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 /// 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 canonicalizeOperands(Instruction *I);
180 void ReassociateExpression(BinaryOperator *I);
181 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
182 Value *OptimizeExpression(BinaryOperator *I,
183 SmallVectorImpl<ValueEntry> &Ops);
184 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
185 Value *OptimizeXor(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
186 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, APInt &ConstOpnd,
188 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
189 APInt &ConstOpnd, Value *&Res);
190 bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
191 SmallVectorImpl<Factor> &Factors);
192 Value *buildMinimalMultiplyDAG(IRBuilder<> &Builder,
193 SmallVectorImpl<Factor> &Factors);
194 Value *OptimizeMul(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
195 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
196 void EraseInst(Instruction *I);
197 void OptimizeInst(Instruction *I);
198 Instruction *canonicalizeNegConstExpr(Instruction *I);
202 XorOpnd::XorOpnd(Value *V) {
203 assert(!isa<ConstantInt>(V) && "No ConstantInt");
205 Instruction *I = dyn_cast<Instruction>(V);
208 if (I && (I->getOpcode() == Instruction::Or ||
209 I->getOpcode() == Instruction::And)) {
210 Value *V0 = I->getOperand(0);
211 Value *V1 = I->getOperand(1);
212 if (isa<ConstantInt>(V0))
215 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) {
216 ConstPart = C->getValue();
218 isOr = (I->getOpcode() == Instruction::Or);
223 // view the operand as "V | 0"
225 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth());
229 char Reassociate::ID = 0;
230 INITIALIZE_PASS(Reassociate, "reassociate",
231 "Reassociate expressions", false, false)
233 // Public interface to the Reassociate pass
234 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
236 /// Return true if V is an instruction of the specified opcode and if it
237 /// only has one use.
238 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
239 if (V->hasOneUse() && isa<Instruction>(V) &&
240 cast<Instruction>(V)->getOpcode() == Opcode &&
241 (!isa<FPMathOperator>(V) ||
242 cast<Instruction>(V)->hasUnsafeAlgebra()))
243 return cast<BinaryOperator>(V);
247 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
249 if (V->hasOneUse() && isa<Instruction>(V) &&
250 (cast<Instruction>(V)->getOpcode() == Opcode1 ||
251 cast<Instruction>(V)->getOpcode() == Opcode2) &&
252 (!isa<FPMathOperator>(V) ||
253 cast<Instruction>(V)->hasUnsafeAlgebra()))
254 return cast<BinaryOperator>(V);
258 static bool isUnmovableInstruction(Instruction *I) {
259 switch (I->getOpcode()) {
260 case Instruction::PHI:
261 case Instruction::LandingPad:
262 case Instruction::Alloca:
263 case Instruction::Load:
264 case Instruction::Invoke:
265 case Instruction::UDiv:
266 case Instruction::SDiv:
267 case Instruction::FDiv:
268 case Instruction::URem:
269 case Instruction::SRem:
270 case Instruction::FRem:
272 case Instruction::Call:
273 return !isa<DbgInfoIntrinsic>(I);
279 void Reassociate::BuildRankMap(Function &F) {
282 // Assign distinct ranks to function arguments.
283 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
284 ValueRankMap[&*I] = ++i;
285 DEBUG(dbgs() << "Calculated Rank[" << I->getName() << "] = " << i << "\n");
288 ReversePostOrderTraversal<Function*> RPOT(&F);
289 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
290 E = RPOT.end(); I != E; ++I) {
292 unsigned BBRank = RankMap[BB] = ++i << 16;
294 // Walk the basic block, adding precomputed ranks for any instructions that
295 // we cannot move. This ensures that the ranks for these instructions are
296 // all different in the block.
297 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
298 if (isUnmovableInstruction(I))
299 ValueRankMap[&*I] = ++BBRank;
303 unsigned Reassociate::getRank(Value *V) {
304 Instruction *I = dyn_cast<Instruction>(V);
306 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
307 return 0; // Otherwise it's a global or constant, rank 0.
310 if (unsigned Rank = ValueRankMap[I])
311 return Rank; // Rank already known?
313 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
314 // we can reassociate expressions for code motion! Since we do not recurse
315 // for PHI nodes, we cannot have infinite recursion here, because there
316 // cannot be loops in the value graph that do not go through PHI nodes.
