1 //===- Reassociate.cpp - Reassociate binary expressions -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass reassociates commutative expressions in an order that is designed
11 // to promote better constant propagation, GCSE, LICM, PRE, etc.
13 // For example: 4 + (x + 5) -> x + (4 + 5)
15 // In the implementation of this algorithm, constants are assigned rank = 0,
16 // function arguments are rank = 1, and other values are assigned ranks
17 // corresponding to the reverse post order traversal of current function
18 // (starting at 2), which effectively gives values in deep loops higher rank
19 // than values not in loops.
21 //===----------------------------------------------------------------------===//
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/ADT/DenseMap.h"
25 #include "llvm/ADT/PostOrderIterator.h"
26 #include "llvm/ADT/STLExtras.h"
27 #include "llvm/ADT/SetVector.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/IR/CFG.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/IRBuilder.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/ValueHandle.h"
38 #include "llvm/Pass.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Transforms/Utils/Local.h"
45 #define DEBUG_TYPE "reassociate"
47 STATISTIC(NumChanged, "Number of insts reassociated");
48 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
49 STATISTIC(NumFactor , "Number of multiplies factored");
55 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {}
57 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) {
58 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start.
63 /// Print out the expression identified in the Ops list.
65 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
66 Module *M = I->getParent()->getParent()->getParent();
67 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
68 << *Ops[0].Op->getType() << '\t';
69 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
71 Ops[i].Op->printAsOperand(dbgs(), false, M);
72 dbgs() << ", #" << Ops[i].Rank << "] ";
78 /// \brief Utility class representing a base and exponent pair which form one
79 /// factor of some product.
84 Factor(Value *Base, unsigned Power) : Base(Base), Power(Power) {}
86 /// \brief Sort factors by their Base.
88 bool operator()(const Factor &LHS, const Factor &RHS) {
89 return LHS.Base < RHS.Base;
93 /// \brief Compare factors for equal bases.
95 bool operator()(const Factor &LHS, const Factor &RHS) {
96 return LHS.Base == RHS.Base;
100 /// \brief Sort factors in descending order by their power.
101 struct PowerDescendingSorter {
102 bool operator()(const Factor &LHS, const Factor &RHS) {
103 return LHS.Power > RHS.Power;
107 /// \brief Compare factors for equal powers.
109 bool operator()(const Factor &LHS, const Factor &RHS) {
110 return LHS.Power == RHS.Power;
115 /// Utility class representing a non-constant Xor-operand. We classify
116 /// non-constant Xor-Operands into two categories:
117 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0
119 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
121 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
122 /// operand as "E | 0"
127 bool isInvalid() const { return SymbolicPart == nullptr; }
128 bool isOrExpr() const { return isOr; }
129 Value *getValue() const { return OrigVal; }
130 Value *getSymbolicPart() const { return SymbolicPart; }
131 unsigned getSymbolicRank() const { return SymbolicRank; }
132 const APInt &getConstPart() const { return ConstPart; }
134 void Invalidate() { SymbolicPart = OrigVal = nullptr; }
135 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
137 // Sort the XorOpnd-Pointer in ascending order of symbolic-value-rank.
138 // The purpose is twofold:
139 // 1) Cluster together the operands sharing the same symbolic-value.
140 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
141 // could potentially shorten crital path, and expose more loop-invariants.
142 // Note that values' rank are basically defined in RPO order (FIXME).
143 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
144 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
145 // "z" in the order of X-Y-Z is better than any other orders.
146 struct PtrSortFunctor {
147 bool operator()(XorOpnd * const &LHS, XorOpnd * const &RHS) {
148 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
155 unsigned SymbolicRank;
161 class Reassociate : public FunctionPass {
162 DenseMap<BasicBlock*, unsigned> RankMap;
163 DenseMap<AssertingVH<Value>, unsigned> ValueRankMap;
164 SetVector<AssertingVH<Instruction> > RedoInsts;
167 static char ID; // Pass identification, replacement for typeid
168 Reassociate() : FunctionPass(ID) {
169 initializeReassociatePass(*PassRegistry::getPassRegistry());
172 bool runOnFunction(Function &F) override;
174 void getAnalysisUsage(AnalysisUsage &AU) const override {
175 AU.setPreservesCFG();
178 void BuildRankMap(Function &F);
179 unsigned getRank(Value *V);
180 void canonicalizeOperands(Instruction *I);
181 void ReassociateExpression(BinaryOperator *I);
182 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
183 Value *OptimizeExpression(BinaryOperator *I,
184 SmallVectorImpl<ValueEntry> &Ops);
185 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
186 Value *OptimizeXor(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
187 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, APInt &ConstOpnd,
189 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
190 APInt &ConstOpnd, Value *&Res);
191 bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
192 SmallVectorImpl<Factor> &Factors);
193 Value *buildMinimalMultiplyDAG(IRBuilder<> &Builder,
194 SmallVectorImpl<Factor> &Factors);
195 Value *OptimizeMul(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
196 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
197 void EraseInst(Instruction *I);
198 void OptimizeInst(Instruction *I);
199 Instruction *canonicalizeNegConstExpr(Instruction *I);
203 XorOpnd::XorOpnd(Value *V) {
204 assert(!isa<ConstantInt>(V) && "No ConstantInt");
206 Instruction *I = dyn_cast<Instruction>(V);
209 if (I && (I->getOpcode() == Instruction::Or ||
210 I->getOpcode() == Instruction::And)) {
211 Value *V0 = I->getOperand(0);
212 Value *V1 = I->getOperand(1);
213 if (isa<ConstantInt>(V0))
216 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) {
217 ConstPart = C->getValue();
219 isOr = (I->getOpcode() == Instruction::Or);
224 // view the operand as "V | 0"
226 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth());
230 char Reassociate::ID = 0;
231 INITIALIZE_PASS(Reassociate, "reassociate",
232 "Reassociate expressions", false, false)
234 // Public interface to the Reassociate pass
235 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
237 /// Return true if V is an instruction of the specified opcode and if it
238 /// only has one use.
239 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
240 if (V->hasOneUse() && isa<Instruction>(V) &&
241 cast<Instruction>(V)->getOpcode() == Opcode &&
242 (!isa<FPMathOperator>(V) ||
243 cast<Instruction>(V)->hasUnsafeAlgebra()))
244 return cast<BinaryOperator>(V);
248 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
250 if (V->hasOneUse() && isa<Instruction>(V) &&
251 (cast<Instruction>(V)->getOpcode() == Opcode1 ||
252 cast<Instruction>(V)->getOpcode() == Opcode2) &&
253 (!isa<FPMathOperator>(V) ||
254 cast<Instruction>(V)->hasUnsafeAlgebra()))
255 return cast<BinaryOperator>(V);
259 void Reassociate::BuildRankMap(Function &F) {
262 // Assign distinct ranks to function arguments.
