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);
197 Instruction *canonicalizeNegConstExpr(Instruction *I);
201 XorOpnd::XorOpnd(Value *V) {
202 assert(!isa<ConstantInt>(V) && "No ConstantInt");
204 Instruction *I = dyn_cast<Instruction>(V);
207 if (I && (I->getOpcode() == Instruction::Or ||
208 I->getOpcode() == Instruction::And)) {
209 Value *V0 = I->getOperand(0);
210 Value *V1 = I->getOperand(1);
211 if (isa<ConstantInt>(V0))
214 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) {
215 ConstPart = C->getValue();
217 isOr = (I->getOpcode() == Instruction::Or);
222 // view the operand as "V | 0"
224 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth());
228 char Reassociate::ID = 0;
229 INITIALIZE_PASS(Reassociate, "reassociate",
230 "Reassociate expressions", false, false)
232 // Public interface to the Reassociate pass
233 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
235 /// isReassociableOp - Return true if V is an instruction of the specified
236 /// opcode and if it only has one use.
237 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
238 if (V->hasOneUse() && isa<Instruction>(V) &&
239 cast<Instruction>(V)->getOpcode() == Opcode)
240 return cast<BinaryOperator>(V);
244 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
246 if (V->hasOneUse() && isa<Instruction>(V) &&
247 (cast<Instruction>(V)->getOpcode() == Opcode1 ||
248 cast<Instruction>(V)->getOpcode() == Opcode2))
249 return cast<BinaryOperator>(V);
253 static bool isUnmovableInstruction(Instruction *I) {
254 switch (I->getOpcode()) {
255 case Instruction::PHI:
256 case Instruction::LandingPad:
257 case Instruction::Alloca:
258 case Instruction::Load:
259 case Instruction::Invoke:
260 case Instruction::UDiv:
261 case Instruction::SDiv:
262 case Instruction::FDiv:
263 case Instruction::URem:
264 case Instruction::SRem:
265 case Instruction::FRem:
267 case Instruction::Call:
268 return !isa<DbgInfoIntrinsic>(I);
274 void Reassociate::BuildRankMap(Function &F) {
277 // Assign distinct ranks to function arguments
278 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
279 ValueRankMap[&*I] = ++i;
281 ReversePostOrderTraversal<Function*> RPOT(&F);
282 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
283 E = RPOT.end(); I != E; ++I) {
285 unsigned BBRank = RankMap[BB] = ++i << 16;
287 // Walk the basic block, adding precomputed ranks for any instructions that
288 // we cannot move. This ensures that the ranks for these instructions are
289 // all different in the block.
290 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
291 if (isUnmovableInstruction(I))
292 ValueRankMap[&*I] = ++BBRank;
296 unsigned Reassociate::getRank(Value *V) {
297 Instruction *I = dyn_cast<Instruction>(V);
299 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
300 return 0; // Otherwise it's a global or constant, rank 0.
303 if (unsigned Rank = ValueRankMap[I])
304 return Rank; // Rank already known?
306 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
307 // we can reassociate expressions for code motion! Since we do not recurse
308 // for PHI nodes, we cannot have infinite recursion here, because there
309 // cannot be loops in the value graph that do not go through PHI nodes.
310 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
311 for (unsigned i = 0, e = I->getNumOperands();
312 i != e && Rank != MaxRank; ++i)
313 Rank = std::max(Rank, getRank(I->getOperand(i)));
315 // If this is a not or neg instruction, do not count it for rank. This
316 // assures us that X and ~X will have the same rank.
317 Type *Ty = V->getType();
318 if ((!Ty->isIntegerTy() && !Ty->isFloatingPointTy()) ||
319 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
320 !BinaryOperator::isFNeg(I)))
323 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = "
326 return ValueRankMap[I] = Rank;
329 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
330 Instruction *InsertBefore, Value *FlagsOp) {
331 if (S1->getType()->isIntegerTy())
332 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
334 BinaryOperator *Res =
335 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
336 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
341 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
342 Instruction *InsertBefore, Value *FlagsOp) {
343 if (S1->getType()->isIntegerTy())
344 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
346 BinaryOperator *Res =
347 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
348 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
353 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
354 Instruction *InsertBefore, Value *FlagsOp) {
355 if (S1->getType()->isIntegerTy())
356 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
358 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
359 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
364 /// LowerNegateToMultiply - Replace 0-X with X*-1.
366 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
367 Type *Ty = Neg->getType();
368 Constant *NegOne = Ty->isIntegerTy() ? ConstantInt::getAllOnesValue(Ty)
369 : ConstantFP::get(Ty, -1.0);
371 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
372 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
374 Neg->replaceAllUsesWith(Res);
375 Res->setDebugLoc(Neg->getDebugLoc());
379 /// CarmichaelShift - Returns k such that lambda(2^Bitwidth) = 2^k, where lambda
380 /// is the Carmichael function. This means that x^(2^k) === 1 mod 2^Bitwidth for
381 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
382 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
383 /// even x in Bitwidth-bit arithmetic.
384 static unsigned CarmichaelShift(unsigned Bitwidth) {
390 /// IncorporateWeight - Add the extra weight 'RHS' to the existing weight 'LHS',
391 /// reducing the combined weight using any special properties of the operation.
392 /// The existing weight LHS represents the computation X op X op ... op X where
393 /// X occurs LHS times. The combined weight represents X op X op ... op X with
394 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
395 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
396 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
397 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
398 // If we were working with infinite precision arithmetic then the combined
399 // weight would be LHS + RHS. But we are using finite precision arithmetic,
400 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
401 // for nilpotent operations and addition, but not for idempotent operations
402 // and multiplication), so it is important to correctly reduce the combined
403 // weight back into range if wrapping would be wrong.
405 // If RHS is zero then the weight didn't change.
406 if (RHS.isMinValue())
408 // If LHS is zero then the combined weight is RHS.
409 if (LHS.isMinValue()) {
413 // From this point on we know that neither LHS nor RHS is zero.
415 if (Instruction::isIdempotent(Opcode)) {
416 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
417 // weight of 1. Keeping weights at zero or one also means that wrapping is
419 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
420 return; // Return a weight of 1.
422 if (Instruction::isNilpotent(Opcode)) {
423 // Nilpotent means X op X === 0, so reduce weights modulo 2.
424 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
425 LHS = 0; // 1 + 1 === 0 modulo 2.
428 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
429 // TODO: Reduce the weight by exploiting nsw/nuw?
434 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
435 "Unknown associative operation!");
436 unsigned Bitwidth = LHS.getBitWidth();
437 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
438 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
439 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
440 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
441 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
442 // which by a happy accident means that they can always be represented using
444 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
445 // the Carmichael number).
447 /// CM - The value of Carmichael's lambda function.
448 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
449 // Any weight W >= Threshold can be replaced with W - CM.
450 APInt Threshold = CM + Bitwidth;
451 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
452 // For Bitwidth 4 or more the following sum does not overflow.
