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/IR/CFG.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/ValueHandle.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Transforms/Utils/Local.h"
44 #define DEBUG_TYPE "reassociate"
46 STATISTIC(NumChanged, "Number of insts reassociated");
47 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
48 STATISTIC(NumFactor , "Number of multiplies factored");
54 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {}
56 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) {
57 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start.
62 /// PrintOps - Print out the expression identified in the Ops list.
64 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
65 Module *M = I->getParent()->getParent()->getParent();
66 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
67 << *Ops[0].Op->getType() << '\t';
68 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
70 Ops[i].Op->printAsOperand(dbgs(), false, M);
71 dbgs() << ", #" << Ops[i].Rank << "] ";
77 /// \brief Utility class representing a base and exponent pair which form one
78 /// factor of some product.
83 Factor(Value *Base, unsigned Power) : Base(Base), Power(Power) {}
85 /// \brief Sort factors by their Base.
87 bool operator()(const Factor &LHS, const Factor &RHS) {
88 return LHS.Base < RHS.Base;
92 /// \brief Compare factors for equal bases.
94 bool operator()(const Factor &LHS, const Factor &RHS) {
95 return LHS.Base == RHS.Base;
99 /// \brief Sort factors in descending order by their power.
100 struct PowerDescendingSorter {
101 bool operator()(const Factor &LHS, const Factor &RHS) {
102 return LHS.Power > RHS.Power;
106 /// \brief Compare factors for equal powers.
108 bool operator()(const Factor &LHS, const Factor &RHS) {
109 return LHS.Power == RHS.Power;
114 /// Utility class representing a non-constant Xor-operand. We classify
115 /// non-constant Xor-Operands into two categories:
116 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0
118 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
120 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
121 /// operand as "E | 0"
126 bool isInvalid() const { return SymbolicPart == nullptr; }
127 bool isOrExpr() const { return isOr; }
128 Value *getValue() const { return OrigVal; }
129 Value *getSymbolicPart() const { return SymbolicPart; }
130 unsigned getSymbolicRank() const { return SymbolicRank; }
131 const APInt &getConstPart() const { return ConstPart; }
133 void Invalidate() { SymbolicPart = OrigVal = nullptr; }
134 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
136 // Sort the XorOpnd-Pointer in ascending order of symbolic-value-rank.
137 // The purpose is twofold:
138 // 1) Cluster together the operands sharing the same symbolic-value.
139 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
140 // could potentially shorten crital path, and expose more loop-invariants.
141 // Note that values' rank are basically defined in RPO order (FIXME).
142 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
143 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
144 // "z" in the order of X-Y-Z is better than any other orders.
145 struct PtrSortFunctor {
146 bool operator()(XorOpnd * const &LHS, XorOpnd * const &RHS) {
147 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
154 unsigned SymbolicRank;
160 class Reassociate : public FunctionPass {
161 DenseMap<BasicBlock*, unsigned> RankMap;
162 DenseMap<AssertingVH<Value>, unsigned> ValueRankMap;
163 SetVector<AssertingVH<Instruction> > RedoInsts;
166 static char ID; // Pass identification, replacement for typeid
167 Reassociate() : FunctionPass(ID) {
168 initializeReassociatePass(*PassRegistry::getPassRegistry());
171 bool runOnFunction(Function &F) override;
173 void getAnalysisUsage(AnalysisUsage &AU) const override {
174 AU.setPreservesCFG();
177 void BuildRankMap(Function &F);
178 unsigned getRank(Value *V);
179 void ReassociateExpression(BinaryOperator *I);
180 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
181 Value *OptimizeExpression(BinaryOperator *I,
182 SmallVectorImpl<ValueEntry> &Ops);
183 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
184 Value *OptimizeXor(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
185 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, APInt &ConstOpnd,
187 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
188 APInt &ConstOpnd, Value *&Res);
189 bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
190 SmallVectorImpl<Factor> &Factors);
191 Value *buildMinimalMultiplyDAG(IRBuilder<> &Builder,
192 SmallVectorImpl<Factor> &Factors);
193 Value *OptimizeMul(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
194 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
195 void EraseInst(Instruction *I);
196 void OptimizeInst(Instruction *I);
200 XorOpnd::XorOpnd(Value *V) {
201 assert(!isa<ConstantInt>(V) && "No ConstantInt");
203 Instruction *I = dyn_cast<Instruction>(V);
206 if (I && (I->getOpcode() == Instruction::Or ||
207 I->getOpcode() == Instruction::And)) {
208 Value *V0 = I->getOperand(0);
209 Value *V1 = I->getOperand(1);
210 if (isa<ConstantInt>(V0))
213 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) {
214 ConstPart = C->getValue();
216 isOr = (I->getOpcode() == Instruction::Or);
221 // view the operand as "V | 0"
223 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth());
227 char Reassociate::ID = 0;
228 INITIALIZE_PASS(Reassociate, "reassociate",
229 "Reassociate expressions", false, false)
231 // Public interface to the Reassociate pass
232 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
234 /// isReassociableOp - Return true if V is an instruction of the specified
235 /// opcode and if it only has one use.
236 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
237 if (V->hasOneUse() && isa<Instruction>(V) &&
238 cast<Instruction>(V)->getOpcode() == Opcode)
239 return cast<BinaryOperator>(V);
243 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
245 if (V->hasOneUse() && isa<Instruction>(V) &&
246 (cast<Instruction>(V)->getOpcode() == Opcode1 ||
247 cast<Instruction>(V)->getOpcode() == Opcode2))
248 return cast<BinaryOperator>(V);
252 static bool isUnmovableInstruction(Instruction *I) {
253 switch (I->getOpcode()) {
254 case Instruction::PHI:
255 case Instruction::LandingPad:
256 case Instruction::Alloca:
257 case Instruction::Load:
258 case Instruction::Invoke:
259 case Instruction::UDiv:
260 case Instruction::SDiv:
261 case Instruction::FDiv:
262 case Instruction::URem:
263 case Instruction::SRem:
264 case Instruction::FRem:
266 case Instruction::Call:
267 return !isa<DbgInfoIntrinsic>(I);
273 void Reassociate::BuildRankMap(Function &F) {
276 // Assign distinct ranks to function arguments
277 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
278 ValueRankMap[&*I] = ++i;
280 ReversePostOrderTraversal<Function*> RPOT(&F);
281 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
282 E = RPOT.end(); I != E; ++I) {
284 unsigned BBRank = RankMap[BB] = ++i << 16;
286 // Walk the basic block, adding precomputed ranks for any instructions that
287 // we cannot move. This ensures that the ranks for these instructions are
288 // all different in the block.
289 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
290 if (isUnmovableInstruction(I))
291 ValueRankMap[&*I] = ++BBRank;
295 unsigned Reassociate::getRank(Value *V) {
296 Instruction *I = dyn_cast<Instruction>(V);
298 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
299 return 0; // Otherwise it's a global or constant, rank 0.
302 if (unsigned Rank = ValueRankMap[I])
303 return Rank; // Rank already known?
305 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
306 // we can reassociate expressions for code motion! Since we do not recurse
307 // for PHI nodes, we cannot have infinite recursion here, because there
308 // cannot be loops in the value graph that do not go through PHI nodes.
