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 (!isa<FPMathOperator>(V) ||
241 cast<Instruction>(V)->hasUnsafeAlgebra()))
242 return cast<BinaryOperator>(V);
246 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
248 if (V->hasOneUse() && isa<Instruction>(V) &&
249 (cast<Instruction>(V)->getOpcode() == Opcode1 ||
250 cast<Instruction>(V)->getOpcode() == Opcode2) &&
251 (!isa<FPMathOperator>(V) ||
252 cast<Instruction>(V)->hasUnsafeAlgebra()))
253 return cast<BinaryOperator>(V);
257 static bool isUnmovableInstruction(Instruction *I) {
258 switch (I->getOpcode()) {
259 case Instruction::PHI:
260 case Instruction::LandingPad:
261 case Instruction::Alloca:
262 case Instruction::Load:
263 case Instruction::Invoke:
264 case Instruction::UDiv:
265 case Instruction::SDiv:
266 case Instruction::FDiv:
267 case Instruction::URem:
268 case Instruction::SRem:
269 case Instruction::FRem:
271 case Instruction::Call:
272 return !isa<DbgInfoIntrinsic>(I);
278 void Reassociate::BuildRankMap(Function &F) {
281 // Assign distinct ranks to function arguments
282 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
283 ValueRankMap[&*I] = ++i;
285 ReversePostOrderTraversal<Function*> RPOT(&F);
286 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
287 E = RPOT.end(); I != E; ++I) {
289 unsigned BBRank = RankMap[BB] = ++i << 16;
291 // Walk the basic block, adding precomputed ranks for any instructions that
292 // we cannot move. This ensures that the ranks for these instructions are
293 // all different in the block.
294 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
295 if (isUnmovableInstruction(I))
296 ValueRankMap[&*I] = ++BBRank;
300 unsigned Reassociate::getRank(Value *V) {
301 Instruction *I = dyn_cast<Instruction>(V);
303 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
304 return 0; // Otherwise it's a global or constant, rank 0.
307 if (unsigned Rank = ValueRankMap[I])
308 return Rank; // Rank already known?
310 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
311 // we can reassociate expressions for code motion! Since we do not recurse
312 // for PHI nodes, we cannot have infinite recursion here, because there
313 // cannot be loops in the value graph that do not go through PHI nodes.
314 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
315 for (unsigned i = 0, e = I->getNumOperands();
316 i != e && Rank != MaxRank; ++i)
317 Rank = std::max(Rank, getRank(I->getOperand(i)));
319 // If this is a not or neg instruction, do not count it for rank. This
320 // assures us that X and ~X will have the same rank.
321 Type *Ty = V->getType();
322 if ((!Ty->isIntegerTy() && !Ty->isFloatingPointTy()) ||
323 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
324 !BinaryOperator::isFNeg(I)))
327 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = "
330 return ValueRankMap[I] = Rank;
333 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
334 Instruction *InsertBefore, Value *FlagsOp) {
335 if (S1->getType()->isIntegerTy())
336 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
338 BinaryOperator *Res =
339 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
340 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
345 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
346 Instruction *InsertBefore, Value *FlagsOp) {
347 if (S1->getType()->isIntegerTy())
348 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
350 BinaryOperator *Res =
351 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
352 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
357 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
358 Instruction *InsertBefore, Value *FlagsOp) {
359 if (S1->getType()->isIntegerTy())
360 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
362 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
363 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
368 /// LowerNegateToMultiply - Replace 0-X with X*-1.
370 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
371 Type *Ty = Neg->getType();
372 Constant *NegOne = Ty->isIntegerTy() ? ConstantInt::getAllOnesValue(Ty)
373 : ConstantFP::get(Ty, -1.0);
375 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
376 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
378 Neg->replaceAllUsesWith(Res);
379 Res->setDebugLoc(Neg->getDebugLoc());
383 /// CarmichaelShift - Returns k such that lambda(2^Bitwidth) = 2^k, where lambda
384 /// is the Carmichael function. This means that x^(2^k) === 1 mod 2^Bitwidth for
385 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
386 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
387 /// even x in Bitwidth-bit arithmetic.
388 static unsigned CarmichaelShift(unsigned Bitwidth) {
394 /// IncorporateWeight - Add the extra weight 'RHS' to the existing weight 'LHS',
395 /// reducing the combined weight using any special properties of the operation.
396 /// The existing weight LHS represents the computation X op X op ... op X where
397 /// X occurs LHS times. The combined weight represents X op X op ... op X with
398 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
399 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
400 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
401 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
402 // If we were working with infinite precision arithmetic then the combined
403 // weight would be LHS + RHS. But we are using finite precision arithmetic,
404 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
405 // for nilpotent operations and addition, but not for idempotent operations
406 // and multiplication), so it is important to correctly reduce the combined
407 // weight back into range if wrapping would be wrong.
409 // If RHS is zero then the weight didn't change.
410 if (RHS.isMinValue())
412 // If LHS is zero then the combined weight is RHS.
413 if (LHS.isMinValue()) {
417 // From this point on we know that neither LHS nor RHS is zero.
419 if (Instruction::isIdempotent(Opcode)) {
420 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
421 // weight of 1. Keeping weights at zero or one also means that wrapping is
423 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
424 return; // Return a weight of 1.
426 if (Instruction::isNilpotent(Opcode)) {
427 // Nilpotent means X op X === 0, so reduce weights modulo 2.
428 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
429 LHS = 0; // 1 + 1 === 0 modulo 2.
432 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
433 // TODO: Reduce the weight by exploiting nsw/nuw?
438 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
439 "Unknown associative operation!");
440 unsigned Bitwidth = LHS.getBitWidth();
441 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
442 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
443 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
444 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
445 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
446 // which by a happy accident means that they can always be represented using
448 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
449 // the Carmichael number).
451 /// CM - The value of Carmichael's lambda function.
452 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
453 // Any weight W >= Threshold can be replaced with W - CM.
454 APInt Threshold = CM + Bitwidth;
455 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
456 // For Bitwidth 4 or more the following sum does not overflow.
458 while (LHS.uge(Threshold))
461 // To avoid problems with overflow do everything the same as above but using
463 unsigned CM = 1U << CarmichaelShift(Bitwidth);
464 unsigned Threshold = CM + Bitwidth;
465 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
466 "Weights not reduced!");
467 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
468 while (Total >= Threshold)
474 typedef std::pair<Value*, APInt> RepeatedValue;
476 /// LinearizeExprTree - Given an associative binary expression, return the leaf
477 /// nodes in Ops along with their weights (how many times the leaf occurs). The
478 /// original expression is the same as
479 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
481 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
485 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
487 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
489 /// This routine may modify the function, in which case it returns 'true'. The
490 /// changes it makes may well be destructive, changing the value computed by 'I'
491 /// to something completely different. Thus if the routine returns 'true' then
492 /// you MUST either replace I with a new expression computed from the Ops array,
493 /// or use RewriteExprTree to put the values back in.
495 /// A leaf node is either not a binary operation of the same kind as the root
496 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
497 /// opcode), or is the same kind of binary operator but has a use which either
498 /// does not belong to the expression, or does belong to the expression but is
499 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
500 /// of the expression, while for non-leaf nodes (except for the root 'I') every
501 /// use is a non-leaf node of the expression.