263 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
264 ValueRankMap[&*I] = ++i;
265 DEBUG(dbgs() << "Calculated Rank[" << I->getName() << "] = " << i << "\n");
268 ReversePostOrderTraversal<Function*> RPOT(&F);
269 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
270 E = RPOT.end(); I != E; ++I) {
272 unsigned BBRank = RankMap[BB] = ++i << 16;
274 // Walk the basic block, adding precomputed ranks for any instructions that
275 // we cannot move. This ensures that the ranks for these instructions are
276 // all different in the block.
277 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
278 if (mayBeMemoryDependent(*I))
279 ValueRankMap[&*I] = ++BBRank;
283 unsigned Reassociate::getRank(Value *V) {
284 Instruction *I = dyn_cast<Instruction>(V);
286 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
287 return 0; // Otherwise it's a global or constant, rank 0.
290 if (unsigned Rank = ValueRankMap[I])
291 return Rank; // Rank already known?
293 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
294 // we can reassociate expressions for code motion! Since we do not recurse
295 // for PHI nodes, we cannot have infinite recursion here, because there
296 // cannot be loops in the value graph that do not go through PHI nodes.
297 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
298 for (unsigned i = 0, e = I->getNumOperands();
299 i != e && Rank != MaxRank; ++i)
300 Rank = std::max(Rank, getRank(I->getOperand(i)));
302 // If this is a not or neg instruction, do not count it for rank. This
303 // assures us that X and ~X will have the same rank.
304 if (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
305 !BinaryOperator::isFNeg(I))
308 DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
310 return ValueRankMap[I] = Rank;
313 // Canonicalize constants to RHS. Otherwise, sort the operands by rank.
314 void Reassociate::canonicalizeOperands(Instruction *I) {
315 assert(isa<BinaryOperator>(I) && "Expected binary operator.");
316 assert(I->isCommutative() && "Expected commutative operator.");
318 Value *LHS = I->getOperand(0);
319 Value *RHS = I->getOperand(1);
320 unsigned LHSRank = getRank(LHS);
321 unsigned RHSRank = getRank(RHS);
323 if (isa<Constant>(RHS))
326 if (isa<Constant>(LHS) || RHSRank < LHSRank)
327 cast<BinaryOperator>(I)->swapOperands();
330 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
331 Instruction *InsertBefore, Value *FlagsOp) {
332 if (S1->getType()->isIntOrIntVectorTy())
333 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
335 BinaryOperator *Res =
336 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
337 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
342 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
343 Instruction *InsertBefore, Value *FlagsOp) {
344 if (S1->getType()->isIntOrIntVectorTy())
345 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
347 BinaryOperator *Res =
348 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
349 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
354 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
355 Instruction *InsertBefore, Value *FlagsOp) {
356 if (S1->getType()->isIntOrIntVectorTy())
357 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
359 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
360 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
365 /// Replace 0-X with X*-1.
366 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
367 Type *Ty = Neg->getType();
368 Constant *NegOne = Ty->isIntOrIntVectorTy() ?
369 ConstantInt::getAllOnesValue(Ty) : 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 /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
380 /// 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 /// 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 /// 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 Changed = 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).second && "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).second && "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 || (isa<FPMathOperator>(Op) &&
666 !cast<Instruction>(Op)->hasUnsafeAlgebra())) &&
667 "Should have been handled above!");
668 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
670 // If this is a multiply expression, turn any internal negations into
671 // multiplies by -1 so they can be reassociated.
672 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
673 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
674 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
675 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
676 BO = LowerNegateToMultiply(BO);
677 DEBUG(dbgs() << *BO << '\n');
678 Worklist.push_back(std::make_pair(BO, Weight));
683 // Failed to morph into an expression of the right type. This really is
685 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
686 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
687 LeafOrder.push_back(Op);
692 // The leaves, repeated according to their weights, represent the linearized
693 // form of the expression.
694 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
695 Value *V = LeafOrder[i];
696 LeafMap::iterator It = Leaves.find(V);
697 if (It == Leaves.end())
698 // Node initially thought to be a leaf wasn't.
700 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
701 APInt Weight = It->second;
702 if (Weight.isMinValue())
703 // Leaf already output or weight reduction eliminated it.
705 // Ensure the leaf is only output once.
707 Ops.push_back(std::make_pair(V, Weight));
710 // For nilpotent operations or addition there may be no operands, for example
711 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
712 // in both cases the weight reduces to 0 causing the value to be skipped.
714 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
715 assert(Identity && "Associative operation without identity!");
716 Ops.emplace_back(Identity, APInt(Bitwidth, 1));
722 /// Now that the operands for this expression tree are
723 /// linearized and optimized, emit them in-order.
724 void Reassociate::RewriteExprTree(BinaryOperator *I,
725 SmallVectorImpl<ValueEntry> &Ops) {
726 assert(Ops.size() > 1 && "Single values should be used directly!");
728 // Since our optimizations should never increase the number of operations, the
729 // new expression can usually be written reusing the existing binary operators
730 // from the original expression tree, without creating any new instructions,
731 // though the rewritten expression may have a completely different topology.
732 // We take care to not change anything if the new expression will be the same
733 // as the original. If more than trivial changes (like commuting operands)
734 // were made then we are obliged to clear out any optional subclass data like
737 /// NodesToRewrite - Nodes from the original expression available for writing
738 /// the new expression into.
739 SmallVector<BinaryOperator*, 8> NodesToRewrite;
740 unsigned Opcode = I->getOpcode();
741 BinaryOperator *Op = I;
743 /// NotRewritable - The operands being written will be the leaves of the new
744 /// expression and must not be used as inner nodes (via NodesToRewrite) by
745 /// mistake. Inner nodes are always reassociable, and usually leaves are not
746 /// (if they were they would have been incorporated into the expression and so
747 /// would not be leaves), so most of the time there is no danger of this. But
748 /// in rare cases a leaf may become reassociable if an optimization kills uses
749 /// of it, or it may momentarily become reassociable during rewriting (below)
750 /// due it being removed as an operand of one of its uses. Ensure that misuse
751 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
752 /// leaves and refusing to reuse any of them as inner nodes.
753 SmallPtrSet<Value*, 8> NotRewritable;
754 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
755 NotRewritable.insert(Ops[i].Op);
757 // ExpressionChanged - Non-null if the rewritten expression differs from the
758 // original in some non-trivial way, requiring the clearing of optional flags.