309 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
310 for (unsigned i = 0, e = I->getNumOperands();
311 i != e && Rank != MaxRank; ++i)
312 Rank = std::max(Rank, getRank(I->getOperand(i)));
314 // If this is a not or neg instruction, do not count it for rank. This
315 // assures us that X and ~X will have the same rank.
316 Type *Ty = V->getType();
317 if ((!Ty->isIntegerTy() && !Ty->isFloatingPointTy()) ||
318 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
319 !BinaryOperator::isFNeg(I)))
322 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = "
325 return ValueRankMap[I] = Rank;
328 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
329 Instruction *InsertBefore, Value *FlagsOp) {
330 if (S1->getType()->isIntegerTy())
331 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
333 BinaryOperator *Res =
334 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
335 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
340 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
341 Instruction *InsertBefore, Value *FlagsOp) {
342 if (S1->getType()->isIntegerTy())
343 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
345 BinaryOperator *Res =
346 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
347 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
352 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
353 Instruction *InsertBefore, Value *FlagsOp) {
354 if (S1->getType()->isIntegerTy())
355 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
357 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
358 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
363 /// LowerNegateToMultiply - Replace 0-X with X*-1.
365 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
366 Type *Ty = Neg->getType();
367 Constant *NegOne = Ty->isIntegerTy() ? ConstantInt::getAllOnesValue(Ty)
368 : ConstantFP::get(Ty, -1.0);
370 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
371 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
373 Neg->replaceAllUsesWith(Res);
374 Res->setDebugLoc(Neg->getDebugLoc());
378 /// CarmichaelShift - Returns k such that lambda(2^Bitwidth) = 2^k, where lambda
379 /// is the Carmichael function. This means that x^(2^k) === 1 mod 2^Bitwidth for
380 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
381 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
382 /// even x in Bitwidth-bit arithmetic.
383 static unsigned CarmichaelShift(unsigned Bitwidth) {
389 /// IncorporateWeight - Add the extra weight 'RHS' to the existing weight 'LHS',
390 /// reducing the combined weight using any special properties of the operation.
391 /// The existing weight LHS represents the computation X op X op ... op X where
392 /// X occurs LHS times. The combined weight represents X op X op ... op X with
393 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
394 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
395 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
396 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
397 // If we were working with infinite precision arithmetic then the combined
398 // weight would be LHS + RHS. But we are using finite precision arithmetic,
399 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
400 // for nilpotent operations and addition, but not for idempotent operations
401 // and multiplication), so it is important to correctly reduce the combined
402 // weight back into range if wrapping would be wrong.
404 // If RHS is zero then the weight didn't change.
405 if (RHS.isMinValue())
407 // If LHS is zero then the combined weight is RHS.
408 if (LHS.isMinValue()) {
412 // From this point on we know that neither LHS nor RHS is zero.
414 if (Instruction::isIdempotent(Opcode)) {
415 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
416 // weight of 1. Keeping weights at zero or one also means that wrapping is
418 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
419 return; // Return a weight of 1.
421 if (Instruction::isNilpotent(Opcode)) {
422 // Nilpotent means X op X === 0, so reduce weights modulo 2.
423 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
424 LHS = 0; // 1 + 1 === 0 modulo 2.
427 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
428 // TODO: Reduce the weight by exploiting nsw/nuw?
433 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
434 "Unknown associative operation!");
435 unsigned Bitwidth = LHS.getBitWidth();
436 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
437 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
438 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
439 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
440 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
441 // which by a happy accident means that they can always be represented using
443 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
444 // the Carmichael number).
446 /// CM - The value of Carmichael's lambda function.
447 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
448 // Any weight W >= Threshold can be replaced with W - CM.
449 APInt Threshold = CM + Bitwidth;
450 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
451 // For Bitwidth 4 or more the following sum does not overflow.
453 while (LHS.uge(Threshold))
456 // To avoid problems with overflow do everything the same as above but using
458 unsigned CM = 1U << CarmichaelShift(Bitwidth);
459 unsigned Threshold = CM + Bitwidth;
460 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
461 "Weights not reduced!");
462 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
463 while (Total >= Threshold)
469 typedef std::pair<Value*, APInt> RepeatedValue;
471 /// LinearizeExprTree - Given an associative binary expression, return the leaf
472 /// nodes in Ops along with their weights (how many times the leaf occurs). The
473 /// original expression is the same as
474 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
476 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
480 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
482 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
484 /// This routine may modify the function, in which case it returns 'true'. The
485 /// changes it makes may well be destructive, changing the value computed by 'I'
486 /// to something completely different. Thus if the routine returns 'true' then
487 /// you MUST either replace I with a new expression computed from the Ops array,
488 /// or use RewriteExprTree to put the values back in.
490 /// A leaf node is either not a binary operation of the same kind as the root
491 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
492 /// opcode), or is the same kind of binary operator but has a use which either
493 /// does not belong to the expression, or does belong to the expression but is
494 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
495 /// of the expression, while for non-leaf nodes (except for the root 'I') every
496 /// use is a non-leaf node of the expression.
499 /// expression graph node names
509 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
510 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
512 /// The expression is maximal: if some instruction is a binary operator of the
513 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
514 /// then the instruction also belongs to the expression, is not a leaf node of
515 /// it, and its operands also belong to the expression (but may be leaf nodes).
517 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
518 /// order to ensure that every non-root node in the expression has *exactly one*
519 /// use by a non-leaf node of the expression. This destruction means that the
520 /// caller MUST either replace 'I' with a new expression or use something like
521 /// RewriteExprTree to put the values back in if the routine indicates that it
522 /// made a change by returning 'true'.
524 /// In the above example either the right operand of A or the left operand of B
525 /// will be replaced by undef. If it is B's operand then this gives:
529 /// + + | A, B - operand of B replaced with undef
535 /// Note that such undef operands can only be reached by passing through 'I'.
536 /// For example, if you visit operands recursively starting from a leaf node
537 /// then you will never see such an undef operand unless you get back to 'I',
538 /// which requires passing through a phi node.
540 /// Note that this routine may also mutate binary operators of the wrong type
541 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
542 /// of the expression) if it can turn them into binary operators of the right
543 /// type and thus make the expression bigger.
545 static bool LinearizeExprTree(BinaryOperator *I,
546 SmallVectorImpl<RepeatedValue> &Ops) {
547 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
548 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
549 unsigned Opcode = I->getOpcode();
550 assert(I->isAssociative() && I->isCommutative() &&
551 "Expected an associative and commutative operation!");
553 // Visit all operands of the expression, keeping track of their weight (the
554 // number of paths from the expression root to the operand, or if you like
555 // the number of times that operand occurs in the linearized expression).
556 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
557 // while A has weight two.
559 // Worklist of non-leaf nodes (their operands are in the expression too) along
560 // with their weights, representing a certain number of paths to the operator.
561 // If an operator occurs in the worklist multiple times then we found multiple
562 // ways to get to it.
563 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
564 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
565 bool MadeChange = false;
567 // Leaves of the expression are values that either aren't the right kind of
568 // operation (eg: a constant, or a multiply in an add tree), or are, but have
569 // some uses that are not inside the expression. For example, in I = X + X,
570 // X = A + B, the value X has two uses (by I) that are in the expression. If
571 // X has any other uses, for example in a return instruction, then we consider
572 // X to be a leaf, and won't analyze it further. When we first visit a value,
573 // if it has more than one use then at first we conservatively consider it to
574 // be a leaf. Later, as the expression is explored, we may discover some more
575 // uses of the value from inside the expression. If all uses turn out to be
576 // from within the expression (and the value is a binary operator of the right
577 // kind) then the value is no longer considered to be a leaf, and its operands
580 // Leaves - Keeps track of the set of putative leaves as well as the number of
581 // paths to each leaf seen so far.