317 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
318 for (unsigned i = 0, e = I->getNumOperands();
319 i != e && Rank != MaxRank; ++i)
320 Rank = std::max(Rank, getRank(I->getOperand(i)));
322 // If this is a not or neg instruction, do not count it for rank. This
323 // assures us that X and ~X will have the same rank.
324 if (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
325 !BinaryOperator::isFNeg(I))
328 DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
330 return ValueRankMap[I] = Rank;
333 // Canonicalize constants to RHS. Otherwise, sort the operands by rank.
334 void Reassociate::canonicalizeOperands(Instruction *I) {
335 assert(isa<BinaryOperator>(I) && "Expected binary operator.");
336 assert(I->isCommutative() && "Expected commutative operator.");
338 Value *LHS = I->getOperand(0);
339 Value *RHS = I->getOperand(1);
340 unsigned LHSRank = getRank(LHS);
341 unsigned RHSRank = getRank(RHS);
343 if (isa<Constant>(RHS))
346 if (isa<Constant>(LHS) || RHSRank < LHSRank)
347 cast<BinaryOperator>(I)->swapOperands();
350 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
351 Instruction *InsertBefore, Value *FlagsOp) {
352 if (S1->getType()->isIntOrIntVectorTy())
353 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
355 BinaryOperator *Res =
356 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
357 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
362 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
363 Instruction *InsertBefore, Value *FlagsOp) {
364 if (S1->getType()->isIntOrIntVectorTy())
365 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
367 BinaryOperator *Res =
368 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
369 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
374 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
375 Instruction *InsertBefore, Value *FlagsOp) {
376 if (S1->getType()->isIntOrIntVectorTy())
377 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
379 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
380 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
385 /// Replace 0-X with X*-1.
386 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
387 Type *Ty = Neg->getType();
388 Constant *NegOne = Ty->isIntOrIntVectorTy() ?
389 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
391 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
392 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
394 Neg->replaceAllUsesWith(Res);
395 Res->setDebugLoc(Neg->getDebugLoc());
399 /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
400 /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for
401 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
402 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
403 /// even x in Bitwidth-bit arithmetic.
404 static unsigned CarmichaelShift(unsigned Bitwidth) {
410 /// Add the extra weight 'RHS' to the existing weight 'LHS',
411 /// reducing the combined weight using any special properties of the operation.
412 /// The existing weight LHS represents the computation X op X op ... op X where
413 /// X occurs LHS times. The combined weight represents X op X op ... op X with
414 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
415 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
416 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
417 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
418 // If we were working with infinite precision arithmetic then the combined
419 // weight would be LHS + RHS. But we are using finite precision arithmetic,
420 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
421 // for nilpotent operations and addition, but not for idempotent operations
422 // and multiplication), so it is important to correctly reduce the combined
423 // weight back into range if wrapping would be wrong.
425 // If RHS is zero then the weight didn't change.
426 if (RHS.isMinValue())
428 // If LHS is zero then the combined weight is RHS.
429 if (LHS.isMinValue()) {
433 // From this point on we know that neither LHS nor RHS is zero.
435 if (Instruction::isIdempotent(Opcode)) {
436 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
437 // weight of 1. Keeping weights at zero or one also means that wrapping is
439 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
440 return; // Return a weight of 1.
442 if (Instruction::isNilpotent(Opcode)) {
443 // Nilpotent means X op X === 0, so reduce weights modulo 2.
444 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
445 LHS = 0; // 1 + 1 === 0 modulo 2.
448 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
449 // TODO: Reduce the weight by exploiting nsw/nuw?
454 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
455 "Unknown associative operation!");
456 unsigned Bitwidth = LHS.getBitWidth();
457 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
458 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
459 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
460 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
461 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
462 // which by a happy accident means that they can always be represented using
464 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
465 // the Carmichael number).
467 /// CM - The value of Carmichael's lambda function.
468 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
469 // Any weight W >= Threshold can be replaced with W - CM.
470 APInt Threshold = CM + Bitwidth;
471 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
472 // For Bitwidth 4 or more the following sum does not overflow.