504 /// expression graph node names
514 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
515 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
517 /// The expression is maximal: if some instruction is a binary operator of the
518 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
519 /// then the instruction also belongs to the expression, is not a leaf node of
520 /// it, and its operands also belong to the expression (but may be leaf nodes).
522 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
523 /// order to ensure that every non-root node in the expression has *exactly one*
524 /// use by a non-leaf node of the expression. This destruction means that the
525 /// caller MUST either replace 'I' with a new expression or use something like
526 /// RewriteExprTree to put the values back in if the routine indicates that it
527 /// made a change by returning 'true'.
529 /// In the above example either the right operand of A or the left operand of B
530 /// will be replaced by undef. If it is B's operand then this gives:
534 /// + + | A, B - operand of B replaced with undef
540 /// Note that such undef operands can only be reached by passing through 'I'.
541 /// For example, if you visit operands recursively starting from a leaf node
542 /// then you will never see such an undef operand unless you get back to 'I',
543 /// which requires passing through a phi node.
545 /// Note that this routine may also mutate binary operators of the wrong type
546 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
547 /// of the expression) if it can turn them into binary operators of the right
548 /// type and thus make the expression bigger.
550 static bool LinearizeExprTree(BinaryOperator *I,
551 SmallVectorImpl<RepeatedValue> &Ops) {
552 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
553 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
554 unsigned Opcode = I->getOpcode();
555 assert(I->isAssociative() && I->isCommutative() &&
556 "Expected an associative and commutative operation!");
558 // Visit all operands of the expression, keeping track of their weight (the
559 // number of paths from the expression root to the operand, or if you like
560 // the number of times that operand occurs in the linearized expression).
561 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
562 // while A has weight two.
564 // Worklist of non-leaf nodes (their operands are in the expression too) along
565 // with their weights, representing a certain number of paths to the operator.
566 // If an operator occurs in the worklist multiple times then we found multiple
567 // ways to get to it.
568 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
569 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
570 bool MadeChange = false;
572 // Leaves of the expression are values that either aren't the right kind of
573 // operation (eg: a constant, or a multiply in an add tree), or are, but have
574 // some uses that are not inside the expression. For example, in I = X + X,
575 // X = A + B, the value X has two uses (by I) that are in the expression. If
576 // X has any other uses, for example in a return instruction, then we consider
577 // X to be a leaf, and won't analyze it further. When we first visit a value,
578 // if it has more than one use then at first we conservatively consider it to
579 // be a leaf. Later, as the expression is explored, we may discover some more
580 // uses of the value from inside the expression. If all uses turn out to be
581 // from within the expression (and the value is a binary operator of the right
582 // kind) then the value is no longer considered to be a leaf, and its operands
585 // Leaves - Keeps track of the set of putative leaves as well as the number of
586 // paths to each leaf seen so far.
587 typedef DenseMap<Value*, APInt> LeafMap;
588 LeafMap Leaves; // Leaf -> Total weight so far.
589 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
592 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
594 while (!Worklist.empty()) {
595 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
596 I = P.first; // We examine the operands of this binary operator.
598 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
599 Value *Op = I->getOperand(OpIdx);
600 APInt Weight = P.second; // Number of paths to this operand.
601 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
602 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
604 // If this is a binary operation of the right kind with only one use then
605 // add its operands to the expression.
606 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
607 assert(Visited.insert(Op) && "Not first visit!");
608 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
609 Worklist.push_back(std::make_pair(BO, Weight));
613 // Appears to be a leaf. Is the operand already in the set of leaves?
614 LeafMap::iterator It = Leaves.find(Op);
615 if (It == Leaves.end()) {
616 // Not in the leaf map. Must be the first time we saw this operand.
617 assert(Visited.insert(Op) && "Not first visit!");
618 if (!Op->hasOneUse()) {
619 // This value has uses not accounted for by the expression, so it is
620 // not safe to modify. Mark it as being a leaf.
621 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
622 LeafOrder.push_back(Op);
626 // No uses outside the expression, try morphing it.
627 } else if (It != Leaves.end()) {
628 // Already in the leaf map.
629 assert(Visited.count(Op) && "In leaf map but not visited!");
631 // Update the number of paths to the leaf.
632 IncorporateWeight(It->second, Weight, Opcode);
634 #if 0 // TODO: Re-enable once PR13021 is fixed.
635 // The leaf already has one use from inside the expression. As we want
636 // exactly one such use, drop this new use of the leaf.
637 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
638 I->setOperand(OpIdx, UndefValue::get(I->getType()));
641 // If the leaf is a binary operation of the right kind and we now see
642 // that its multiple original uses were in fact all by nodes belonging
643 // to the expression, then no longer consider it to be a leaf and add
644 // its operands to the expression.
645 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
646 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
647 Worklist.push_back(std::make_pair(BO, It->second));
653 // If we still have uses that are not accounted for by the expression
654 // then it is not safe to modify the value.
655 if (!Op->hasOneUse())
658 // No uses outside the expression, try morphing it.
660 Leaves.erase(It); // Since the value may be morphed below.
663 // At this point we have a value which, first of all, is not a binary
664 // expression of the right kind, and secondly, is only used inside the
665 // expression. This means that it can safely be modified. See if we
666 // can usefully morph it into an expression of the right kind.
667 assert((!isa<Instruction>(Op) ||
668 cast<Instruction>(Op)->getOpcode() != Opcode
669 || (isa<FPMathOperator>(Op) &&
670 !cast<Instruction>(Op)->hasUnsafeAlgebra())) &&
671 "Should have been handled above!");
672 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
674 // If this is a multiply expression, turn any internal negations into
675 // multiplies by -1 so they can be reassociated.
676 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
677 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
678 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
679 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
680 BO = LowerNegateToMultiply(BO);
681 DEBUG(dbgs() << *BO << '\n');
682 Worklist.push_back(std::make_pair(BO, Weight));
687 // Failed to morph into an expression of the right type. This really is
689 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
690 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
691 LeafOrder.push_back(Op);
696 // The leaves, repeated according to their weights, represent the linearized
697 // form of the expression.
698 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
699 Value *V = LeafOrder[i];
700 LeafMap::iterator It = Leaves.find(V);
701 if (It == Leaves.end())
702 // Node initially thought to be a leaf wasn't.
704 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
705 APInt Weight = It->second;
706 if (Weight.isMinValue())
707 // Leaf already output or weight reduction eliminated it.
709 // Ensure the leaf is only output once.
711 Ops.push_back(std::make_pair(V, Weight));
714 // For nilpotent operations or addition there may be no operands, for example
715 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
716 // in both cases the weight reduces to 0 causing the value to be skipped.
718 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
719 assert(Identity && "Associative operation without identity!");
720 Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1)));
726 // RewriteExprTree - Now that the operands for this expression tree are
727 // linearized and optimized, emit them in-order.
728 void Reassociate::RewriteExprTree(BinaryOperator *I,
729 SmallVectorImpl<ValueEntry> &Ops) {
730 assert(Ops.size() > 1 && "Single values should be used directly!");
732 // Since our optimizations should never increase the number of operations, the
733 // new expression can usually be written reusing the existing binary operators
734 // from the original expression tree, without creating any new instructions,
735 // though the rewritten expression may have a completely different topology.