582 typedef DenseMap<Value*, APInt> LeafMap;
583 LeafMap Leaves; // Leaf -> Total weight so far.
584 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
587 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
589 while (!Worklist.empty()) {
590 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
591 I = P.first; // We examine the operands of this binary operator.
593 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
594 Value *Op = I->getOperand(OpIdx);
595 APInt Weight = P.second; // Number of paths to this operand.
596 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
597 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
599 // If this is a binary operation of the right kind with only one use then
600 // add its operands to the expression.
601 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
602 assert(Visited.insert(Op) && "Not first visit!");
603 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
604 Worklist.push_back(std::make_pair(BO, Weight));
608 // Appears to be a leaf. Is the operand already in the set of leaves?
609 LeafMap::iterator It = Leaves.find(Op);
610 if (It == Leaves.end()) {
611 // Not in the leaf map. Must be the first time we saw this operand.
612 assert(Visited.insert(Op) && "Not first visit!");
613 if (!Op->hasOneUse()) {
614 // This value has uses not accounted for by the expression, so it is
615 // not safe to modify. Mark it as being a leaf.
616 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
617 LeafOrder.push_back(Op);
621 // No uses outside the expression, try morphing it.
622 } else if (It != Leaves.end()) {
623 // Already in the leaf map.
624 assert(Visited.count(Op) && "In leaf map but not visited!");
626 // Update the number of paths to the leaf.
627 IncorporateWeight(It->second, Weight, Opcode);
629 #if 0 // TODO: Re-enable once PR13021 is fixed.
630 // The leaf already has one use from inside the expression. As we want
631 // exactly one such use, drop this new use of the leaf.
632 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
633 I->setOperand(OpIdx, UndefValue::get(I->getType()));
636 // If the leaf is a binary operation of the right kind and we now see
637 // that its multiple original uses were in fact all by nodes belonging
638 // to the expression, then no longer consider it to be a leaf and add
639 // its operands to the expression.
640 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
641 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
642 Worklist.push_back(std::make_pair(BO, It->second));
648 // If we still have uses that are not accounted for by the expression
649 // then it is not safe to modify the value.
650 if (!Op->hasOneUse())
653 // No uses outside the expression, try morphing it.
655 Leaves.erase(It); // Since the value may be morphed below.
658 // At this point we have a value which, first of all, is not a binary
659 // expression of the right kind, and secondly, is only used inside the
660 // expression. This means that it can safely be modified. See if we
661 // can usefully morph it into an expression of the right kind.
662 assert((!isa<Instruction>(Op) ||
663 cast<Instruction>(Op)->getOpcode() != Opcode) &&
664 "Should have been handled above!");
665 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
667 // If this is a multiply expression, turn any internal negations into
668 // multiplies by -1 so they can be reassociated.
669 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
670 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
671 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
672 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
673 BO = LowerNegateToMultiply(BO);
674 DEBUG(dbgs() << *BO << '\n');
675 Worklist.push_back(std::make_pair(BO, Weight));
680 // Failed to morph into an expression of the right type. This really is
682 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
683 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
684 LeafOrder.push_back(Op);
689 // The leaves, repeated according to their weights, represent the linearized
690 // form of the expression.
691 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
692 Value *V = LeafOrder[i];
693 LeafMap::iterator It = Leaves.find(V);
694 if (It == Leaves.end())
695 // Node initially thought to be a leaf wasn't.
697 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
698 APInt Weight = It->second;
699 if (Weight.isMinValue())
700 // Leaf already output or weight reduction eliminated it.
702 // Ensure the leaf is only output once.
704 Ops.push_back(std::make_pair(V, Weight));
707 // For nilpotent operations or addition there may be no operands, for example
708 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
709 // in both cases the weight reduces to 0 causing the value to be skipped.
711 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
712 assert(Identity && "Associative operation without identity!");
713 Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1)));
719 // RewriteExprTree - Now that the operands for this expression tree are
720 // linearized and optimized, emit them in-order.
721 void Reassociate::RewriteExprTree(BinaryOperator *I,
722 SmallVectorImpl<ValueEntry> &Ops) {
723 assert(Ops.size() > 1 && "Single values should be used directly!");
725 // Since our optimizations should never increase the number of operations, the
726 // new expression can usually be written reusing the existing binary operators
727 // from the original expression tree, without creating any new instructions,
728 // though the rewritten expression may have a completely different topology.
729 // We take care to not change anything if the new expression will be the same
730 // as the original. If more than trivial changes (like commuting operands)
731 // were made then we are obliged to clear out any optional subclass data like
734 /// NodesToRewrite - Nodes from the original expression available for writing
735 /// the new expression into.
736 SmallVector<BinaryOperator*, 8> NodesToRewrite;
737 unsigned Opcode = I->getOpcode();
738 BinaryOperator *Op = I;
740 /// NotRewritable - The operands being written will be the leaves of the new
741 /// expression and must not be used as inner nodes (via NodesToRewrite) by
742 /// mistake. Inner nodes are always reassociable, and usually leaves are not
743 /// (if they were they would have been incorporated into the expression and so
744 /// would not be leaves), so most of the time there is no danger of this. But
745 /// in rare cases a leaf may become reassociable if an optimization kills uses
746 /// of it, or it may momentarily become reassociable during rewriting (below)
747 /// due it being removed as an operand of one of its uses. Ensure that misuse
748 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
749 /// leaves and refusing to reuse any of them as inner nodes.
750 SmallPtrSet<Value*, 8> NotRewritable;
751 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
752 NotRewritable.insert(Ops[i].Op);
754 // ExpressionChanged - Non-null if the rewritten expression differs from the
755 // original in some non-trivial way, requiring the clearing of optional flags.
756 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
757 BinaryOperator *ExpressionChanged = nullptr;
758 for (unsigned i = 0; ; ++i) {
759 // The last operation (which comes earliest in the IR) is special as both
760 // operands will come from Ops, rather than just one with the other being
762 if (i+2 == Ops.size()) {
763 Value *NewLHS = Ops[i].Op;
764 Value *NewRHS = Ops[i+1].Op;
765 Value *OldLHS = Op->getOperand(0);
766 Value *OldRHS = Op->getOperand(1);
768 if (NewLHS == OldLHS && NewRHS == OldRHS)
769 // Nothing changed, leave it alone.
772 if (NewLHS == OldRHS && NewRHS == OldLHS) {
773 // The order of the operands was reversed. Swap them.
774 DEBUG(dbgs() << "RA: " << *Op << '\n');
776 DEBUG(dbgs() << "TO: " << *Op << '\n');
782 // The new operation differs non-trivially from the original. Overwrite
783 // the old operands with the new ones.