474 while (LHS.uge(Threshold))
477 // To avoid problems with overflow do everything the same as above but using
479 unsigned CM = 1U << CarmichaelShift(Bitwidth);
480 unsigned Threshold = CM + Bitwidth;
481 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
482 "Weights not reduced!");
483 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
484 while (Total >= Threshold)
490 typedef std::pair<Value*, APInt> RepeatedValue;
492 /// Given an associative binary expression, return the leaf
493 /// nodes in Ops along with their weights (how many times the leaf occurs). The
494 /// original expression is the same as
495 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
497 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
501 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
503 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
505 /// This routine may modify the function, in which case it returns 'true'. The
506 /// changes it makes may well be destructive, changing the value computed by 'I'
507 /// to something completely different. Thus if the routine returns 'true' then
508 /// you MUST either replace I with a new expression computed from the Ops array,
509 /// or use RewriteExprTree to put the values back in.
511 /// A leaf node is either not a binary operation of the same kind as the root
512 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
513 /// opcode), or is the same kind of binary operator but has a use which either
514 /// does not belong to the expression, or does belong to the expression but is
515 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
516 /// of the expression, while for non-leaf nodes (except for the root 'I') every
517 /// use is a non-leaf node of the expression.
520 /// expression graph node names
530 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
531 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
533 /// The expression is maximal: if some instruction is a binary operator of the
534 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
535 /// then the instruction also belongs to the expression, is not a leaf node of
536 /// it, and its operands also belong to the expression (but may be leaf nodes).
538 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
539 /// order to ensure that every non-root node in the expression has *exactly one*
540 /// use by a non-leaf node of the expression. This destruction means that the
541 /// caller MUST either replace 'I' with a new expression or use something like
542 /// RewriteExprTree to put the values back in if the routine indicates that it
543 /// made a change by returning 'true'.
545 /// In the above example either the right operand of A or the left operand of B
546 /// will be replaced by undef. If it is B's operand then this gives:
550 /// + + | A, B - operand of B replaced with undef
556 /// Note that such undef operands can only be reached by passing through 'I'.
557 /// For example, if you visit operands recursively starting from a leaf node
558 /// then you will never see such an undef operand unless you get back to 'I',
559 /// which requires passing through a phi node.
561 /// Note that this routine may also mutate binary operators of the wrong type
562 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
563 /// of the expression) if it can turn them into binary operators of the right
564 /// type and thus make the expression bigger.
566 static bool LinearizeExprTree(BinaryOperator *I,
567 SmallVectorImpl<RepeatedValue> &Ops) {
568 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
569 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
570 unsigned Opcode = I->getOpcode();
571 assert(I->isAssociative() && I->isCommutative() &&
572 "Expected an associative and commutative operation!");
574 // Visit all operands of the expression, keeping track of their weight (the
575 // number of paths from the expression root to the operand, or if you like
576 // the number of times that operand occurs in the linearized expression).
577 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
578 // while A has weight two.
580 // Worklist of non-leaf nodes (their operands are in the expression too) along
581 // with their weights, representing a certain number of paths to the operator.
582 // If an operator occurs in the worklist multiple times then we found multiple
583 // ways to get to it.
584 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
585 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
586 bool Changed = false;
588 // Leaves of the expression are values that either aren't the right kind of
589 // operation (eg: a constant, or a multiply in an add tree), or are, but have
590 // some uses that are not inside the expression. For example, in I = X + X,
591 // X = A + B, the value X has two uses (by I) that are in the expression. If
592 // X has any other uses, for example in a return instruction, then we consider
593 // X to be a leaf, and won't analyze it further. When we first visit a value,
594 // if it has more than one use then at first we conservatively consider it to
595 // be a leaf. Later, as the expression is explored, we may discover some more
596 // uses of the value from inside the expression. If all uses turn out to be
597 // from within the expression (and the value is a binary operator of the right
598 // kind) then the value is no longer considered to be a leaf, and its operands
601 // Leaves - Keeps track of the set of putative leaves as well as the number of
602 // paths to each leaf seen so far.
603 typedef DenseMap<Value*, APInt> LeafMap;
604 LeafMap Leaves; // Leaf -> Total weight so far.
605 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
608 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
610 while (!Worklist.empty()) {
611 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
612 I = P.first; // We examine the operands of this binary operator.
614 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
615 Value *Op = I->getOperand(OpIdx);
616 APInt Weight = P.second; // Number of paths to this operand.
617 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
618 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
620 // If this is a binary operation of the right kind with only one use then
621 // add its operands to the expression.
622 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
623 assert(Visited.insert(Op).second && "Not first visit!");
624 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
625 Worklist.push_back(std::make_pair(BO, Weight));
629 // Appears to be a leaf. Is the operand already in the set of leaves?