736 // We take care to not change anything if the new expression will be the same
737 // as the original. If more than trivial changes (like commuting operands)
738 // were made then we are obliged to clear out any optional subclass data like
741 /// NodesToRewrite - Nodes from the original expression available for writing
742 /// the new expression into.
743 SmallVector<BinaryOperator*, 8> NodesToRewrite;
744 unsigned Opcode = I->getOpcode();
745 BinaryOperator *Op = I;
747 /// NotRewritable - The operands being written will be the leaves of the new
748 /// expression and must not be used as inner nodes (via NodesToRewrite) by
749 /// mistake. Inner nodes are always reassociable, and usually leaves are not
750 /// (if they were they would have been incorporated into the expression and so
751 /// would not be leaves), so most of the time there is no danger of this. But
752 /// in rare cases a leaf may become reassociable if an optimization kills uses
753 /// of it, or it may momentarily become reassociable during rewriting (below)
754 /// due it being removed as an operand of one of its uses. Ensure that misuse
755 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
756 /// leaves and refusing to reuse any of them as inner nodes.
757 SmallPtrSet<Value*, 8> NotRewritable;
758 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
759 NotRewritable.insert(Ops[i].Op);
761 // ExpressionChanged - Non-null if the rewritten expression differs from the
762 // original in some non-trivial way, requiring the clearing of optional flags.
763 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
764 BinaryOperator *ExpressionChanged = nullptr;
765 for (unsigned i = 0; ; ++i) {
766 // The last operation (which comes earliest in the IR) is special as both
767 // operands will come from Ops, rather than just one with the other being
769 if (i+2 == Ops.size()) {
770 Value *NewLHS = Ops[i].Op;
771 Value *NewRHS = Ops[i+1].Op;
772 Value *OldLHS = Op->getOperand(0);
773 Value *OldRHS = Op->getOperand(1);
775 if (NewLHS == OldLHS && NewRHS == OldRHS)
776 // Nothing changed, leave it alone.
779 if (NewLHS == OldRHS && NewRHS == OldLHS) {
780 // The order of the operands was reversed. Swap them.
781 DEBUG(dbgs() << "RA: " << *Op << '\n');
783 DEBUG(dbgs() << "TO: " << *Op << '\n');
789 // The new operation differs non-trivially from the original. Overwrite
790 // the old operands with the new ones.
791 DEBUG(dbgs() << "RA: " << *Op << '\n');
792 if (NewLHS != OldLHS) {
793 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
794 if (BO && !NotRewritable.count(BO))
795 NodesToRewrite.push_back(BO);
796 Op->setOperand(0, NewLHS);
798 if (NewRHS != OldRHS) {
799 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
800 if (BO && !NotRewritable.count(BO))
801 NodesToRewrite.push_back(BO);
802 Op->setOperand(1, NewRHS);
804 DEBUG(dbgs() << "TO: " << *Op << '\n');
806 ExpressionChanged = Op;
813 // Not the last operation. The left-hand side will be a sub-expression
814 // while the right-hand side will be the current element of Ops.
815 Value *NewRHS = Ops[i].Op;
816 if (NewRHS != Op->getOperand(1)) {
817 DEBUG(dbgs() << "RA: " << *Op << '\n');
818 if (NewRHS == Op->getOperand(0)) {
819 // The new right-hand side was already present as the left operand. If
820 // we are lucky then swapping the operands will sort out both of them.
823 // Overwrite with the new right-hand side.
824 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
825 if (BO && !NotRewritable.count(BO))
826 NodesToRewrite.push_back(BO);
827 Op->setOperand(1, NewRHS);
828 ExpressionChanged = Op;
830 DEBUG(dbgs() << "TO: " << *Op << '\n');
835 // Now deal with the left-hand side. If this is already an operation node
836 // from the original expression then just rewrite the rest of the expression
838 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
839 if (BO && !NotRewritable.count(BO)) {
844 // Otherwise, grab a spare node from the original expression and use that as
845 // the left-hand side. If there are no nodes left then the optimizers made
846 // an expression with more nodes than the original! This usually means that
847 // they did something stupid but it might mean that the problem was just too
848 // hard (finding the mimimal number of multiplications needed to realize a
849 // multiplication expression is NP-complete). Whatever the reason, smart or
850 // stupid, create a new node if there are none left.
851 BinaryOperator *NewOp;
852 if (NodesToRewrite.empty()) {
853 Constant *Undef = UndefValue::get(I->getType());
854 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
855 Undef, Undef, "", I);
856 if (NewOp->getType()->isFloatingPointTy())
857 NewOp->setFastMathFlags(I->getFastMathFlags());
859 NewOp = NodesToRewrite.pop_back_val();
862 DEBUG(dbgs() << "RA: " << *Op << '\n');
863 Op->setOperand(0, NewOp);
864 DEBUG(dbgs() << "TO: " << *Op << '\n');
865 ExpressionChanged = Op;
871 // If the expression changed non-trivially then clear out all subclass data
872 // starting from the operator specified in ExpressionChanged, and compactify
873 // the operators to just before the expression root to guarantee that the
874 // expression tree is dominated by all of Ops.
875 if (ExpressionChanged)
877 // Preserve FastMathFlags.
878 if (isa<FPMathOperator>(I)) {
879 FastMathFlags Flags = I->getFastMathFlags();
880 ExpressionChanged->clearSubclassOptionalData();
881 ExpressionChanged->setFastMathFlags(Flags);
883 ExpressionChanged->clearSubclassOptionalData();
885 if (ExpressionChanged == I)
887 ExpressionChanged->moveBefore(I);
888 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
891 // Throw away any left over nodes from the original expression.
892 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
893 RedoInsts.insert(NodesToRewrite[i]);
896 /// NegateValue - Insert instructions before the instruction pointed to by BI,
897 /// that computes the negative version of the value specified. The negative
898 /// version of the value is returned, and BI is left pointing at the instruction
899 /// that should be processed next by the reassociation pass.
900 static Value *NegateValue(Value *V, Instruction *BI) {
901 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
902 return ConstantExpr::getFNeg(C);
903 if (Constant *C = dyn_cast<Constant>(V))
904 return ConstantExpr::getNeg(C);
906 // We are trying to expose opportunity for reassociation. One of the things
907 // that we want to do to achieve this is to push a negation as deep into an
908 // expression chain as possible, to expose the add instructions. In practice,
909 // this means that we turn this:
910 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
911 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
912 // the constants. We assume that instcombine will clean up the mess later if
913 // we introduce tons of unnecessary negation instructions.
915 if (BinaryOperator *I =
916 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
917 // Push the negates through the add.
918 I->setOperand(0, NegateValue(I->getOperand(0), BI));
919 I->setOperand(1, NegateValue(I->getOperand(1), BI));
921 // We must move the add instruction here, because the neg instructions do
922 // not dominate the old add instruction in general. By moving it, we are
923 // assured that the neg instructions we just inserted dominate the
924 // instruction we are about to insert after them.