454 while (LHS.uge(Threshold))
457 // To avoid problems with overflow do everything the same as above but using
459 unsigned CM = 1U << CarmichaelShift(Bitwidth);
460 unsigned Threshold = CM + Bitwidth;
461 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
462 "Weights not reduced!");
463 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
464 while (Total >= Threshold)
470 typedef std::pair<Value*, APInt> RepeatedValue;
472 /// LinearizeExprTree - Given an associative binary expression, return the leaf
473 /// nodes in Ops along with their weights (how many times the leaf occurs). The
474 /// original expression is the same as
475 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
477 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
481 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
483 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
485 /// This routine may modify the function, in which case it returns 'true'. The
486 /// changes it makes may well be destructive, changing the value computed by 'I'
487 /// to something completely different. Thus if the routine returns 'true' then
488 /// you MUST either replace I with a new expression computed from the Ops array,
489 /// or use RewriteExprTree to put the values back in.
491 /// A leaf node is either not a binary operation of the same kind as the root
492 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
493 /// opcode), or is the same kind of binary operator but has a use which either
494 /// does not belong to the expression, or does belong to the expression but is
495 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
496 /// of the expression, while for non-leaf nodes (except for the root 'I') every
497 /// use is a non-leaf node of the expression.
500 /// expression graph node names
510 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
511 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
513 /// The expression is maximal: if some instruction is a binary operator of the
514 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
515 /// then the instruction also belongs to the expression, is not a leaf node of
516 /// it, and its operands also belong to the expression (but may be leaf nodes).
518 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
519 /// order to ensure that every non-root node in the expression has *exactly one*
520 /// use by a non-leaf node of the expression. This destruction means that the
521 /// caller MUST either replace 'I' with a new expression or use something like
522 /// RewriteExprTree to put the values back in if the routine indicates that it
523 /// made a change by returning 'true'.
525 /// In the above example either the right operand of A or the left operand of B
526 /// will be replaced by undef. If it is B's operand then this gives:
530 /// + + | A, B - operand of B replaced with undef
536 /// Note that such undef operands can only be reached by passing through 'I'.
537 /// For example, if you visit operands recursively starting from a leaf node
538 /// then you will never see such an undef operand unless you get back to 'I',
539 /// which requires passing through a phi node.
541 /// Note that this routine may also mutate binary operators of the wrong type
542 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
543 /// of the expression) if it can turn them into binary operators of the right
544 /// type and thus make the expression bigger.
546 static bool LinearizeExprTree(BinaryOperator *I,
547 SmallVectorImpl<RepeatedValue> &Ops) {
548 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
549 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
550 unsigned Opcode = I->getOpcode();
551 assert(I->isAssociative() && I->isCommutative() &&
552 "Expected an associative and commutative operation!");
554 // Visit all operands of the expression, keeping track of their weight (the
555 // number of paths from the expression root to the operand, or if you like
556 // the number of times that operand occurs in the linearized expression).
557 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
558 // while A has weight two.
560 // Worklist of non-leaf nodes (their operands are in the expression too) along
561 // with their weights, representing a certain number of paths to the operator.
562 // If an operator occurs in the worklist multiple times then we found multiple
563 // ways to get to it.
564 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
565 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
566 bool MadeChange = false;
568 // Leaves of the expression are values that either aren't the right kind of
569 // operation (eg: a constant, or a multiply in an add tree), or are, but have
570 // some uses that are not inside the expression. For example, in I = X + X,
571 // X = A + B, the value X has two uses (by I) that are in the expression. If
572 // X has any other uses, for example in a return instruction, then we consider
573 // X to be a leaf, and won't analyze it further. When we first visit a value,
574 // if it has more than one use then at first we conservatively consider it to
575 // be a leaf. Later, as the expression is explored, we may discover some more
576 // uses of the value from inside the expression. If all uses turn out to be
577 // from within the expression (and the value is a binary operator of the right
578 // kind) then the value is no longer considered to be a leaf, and its operands
581 // Leaves - Keeps track of the set of putative leaves as well as the number of
582 // paths to each leaf seen so far.
583 typedef DenseMap<Value*, APInt> LeafMap;
584 LeafMap Leaves; // Leaf -> Total weight so far.
585 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
588 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
590 while (!Worklist.empty()) {
591 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
592 I = P.first; // We examine the operands of this binary operator.
594 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
595 Value *Op = I->getOperand(OpIdx);
596 APInt Weight = P.second; // Number of paths to this operand.
597 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
598 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
600 // If this is a binary operation of the right kind with only one use then
601 // add its operands to the expression.
602 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
603 assert(Visited.insert(Op) && "Not first visit!");
604 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
605 Worklist.push_back(std::make_pair(BO, Weight));
609 // Appears to be a leaf. Is the operand already in the set of leaves?
610 LeafMap::iterator It = Leaves.find(Op);
611 if (It == Leaves.end()) {
612 // Not in the leaf map. Must be the first time we saw this operand.
613 assert(Visited.insert(Op) && "Not first visit!");
614 if (!Op->hasOneUse()) {
615 // This value has uses not accounted for by the expression, so it is
616 // not safe to modify. Mark it as being a leaf.
617 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
618 LeafOrder.push_back(Op);
622 // No uses outside the expression, try morphing it.
623 } else if (It != Leaves.end()) {
624 // Already in the leaf map.
625 assert(Visited.count(Op) && "In leaf map but not visited!");
627 // Update the number of paths to the leaf.
628 IncorporateWeight(It->second, Weight, Opcode);
630 #if 0 // TODO: Re-enable once PR13021 is fixed.
631 // The leaf already has one use from inside the expression. As we want
632 // exactly one such use, drop this new use of the leaf.
633 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
634 I->setOperand(OpIdx, UndefValue::get(I->getType()));
637 // If the leaf is a binary operation of the right kind and we now see
638 // that its multiple original uses were in fact all by nodes belonging
639 // to the expression, then no longer consider it to be a leaf and add
640 // its operands to the expression.
641 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
642 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
643 Worklist.push_back(std::make_pair(BO, It->second));
649 // If we still have uses that are not accounted for by the expression
650 // then it is not safe to modify the value.
651 if (!Op->hasOneUse())
654 // No uses outside the expression, try morphing it.
656 Leaves.erase(It); // Since the value may be morphed below.
659 // At this point we have a value which, first of all, is not a binary
660 // expression of the right kind, and secondly, is only used inside the
661 // expression. This means that it can safely be modified. See if we
662 // can usefully morph it into an expression of the right kind.
663 assert((!isa<Instruction>(Op) ||
664 cast<Instruction>(Op)->getOpcode() != Opcode) &&
665 "Should have been handled above!");
666 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
668 // If this is a multiply expression, turn any internal negations into
669 // multiplies by -1 so they can be reassociated.
670 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
671 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
672 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
673 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
674 BO = LowerNegateToMultiply(BO);
675 DEBUG(dbgs() << *BO << '\n');
676 Worklist.push_back(std::make_pair(BO, Weight));
681 // Failed to morph into an expression of the right type. This really is
683 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
684 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
685 LeafOrder.push_back(Op);
690 // The leaves, repeated according to their weights, represent the linearized
691 // form of the expression.
692 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
693 Value *V = LeafOrder[i];
694 LeafMap::iterator It = Leaves.find(V);
695 if (It == Leaves.end())
696 // Node initially thought to be a leaf wasn't.
698 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
699 APInt Weight = It->second;
700 if (Weight.isMinValue())
701 // Leaf already output or weight reduction eliminated it.
703 // Ensure the leaf is only output once.
