->isIntOrIntVectorTy())
359 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
361 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
362 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
367 /// Replace 0-X with X*-1.
368 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
369 Type *Ty = Neg->getType();
370 Constant *NegOne = Ty->isIntOrIntVectorTy() ?
371 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
373 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
374 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
376 Neg->replaceAllUsesWith(Res);
377 Res->setDebugLoc(Neg->getDebugLoc());
381 /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
382 /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for
383 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
384 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
385 /// even x in Bitwidth-bit arithmetic.
386 static unsigned CarmichaelShift(unsigned Bitwidth) {
392 /// Add the extra weight 'RHS' to the existing weight 'LHS',
393 /// reducing the combined weight using any special properties of the operation.
394 /// The existing weight LHS represents the computation X op X op ... op X where
395 /// X occurs LHS times. The combined weight represents X op X op ... op X with
396 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
397 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
398 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
399 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
400 // If we were working with infinite precision arithmetic then the combined
401 // weight would be LHS + RHS. But we are using finite precision arithmetic,
402 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
403 // for nilpotent operations and addition, but not for idempotent operations
404 // and multiplication), so it is important to correctly reduce the combined
405 // weight back into range if wrapping would be wrong.
407 // If RHS is zero then the weight didn't change.
408 if (RHS.isMinValue())
410 // If LHS is zero then the combined weight is RHS.
411 if (LHS.isMinValue()) {
415 // From this point on we know that neither LHS nor RHS is zero.
417 if (Instruction::isIdempotent(Opcode)) {
418 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
419 // weight of 1. Keeping weights at zero or one also means that wrapping is
421 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
422 return; // Return a weight of 1.
424 if (Instruction::isNilpotent(Opcode)) {
425 // Nilpotent means X op X === 0, so reduce weights modulo 2.
426 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
427 LHS = 0; // 1 + 1 === 0 modulo 2.
430 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
431 // TODO: Reduce the weight by exploiting nsw/nuw?
436 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
437 "Unknown associative operation!");
438 unsigned Bitwidth = LHS.getBitWidth();
439 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
440 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
441 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
442 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
443 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
444 // which by a happy accident means that they can always be represented using
446 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
447 // the Carmichael number).
449 /// CM - The value of Carmichael's lambda function.
450 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
451 // Any weight W >= Threshold can be replaced with W - CM.
452 APInt Threshold = CM + Bitwidth;
453 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
454 // For Bitwidth 4 or more the following sum does not overflow.
456 while (LHS.uge(Threshold))
459 // To avoid problems with overflow do everything the same as above but using
461 unsigned CM = 1U << CarmichaelShift(Bitwidth);
462 unsigned Threshold = CM + Bitwidth;
463 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
464 "Weights not reduced!");
465 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
466 while (Total >= Threshold)
472 typedef std::pair<Value*, APInt> RepeatedValue;
474 /// Given an associative binary expression, return the leaf
475 /// nodes in Ops along with their weights (how many times the leaf occurs). The
476 /// original expression is the same as
477 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
479 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
483 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
485 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
487 /// This routine may modify the function, in which case it returns 'true'. The
488 /// changes it makes may well be destructive, changing the value computed by 'I'
489 /// to something completely different. Thus if the routine returns 'true' then
490 /// you MUST either replace I with a new expression computed from the Ops array,
491 /// or use RewriteExprTree to put the values back in.
493 /// A leaf node is either not a binary operation of the same kind as the root
494 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
495 /// opcode), or is the same kind of binary operator but has a use which either
496 /// does not belong to the expression, or does belong to the expression but is
497 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
498 /// of the expression, while for non-leaf nodes (except for the root 'I') every
499 /// use is a non-leaf node of the expression.
502 /// expression graph node names
512 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
513 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
515 /// The expression is maximal: if some instruction is a binary operator of the
516 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
517 /// then the instruction also belongs to the expression, is not a leaf node of
518 /// it, and its operands also belong to the expression (but may be leaf nodes).
520 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
521 /// order to ensure that every non-root node in the expression has *exactly one*
522 /// use by a non-leaf node of the expression. This destruction means that the
523 /// caller MUST either replace 'I' with a new expression or use something like
524 /// RewriteExprTree to put the values back in if the routine indicates that it
525 /// made a change by returning 'true'.
527 /// In the above example either the right operand of A or the left operand of B
528 /// will be replaced by undef. If it is B's operand then this gives:
532 /// + + | A, B - operand of B replaced with undef
538 /// Note that such undef operands can only be reached by passing through 'I'.
539 /// For example, if you visit operands recursively starting from a leaf node
540 /// then you will never see such an undef operand unless you get back to 'I',
541 /// which requires passing through a phi node.
543 /// Note that this routine may also mutate binary operators of the wrong type
544 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
545 /// of the expression) if it can turn them into binary operators of the right
546 /// type and thus make the expression bigger.
548 static bool LinearizeExprTree(BinaryOperator *I,
549 SmallVectorImpl<RepeatedValue> &Ops) {
550 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
551 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
552 unsigned Opcode = I->getOpcode();
553 assert(I->isAssociative() && I->isCommutative() &&
554 "Expected an associative and commutative operation!");
556 // Visit all operands of the expression, keeping track of their weight (the
557 // number of paths from the expression root to the operand, or if you like
558 // the number of times that operand occurs in the linearized expression).
559 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
560 // while A has weight two.
562 // Worklist of non-leaf nodes (their operands are in the expression too) along
563 // with their weights, representing a certain number of paths to the operator.
564 // If an operator occurs in the worklist multiple times then we found multiple
565 // ways to get to it.
566 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
567 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
568 bool Changed = false;
570 // Leaves of the expression are values that either aren't the right kind of
571 // operation (eg: a constant, or a multiply in an add tree), or are, but have
572 // some uses that are not inside the expression. For example, in I = X + X,
573 // X = A + B, the value X has two uses (by I) that are in the expression. If
574 // X has any other uses, for example in a return instruction, then we consider
575 // X to be a leaf, and won't analyze it further. When we first visit a value,
576 // if it has more than one use then at first we conservatively consider it to
577 // be a leaf. Later, as the expression is explored, we may discover some more
578 // uses of the value from inside the expression. If all uses turn out to be
579 // from within the expression (and the value is a binary operator of the right
580 // kind) then the value is no longer considered to be a leaf, and its operands
583 // Leaves - Keeps track of the set of putative leaves as well as the number of
584 // paths to each leaf seen so far.