759 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
760 BinaryOperator *ExpressionChanged = nullptr;
761 for (unsigned i = 0; ; ++i) {
762 // The last operation (which comes earliest in the IR) is special as both
763 // operands will come from Ops, rather than just one with the other being
765 if (i+2 == Ops.size()) {
766 Value *NewLHS = Ops[i].Op;
767 Value *NewRHS = Ops[i+1].Op;
768 Value *OldLHS = Op->getOperand(0);
769 Value *OldRHS = Op->getOperand(1);
771 if (NewLHS == OldLHS && NewRHS == OldRHS)
772 // Nothing changed, leave it alone.
775 if (NewLHS == OldRHS && NewRHS == OldLHS) {
776 // The order of the operands was reversed. Swap them.
777 DEBUG(dbgs() << "RA: " << *Op << '\n');
779 DEBUG(dbgs() << "TO: " << *Op << '\n');
785 // The new operation differs non-trivially from the original. Overwrite
786 // the old operands with the new ones.
787 DEBUG(dbgs() << "RA: " << *Op << '\n');
788 if (NewLHS != OldLHS) {
789 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
790 if (BO && !NotRewritable.count(BO))
791 NodesToRewrite.push_back(BO);
792 Op->setOperand(0, NewLHS);
794 if (NewRHS != OldRHS) {
795 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
796 if (BO && !NotRewritable.count(BO))
797 NodesToRewrite.push_back(BO);
798 Op->setOperand(1, NewRHS);
800 DEBUG(dbgs() << "TO: " << *Op << '\n');
802 ExpressionChanged = Op;
809 // Not the last operation. The left-hand side will be a sub-expression
810 // while the right-hand side will be the current element of Ops.
811 Value *NewRHS = Ops[i].Op;
812 if (NewRHS != Op->getOperand(1)) {
813 DEBUG(dbgs() << "RA: " << *Op << '\n');
814 if (NewRHS == Op->getOperand(0)) {
815 // The new right-hand side was already present as the left operand. If
816 // we are lucky then swapping the operands will sort out both of them.
819 // Overwrite with the new right-hand side.
820 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
821 if (BO && !NotRewritable.count(BO))
822 NodesToRewrite.push_back(BO);
823 Op->setOperand(1, NewRHS);
824 ExpressionChanged = Op;
826 DEBUG(dbgs() << "TO: " << *Op << '\n');
831 // Now deal with the left-hand side. If this is already an operation node
832 // from the original expression then just rewrite the rest of the expression
834 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
835 if (BO && !NotRewritable.count(BO)) {
840 // Otherwise, grab a spare node from the original expression and use that as
841 // the left-hand side. If there are no nodes left then the optimizers made
842 // an expression with more nodes than the original! This usually means that
843 // they did something stupid but it might mean that the problem was just too
844 // hard (finding the mimimal number of multiplications needed to realize a
845 // multiplication expression is NP-complete). Whatever the reason, smart or
846 // stupid, create a new node if there are none left.
847 BinaryOperator *NewOp;
848 if (NodesToRewrite.empty()) {
849 Constant *Undef = UndefValue::get(I->getType());
850 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
851 Undef, Undef, "", I);
852 if (NewOp->getType()->isFPOrFPVectorTy())
853 NewOp->setFastMathFlags(I->getFastMathFlags());
855 NewOp = NodesToRewrite.pop_back_val();
858 DEBUG(dbgs() << "RA: " << *Op << '\n');
859 Op->setOperand(0, NewOp);
860 DEBUG(dbgs() << "TO: " << *Op << '\n');
861 ExpressionChanged = Op;
867 // If the expression changed non-trivially then clear out all subclass data
868 // starting from the operator specified in ExpressionChanged, and compactify
869 // the operators to just before the expression root to guarantee that the
870 // expression tree is dominated by all of Ops.
871 if (ExpressionChanged)
873 // Preserve FastMathFlags.
874 if (isa<FPMathOperator>(I)) {
875 FastMathFlags Flags = I->getFastMathFlags();
876 ExpressionChanged->clearSubclassOptionalData();
877 ExpressionChanged->setFastMathFlags(Flags);
879 ExpressionChanged->clearSubclassOptionalData();
881 if (ExpressionChanged == I)
883 ExpressionChanged->moveBefore(I);
884 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
887 // Throw away any left over nodes from the original expression.
888 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
889 RedoInsts.insert(NodesToRewrite[i]);
892 /// Insert instructions before the instruction pointed to by BI,
893 /// that computes the negative version of the value specified. The negative
894 /// version of the value is returned, and BI is left pointing at the instruction
895 /// that should be processed next by the reassociation pass.
896 static Value *NegateValue(Value *V, Instruction *BI) {
897 if (Constant *C = dyn_cast<Constant>(V)) {
898 if (C->getType()->isFPOrFPVectorTy()) {
899 return ConstantExpr::getFNeg(C);
901 return ConstantExpr::getNeg(C);
905 // We are trying to expose opportunity for reassociation. One of the things
906 // that we want to do to achieve this is to push a negation as deep into an
907 // expression chain as possible, to expose the add instructions. In practice,
908 // this means that we turn this:
909 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
910 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
911 // the constants. We assume that instcombine will clean up the mess later if
912 // we introduce tons of unnecessary negation instructions.
914 if (BinaryOperator *I =
915 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
916 // Push the negates through the add.
917 I->setOperand(0, NegateValue(I->getOperand(0), BI));
918 I->setOperand(1, NegateValue(I->getOperand(1), BI));
919 if (I->getOpcode() == Instruction::Add) {
920 I->setHasNoUnsignedWrap(false);
921 I->setHasNoSignedWrap(false);
924 // We must move the add instruction here, because the neg instructions do
925 // not dominate the old add instruction in general. By moving it, we are
926 // assured that the neg instructions we just inserted dominate the
927 // instruction we are about to insert after them.
930 I->setName(I->getName()+".neg");
934 // Okay, we need to materialize a negated version of V with an instruction.
935 // Scan the use lists of V to see if we have one already.
936 for (User *U : V->users()) {
937 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
940 // We found one! Now we have to make sure that the definition dominates
941 // this use. We do this by moving it to the entry block (if it is a
942 // non-instruction value) or right after the definition. These negates will
943 // be zapped by reassociate later, so we don't need much finesse here.
944 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
946 // Verify that the negate is in this function, V might be a constant expr.
947 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
950 BasicBlock::iterator InsertPt;
951 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
952 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
953 InsertPt = II->getNormalDest()->begin();
955 InsertPt = InstInput;
958 while (isa<PHINode>(InsertPt)) ++InsertPt;
960 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
962 TheNeg->moveBefore(InsertPt);
963 if (TheNeg->getOpcode() == Instruction::Sub) {
964 TheNeg->setHasNoUnsignedWrap(false);
965 TheNeg->setHasNoSignedWrap(false);
967 TheNeg->andIRFlags(BI);
972 // Insert a 'neg' instruction that subtracts the value from zero to get the
974 return CreateNeg(V, V->getName() + ".neg", BI, BI);
977 /// Return true if we should break up this subtract of X-Y into (X + -Y).