705 Ops.push_back(std::make_pair(V, Weight));
708 // For nilpotent operations or addition there may be no operands, for example
709 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
710 // in both cases the weight reduces to 0 causing the value to be skipped.
712 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
713 assert(Identity && "Associative operation without identity!");
714 Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1)));
720 // RewriteExprTree - Now that the operands for this expression tree are
721 // linearized and optimized, emit them in-order.
722 void Reassociate::RewriteExprTree(BinaryOperator *I,
723 SmallVectorImpl<ValueEntry> &Ops) {
724 assert(Ops.size() > 1 && "Single values should be used directly!");
726 // Since our optimizations should never increase the number of operations, the
727 // new expression can usually be written reusing the existing binary operators
728 // from the original expression tree, without creating any new instructions,
729 // though the rewritten expression may have a completely different topology.
730 // We take care to not change anything if the new expression will be the same
731 // as the original. If more than trivial changes (like commuting operands)
732 // were made then we are obliged to clear out any optional subclass data like
735 /// NodesToRewrite - Nodes from the original expression available for writing
736 /// the new expression into.
737 SmallVector<BinaryOperator*, 8> NodesToRewrite;
738 unsigned Opcode = I->getOpcode();
739 BinaryOperator *Op = I;
741 /// NotRewritable - The operands being written will be the leaves of the new
742 /// expression and must not be used as inner nodes (via NodesToRewrite) by
743 /// mistake. Inner nodes are always reassociable, and usually leaves are not
744 /// (if they were they would have been incorporated into the expression and so
745 /// would not be leaves), so most of the time there is no danger of this. But
746 /// in rare cases a leaf may become reassociable if an optimization kills uses
747 /// of it, or it may momentarily become reassociable during rewriting (below)
748 /// due it being removed as an operand of one of its uses. Ensure that misuse
749 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
750 /// leaves and refusing to reuse any of them as inner nodes.
751 SmallPtrSet<Value*, 8> NotRewritable;
752 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
753 NotRewritable.insert(Ops[i].Op);
755 // ExpressionChanged - Non-null if the rewritten expression differs from the
756 // original in some non-trivial way, requiring the clearing of optional flags.
757 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
758 BinaryOperator *ExpressionChanged = nullptr;
759 for (unsigned i = 0; ; ++i) {
760 // The last operation (which comes earliest in the IR) is special as both
761 // operands will come from Ops, rather than just one with the other being
763 if (i+2 == Ops.size()) {
764 Value *NewLHS = Ops[i].Op;
765 Value *NewRHS = Ops[i+1].Op;
766 Value *OldLHS = Op->getOperand(0);
767 Value *OldRHS = Op->getOperand(1);
769 if (NewLHS == OldLHS && NewRHS == OldRHS)
770 // Nothing changed, leave it alone.
773 if (NewLHS == OldRHS && NewRHS == OldLHS) {
774 // The order of the operands was reversed. Swap them.
775 DEBUG(dbgs() << "RA: " << *Op << '\n');
777 DEBUG(dbgs() << "TO: " << *Op << '\n');
783 // The new operation differs non-trivially from the original. Overwrite
784 // the old operands with the new ones.
785 DEBUG(dbgs() << "RA: " << *Op << '\n');
786 if (NewLHS != OldLHS) {
787 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
788 if (BO && !NotRewritable.count(BO))
789 NodesToRewrite.push_back(BO);
790 Op->setOperand(0, NewLHS);
792 if (NewRHS != OldRHS) {
793 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
794 if (BO && !NotRewritable.count(BO))
795 NodesToRewrite.push_back(BO);
796 Op->setOperand(1, NewRHS);
798 DEBUG(dbgs() << "TO: " << *Op << '\n');
800 ExpressionChanged = Op;
807 // Not the last operation. The left-hand side will be a sub-expression
808 // while the right-hand side will be the current element of Ops.
809 Value *NewRHS = Ops[i].Op;
810 if (NewRHS != Op->getOperand(1)) {
811 DEBUG(dbgs() << "RA: " << *Op << '\n');
812 if (NewRHS == Op->getOperand(0)) {
813 // The new right-hand side was already present as the left operand. If
814 // we are lucky then swapping the operands will sort out both of them.
817 // Overwrite with the new right-hand side.
818 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
819 if (BO && !NotRewritable.count(BO))
820 NodesToRewrite.push_back(BO);
821 Op->setOperand(1, NewRHS);
822 ExpressionChanged = Op;
824 DEBUG(dbgs() << "TO: " << *Op << '\n');
829 // Now deal with the left-hand side. If this is already an operation node
830 // from the original expression then just rewrite the rest of the expression
832 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
833 if (BO && !NotRewritable.count(BO)) {
838 // Otherwise, grab a spare node from the original expression and use that as
839 // the left-hand side. If there are no nodes left then the optimizers made
840 // an expression with more nodes than the original! This usually means that
841 // they did something stupid but it might mean that the problem was just too
842 // hard (finding the mimimal number of multiplications needed to realize a
843 // multiplication expression is NP-complete). Whatever the reason, smart or
844 // stupid, create a new node if there are none left.
845 BinaryOperator *NewOp;
846 if (NodesToRewrite.empty()) {
847 Constant *Undef = UndefValue::get(I->getType());
848 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
849 Undef, Undef, "", I);
850 if (NewOp->getType()->isFloatingPointTy())
851 NewOp->setFastMathFlags(I->getFastMathFlags());
853 NewOp = NodesToRewrite.pop_back_val();
856 DEBUG(dbgs() << "RA: " << *Op << '\n');
857 Op->setOperand(0, NewOp);
858 DEBUG(dbgs() << "TO: " << *Op << '\n');
859 ExpressionChanged = Op;
865 // If the expression changed non-trivially then clear out all subclass data
866 // starting from the operator specified in ExpressionChanged, and compactify
867 // the operators to just before the expression root to guarantee that the
868 // expression tree is dominated by all of Ops.
869 if (ExpressionChanged)
871 // Preserve FastMathFlags.
872 if (isa<FPMathOperator>(I)) {
873 FastMathFlags Flags = I->getFastMathFlags();
874 ExpressionChanged->clearSubclassOptionalData();
875 ExpressionChanged->setFastMathFlags(Flags);
877 ExpressionChanged->clearSubclassOptionalData();
879 if (ExpressionChanged == I)
881 ExpressionChanged->moveBefore(I);
882 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
885 // Throw away any left over nodes from the original expression.
886 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
887 RedoInsts.insert(NodesToRewrite[i]);
890 /// NegateValue - Insert instructions before the instruction pointed to by BI,
891 /// that computes the negative version of the value specified. The negative
892 /// version of the value is returned, and BI is left pointing at the instruction
893 /// that should be processed next by the reassociation pass.
894 static Value *NegateValue(Value *V, Instruction *BI) {
895 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
896 return ConstantExpr::getFNeg(C);
897 if (Constant *C = dyn_cast<Constant>(V))
898 return ConstantExpr::getNeg(C);
900 // We are trying to expose opportunity for reassociation. One of the things
901 // that we want to do to achieve this is to push a negation as deep into an
902 // expression chain as possible, to expose the add instructions. In practice,
903 // this means that we turn this:
904 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
905 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
906 // the constants. We assume that instcombine will clean up the mess later if
907 // we introduce tons of unnecessary negation instructions.