927 I->setName(I->getName()+".neg");
931 // Okay, we need to materialize a negated version of V with an instruction.
932 // Scan the use lists of V to see if we have one already.
933 for (User *U : V->users()) {
934 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
937 // We found one! Now we have to make sure that the definition dominates
938 // this use. We do this by moving it to the entry block (if it is a
939 // non-instruction value) or right after the definition. These negates will
940 // be zapped by reassociate later, so we don't need much finesse here.
941 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
943 // Verify that the negate is in this function, V might be a constant expr.
944 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
947 BasicBlock::iterator InsertPt;
948 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
949 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
950 InsertPt = II->getNormalDest()->begin();
952 InsertPt = InstInput;
955 while (isa<PHINode>(InsertPt)) ++InsertPt;
957 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
959 TheNeg->moveBefore(InsertPt);
963 // Insert a 'neg' instruction that subtracts the value from zero to get the
965 return CreateNeg(V, V->getName() + ".neg", BI, BI);
968 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
969 /// X-Y into (X + -Y).
970 static bool ShouldBreakUpSubtract(Instruction *Sub) {
971 // If this is a negation, we can't split it up!
972 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
975 // Don't breakup X - undef.
976 if (isa<UndefValue>(Sub->getOperand(1)))
979 // Don't bother to break this up unless either the LHS is an associable add or
980 // subtract or if this is only used by one.
981 Value *V0 = Sub->getOperand(0);
982 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
983 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
985 Value *V1 = Sub->getOperand(1);
986 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
987 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
989 Value *VB = Sub->user_back();
990 if (Sub->hasOneUse() &&
991 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
992 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
998 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
999 /// only used by an add, transform this into (X+(0-Y)) to promote better
1001 static BinaryOperator *BreakUpSubtract(Instruction *Sub) {
1002 // Convert a subtract into an add and a neg instruction. This allows sub
1003 // instructions to be commuted with other add instructions.
1005 // Calculate the negative value of Operand 1 of the sub instruction,
1006 // and set it as the RHS of the add instruction we just made.
1008 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
1009 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
1010 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
1011 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
1014 // Everyone now refers to the add instruction.
1015 Sub->replaceAllUsesWith(New);
1016 New->setDebugLoc(Sub->getDebugLoc());
1018 DEBUG(dbgs() << "Negated: " << *New << '\n');
1022 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
1023 /// by one, change this into a multiply by a constant to assist with further
1025 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1026 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
1027 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
1029 BinaryOperator *Mul =
1030 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
1031 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
1034 // Everyone now refers to the mul instruction.
1035 Shl->replaceAllUsesWith(Mul);
1036 Mul->setDebugLoc(Shl->getDebugLoc());
1038 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
1039 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
1041 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
1042 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
1044 Mul->setHasNoSignedWrap(true);
1045 Mul->setHasNoUnsignedWrap(NUW);
1049 /// FindInOperandList - Scan backwards and forwards among values with the same
1050 /// rank as element i to see if X exists. If X does not exist, return i. This
1051 /// is useful when scanning for 'x' when we see '-x' because they both get the
1053 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
1055 unsigned XRank = Ops[i].Rank;
1056 unsigned e = Ops.size();
1057 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
1060 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1061 if (Instruction *I2 = dyn_cast<Instruction>(X))
1062 if (I1->isIdenticalTo(I2))
1066 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
1069 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1070 if (Instruction *I2 = dyn_cast<Instruction>(X))
1071 if (I1->isIdenticalTo(I2))
1077 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
1078 /// and returning the result. Insert the tree before I.
1079 static Value *EmitAddTreeOfValues(Instruction *I,
1080 SmallVectorImpl<WeakVH> &Ops){
1081 if (Ops.size() == 1) return Ops.back();
1083 Value *V1 = Ops.back();
1085 Value *V2 = EmitAddTreeOfValues(I, Ops);
1086 return CreateAdd(V2, V1, "tmp", I, I);
1089 /// RemoveFactorFromExpression - If V is an expression tree that is a
1090 /// multiplication sequence, and if this sequence contains a multiply by Factor,
1091 /// remove Factor from the tree and return the new tree.
1092 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
1093 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1097 SmallVector<RepeatedValue, 8> Tree;
1098 MadeChange |= LinearizeExprTree(BO, Tree);
1099 SmallVector<ValueEntry, 8> Factors;
1100 Factors.reserve(Tree.size());
1101 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1102 RepeatedValue E = Tree[i];
1103 Factors.append(E.second.getZExtValue(),
1104 ValueEntry(getRank(E.first), E.first));
1107 bool FoundFactor = false;
1108 bool NeedsNegate = false;
1109 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1110 if (Factors[i].Op == Factor) {
1112 Factors.erase(Factors.begin()+i);
1116 // If this is a negative version of this factor, remove it.
1117 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1118 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1119 if (FC1->getValue() == -FC2->getValue()) {
1120 FoundFactor = NeedsNegate = true;
1121 Factors.erase(Factors.begin()+i);
1124 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1125 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1126 APFloat F1(FC1->getValueAPF());
1127 APFloat F2(FC2->getValueAPF());
1129 if (F1.compare(F2) == APFloat::cmpEqual) {
1130 FoundFactor = NeedsNegate = true;
1131 Factors.erase(Factors.begin() + i);
1139 // Make sure to restore the operands to the expression tree.
1140 RewriteExprTree(BO, Factors);
1144 BasicBlock::iterator InsertPt = BO; ++InsertPt;
1146 // If this was just a single multiply, remove the multiply and return the only
1147 // remaining operand.
1148 if (Factors.size() == 1) {
1149 RedoInsts.insert(BO);
1152 RewriteExprTree(BO, Factors);
1157 V = CreateNeg(V, "neg", InsertPt, BO);
1162 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
1163 /// add its operands as factors, otherwise add V to the list of factors.
1165 /// Ops is the top-level list of add operands we're trying to factor.
1166 static void FindSingleUseMultiplyFactors(Value *V,
1167 SmallVectorImpl<Value*> &Factors,
1168 const SmallVectorImpl<ValueEntry> &Ops) {
1169 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1171 Factors.push_back(V);
1175 // Otherwise, add the LHS and RHS to the list of factors.
1176 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
1177 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
1180 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
1181 /// instruction. This optimizes based on identities. If it can be reduced to
1182 /// a single Value, it is returned, otherwise the Ops list is mutated as
1184 static Value *OptimizeAndOrXor(unsigned Opcode,
1185 SmallVectorImpl<ValueEntry> &Ops) {
1186 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1187 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1188 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1189 // First, check for X and ~X in the operand list.
1190 assert(i < Ops.size());
1191 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1192 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1193 unsigned FoundX = FindInOperandList(Ops, i, X);
1195 if (Opcode == Instruction::And) // ...&X&~X = 0
1196 return Constant::getNullValue(X->getType());
1198 if (Opcode == Instruction::Or) // ...|X|~X = -1
1199 return Constant::getAllOnesValue(X->getType());
1203 // Next, check for duplicate pairs of values, which we assume are next to
1204 // each other, due to our sorting criteria.
1205 assert(i < Ops.size());
1206 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1207 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1208 // Drop duplicate values for And and Or.