978 static bool ShouldBreakUpSubtract(Instruction *Sub) {
979 // If this is a negation, we can't split it up!
980 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
983 // Don't breakup X - undef.
984 if (isa<UndefValue>(Sub->getOperand(1)))
987 // Don't bother to break this up unless either the LHS is an associable add or
988 // subtract or if this is only used by one.
989 Value *V0 = Sub->getOperand(0);
990 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
991 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
993 Value *V1 = Sub->getOperand(1);
994 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
995 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
997 Value *VB = Sub->user_back();
998 if (Sub->hasOneUse() &&
999 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
1000 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
1006 /// If we have (X-Y), and if either X is an add, or if this is only used by an
1007 /// add, transform this into (X+(0-Y)) to promote better reassociation.
1008 static BinaryOperator *BreakUpSubtract(Instruction *Sub) {
1009 // Convert a subtract into an add and a neg instruction. This allows sub
1010 // instructions to be commuted with other add instructions.
1012 // Calculate the negative value of Operand 1 of the sub instruction,
1013 // and set it as the RHS of the add instruction we just made.
1015 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
1016 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
1017 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
1018 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
1021 // Everyone now refers to the add instruction.
1022 Sub->replaceAllUsesWith(New);
1023 New->setDebugLoc(Sub->getDebugLoc());
1025 DEBUG(dbgs() << "Negated: " << *New << '\n');
1029 /// If this is a shift of a reassociable multiply or is used by one, change
1030 /// this into a multiply by a constant to assist with further reassociation.
1031 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1032 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
1033 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
1035 BinaryOperator *Mul =
1036 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
1037 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
1040 // Everyone now refers to the mul instruction.
1041 Shl->replaceAllUsesWith(Mul);
1042 Mul->setDebugLoc(Shl->getDebugLoc());
1044 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
1045 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
1047 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
1048 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
1050 Mul->setHasNoSignedWrap(true);
1051 Mul->setHasNoUnsignedWrap(NUW);
1055 /// Scan backwards and forwards among values with the same rank as element i
1056 /// to see if X exists. If X does not exist, return i. This is useful when
1057 /// scanning for 'x' when we see '-x' because they both get the same rank.
1058 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
1060 unsigned XRank = Ops[i].Rank;
1061 unsigned e = Ops.size();
1062 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
1065 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1066 if (Instruction *I2 = dyn_cast<Instruction>(X))
1067 if (I1->isIdenticalTo(I2))
1071 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
1074 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1075 if (Instruction *I2 = dyn_cast<Instruction>(X))
1076 if (I1->isIdenticalTo(I2))
1082 /// Emit a tree of add instructions, summing Ops together
1083 /// and returning the result. Insert the tree before I.
1084 static Value *EmitAddTreeOfValues(Instruction *I,
1085 SmallVectorImpl<WeakVH> &Ops){
1086 if (Ops.size() == 1) return Ops.back();
1088 Value *V1 = Ops.back();
1090 Value *V2 = EmitAddTreeOfValues(I, Ops);
1091 return CreateAdd(V2, V1, "tmp", I, I);
1094 /// If V is an expression tree that is a multiplication sequence,
1095 /// and if this sequence contains a multiply by Factor,
1096 /// remove Factor from the tree and return the new tree.
1097 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
1098 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1102 SmallVector<RepeatedValue, 8> Tree;
1103 MadeChange |= LinearizeExprTree(BO, Tree);
1104 SmallVector<ValueEntry, 8> Factors;
1105 Factors.reserve(Tree.size());
1106 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1107 RepeatedValue E = Tree[i];
1108 Factors.append(E.second.getZExtValue(),
1109 ValueEntry(getRank(E.first), E.first));
1112 bool FoundFactor = false;
1113 bool NeedsNegate = false;
1114 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1115 if (Factors[i].Op == Factor) {
1117 Factors.erase(Factors.begin()+i);
1121 // If this is a negative version of this factor, remove it.
1122 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1123 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1124 if (FC1->getValue() == -FC2->getValue()) {
1125 FoundFactor = NeedsNegate = true;
1126 Factors.erase(Factors.begin()+i);
1129 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1130 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1131 APFloat F1(FC1->getValueAPF());
1132 APFloat F2(FC2->getValueAPF());
1134 if (F1.compare(F2) == APFloat::cmpEqual) {
1135 FoundFactor = NeedsNegate = true;
1136 Factors.erase(Factors.begin() + i);
1144 // Make sure to restore the operands to the expression tree.
1145 RewriteExprTree(BO, Factors);
1149 BasicBlock::iterator InsertPt = BO; ++InsertPt;
1151 // If this was just a single multiply, remove the multiply and return the only
1152 // remaining operand.
1153 if (Factors.size() == 1) {
1154 RedoInsts.insert(BO);
1157 RewriteExprTree(BO, Factors);
1162 V = CreateNeg(V, "neg", InsertPt, BO);
1167 /// If V is a single-use multiply, recursively add its operands as factors,
1168 /// otherwise add V to the list of factors.
1170 /// Ops is the top-level list of add operands we're trying to factor.
1171 static void FindSingleUseMultiplyFactors(Value *V,
1172 SmallVectorImpl<Value*> &Factors,
1173 const SmallVectorImpl<ValueEntry> &Ops) {
1174 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1176 Factors.push_back(V);
1180 // Otherwise, add the LHS and RHS to the list of factors.
1181 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
1182 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
1185 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1186 /// This optimizes based on identities. If it can be reduced to a single Value,
1187 /// it is returned, otherwise the Ops list is mutated as necessary.
1188 static Value *OptimizeAndOrXor(unsigned Opcode,
1189 SmallVectorImpl<ValueEntry> &Ops) {
1190 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1191 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1192 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1193 // First, check for X and ~X in the operand list.
1194 assert(i < Ops.size());
1195 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1196 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1197 unsigned FoundX = FindInOperandList(Ops, i, X);
1199 if (Opcode == Instruction::And) // ...&X&~X = 0
1200 return Constant::getNullValue(X->getType());
1202 if (Opcode == Instruction::Or) // ...|X|~X = -1
1203 return Constant::getAllOnesValue(X->getType());
1207 // Next, check for duplicate pairs of values, which we assume are next to
1208 // each other, due to our sorting criteria.
1209 assert(i < Ops.size());
1210 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1211 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1212 // Drop duplicate values for And and Or.
1213 Ops.erase(Ops.begin()+i);
1219 // Drop pairs of values for Xor.