784 DEBUG(dbgs() << "RA: " << *Op << '\n');
785 if (NewLHS != OldLHS) {
786 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
787 if (BO && !NotRewritable.count(BO))
788 NodesToRewrite.push_back(BO);
789 Op->setOperand(0, NewLHS);
791 if (NewRHS != OldRHS) {
792 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
793 if (BO && !NotRewritable.count(BO))
794 NodesToRewrite.push_back(BO);
795 Op->setOperand(1, NewRHS);
797 DEBUG(dbgs() << "TO: " << *Op << '\n');
799 ExpressionChanged = Op;
806 // Not the last operation. The left-hand side will be a sub-expression
807 // while the right-hand side will be the current element of Ops.
808 Value *NewRHS = Ops[i].Op;
809 if (NewRHS != Op->getOperand(1)) {
810 DEBUG(dbgs() << "RA: " << *Op << '\n');
811 if (NewRHS == Op->getOperand(0)) {
812 // The new right-hand side was already present as the left operand. If
813 // we are lucky then swapping the operands will sort out both of them.
816 // Overwrite with the new right-hand side.
817 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
818 if (BO && !NotRewritable.count(BO))
819 NodesToRewrite.push_back(BO);
820 Op->setOperand(1, NewRHS);
821 ExpressionChanged = Op;
823 DEBUG(dbgs() << "TO: " << *Op << '\n');
828 // Now deal with the left-hand side. If this is already an operation node
829 // from the original expression then just rewrite the rest of the expression
831 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
832 if (BO && !NotRewritable.count(BO)) {
837 // Otherwise, grab a spare node from the original expression and use that as
838 // the left-hand side. If there are no nodes left then the optimizers made
839 // an expression with more nodes than the original! This usually means that
840 // they did something stupid but it might mean that the problem was just too
841 // hard (finding the mimimal number of multiplications needed to realize a
842 // multiplication expression is NP-complete). Whatever the reason, smart or
843 // stupid, create a new node if there are none left.
844 BinaryOperator *NewOp;
845 if (NodesToRewrite.empty()) {
846 Constant *Undef = UndefValue::get(I->getType());
847 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
848 Undef, Undef, "", I);
849 if (NewOp->getType()->isFloatingPointTy())
850 NewOp->setFastMathFlags(I->getFastMathFlags());
852 NewOp = NodesToRewrite.pop_back_val();
855 DEBUG(dbgs() << "RA: " << *Op << '\n');
856 Op->setOperand(0, NewOp);
857 DEBUG(dbgs() << "TO: " << *Op << '\n');
858 ExpressionChanged = Op;
864 // If the expression changed non-trivially then clear out all subclass data
865 // starting from the operator specified in ExpressionChanged, and compactify
866 // the operators to just before the expression root to guarantee that the
867 // expression tree is dominated by all of Ops.
868 if (ExpressionChanged)
870 // Preserve FastMathFlags.
871 if (isa<FPMathOperator>(I)) {
872 FastMathFlags Flags = I->getFastMathFlags();
873 ExpressionChanged->clearSubclassOptionalData();
874 ExpressionChanged->setFastMathFlags(Flags);
876 ExpressionChanged->clearSubclassOptionalData();
878 if (ExpressionChanged == I)
880 ExpressionChanged->moveBefore(I);
881 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
884 // Throw away any left over nodes from the original expression.
885 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
886 RedoInsts.insert(NodesToRewrite[i]);
889 /// NegateValue - Insert instructions before the instruction pointed to by BI,
890 /// that computes the negative version of the value specified. The negative
891 /// version of the value is returned, and BI is left pointing at the instruction
892 /// that should be processed next by the reassociation pass.
893 static Value *NegateValue(Value *V, Instruction *BI) {
894 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
895 return ConstantExpr::getFNeg(C);
896 if (Constant *C = dyn_cast<Constant>(V))
897 return ConstantExpr::getNeg(C);
899 // We are trying to expose opportunity for reassociation. One of the things
900 // that we want to do to achieve this is to push a negation as deep into an
901 // expression chain as possible, to expose the add instructions. In practice,
902 // this means that we turn this:
903 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
904 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
905 // the constants. We assume that instcombine will clean up the mess later if
906 // we introduce tons of unnecessary negation instructions.
908 if (BinaryOperator *I =
909 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
910 // Push the negates through the add.
911 I->setOperand(0, NegateValue(I->getOperand(0), BI));
912 I->setOperand(1, NegateValue(I->getOperand(1), BI));
914 // We must move the add instruction here, because the neg instructions do
915 // not dominate the old add instruction in general. By moving it, we are
916 // assured that the neg instructions we just inserted dominate the
917 // instruction we are about to insert after them.
920 I->setName(I->getName()+".neg");
924 // Okay, we need to materialize a negated version of V with an instruction.
925 // Scan the use lists of V to see if we have one already.
926 for (User *U : V->users()) {
927 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
930 // We found one! Now we have to make sure that the definition dominates
931 // this use. We do this by moving it to the entry block (if it is a
932 // non-instruction value) or right after the definition. These negates will
933 // be zapped by reassociate later, so we don't need much finesse here.
934 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
936 // Verify that the negate is in this function, V might be a constant expr.
937 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
940 BasicBlock::iterator InsertPt;
941 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
942 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
943 InsertPt = II->getNormalDest()->begin();
945 InsertPt = InstInput;
948 while (isa<PHINode>(InsertPt)) ++InsertPt;
950 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
952 TheNeg->moveBefore(InsertPt);
956 // Insert a 'neg' instruction that subtracts the value from zero to get the
958 return CreateNeg(V, V->getName() + ".neg", BI, BI);
961 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
962 /// X-Y into (X + -Y).
963 static bool ShouldBreakUpSubtract(Instruction *Sub) {
964 // If this is a negation, we can't split it up!
965 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
968 // Don't bother to break this up unless either the LHS is an associable add or
969 // subtract or if this is only used by one.
970 Value *V0 = Sub->getOperand(0);
971 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
972 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
974 Value *V1 = Sub->getOperand(1);
975 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
976 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
978 Value *VB = Sub->user_back();
979 if (Sub->hasOneUse() &&
980 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
981 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
987 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
988 /// only used by an add, transform this into (X+(0-Y)) to promote better
990 static BinaryOperator *BreakUpSubtract(Instruction *Sub) {
991 // Convert a subtract into an add and a neg instruction. This allows sub
992 // instructions to be commuted with other add instructions.
994 // Calculate the negative value of Operand 1 of the sub instruction,
995 // and set it as the RHS of the add instruction we just made.
997 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
998 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
999 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
1000 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
1003 // Everyone now refers to the add instruction.
1004 Sub->replaceAllUsesWith(New);
1005 New->setDebugLoc(Sub->getDebugLoc());
1007 DEBUG(dbgs() << "Negated: " << *New << '\n');
1011 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
1012 /// by one, change this into a multiply by a constant to assist with further
1014 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1015 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
1016 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
1018 BinaryOperator *Mul =
1019 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
1020 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
1022 Shl->replaceAllUsesWith(Mul);
1023 Mul->setDebugLoc(Shl->getDebugLoc());
1027 /// FindInOperandList - Scan backwards and forwards among values with the same
1028 /// rank as element i to see if X exists. If X does not exist, return i. This
1029 /// is useful when scanning for 'x' when we see '-x' because they both get the
1031 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
1033 unsigned XRank = Ops[i].Rank;
1034 unsigned e = Ops.size();
1035 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
1039 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
1045 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
1046 /// and returning the result. Insert the tree before I.