630 LeafMap::iterator It = Leaves.find(Op);
631 if (It == Leaves.end()) {
632 // Not in the leaf map. Must be the first time we saw this operand.
633 assert(Visited.insert(Op).second && "Not first visit!");
634 if (!Op->hasOneUse()) {
635 // This value has uses not accounted for by the expression, so it is
636 // not safe to modify. Mark it as being a leaf.
637 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
638 LeafOrder.push_back(Op);
642 // No uses outside the expression, try morphing it.
643 } else if (It != Leaves.end()) {
644 // Already in the leaf map.
645 assert(Visited.count(Op) && "In leaf map but not visited!");
647 // Update the number of paths to the leaf.
648 IncorporateWeight(It->second, Weight, Opcode);
650 #if 0 // TODO: Re-enable once PR13021 is fixed.
651 // The leaf already has one use from inside the expression. As we want
652 // exactly one such use, drop this new use of the leaf.
653 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
654 I->setOperand(OpIdx, UndefValue::get(I->getType()));
657 // If the leaf is a binary operation of the right kind and we now see
658 // that its multiple original uses were in fact all by nodes belonging
659 // to the expression, then no longer consider it to be a leaf and add
660 // its operands to the expression.
661 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
662 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
663 Worklist.push_back(std::make_pair(BO, It->second));
669 // If we still have uses that are not accounted for by the expression
670 // then it is not safe to modify the value.
671 if (!Op->hasOneUse())
674 // No uses outside the expression, try morphing it.
676 Leaves.erase(It); // Since the value may be morphed below.
679 // At this point we have a value which, first of all, is not a binary
680 // expression of the right kind, and secondly, is only used inside the
681 // expression. This means that it can safely be modified. See if we
682 // can usefully morph it into an expression of the right kind.
683 assert((!isa<Instruction>(Op) ||
684 cast<Instruction>(Op)->getOpcode() != Opcode
685 || (isa<FPMathOperator>(Op) &&
686 !cast<Instruction>(Op)->hasUnsafeAlgebra())) &&
687 "Should have been handled above!");
688 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
690 // If this is a multiply expression, turn any internal negations into
691 // multiplies by -1 so they can be reassociated.
692 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
693 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
694 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
695 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
696 BO = LowerNegateToMultiply(BO);
697 DEBUG(dbgs() << *BO << '\n');
698 Worklist.push_back(std::make_pair(BO, Weight));
703 // Failed to morph into an expression of the right type. This really is
705 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
706 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
707 LeafOrder.push_back(Op);
712 // The leaves, repeated according to their weights, represent the linearized
713 // form of the expression.
714 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
715 Value *V = LeafOrder[i];
716 LeafMap::iterator It = Leaves.find(V);
717 if (It == Leaves.end())
718 // Node initially thought to be a leaf wasn't.
720 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
721 APInt Weight = It->second;
722 if (Weight.isMinValue())
723 // Leaf already output or weight reduction eliminated it.
725 // Ensure the leaf is only output once.
727 Ops.push_back(std::make_pair(V, Weight));
730 // For nilpotent operations or addition there may be no operands, for example
731 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
732 // in both cases the weight reduces to 0 causing the value to be skipped.
734 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
735 assert(Identity && "Associative operation without identity!");
736 Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1)));
742 /// Now that the operands for this expression tree are
743 /// linearized and optimized, emit them in-order.
744 void Reassociate::RewriteExprTree(BinaryOperator *I,
745 SmallVectorImpl<ValueEntry> &Ops) {
746 assert(Ops.size() > 1 && "Single values should be used directly!");
748 // Since our optimizations should never increase the number of operations, the
749 // new expression can usually be written reusing the existing binary operators
750 // from the original expression tree, without creating any new instructions,
751 // though the rewritten expression may have a completely different topology.
752 // We take care to not change anything if the new expression will be the same
753 // as the original. If more than trivial changes (like commuting operands)
754 // were made then we are obliged to clear out any optional subclass data like
757 /// NodesToRewrite - Nodes from the original expression available for writing
758 /// the new expression into.