1209 Ops.erase(Ops.begin()+i);
1215 // Drop pairs of values for Xor.
1216 assert(Opcode == Instruction::Xor);
1218 return Constant::getNullValue(Ops[0].Op->getType());
1221 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1229 /// Helper funciton of CombineXorOpnd(). It creates a bitwise-and
1230 /// instruction with the given two operands, and return the resulting
1231 /// instruction. There are two special cases: 1) if the constant operand is 0,
1232 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1234 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1235 const APInt &ConstOpnd) {
1236 if (ConstOpnd != 0) {
1237 if (!ConstOpnd.isAllOnesValue()) {
1238 LLVMContext &Ctx = Opnd->getType()->getContext();
1240 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1241 "and.ra", InsertBefore);
1242 I->setDebugLoc(InsertBefore->getDebugLoc());
1250 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1251 // into "R ^ C", where C would be 0, and R is a symbolic value.
1253 // If it was successful, true is returned, and the "R" and "C" is returned
1254 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1255 // and both "Res" and "ConstOpnd" remain unchanged.
1257 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1258 APInt &ConstOpnd, Value *&Res) {
1259 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1260 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1261 // = (x & ~c1) ^ (c1 ^ c2)
1262 // It is useful only when c1 == c2.
1263 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) {
1264 if (!Opnd1->getValue()->hasOneUse())
1267 const APInt &C1 = Opnd1->getConstPart();
1268 if (C1 != ConstOpnd)
1271 Value *X = Opnd1->getSymbolicPart();
1272 Res = createAndInstr(I, X, ~C1);
1273 // ConstOpnd was C2, now C1 ^ C2.
1276 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1277 RedoInsts.insert(T);
1284 // Helper function of OptimizeXor(). It tries to simplify
1285 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1288 // If it was successful, true is returned, and the "R" and "C" is returned
1289 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1290 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1291 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1292 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
1293 APInt &ConstOpnd, Value *&Res) {
1294 Value *X = Opnd1->getSymbolicPart();
1295 if (X != Opnd2->getSymbolicPart())
1298 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1299 int DeadInstNum = 1;
1300 if (Opnd1->getValue()->hasOneUse())
1302 if (Opnd2->getValue()->hasOneUse())
1306 // (x | c1) ^ (x & c2)
1307 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1308 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1309 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1311 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1312 if (Opnd2->isOrExpr())
1313 std::swap(Opnd1, Opnd2);
1315 const APInt &C1 = Opnd1->getConstPart();
1316 const APInt &C2 = Opnd2->getConstPart();
1317 APInt C3((~C1) ^ C2);
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 } else if (Opnd1->isOrExpr()) {
1330 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1332 const APInt &C1 = Opnd1->getConstPart();
1333 const APInt &C2 = Opnd2->getConstPart();
1336 // Do not increase code size
1337 if (C3 != 0 && !C3.isAllOnesValue()) {
1338 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1339 if (NewInstNum > DeadInstNum)
1343 Res = createAndInstr(I, X, C3);
1346 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1348 const APInt &C1 = Opnd1->getConstPart();
1349 const APInt &C2 = Opnd2->getConstPart();
1351 Res = createAndInstr(I, X, C3);
1354 // Put the original operands in the Redo list; hope they will be deleted
1356 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1357 RedoInsts.insert(T);
1358 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1359 RedoInsts.insert(T);
1364 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1365 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1367 Value *Reassociate::OptimizeXor(Instruction *I,
1368 SmallVectorImpl<ValueEntry> &Ops) {
1369 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1372 if (Ops.size() == 1)
1375 SmallVector<XorOpnd, 8> Opnds;
1376 SmallVector<XorOpnd*, 8> OpndPtrs;
1377 Type *Ty = Ops[0].Op->getType();
1378 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1380 // Step 1: Convert ValueEntry to XorOpnd
1381 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1382 Value *V = Ops[i].Op;
1383 if (!isa<ConstantInt>(V)) {
1385 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1388 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1391 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1392 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1393 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1394 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1395 // when new elements are added to the vector.
1396 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1397 OpndPtrs.push_back(&Opnds[i]);
1399 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1400 // the same symbolic value cluster together. For instance, the input operand
1401 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1402 // ("x | 123", "x & 789", "y & 456").
1403 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor());
1405 // Step 3: Combine adjacent operands
1406 XorOpnd *PrevOpnd = nullptr;
1407 bool Changed = false;
1408 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1409 XorOpnd *CurrOpnd = OpndPtrs[i];
1410 // The combined value
1413 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1414 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1417 *CurrOpnd = XorOpnd(CV);
1419 CurrOpnd->Invalidate();
1424 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1425 PrevOpnd = CurrOpnd;
1429 // step 3.2: When previous and current operands share the same symbolic
1430 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1432 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1433 // Remove previous operand
1434 PrevOpnd->Invalidate();
1436 *CurrOpnd = XorOpnd(CV);
1437 PrevOpnd = CurrOpnd;
1439 CurrOpnd->Invalidate();
1446 // Step 4: Reassemble the Ops
1449 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1450 XorOpnd &O = Opnds[i];
1453 ValueEntry VE(getRank(O.getValue()), O.getValue());
1456 if (ConstOpnd != 0) {
1457 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1458 ValueEntry VE(getRank(C), C);
1461 int Sz = Ops.size();
1463 return Ops.back().Op;
1465 assert(ConstOpnd == 0);
1466 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1473 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
1474 /// optimizes based on identities. If it can be reduced to a single Value, it
1475 /// is returned, otherwise the Ops list is mutated as necessary.
1476 Value *Reassociate::OptimizeAdd(Instruction *I,
1477 SmallVectorImpl<ValueEntry> &Ops) {
1478 // Scan the operand lists looking for X and -X pairs. If we find any, we
1479 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1481 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1483 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1484 Value *TheOp = Ops[i].Op;
1485 // Check to see if we've seen this operand before. If so, we factor all
1486 // instances of the operand together. Due to our sorting criteria, we know
1487 // that these need to be next to each other in the vector.
1488 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1489 // Rescan the list, remove all instances of this operand from the expr.
1490 unsigned NumFound = 0;
1492 Ops.erase(Ops.begin()+i);
1494 } while (i != Ops.size() && Ops[i].Op == TheOp);
1496 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1499 // Insert a new multiply.
1500 Type *Ty = TheOp->getType();
1501 Constant *C = Ty->isIntegerTy() ? ConstantInt::get(Ty, NumFound)
1502 : ConstantFP::get(Ty, NumFound);
1503 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1505 // Now that we have inserted a multiply, optimize it. This allows us to
1506 // handle cases that require multiple factoring steps, such as this:
1507 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1508 RedoInsts.insert(Mul);
1510 // If every add operand was a duplicate, return the multiply.
1514 // Otherwise, we had some input that didn't have the dupe, such as
1515 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1516 // things being added by this operation.
1517 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1524 // Check for X and -X or X and ~X in the operand list.
1525 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1526 !BinaryOperator::isNot(TheOp))
1530 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1531 X = BinaryOperator::getNegArgument(TheOp);
1532 else if (BinaryOperator::isNot(TheOp))
1533 X = BinaryOperator::getNotArgument(TheOp);
1535 unsigned FoundX = FindInOperandList(Ops, i, X);
1539 // Remove X and -X from the operand list.