1220 assert(Opcode == Instruction::Xor);
1222 return Constant::getNullValue(Ops[0].Op->getType());
1225 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1233 /// Helper funciton of CombineXorOpnd(). It creates a bitwise-and
1234 /// instruction with the given two operands, and return the resulting
1235 /// instruction. There are two special cases: 1) if the constant operand is 0,
1236 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1238 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1239 const APInt &ConstOpnd) {
1240 if (ConstOpnd != 0) {
1241 if (!ConstOpnd.isAllOnesValue()) {
1242 LLVMContext &Ctx = Opnd->getType()->getContext();
1244 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1245 "and.ra", InsertBefore);
1246 I->setDebugLoc(InsertBefore->getDebugLoc());
1254 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1255 // into "R ^ C", where C would be 0, and R is a symbolic value.
1257 // If it was successful, true is returned, and the "R" and "C" is returned
1258 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1259 // and both "Res" and "ConstOpnd" remain unchanged.
1261 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1262 APInt &ConstOpnd, Value *&Res) {
1263 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1264 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1265 // = (x & ~c1) ^ (c1 ^ c2)
1266 // It is useful only when c1 == c2.
1267 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) {
1268 if (!Opnd1->getValue()->hasOneUse())
1271 const APInt &C1 = Opnd1->getConstPart();
1272 if (C1 != ConstOpnd)
1275 Value *X = Opnd1->getSymbolicPart();
1276 Res = createAndInstr(I, X, ~C1);
1277 // ConstOpnd was C2, now C1 ^ C2.
1280 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1281 RedoInsts.insert(T);
1288 // Helper function of OptimizeXor(). It tries to simplify
1289 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1292 // If it was successful, true is returned, and the "R" and "C" is returned
1293 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1294 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1295 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1296 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
1297 APInt &ConstOpnd, Value *&Res) {
1298 Value *X = Opnd1->getSymbolicPart();
1299 if (X != Opnd2->getSymbolicPart())
1302 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1303 int DeadInstNum = 1;
1304 if (Opnd1->getValue()->hasOneUse())
1306 if (Opnd2->getValue()->hasOneUse())
1310 // (x | c1) ^ (x & c2)
1311 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1312 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1313 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1315 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1316 if (Opnd2->isOrExpr())
1317 std::swap(Opnd1, Opnd2);
1319 const APInt &C1 = Opnd1->getConstPart();
1320 const APInt &C2 = Opnd2->getConstPart();
1321 APInt C3((~C1) ^ C2);
1323 // Do not increase code size!
1324 if (C3 != 0 && !C3.isAllOnesValue()) {
1325 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1326 if (NewInstNum > DeadInstNum)
1330 Res = createAndInstr(I, X, C3);
1333 } else if (Opnd1->isOrExpr()) {
1334 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1336 const APInt &C1 = Opnd1->getConstPart();
1337 const APInt &C2 = Opnd2->getConstPart();
1340 // Do not increase code size
1341 if (C3 != 0 && !C3.isAllOnesValue()) {
1342 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1343 if (NewInstNum > DeadInstNum)
1347 Res = createAndInstr(I, X, C3);
1350 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1352 const APInt &C1 = Opnd1->getConstPart();
1353 const APInt &C2 = Opnd2->getConstPart();
1355 Res = createAndInstr(I, X, C3);
1358 // Put the original operands in the Redo list; hope they will be deleted
1360 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1361 RedoInsts.insert(T);
1362 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1363 RedoInsts.insert(T);
1368 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1369 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1371 Value *Reassociate::OptimizeXor(Instruction *I,
1372 SmallVectorImpl<ValueEntry> &Ops) {
1373 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1376 if (Ops.size() == 1)
1379 SmallVector<XorOpnd, 8> Opnds;
1380 SmallVector<XorOpnd*, 8> OpndPtrs;
1381 Type *Ty = Ops[0].Op->getType();
1382 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1384 // Step 1: Convert ValueEntry to XorOpnd
1385 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1386 Value *V = Ops[i].Op;
1387 if (!isa<ConstantInt>(V)) {
1389 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1392 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1395 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1396 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1397 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1398 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1399 // when new elements are added to the vector.
1400 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1401 OpndPtrs.push_back(&Opnds[i]);
1403 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1404 // the same symbolic value cluster together. For instance, the input operand
1405 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1406 // ("x | 123", "x & 789", "y & 456").
1407 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor());
1409 // Step 3: Combine adjacent operands
1410 XorOpnd *PrevOpnd = nullptr;
1411 bool Changed = false;
1412 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1413 XorOpnd *CurrOpnd = OpndPtrs[i];
1414 // The combined value
1417 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1418 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1421 *CurrOpnd = XorOpnd(CV);
1423 CurrOpnd->Invalidate();
1428 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1429 PrevOpnd = CurrOpnd;
1433 // step 3.2: When previous and current operands share the same symbolic
1434 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1436 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1437 // Remove previous operand
1438 PrevOpnd->Invalidate();
1440 *CurrOpnd = XorOpnd(CV);
1441 PrevOpnd = CurrOpnd;
1443 CurrOpnd->Invalidate();
1450 // Step 4: Reassemble the Ops
1453 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1454 XorOpnd &O = Opnds[i];
1457 ValueEntry VE(getRank(O.getValue()), O.getValue());
1460 if (ConstOpnd != 0) {
1461 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1462 ValueEntry VE(getRank(C), C);
1465 int Sz = Ops.size();
1467 return Ops.back().Op;
1469 assert(ConstOpnd == 0);
1470 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1477 /// Optimize a series of operands to an 'add' instruction. This
1478 /// optimizes based on identities. If it can be reduced to a single Value, it
1479 /// is returned, otherwise the Ops list is mutated as necessary.
1480 Value *Reassociate::OptimizeAdd(Instruction *I,
1481 SmallVectorImpl<ValueEntry> &Ops) {
1482 // Scan the operand lists looking for X and -X pairs. If we find any, we
1483 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1485 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1487 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1488 Value *TheOp = Ops[i].Op;
1489 // Check to see if we've seen this operand before. If so, we factor all
1490 // instances of the operand together. Due to our sorting criteria, we know
1491 // that these need to be next to each other in the vector.
1492 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1493 // Rescan the list, remove all instances of this operand from the expr.
1494 unsigned NumFound = 0;
1496 Ops.erase(Ops.begin()+i);
1498 } while (i != Ops.size() && Ops[i].Op == TheOp);
1500 DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1503 // Insert a new multiply.
1504 Type *Ty = TheOp->getType();
1505 Constant *C = Ty->isIntOrIntVectorTy() ?
1506 ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1507 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1509 // Now that we have inserted a multiply, optimize it. This allows us to
1510 // handle cases that require multiple factoring steps, such as this:
1511 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1512 RedoInsts.insert(Mul);
1514 // If every add operand was a duplicate, return the multiply.