909 if (BinaryOperator *I =
910 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
911 // Push the negates through the add.
912 I->setOperand(0, NegateValue(I->getOperand(0), BI));
913 I->setOperand(1, NegateValue(I->getOperand(1), BI));
915 // We must move the add instruction here, because the neg instructions do
916 // not dominate the old add instruction in general. By moving it, we are
917 // assured that the neg instructions we just inserted dominate the
918 // instruction we are about to insert after them.
921 I->setName(I->getName()+".neg");
925 // Okay, we need to materialize a negated version of V with an instruction.
926 // Scan the use lists of V to see if we have one already.
927 for (User *U : V->users()) {
928 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
931 // We found one! Now we have to make sure that the definition dominates
932 // this use. We do this by moving it to the entry block (if it is a
933 // non-instruction value) or right after the definition. These negates will
934 // be zapped by reassociate later, so we don't need much finesse here.
935 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
937 // Verify that the negate is in this function, V might be a constant expr.
938 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
941 BasicBlock::iterator InsertPt;
942 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
943 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
944 InsertPt = II->getNormalDest()->begin();
946 InsertPt = InstInput;
949 while (isa<PHINode>(InsertPt)) ++InsertPt;
951 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
953 TheNeg->moveBefore(InsertPt);
957 // Insert a 'neg' instruction that subtracts the value from zero to get the
959 return CreateNeg(V, V->getName() + ".neg", BI, BI);
962 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
963 /// X-Y into (X + -Y).
964 static bool ShouldBreakUpSubtract(Instruction *Sub) {
965 // If this is a negation, we can't split it up!
966 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
969 // Don't breakup X - undef.
970 if (isa<UndefValue>(Sub->getOperand(1)))
973 // Don't bother to break this up unless either the LHS is an associable add or
974 // subtract or if this is only used by one.
975 Value *V0 = Sub->getOperand(0);
976 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
977 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
979 Value *V1 = Sub->getOperand(1);
980 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
981 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
983 Value *VB = Sub->user_back();
984 if (Sub->hasOneUse() &&
985 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
986 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
992 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
993 /// only used by an add, transform this into (X+(0-Y)) to promote better
995 static BinaryOperator *BreakUpSubtract(Instruction *Sub) {
996 // Convert a subtract into an add and a neg instruction. This allows sub
997 // instructions to be commuted with other add instructions.
999 // Calculate the negative value of Operand 1 of the sub instruction,
1000 // and set it as the RHS of the add instruction we just made.
1002 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
1003 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
1004 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
1005 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
1008 // Everyone now refers to the add instruction.
1009 Sub->replaceAllUsesWith(New);
1010 New->setDebugLoc(Sub->getDebugLoc());
1012 DEBUG(dbgs() << "Negated: " << *New << '\n');
1016 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
1017 /// by one, change this into a multiply by a constant to assist with further
1019 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1020 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
1021 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
1023 BinaryOperator *Mul =
1024 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
1025 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
1027 Shl->replaceAllUsesWith(Mul);
1028 Mul->setDebugLoc(Shl->getDebugLoc());
1032 /// FindInOperandList - Scan backwards and forwards among values with the same
1033 /// rank as element i to see if X exists. If X does not exist, return i. This
1034 /// is useful when scanning for 'x' when we see '-x' because they both get the
1036 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
1038 unsigned XRank = Ops[i].Rank;
1039 unsigned e = Ops.size();
1040 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
1043 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1044 if (Instruction *I2 = dyn_cast<Instruction>(X))
1045 if (I1->isIdenticalTo(I2))
1049 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
1052 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1053 if (Instruction *I2 = dyn_cast<Instruction>(X))
1054 if (I1->isIdenticalTo(I2))
1060 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
1061 /// and returning the result. Insert the tree before I.
1062 static Value *EmitAddTreeOfValues(Instruction *I,
1063 SmallVectorImpl<WeakVH> &Ops){
1064 if (Ops.size() == 1) return Ops.back();
1066 Value *V1 = Ops.back();
1068 Value *V2 = EmitAddTreeOfValues(I, Ops);
1069 return CreateAdd(V2, V1, "tmp", I, I);
1072 /// RemoveFactorFromExpression - If V is an expression tree that is a
1073 /// multiplication sequence, and if this sequence contains a multiply by Factor,
1074 /// remove Factor from the tree and return the new tree.
1075 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
1076 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1080 SmallVector<RepeatedValue, 8> Tree;
1081 MadeChange |= LinearizeExprTree(BO, Tree);
1082 SmallVector<ValueEntry, 8> Factors;
1083 Factors.reserve(Tree.size());
1084 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1085 RepeatedValue E = Tree[i];
1086 Factors.append(E.second.getZExtValue(),
1087 ValueEntry(getRank(E.first), E.first));
1090 bool FoundFactor = false;
1091 bool NeedsNegate = false;
1092 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1093 if (Factors[i].Op == Factor) {
1095 Factors.erase(Factors.begin()+i);
1099 // If this is a negative version of this factor, remove it.
1100 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1101 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1102 if (FC1->getValue() == -FC2->getValue()) {
1103 FoundFactor = NeedsNegate = true;
1104 Factors.erase(Factors.begin()+i);
1107 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1108 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1109 APFloat F1(FC1->getValueAPF());
1110 APFloat F2(FC2->getValueAPF());
1112 if (F1.compare(F2) == APFloat::cmpEqual) {
1113 FoundFactor = NeedsNegate = true;
1114 Factors.erase(Factors.begin() + i);
1122 // Make sure to restore the operands to the expression tree.
1123 RewriteExprTree(BO, Factors);
1127 BasicBlock::iterator InsertPt = BO; ++InsertPt;
1129 // If this was just a single multiply, remove the multiply and return the only
1130 // remaining operand.
1131 if (Factors.size() == 1) {
1132 RedoInsts.insert(BO);
1135 RewriteExprTree(BO, Factors);
1140 V = CreateNeg(V, "neg", InsertPt, BO);
1145 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
1146 /// add its operands as factors, otherwise add V to the list of factors.
1148 /// Ops is the top-level list of add operands we're trying to factor.
1149 static void FindSingleUseMultiplyFactors(Value *V,
1150 SmallVectorImpl<Value*> &Factors,
1151 const SmallVectorImpl<ValueEntry> &Ops) {
1152 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1154 Factors.push_back(V);
1158 // Otherwise, add the LHS and RHS to the list of factors.
1159 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
1160 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
1163 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
1164 /// instruction. This optimizes based on identities. If it can be reduced to
1165 /// a single Value, it is returned, otherwise the Ops list is mutated as
1167 static Value *OptimizeAndOrXor(unsigned Opcode,
1168 SmallVectorImpl<ValueEntry> &Ops) {
1169 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1170 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1171 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1172 // First, check for X and ~X in the operand list.
1173 assert(i < Ops.size());
1174 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1175 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1176 unsigned FoundX = FindInOperandList(Ops, i, X);
1178 if (Opcode == Instruction::And) // ...&X&~X = 0
1179 return Constant::getNullValue(X->getType());
1181 if (Opcode == Instruction::Or) // ...|X|~X = -1
1182 return Constant::getAllOnesValue(X->getType());
1186 // Next, check for duplicate pairs of values, which we assume are next to
1187 // each other, due to our sorting criteria.