1540 if (Ops.size() == 2 &&
1541 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1542 return Constant::getNullValue(X->getType());
1544 // Remove X and ~X from the operand list.
1545 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1546 return Constant::getAllOnesValue(X->getType());
1548 Ops.erase(Ops.begin()+i);
1552 --i; // Need to back up an extra one.
1553 Ops.erase(Ops.begin()+FoundX);
1555 --i; // Revisit element.
1556 e -= 2; // Removed two elements.
1558 // if X and ~X we append -1 to the operand list.
1559 if (BinaryOperator::isNot(TheOp)) {
1560 Value *V = Constant::getAllOnesValue(X->getType());
1561 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1566 // Scan the operand list, checking to see if there are any common factors
1567 // between operands. Consider something like A*A+A*B*C+D. We would like to
1568 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1569 // To efficiently find this, we count the number of times a factor occurs
1570 // for any ADD operands that are MULs.
1571 DenseMap<Value*, unsigned> FactorOccurrences;
1573 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1574 // where they are actually the same multiply.
1575 unsigned MaxOcc = 0;
1576 Value *MaxOccVal = nullptr;
1577 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1578 BinaryOperator *BOp =
1579 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1583 // Compute all of the factors of this added value.
1584 SmallVector<Value*, 8> Factors;
1585 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1586 assert(Factors.size() > 1 && "Bad linearize!");
1588 // Add one to FactorOccurrences for each unique factor in this op.
1589 SmallPtrSet<Value*, 8> Duplicates;
1590 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1591 Value *Factor = Factors[i];
1592 if (!Duplicates.insert(Factor))
1595 unsigned Occ = ++FactorOccurrences[Factor];
1601 // If Factor is a negative constant, add the negated value as a factor
1602 // because we can percolate the negate out. Watch for minint, which
1603 // cannot be positivified.
1604 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1605 if (CI->isNegative() && !CI->isMinValue(true)) {
1606 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1607 assert(!Duplicates.count(Factor) &&
1608 "Shouldn't have two constant factors, missed a canonicalize");
1609 unsigned Occ = ++FactorOccurrences[Factor];
1615 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1616 if (CF->isNegative()) {
1617 APFloat F(CF->getValueAPF());
1619 Factor = ConstantFP::get(CF->getContext(), F);
1620 assert(!Duplicates.count(Factor) &&
1621 "Shouldn't have two constant factors, missed a canonicalize");
1622 unsigned Occ = ++FactorOccurrences[Factor];
1632 // If any factor occurred more than one time, we can pull it out.
1634 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1637 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1638 // this, we could otherwise run into situations where removing a factor
1639 // from an expression will drop a use of maxocc, and this can cause
1640 // RemoveFactorFromExpression on successive values to behave differently.
1641 Instruction *DummyInst =
1642 I->getType()->isIntegerTy()
1643 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1644 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1646 SmallVector<WeakVH, 4> NewMulOps;
1647 for (unsigned i = 0; i != Ops.size(); ++i) {
1648 // Only try to remove factors from expressions we're allowed to.
1649 BinaryOperator *BOp =
1650 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1654 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1655 // The factorized operand may occur several times. Convert them all in
1657 for (unsigned j = Ops.size(); j != i;) {
1659 if (Ops[j].Op == Ops[i].Op) {
1660 NewMulOps.push_back(V);
1661 Ops.erase(Ops.begin()+j);
1668 // No need for extra uses anymore.
1671 unsigned NumAddedValues = NewMulOps.size();
1672 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1674 // Now that we have inserted the add tree, optimize it. This allows us to
1675 // handle cases that require multiple factoring steps, such as this:
1676 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1677 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1678 (void)NumAddedValues;
1679 if (Instruction *VI = dyn_cast<Instruction>(V))
1680 RedoInsts.insert(VI);
1682 // Create the multiply.
1683 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1685 // Rerun associate on the multiply in case the inner expression turned into
1686 // a multiply. We want to make sure that we keep things in canonical form.
1687 RedoInsts.insert(V2);
1689 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1690 // entire result expression is just the multiply "A*(B+C)".
1694 // Otherwise, we had some input that didn't have the factor, such as
1695 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1696 // things being added by this operation.
1697 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1703 /// \brief Build up a vector of value/power pairs factoring a product.
1705 /// Given a series of multiplication operands, build a vector of factors and
1706 /// the powers each is raised to when forming the final product. Sort them in
1707 /// the order of descending power.
1709 /// (x*x) -> [(x, 2)]
1710 /// ((x*x)*x) -> [(x, 3)]
1711 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1713 /// \returns Whether any factors have a power greater than one.
1714 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1715 SmallVectorImpl<Factor> &Factors) {
1716 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1717 // Compute the sum of powers of simplifiable factors.
1718 unsigned FactorPowerSum = 0;
1719 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1720 Value *Op = Ops[Idx-1].Op;
1722 // Count the number of occurrences of this value.
1724 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1726 // Track for simplification all factors which occur 2 or more times.
1728 FactorPowerSum += Count;
1731 // We can only simplify factors if the sum of the powers of our simplifiable
1732 // factors is 4 or higher. When that is the case, we will *always* have
1733 // a simplification. This is an important invariant to prevent cyclicly
1734 // trying to simplify already minimal formations.
1735 if (FactorPowerSum < 4)
1738 // Now gather the simplifiable factors, removing them from Ops.
1740 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1741 Value *Op = Ops[Idx-1].Op;
1743 // Count the number of occurrences of this value.
1745 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1749 // Move an even number of occurrences to Factors.
1752 FactorPowerSum += Count;
1753 Factors.push_back(Factor(Op, Count));
1754 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1757 // None of the adjustments above should have reduced the sum of factor powers
1758 // below our mininum of '4'.
1759 assert(FactorPowerSum >= 4);
1761 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1765 /// \brief Build a tree of multiplies, computing the product of Ops.
1766 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1767 SmallVectorImpl<Value*> &Ops) {
1768 if (Ops.size() == 1)
1771 Value *LHS = Ops.pop_back_val();
1773 if (LHS->getType()->isIntegerTy())
1774 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1776 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1777 } while (!Ops.empty());
1782 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1784 /// Given a vector of values raised to various powers, where no two values are
1785 /// equal and the powers are sorted in decreasing order, compute the minimal
1786 /// DAG of multiplies to compute the final product, and return that product
1788 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1789 SmallVectorImpl<Factor> &Factors) {
1790 assert(Factors[0].Power);
1791 SmallVector<Value *, 4> OuterProduct;
1792 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1793 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1794 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1799 // We want to multiply across all the factors with the same power so that
1800 // we can raise them to that power as a single entity. Build a mini tree
1802 SmallVector<Value *, 4> InnerProduct;
1803 InnerProduct.push_back(Factors[LastIdx].Base);
1805 InnerProduct.push_back(Factors[Idx].Base);
1807 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1809 // Reset the base value of the first factor to the new expression tree.
1810 // We'll remove all the factors with the same power in a second pass.