585 typedef DenseMap<Value*, APInt> LeafMap;
586 LeafMap Leaves; // Leaf -> Total weight so far.
587 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
590 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
592 while (!Worklist.empty()) {
593 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
594 I = P.first; // We examine the operands of this binary operator.
596 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
597 Value *Op = I->getOperand(OpIdx);
598 APInt Weight = P.second; // Number of paths to this operand.
599 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
600 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
602 // If this is a binary operation of the right kind with only one use then
603 // add its operands to the expression.
604 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
605 assert(Visited.insert(Op).second && "Not first visit!");
606 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
607 Worklist.push_back(std::make_pair(BO, Weight));
611 // Appears to be a leaf. Is the operand already in the set of leaves?
612 LeafMap::iterator It = Leaves.find(Op);
613 if (It == Leaves.end()) {
614 // Not in the leaf map. Must be the first time we saw this operand.
615 assert(Visited.insert(Op).second && "Not first visit!");
616 if (!Op->hasOneUse()) {
617 // This value has uses not accounted for by the expression, so it is
618 // not safe to modify. Mark it as being a leaf.
619 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
620 LeafOrder.push_back(Op);
624 // No uses outside the expression, try morphing it.
625 } else if (It != Leaves.end()) {
626 // Already in the leaf map.
627 assert(Visited.count(Op) && "In leaf map but not visited!");
629 // Update the number of paths to the leaf.
630 IncorporateWeight(It->second, Weight, Opcode);
632 #if 0 // TODO: Re-enable once PR13021 is fixed.
633 // The leaf already has one use from inside the expression. As we want
634 // exactly one such use, drop this new use of the leaf.
635 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
636 I->setOperand(OpIdx, UndefValue::get(I->getType()));
639 // If the leaf is a binary operation of the right kind and we now see
640 // that its multiple original uses were in fact all by nodes belonging
641 // to the expression, then no longer consider it to be a leaf and add
642 // its operands to the expression.
643 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
644 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
645 Worklist.push_back(std::make_pair(BO, It->second));
651 // If we still have uses that are not accounted for by the expression
652 // then it is not safe to modify the value.
653 if (!Op->hasOneUse())
656 // No uses outside the expression, try morphing it.
658 Leaves.erase(It); // Since the value may be morphed below.
661 // At this point we have a value which, first of all, is not a binary
662 // expression of the right kind, and secondly, is only used inside the
663 // expression. This means that it can safely be modified. See if we
664 // can usefully morph it into an expression of the right kind.
665 assert((!isa<Instruction>(Op) ||
666 cast<Instruction>(Op)->getOpcode() != Opcode
667 || (isa<FPMathOperator>(Op) &&
668 !cast<Instruction>(Op)->hasUnsafeAlgebra())) &&
669 "Should have been handled above!");
670 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
672 // If this is a multiply expression, turn any internal negations into
673 // multiplies by -1 so they can be reassociated.
674 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
675 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
676 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
677 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
678 BO = LowerNegateToMultiply(BO);
679 DEBUG(dbgs() << *BO << '\n');
680 Worklist.push_back(std::make_pair(BO, Weight));
685 // Failed to morph into an expression of the right type. This really is
687 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
688 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
689 LeafOrder.push_back(Op);
694 // The leaves, repeated according to their weights, represent the linearized
695 // form of the expression.
696 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
697 Value *V = LeafOrder[i];
698 LeafMap::iterator It = Leaves.find(V);
699 if (It == Leaves.end())
700 // Node initially thought to be a leaf wasn't.
702 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
703 APInt Weight = It->second;
704 if (Weight.isMinValue())
705 // Leaf already output or weight reduction eliminated it.
707 // Ensure the leaf is only output once.
709 Ops.push_back(std::make_pair(V, Weight));
712 // For nilpotent operations or addition there may be no operands, for example
713 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
714 // in both cases the weight reduces to 0 causing the value to be skipped.
716 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
717 assert(Identity && "Associative operation without identity!");
718 Ops.emplace_back(Identity, APInt(Bitwidth, 1));
724 /// Now that the operands for this expression tree are
725 /// linearized and optimized, emit them in-order.
726 void Reassociate::RewriteExprTree(BinaryOperator *I,
727 SmallVectorImpl<ValueEntry> &Ops) {
728 assert(Ops.size() > 1 && "Single values should be used directly!");
730 // Since our optimizations should never increase the number of operations, the
731 // new expression can usually be written reusing the existing binary operators
732 // from the original expression tree, without creating any new instructions,
733 // though the rewritten expression may have a completely different topology.
734 // We take care to not change anything if the new expression will be the same
735 // as the original. If more than trivial changes (like commuting operands)
736 // were made then we are obliged to clear out any optional subclass data like
739 /// NodesToRewrite - Nodes from the original expression available for writing
740 /// the new expression into.
741 SmallVector<BinaryOperator*, 8> NodesToRewrite;
742 unsigned Opcode = I->getOpcode();
743 BinaryOperator *Op = I;
745 /// NotRewritable - The operands being written will be the leaves of the new
746 /// expression and must not be used as inner nodes (via NodesToRewrite) by
747 /// mistake. Inner nodes are always reassociable, and usually leaves are not
748 /// (if they were they would have been incorporated into the expression and so
749 /// would not be leaves), so most of the time there is no danger of this. But
750 /// in rare cases a leaf may become reassociable if an optimization kills uses
751 /// of it, or it may momentarily become reassociable during rewriting (below)
752 /// due it being removed as an operand of one of its uses. Ensure that misuse
753 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
754 /// leaves and refusing to reuse any of them as inner nodes.
755 SmallPtrSet<Value*, 8> NotRewritable;
756 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
757 NotRewritable.insert(Ops[i].Op);
759 // ExpressionChanged - Non-null if the rewritten expression differs from the
760 // original in some non-trivial way, requiring the clearing of optional flags.
761 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
762 BinaryOperator *ExpressionChanged = nullptr;
763 for (unsigned i = 0; ; ++i) {
764 // The last operation (which comes earliest in the IR) is special as both
765 // operands will come from Ops, rather than just one with the other being
767 if (i+2 == Ops.size()) {
768 Value *NewLHS = Ops[i].Op;
769 Value *NewRHS = Ops[i+1].Op;
770 Value *OldLHS = Op->getOperand(0);
771 Value *OldRHS = Op->getOperand(1);
773 if (NewLHS == OldLHS && NewRHS == OldRHS)
774 // Nothing changed, leave it alone.
777 if (NewLHS == OldRHS && NewRHS == OldLHS) {
778 // The order of the operands was reversed. Swap them.
779 DEBUG(dbgs() << "RA: " << *Op << '\n');
781 DEBUG(dbgs() << "TO: " << *Op << '\n');
787 // The new operation differs non-trivially from the original. Overwrite
788 // the old operands with the new ones.
