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 optimizeFAddNegExpr(ConstantFP *ConstOperand, Instruction *I,
198 void OptimizeInst(Instruction *I);
202 XorOpnd::XorOpnd(Value *V) {
203 assert(!isa<ConstantInt>(V) && "No ConstantInt");
205 Instruction *I = dyn_cast<Instruction>(V);
208 if (I && (I->getOpcode() == Instruction::Or ||
209 I->getOpcode() == Instruction::And)) {
210 Value *V0 = I->getOperand(0);
211 Value *V1 = I->getOperand(1);
212 if (isa<ConstantInt>(V0))
215 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) {
216 ConstPart = C->getValue();
218 isOr = (I->getOpcode() == Instruction::Or);
223 // view the operand as "V | 0"
225 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth());
229 char Reassociate::ID = 0;
230 INITIALIZE_PASS(Reassociate, "reassociate",
231 "Reassociate expressions", false, false)
233 // Public interface to the Reassociate pass
234 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
236 /// isReassociableOp - Return true if V is an instruction of the specified
237 /// opcode and if it only has one use.
238 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
239 if (V->hasOneUse() && isa<Instruction>(V) &&
240 cast<Instruction>(V)->getOpcode() == Opcode)
241 return cast<BinaryOperator>(V);
245 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
247 if (V->hasOneUse() && isa<Instruction>(V) &&
248 (cast<Instruction>(V)->getOpcode() == Opcode1 ||
249 cast<Instruction>(V)->getOpcode() == Opcode2))
250 return cast<BinaryOperator>(V);
254 static bool isUnmovableInstruction(Instruction *I) {
255 switch (I->getOpcode()) {
256 case Instruction::PHI:
257 case Instruction::LandingPad:
258 case Instruction::Alloca:
259 case Instruction::Load:
260 case Instruction::Invoke:
261 case Instruction::UDiv:
262 case Instruction::SDiv:
263 case Instruction::FDiv:
264 case Instruction::URem:
265 case Instruction::SRem:
266 case Instruction::FRem:
268 case Instruction::Call:
269 return !isa<DbgInfoIntrinsic>(I);
275 void Reassociate::BuildRankMap(Function &F) {
278 // Assign distinct ranks to function arguments
279 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
280 ValueRankMap[&*I] = ++i;
282 ReversePostOrderTraversal<Function*> RPOT(&F);
283 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
284 E = RPOT.end(); I != E; ++I) {
286 unsigned BBRank = RankMap[BB] = ++i << 16;
288 // Walk the basic block, adding precomputed ranks for any instructions that
289 // we cannot move. This ensures that the ranks for these instructions are
290 // all different in the block.
291 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
292 if (isUnmovableInstruction(I))
293 ValueRankMap[&*I] = ++BBRank;
297 unsigned Reassociate::getRank(Value *V) {
298 Instruction *I = dyn_cast<Instruction>(V);
300 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
301 return 0; // Otherwise it's a global or constant, rank 0.
304 if (unsigned Rank = ValueRankMap[I])
305 return Rank; // Rank already known?
307 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
308 // we can reassociate expressions for code motion! Since we do not recurse
309 // for PHI nodes, we cannot have infinite recursion here, because there
310 // cannot be loops in the value graph that do not go through PHI nodes.
311 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
312 for (unsigned i = 0, e = I->getNumOperands();
313 i != e && Rank != MaxRank; ++i)
314 Rank = std::max(Rank, getRank(I->getOperand(i)));
316 // If this is a not or neg instruction, do not count it for rank. This
317 // assures us that X and ~X will have the same rank.
318 Type *Ty = V->getType();
319 if ((!Ty->isIntegerTy() && !Ty->isFloatingPointTy()) ||
320 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
321 !BinaryOperator::isFNeg(I)))
324 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = "
327 return ValueRankMap[I] = Rank;
330 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
331 Instruction *InsertBefore, Value *FlagsOp) {
332 if (S1->getType()->isIntegerTy())
333 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
335 BinaryOperator *Res =
336 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
337 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
342 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
343 Instruction *InsertBefore, Value *FlagsOp) {
344 if (S1->getType()->isIntegerTy())
345 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
347 BinaryOperator *Res =
348 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
349 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
354 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
355 Instruction *InsertBefore, Value *FlagsOp) {
356 if (S1->getType()->isIntegerTy())
357 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
359 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
360 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
365 /// LowerNegateToMultiply - Replace 0-X with X*-1.
367 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
368 Type *Ty = Neg->getType();
369 Constant *NegOne = Ty->isIntegerTy() ? ConstantInt::getAllOnesValue(Ty)
370 : ConstantFP::get(Ty, -1.0);
372 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
373 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
375 Neg->replaceAllUsesWith(Res);
376 Res->setDebugLoc(Neg->getDebugLoc());
380 /// CarmichaelShift - Returns k such that lambda(2^Bitwidth) = 2^k, where lambda
381 /// is the Carmichael function. This means that x^(2^k) === 1 mod 2^Bitwidth for
382 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
383 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
384 /// even x in Bitwidth-bit arithmetic.
385 static unsigned CarmichaelShift(unsigned Bitwidth) {
391 /// IncorporateWeight - Add the extra weight 'RHS' to the existing weight 'LHS',
392 /// reducing the combined weight using any special properties of the operation.
393 /// The existing weight LHS represents the computation X op X op ... op X where
394 /// X occurs LHS times. The combined weight represents X op X op ... op X with
395 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
396 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
397 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
398 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
399 // If we were working with infinite precision arithmetic then the combined
400 // weight would be LHS + RHS. But we are using finite precision arithmetic,
401 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
402 // for nilpotent operations and addition, but not for idempotent operations
403 // and multiplication), so it is important to correctly reduce the combined
404 // weight back into range if wrapping would be wrong.
406 // If RHS is zero then the weight didn't change.
407 if (RHS.isMinValue())
409 // If LHS is zero then the combined weight is RHS.
410 if (LHS.isMinValue()) {
414 // From this point on we know that neither LHS nor RHS is zero.
416 if (Instruction::isIdempotent(Opcode)) {
417 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
418 // weight of 1. Keeping weights at zero or one also means that wrapping is
420 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
421 return; // Return a weight of 1.
423 if (Instruction::isNilpotent(Opcode)) {
424 // Nilpotent means X op X === 0, so reduce weights modulo 2.
425 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
426 LHS = 0; // 1 + 1 === 0 modulo 2.
429 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
430 // TODO: Reduce the weight by exploiting nsw/nuw?
435 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
436 "Unknown associative operation!");
437 unsigned Bitwidth = LHS.getBitWidth();
438 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
439 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
440 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
441 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
442 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
443 // which by a happy accident means that they can always be represented using
445 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
446 // the Carmichael number).
448 /// CM - The value of Carmichael's lambda function.
449 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
450 // Any weight W >= Threshold can be replaced with W - CM.
451 APInt Threshold = CM + Bitwidth;
452 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
453 // For Bitwidth 4 or more the following sum does not overflow.