759 SmallVector<BinaryOperator*, 8> NodesToRewrite;
760 unsigned Opcode = I->getOpcode();
761 BinaryOperator *Op = I;
763 /// NotRewritable - The operands being written will be the leaves of the new
764 /// expression and must not be used as inner nodes (via NodesToRewrite) by
765 /// mistake. Inner nodes are always reassociable, and usually leaves are not
766 /// (if they were they would have been incorporated into the expression and so
767 /// would not be leaves), so most of the time there is no danger of this. But
768 /// in rare cases a leaf may become reassociable if an optimization kills uses
769 /// of it, or it may momentarily become reassociable during rewriting (below)
770 /// due it being removed as an operand of one of its uses. Ensure that misuse
771 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
772 /// leaves and refusing to reuse any of them as inner nodes.
773 SmallPtrSet<Value*, 8> NotRewritable;
774 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
775 NotRewritable.insert(Ops[i].Op);
777 // ExpressionChanged - Non-null if the rewritten expression differs from the
778 // original in some non-trivial way, requiring the clearing of optional flags.
779 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
780 BinaryOperator *ExpressionChanged = nullptr;
781 for (unsigned i = 0; ; ++i) {
782 // The last operation (which comes earliest in the IR) is special as both
783 // operands will come from Ops, rather than just one with the other being
785 if (i+2 == Ops.size()) {
786 Value *NewLHS = Ops[i].Op;
787 Value *NewRHS = Ops[i+1].Op;
788 Value *OldLHS = Op->getOperand(0);
789 Value *OldRHS = Op->getOperand(1);
791 if (NewLHS == OldLHS && NewRHS == OldRHS)
792 // Nothing changed, leave it alone.
795 if (NewLHS == OldRHS && NewRHS == OldLHS) {
796 // The order of the operands was reversed. Swap them.
797 DEBUG(dbgs() << "RA: " << *Op << '\n');
799 DEBUG(dbgs() << "TO: " << *Op << '\n');
805 // The new operation differs non-trivially from the original. Overwrite
806 // the old operands with the new ones.
807 DEBUG(dbgs() << "RA: " << *Op << '\n');
808 if (NewLHS != OldLHS) {
809 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
810 if (BO && !NotRewritable.count(BO))
811 NodesToRewrite.push_back(BO);
812 Op->setOperand(0, NewLHS);
814 if (NewRHS != OldRHS) {
815 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
816 if (BO && !NotRewritable.count(BO))
817 NodesToRewrite.push_back(BO);
818 Op->setOperand(1, NewRHS);
820 DEBUG(dbgs() << "TO: " << *Op << '\n');
822 ExpressionChanged = Op;
829 // Not the last operation. The left-hand side will be a sub-expression
830 // while the right-hand side will be the current element of Ops.
831 Value *NewRHS = Ops[i].Op;
832 if (NewRHS != Op->getOperand(1)) {
833 DEBUG(dbgs() << "RA: " << *Op << '\n');
834 if (NewRHS == Op->getOperand(0)) {
835 // The new right-hand side was already present as the left operand. If
836 // we are lucky then swapping the operands will sort out both of them.
839 // Overwrite with the new right-hand side.
840 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
841 if (BO && !NotRewritable.count(BO))
842 NodesToRewrite.push_back(BO);
843 Op->setOperand(1, NewRHS);
844 ExpressionChanged = Op;
846 DEBUG(dbgs() << "TO: " << *Op << '\n');
851 // Now deal with the left-hand side. If this is already an operation node
852 // from the original expression then just rewrite the rest of the expression
854 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
855 if (BO && !NotRewritable.count(BO)) {
860 // Otherwise, grab a spare node from the original expression and use that as
861 // the left-hand side. If there are no nodes left then the optimizers made
862 // an expression with more nodes than the original! This usually means that
863 // they did something stupid but it might mean that the problem was just too
864 // hard (finding the mimimal number of multiplications needed to realize a
865 // multiplication expression is NP-complete). Whatever the reason, smart or
866 // stupid, create a new node if there are none left.
867 BinaryOperator *NewOp;
868 if (NodesToRewrite.empty()) {
869 Constant *Undef = UndefValue::get(I->getType());
870 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
871 Undef, Undef, "", I);
872 if (NewOp->getType()->isFPOrFPVectorTy())
873 NewOp->setFastMathFlags(I->getFastMathFlags());
875 NewOp = NodesToRewrite.pop_back_val();
878 DEBUG(dbgs() << "RA: " << *Op << '\n');
879 Op->setOperand(0, NewOp);
880 DEBUG(dbgs() << "TO: " << *Op << '\n');
881 ExpressionChanged = Op;
887 // If the expression changed non-trivially then clear out all subclass data
888 // starting from the operator specified in ExpressionChanged, and compactify
889 // the operators to just before the expression root to guarantee that the
890 // expression tree is dominated by all of Ops.