789 DEBUG(dbgs() << "RA: " << *Op << '\n');
790 if (NewLHS != OldLHS) {
791 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
792 if (BO && !NotRewritable.count(BO))
793 NodesToRewrite.push_back(BO);
794 Op->setOperand(0, NewLHS);
796 if (NewRHS != OldRHS) {
797 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
798 if (BO && !NotRewritable.count(BO))
799 NodesToRewrite.push_back(BO);
800 Op->setOperand(1, NewRHS);
802 DEBUG(dbgs() << "TO: " << *Op << '\n');
804 ExpressionChanged = Op;
811 // Not the last operation. The left-hand side will be a sub-expression
812 // while the right-hand side will be the current element of Ops.
813 Value *NewRHS = Ops[i].Op;
814 if (NewRHS != Op->getOperand(1)) {
815 DEBUG(dbgs() << "RA: " << *Op << '\n');
816 if (NewRHS == Op->getOperand(0)) {
817 // The new right-hand side was already present as the left operand. If
818 // we are lucky then swapping the operands will sort out both of them.
821 // Overwrite with the new right-hand side.
822 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
823 if (BO && !NotRewritable.count(BO))
824 NodesToRewrite.push_back(BO);
825 Op->setOperand(1, NewRHS);
826 ExpressionChanged = Op;
828 DEBUG(dbgs() << "TO: " << *Op << '\n');
833 // Now deal with the left-hand side. If this is already an operation node
834 // from the original expression then just rewrite the rest of the expression
836 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
837 if (BO && !NotRewritable.count(BO)) {
842 // Otherwise, grab a spare node from the original expression and use that as
843 // the left-hand side. If there are no nodes left then the optimizers made
844 // an expression with more nodes than the original! This usually means that
845 // they did something stupid but it might mean that the problem was just too
846 // hard (finding the mimimal number of multiplications needed to realize a
847 // multiplication expression is NP-complete). Whatever the reason, smart or
848 // stupid, create a new node if there are none left.
849 BinaryOperator *NewOp;
850 if (NodesToRewrite.empty()) {
851 Constant *Undef = UndefValue::get(I->getType());
852 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
853 Undef, Undef, "", I);
854 if (NewOp->getType()->isFPOrFPVectorTy())
855 NewOp->setFastMathFlags(I->getFastMathFlags());
857 NewOp = NodesToRewrite.pop_back_val();
860 DEBUG(dbgs() << "RA: " << *Op << '\n');
861 Op->setOperand(0, NewOp);
862 DEBUG(dbgs() << "TO: " << *Op << '\n');
863 ExpressionChanged = Op;
869 // If the expression changed non-trivially then clear out all subclass data
870 // starting from the operator specified in ExpressionChanged, and compactify
871 // the operators to just before the expression root to guarantee that the
872 // expression tree is dominated by all of Ops.
873 if (ExpressionChanged)
875 // Preserve FastMathFlags.
876 if (isa<FPMathOperator>(I)) {
877 FastMathFlags Flags = I->getFastMathFlags();
878 ExpressionChanged->clearSubclassOptionalData();
879 ExpressionChanged->setFastMathFlags(Flags);
881 ExpressionChanged->clearSubclassOptionalData();
883 if (ExpressionChanged == I)
885 ExpressionChanged->moveBefore(I);
886 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
889 // Throw away any left over nodes from the original expression.
890 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
891 RedoInsts.insert(NodesToRewrite[i]);
894 /// Insert instructions before the instruction pointed to by BI,
895 /// that computes the negative version of the value specified. The negative
896 /// version of the value is returned, and BI is left pointing at the instruction
897 /// that should be processed next by the reassociation pass.
898 static Value *NegateValue(Value *V, Instruction *BI) {
899 if (Constant *C = dyn_cast<Constant>(V)) {
900 if (C->getType()->isFPOrFPVectorTy()) {
901 return ConstantExpr::getFNeg(C);
903 return ConstantExpr::getNeg(C);
907 // We are trying to expose opportunity for reassociation. One of the things
908 // that we want to do to achieve this is to push a negation as deep into an
909 // expression chain as possible, to expose the add instructions. In practice,
910 // this means that we turn this:
911 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
912 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
913 // the constants. We assume that instcombine will clean up the mess later if
914 // we introduce tons of unnecessary negation instructions.
916 if (BinaryOperator *I =
917 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
918 // Push the negates through the add.
919 I->setOperand(0, NegateValue(I->getOperand(0), BI));
920 I->setOperand(1, NegateValue(I->getOperand(1), BI));
921 if (I->getOpcode() == Instruction::Add) {
922 I->setHasNoUnsignedWrap(false);
923 I->setHasNoSignedWrap(false);
926 // We must move the add instruction here, because the neg instructions do
927 // not dominate the old add instruction in general. By moving it, we are
928 // assured that the neg instructions we just inserted dominate the
929 // instruction we are about to insert after them.
932 I->setName(I->getName()+".neg");
936 // Okay, we need to materialize a negated version of V with an instruction.
937 // Scan the use lists of V to see if we have one already.
938 for (User *U : V->users()) {
939 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
942 // We found one! Now we have to make sure that the definition dominates
943 // this use. We do this by moving it to the entry block (if it is a
944 // non-instruction value) or right after the definition. These negates will
945 // be zapped by reassociate later, so we don't need much finesse here.
946 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
948 // Verify that the negate is in this function, V might be a constant expr.
949 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
952 BasicBlock::iterator InsertPt;
953 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
954 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
955 InsertPt = II->getNormalDest()->begin();
956 } else if (auto *CPI = dyn_cast<CatchPadInst>(InstInput)) {
957 InsertPt = CPI->getNormalDest()->begin();
959 InsertPt = InstInput;
962 while (isa<PHINode>(InsertPt)) ++InsertPt;
964 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
966 TheNeg->moveBefore(InsertPt);
967 if (TheNeg->getOpcode() == Instruction::Sub) {
968 TheNeg->setHasNoUnsignedWrap(false);
969 TheNeg->setHasNoSignedWrap(false);
971 TheNeg->andIRFlags(BI);
976 // Insert a 'neg' instruction that subtracts the value from zero to get the
978 return CreateNeg(V, V->getName() + ".neg", BI, BI);
981 /// Return true if we should break up this subtract of X-Y into (X + -Y).
982 static bool ShouldBreakUpSubtract(Instruction *Sub) {
983 // If this is a negation, we can't split it up!
984 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
987 // Don't breakup X - undef.
988 if (isa<UndefValue>(Sub->getOperand(1)))
991 // Don't bother to break this up unless either the LHS is an associable add or
992 // subtract or if this is only used by one.