455 while (LHS.uge(Threshold))
458 // To avoid problems with overflow do everything the same as above but using
460 unsigned CM = 1U << CarmichaelShift(Bitwidth);
461 unsigned Threshold = CM + Bitwidth;
462 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
463 "Weights not reduced!");
464 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
465 while (Total >= Threshold)
471 typedef std::pair<Value*, APInt> RepeatedValue;
473 /// LinearizeExprTree - Given an associative binary expression, return the leaf
474 /// nodes in Ops along with their weights (how many times the leaf occurs). The
475 /// original expression is the same as
476 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
478 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
482 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
484 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
486 /// This routine may modify the function, in which case it returns 'true'. The
487 /// changes it makes may well be destructive, changing the value computed by 'I'
488 /// to something completely different. Thus if the routine returns 'true' then
489 /// you MUST either replace I with a new expression computed from the Ops array,
490 /// or use RewriteExprTree to put the values back in.
492 /// A leaf node is either not a binary operation of the same kind as the root
493 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
494 /// opcode), or is the same kind of binary operator but has a use which either
495 /// does not belong to the expression, or does belong to the expression but is
496 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
497 /// of the expression, while for non-leaf nodes (except for the root 'I') every
498 /// use is a non-leaf node of the expression.
501 /// expression graph node names
511 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
512 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
514 /// The expression is maximal: if some instruction is a binary operator of the
515 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
516 /// then the instruction also belongs to the expression, is not a leaf node of
517 /// it, and its operands also belong to the expression (but may be leaf nodes).
519 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
520 /// order to ensure that every non-root node in the expression has *exactly one*
521 /// use by a non-leaf node of the expression. This destruction means that the
522 /// caller MUST either replace 'I' with a new expression or use something like
523 /// RewriteExprTree to put the values back in if the routine indicates that it
524 /// made a change by returning 'true'.
526 /// In the above example either the right operand of A or the left operand of B
527 /// will be replaced by undef. If it is B's operand then this gives:
531 /// + + | A, B - operand of B replaced with undef
537 /// Note that such undef operands can only be reached by passing through 'I'.
538 /// For example, if you visit operands recursively starting from a leaf node
539 /// then you will never see such an undef operand unless you get back to 'I',
540 /// which requires passing through a phi node.
542 /// Note that this routine may also mutate binary operators of the wrong type
543 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
544 /// of the expression) if it can turn them into binary operators of the right
545 /// type and thus make the expression bigger.
547 static bool LinearizeExprTree(BinaryOperator *I,
548 SmallVectorImpl<RepeatedValue> &Ops) {
549 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
550 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
551 unsigned Opcode = I->getOpcode();
552 assert(I->isAssociative() && I->isCommutative() &&
553 "Expected an associative and commutative operation!");
555 // Visit all operands of the expression, keeping track of their weight (the
556 // number of paths from the expression root to the operand, or if you like
557 // the number of times that operand occurs in the linearized expression).
558 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
559 // while A has weight two.
561 // Worklist of non-leaf nodes (their operands are in the expression too) along
562 // with their weights, representing a certain number of paths to the operator.
563 // If an operator occurs in the worklist multiple times then we found multiple
564 // ways to get to it.
565 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
566 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
567 bool MadeChange = false;
569 // Leaves of the expression are values that either aren't the right kind of
570 // operation (eg: a constant, or a multiply in an add tree), or are, but have
571 // some uses that are not inside the expression. For example, in I = X + X,
572 // X = A + B, the value X has two uses (by I) that are in the expression. If
573 // X has any other uses, for example in a return instruction, then we consider
574 // X to be a leaf, and won't analyze it further. When we first visit a value,
575 // if it has more than one use then at first we conservatively consider it to
576 // be a leaf. Later, as the expression is explored, we may discover some more
577 // uses of the value from inside the expression. If all uses turn out to be
578 // from within the expression (and the value is a binary operator of the right
579 // kind) then the value is no longer considered to be a leaf, and its operands
582 // Leaves - Keeps track of the set of putative leaves as well as the number of
583 // paths to each leaf seen so far.
584 typedef DenseMap<Value*, APInt> LeafMap;
585 LeafMap Leaves; // Leaf -> Total weight so far.
586 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
589 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
591 while (!Worklist.empty()) {
592 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
593 I = P.first; // We examine the operands of this binary operator.
595 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
596 Value *Op = I->getOperand(OpIdx);
597 APInt Weight = P.second; // Number of paths to this operand.
598 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
599 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
601 // If this is a binary operation of the right kind with only one use then
602 // add its operands to the expression.
603 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
604 assert(Visited.insert(Op) && "Not first visit!");
605 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
606 Worklist.push_back(std::make_pair(BO, Weight));
610 // Appears to be a leaf. Is the operand already in the set of leaves?
611 LeafMap::iterator It = Leaves.find(Op);
612 if (It == Leaves.end()) {
613 // Not in the leaf map. Must be the first time we saw this operand.
614 assert(Visited.insert(Op) && "Not first visit!");
615 if (!Op->hasOneUse()) {
616 // This value has uses not accounted for by the expression, so it is
617 // not safe to modify. Mark it as being a leaf.
618 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
619 LeafOrder.push_back(Op);
623 // No uses outside the expression, try morphing it.
624 } else if (It != Leaves.end()) {
625 // Already in the leaf map.
626 assert(Visited.count(Op) && "In leaf map but not visited!");
628 // Update the number of paths to the leaf.
629 IncorporateWeight(It->second, Weight, Opcode);
631 #if 0 // TODO: Re-enable once PR13021 is fixed.
632 // The leaf already has one use from inside the expression. As we want
633 // exactly one such use, drop this new use of the leaf.
634 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
635 I->setOperand(OpIdx, UndefValue::get(I->getType()));
638 // If the leaf is a binary operation of the right kind and we now see
639 // that its multiple original uses were in fact all by nodes belonging
640 // to the expression, then no longer consider it to be a leaf and add
641 // its operands to the expression.
642 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
643 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
644 Worklist.push_back(std::make_pair(BO, It->second));
650 // If we still have uses that are not accounted for by the expression
651 // then it is not safe to modify the value.
652 if (!Op->hasOneUse())
655 // No uses outside the expression, try morphing it.
657 Leaves.erase(It); // Since the value may be morphed below.
660 // At this point we have a value which, first of all, is not a binary
661 // expression of the right kind, and secondly, is only used inside the
662 // expression. This means that it can safely be modified. See if we
663 // can usefully morph it into an expression of the right kind.
664 assert((!isa<Instruction>(Op) ||
665 cast<Instruction>(Op)->getOpcode() != Opcode) &&
666 "Should have been handled above!");
667 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
669 // If this is a multiply expression, turn any internal negations into
670 // multiplies by -1 so they can be reassociated.
671 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
672 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
673 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
674 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
675 BO = LowerNegateToMultiply(BO);
676 DEBUG(dbgs() << *BO << '\n');
677 Worklist.push_back(std::make_pair(BO, Weight));
682 // Failed to morph into an expression of the right type. This really is
684 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
685 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
686 LeafOrder.push_back(Op);
691 // The leaves, repeated according to their weights, represent the linearized
692 // form of the expression.