1047 static Value *EmitAddTreeOfValues(Instruction *I,
1048 SmallVectorImpl<WeakVH> &Ops){
1049 if (Ops.size() == 1) return Ops.back();
1051 Value *V1 = Ops.back();
1053 Value *V2 = EmitAddTreeOfValues(I, Ops);
1054 return CreateAdd(V2, V1, "tmp", I, I);
1057 /// RemoveFactorFromExpression - If V is an expression tree that is a
1058 /// multiplication sequence, and if this sequence contains a multiply by Factor,
1059 /// remove Factor from the tree and return the new tree.
1060 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
1061 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1065 SmallVector<RepeatedValue, 8> Tree;
1066 MadeChange |= LinearizeExprTree(BO, Tree);
1067 SmallVector<ValueEntry, 8> Factors;
1068 Factors.reserve(Tree.size());
1069 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1070 RepeatedValue E = Tree[i];
1071 Factors.append(E.second.getZExtValue(),
1072 ValueEntry(getRank(E.first), E.first));
1075 bool FoundFactor = false;
1076 bool NeedsNegate = false;
1077 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1078 if (Factors[i].Op == Factor) {
1080 Factors.erase(Factors.begin()+i);
1084 // If this is a negative version of this factor, remove it.
1085 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1086 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1087 if (FC1->getValue() == -FC2->getValue()) {
1088 FoundFactor = NeedsNegate = true;
1089 Factors.erase(Factors.begin()+i);
1092 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1093 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1094 APFloat F1(FC1->getValueAPF());
1095 APFloat F2(FC2->getValueAPF());
1097 if (F1.compare(F2) == APFloat::cmpEqual) {
1098 FoundFactor = NeedsNegate = true;
1099 Factors.erase(Factors.begin() + i);
1107 // Make sure to restore the operands to the expression tree.
1108 RewriteExprTree(BO, Factors);
1112 BasicBlock::iterator InsertPt = BO; ++InsertPt;
1114 // If this was just a single multiply, remove the multiply and return the only
1115 // remaining operand.
1116 if (Factors.size() == 1) {
1117 RedoInsts.insert(BO);
1120 RewriteExprTree(BO, Factors);
1125 V = CreateNeg(V, "neg", InsertPt, BO);
1130 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
1131 /// add its operands as factors, otherwise add V to the list of factors.
1133 /// Ops is the top-level list of add operands we're trying to factor.
1134 static void FindSingleUseMultiplyFactors(Value *V,
1135 SmallVectorImpl<Value*> &Factors,
1136 const SmallVectorImpl<ValueEntry> &Ops) {
1137 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1139 Factors.push_back(V);
1143 // Otherwise, add the LHS and RHS to the list of factors.
1144 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
1145 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
1148 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
1149 /// instruction. This optimizes based on identities. If it can be reduced to
1150 /// a single Value, it is returned, otherwise the Ops list is mutated as
1152 static Value *OptimizeAndOrXor(unsigned Opcode,
1153 SmallVectorImpl<ValueEntry> &Ops) {
1154 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1155 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1156 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1157 // First, check for X and ~X in the operand list.
1158 assert(i < Ops.size());
1159 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1160 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1161 unsigned FoundX = FindInOperandList(Ops, i, X);
1163 if (Opcode == Instruction::And) // ...&X&~X = 0
1164 return Constant::getNullValue(X->getType());
1166 if (Opcode == Instruction::Or) // ...|X|~X = -1
1167 return Constant::getAllOnesValue(X->getType());
1171 // Next, check for duplicate pairs of values, which we assume are next to
1172 // each other, due to our sorting criteria.
1173 assert(i < Ops.size());
1174 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1175 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1176 // Drop duplicate values for And and Or.
1177 Ops.erase(Ops.begin()+i);
1183 // Drop pairs of values for Xor.
1184 assert(Opcode == Instruction::Xor);
1186 return Constant::getNullValue(Ops[0].Op->getType());
1189 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1197 /// Helper funciton of CombineXorOpnd(). It creates a bitwise-and
1198 /// instruction with the given two operands, and return the resulting
1199 /// instruction. There are two special cases: 1) if the constant operand is 0,
1200 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1202 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1203 const APInt &ConstOpnd) {
1204 if (ConstOpnd != 0) {
1205 if (!ConstOpnd.isAllOnesValue()) {
1206 LLVMContext &Ctx = Opnd->getType()->getContext();
1208 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1209 "and.ra", InsertBefore);
1210 I->setDebugLoc(InsertBefore->getDebugLoc());
1218 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1219 // into "R ^ C", where C would be 0, and R is a symbolic value.
1221 // If it was successful, true is returned, and the "R" and "C" is returned
1222 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1223 // and both "Res" and "ConstOpnd" remain unchanged.
1225 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1226 APInt &ConstOpnd, Value *&Res) {
1227 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1228 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1229 // = (x & ~c1) ^ (c1 ^ c2)
1230 // It is useful only when c1 == c2.
1231 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) {
1232 if (!Opnd1->getValue()->hasOneUse())
1235 const APInt &C1 = Opnd1->getConstPart();
1236 if (C1 != ConstOpnd)
1239 Value *X = Opnd1->getSymbolicPart();
1240 Res = createAndInstr(I, X, ~C1);
1241 // ConstOpnd was C2, now C1 ^ C2.
1244 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1245 RedoInsts.insert(T);
1252 // Helper function of OptimizeXor(). It tries to simplify
1253 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1256 // If it was successful, true is returned, and the "R" and "C" is returned
1257 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1258 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1259 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1260 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
1261 APInt &ConstOpnd, Value *&Res) {
1262 Value *X = Opnd1->getSymbolicPart();
1263 if (X != Opnd2->getSymbolicPart())
1266 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1267 int DeadInstNum = 1;
1268 if (Opnd1->getValue()->hasOneUse())
1270 if (Opnd2->getValue()->hasOneUse())
1274 // (x | c1) ^ (x & c2)
1275 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1276 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1277 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1279 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1280 if (Opnd2->isOrExpr())
1281 std::swap(Opnd1, Opnd2);
1283 const APInt &C1 = Opnd1->getConstPart();
1284 const APInt &C2 = Opnd2->getConstPart();
1285 APInt C3((~C1) ^ C2);
1287 // Do not increase code size!
1288 if (C3 != 0 && !C3.isAllOnesValue()) {
1289 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1290 if (NewInstNum > DeadInstNum)
1294 Res = createAndInstr(I, X, C3);
1297 } else if (Opnd1->isOrExpr()) {
1298 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1300 const APInt &C1 = Opnd1->getConstPart();
1301 const APInt &C2 = Opnd2->getConstPart();
1304 // Do not increase code size
1305 if (C3 != 0 && !C3.isAllOnesValue()) {
1306 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1307 if (NewInstNum > DeadInstNum)
1311 Res = createAndInstr(I, X, C3);
1314 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1316 const APInt &C1 = Opnd1->getConstPart();
1317 const APInt &C2 = Opnd2->getConstPart();
1319 Res = createAndInstr(I, X, C3);
1322 // Put the original operands in the Redo list; hope they will be deleted
1324 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1325 RedoInsts.insert(T);
1326 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1327 RedoInsts.insert(T);
1332 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1333 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1335 Value *Reassociate::OptimizeXor(Instruction *I,
1336 SmallVectorImpl<ValueEntry> &Ops) {
1337 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1340 if (Ops.size() == 1)
1343 SmallVector<XorOpnd, 8> Opnds;
1344 SmallVector<XorOpnd*, 8> OpndPtrs;
1345 Type *Ty = Ops[0].Op->getType();
1346 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1348 // Step 1: Convert ValueEntry to XorOpnd
1349 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1350 Value *V = Ops[i].Op;
1351 if (!isa<ConstantInt>(V)) {
1353 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1356 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1359 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1360 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1361 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1362 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1363 // when new elements are added to the vector.