993 Value *V0 = Sub->getOperand(0);
994 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
995 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
997 Value *V1 = Sub->getOperand(1);
998 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
999 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
1001 Value *VB = Sub->user_back();
1002 if (Sub->hasOneUse() &&
1003 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
1004 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
1010 /// If we have (X-Y), and if either X is an add, or if this is only used by an
1011 /// add, transform this into (X+(0-Y)) to promote better reassociation.
1012 static BinaryOperator *BreakUpSubtract(Instruction *Sub) {
1013 // Convert a subtract into an add and a neg instruction. This allows sub
1014 // instructions to be commuted with other add instructions.
1016 // Calculate the negative value of Operand 1 of the sub instruction,
1017 // and set it as the RHS of the add instruction we just made.
1019 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
1020 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
1021 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
1022 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
1025 // Everyone now refers to the add instruction.
1026 Sub->replaceAllUsesWith(New);
1027 New->setDebugLoc(Sub->getDebugLoc());
1029 DEBUG(dbgs() << "Negated: " << *New << '\n');
1033 /// If this is a shift of a reassociable multiply or is used by one, change
1034 /// this into a multiply by a constant to assist with further reassociation.
1035 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1036 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
1037 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
1039 BinaryOperator *Mul =
1040 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
1041 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
1044 // Everyone now refers to the mul instruction.
1045 Shl->replaceAllUsesWith(Mul);
1046 Mul->setDebugLoc(Shl->getDebugLoc());
1048 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
1049 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
1051 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
1052 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
1054 Mul->setHasNoSignedWrap(true);
1055 Mul->setHasNoUnsignedWrap(NUW);
1059 /// Scan backwards and forwards among values with the same rank as element i
1060 /// to see if X exists. If X does not exist, return i. This is useful when
1061 /// scanning for 'x' when we see '-x' because they both get the same rank.
1062 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
1064 unsigned XRank = Ops[i].Rank;
1065 unsigned e = Ops.size();
1066 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
1069 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1070 if (Instruction *I2 = dyn_cast<Instruction>(X))
1071 if (I1->isIdenticalTo(I2))
1075 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
1078 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1079 if (Instruction *I2 = dyn_cast<Instruction>(X))
1080 if (I1->isIdenticalTo(I2))
1086 /// Emit a tree of add instructions, summing Ops together
1087 /// and returning the result. Insert the tree before I.
1088 static Value *EmitAddTreeOfValues(Instruction *I,
1089 SmallVectorImpl<WeakVH> &Ops){
1090 if (Ops.size() == 1) return Ops.back();
1092 Value *V1 = Ops.back();
1094 Value *V2 = EmitAddTreeOfValues(I, Ops);
1095 return CreateAdd(V2, V1, "tmp", I, I);
1098 /// If V is an expression tree that is a multiplication sequence,
1099 /// and if this sequence contains a multiply by Factor,
1100 /// remove Factor from the tree and return the new tree.
1101 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
1102 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1106 SmallVector<RepeatedValue, 8> Tree;
1107 MadeChange |= LinearizeExprTree(BO, Tree);
1108 SmallVector<ValueEntry, 8> Factors;
1109 Factors.reserve(Tree.size());
1110 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1111 RepeatedValue E = Tree[i];
1112 Factors.append(E.second.getZExtValue(),
1113 ValueEntry(getRank(E.first), E.first));
1116 bool FoundFactor = false;
1117 bool NeedsNegate = false;
1118 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1119 if (Factors[i].Op == Factor) {
1121 Factors.erase(Factors.begin()+i);
1125 // If this is a negative version of this factor, remove it.
1126 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1127 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1128 if (FC1->getValue() == -FC2->getValue()) {
1129 FoundFactor = NeedsNegate = true;
1130 Factors.erase(Factors.begin()+i);
1133 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1134 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1135 APFloat F1(FC1->getValueAPF());
1136 APFloat F2(FC2->getValueAPF());
1138 if (F1.compare(F2) == APFloat::cmpEqual) {
1139 FoundFactor = NeedsNegate = true;
1140 Factors.erase(Factors.begin() + i);
1148 // Make sure to restore the operands to the expression tree.
1149 RewriteExprTree(BO, Factors);
1153 BasicBlock::iterator InsertPt = BO; ++InsertPt;
1155 // If this was just a single multiply, remove the multiply and return the only
1156 // remaining operand.
1157 if (Factors.size() == 1) {
1158 RedoInsts.insert(BO);
1161 RewriteExprTree(BO, Factors);
1166 V = CreateNeg(V, "neg", InsertPt, BO);
1171 /// If V is a single-use multiply, recursively add its operands as factors,
1172 /// otherwise add V to the list of factors.
1174 /// Ops is the top-level list of add operands we're trying to factor.
1175 static void FindSingleUseMultiplyFactors(Value *V,
1176 SmallVectorImpl<Value*> &Factors,
1177 const SmallVectorImpl<ValueEntry> &Ops) {
1178 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1180 Factors.push_back(V);
1184 // Otherwise, add the LHS and RHS to the list of factors.
1185 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
1186 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
1189 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1190 /// This optimizes based on identities. If it can be reduced to a single Value,
1191 /// it is returned, otherwise the Ops list is mutated as necessary.
1192 static Value *OptimizeAndOrXor(unsigned Opcode,
1193 SmallVectorImpl<ValueEntry> &Ops) {
1194 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1195 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1196 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1197 // First, check for X and ~X in the operand list.
1198 assert(i < Ops.size());
1199 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1200 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1201 unsigned FoundX = FindInOperandList(Ops, i, X);
1203 if (Opcode == Instruction::And) // ...&X&~X = 0
1204 return Constant::getNullValue(X->getType());
1206 if (Opcode == Instruction::Or) // ...|X|~X = -1
1207 return Constant::getAllOnesValue(X->getType());
1211 // Next, check for duplicate pairs of values, which we assume are next to
1212 // each other, due to our sorting criteria.
1213 assert(i < Ops.size());
1214 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1215 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1216 // Drop duplicate values for And and Or.
1217 Ops.erase(Ops.begin()+i);
1223 // Drop pairs of values for Xor.
1224 assert(Opcode == Instruction::Xor);
1226 return Constant::getNullValue(Ops[0].Op->getType());
1229 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1237 /// Helper funciton of CombineXorOpnd(). It creates a bitwise-and
1238 /// instruction with the given two operands, and return the resulting
1239 /// instruction. There are two special cases: 1) if the constant operand is 0,
1240 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1242 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1243 const APInt &ConstOpnd) {
1244 if (ConstOpnd != 0) {
1245 if (!ConstOpnd.isAllOnesValue()) {
1246 LLVMContext &Ctx = Opnd->getType()->getContext();
1248 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1249 "and.ra", InsertBefore);
1250 I->setDebugLoc(InsertBefore->getDebugLoc());
1258 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1259 // into "R ^ C", where C would be 0, and R is a symbolic value.