693 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
694 Value *V = LeafOrder[i];
695 LeafMap::iterator It = Leaves.find(V);
696 if (It == Leaves.end())
697 // Node initially thought to be a leaf wasn't.
699 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
700 APInt Weight = It->second;
701 if (Weight.isMinValue())
702 // Leaf already output or weight reduction eliminated it.
704 // Ensure the leaf is only output once.
706 Ops.push_back(std::make_pair(V, Weight));
709 // For nilpotent operations or addition there may be no operands, for example
710 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
711 // in both cases the weight reduces to 0 causing the value to be skipped.
713 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
714 assert(Identity && "Associative operation without identity!");
715 Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1)));
721 // RewriteExprTree - Now that the operands for this expression tree are
722 // linearized and optimized, emit them in-order.
723 void Reassociate::RewriteExprTree(BinaryOperator *I,
724 SmallVectorImpl<ValueEntry> &Ops) {
725 assert(Ops.size() > 1 && "Single values should be used directly!");
727 // Since our optimizations should never increase the number of operations, the
728 // new expression can usually be written reusing the existing binary operators
729 // from the original expression tree, without creating any new instructions,
730 // though the rewritten expression may have a completely different topology.
731 // We take care to not change anything if the new expression will be the same
732 // as the original. If more than trivial changes (like commuting operands)
733 // were made then we are obliged to clear out any optional subclass data like
736 /// NodesToRewrite - Nodes from the original expression available for writing
737 /// the new expression into.
738 SmallVector<BinaryOperator*, 8> NodesToRewrite;
739 unsigned Opcode = I->getOpcode();
740 BinaryOperator *Op = I;
742 /// NotRewritable - The operands being written will be the leaves of the new
743 /// expression and must not be used as inner nodes (via NodesToRewrite) by
744 /// mistake. Inner nodes are always reassociable, and usually leaves are not
745 /// (if they were they would have been incorporated into the expression and so
746 /// would not be leaves), so most of the time there is no danger of this. But
747 /// in rare cases a leaf may become reassociable if an optimization kills uses
748 /// of it, or it may momentarily become reassociable during rewriting (below)
749 /// due it being removed as an operand of one of its uses. Ensure that misuse
750 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
751 /// leaves and refusing to reuse any of them as inner nodes.
752 SmallPtrSet<Value*, 8> NotRewritable;
753 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
754 NotRewritable.insert(Ops[i].Op);
756 // ExpressionChanged - Non-null if the rewritten expression differs from the
757 // original in some non-trivial way, requiring the clearing of optional flags.
758 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
759 BinaryOperator *ExpressionChanged = nullptr;
760 for (unsigned i = 0; ; ++i) {
761 // The last operation (which comes earliest in the IR) is special as both
762 // operands will come from Ops, rather than just one with the other being
764 if (i+2 == Ops.size()) {
765 Value *NewLHS = Ops[i].Op;
766 Value *NewRHS = Ops[i+1].Op;
767 Value *OldLHS = Op->getOperand(0);
768 Value *OldRHS = Op->getOperand(1);
770 if (NewLHS == OldLHS && NewRHS == OldRHS)
771 // Nothing changed, leave it alone.
774 if (NewLHS == OldRHS && NewRHS == OldLHS) {
775 // The order of the operands was reversed. Swap them.
776 DEBUG(dbgs() << "RA: " << *Op << '\n');
778 DEBUG(dbgs() << "TO: " << *Op << '\n');
784 // The new operation differs non-trivially from the original. Overwrite
785 // the old operands with the new ones.
786 DEBUG(dbgs() << "RA: " << *Op << '\n');
787 if (NewLHS != OldLHS) {
788 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
789 if (BO && !NotRewritable.count(BO))
790 NodesToRewrite.push_back(BO);
791 Op->setOperand(0, NewLHS);
793 if (NewRHS != OldRHS) {
794 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
795 if (BO && !NotRewritable.count(BO))
796 NodesToRewrite.push_back(BO);
797 Op->setOperand(1, NewRHS);
799 DEBUG(dbgs() << "TO: " << *Op << '\n');
801 ExpressionChanged = Op;
808 // Not the last operation. The left-hand side will be a sub-expression
809 // while the right-hand side will be the current element of Ops.
810 Value *NewRHS = Ops[i].Op;
811 if (NewRHS != Op->getOperand(1)) {
812 DEBUG(dbgs() << "RA: " << *Op << '\n');
813 if (NewRHS == Op->getOperand(0)) {
814 // The new right-hand side was already present as the left operand. If
815 // we are lucky then swapping the operands will sort out both of them.
818 // Overwrite with the new right-hand side.
819 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
820 if (BO && !NotRewritable.count(BO))
821 NodesToRewrite.push_back(BO);
822 Op->setOperand(1, NewRHS);
823 ExpressionChanged = Op;
825 DEBUG(dbgs() << "TO: " << *Op << '\n');
830 // Now deal with the left-hand side. If this is already an operation node
831 // from the original expression then just rewrite the rest of the expression
833 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
834 if (BO && !NotRewritable.count(BO)) {
839 // Otherwise, grab a spare node from the original expression and use that as
840 // the left-hand side. If there are no nodes left then the optimizers made
841 // an expression with more nodes than the original! This usually means that
842 // they did something stupid but it might mean that the problem was just too
843 // hard (finding the mimimal number of multiplications needed to realize a
844 // multiplication expression is NP-complete). Whatever the reason, smart or
845 // stupid, create a new node if there are none left.
846 BinaryOperator *NewOp;
847 if (NodesToRewrite.empty()) {
848 Constant *Undef = UndefValue::get(I->getType());
849 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
850 Undef, Undef, "", I);
851 if (NewOp->getType()->isFloatingPointTy())
852 NewOp->setFastMathFlags(I->getFastMathFlags());
854 NewOp = NodesToRewrite.pop_back_val();
857 DEBUG(dbgs() << "RA: " << *Op << '\n');
858 Op->setOperand(0, NewOp);
859 DEBUG(dbgs() << "TO: " << *Op << '\n');
860 ExpressionChanged = Op;
866 // If the expression changed non-trivially then clear out all subclass data
867 // starting from the operator specified in ExpressionChanged, and compactify
868 // the operators to just before the expression root to guarantee that the
869 // expression tree is dominated by all of Ops.
870 if (ExpressionChanged)
872 // Preserve FastMathFlags.
873 if (isa<FPMathOperator>(I)) {
874 FastMathFlags Flags = I->getFastMathFlags();
875 ExpressionChanged->clearSubclassOptionalData();
876 ExpressionChanged->setFastMathFlags(Flags);
878 ExpressionChanged->clearSubclassOptionalData();
880 if (ExpressionChanged == I)
882 ExpressionChanged->moveBefore(I);
883 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
886 // Throw away any left over nodes from the original expression.
887 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
888 RedoInsts.insert(NodesToRewrite[i]);
891 /// NegateValue - Insert instructions before the instruction pointed to by BI,
892 /// that computes the negative version of the value specified. The negative
893 /// version of the value is returned, and BI is left pointing at the instruction
894 /// that should be processed next by the reassociation pass.