1518 // Otherwise, we had some input that didn't have the dupe, such as
1519 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1520 // things being added by this operation.
1521 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1528 // Check for X and -X or X and ~X in the operand list.
1529 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1530 !BinaryOperator::isNot(TheOp))
1534 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1535 X = BinaryOperator::getNegArgument(TheOp);
1536 else if (BinaryOperator::isNot(TheOp))
1537 X = BinaryOperator::getNotArgument(TheOp);
1539 unsigned FoundX = FindInOperandList(Ops, i, X);
1543 // Remove X and -X from the operand list.
1544 if (Ops.size() == 2 &&
1545 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1546 return Constant::getNullValue(X->getType());
1548 // Remove X and ~X from the operand list.
1549 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1550 return Constant::getAllOnesValue(X->getType());
1552 Ops.erase(Ops.begin()+i);
1556 --i; // Need to back up an extra one.
1557 Ops.erase(Ops.begin()+FoundX);
1559 --i; // Revisit element.
1560 e -= 2; // Removed two elements.
1562 // if X and ~X we append -1 to the operand list.
1563 if (BinaryOperator::isNot(TheOp)) {
1564 Value *V = Constant::getAllOnesValue(X->getType());
1565 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1570 // Scan the operand list, checking to see if there are any common factors
1571 // between operands. Consider something like A*A+A*B*C+D. We would like to
1572 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1573 // To efficiently find this, we count the number of times a factor occurs
1574 // for any ADD operands that are MULs.
1575 DenseMap<Value*, unsigned> FactorOccurrences;
1577 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1578 // where they are actually the same multiply.
1579 unsigned MaxOcc = 0;
1580 Value *MaxOccVal = nullptr;
1581 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1582 BinaryOperator *BOp =
1583 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1587 // Compute all of the factors of this added value.
1588 SmallVector<Value*, 8> Factors;
1589 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1590 assert(Factors.size() > 1 && "Bad linearize!");
1592 // Add one to FactorOccurrences for each unique factor in this op.
1593 SmallPtrSet<Value*, 8> Duplicates;
1594 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1595 Value *Factor = Factors[i];
1596 if (!Duplicates.insert(Factor).second)
1599 unsigned Occ = ++FactorOccurrences[Factor];
1605 // If Factor is a negative constant, add the negated value as a factor
1606 // because we can percolate the negate out. Watch for minint, which
1607 // cannot be positivified.
1608 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1609 if (CI->isNegative() && !CI->isMinValue(true)) {
1610 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1611 assert(!Duplicates.count(Factor) &&
1612 "Shouldn't have two constant factors, missed a canonicalize");
1613 unsigned Occ = ++FactorOccurrences[Factor];
1619 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1620 if (CF->isNegative()) {
1621 APFloat F(CF->getValueAPF());
1623 Factor = ConstantFP::get(CF->getContext(), F);
1624 assert(!Duplicates.count(Factor) &&
1625 "Shouldn't have two constant factors, missed a canonicalize");
1626 unsigned Occ = ++FactorOccurrences[Factor];
1636 // If any factor occurred more than one time, we can pull it out.
1638 DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1641 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1642 // this, we could otherwise run into situations where removing a factor
1643 // from an expression will drop a use of maxocc, and this can cause
1644 // RemoveFactorFromExpression on successive values to behave differently.
1645 Instruction *DummyInst =
1646 I->getType()->isIntOrIntVectorTy()
1647 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1648 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1650 SmallVector<WeakVH, 4> NewMulOps;
1651 for (unsigned i = 0; i != Ops.size(); ++i) {
1652 // Only try to remove factors from expressions we're allowed to.
1653 BinaryOperator *BOp =
1654 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1658 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1659 // The factorized operand may occur several times. Convert them all in
1661 for (unsigned j = Ops.size(); j != i;) {
1663 if (Ops[j].Op == Ops[i].Op) {
1664 NewMulOps.push_back(V);
1665 Ops.erase(Ops.begin()+j);
1672 // No need for extra uses anymore.
1675 unsigned NumAddedValues = NewMulOps.size();
1676 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1678 // Now that we have inserted the add tree, optimize it. This allows us to
1679 // handle cases that require multiple factoring steps, such as this:
1680 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1681 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1682 (void)NumAddedValues;
1683 if (Instruction *VI = dyn_cast<Instruction>(V))
1684 RedoInsts.insert(VI);
1686 // Create the multiply.
1687 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1689 // Rerun associate on the multiply in case the inner expression turned into
1690 // a multiply. We want to make sure that we keep things in canonical form.
1691 RedoInsts.insert(V2);
1693 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1694 // entire result expression is just the multiply "A*(B+C)".
1698 // Otherwise, we had some input that didn't have the factor, such as
1699 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1700 // things being added by this operation.
1701 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1707 /// \brief Build up a vector of value/power pairs factoring a product.
1709 /// Given a series of multiplication operands, build a vector of factors and
1710 /// the powers each is raised to when forming the final product. Sort them in
1711 /// the order of descending power.
1713 /// (x*x) -> [(x, 2)]
1714 /// ((x*x)*x) -> [(x, 3)]
1715 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1717 /// \returns Whether any factors have a power greater than one.
1718 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1719 SmallVectorImpl<Factor> &Factors) {
1720 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1721 // Compute the sum of powers of simplifiable factors.
1722 unsigned FactorPowerSum = 0;
1723 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1724 Value *Op = Ops[Idx-1].Op;
1726 // Count the number of occurrences of this value.
1728 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1730 // Track for simplification all factors which occur 2 or more times.
1732 FactorPowerSum += Count;
1735 // We can only simplify factors if the sum of the powers of our simplifiable
1736 // factors is 4 or higher. When that is the case, we will *always* have
1737 // a simplification. This is an important invariant to prevent cyclicly
1738 // trying to simplify already minimal formations.
1739 if (FactorPowerSum < 4)
1742 // Now gather the simplifiable factors, removing them from Ops.
1744 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1745 Value *Op = Ops[Idx-1].Op;
1747 // Count the number of occurrences of this value.
1749 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1753 // Move an even number of occurrences to Factors.
1756 FactorPowerSum += Count;
1757 Factors.push_back(Factor(Op, Count));
1758 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1761 // None of the adjustments above should have reduced the sum of factor powers
1762 // below our mininum of '4'.
1763 assert(FactorPowerSum >= 4);
1765 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1769 /// \brief Build a tree of multiplies, computing the product of Ops.