1364 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1365 OpndPtrs.push_back(&Opnds[i]);
1367 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1368 // the same symbolic value cluster together. For instance, the input operand
1369 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1370 // ("x | 123", "x & 789", "y & 456").
1371 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor());
1373 // Step 3: Combine adjacent operands
1374 XorOpnd *PrevOpnd = nullptr;
1375 bool Changed = false;
1376 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1377 XorOpnd *CurrOpnd = OpndPtrs[i];
1378 // The combined value
1381 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1382 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1385 *CurrOpnd = XorOpnd(CV);
1387 CurrOpnd->Invalidate();
1392 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1393 PrevOpnd = CurrOpnd;
1397 // step 3.2: When previous and current operands share the same symbolic
1398 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1400 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1401 // Remove previous operand
1402 PrevOpnd->Invalidate();
1404 *CurrOpnd = XorOpnd(CV);
1405 PrevOpnd = CurrOpnd;
1407 CurrOpnd->Invalidate();
1414 // Step 4: Reassemble the Ops
1417 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1418 XorOpnd &O = Opnds[i];
1421 ValueEntry VE(getRank(O.getValue()), O.getValue());
1424 if (ConstOpnd != 0) {
1425 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1426 ValueEntry VE(getRank(C), C);
1429 int Sz = Ops.size();
1431 return Ops.back().Op;
1433 assert(ConstOpnd == 0);
1434 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1441 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
1442 /// optimizes based on identities. If it can be reduced to a single Value, it
1443 /// is returned, otherwise the Ops list is mutated as necessary.
1444 Value *Reassociate::OptimizeAdd(Instruction *I,
1445 SmallVectorImpl<ValueEntry> &Ops) {
1446 // Scan the operand lists looking for X and -X pairs. If we find any, we
1447 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1449 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1451 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1452 Value *TheOp = Ops[i].Op;
1453 // Check to see if we've seen this operand before. If so, we factor all
1454 // instances of the operand together. Due to our sorting criteria, we know
1455 // that these need to be next to each other in the vector.
1456 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1457 // Rescan the list, remove all instances of this operand from the expr.
1458 unsigned NumFound = 0;
1460 Ops.erase(Ops.begin()+i);
1462 } while (i != Ops.size() && Ops[i].Op == TheOp);
1464 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1467 // Insert a new multiply.
1468 Type *Ty = TheOp->getType();
1469 Constant *C = Ty->isIntegerTy() ? ConstantInt::get(Ty, NumFound)
1470 : ConstantFP::get(Ty, NumFound);
1471 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1473 // Now that we have inserted a multiply, optimize it. This allows us to
1474 // handle cases that require multiple factoring steps, such as this:
1475 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1476 RedoInsts.insert(Mul);
1478 // If every add operand was a duplicate, return the multiply.
1482 // Otherwise, we had some input that didn't have the dupe, such as
1483 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1484 // things being added by this operation.
1485 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1492 // Check for X and -X or X and ~X in the operand list.
1493 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1494 !BinaryOperator::isNot(TheOp))
1498 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1499 X = BinaryOperator::getNegArgument(TheOp);
1500 else if (BinaryOperator::isNot(TheOp))
1501 X = BinaryOperator::getNotArgument(TheOp);
1503 unsigned FoundX = FindInOperandList(Ops, i, X);
1507 // Remove X and -X from the operand list.
1508 if (Ops.size() == 2 &&
1509 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1510 return Constant::getNullValue(X->getType());
1512 // Remove X and ~X from the operand list.
1513 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1514 return Constant::getAllOnesValue(X->getType());
1516 Ops.erase(Ops.begin()+i);
1520 --i; // Need to back up an extra one.
1521 Ops.erase(Ops.begin()+FoundX);
1523 --i; // Revisit element.
1524 e -= 2; // Removed two elements.
1526 // if X and ~X we append -1 to the operand list.
1527 if (BinaryOperator::isNot(TheOp)) {
1528 Value *V = Constant::getAllOnesValue(X->getType());
1529 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1534 // Scan the operand list, checking to see if there are any common factors
1535 // between operands. Consider something like A*A+A*B*C+D. We would like to
1536 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1537 // To efficiently find this, we count the number of times a factor occurs
1538 // for any ADD operands that are MULs.
1539 DenseMap<Value*, unsigned> FactorOccurrences;
1541 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1542 // where they are actually the same multiply.
1543 unsigned MaxOcc = 0;
1544 Value *MaxOccVal = nullptr;
1545 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1546 BinaryOperator *BOp =
1547 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1551 // Compute all of the factors of this added value.
1552 SmallVector<Value*, 8> Factors;
1553 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1554 assert(Factors.size() > 1 && "Bad linearize!");
1556 // Add one to FactorOccurrences for each unique factor in this op.
1557 SmallPtrSet<Value*, 8> Duplicates;
1558 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1559 Value *Factor = Factors[i];
1560 if (!Duplicates.insert(Factor))
1563 unsigned Occ = ++FactorOccurrences[Factor];
1569 // If Factor is a negative constant, add the negated value as a factor
1570 // because we can percolate the negate out. Watch for minint, which
1571 // cannot be positivified.
1572 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1573 if (CI->isNegative() && !CI->isMinValue(true)) {
1574 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1575 assert(!Duplicates.count(Factor) &&
1576 "Shouldn't have two constant factors, missed a canonicalize");
1577 unsigned Occ = ++FactorOccurrences[Factor];
1583 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1584 if (CF->isNegative()) {
1585 APFloat F(CF->getValueAPF());
1587 Factor = ConstantFP::get(CF->getContext(), F);
1588 assert(!Duplicates.count(Factor) &&
1589 "Shouldn't have two constant factors, missed a canonicalize");
1590 unsigned Occ = ++FactorOccurrences[Factor];
1600 // If any factor occurred more than one time, we can pull it out.
1602 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1605 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1606 // this, we could otherwise run into situations where removing a factor
1607 // from an expression will drop a use of maxocc, and this can cause
1608 // RemoveFactorFromExpression on successive values to behave differently.
1609 Instruction *DummyInst =
1610 I->getType()->isIntegerTy()
1611 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1612 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1614 SmallVector<WeakVH, 4> NewMulOps;
1615 for (unsigned i = 0; i != Ops.size(); ++i) {
1616 // Only try to remove factors from expressions we're allowed to.
1617 BinaryOperator *BOp =
1618 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1622 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1623 // The factorized operand may occur several times. Convert them all in
1625 for (unsigned j = Ops.size(); j != i;) {
1627 if (Ops[j].Op == Ops[i].Op) {
1628 NewMulOps.push_back(V);
1629 Ops.erase(Ops.begin()+j);
1636 // No need for extra uses anymore.