895 static Value *NegateValue(Value *V, Instruction *BI) {
896 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
897 return ConstantExpr::getFNeg(C);
898 if (Constant *C = dyn_cast<Constant>(V))
899 return ConstantExpr::getNeg(C);
901 // We are trying to expose opportunity for reassociation. One of the things
902 // that we want to do to achieve this is to push a negation as deep into an
903 // expression chain as possible, to expose the add instructions. In practice,
904 // this means that we turn this:
905 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
906 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
907 // the constants. We assume that instcombine will clean up the mess later if
908 // we introduce tons of unnecessary negation instructions.
910 if (BinaryOperator *I =
911 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
912 // Push the negates through the add.
913 I->setOperand(0, NegateValue(I->getOperand(0), BI));
914 I->setOperand(1, NegateValue(I->getOperand(1), BI));
916 // We must move the add instruction here, because the neg instructions do
917 // not dominate the old add instruction in general. By moving it, we are
918 // assured that the neg instructions we just inserted dominate the
919 // instruction we are about to insert after them.
922 I->setName(I->getName()+".neg");
926 // Okay, we need to materialize a negated version of V with an instruction.
927 // Scan the use lists of V to see if we have one already.
928 for (User *U : V->users()) {
929 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
932 // We found one! Now we have to make sure that the definition dominates
933 // this use. We do this by moving it to the entry block (if it is a
934 // non-instruction value) or right after the definition. These negates will
935 // be zapped by reassociate later, so we don't need much finesse here.
936 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
938 // Verify that the negate is in this function, V might be a constant expr.
939 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
942 BasicBlock::iterator InsertPt;
943 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
944 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
945 InsertPt = II->getNormalDest()->begin();
947 InsertPt = InstInput;
950 while (isa<PHINode>(InsertPt)) ++InsertPt;
952 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
954 TheNeg->moveBefore(InsertPt);
958 // Insert a 'neg' instruction that subtracts the value from zero to get the
960 return CreateNeg(V, V->getName() + ".neg", BI, BI);
963 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
964 /// X-Y into (X + -Y).
965 static bool ShouldBreakUpSubtract(Instruction *Sub) {
966 // If this is a negation, we can't split it up!
967 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
970 // Don't breakup X - undef.
971 if (isa<UndefValue>(Sub->getOperand(1)))
974 // Don't bother to break this up unless either the LHS is an associable add or
975 // subtract or if this is only used by one.
976 Value *V0 = Sub->getOperand(0);
977 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
978 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
980 Value *V1 = Sub->getOperand(1);
981 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
982 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
984 Value *VB = Sub->user_back();
985 if (Sub->hasOneUse() &&
986 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
987 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
993 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
994 /// only used by an add, transform this into (X+(0-Y)) to promote better
996 static BinaryOperator *BreakUpSubtract(Instruction *Sub) {
997 // Convert a subtract into an add and a neg instruction. This allows sub
998 // instructions to be commuted with other add instructions.
1000 // Calculate the negative value of Operand 1 of the sub instruction,
1001 // and set it as the RHS of the add instruction we just made.
1003 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
1004 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
1005 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
1006 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
1009 // Everyone now refers to the add instruction.
1010 Sub->replaceAllUsesWith(New);
1011 New->setDebugLoc(Sub->getDebugLoc());
1013 DEBUG(dbgs() << "Negated: " << *New << '\n');
1017 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
1018 /// by one, change this into a multiply by a constant to assist with further
1020 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1021 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
1022 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
1024 BinaryOperator *Mul =
1025 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
1026 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
1028 Shl->replaceAllUsesWith(Mul);
1029 Mul->setDebugLoc(Shl->getDebugLoc());
1033 /// FindInOperandList - Scan backwards and forwards among values with the same
1034 /// rank as element i to see if X exists. If X does not exist, return i. This
1035 /// is useful when scanning for 'x' when we see '-x' because they both get the
1037 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
1039 unsigned XRank = Ops[i].Rank;
1040 unsigned e = Ops.size();
1041 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
1044 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1045 if (Instruction *I2 = dyn_cast<Instruction>(X))
1046 if (I1->isIdenticalTo(I2))
1050 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
1053 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1054 if (Instruction *I2 = dyn_cast<Instruction>(X))
1055 if (I1->isIdenticalTo(I2))
1061 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
1062 /// and returning the result. Insert the tree before I.
1063 static Value *EmitAddTreeOfValues(Instruction *I,
1064 SmallVectorImpl<WeakVH> &Ops){
1065 if (Ops.size() == 1) return Ops.back();
1067 Value *V1 = Ops.back();
1069 Value *V2 = EmitAddTreeOfValues(I, Ops);
1070 return CreateAdd(V2, V1, "tmp", I, I);
1073 /// RemoveFactorFromExpression - If V is an expression tree that is a
1074 /// multiplication sequence, and if this sequence contains a multiply by Factor,
1075 /// remove Factor from the tree and return the new tree.
1076 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
1077 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1081 SmallVector<RepeatedValue, 8> Tree;
1082 MadeChange |= LinearizeExprTree(BO, Tree);
1083 SmallVector<ValueEntry, 8> Factors;
1084 Factors.reserve(Tree.size());
1085 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1086 RepeatedValue E = Tree[i];
1087 Factors.append(E.second.getZExtValue(),
1088 ValueEntry(getRank(E.first), E.first));
1091 bool FoundFactor = false;
1092 bool NeedsNegate = false;
1093 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1094 if (Factors[i].Op == Factor) {
1096 Factors.erase(Factors.begin()+i);
1100 // If this is a negative version of this factor, remove it.
1101 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1102 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1103 if (FC1->getValue() == -FC2->getValue()) {
1104 FoundFactor = NeedsNegate = true;
1105 Factors.erase(Factors.begin()+i);
1108 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1109 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1110 APFloat F1(FC1->getValueAPF());
1111 APFloat F2(FC2->getValueAPF());
1113 if (F1.compare(F2) == APFloat::cmpEqual) {
1114 FoundFactor = NeedsNegate = true;
1115 Factors.erase(Factors.begin() + i);
1123 // Make sure to restore the operands to the expression tree.
1124 RewriteExprTree(BO, Factors);
1128 BasicBlock::iterator InsertPt = BO; ++InsertPt;
1130 // If this was just a single multiply, remove the multiply and return the only
1131 // remaining operand.
1132 if (Factors.size() == 1) {
1133 RedoInsts.insert(BO);
1136 RewriteExprTree(BO, Factors);
1141 V = CreateNeg(V, "neg", InsertPt, BO);
1146 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
1147 /// add its operands as factors, otherwise add V to the list of factors.
1149 /// Ops is the top-level list of add operands we're trying to factor.
1150 static void FindSingleUseMultiplyFactors(Value *V,
1151 SmallVectorImpl<Value*> &Factors,
1152 const SmallVectorImpl<ValueEntry> &Ops) {
1153 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1155 Factors.push_back(V);
1159 // Otherwise, add the LHS and RHS to the list of factors.
1160 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
1161 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
1164 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
1165 /// instruction. This optimizes based on identities. If it can be reduced to
1166 /// a single Value, it is returned, otherwise the Ops list is mutated as
1168 static Value *OptimizeAndOrXor(unsigned Opcode,
1169 SmallVectorImpl<ValueEntry> &Ops) {
1170 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1171 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1172 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1173 // First, check for X and ~X in the operand list.