1770 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1771 SmallVectorImpl<Value*> &Ops) {
1772 if (Ops.size() == 1)
1775 Value *LHS = Ops.pop_back_val();
1777 if (LHS->getType()->isIntOrIntVectorTy())
1778 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1780 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1781 } while (!Ops.empty());
1786 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1788 /// Given a vector of values raised to various powers, where no two values are
1789 /// equal and the powers are sorted in decreasing order, compute the minimal
1790 /// DAG of multiplies to compute the final product, and return that product
1792 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1793 SmallVectorImpl<Factor> &Factors) {
1794 assert(Factors[0].Power);
1795 SmallVector<Value *, 4> OuterProduct;
1796 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1797 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1798 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1803 // We want to multiply across all the factors with the same power so that
1804 // we can raise them to that power as a single entity. Build a mini tree
1806 SmallVector<Value *, 4> InnerProduct;
1807 InnerProduct.push_back(Factors[LastIdx].Base);
1809 InnerProduct.push_back(Factors[Idx].Base);
1811 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1813 // Reset the base value of the first factor to the new expression tree.
1814 // We'll remove all the factors with the same power in a second pass.
1815 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1816 if (Instruction *MI = dyn_cast<Instruction>(M))
1817 RedoInsts.insert(MI);
1821 // Unique factors with equal powers -- we've folded them into the first one's
1823 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1824 Factor::PowerEqual()),
1827 // Iteratively collect the base of each factor with an add power into the
1828 // outer product, and halve each power in preparation for squaring the
1830 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1831 if (Factors[Idx].Power & 1)
1832 OuterProduct.push_back(Factors[Idx].Base);
1833 Factors[Idx].Power >>= 1;
1835 if (Factors[0].Power) {
1836 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1837 OuterProduct.push_back(SquareRoot);
1838 OuterProduct.push_back(SquareRoot);
1840 if (OuterProduct.size() == 1)
1841 return OuterProduct.front();
1843 Value *V = buildMultiplyTree(Builder, OuterProduct);
1847 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1848 SmallVectorImpl<ValueEntry> &Ops) {
1849 // We can only optimize the multiplies when there is a chain of more than
1850 // three, such that a balanced tree might require fewer total multiplies.
1854 // Try to turn linear trees of multiplies without other uses of the
1855 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1857 SmallVector<Factor, 4> Factors;
1858 if (!collectMultiplyFactors(Ops, Factors))
1859 return nullptr; // All distinct factors, so nothing left for us to do.
1861 IRBuilder<> Builder(I);
1862 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1866 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1867 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1871 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1872 SmallVectorImpl<ValueEntry> &Ops) {
1873 // Now that we have the linearized expression tree, try to optimize it.
1874 // Start by folding any constants that we found.
1875 Constant *Cst = nullptr;
1876 unsigned Opcode = I->getOpcode();
1877 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1878 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1879 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1881 // If there was nothing but constants then we are done.
1885 // Put the combined constant back at the end of the operand list, except if
1886 // there is no point. For example, an add of 0 gets dropped here, while a
1887 // multiplication by zero turns the whole expression into zero.
1888 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1889 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1891 Ops.push_back(ValueEntry(0, Cst));
1894 if (Ops.size() == 1) return Ops[0].Op;
1896 // Handle destructive annihilation due to identities between elements in the
1897 // argument list here.
1898 unsigned NumOps = Ops.size();
1901 case Instruction::And:
1902 case Instruction::Or:
1903 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1907 case Instruction::Xor:
1908 if (Value *Result = OptimizeXor(I, Ops))
1912 case Instruction::Add:
1913 case Instruction::FAdd:
1914 if (Value *Result = OptimizeAdd(I, Ops))
1918 case Instruction::Mul:
1919 case Instruction::FMul:
1920 if (Value *Result = OptimizeMul(I, Ops))
1925 if (Ops.size() != NumOps)
1926 return OptimizeExpression(I, Ops);
1930 /// Zap the given instruction, adding interesting operands to the work list.
1931 void Reassociate::EraseInst(Instruction *I) {
1932 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1933 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1934 // Erase the dead instruction.
1935 ValueRankMap.erase(I);
1936 RedoInsts.remove(I);
1937 I->eraseFromParent();
1938 // Optimize its operands.
1939 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1940 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1941 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1942 // If this is a node in an expression tree, climb to the expression root
1943 // and add that since that's where optimization actually happens.
1944 unsigned Opcode = Op->getOpcode();
1945 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1946 Visited.insert(Op).second)
1947 Op = Op->user_back();
1948 RedoInsts.insert(Op);
1952 // Canonicalize expressions of the following form:
1953 // x + (-Constant * y) -> x - (Constant * y)
1954 // x - (-Constant * y) -> x + (Constant * y)
1955 Instruction *Reassociate::canonicalizeNegConstExpr(Instruction *I) {
1956 if (!I->hasOneUse() || I->getType()->isVectorTy())
1959 // Must be a fmul or fdiv instruction.
1960 unsigned Opcode = I->getOpcode();
1961 if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv)
1964 auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0));
1965 auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1));
1967 // Both operands are constant, let it get constant folded away.
1971 ConstantFP *CF = C0 ? C0 : C1;
1973 // Must have one constant operand.
1977 // Must be a negative ConstantFP.
1978 if (!CF->isNegative())
1981 // User must be a binary operator with one or more uses.
1982 Instruction *User = I->user_back();
1983 if (!isa<BinaryOperator>(User) || !User->hasNUsesOrMore(1))
1986 unsigned UserOpcode = User->getOpcode();
1987 if (UserOpcode != Instruction::FAdd && 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 APFloat Val = CF->getValueAPF();
1998 I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val));
2000 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
2001 // ((-Const*y) + x) -> (x + (-Const*y)).
2002 if (User->getOperand(0) == I && User->isCommutative())
2003 cast<BinaryOperator>(User)->swapOperands();
2005 Value *Op0 = User->getOperand(0);
2006 Value *Op1 = User->getOperand(1);
2008 switch (UserOpcode) {
2010 llvm_unreachable("Unexpected Opcode!");
2011 case Instruction::FAdd:
2012 NI = BinaryOperator::CreateFSub(Op0, Op1);
2013 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2015 case Instruction::FSub:
2016 NI = BinaryOperator::CreateFAdd(Op0, Op1);
2017 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2021 NI->insertBefore(User);
2022 NI->setName(User->getName());
2023 User->replaceAllUsesWith(NI);
2024 NI->setDebugLoc(I->getDebugLoc());
2025 RedoInsts.insert(I);
2030 /// Inspect and optimize the given instruction. Note that erasing
2031 /// instructions is not allowed.
2032 void Reassociate::OptimizeInst(Instruction *I) {
2033 // Only consider operations that we understand.
2034 if (!isa<BinaryOperator>(I))
2037 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
2038 // If an operand of this shift is a reassociable multiply, or if the shift
2039 // is used by a reassociable multiply or add, turn into a multiply.