1188 assert(i < Ops.size());
1189 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1190 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1191 // Drop duplicate values for And and Or.
1192 Ops.erase(Ops.begin()+i);
1198 // Drop pairs of values for Xor.
1199 assert(Opcode == Instruction::Xor);
1201 return Constant::getNullValue(Ops[0].Op->getType());
1204 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1212 /// Helper funciton of CombineXorOpnd(). It creates a bitwise-and
1213 /// instruction with the given two operands, and return the resulting
1214 /// instruction. There are two special cases: 1) if the constant operand is 0,
1215 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1217 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1218 const APInt &ConstOpnd) {
1219 if (ConstOpnd != 0) {
1220 if (!ConstOpnd.isAllOnesValue()) {
1221 LLVMContext &Ctx = Opnd->getType()->getContext();
1223 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1224 "and.ra", InsertBefore);
1225 I->setDebugLoc(InsertBefore->getDebugLoc());
1233 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1234 // into "R ^ C", where C would be 0, and R is a symbolic value.
1236 // If it was successful, true is returned, and the "R" and "C" is returned
1237 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1238 // and both "Res" and "ConstOpnd" remain unchanged.
1240 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1241 APInt &ConstOpnd, Value *&Res) {
1242 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1243 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1244 // = (x & ~c1) ^ (c1 ^ c2)
1245 // It is useful only when c1 == c2.
1246 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) {
1247 if (!Opnd1->getValue()->hasOneUse())
1250 const APInt &C1 = Opnd1->getConstPart();
1251 if (C1 != ConstOpnd)
1254 Value *X = Opnd1->getSymbolicPart();
1255 Res = createAndInstr(I, X, ~C1);
1256 // ConstOpnd was C2, now C1 ^ C2.
1259 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1260 RedoInsts.insert(T);
1267 // Helper function of OptimizeXor(). It tries to simplify
1268 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1271 // If it was successful, true is returned, and the "R" and "C" is returned
1272 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1273 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1274 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1275 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
1276 APInt &ConstOpnd, Value *&Res) {
1277 Value *X = Opnd1->getSymbolicPart();
1278 if (X != Opnd2->getSymbolicPart())
1281 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1282 int DeadInstNum = 1;
1283 if (Opnd1->getValue()->hasOneUse())
1285 if (Opnd2->getValue()->hasOneUse())
1289 // (x | c1) ^ (x & c2)
1290 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1291 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1292 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1294 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1295 if (Opnd2->isOrExpr())
1296 std::swap(Opnd1, Opnd2);
1298 const APInt &C1 = Opnd1->getConstPart();
1299 const APInt &C2 = Opnd2->getConstPart();
1300 APInt C3((~C1) ^ C2);
1302 // Do not increase code size!
1303 if (C3 != 0 && !C3.isAllOnesValue()) {
1304 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1305 if (NewInstNum > DeadInstNum)
1309 Res = createAndInstr(I, X, C3);
1312 } else if (Opnd1->isOrExpr()) {
1313 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1315 const APInt &C1 = Opnd1->getConstPart();
1316 const APInt &C2 = Opnd2->getConstPart();
1319 // Do not increase code size
1320 if (C3 != 0 && !C3.isAllOnesValue()) {
1321 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1322 if (NewInstNum > DeadInstNum)
1326 Res = createAndInstr(I, X, C3);
1329 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1331 const APInt &C1 = Opnd1->getConstPart();
1332 const APInt &C2 = Opnd2->getConstPart();
1334 Res = createAndInstr(I, X, C3);
1337 // Put the original operands in the Redo list; hope they will be deleted
1339 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1340 RedoInsts.insert(T);
1341 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1342 RedoInsts.insert(T);
1347 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1348 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1350 Value *Reassociate::OptimizeXor(Instruction *I,
1351 SmallVectorImpl<ValueEntry> &Ops) {
1352 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1355 if (Ops.size() == 1)
1358 SmallVector<XorOpnd, 8> Opnds;
1359 SmallVector<XorOpnd*, 8> OpndPtrs;
1360 Type *Ty = Ops[0].Op->getType();
1361 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1363 // Step 1: Convert ValueEntry to XorOpnd
1364 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1365 Value *V = Ops[i].Op;
1366 if (!isa<ConstantInt>(V)) {
1368 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1371 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1374 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1375 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1376 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1377 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1378 // when new elements are added to the vector.
1379 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1380 OpndPtrs.push_back(&Opnds[i]);
1382 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1383 // the same symbolic value cluster together. For instance, the input operand
1384 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1385 // ("x | 123", "x & 789", "y & 456").
1386 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor());
1388 // Step 3: Combine adjacent operands
1389 XorOpnd *PrevOpnd = nullptr;
1390 bool Changed = false;
1391 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1392 XorOpnd *CurrOpnd = OpndPtrs[i];
1393 // The combined value
1396 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1397 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1400 *CurrOpnd = XorOpnd(CV);
1402 CurrOpnd->Invalidate();
1407 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1408 PrevOpnd = CurrOpnd;
1412 // step 3.2: When previous and current operands share the same symbolic
1413 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1415 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1416 // Remove previous operand
1417 PrevOpnd->Invalidate();
1419 *CurrOpnd = XorOpnd(CV);
1420 PrevOpnd = CurrOpnd;
1422 CurrOpnd->Invalidate();
1429 // Step 4: Reassemble the Ops
1432 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1433 XorOpnd &O = Opnds[i];
1436 ValueEntry VE(getRank(O.getValue()), O.getValue());
1439 if (ConstOpnd != 0) {
1440 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1441 ValueEntry VE(getRank(C), C);
1444 int Sz = Ops.size();
1446 return Ops.back().Op;
1448 assert(ConstOpnd == 0);
1449 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1456 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
1457 /// optimizes based on identities. If it can be reduced to a single Value, it
1458 /// is returned, otherwise the Ops list is mutated as necessary.
1459 Value *Reassociate::OptimizeAdd(Instruction *I,
1460 SmallVectorImpl<ValueEntry> &Ops) {
1461 // Scan the operand lists looking for X and -X pairs. If we find any, we
1462 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1464 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1466 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1467 Value *TheOp = Ops[i].Op;
1468 // Check to see if we've seen this operand before. If so, we factor all
1469 // instances of the operand together. Due to our sorting criteria, we know
1470 // that these need to be next to each other in the vector.
1471 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1472 // Rescan the list, remove all instances of this operand from the expr.
1473 unsigned NumFound = 0;
1475 Ops.erase(Ops.begin()+i);
1477 } while (i != Ops.size() && Ops[i].Op == TheOp);
1479 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1482 // Insert a new multiply.
1483 Type *Ty = TheOp->getType();
1484 Constant *C = Ty->isIntegerTy() ? ConstantInt::get(Ty, NumFound)
1485 : ConstantFP::get(Ty, NumFound);
1486 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1488 // Now that we have inserted a multiply, optimize it. This allows us to
1489 // handle cases that require multiple factoring steps, such as this:
1490 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1491 RedoInsts.insert(Mul);
1493 // If every add operand was a duplicate, return the multiply.
1497 // Otherwise, we had some input that didn't have the dupe, such as
1498 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1499 // things being added by this operation.
1500 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1507 // Check for X and -X or X and ~X in the operand list.