891 if (ExpressionChanged)
893 // Preserve FastMathFlags.
894 if (isa<FPMathOperator>(I)) {
895 FastMathFlags Flags = I->getFastMathFlags();
896 ExpressionChanged->clearSubclassOptionalData();
897 ExpressionChanged->setFastMathFlags(Flags);
899 ExpressionChanged->clearSubclassOptionalData();
901 if (ExpressionChanged == I)
903 ExpressionChanged->moveBefore(I);
904 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
907 // Throw away any left over nodes from the original expression.
908 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
909 RedoInsts.insert(NodesToRewrite[i]);
912 /// Insert instructions before the instruction pointed to by BI,
913 /// that computes the negative version of the value specified. The negative
914 /// version of the value is returned, and BI is left pointing at the instruction
915 /// that should be processed next by the reassociation pass.
916 static Value *NegateValue(Value *V, Instruction *BI) {
917 if (Constant *C = dyn_cast<Constant>(V)) {
918 if (C->getType()->isFPOrFPVectorTy()) {
919 return ConstantExpr::getFNeg(C);
921 return ConstantExpr::getNeg(C);
925 // We are trying to expose opportunity for reassociation. One of the things
926 // that we want to do to achieve this is to push a negation as deep into an
927 // expression chain as possible, to expose the add instructions. In practice,
928 // this means that we turn this:
929 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
930 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
931 // the constants. We assume that instcombine will clean up the mess later if
932 // we introduce tons of unnecessary negation instructions.
934 if (BinaryOperator *I =
935 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
936 // Push the negates through the add.
937 I->setOperand(0, NegateValue(I->getOperand(0), BI));
938 I->setOperand(1, NegateValue(I->getOperand(1), BI));
940 // We must move the add instruction here, because the neg instructions do
941 // not dominate the old add instruction in general. By moving it, we are
942 // assured that the neg instructions we just inserted dominate the
943 // instruction we are about to insert after them.
946 I->setName(I->getName()+".neg");
950 // Okay, we need to materialize a negated version of V with an instruction.
951 // Scan the use lists of V to see if we have one already.
952 for (User *U : V->users()) {
953 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
956 // We found one! Now we have to make sure that the definition dominates
957 // this use. We do this by moving it to the entry block (if it is a
958 // non-instruction value) or right after the definition. These negates will
959 // be zapped by reassociate later, so we don't need much finesse here.
960 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
962 // Verify that the negate is in this function, V might be a constant expr.
963 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
966 BasicBlock::iterator InsertPt;
967 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
968 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
969 InsertPt = II->getNormalDest()->begin();
971 InsertPt = InstInput;
974 while (isa<PHINode>(InsertPt)) ++InsertPt;
976 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
978 TheNeg->moveBefore(InsertPt);
982 // Insert a 'neg' instruction that subtracts the value from zero to get the
984 return CreateNeg(V, V->getName() + ".neg", BI, BI);
987 /// Return true if we should break up this subtract of X-Y into (X + -Y).
988 static bool ShouldBreakUpSubtract(Instruction *Sub) {
989 // If this is a negation, we can't split it up!
990 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
993 // Don't breakup X - undef.
994 if (isa<UndefValue>(Sub->getOperand(1)))
997 // Don't bother to break this up unless either the LHS is an associable add or
998 // subtract or if this is only used by one.
999 Value *V0 = Sub->getOperand(0);
1000 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
1001 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
1003 Value *V1 = Sub->getOperand(1);
1004 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
1005 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
1007 Value *VB = Sub->user_back();
1008 if (Sub->hasOneUse() &&
1009 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
1010 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
1016 /// If we have (X-Y), and if either X is an add, or if this is only used by an
1017 /// add, transform this into (X+(0-Y)) to promote better reassociation.
1018 static BinaryOperator *BreakUpSubtract(Instruction *Sub) {
1019 // Convert a subtract into an add and a neg instruction. This allows sub
1020 // instructions to be commuted with other add instructions.
1022 // Calculate the negative value of Operand 1 of the sub instruction,
1023 // and set it as the RHS of the add instruction we just made.