1261 // If it was successful, true is returned, and the "R" and "C" is returned
1262 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1263 // and both "Res" and "ConstOpnd" remain unchanged.
1265 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1266 APInt &ConstOpnd, Value *&Res) {
1267 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1268 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1269 // = (x & ~c1) ^ (c1 ^ c2)
1270 // It is useful only when c1 == c2.
1271 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) {
1272 if (!Opnd1->getValue()->hasOneUse())
1275 const APInt &C1 = Opnd1->getConstPart();
1276 if (C1 != ConstOpnd)
1279 Value *X = Opnd1->getSymbolicPart();
1280 Res = createAndInstr(I, X, ~C1);
1281 // ConstOpnd was C2, now C1 ^ C2.
1284 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1285 RedoInsts.insert(T);
1292 // Helper function of OptimizeXor(). It tries to simplify
1293 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1296 // If it was successful, true is returned, and the "R" and "C" is returned
1297 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1298 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1299 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1300 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
1301 APInt &ConstOpnd, Value *&Res) {
1302 Value *X = Opnd1->getSymbolicPart();
1303 if (X != Opnd2->getSymbolicPart())
1306 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1307 int DeadInstNum = 1;
1308 if (Opnd1->getValue()->hasOneUse())
1310 if (Opnd2->getValue()->hasOneUse())
1314 // (x | c1) ^ (x & c2)
1315 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1316 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1317 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1319 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1320 if (Opnd2->isOrExpr())
1321 std::swap(Opnd1, Opnd2);
1323 const APInt &C1 = Opnd1->getConstPart();
1324 const APInt &C2 = Opnd2->getConstPart();
1325 APInt C3((~C1) ^ C2);
1327 // Do not increase code size!
1328 if (C3 != 0 && !C3.isAllOnesValue()) {
1329 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1330 if (NewInstNum > DeadInstNum)
1334 Res = createAndInstr(I, X, C3);
1337 } else if (Opnd1->isOrExpr()) {
1338 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1340 const APInt &C1 = Opnd1->getConstPart();
1341 const APInt &C2 = Opnd2->getConstPart();
1344 // Do not increase code size
1345 if (C3 != 0 && !C3.isAllOnesValue()) {
1346 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1347 if (NewInstNum > DeadInstNum)
1351 Res = createAndInstr(I, X, C3);
1354 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1356 const APInt &C1 = Opnd1->getConstPart();
1357 const APInt &C2 = Opnd2->getConstPart();
1359 Res = createAndInstr(I, X, C3);
1362 // Put the original operands in the Redo list; hope they will be deleted
1364 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1365 RedoInsts.insert(T);
1366 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1367 RedoInsts.insert(T);
1372 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1373 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1375 Value *Reassociate::OptimizeXor(Instruction *I,
1376 SmallVectorImpl<ValueEntry> &Ops) {
1377 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1380 if (Ops.size() == 1)
1383 SmallVector<XorOpnd, 8> Opnds;
1384 SmallVector<XorOpnd*, 8> OpndPtrs;
1385 Type *Ty = Ops[0].Op->getType();
1386 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1388 // Step 1: Convert ValueEntry to XorOpnd
1389 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1390 Value *V = Ops[i].Op;
1391 if (!isa<ConstantInt>(V)) {
1393 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1396 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1399 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1400 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1401 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1402 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1403 // when new elements are added to the vector.
1404 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1405 OpndPtrs.push_back(&Opnds[i]);
1407 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1408 // the same symbolic value cluster together. For instance, the input operand
1409 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1410 // ("x | 123", "x & 789", "y & 456").
1411 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor());
1413 // Step 3: Combine adjacent operands
1414 XorOpnd *PrevOpnd = nullptr;
1415 bool Changed = false;
1416 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1417 XorOpnd *CurrOpnd = OpndPtrs[i];
1418 // The combined value
1421 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1422 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1425 *CurrOpnd = XorOpnd(CV);
1427 CurrOpnd->Invalidate();
1432 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1433 PrevOpnd = CurrOpnd;
1437 // step 3.2: When previous and current operands share the same symbolic
1438 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1440 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1441 // Remove previous operand
1442 PrevOpnd->Invalidate();
1444 *CurrOpnd = XorOpnd(CV);
1445 PrevOpnd = CurrOpnd;
1447 CurrOpnd->Invalidate();
1454 // Step 4: Reassemble the Ops
1457 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1458 XorOpnd &O = Opnds[i];
1461 ValueEntry VE(getRank(O.getValue()), O.getValue());
1464 if (ConstOpnd != 0) {
1465 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1466 ValueEntry VE(getRank(C), C);
1469 int Sz = Ops.size();
1471 return Ops.back().Op;
1473 assert(ConstOpnd == 0);
1474 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1481 /// Optimize a series of operands to an 'add' instruction. This
1482 /// optimizes based on identities. If it can be reduced to a single Value, it
1483 /// is returned, otherwise the Ops list is mutated as necessary.
1484 Value *Reassociate::OptimizeAdd(Instruction *I,
1485 SmallVectorImpl<ValueEntry> &Ops) {
1486 // Scan the operand lists looking for X and -X pairs. If we find any, we
1487 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1489 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1491 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1492 Value *TheOp = Ops[i].Op;
1493 // Check to see if we've seen this operand before. If so, we factor all
1494 // instances of the operand together. Due to our sorting criteria, we know
1495 // that these need to be next to each other in the vector.
1496 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1497 // Rescan the list, remove all instances of this operand from the expr.
1498 unsigned NumFound = 0;
1500 Ops.erase(Ops.begin()+i);
1502 } while (i != Ops.size() && Ops[i].Op == TheOp);
1504 DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1507 // Insert a new multiply.
1508 Type *Ty = TheOp->getType();
1509 Constant *C = Ty->isIntOrIntVectorTy() ?
1510 ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1511 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1513 // Now that we have inserted a multiply, optimize it. This allows us to
1514 // handle cases that require multiple factoring steps, such as this:
1515 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1516 RedoInsts.insert(Mul);
1518 // If every add operand was a duplicate, return the multiply.
1522 // Otherwise, we had some input that didn't have the dupe, such as
1523 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1524 // things being added by this operation.
1525 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1532 // Check for X and -X or X and ~X in the operand list.