2040 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
2042 (isReassociableOp(I->user_back(), Instruction::Mul) ||
2043 isReassociableOp(I->user_back(), Instruction::Add)))) {
2044 Instruction *NI = ConvertShiftToMul(I);
2045 RedoInsts.insert(I);
2050 // Canonicalize negative constants out of expressions.
2051 if (Instruction *Res = canonicalizeNegConstExpr(I))
2054 // Commute binary operators, to canonicalize the order of their operands.
2055 // This can potentially expose more CSE opportunities, and makes writing other
2056 // transformations simpler.
2057 if (I->isCommutative())
2058 canonicalizeOperands(I);
2060 // TODO: We should optimize vector Xor instructions, but they are
2061 // currently unsupported.
2062 if (I->getType()->isVectorTy() && I->getOpcode() == Instruction::Xor)
2065 // Don't optimize floating point instructions that don't have unsafe algebra.
2066 if (I->getType()->isFloatingPointTy() && !I->hasUnsafeAlgebra())
2069 // Do not reassociate boolean (i1) expressions. We want to preserve the
2070 // original order of evaluation for short-circuited comparisons that
2071 // SimplifyCFG has folded to AND/OR expressions. If the expression
2072 // is not further optimized, it is likely to be transformed back to a
2073 // short-circuited form for code gen, and the source order may have been
2074 // optimized for the most likely conditions.
2075 if (I->getType()->isIntegerTy(1))
2078 // If this is a subtract instruction which is not already in negate form,
2079 // see if we can convert it to X+-Y.
2080 if (I->getOpcode() == Instruction::Sub) {
2081 if (ShouldBreakUpSubtract(I)) {
2082 Instruction *NI = BreakUpSubtract(I);
2083 RedoInsts.insert(I);
2086 } else if (BinaryOperator::isNeg(I)) {
2087 // Otherwise, this is a negation. See if the operand is a multiply tree
2088 // and if this is not an inner node of a multiply tree.
2089 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2091 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2092 Instruction *NI = LowerNegateToMultiply(I);
2093 RedoInsts.insert(I);
2098 } else if (I->getOpcode() == Instruction::FSub) {
2099 if (ShouldBreakUpSubtract(I)) {
2100 Instruction *NI = BreakUpSubtract(I);
2101 RedoInsts.insert(I);
2104 } else if (BinaryOperator::isFNeg(I)) {
2105 // Otherwise, this is a negation. See if the operand is a multiply tree
2106 // and if this is not an inner node of a multiply tree.
2107 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2109 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2110 Instruction *NI = LowerNegateToMultiply(I);
2111 RedoInsts.insert(I);
2118 // If this instruction is an associative binary operator, process it.
2119 if (!I->isAssociative()) return;
2120 BinaryOperator *BO = cast<BinaryOperator>(I);
2122 // If this is an interior node of a reassociable tree, ignore it until we
2123 // get to the root of the tree, to avoid N^2 analysis.
2124 unsigned Opcode = BO->getOpcode();
2125 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode)
2128 // If this is an add tree that is used by a sub instruction, ignore it
2129 // until we process the subtract.
2130 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2131 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2133 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2134 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2137 ReassociateExpression(BO);
2140 void Reassociate::ReassociateExpression(BinaryOperator *I) {
2141 // First, walk the expression tree, linearizing the tree, collecting the
2142 // operand information.
2143 SmallVector<RepeatedValue, 8> Tree;
2144 MadeChange |= LinearizeExprTree(I, Tree);
2145 SmallVector<ValueEntry, 8> Ops;
2146 Ops.reserve(Tree.size());
2147 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2148 RepeatedValue E = Tree[i];
2149 Ops.append(E.second.getZExtValue(),
2150 ValueEntry(getRank(E.first), E.first));
2153 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2155 // Now that we have linearized the tree to a list and have gathered all of
2156 // the operands and their ranks, sort the operands by their rank. Use a
2157 // stable_sort so that values with equal ranks will have their relative
2158 // positions maintained (and so the compiler is deterministic). Note that
2159 // this sorts so that the highest ranking values end up at the beginning of
2161 std::stable_sort(Ops.begin(), Ops.end());
2163 // Now that we have the expression tree in a convenient
2164 // sorted form, optimize it globally if possible.
2165 if (Value *V = OptimizeExpression(I, Ops)) {
2167 // Self-referential expression in unreachable code.
2169 // This expression tree simplified to something that isn't a tree,
2171 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2172 I->replaceAllUsesWith(V);
2173 if (Instruction *VI = dyn_cast<Instruction>(V))
2174 VI->setDebugLoc(I->getDebugLoc());
2175 RedoInsts.insert(I);
2180 // We want to sink immediates as deeply as possible except in the case where
2181 // this is a multiply tree used only by an add, and the immediate is a -1.
2182 // In this case we reassociate to put the negation on the outside so that we
2183 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2184 if (I->hasOneUse()) {
2185 if (I->getOpcode() == Instruction::Mul &&
2186 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2187 isa<ConstantInt>(Ops.back().Op) &&
2188 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
2189 ValueEntry Tmp = Ops.pop_back_val();
2190 Ops.insert(Ops.begin(), Tmp);
2191 } else if (I->getOpcode() == Instruction::FMul &&
2192 cast<Instruction>(I->user_back())->getOpcode() ==
2193 Instruction::FAdd &&
2194 isa<ConstantFP>(Ops.back().Op) &&
2195 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2196 ValueEntry Tmp = Ops.pop_back_val();
2197 Ops.insert(Ops.begin(), Tmp);
2201 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2203 if (Ops.size() == 1) {
2205 // Self-referential expression in unreachable code.
2208 // This expression tree simplified to something that isn't a tree,
2210 I->replaceAllUsesWith(Ops[0].Op);
2211 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2212 OI->setDebugLoc(I->getDebugLoc());
2213 RedoInsts.insert(I);
2217 // Now that we ordered and optimized the expressions, splat them back into
2218 // the expression tree, removing any unneeded nodes.
2219 RewriteExprTree(I, Ops);
2222 bool Reassociate::runOnFunction(Function &F) {
2223 if (skipOptnoneFunction(F))
2226 // Calculate the rank map for F
2230 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
2231 // Optimize every instruction in the basic block.
2232 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; )
2233 if (isInstructionTriviallyDead(II)) {
2237 assert(II->getParent() == BI && "Moved to a different block!");
2241 // If this produced extra instructions to optimize, handle them now.
2242 while (!RedoInsts.empty()) {
2243 Instruction *I = RedoInsts.pop_back_val();
2244 if (isInstructionTriviallyDead(I))
2251 // We are done with the rank map.
2253 ValueRankMap.clear();