1025 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
1026 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
1027 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
1028 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
1031 // Everyone now refers to the add instruction.
1032 Sub->replaceAllUsesWith(New);
1033 New->setDebugLoc(Sub->getDebugLoc());
1035 DEBUG(dbgs() << "Negated: " << *New << '\n');
1039 /// If this is a shift of a reassociable multiply or is used by one, change
1040 /// this into a multiply by a constant to assist with further reassociation.
1041 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1042 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
1043 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
1045 BinaryOperator *Mul =
1046 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
1047 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
1050 // Everyone now refers to the mul instruction.
1051 Shl->replaceAllUsesWith(Mul);
1052 Mul->setDebugLoc(Shl->getDebugLoc());
1054 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
1055 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
1057 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
1058 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
1060 Mul->setHasNoSignedWrap(true);
1061 Mul->setHasNoUnsignedWrap(NUW);
1065 /// Scan backwards and forwards among values with the same rank as element i
1066 /// to see if X exists. If X does not exist, return i. This is useful when
1067 /// scanning for 'x' when we see '-x' because they both get the same rank.
1068 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
1070 unsigned XRank = Ops[i].Rank;
1071 unsigned e = Ops.size();
1072 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
1075 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1076 if (Instruction *I2 = dyn_cast<Instruction>(X))
1077 if (I1->isIdenticalTo(I2))
1081 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
1084 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1085 if (Instruction *I2 = dyn_cast<Instruction>(X))
1086 if (I1->isIdenticalTo(I2))
1092 /// Emit a tree of add instructions, summing Ops together
1093 /// and returning the result. Insert the tree before I.
1094 static Value *EmitAddTreeOfValues(Instruction *I,
1095 SmallVectorImpl<WeakVH> &Ops){
1096 if (Ops.size() == 1) return Ops.back();
1098 Value *V1 = Ops.back();
1100 Value *V2 = EmitAddTreeOfValues(I, Ops);
1101 return CreateAdd(V2, V1, "tmp", I, I);
1104 /// If V is an expression tree that is a multiplication sequence,
1105 /// and if this sequence contains a multiply by Factor,
1106 /// remove Factor from the tree and return the new tree.
1107 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
1108 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1112 SmallVector<RepeatedValue, 8> Tree;
1113 MadeChange |= LinearizeExprTree(BO, Tree);
1114 SmallVector<ValueEntry, 8> Factors;
1115 Factors.reserve(Tree.size());
1116 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1117 RepeatedValue E = Tree[i];
1118 Factors.append(E.second.getZExtValue(),
1119 ValueEntry(getRank(E.first), E.first));
1122 bool FoundFactor = false;
1123 bool NeedsNegate = false;
1124 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1125 if (Factors[i].Op == Factor) {
1127 Factors.erase(Factors.begin()+i);
1131 // If this is a negative version of this factor, remove it.
1132 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1133 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1134 if (FC1->getValue() == -FC2->getValue()) {
1135 FoundFactor = NeedsNegate = true;
1136 Factors.erase(Factors.begin()+i);
1139 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1140 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1141 APFloat F1(FC1->getValueAPF());
1142 APFloat F2(FC2->getValueAPF());
1144 if (F1.compare(F2) == APFloat::cmpEqual) {
1145 FoundFactor = NeedsNegate = true;
1146 Factors.erase(Factors.begin() + i);
1154 // Make sure to restore the operands to the expression tree.
1155 RewriteExprTree(BO, Factors);
1159 BasicBlock::iterator InsertPt = BO; ++InsertPt;
1161 // If this was just a single multiply, remove the multiply and return the only
1162 // remaining operand.
1163 if (Factors.size() == 1) {
1164 RedoInsts.insert(BO);
1167 RewriteExprTree(BO, Factors);
1172 V = CreateNeg(V, "neg", InsertPt, BO);
1177 /// If V is a single-use multiply, recursively add its operands as factors,
1178 /// otherwise add V to the list of factors.
1180 /// Ops is the top-level list of add operands we're trying to factor.
1181 static void FindSingleUseMultiplyFactors(Value *V,
1182 SmallVectorImpl<Value*> &Factors,
1183 const SmallVectorImpl<ValueEntry> &Ops) {
1184 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1186 Factors.push_back(V);
1190 // Otherwise, add the LHS and RHS to the list of factors.