1811 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1812 if (Instruction *MI = dyn_cast<Instruction>(M))
1813 RedoInsts.insert(MI);
1817 // Unique factors with equal powers -- we've folded them into the first one's
1819 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1820 Factor::PowerEqual()),
1823 // Iteratively collect the base of each factor with an add power into the
1824 // outer product, and halve each power in preparation for squaring the
1826 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1827 if (Factors[Idx].Power & 1)
1828 OuterProduct.push_back(Factors[Idx].Base);
1829 Factors[Idx].Power >>= 1;
1831 if (Factors[0].Power) {
1832 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1833 OuterProduct.push_back(SquareRoot);
1834 OuterProduct.push_back(SquareRoot);
1836 if (OuterProduct.size() == 1)
1837 return OuterProduct.front();
1839 Value *V = buildMultiplyTree(Builder, OuterProduct);
1843 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1844 SmallVectorImpl<ValueEntry> &Ops) {
1845 // We can only optimize the multiplies when there is a chain of more than
1846 // three, such that a balanced tree might require fewer total multiplies.
1850 // Try to turn linear trees of multiplies without other uses of the
1851 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1853 SmallVector<Factor, 4> Factors;
1854 if (!collectMultiplyFactors(Ops, Factors))
1855 return nullptr; // All distinct factors, so nothing left for us to do.
1857 IRBuilder<> Builder(I);
1858 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1862 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1863 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1867 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1868 SmallVectorImpl<ValueEntry> &Ops) {
1869 // Now that we have the linearized expression tree, try to optimize it.
1870 // Start by folding any constants that we found.
1871 Constant *Cst = nullptr;
1872 unsigned Opcode = I->getOpcode();
1873 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1874 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1875 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1877 // If there was nothing but constants then we are done.
1881 // Put the combined constant back at the end of the operand list, except if
1882 // there is no point. For example, an add of 0 gets dropped here, while a
1883 // multiplication by zero turns the whole expression into zero.
1884 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1885 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1887 Ops.push_back(ValueEntry(0, Cst));
1890 if (Ops.size() == 1) return Ops[0].Op;
1892 // Handle destructive annihilation due to identities between elements in the
1893 // argument list here.
1894 unsigned NumOps = Ops.size();
1897 case Instruction::And:
1898 case Instruction::Or:
1899 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1903 case Instruction::Xor:
1904 if (Value *Result = OptimizeXor(I, Ops))
1908 case Instruction::Add:
1909 case Instruction::FAdd:
1910 if (Value *Result = OptimizeAdd(I, Ops))
1914 case Instruction::Mul:
1915 case Instruction::FMul:
1916 if (Value *Result = OptimizeMul(I, Ops))
1921 if (Ops.size() != NumOps)
1922 return OptimizeExpression(I, Ops);
1926 /// EraseInst - Zap the given instruction, adding interesting operands to the
1928 void Reassociate::EraseInst(Instruction *I) {
1929 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1930 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1931 // Erase the dead instruction.
1932 ValueRankMap.erase(I);
1933 RedoInsts.remove(I);
1934 I->eraseFromParent();
1935 // Optimize its operands.
1936 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1937 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1938 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1939 // If this is a node in an expression tree, climb to the expression root
1940 // and add that since that's where optimization actually happens.
1941 unsigned Opcode = Op->getOpcode();
1942 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1944 Op = Op->user_back();
1945 RedoInsts.insert(Op);
1949 // Canonicalize expressions of the following form:
1950 // x + (-Constant * y) -> x - (Constant * y)
1951 // x - (-Constant * y) -> x + (Constant * y)
1952 Instruction *Reassociate::canonicalizeNegConstExpr(Instruction *I) {
1953 if (!I->hasOneUse() || I->getType()->isVectorTy())
1956 // Must be a mul instruction.
1957 unsigned Opcode = I->getOpcode();
1958 if (Opcode != Instruction::Mul && Opcode != Instruction::FMul &&
1959 Opcode != Instruction::FDiv)
1962 // Must have at least one constant operand.
1963 Constant *C0 = dyn_cast<Constant>(I->getOperand(0));
1964 Constant *C1 = dyn_cast<Constant>(I->getOperand(1));
1968 // Must be a negative ConstantInt or ConstantFP.
1969 Constant *C = C0 ? C0 : C1;
1970 unsigned ConstIdx = C0 ? 0 : 1;
1971 if (auto *CI = dyn_cast<ConstantInt>(C)) {
1972 if (!CI->isNegative())
1974 } else if (auto *CF = dyn_cast<ConstantFP>(C)) {
1975 if (!CF->isNegative())
1980 // User must be a binary operator with one or more uses.
1981 Instruction *User = I->user_back();
1982 if (!isa<BinaryOperator>(User) || !User->getNumUses())
1985 unsigned UserOpcode = User->getOpcode();
1986 if (UserOpcode != Instruction::Add && UserOpcode != Instruction::FAdd &&
1987 UserOpcode != Instruction::Sub && UserOpcode != Instruction::FSub)
1990 // Subtraction is not commutative. Explicitly, the following transform is
1991 // not valid: (-Constant * y) - x -> x + (Constant * y)
1992 if (!User->isCommutative() && User->getOperand(1) != I)
1995 // Change the sign of the constant.
1996 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1997 I->setOperand(ConstIdx, ConstantInt::get(CI->getContext(), -CI->getValue()));
1999 ConstantFP *CF = cast<ConstantFP>(C);
2000 APFloat Val = CF->getValueAPF();
2002 I->setOperand(ConstIdx, ConstantFP::get(CF->getContext(), Val));
2005 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
2006 // ((-Const*y) + x) -> (x + (-Const*y)).
2007 if (User->getOperand(0) == I && User->isCommutative())
2008 cast<BinaryOperator>(User)->swapOperands();
2010 Value *Op0 = User->getOperand(0);
2011 Value *Op1 = User->getOperand(1);
2013 switch(UserOpcode) {
2015 llvm_unreachable("Unexpected Opcode!");
2016 case Instruction::Add:
2017 NI = BinaryOperator::CreateSub(Op0, Op1);
2019 case Instruction::Sub:
2020 NI = BinaryOperator::CreateAdd(Op0, Op1);
2022 case Instruction::FAdd:
2023 NI = BinaryOperator::CreateFSub(Op0, Op1);
2024 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2026 case Instruction::FSub:
2027 NI = BinaryOperator::CreateFAdd(Op0, Op1);
2028 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2032 NI->insertBefore(User);
2033 NI->setName(User->getName());
2034 User->replaceAllUsesWith(NI);
2035 NI->setDebugLoc(I->getDebugLoc());
2036 RedoInsts.insert(I);
2041 /// OptimizeInst - Inspect and optimize the given instruction. Note that erasing
2042 /// instructions is not allowed.
2043 void Reassociate::OptimizeInst(Instruction *I) {
2044 // Only consider operations that we understand.
2045 if (!isa<BinaryOperator>(I))
2048 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
2049 // If an operand of this shift is a reassociable multiply, or if the shift
2050 // is used by a reassociable multiply or add, turn into a multiply.