1508 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1509 !BinaryOperator::isNot(TheOp))
1513 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1514 X = BinaryOperator::getNegArgument(TheOp);
1515 else if (BinaryOperator::isNot(TheOp))
1516 X = BinaryOperator::getNotArgument(TheOp);
1518 unsigned FoundX = FindInOperandList(Ops, i, X);
1522 // Remove X and -X from the operand list.
1523 if (Ops.size() == 2 &&
1524 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1525 return Constant::getNullValue(X->getType());
1527 // Remove X and ~X from the operand list.
1528 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1529 return Constant::getAllOnesValue(X->getType());
1531 Ops.erase(Ops.begin()+i);
1535 --i; // Need to back up an extra one.
1536 Ops.erase(Ops.begin()+FoundX);
1538 --i; // Revisit element.
1539 e -= 2; // Removed two elements.
1541 // if X and ~X we append -1 to the operand list.
1542 if (BinaryOperator::isNot(TheOp)) {
1543 Value *V = Constant::getAllOnesValue(X->getType());
1544 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1549 // Scan the operand list, checking to see if there are any common factors
1550 // between operands. Consider something like A*A+A*B*C+D. We would like to
1551 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1552 // To efficiently find this, we count the number of times a factor occurs
1553 // for any ADD operands that are MULs.
1554 DenseMap<Value*, unsigned> FactorOccurrences;
1556 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1557 // where they are actually the same multiply.
1558 unsigned MaxOcc = 0;
1559 Value *MaxOccVal = nullptr;
1560 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1561 BinaryOperator *BOp =
1562 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1566 // Compute all of the factors of this added value.
1567 SmallVector<Value*, 8> Factors;
1568 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1569 assert(Factors.size() > 1 && "Bad linearize!");
1571 // Add one to FactorOccurrences for each unique factor in this op.
1572 SmallPtrSet<Value*, 8> Duplicates;
1573 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1574 Value *Factor = Factors[i];
1575 if (!Duplicates.insert(Factor))
1578 unsigned Occ = ++FactorOccurrences[Factor];
1584 // If Factor is a negative constant, add the negated value as a factor
1585 // because we can percolate the negate out. Watch for minint, which
1586 // cannot be positivified.
1587 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1588 if (CI->isNegative() && !CI->isMinValue(true)) {
1589 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1590 assert(!Duplicates.count(Factor) &&
1591 "Shouldn't have two constant factors, missed a canonicalize");
1592 unsigned Occ = ++FactorOccurrences[Factor];
1598 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1599 if (CF->isNegative()) {
1600 APFloat F(CF->getValueAPF());
1602 Factor = ConstantFP::get(CF->getContext(), F);
1603 assert(!Duplicates.count(Factor) &&
1604 "Shouldn't have two constant factors, missed a canonicalize");
1605 unsigned Occ = ++FactorOccurrences[Factor];
1615 // If any factor occurred more than one time, we can pull it out.
1617 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1620 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1621 // this, we could otherwise run into situations where removing a factor
1622 // from an expression will drop a use of maxocc, and this can cause
1623 // RemoveFactorFromExpression on successive values to behave differently.
1624 Instruction *DummyInst =
1625 I->getType()->isIntegerTy()
1626 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1627 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1629 SmallVector<WeakVH, 4> NewMulOps;
1630 for (unsigned i = 0; i != Ops.size(); ++i) {
1631 // Only try to remove factors from expressions we're allowed to.
1632 BinaryOperator *BOp =
1633 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1637 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1638 // The factorized operand may occur several times. Convert them all in
1640 for (unsigned j = Ops.size(); j != i;) {
1642 if (Ops[j].Op == Ops[i].Op) {
1643 NewMulOps.push_back(V);
1644 Ops.erase(Ops.begin()+j);
1651 // No need for extra uses anymore.
1654 unsigned NumAddedValues = NewMulOps.size();
1655 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1657 // Now that we have inserted the add tree, optimize it. This allows us to
1658 // handle cases that require multiple factoring steps, such as this:
1659 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1660 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1661 (void)NumAddedValues;
1662 if (Instruction *VI = dyn_cast<Instruction>(V))
1663 RedoInsts.insert(VI);
1665 // Create the multiply.
1666 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1668 // Rerun associate on the multiply in case the inner expression turned into
1669 // a multiply. We want to make sure that we keep things in canonical form.
1670 RedoInsts.insert(V2);
1672 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1673 // entire result expression is just the multiply "A*(B+C)".
1677 // Otherwise, we had some input that didn't have the factor, such as
1678 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1679 // things being added by this operation.
1680 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1686 /// \brief Build up a vector of value/power pairs factoring a product.
1688 /// Given a series of multiplication operands, build a vector of factors and
1689 /// the powers each is raised to when forming the final product. Sort them in
1690 /// the order of descending power.
1692 /// (x*x) -> [(x, 2)]
1693 /// ((x*x)*x) -> [(x, 3)]
1694 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1696 /// \returns Whether any factors have a power greater than one.
1697 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1698 SmallVectorImpl<Factor> &Factors) {
1699 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1700 // Compute the sum of powers of simplifiable factors.
1701 unsigned FactorPowerSum = 0;
1702 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1703 Value *Op = Ops[Idx-1].Op;
1705 // Count the number of occurrences of this value.
1707 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1709 // Track for simplification all factors which occur 2 or more times.
1711 FactorPowerSum += Count;
1714 // We can only simplify factors if the sum of the powers of our simplifiable
1715 // factors is 4 or higher. When that is the case, we will *always* have
1716 // a simplification. This is an important invariant to prevent cyclicly
1717 // trying to simplify already minimal formations.
1718 if (FactorPowerSum < 4)
1721 // Now gather the simplifiable factors, removing them from Ops.
1723 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1724 Value *Op = Ops[Idx-1].Op;
1726 // Count the number of occurrences of this value.
1728 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1732 // Move an even number of occurrences to Factors.
1735 FactorPowerSum += Count;
1736 Factors.push_back(Factor(Op, Count));
1737 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1740 // None of the adjustments above should have reduced the sum of factor powers
1741 // below our mininum of '4'.
1742 assert(FactorPowerSum >= 4);
1744 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1748 /// \brief Build a tree of multiplies, computing the product of Ops.
1749 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1750 SmallVectorImpl<Value*> &Ops) {
1751 if (Ops.size() == 1)
1754 Value *LHS = Ops.pop_back_val();
1756 if (LHS->getType()->isIntegerTy())
1757 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1759 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1760 } while (!Ops.empty());
1765 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1767 /// Given a vector of values raised to various powers, where no two values are
1768 /// equal and the powers are sorted in decreasing order, compute the minimal
1769 /// DAG of multiplies to compute the final product, and return that product
1771 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1772 SmallVectorImpl<Factor> &Factors) {
1773 assert(Factors[0].Power);
1774 SmallVector<Value *, 4> OuterProduct;
1775 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1776 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1777 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1782 // We want to multiply across all the factors with the same power so that
1783 // we can raise them to that power as a single entity. Build a mini tree
1785 SmallVector<Value *, 4> InnerProduct;
1786 InnerProduct.push_back(Factors[LastIdx].Base);
1788 InnerProduct.push_back(Factors[Idx].Base);
1790 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1792 // Reset the base value of the first factor to the new expression tree.
1793 // We'll remove all the factors with the same power in a second pass.