1533 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1534 !BinaryOperator::isNot(TheOp))
1538 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1539 X = BinaryOperator::getNegArgument(TheOp);
1540 else if (BinaryOperator::isNot(TheOp))
1541 X = BinaryOperator::getNotArgument(TheOp);
1543 unsigned FoundX = FindInOperandList(Ops, i, X);
1547 // Remove X and -X from the operand list.
1548 if (Ops.size() == 2 &&
1549 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1550 return Constant::getNullValue(X->getType());
1552 // Remove X and ~X from the operand list.
1553 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1554 return Constant::getAllOnesValue(X->getType());
1556 Ops.erase(Ops.begin()+i);
1560 --i; // Need to back up an extra one.
1561 Ops.erase(Ops.begin()+FoundX);
1563 --i; // Revisit element.
1564 e -= 2; // Removed two elements.
1566 // if X and ~X we append -1 to the operand list.
1567 if (BinaryOperator::isNot(TheOp)) {
1568 Value *V = Constant::getAllOnesValue(X->getType());
1569 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1574 // Scan the operand list, checking to see if there are any common factors
1575 // between operands. Consider something like A*A+A*B*C+D. We would like to
1576 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1577 // To efficiently find this, we count the number of times a factor occurs
1578 // for any ADD operands that are MULs.
1579 DenseMap<Value*, unsigned> FactorOccurrences;
1581 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1582 // where they are actually the same multiply.
1583 unsigned MaxOcc = 0;
1584 Value *MaxOccVal = nullptr;
1585 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1586 BinaryOperator *BOp =
1587 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1591 // Compute all of the factors of this added value.
1592 SmallVector<Value*, 8> Factors;
1593 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1594 assert(Factors.size() > 1 && "Bad linearize!");
1596 // Add one to FactorOccurrences for each unique factor in this op.
1597 SmallPtrSet<Value*, 8> Duplicates;
1598 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1599 Value *Factor = Factors[i];
1600 if (!Duplicates.insert(Factor).second)
1603 unsigned Occ = ++FactorOccurrences[Factor];
1609 // If Factor is a negative constant, add the negated value as a factor
1610 // because we can percolate the negate out. Watch for minint, which
1611 // cannot be positivified.
1612 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1613 if (CI->isNegative() && !CI->isMinValue(true)) {
1614 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1615 assert(!Duplicates.count(Factor) &&
1616 "Shouldn't have two constant factors, missed a canonicalize");
1617 unsigned Occ = ++FactorOccurrences[Factor];
1623 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1624 if (CF->isNegative()) {
1625 APFloat F(CF->getValueAPF());
1627 Factor = ConstantFP::get(CF->getContext(), F);
1628 assert(!Duplicates.count(Factor) &&
1629 "Shouldn't have two constant factors, missed a canonicalize");
1630 unsigned Occ = ++FactorOccurrences[Factor];
1640 // If any factor occurred more than one time, we can pull it out.
1642 DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1645 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1646 // this, we could otherwise run into situations where removing a factor
1647 // from an expression will drop a use of maxocc, and this can cause
1648 // RemoveFactorFromExpression on successive values to behave differently.
1649 Instruction *DummyInst =
1650 I->getType()->isIntOrIntVectorTy()
1651 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1652 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1654 SmallVector<WeakVH, 4> NewMulOps;
1655 for (unsigned i = 0; i != Ops.size(); ++i) {
1656 // Only try to remove factors from expressions we're allowed to.
1657 BinaryOperator *BOp =
1658 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1662 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1663 // The factorized operand may occur several times. Convert them all in
1665 for (unsigned j = Ops.size(); j != i;) {
1667 if (Ops[j].Op == Ops[i].Op) {
1668 NewMulOps.push_back(V);
1669 Ops.erase(Ops.begin()+j);
1676 // No need for extra uses anymore.
1679 unsigned NumAddedValues = NewMulOps.size();
1680 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1682 // Now that we have inserted the add tree, optimize it. This allows us to
1683 // handle cases that require multiple factoring steps, such as this:
1684 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1685 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1686 (void)NumAddedValues;
1687 if (Instruction *VI = dyn_cast<Instruction>(V))
1688 RedoInsts.insert(VI);
1690 // Create the multiply.
1691 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1693 // Rerun associate on the multiply in case the inner expression turned into
1694 // a multiply. We want to make sure that we keep things in canonical form.
1695 RedoInsts.insert(V2);
1697 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1698 // entire result expression is just the multiply "A*(B+C)".
1702 // Otherwise, we had some input that didn't have the factor, such as
1703 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1704 // things being added by this operation.
1705 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1711 /// \brief Build up a vector of value/power pairs factoring a product.
1713 /// Given a series of multiplication operands, build a vector of factors and
1714 /// the powers each is raised to when forming the final product. Sort them in
1715 /// the order of descending power.
1717 /// (x*x) -> [(x, 2)]
1718 /// ((x*x)*x) -> [(x, 3)]
1719 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1721 /// \returns Whether any factors have a power greater than one.
1722 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1723 SmallVectorImpl<Factor> &Factors) {
1724 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1725 // Compute the sum of powers of simplifiable factors.
1726 unsigned FactorPowerSum = 0;
1727 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1728 Value *Op = Ops[Idx-1].Op;
1730 // Count the number of occurrences of this value.
1732 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1734 // Track for simplification all factors which occur 2 or more times.
1736 FactorPowerSum += Count;
1739 // We can only simplify factors if the sum of the powers of our simplifiable
1740 // factors is 4 or higher. When that is the case, we will *always* have
1741 // a simplification. This is an important invariant to prevent cyclicly
1742 // trying to simplify already minimal formations.
1743 if (FactorPowerSum < 4)
1746 // Now gather the simplifiable factors, removing them from Ops.
1748 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1749 Value *Op = Ops[Idx-1].Op;
1751 // Count the number of occurrences of this value.
1753 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1757 // Move an even number of occurrences to Factors.
1760 FactorPowerSum += Count;
1761 Factors.push_back(Factor(Op, Count));
1762 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1765 // None of the adjustments above should have reduced the sum of factor powers
1766 // below our mininum of '4'.
1767 assert(FactorPowerSum >= 4);
1769 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1773 /// \brief Build a tree of multiplies, computing the product of Ops.
1774 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1775 SmallVectorImpl<Value*> &Ops) {
1776 if (Ops.size() == 1)
1779 Value *LHS = Ops.pop_back_val();
1781 if (LHS->getType()->isIntOrIntVectorTy())
1782 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1784 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1785 } while (!Ops.empty());
1790 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1792 /// Given a vector of values raised to various powers, where no two values are
1793 /// equal and the powers are sorted in decreasing order, compute the minimal
1794 /// DAG of multiplies to compute the final product, and return that product
1796 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1797 SmallVectorImpl<Factor> &Factors) {
1798 assert(Factors[0].Power);
1799 SmallVector<Value *, 4> OuterProduct;
1800 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1801 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1802 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1807 // We want to multiply across all the factors with the same power so that
1808 // we can raise them to that power as a single entity. Build a mini tree
1810 SmallVector<Value *, 4> InnerProduct;
1811 InnerProduct.push_back(Factors[LastIdx].Base);
1813 InnerProduct.push_back(Factors[Idx].Base);
1815 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1817 // Reset the base value of the first factor to the new expression tree.