1174 assert(i < Ops.size());
1175 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1176 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1177 unsigned FoundX = FindInOperandList(Ops, i, X);
1179 if (Opcode == Instruction::And) // ...&X&~X = 0
1180 return Constant::getNullValue(X->getType());
1182 if (Opcode == Instruction::Or) // ...|X|~X = -1
1183 return Constant::getAllOnesValue(X->getType());
1187 // Next, check for duplicate pairs of values, which we assume are next to
1188 // each other, due to our sorting criteria.
1189 assert(i < Ops.size());
1190 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1191 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1192 // Drop duplicate values for And and Or.
1193 Ops.erase(Ops.begin()+i);
1199 // Drop pairs of values for Xor.
1200 assert(Opcode == Instruction::Xor);
1202 return Constant::getNullValue(Ops[0].Op->getType());
1205 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1213 /// Helper funciton of CombineXorOpnd(). It creates a bitwise-and
1214 /// instruction with the given two operands, and return the resulting
1215 /// instruction. There are two special cases: 1) if the constant operand is 0,
1216 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1218 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1219 const APInt &ConstOpnd) {
1220 if (ConstOpnd != 0) {
1221 if (!ConstOpnd.isAllOnesValue()) {
1222 LLVMContext &Ctx = Opnd->getType()->getContext();
1224 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1225 "and.ra", InsertBefore);
1226 I->setDebugLoc(InsertBefore->getDebugLoc());
1234 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1235 // into "R ^ C", where C would be 0, and R is a symbolic value.
1237 // If it was successful, true is returned, and the "R" and "C" is returned
1238 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1239 // and both "Res" and "ConstOpnd" remain unchanged.
1241 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1242 APInt &ConstOpnd, Value *&Res) {
1243 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1244 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1245 // = (x & ~c1) ^ (c1 ^ c2)
1246 // It is useful only when c1 == c2.
1247 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) {
1248 if (!Opnd1->getValue()->hasOneUse())
1251 const APInt &C1 = Opnd1->getConstPart();
1252 if (C1 != ConstOpnd)
1255 Value *X = Opnd1->getSymbolicPart();
1256 Res = createAndInstr(I, X, ~C1);
1257 // ConstOpnd was C2, now C1 ^ C2.
1260 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1261 RedoInsts.insert(T);
1268 // Helper function of OptimizeXor(). It tries to simplify
1269 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1272 // If it was successful, true is returned, and the "R" and "C" is returned
1273 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1274 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1275 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1276 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
1277 APInt &ConstOpnd, Value *&Res) {
1278 Value *X = Opnd1->getSymbolicPart();
1279 if (X != Opnd2->getSymbolicPart())
1282 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1283 int DeadInstNum = 1;
1284 if (Opnd1->getValue()->hasOneUse())
1286 if (Opnd2->getValue()->hasOneUse())
1290 // (x | c1) ^ (x & c2)
1291 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1292 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1293 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1295 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1296 if (Opnd2->isOrExpr())
1297 std::swap(Opnd1, Opnd2);
1299 const APInt &C1 = Opnd1->getConstPart();
1300 const APInt &C2 = Opnd2->getConstPart();
1301 APInt C3((~C1) ^ C2);
1303 // Do not increase code size!
1304 if (C3 != 0 && !C3.isAllOnesValue()) {
1305 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1306 if (NewInstNum > DeadInstNum)
1310 Res = createAndInstr(I, X, C3);
1313 } else if (Opnd1->isOrExpr()) {
1314 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1316 const APInt &C1 = Opnd1->getConstPart();
1317 const APInt &C2 = Opnd2->getConstPart();
1320 // Do not increase code size
1321 if (C3 != 0 && !C3.isAllOnesValue()) {
1322 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1323 if (NewInstNum > DeadInstNum)
1327 Res = createAndInstr(I, X, C3);
1330 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1332 const APInt &C1 = Opnd1->getConstPart();
1333 const APInt &C2 = Opnd2->getConstPart();
1335 Res = createAndInstr(I, X, C3);
1338 // Put the original operands in the Redo list; hope they will be deleted
1340 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1341 RedoInsts.insert(T);
1342 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1343 RedoInsts.insert(T);
1348 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1349 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1351 Value *Reassociate::OptimizeXor(Instruction *I,
1352 SmallVectorImpl<ValueEntry> &Ops) {
1353 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1356 if (Ops.size() == 1)
1359 SmallVector<XorOpnd, 8> Opnds;
1360 SmallVector<XorOpnd*, 8> OpndPtrs;
1361 Type *Ty = Ops[0].Op->getType();
1362 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1364 // Step 1: Convert ValueEntry to XorOpnd
1365 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1366 Value *V = Ops[i].Op;
1367 if (!isa<ConstantInt>(V)) {
1369 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1372 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1375 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1376 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1377 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1378 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1379 // when new elements are added to the vector.
1380 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1381 OpndPtrs.push_back(&Opnds[i]);
1383 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1384 // the same symbolic value cluster together. For instance, the input operand
1385 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1386 // ("x | 123", "x & 789", "y & 456").
1387 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor());
1389 // Step 3: Combine adjacent operands
1390 XorOpnd *PrevOpnd = nullptr;
1391 bool Changed = false;
1392 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1393 XorOpnd *CurrOpnd = OpndPtrs[i];
1394 // The combined value
1397 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1398 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1401 *CurrOpnd = XorOpnd(CV);
1403 CurrOpnd->Invalidate();
1408 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1409 PrevOpnd = CurrOpnd;
1413 // step 3.2: When previous and current operands share the same symbolic
1414 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1416 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1417 // Remove previous operand
1418 PrevOpnd->Invalidate();
1420 *CurrOpnd = XorOpnd(CV);
1421 PrevOpnd = CurrOpnd;
1423 CurrOpnd->Invalidate();
1430 // Step 4: Reassemble the Ops
1433 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1434 XorOpnd &O = Opnds[i];
1437 ValueEntry VE(getRank(O.getValue()), O.getValue());
1440 if (ConstOpnd != 0) {
1441 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1442 ValueEntry VE(getRank(C), C);
1445 int Sz = Ops.size();
1447 return Ops.back().Op;
1449 assert(ConstOpnd == 0);
1450 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1457 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
1458 /// optimizes based on identities. If it can be reduced to a single Value, it
1459 /// is returned, otherwise the Ops list is mutated as necessary.
1460 Value *Reassociate::OptimizeAdd(Instruction *I,
1461 SmallVectorImpl<ValueEntry> &Ops) {
1462 // Scan the operand lists looking for X and -X pairs. If we find any, we
1463 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1465 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1467 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1468 Value *TheOp = Ops[i].Op;
1469 // Check to see if we've seen this operand before. If so, we factor all
1470 // instances of the operand together. Due to our sorting criteria, we know
1471 // that these need to be next to each other in the vector.
1472 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1473 // Rescan the list, remove all instances of this operand from the expr.