1639 unsigned NumAddedValues = NewMulOps.size();
1640 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1642 // Now that we have inserted the add tree, optimize it. This allows us to
1643 // handle cases that require multiple factoring steps, such as this:
1644 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1645 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1646 (void)NumAddedValues;
1647 if (Instruction *VI = dyn_cast<Instruction>(V))
1648 RedoInsts.insert(VI);
1650 // Create the multiply.
1651 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1653 // Rerun associate on the multiply in case the inner expression turned into
1654 // a multiply. We want to make sure that we keep things in canonical form.
1655 RedoInsts.insert(V2);
1657 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1658 // entire result expression is just the multiply "A*(B+C)".
1662 // Otherwise, we had some input that didn't have the factor, such as
1663 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1664 // things being added by this operation.
1665 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1671 /// \brief Build up a vector of value/power pairs factoring a product.
1673 /// Given a series of multiplication operands, build a vector of factors and
1674 /// the powers each is raised to when forming the final product. Sort them in
1675 /// the order of descending power.
1677 /// (x*x) -> [(x, 2)]
1678 /// ((x*x)*x) -> [(x, 3)]
1679 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1681 /// \returns Whether any factors have a power greater than one.
1682 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1683 SmallVectorImpl<Factor> &Factors) {
1684 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1685 // Compute the sum of powers of simplifiable factors.
1686 unsigned FactorPowerSum = 0;
1687 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1688 Value *Op = Ops[Idx-1].Op;
1690 // Count the number of occurrences of this value.
1692 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1694 // Track for simplification all factors which occur 2 or more times.
1696 FactorPowerSum += Count;
1699 // We can only simplify factors if the sum of the powers of our simplifiable
1700 // factors is 4 or higher. When that is the case, we will *always* have
1701 // a simplification. This is an important invariant to prevent cyclicly
1702 // trying to simplify already minimal formations.
1703 if (FactorPowerSum < 4)
1706 // Now gather the simplifiable factors, removing them from Ops.
1708 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1709 Value *Op = Ops[Idx-1].Op;
1711 // Count the number of occurrences of this value.
1713 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1717 // Move an even number of occurrences to Factors.
1720 FactorPowerSum += Count;
1721 Factors.push_back(Factor(Op, Count));
1722 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1725 // None of the adjustments above should have reduced the sum of factor powers
1726 // below our mininum of '4'.
1727 assert(FactorPowerSum >= 4);
1729 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1733 /// \brief Build a tree of multiplies, computing the product of Ops.
1734 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1735 SmallVectorImpl<Value*> &Ops) {
1736 if (Ops.size() == 1)
1739 Value *LHS = Ops.pop_back_val();
1741 if (LHS->getType()->isIntegerTy())
1742 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1744 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1745 } while (!Ops.empty());
1750 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1752 /// Given a vector of values raised to various powers, where no two values are
1753 /// equal and the powers are sorted in decreasing order, compute the minimal
1754 /// DAG of multiplies to compute the final product, and return that product
1756 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1757 SmallVectorImpl<Factor> &Factors) {
1758 assert(Factors[0].Power);
1759 SmallVector<Value *, 4> OuterProduct;
1760 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1761 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1762 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1767 // We want to multiply across all the factors with the same power so that
1768 // we can raise them to that power as a single entity. Build a mini tree
1770 SmallVector<Value *, 4> InnerProduct;
1771 InnerProduct.push_back(Factors[LastIdx].Base);
1773 InnerProduct.push_back(Factors[Idx].Base);
1775 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1777 // Reset the base value of the first factor to the new expression tree.
1778 // We'll remove all the factors with the same power in a second pass.
1779 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1780 if (Instruction *MI = dyn_cast<Instruction>(M))
1781 RedoInsts.insert(MI);
1785 // Unique factors with equal powers -- we've folded them into the first one's
1787 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1788 Factor::PowerEqual()),
1791 // Iteratively collect the base of each factor with an add power into the
1792 // outer product, and halve each power in preparation for squaring the
1794 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1795 if (Factors[Idx].Power & 1)
1796 OuterProduct.push_back(Factors[Idx].Base);
1797 Factors[Idx].Power >>= 1;
1799 if (Factors[0].Power) {
1800 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1801 OuterProduct.push_back(SquareRoot);
1802 OuterProduct.push_back(SquareRoot);
1804 if (OuterProduct.size() == 1)
1805 return OuterProduct.front();
1807 Value *V = buildMultiplyTree(Builder, OuterProduct);
1811 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1812 SmallVectorImpl<ValueEntry> &Ops) {
1813 // We can only optimize the multiplies when there is a chain of more than
1814 // three, such that a balanced tree might require fewer total multiplies.
1818 // Try to turn linear trees of multiplies without other uses of the
1819 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1821 SmallVector<Factor, 4> Factors;
1822 if (!collectMultiplyFactors(Ops, Factors))
1823 return nullptr; // All distinct factors, so nothing left for us to do.
1825 IRBuilder<> Builder(I);
1826 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1830 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1831 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1835 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1836 SmallVectorImpl<ValueEntry> &Ops) {
1837 // Now that we have the linearized expression tree, try to optimize it.
1838 // Start by folding any constants that we found.
1839 Constant *Cst = nullptr;
1840 unsigned Opcode = I->getOpcode();
1841 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1842 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1843 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1845 // If there was nothing but constants then we are done.
1849 // Put the combined constant back at the end of the operand list, except if
1850 // there is no point. For example, an add of 0 gets dropped here, while a
1851 // multiplication by zero turns the whole expression into zero.
1852 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1853 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1855 Ops.push_back(ValueEntry(0, Cst));
1858 if (Ops.size() == 1) return Ops[0].Op;
1860 // Handle destructive annihilation due to identities between elements in the
1861 // argument list here.
1862 unsigned NumOps = Ops.size();
1865 case Instruction::And:
1866 case Instruction::Or:
1867 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1871 case Instruction::Xor:
1872 if (Value *Result = OptimizeXor(I, Ops))
1876 case Instruction::Add:
1877 case Instruction::FAdd:
1878 if (Value *Result = OptimizeAdd(I, Ops))
1882 case Instruction::Mul:
1883 case Instruction::FMul:
1884 if (Value *Result = OptimizeMul(I, Ops))
1889 if (Ops.size() != NumOps)
1890 return OptimizeExpression(I, Ops);
1894 /// EraseInst - Zap the given instruction, adding interesting operands to the
1896 void Reassociate::EraseInst(Instruction *I) {
1897 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1898 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1899 // Erase the dead instruction.
1900 ValueRankMap.erase(I);
1901 RedoInsts.remove(I);
1902 I->eraseFromParent();
1903 // Optimize its operands.
1904 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1905 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1906 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1907 // If this is a node in an expression tree, climb to the expression root
1908 // and add that since that's where optimization actually happens.
1909 unsigned Opcode = Op->getOpcode();
1910 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1912 Op = Op->user_back();
1913 RedoInsts.insert(Op);
1917 /// OptimizeInst - Inspect and optimize the given instruction. Note that erasing
1918 /// instructions is not allowed.
1919 void Reassociate::OptimizeInst(Instruction *I) {
1920 // Only consider operations that we understand.
1921 if (!isa<BinaryOperator>(I))
1924 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
1925 // If an operand of this shift is a reassociable multiply, or if the shift
1926 // is used by a reassociable multiply or add, turn into a multiply.