1818 // We'll remove all the factors with the same power in a second pass.
1819 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1820 if (Instruction *MI = dyn_cast<Instruction>(M))
1821 RedoInsts.insert(MI);
1825 // Unique factors with equal powers -- we've folded them into the first one's
1827 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1828 Factor::PowerEqual()),
1831 // Iteratively collect the base of each factor with an add power into the
1832 // outer product, and halve each power in preparation for squaring the
1834 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1835 if (Factors[Idx].Power & 1)
1836 OuterProduct.push_back(Factors[Idx].Base);
1837 Factors[Idx].Power >>= 1;
1839 if (Factors[0].Power) {
1840 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1841 OuterProduct.push_back(SquareRoot);
1842 OuterProduct.push_back(SquareRoot);
1844 if (OuterProduct.size() == 1)
1845 return OuterProduct.front();
1847 Value *V = buildMultiplyTree(Builder, OuterProduct);
1851 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1852 SmallVectorImpl<ValueEntry> &Ops) {
1853 // We can only optimize the multiplies when there is a chain of more than
1854 // three, such that a balanced tree might require fewer total multiplies.
1858 // Try to turn linear trees of multiplies without other uses of the
1859 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1861 SmallVector<Factor, 4> Factors;
1862 if (!collectMultiplyFactors(Ops, Factors))
1863 return nullptr; // All distinct factors, so nothing left for us to do.
1865 IRBuilder<> Builder(I);
1866 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1870 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1871 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1875 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1876 SmallVectorImpl<ValueEntry> &Ops) {
1877 // Now that we have the linearized expression tree, try to optimize it.
1878 // Start by folding any constants that we found.
1879 Constant *Cst = nullptr;
1880 unsigned Opcode = I->getOpcode();
1881 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1882 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1883 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1885 // If there was nothing but constants then we are done.
1889 // Put the combined constant back at the end of the operand list, except if
1890 // there is no point. For example, an add of 0 gets dropped here, while a
1891 // multiplication by zero turns the whole expression into zero.
1892 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1893 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1895 Ops.push_back(ValueEntry(0, Cst));
1898 if (Ops.size() == 1) return Ops[0].Op;
1900 // Handle destructive annihilation due to identities between elements in the
1901 // argument list here.
1902 unsigned NumOps = Ops.size();
1905 case Instruction::And:
1906 case Instruction::Or:
1907 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1911 case Instruction::Xor:
1912 if (Value *Result = OptimizeXor(I, Ops))
1916 case Instruction::Add:
1917 case Instruction::FAdd:
1918 if (Value *Result = OptimizeAdd(I, Ops))
1922 case Instruction::Mul:
1923 case Instruction::FMul:
1924 if (Value *Result = OptimizeMul(I, Ops))
1929 if (Ops.size() != NumOps)
1930 return OptimizeExpression(I, Ops);
1934 /// Zap the given instruction, adding interesting operands to the work list.
1935 void Reassociate::EraseInst(Instruction *I) {
1936 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1937 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1938 // Erase the dead instruction.
1939 ValueRankMap.erase(I);
1940 RedoInsts.remove(I);
1941 I->eraseFromParent();
1942 // Optimize its operands.
1943 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1944 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1945 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1946 // If this is a node in an expression tree, climb to the expression root
1947 // and add that since that's where optimization actually happens.
1948 unsigned Opcode = Op->getOpcode();
1949 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1950 Visited.insert(Op).second)
1951 Op = Op->user_back();
1952 RedoInsts.insert(Op);
1956 // Canonicalize expressions of the following form:
1957 // x + (-Constant * y) -> x - (Constant * y)
1958 // x - (-Constant * y) -> x + (Constant * y)
1959 Instruction *Reassociate::canonicalizeNegConstExpr(Instruction *I) {
1960 if (!I->hasOneUse() || I->getType()->isVectorTy())
1963 // Must be a fmul or fdiv instruction.
1964 unsigned Opcode = I->getOpcode();
1965 if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv)
1968 auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0));
1969 auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1));
1971 // Both operands are constant, let it get constant folded away.
1975 ConstantFP *CF = C0 ? C0 : C1;
1977 // Must have one constant operand.
1981 // Must be a negative ConstantFP.
1982 if (!CF->isNegative())
1985 // User must be a binary operator with one or more uses.
1986 Instruction *User = I->user_back();
1987 if (!isa<BinaryOperator>(User) || !User->hasNUsesOrMore(1))
1990 unsigned UserOpcode = User->getOpcode();
1991 if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub)
1994 // Subtraction is not commutative. Explicitly, the following transform is
1995 // not valid: (-Constant * y) - x -> x + (Constant * y)
1996 if (!User->isCommutative() && User->getOperand(1) != I)
1999 // Change the sign of the constant.
2000 APFloat Val = CF->getValueAPF();
2002 I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val));
2004 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
2005 // ((-Const*y) + x) -> (x + (-Const*y)).
2006 if (User->getOperand(0) == I && User->isCommutative())
2007 cast<BinaryOperator>(User)->swapOperands();
2009 Value *Op0 = User->getOperand(0);
2010 Value *Op1 = User->getOperand(1);
2012 switch (UserOpcode) {
2014 llvm_unreachable("Unexpected Opcode!");
2015 case Instruction::FAdd:
2016 NI = BinaryOperator::CreateFSub(Op0, Op1);
2017 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2019 case Instruction::FSub:
2020 NI = BinaryOperator::CreateFAdd(Op0, Op1);
2021 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2025 NI->insertBefore(User);
2026 NI->setName(User->getName());
2027 User->replaceAllUsesWith(NI);
2028 NI->setDebugLoc(I->getDebugLoc());
2029 RedoInsts.insert(I);
2034 /// Inspect and optimize the given instruction. Note that erasing
2035 /// instructions is not allowed.
2036 void Reassociate::OptimizeInst(Instruction *I) {
2037 // Only consider operations that we understand.
2038 if (!isa<BinaryOperator>(I))
2041 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
2042 // If an operand of this shift is a reassociable multiply, or if the shift
2043 // is used by a reassociable multiply or add, turn into a multiply.