1474 unsigned NumFound = 0;
1476 Ops.erase(Ops.begin()+i);
1478 } while (i != Ops.size() && Ops[i].Op == TheOp);
1480 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1483 // Insert a new multiply.
1484 Type *Ty = TheOp->getType();
1485 Constant *C = Ty->isIntegerTy() ? ConstantInt::get(Ty, NumFound)
1486 : ConstantFP::get(Ty, NumFound);
1487 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1489 // Now that we have inserted a multiply, optimize it. This allows us to
1490 // handle cases that require multiple factoring steps, such as this:
1491 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1492 RedoInsts.insert(Mul);
1494 // If every add operand was a duplicate, return the multiply.
1498 // Otherwise, we had some input that didn't have the dupe, such as
1499 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1500 // things being added by this operation.
1501 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1508 // Check for X and -X or X and ~X in the operand list.
1509 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1510 !BinaryOperator::isNot(TheOp))
1514 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1515 X = BinaryOperator::getNegArgument(TheOp);
1516 else if (BinaryOperator::isNot(TheOp))
1517 X = BinaryOperator::getNotArgument(TheOp);
1519 unsigned FoundX = FindInOperandList(Ops, i, X);
1523 // Remove X and -X from the operand list.
1524 if (Ops.size() == 2 &&
1525 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1526 return Constant::getNullValue(X->getType());
1528 // Remove X and ~X from the operand list.
1529 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1530 return Constant::getAllOnesValue(X->getType());
1532 Ops.erase(Ops.begin()+i);
1536 --i; // Need to back up an extra one.
1537 Ops.erase(Ops.begin()+FoundX);
1539 --i; // Revisit element.
1540 e -= 2; // Removed two elements.
1542 // if X and ~X we append -1 to the operand list.
1543 if (BinaryOperator::isNot(TheOp)) {
1544 Value *V = Constant::getAllOnesValue(X->getType());
1545 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1550 // Scan the operand list, checking to see if there are any common factors
1551 // between operands. Consider something like A*A+A*B*C+D. We would like to
1552 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1553 // To efficiently find this, we count the number of times a factor occurs
1554 // for any ADD operands that are MULs.
1555 DenseMap<Value*, unsigned> FactorOccurrences;
1557 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1558 // where they are actually the same multiply.
1559 unsigned MaxOcc = 0;
1560 Value *MaxOccVal = nullptr;
1561 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1562 BinaryOperator *BOp =
1563 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1567 // Compute all of the factors of this added value.
1568 SmallVector<Value*, 8> Factors;
1569 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1570 assert(Factors.size() > 1 && "Bad linearize!");
1572 // Add one to FactorOccurrences for each unique factor in this op.
1573 SmallPtrSet<Value*, 8> Duplicates;
1574 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1575 Value *Factor = Factors[i];
1576 if (!Duplicates.insert(Factor))
1579 unsigned Occ = ++FactorOccurrences[Factor];
1585 // If Factor is a negative constant, add the negated value as a factor
1586 // because we can percolate the negate out. Watch for minint, which
1587 // cannot be positivified.
1588 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1589 if (CI->isNegative() && !CI->isMinValue(true)) {
1590 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1591 assert(!Duplicates.count(Factor) &&
1592 "Shouldn't have two constant factors, missed a canonicalize");
1593 unsigned Occ = ++FactorOccurrences[Factor];
1599 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1600 if (CF->isNegative()) {
1601 APFloat F(CF->getValueAPF());
1603 Factor = ConstantFP::get(CF->getContext(), F);
1604 assert(!Duplicates.count(Factor) &&
1605 "Shouldn't have two constant factors, missed a canonicalize");
1606 unsigned Occ = ++FactorOccurrences[Factor];
1616 // If any factor occurred more than one time, we can pull it out.
1618 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1621 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1622 // this, we could otherwise run into situations where removing a factor
1623 // from an expression will drop a use of maxocc, and this can cause
1624 // RemoveFactorFromExpression on successive values to behave differently.
1625 Instruction *DummyInst =
1626 I->getType()->isIntegerTy()
1627 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1628 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1630 SmallVector<WeakVH, 4> NewMulOps;
1631 for (unsigned i = 0; i != Ops.size(); ++i) {
1632 // Only try to remove factors from expressions we're allowed to.
1633 BinaryOperator *BOp =
1634 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1638 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1639 // The factorized operand may occur several times. Convert them all in
1641 for (unsigned j = Ops.size(); j != i;) {
1643 if (Ops[j].Op == Ops[i].Op) {
1644 NewMulOps.push_back(V);
1645 Ops.erase(Ops.begin()+j);
1652 // No need for extra uses anymore.
1655 unsigned NumAddedValues = NewMulOps.size();
1656 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1658 // Now that we have inserted the add tree, optimize it. This allows us to
1659 // handle cases that require multiple factoring steps, such as this:
1660 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1661 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1662 (void)NumAddedValues;
1663 if (Instruction *VI = dyn_cast<Instruction>(V))
1664 RedoInsts.insert(VI);
1666 // Create the multiply.
1667 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1669 // Rerun associate on the multiply in case the inner expression turned into
1670 // a multiply. We want to make sure that we keep things in canonical form.
1671 RedoInsts.insert(V2);
1673 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1674 // entire result expression is just the multiply "A*(B+C)".
1678 // Otherwise, we had some input that didn't have the factor, such as
1679 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1680 // things being added by this operation.
1681 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1687 /// \brief Build up a vector of value/power pairs factoring a product.
1689 /// Given a series of multiplication operands, build a vector of factors and
1690 /// the powers each is raised to when forming the final product. Sort them in
1691 /// the order of descending power.
1693 /// (x*x) -> [(x, 2)]
1694 /// ((x*x)*x) -> [(x, 3)]
1695 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1697 /// \returns Whether any factors have a power greater than one.
1698 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1699 SmallVectorImpl<Factor> &Factors) {
1700 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1701 // Compute the sum of powers of simplifiable factors.
1702 unsigned FactorPowerSum = 0;
1703 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1704 Value *Op = Ops[Idx-1].Op;
1706 // Count the number of occurrences of this value.
1708 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1710 // Track for simplification all factors which occur 2 or more times.
1712 FactorPowerSum += Count;
1715 // We can only simplify factors if the sum of the powers of our simplifiable
1716 // factors is 4 or higher. When that is the case, we will *always* have
1717 // a simplification. This is an important invariant to prevent cyclicly
1718 // trying to simplify already minimal formations.
1719 if (FactorPowerSum < 4)
1722 // Now gather the simplifiable factors, removing them from Ops.
1724 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1725 Value *Op = Ops[Idx-1].Op;
1727 // Count the number of occurrences of this value.
1729 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1733 // Move an even number of occurrences to Factors.
1736 FactorPowerSum += Count;
1737 Factors.push_back(Factor(Op, Count));
1738 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1741 // None of the adjustments above should have reduced the sum of factor powers
1742 // below our mininum of '4'.
1743 assert(FactorPowerSum >= 4);
1745 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1749 /// \brief Build a tree of multiplies, computing the product of Ops.