1927 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
1929 (isReassociableOp(I->user_back(), Instruction::Mul) ||
1930 isReassociableOp(I->user_back(), Instruction::Add)))) {
1931 Instruction *NI = ConvertShiftToMul(I);
1932 RedoInsts.insert(I);
1937 // Commute floating point binary operators, to canonicalize the order of their
1938 // operands. This can potentially expose more CSE opportunities, and makes
1939 // writing other transformations simpler.
1940 if (I->getType()->isFloatingPointTy() || I->getType()->isVectorTy()) {
1942 // FAdd and FMul can be commuted.
1943 if (I->getOpcode() == Instruction::FMul ||
1944 I->getOpcode() == Instruction::FAdd) {
1945 Value *LHS = I->getOperand(0);
1946 Value *RHS = I->getOperand(1);
1947 unsigned LHSRank = getRank(LHS);
1948 unsigned RHSRank = getRank(RHS);
1950 // Sort the operands by rank.
1951 if (RHSRank < LHSRank) {
1952 I->setOperand(0, RHS);
1953 I->setOperand(1, LHS);
1957 // FIXME: We should commute vector instructions as well. However, this
1958 // requires further analysis to determine the effect on later passes.
1960 // Don't try to optimize vector instructions or anything that doesn't have
1962 if (I->getType()->isVectorTy() || !I->hasUnsafeAlgebra())
1966 // Do not reassociate boolean (i1) expressions. We want to preserve the
1967 // original order of evaluation for short-circuited comparisons that
1968 // SimplifyCFG has folded to AND/OR expressions. If the expression
1969 // is not further optimized, it is likely to be transformed back to a
1970 // short-circuited form for code gen, and the source order may have been
1971 // optimized for the most likely conditions.
1972 if (I->getType()->isIntegerTy(1))
1975 // If this is a subtract instruction which is not already in negate form,
1976 // see if we can convert it to X+-Y.
1977 if (I->getOpcode() == Instruction::Sub) {
1978 if (ShouldBreakUpSubtract(I)) {
1979 Instruction *NI = BreakUpSubtract(I);
1980 RedoInsts.insert(I);
1983 } else if (BinaryOperator::isNeg(I)) {
1984 // Otherwise, this is a negation. See if the operand is a multiply tree
1985 // and if this is not an inner node of a multiply tree.
1986 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
1988 !isReassociableOp(I->user_back(), Instruction::Mul))) {
1989 Instruction *NI = LowerNegateToMultiply(I);
1990 RedoInsts.insert(I);
1995 } else if (I->getOpcode() == Instruction::FSub) {
1996 if (ShouldBreakUpSubtract(I)) {
1997 Instruction *NI = BreakUpSubtract(I);
1998 RedoInsts.insert(I);
2001 } else if (BinaryOperator::isFNeg(I)) {
2002 // Otherwise, this is a negation. See if the operand is a multiply tree
2003 // and if this is not an inner node of a multiply tree.
2004 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2006 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2007 Instruction *NI = LowerNegateToMultiply(I);
2008 RedoInsts.insert(I);
2015 // If this instruction is an associative binary operator, process it.
2016 if (!I->isAssociative()) return;
2017 BinaryOperator *BO = cast<BinaryOperator>(I);
2019 // If this is an interior node of a reassociable tree, ignore it until we
2020 // get to the root of the tree, to avoid N^2 analysis.
2021 unsigned Opcode = BO->getOpcode();
2022 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode)
2025 // If this is an add tree that is used by a sub instruction, ignore it
2026 // until we process the subtract.
2027 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2028 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2030 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2031 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2034 ReassociateExpression(BO);
2037 void Reassociate::ReassociateExpression(BinaryOperator *I) {
2038 assert(!I->getType()->isVectorTy() &&
2039 "Reassociation of vector instructions is not supported.");
2041 // First, walk the expression tree, linearizing the tree, collecting the
2042 // operand information.
2043 SmallVector<RepeatedValue, 8> Tree;
2044 MadeChange |= LinearizeExprTree(I, Tree);
2045 SmallVector<ValueEntry, 8> Ops;
2046 Ops.reserve(Tree.size());
2047 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2048 RepeatedValue E = Tree[i];
2049 Ops.append(E.second.getZExtValue(),
2050 ValueEntry(getRank(E.first), E.first));
2053 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2055 // Now that we have linearized the tree to a list and have gathered all of
2056 // the operands and their ranks, sort the operands by their rank. Use a
2057 // stable_sort so that values with equal ranks will have their relative
2058 // positions maintained (and so the compiler is deterministic). Note that
2059 // this sorts so that the highest ranking values end up at the beginning of
2061 std::stable_sort(Ops.begin(), Ops.end());
2063 // OptimizeExpression - Now that we have the expression tree in a convenient
2064 // sorted form, optimize it globally if possible.
2065 if (Value *V = OptimizeExpression(I, Ops)) {
2067 // Self-referential expression in unreachable code.
2069 // This expression tree simplified to something that isn't a tree,
2071 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2072 I->replaceAllUsesWith(V);
2073 if (Instruction *VI = dyn_cast<Instruction>(V))
2074 VI->setDebugLoc(I->getDebugLoc());
2075 RedoInsts.insert(I);
2080 // We want to sink immediates as deeply as possible except in the case where
2081 // this is a multiply tree used only by an add, and the immediate is a -1.
2082 // In this case we reassociate to put the negation on the outside so that we
2083 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2084 if (I->hasOneUse()) {
2085 if (I->getOpcode() == Instruction::Mul &&
2086 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2087 isa<ConstantInt>(Ops.back().Op) &&
2088 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
2089 ValueEntry Tmp = Ops.pop_back_val();
2090 Ops.insert(Ops.begin(), Tmp);
2091 } else if (I->getOpcode() == Instruction::FMul &&
2092 cast<Instruction>(I->user_back())->getOpcode() ==
2093 Instruction::FAdd &&
2094 isa<ConstantFP>(Ops.back().Op) &&
2095 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2096 ValueEntry Tmp = Ops.pop_back_val();
2097 Ops.insert(Ops.begin(), Tmp);
2101 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2103 if (Ops.size() == 1) {
2105 // Self-referential expression in unreachable code.
2108 // This expression tree simplified to something that isn't a tree,
2110 I->replaceAllUsesWith(Ops[0].Op);
2111 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2112 OI->setDebugLoc(I->getDebugLoc());
2113 RedoInsts.insert(I);
2117 // Now that we ordered and optimized the expressions, splat them back into
2118 // the expression tree, removing any unneeded nodes.
2119 RewriteExprTree(I, Ops);
2122 bool Reassociate::runOnFunction(Function &F) {
2123 if (skipOptnoneFunction(F))
2126 // Calculate the rank map for F
2130 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
2131 // Optimize every instruction in the basic block.
2132 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; )
2133 if (isInstructionTriviallyDead(II)) {
2137 assert(II->getParent() == BI && "Moved to a different block!");
2141 // If this produced extra instructions to optimize, handle them now.
2142 while (!RedoInsts.empty()) {
2143 Instruction *I = RedoInsts.pop_back_val();
2144 if (isInstructionTriviallyDead(I))
2151 // We are done with the rank map.
2153 ValueRankMap.clear();