1794 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1795 if (Instruction *MI = dyn_cast<Instruction>(M))
1796 RedoInsts.insert(MI);
1800 // Unique factors with equal powers -- we've folded them into the first one's
1802 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1803 Factor::PowerEqual()),
1806 // Iteratively collect the base of each factor with an add power into the
1807 // outer product, and halve each power in preparation for squaring the
1809 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1810 if (Factors[Idx].Power & 1)
1811 OuterProduct.push_back(Factors[Idx].Base);
1812 Factors[Idx].Power >>= 1;
1814 if (Factors[0].Power) {
1815 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1816 OuterProduct.push_back(SquareRoot);
1817 OuterProduct.push_back(SquareRoot);
1819 if (OuterProduct.size() == 1)
1820 return OuterProduct.front();
1822 Value *V = buildMultiplyTree(Builder, OuterProduct);
1826 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1827 SmallVectorImpl<ValueEntry> &Ops) {
1828 // We can only optimize the multiplies when there is a chain of more than
1829 // three, such that a balanced tree might require fewer total multiplies.
1833 // Try to turn linear trees of multiplies without other uses of the
1834 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1836 SmallVector<Factor, 4> Factors;
1837 if (!collectMultiplyFactors(Ops, Factors))
1838 return nullptr; // All distinct factors, so nothing left for us to do.
1840 IRBuilder<> Builder(I);
1841 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1845 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1846 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1850 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1851 SmallVectorImpl<ValueEntry> &Ops) {
1852 // Now that we have the linearized expression tree, try to optimize it.
1853 // Start by folding any constants that we found.
1854 Constant *Cst = nullptr;
1855 unsigned Opcode = I->getOpcode();
1856 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1857 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1858 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1860 // If there was nothing but constants then we are done.
1864 // Put the combined constant back at the end of the operand list, except if
1865 // there is no point. For example, an add of 0 gets dropped here, while a
1866 // multiplication by zero turns the whole expression into zero.
1867 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1868 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1870 Ops.push_back(ValueEntry(0, Cst));
1873 if (Ops.size() == 1) return Ops[0].Op;
1875 // Handle destructive annihilation due to identities between elements in the
1876 // argument list here.
1877 unsigned NumOps = Ops.size();
1880 case Instruction::And:
1881 case Instruction::Or:
1882 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1886 case Instruction::Xor:
1887 if (Value *Result = OptimizeXor(I, Ops))
1891 case Instruction::Add:
1892 case Instruction::FAdd:
1893 if (Value *Result = OptimizeAdd(I, Ops))
1897 case Instruction::Mul:
1898 case Instruction::FMul:
1899 if (Value *Result = OptimizeMul(I, Ops))
1904 if (Ops.size() != NumOps)
1905 return OptimizeExpression(I, Ops);
1909 /// EraseInst - Zap the given instruction, adding interesting operands to the
1911 void Reassociate::EraseInst(Instruction *I) {
1912 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1913 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1914 // Erase the dead instruction.
1915 ValueRankMap.erase(I);
1916 RedoInsts.remove(I);
1917 I->eraseFromParent();
1918 // Optimize its operands.
1919 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1920 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1921 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1922 // If this is a node in an expression tree, climb to the expression root
1923 // and add that since that's where optimization actually happens.
1924 unsigned Opcode = Op->getOpcode();
1925 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1927 Op = Op->user_back();
1928 RedoInsts.insert(Op);
1932 // Canonicalize expressions of the following form:
1933 // x + (-Constant * y) -> x - (Constant * y)
1934 // x - (-Constant * y) -> x + (Constant * y)
1935 Instruction *Reassociate::canonicalizeNegConstExpr(Instruction *I) {
1936 if (!I->hasOneUse() || I->getType()->isVectorTy())
1939 // Must have at least one constant operand.
1940 Constant *C0 = dyn_cast<Constant>(I->getOperand(0));
1941 Constant *C1 = dyn_cast<Constant>(I->getOperand(1));
1945 // Must be a negative ConstantInt or ConstantFP.
1946 Constant *C = C0 ? C0 : C1;
1947 unsigned ConstIdx = C0 ? 0 : 1;
1948 if (auto *CI = dyn_cast<ConstantInt>(C)) {
1949 if (!CI->isNegative())
1951 } else if (auto *CF = dyn_cast<ConstantFP>(C)) {
1952 if (!CF->isNegative())
1957 // User must be a binary operator with one or more uses.
1958 Instruction *User = I->user_back();
1959 if (!isa<BinaryOperator>(User) || !User->getNumUses())
1962 // Must be a binary operator with higher precedence that add/sub.
1963 switch(I->getOpcode()) {
1966 case Instruction::Mul:
1967 case Instruction::FMul:
1968 case Instruction::UDiv:
1969 case Instruction::SDiv:
1970 case Instruction::FDiv:
1971 case Instruction::URem:
1972 case Instruction::SRem:
1973 case Instruction::FRem:
1977 unsigned UserOpcode = User->getOpcode();
1978 if (UserOpcode != Instruction::Add && UserOpcode != Instruction::FAdd &&
1979 UserOpcode != Instruction::Sub && UserOpcode != Instruction::FSub)
1982 // Subtraction is not commutative. Explicitly, the following transform is
1983 // not valid: (-Constant * y) - x -> x + (Constant * y)
1984 if (!User->isCommutative() && User->getOperand(1) != I)
1987 // Change the sign of the constant.
1988 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1989 I->setOperand(ConstIdx, ConstantInt::get(CI->getContext(), -CI->getValue()));
1991 ConstantFP *CF = cast<ConstantFP>(C);
1992 APFloat Val = CF->getValueAPF();
1994 I->setOperand(ConstIdx, ConstantFP::get(CF->getContext(), Val));
1997 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
1998 // ((-Const*y) + x) -> (x + (-Const*y)).
1999 if (User->getOperand(0) == I && User->isCommutative())
2000 cast<BinaryOperator>(User)->swapOperands();
2002 Value *Op0 = User->getOperand(0);
2003 Value *Op1 = User->getOperand(1);
2005 switch(UserOpcode) {
2007 llvm_unreachable("Unexpected Opcode!");
2008 case Instruction::Add:
2009 NI = BinaryOperator::CreateSub(Op0, Op1);
2011 case Instruction::Sub:
2012 NI = BinaryOperator::CreateAdd(Op0, Op1);
2014 case Instruction::FAdd:
2015 NI = BinaryOperator::CreateFSub(Op0, Op1);
2016 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2018 case Instruction::FSub:
2019 NI = BinaryOperator::CreateFAdd(Op0, Op1);
2020 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2024 NI->insertBefore(User);
2025 NI->setName(User->getName());
2026 User->replaceAllUsesWith(NI);
2027 NI->setDebugLoc(I->getDebugLoc());
2028 RedoInsts.insert(I);
2033 /// OptimizeInst - Inspect and optimize the given instruction. Note that erasing
2034 /// instructions is not allowed.
2035 void Reassociate::OptimizeInst(Instruction *I) {
2036 // Only consider operations that we understand.
2037 if (!isa<BinaryOperator>(I))
2040 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
2041 // If an operand of this shift is a reassociable multiply, or if the shift
2042 // is used by a reassociable multiply or add, turn into a multiply.