1750 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1751 SmallVectorImpl<Value*> &Ops) {
1752 if (Ops.size() == 1)
1755 Value *LHS = Ops.pop_back_val();
1757 if (LHS->getType()->isIntegerTy())
1758 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1760 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1761 } while (!Ops.empty());
1766 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1768 /// Given a vector of values raised to various powers, where no two values are
1769 /// equal and the powers are sorted in decreasing order, compute the minimal
1770 /// DAG of multiplies to compute the final product, and return that product
1772 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1773 SmallVectorImpl<Factor> &Factors) {
1774 assert(Factors[0].Power);
1775 SmallVector<Value *, 4> OuterProduct;
1776 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1777 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1778 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1783 // We want to multiply across all the factors with the same power so that
1784 // we can raise them to that power as a single entity. Build a mini tree
1786 SmallVector<Value *, 4> InnerProduct;
1787 InnerProduct.push_back(Factors[LastIdx].Base);
1789 InnerProduct.push_back(Factors[Idx].Base);
1791 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1793 // Reset the base value of the first factor to the new expression tree.
1794 // We'll remove all the factors with the same power in a second pass.
1795 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1796 if (Instruction *MI = dyn_cast<Instruction>(M))
1797 RedoInsts.insert(MI);
1801 // Unique factors with equal powers -- we've folded them into the first one's
1803 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1804 Factor::PowerEqual()),
1807 // Iteratively collect the base of each factor with an add power into the
1808 // outer product, and halve each power in preparation for squaring the
1810 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1811 if (Factors[Idx].Power & 1)
1812 OuterProduct.push_back(Factors[Idx].Base);
1813 Factors[Idx].Power >>= 1;
1815 if (Factors[0].Power) {
1816 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1817 OuterProduct.push_back(SquareRoot);
1818 OuterProduct.push_back(SquareRoot);
1820 if (OuterProduct.size() == 1)
1821 return OuterProduct.front();
1823 Value *V = buildMultiplyTree(Builder, OuterProduct);
1827 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1828 SmallVectorImpl<ValueEntry> &Ops) {
1829 // We can only optimize the multiplies when there is a chain of more than
1830 // three, such that a balanced tree might require fewer total multiplies.
1834 // Try to turn linear trees of multiplies without other uses of the
1835 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1837 SmallVector<Factor, 4> Factors;
1838 if (!collectMultiplyFactors(Ops, Factors))
1839 return nullptr; // All distinct factors, so nothing left for us to do.
1841 IRBuilder<> Builder(I);
1842 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1846 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1847 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1851 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1852 SmallVectorImpl<ValueEntry> &Ops) {
1853 // Now that we have the linearized expression tree, try to optimize it.
1854 // Start by folding any constants that we found.
1855 Constant *Cst = nullptr;
1856 unsigned Opcode = I->getOpcode();
1857 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1858 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1859 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1861 // If there was nothing but constants then we are done.
1865 // Put the combined constant back at the end of the operand list, except if
1866 // there is no point. For example, an add of 0 gets dropped here, while a
1867 // multiplication by zero turns the whole expression into zero.
1868 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1869 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1871 Ops.push_back(ValueEntry(0, Cst));
1874 if (Ops.size() == 1) return Ops[0].Op;
1876 // Handle destructive annihilation due to identities between elements in the
1877 // argument list here.
1878 unsigned NumOps = Ops.size();
1881 case Instruction::And:
1882 case Instruction::Or:
1883 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1887 case Instruction::Xor:
1888 if (Value *Result = OptimizeXor(I, Ops))
1892 case Instruction::Add:
1893 case Instruction::FAdd:
1894 if (Value *Result = OptimizeAdd(I, Ops))
1898 case Instruction::Mul:
1899 case Instruction::FMul:
1900 if (Value *Result = OptimizeMul(I, Ops))
1905 if (Ops.size() != NumOps)
1906 return OptimizeExpression(I, Ops);
1910 /// EraseInst - Zap the given instruction, adding interesting operands to the
1912 void Reassociate::EraseInst(Instruction *I) {
1913 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1914 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1915 // Erase the dead instruction.
1916 ValueRankMap.erase(I);
1917 RedoInsts.remove(I);
1918 I->eraseFromParent();
1919 // Optimize its operands.
1920 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1921 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1922 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1923 // If this is a node in an expression tree, climb to the expression root
1924 // and add that since that's where optimization actually happens.
1925 unsigned Opcode = Op->getOpcode();
1926 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1928 Op = Op->user_back();
1929 RedoInsts.insert(Op);
1933 void Reassociate::optimizeFAddNegExpr(ConstantFP *ConstOperand, Instruction *I,
1935 // Change the sign of the constant.
1936 APFloat Val = ConstOperand->getValueAPF();
1938 I->setOperand(0, ConstantFP::get(ConstOperand->getContext(), Val));
1940 assert(I->hasOneUse() && "Only a single use can be replaced.");
1941 Instruction *Parent = I->user_back();
1943 Value *OtherOperand = Parent->getOperand(1 - OperandNr);
1945 unsigned Opcode = Parent->getOpcode();
1946 assert(Opcode == Instruction::FAdd ||
1947 (Opcode == Instruction::FSub && Parent->getOperand(1) == I));
1949 BinaryOperator *NI = Opcode == Instruction::FAdd
1950 ? BinaryOperator::CreateFSub(OtherOperand, I)
1951 : BinaryOperator::CreateFAdd(OtherOperand, I);
1952 NI->setFastMathFlags(cast<FPMathOperator>(Parent)->getFastMathFlags());
1953 NI->insertBefore(Parent);
1954 NI->setName(Parent->getName() + ".repl");
1955 Parent->replaceAllUsesWith(NI);
1956 NI->setDebugLoc(I->getDebugLoc());
1960 /// OptimizeInst - Inspect and optimize the given instruction. Note that erasing
1961 /// instructions is not allowed.
1962 void Reassociate::OptimizeInst(Instruction *I) {
1963 // Only consider operations that we understand.
1964 if (!isa<BinaryOperator>(I))
1967 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
1968 // If an operand of this shift is a reassociable multiply, or if the shift
1969 // is used by a reassociable multiply or add, turn into a multiply.
1970 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
1972 (isReassociableOp(I->user_back(), Instruction::Mul) ||
1973 isReassociableOp(I->user_back(), Instruction::Add)))) {
1974 Instruction *NI = ConvertShiftToMul(I);
1975 RedoInsts.insert(I);
1980 // Commute floating point binary operators, to canonicalize the order of their
1981 // operands. This can potentially expose more CSE opportunities, and makes
1982 // writing other transformations simpler.
1983 if (I->getType()->isFloatingPointTy() || I->getType()->isVectorTy()) {
1985 // FAdd and FMul can be commuted.
1986 unsigned Opcode = I->getOpcode();
1987 if (Opcode == Instruction::FMul || Opcode == Instruction::FAdd) {
1988 Value *LHS = I->getOperand(0);
1989 Value *RHS = I->getOperand(1);
1990 unsigned LHSRank = getRank(LHS);
1991 unsigned RHSRank = getRank(RHS);
1993 // Sort the operands by rank.