2044 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
2046 (isReassociableOp(I->user_back(), Instruction::Mul) ||
2047 isReassociableOp(I->user_back(), Instruction::Add)))) {
2048 Instruction *NI = ConvertShiftToMul(I);
2049 RedoInsts.insert(I);
2054 // Canonicalize negative constants out of expressions.
2055 if (Instruction *Res = canonicalizeNegConstExpr(I))
2058 // Commute binary operators, to canonicalize the order of their operands.
2059 // This can potentially expose more CSE opportunities, and makes writing other
2060 // transformations simpler.
2061 if (I->isCommutative())
2062 canonicalizeOperands(I);
2064 // TODO: We should optimize vector Xor instructions, but they are
2065 // currently unsupported.
2066 if (I->getType()->isVectorTy() && I->getOpcode() == Instruction::Xor)
2069 // Don't optimize floating point instructions that don't have unsafe algebra.
2070 if (I->getType()->isFloatingPointTy() && !I->hasUnsafeAlgebra())
2073 // Do not reassociate boolean (i1) expressions. We want to preserve the
2074 // original order of evaluation for short-circuited comparisons that
2075 // SimplifyCFG has folded to AND/OR expressions. If the expression
2076 // is not further optimized, it is likely to be transformed back to a
2077 // short-circuited form for code gen, and the source order may have been
2078 // optimized for the most likely conditions.
2079 if (I->getType()->isIntegerTy(1))
2082 // If this is a subtract instruction which is not already in negate form,
2083 // see if we can convert it to X+-Y.
2084 if (I->getOpcode() == Instruction::Sub) {
2085 if (ShouldBreakUpSubtract(I)) {
2086 Instruction *NI = BreakUpSubtract(I);
2087 RedoInsts.insert(I);
2090 } else if (BinaryOperator::isNeg(I)) {
2091 // Otherwise, this is a negation. See if the operand is a multiply tree
2092 // and if this is not an inner node of a multiply tree.
2093 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2095 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2096 Instruction *NI = LowerNegateToMultiply(I);
2097 RedoInsts.insert(I);
2102 } else if (I->getOpcode() == Instruction::FSub) {
2103 if (ShouldBreakUpSubtract(I)) {
2104 Instruction *NI = BreakUpSubtract(I);
2105 RedoInsts.insert(I);
2108 } else if (BinaryOperator::isFNeg(I)) {
2109 // Otherwise, this is a negation. See if the operand is a multiply tree
2110 // and if this is not an inner node of a multiply tree.
2111 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2113 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2114 Instruction *NI = LowerNegateToMultiply(I);
2115 RedoInsts.insert(I);
2122 // If this instruction is an associative binary operator, process it.
2123 if (!I->isAssociative()) return;
2124 BinaryOperator *BO = cast<BinaryOperator>(I);
2126 // If this is an interior node of a reassociable tree, ignore it until we
2127 // get to the root of the tree, to avoid N^2 analysis.
2128 unsigned Opcode = BO->getOpcode();
2129 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode)
2132 // If this is an add tree that is used by a sub instruction, ignore it
2133 // until we process the subtract.
2134 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2135 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2137 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2138 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2141 ReassociateExpression(BO);
2144 void Reassociate::ReassociateExpression(BinaryOperator *I) {
2145 // First, walk the expression tree, linearizing the tree, collecting the
2146 // operand information.
2147 SmallVector<RepeatedValue, 8> Tree;
2148 MadeChange |= LinearizeExprTree(I, Tree);
2149 SmallVector<ValueEntry, 8> Ops;
2150 Ops.reserve(Tree.size());
2151 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2152 RepeatedValue E = Tree[i];
2153 Ops.append(E.second.getZExtValue(),
2154 ValueEntry(getRank(E.first), E.first));
2157 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2159 // Now that we have linearized the tree to a list and have gathered all of
2160 // the operands and their ranks, sort the operands by their rank. Use a
2161 // stable_sort so that values with equal ranks will have their relative
2162 // positions maintained (and so the compiler is deterministic). Note that
2163 // this sorts so that the highest ranking values end up at the beginning of
2165 std::stable_sort(Ops.begin(), Ops.end());
2167 // Now that we have the expression tree in a convenient
2168 // sorted form, optimize it globally if possible.
2169 if (Value *V = OptimizeExpression(I, Ops)) {
2171 // Self-referential expression in unreachable code.
2173 // This expression tree simplified to something that isn't a tree,
2175 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2176 I->replaceAllUsesWith(V);
2177 if (Instruction *VI = dyn_cast<Instruction>(V))
2178 VI->setDebugLoc(I->getDebugLoc());
2179 RedoInsts.insert(I);
2184 // We want to sink immediates as deeply as possible except in the case where
2185 // this is a multiply tree used only by an add, and the immediate is a -1.
2186 // In this case we reassociate to put the negation on the outside so that we
2187 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2188 if (I->hasOneUse()) {
2189 if (I->getOpcode() == Instruction::Mul &&
2190 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2191 isa<ConstantInt>(Ops.back().Op) &&
2192 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
2193 ValueEntry Tmp = Ops.pop_back_val();
2194 Ops.insert(Ops.begin(), Tmp);
2195 } else if (I->getOpcode() == Instruction::FMul &&
2196 cast<Instruction>(I->user_back())->getOpcode() ==
2197 Instruction::FAdd &&
2198 isa<ConstantFP>(Ops.back().Op) &&
2199 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2200 ValueEntry Tmp = Ops.pop_back_val();
2201 Ops.insert(Ops.begin(), Tmp);
2205 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2207 if (Ops.size() == 1) {
2209 // Self-referential expression in unreachable code.
2212 // This expression tree simplified to something that isn't a tree,
2214 I->replaceAllUsesWith(Ops[0].Op);
2215 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2216 OI->setDebugLoc(I->getDebugLoc());
2217 RedoInsts.insert(I);
2221 // Now that we ordered and optimized the expressions, splat them back into
2222 // the expression tree, removing any unneeded nodes.
2223 RewriteExprTree(I, Ops);
2226 bool Reassociate::runOnFunction(Function &F) {
2227 if (skipOptnoneFunction(F))
2230 // Calculate the rank map for F
2234 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
2235 // Optimize every instruction in the basic block.
2236 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; )
2237 if (isInstructionTriviallyDead(II)) {
2241 assert(II->getParent() == BI && "Moved to a different block!");
2245 // If this produced extra instructions to optimize, handle them now.
2246 while (!RedoInsts.empty()) {
2247 Instruction *I = RedoInsts.pop_back_val();
2248 if (isInstructionTriviallyDead(I))
2255 // We are done with the rank map.
2257 ValueRankMap.clear();