2043 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
2045 (isReassociableOp(I->user_back(), Instruction::Mul) ||
2046 isReassociableOp(I->user_back(), Instruction::Add)))) {
2047 Instruction *NI = ConvertShiftToMul(I);
2048 RedoInsts.insert(I);
2053 // Canonicalize negative constants out of expressions.
2054 if (Instruction *Res = canonicalizeNegConstExpr(I))
2057 // Commute floating point binary operators, to canonicalize the order of their
2058 // operands. This can potentially expose more CSE opportunities, and makes
2059 // writing other transformations simpler.
2060 if (I->getType()->isFloatingPointTy() || I->getType()->isVectorTy()) {
2062 // FAdd and FMul can be commuted.
2063 unsigned Opcode = I->getOpcode();
2064 if (Opcode == Instruction::FMul || Opcode == Instruction::FAdd) {
2065 Value *LHS = I->getOperand(0);
2066 Value *RHS = I->getOperand(1);
2067 unsigned LHSRank = getRank(LHS);
2068 unsigned RHSRank = getRank(RHS);
2070 // Sort the operands by rank.
2071 if (RHSRank < LHSRank) {
2072 I->setOperand(0, RHS);
2073 I->setOperand(1, LHS);
2077 // FIXME: We should commute vector instructions as well. However, this
2078 // requires further analysis to determine the effect on later passes.
2080 // Don't try to optimize vector instructions or anything that doesn't have
2082 if (I->getType()->isVectorTy() || !I->hasUnsafeAlgebra())
2086 // Do not reassociate boolean (i1) expressions. We want to preserve the
2087 // original order of evaluation for short-circuited comparisons that
2088 // SimplifyCFG has folded to AND/OR expressions. If the expression
2089 // is not further optimized, it is likely to be transformed back to a
2090 // short-circuited form for code gen, and the source order may have been
2091 // optimized for the most likely conditions.
2092 if (I->getType()->isIntegerTy(1))
2095 // If this is a subtract instruction which is not already in negate form,
2096 // see if we can convert it to X+-Y.
2097 if (I->getOpcode() == Instruction::Sub) {
2098 if (ShouldBreakUpSubtract(I)) {
2099 Instruction *NI = BreakUpSubtract(I);
2100 RedoInsts.insert(I);
2103 } else if (BinaryOperator::isNeg(I)) {
2104 // Otherwise, this is a negation. See if the operand is a multiply tree
2105 // and if this is not an inner node of a multiply tree.
2106 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2108 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2109 Instruction *NI = LowerNegateToMultiply(I);
2110 RedoInsts.insert(I);
2115 } else if (I->getOpcode() == Instruction::FSub) {
2116 if (ShouldBreakUpSubtract(I)) {
2117 Instruction *NI = BreakUpSubtract(I);
2118 RedoInsts.insert(I);
2121 } else if (BinaryOperator::isFNeg(I)) {
2122 // Otherwise, this is a negation. See if the operand is a multiply tree
2123 // and if this is not an inner node of a multiply tree.
2124 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2126 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2127 Instruction *NI = LowerNegateToMultiply(I);
2128 RedoInsts.insert(I);
2135 // If this instruction is an associative binary operator, process it.
2136 if (!I->isAssociative()) return;
2137 BinaryOperator *BO = cast<BinaryOperator>(I);
2139 // If this is an interior node of a reassociable tree, ignore it until we
2140 // get to the root of the tree, to avoid N^2 analysis.
2141 unsigned Opcode = BO->getOpcode();
2142 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode)
2145 // If this is an add tree that is used by a sub instruction, ignore it
2146 // until we process the subtract.
2147 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2148 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2150 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2151 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2154 ReassociateExpression(BO);
2157 void Reassociate::ReassociateExpression(BinaryOperator *I) {
2158 assert(!I->getType()->isVectorTy() &&
2159 "Reassociation of vector instructions is not supported.");
2161 // First, walk the expression tree, linearizing the tree, collecting the
2162 // operand information.
2163 SmallVector<RepeatedValue, 8> Tree;
2164 MadeChange |= LinearizeExprTree(I, Tree);
2165 SmallVector<ValueEntry, 8> Ops;
2166 Ops.reserve(Tree.size());
2167 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2168 RepeatedValue E = Tree[i];
2169 Ops.append(E.second.getZExtValue(),
2170 ValueEntry(getRank(E.first), E.first));
2173 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2175 // Now that we have linearized the tree to a list and have gathered all of
2176 // the operands and their ranks, sort the operands by their rank. Use a
2177 // stable_sort so that values with equal ranks will have their relative
2178 // positions maintained (and so the compiler is deterministic). Note that
2179 // this sorts so that the highest ranking values end up at the beginning of
2181 std::stable_sort(Ops.begin(), Ops.end());
2183 // OptimizeExpression - Now that we have the expression tree in a convenient
2184 // sorted form, optimize it globally if possible.
2185 if (Value *V = OptimizeExpression(I, Ops)) {
2187 // Self-referential expression in unreachable code.
2189 // This expression tree simplified to something that isn't a tree,
2191 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2192 I->replaceAllUsesWith(V);
2193 if (Instruction *VI = dyn_cast<Instruction>(V))
2194 VI->setDebugLoc(I->getDebugLoc());
2195 RedoInsts.insert(I);
2200 // We want to sink immediates as deeply as possible except in the case where
2201 // this is a multiply tree used only by an add, and the immediate is a -1.
2202 // In this case we reassociate to put the negation on the outside so that we
2203 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2204 if (I->hasOneUse()) {
2205 if (I->getOpcode() == Instruction::Mul &&
2206 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2207 isa<ConstantInt>(Ops.back().Op) &&
2208 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
2209 ValueEntry Tmp = Ops.pop_back_val();
2210 Ops.insert(Ops.begin(), Tmp);
2211 } else if (I->getOpcode() == Instruction::FMul &&
2212 cast<Instruction>(I->user_back())->getOpcode() ==
2213 Instruction::FAdd &&
2214 isa<ConstantFP>(Ops.back().Op) &&
2215 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2216 ValueEntry Tmp = Ops.pop_back_val();
2217 Ops.insert(Ops.begin(), Tmp);
2221 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2223 if (Ops.size() == 1) {
2225 // Self-referential expression in unreachable code.
2228 // This expression tree simplified to something that isn't a tree,
2230 I->replaceAllUsesWith(Ops[0].Op);
2231 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2232 OI->setDebugLoc(I->getDebugLoc());
2233 RedoInsts.insert(I);
2237 // Now that we ordered and optimized the expressions, splat them back into
2238 // the expression tree, removing any unneeded nodes.
2239 RewriteExprTree(I, Ops);
2242 bool Reassociate::runOnFunction(Function &F) {
2243 if (skipOptnoneFunction(F))
2246 // Calculate the rank map for F
2250 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
2251 // Optimize every instruction in the basic block.
2252 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; )
2253 if (isInstructionTriviallyDead(II)) {
2257 assert(II->getParent() == BI && "Moved to a different block!");
2261 // If this produced extra instructions to optimize, handle them now.
2262 while (!RedoInsts.empty()) {
2263 Instruction *I = RedoInsts.pop_back_val();
2264 if (isInstructionTriviallyDead(I))
2271 // We are done with the rank map.
2273 ValueRankMap.clear();