1994 if (RHSRank < LHSRank) {
1995 I->setOperand(0, RHS);
1996 I->setOperand(1, LHS);
2000 // Reassociate: x + -ConstantFP * y -> x - ConstantFP * y
2001 // The FMul can also be an FDiv, and FAdd can be a FSub.
2002 if (Opcode == Instruction::FMul || Opcode == Instruction::FDiv) {
2003 if (ConstantFP *LHSConst = dyn_cast<ConstantFP>(I->getOperand(0))) {
2004 if (LHSConst->isNegative() && I->hasOneUse()) {
2005 Instruction *Parent = I->user_back();
2006 if (Parent->getOpcode() == Instruction::FAdd) {
2007 if (Parent->getOperand(0) == I)
2008 optimizeFAddNegExpr(LHSConst, I, 0);
2009 else if (Parent->getOperand(1) == I)
2010 optimizeFAddNegExpr(LHSConst, I, 1);
2011 } else if (Parent->getOpcode() == Instruction::FSub)
2012 if (Parent->getOperand(1) == I)
2013 optimizeFAddNegExpr(LHSConst, I, 1);
2018 // FIXME: We should commute vector instructions as well. However, this
2019 // requires further analysis to determine the effect on later passes.
2021 // Don't try to optimize vector instructions or anything that doesn't have
2023 if (I->getType()->isVectorTy() || !I->hasUnsafeAlgebra())
2027 // Do not reassociate boolean (i1) expressions. We want to preserve the
2028 // original order of evaluation for short-circuited comparisons that
2029 // SimplifyCFG has folded to AND/OR expressions. If the expression
2030 // is not further optimized, it is likely to be transformed back to a
2031 // short-circuited form for code gen, and the source order may have been
2032 // optimized for the most likely conditions.
2033 if (I->getType()->isIntegerTy(1))
2036 // If this is a subtract instruction which is not already in negate form,
2037 // see if we can convert it to X+-Y.
2038 if (I->getOpcode() == Instruction::Sub) {
2039 if (ShouldBreakUpSubtract(I)) {
2040 Instruction *NI = BreakUpSubtract(I);
2041 RedoInsts.insert(I);
2044 } else if (BinaryOperator::isNeg(I)) {
2045 // Otherwise, this is a negation. See if the operand is a multiply tree
2046 // and if this is not an inner node of a multiply tree.
2047 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2049 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2050 Instruction *NI = LowerNegateToMultiply(I);
2051 RedoInsts.insert(I);
2056 } else if (I->getOpcode() == Instruction::FSub) {
2057 if (ShouldBreakUpSubtract(I)) {
2058 Instruction *NI = BreakUpSubtract(I);
2059 RedoInsts.insert(I);
2062 } else if (BinaryOperator::isFNeg(I)) {
2063 // Otherwise, this is a negation. See if the operand is a multiply tree
2064 // and if this is not an inner node of a multiply tree.
2065 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2067 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2068 Instruction *NI = LowerNegateToMultiply(I);
2069 RedoInsts.insert(I);
2076 // If this instruction is an associative binary operator, process it.
2077 if (!I->isAssociative()) return;
2078 BinaryOperator *BO = cast<BinaryOperator>(I);
2080 // If this is an interior node of a reassociable tree, ignore it until we
2081 // get to the root of the tree, to avoid N^2 analysis.
2082 unsigned Opcode = BO->getOpcode();
2083 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode)
2086 // If this is an add tree that is used by a sub instruction, ignore it
2087 // until we process the subtract.
2088 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2089 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2091 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2092 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2095 ReassociateExpression(BO);
2098 void Reassociate::ReassociateExpression(BinaryOperator *I) {
2099 assert(!I->getType()->isVectorTy() &&
2100 "Reassociation of vector instructions is not supported.");
2102 // First, walk the expression tree, linearizing the tree, collecting the
2103 // operand information.
2104 SmallVector<RepeatedValue, 8> Tree;
2105 MadeChange |= LinearizeExprTree(I, Tree);
2106 SmallVector<ValueEntry, 8> Ops;
2107 Ops.reserve(Tree.size());
2108 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2109 RepeatedValue E = Tree[i];
2110 Ops.append(E.second.getZExtValue(),
2111 ValueEntry(getRank(E.first), E.first));
2114 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2116 // Now that we have linearized the tree to a list and have gathered all of
2117 // the operands and their ranks, sort the operands by their rank. Use a
2118 // stable_sort so that values with equal ranks will have their relative
2119 // positions maintained (and so the compiler is deterministic). Note that
2120 // this sorts so that the highest ranking values end up at the beginning of
2122 std::stable_sort(Ops.begin(), Ops.end());
2124 // OptimizeExpression - Now that we have the expression tree in a convenient
2125 // sorted form, optimize it globally if possible.
2126 if (Value *V = OptimizeExpression(I, Ops)) {
2128 // Self-referential expression in unreachable code.
2130 // This expression tree simplified to something that isn't a tree,
2132 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2133 I->replaceAllUsesWith(V);
2134 if (Instruction *VI = dyn_cast<Instruction>(V))
2135 VI->setDebugLoc(I->getDebugLoc());
2136 RedoInsts.insert(I);
2141 // We want to sink immediates as deeply as possible except in the case where
2142 // this is a multiply tree used only by an add, and the immediate is a -1.
2143 // In this case we reassociate to put the negation on the outside so that we
2144 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2145 if (I->hasOneUse()) {
2146 if (I->getOpcode() == Instruction::Mul &&
2147 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2148 isa<ConstantInt>(Ops.back().Op) &&
2149 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
2150 ValueEntry Tmp = Ops.pop_back_val();
2151 Ops.insert(Ops.begin(), Tmp);
2152 } else if (I->getOpcode() == Instruction::FMul &&
2153 cast<Instruction>(I->user_back())->getOpcode() ==
2154 Instruction::FAdd &&
2155 isa<ConstantFP>(Ops.back().Op) &&
2156 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2157 ValueEntry Tmp = Ops.pop_back_val();
2158 Ops.insert(Ops.begin(), Tmp);
2162 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2164 if (Ops.size() == 1) {
2166 // Self-referential expression in unreachable code.
2169 // This expression tree simplified to something that isn't a tree,
2171 I->replaceAllUsesWith(Ops[0].Op);
2172 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2173 OI->setDebugLoc(I->getDebugLoc());
2174 RedoInsts.insert(I);
2178 // Now that we ordered and optimized the expressions, splat them back into
2179 // the expression tree, removing any unneeded nodes.
2180 RewriteExprTree(I, Ops);
2183 bool Reassociate::runOnFunction(Function &F) {
2184 if (skipOptnoneFunction(F))
2187 // Calculate the rank map for F
2191 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
2192 // Optimize every instruction in the basic block.
2193 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; )
2194 if (isInstructionTriviallyDead(II)) {
2198 assert(II->getParent() == BI && "Moved to a different block!");
2202 // If this produced extra instructions to optimize, handle them now.
2203 while (!RedoInsts.empty()) {
2204 Instruction *I = RedoInsts.pop_back_val();
2205 if (isInstructionTriviallyDead(I))
2212 // We are done with the rank map.
2214 ValueRankMap.clear();