2051 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
2053 (isReassociableOp(I->user_back(), Instruction::Mul) ||
2054 isReassociableOp(I->user_back(), Instruction::Add)))) {
2055 Instruction *NI = ConvertShiftToMul(I);
2056 RedoInsts.insert(I);
2061 // Canonicalize negative constants out of expressions.
2062 if (Instruction *Res = canonicalizeNegConstExpr(I))
2065 // Commute floating point binary operators, to canonicalize the order of their
2066 // operands. This can potentially expose more CSE opportunities, and makes
2067 // writing other transformations simpler.
2068 if (I->getType()->isFloatingPointTy() || I->getType()->isVectorTy()) {
2070 // FAdd and FMul can be commuted.
2071 unsigned Opcode = I->getOpcode();
2072 if (Opcode == Instruction::FMul || Opcode == Instruction::FAdd) {
2073 Value *LHS = I->getOperand(0);
2074 Value *RHS = I->getOperand(1);
2075 unsigned LHSRank = getRank(LHS);
2076 unsigned RHSRank = getRank(RHS);
2078 // Sort the operands by rank.
2079 if (RHSRank < LHSRank) {
2080 I->setOperand(0, RHS);
2081 I->setOperand(1, LHS);
2085 // FIXME: We should commute vector instructions as well. However, this
2086 // requires further analysis to determine the effect on later passes.
2088 // Don't try to optimize vector instructions or anything that doesn't have
2090 if (I->getType()->isVectorTy() || !I->hasUnsafeAlgebra())
2094 // Do not reassociate boolean (i1) expressions. We want to preserve the
2095 // original order of evaluation for short-circuited comparisons that
2096 // SimplifyCFG has folded to AND/OR expressions. If the expression
2097 // is not further optimized, it is likely to be transformed back to a
2098 // short-circuited form for code gen, and the source order may have been
2099 // optimized for the most likely conditions.
2100 if (I->getType()->isIntegerTy(1))
2103 // If this is a subtract instruction which is not already in negate form,
2104 // see if we can convert it to X+-Y.
2105 if (I->getOpcode() == Instruction::Sub) {
2106 if (ShouldBreakUpSubtract(I)) {
2107 Instruction *NI = BreakUpSubtract(I);
2108 RedoInsts.insert(I);
2111 } else if (BinaryOperator::isNeg(I)) {
2112 // Otherwise, this is a negation. See if the operand is a multiply tree
2113 // and if this is not an inner node of a multiply tree.
2114 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2116 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2117 Instruction *NI = LowerNegateToMultiply(I);
2118 RedoInsts.insert(I);
2123 } else if (I->getOpcode() == Instruction::FSub) {
2124 if (ShouldBreakUpSubtract(I)) {
2125 Instruction *NI = BreakUpSubtract(I);
2126 RedoInsts.insert(I);
2129 } else if (BinaryOperator::isFNeg(I)) {
2130 // Otherwise, this is a negation. See if the operand is a multiply tree
2131 // and if this is not an inner node of a multiply tree.
2132 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2134 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2135 Instruction *NI = LowerNegateToMultiply(I);
2136 RedoInsts.insert(I);
2143 // If this instruction is an associative binary operator, process it.
2144 if (!I->isAssociative()) return;
2145 BinaryOperator *BO = cast<BinaryOperator>(I);
2147 // If this is an interior node of a reassociable tree, ignore it until we
2148 // get to the root of the tree, to avoid N^2 analysis.
2149 unsigned Opcode = BO->getOpcode();
2150 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode)
2153 // If this is an add tree that is used by a sub instruction, ignore it
2154 // until we process the subtract.
2155 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2156 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2158 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2159 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2162 ReassociateExpression(BO);
2165 void Reassociate::ReassociateExpression(BinaryOperator *I) {
2166 assert(!I->getType()->isVectorTy() &&
2167 "Reassociation of vector instructions is not supported.");
2169 // First, walk the expression tree, linearizing the tree, collecting the
2170 // operand information.
2171 SmallVector<RepeatedValue, 8> Tree;
2172 MadeChange |= LinearizeExprTree(I, Tree);
2173 SmallVector<ValueEntry, 8> Ops;
2174 Ops.reserve(Tree.size());
2175 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2176 RepeatedValue E = Tree[i];
2177 Ops.append(E.second.getZExtValue(),
2178 ValueEntry(getRank(E.first), E.first));
2181 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2183 // Now that we have linearized the tree to a list and have gathered all of
2184 // the operands and their ranks, sort the operands by their rank. Use a
2185 // stable_sort so that values with equal ranks will have their relative
2186 // positions maintained (and so the compiler is deterministic). Note that
2187 // this sorts so that the highest ranking values end up at the beginning of
2189 std::stable_sort(Ops.begin(), Ops.end());
2191 // OptimizeExpression - Now that we have the expression tree in a convenient
2192 // sorted form, optimize it globally if possible.
2193 if (Value *V = OptimizeExpression(I, Ops)) {
2195 // Self-referential expression in unreachable code.
2197 // This expression tree simplified to something that isn't a tree,
2199 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2200 I->replaceAllUsesWith(V);
2201 if (Instruction *VI = dyn_cast<Instruction>(V))
2202 VI->setDebugLoc(I->getDebugLoc());
2203 RedoInsts.insert(I);
2208 // We want to sink immediates as deeply as possible except in the case where
2209 // this is a multiply tree used only by an add, and the immediate is a -1.
2210 // In this case we reassociate to put the negation on the outside so that we
2211 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2212 if (I->hasOneUse()) {
2213 if (I->getOpcode() == Instruction::Mul &&
2214 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2215 isa<ConstantInt>(Ops.back().Op) &&
2216 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
2217 ValueEntry Tmp = Ops.pop_back_val();
2218 Ops.insert(Ops.begin(), Tmp);
2219 } else if (I->getOpcode() == Instruction::FMul &&
2220 cast<Instruction>(I->user_back())->getOpcode() ==
2221 Instruction::FAdd &&
2222 isa<ConstantFP>(Ops.back().Op) &&
2223 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2224 ValueEntry Tmp = Ops.pop_back_val();
2225 Ops.insert(Ops.begin(), Tmp);
2229 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2231 if (Ops.size() == 1) {
2233 // Self-referential expression in unreachable code.
2236 // This expression tree simplified to something that isn't a tree,
2238 I->replaceAllUsesWith(Ops[0].Op);
2239 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2240 OI->setDebugLoc(I->getDebugLoc());
2241 RedoInsts.insert(I);
2245 // Now that we ordered and optimized the expressions, splat them back into
2246 // the expression tree, removing any unneeded nodes.
2247 RewriteExprTree(I, Ops);
2250 bool Reassociate::runOnFunction(Function &F) {
2251 if (skipOptnoneFunction(F))
2254 // Calculate the rank map for F
2258 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
2259 // Optimize every instruction in the basic block.
2260 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; )
2261 if (isInstructionTriviallyDead(II)) {
2265 assert(II->getParent() == BI && "Moved to a different block!");
2269 // If this produced extra instructions to optimize, handle them now.
2270 while (!RedoInsts.empty()) {
2271 Instruction *I = RedoInsts.pop_back_val();
2272 if (isInstructionTriviallyDead(I))
2279 // We are done with the rank map.
2281 ValueRankMap.clear();