1191 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
1192 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
1195 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1196 /// This optimizes based on identities. If it can be reduced to a single Value,
1197 /// it is returned, otherwise the Ops list is mutated as necessary.
1198 static Value *OptimizeAndOrXor(unsigned Opcode,
1199 SmallVectorImpl<ValueEntry> &Ops) {
1200 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1201 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1202 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1203 // First, check for X and ~X in the operand list.
1204 assert(i < Ops.size());
1205 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1206 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1207 unsigned FoundX = FindInOperandList(Ops, i, X);
1209 if (Opcode == Instruction::And) // ...&X&~X = 0
1210 return Constant::getNullValue(X->getType());
1212 if (Opcode == Instruction::Or) // ...|X|~X = -1
1213 return Constant::getAllOnesValue(X->getType());
1217 // Next, check for duplicate pairs of values, which we assume are next to
1218 // each other, due to our sorting criteria.
1219 assert(i < Ops.size());
1220 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1221 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1222 // Drop duplicate values for And and Or.
1223 Ops.erase(Ops.begin()+i);
1229 // Drop pairs of values for Xor.
1230 assert(Opcode == Instruction::Xor);
1232 return Constant::getNullValue(Ops[0].Op->getType());
1235 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1243 /// Helper funciton of CombineXorOpnd(). It creates a bitwise-and
1244 /// instruction with the given two operands, and return the resulting
1245 /// instruction. There are two special cases: 1) if the constant operand is 0,
1246 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1248 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1249 const APInt &ConstOpnd) {
1250 if (ConstOpnd != 0) {
1251 if (!ConstOpnd.isAllOnesValue()) {
1252 LLVMContext &Ctx = Opnd->getType()->getContext();
1254 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1255 "and.ra", InsertBefore);
1256 I->setDebugLoc(InsertBefore->getDebugLoc());
1264 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1265 // into "R ^ C", where C would be 0, and R is a symbolic value.
1267 // If it was successful, true is returned, and the "R" and "C" is returned
1268 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1269 // and both "Res" and "ConstOpnd" remain unchanged.
1271 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1272 APInt &ConstOpnd, Value *&Res) {
1273 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1274 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1275 // = (x & ~c1) ^ (c1 ^ c2)
1276 // It is useful only when c1 == c2.
1277 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) {
1278 if (!Opnd1->getValue()->hasOneUse())
1281 const APInt &C1 = Opnd1->getConstPart();
1282 if (C1 != ConstOpnd)
1285 Value *X = Opnd1->getSymbolicPart();
1286 Res = createAndInstr(I, X, ~C1);
1287 // ConstOpnd was C2, now C1 ^ C2.
1290 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1291 RedoInsts.insert(T);
1298 // Helper function of OptimizeXor(). It tries to simplify
1299 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1302 // If it was successful, true is returned, and the "R" and "C" is returned
1303 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1304 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1305 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1306 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
1307 APInt &ConstOpnd, Value *&Res) {
1308 Value *X = Opnd1->getSymbolicPart();
1309 if (X != Opnd2->getSymbolicPart())
1312 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1313 int DeadInstNum = 1;
1314 if (Opnd1->getValue()->hasOneUse())
1316 if (Opnd2->getValue()->hasOneUse())
1320 // (x | c1) ^ (x & c2)
1321 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1322 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1323 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1325 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1326 if (Opnd2->isOrExpr())
1327 std::swap(Opnd1, Opnd2);
1329 const APInt &C1 = Opnd1->getConstPart();
1330 const APInt &C2 = Opnd2->getConstPart();
1331 APInt C3((~C1) ^ C2);
1333 // Do not increase code size!
1334 if (C3 != 0 && !C3.isAllOnesValue()) {
1335 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1336 if (NewInstNum > DeadInstNum)
1340 Res = createAndInstr(I, X, C3);
1343 } else if (Opnd1->isOrExpr()) {
1344 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1346 const APInt &C1 = Opnd1->getConstPart();
1347 const APInt &C2 = Opnd2->getConstPart();
1350 // Do not increase code size
1351 if (C3 != 0 && !C3.isAllOnesValue()) {
1352 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1353 if (NewInstNum > DeadInstNum)
1357 Res = createAndInstr(I, X, C3);
1360 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1362 const APInt &C1 = Opnd1->getConstPart();
1363 const APInt &C2 = Opnd2->getConstPart();
1365 Res = createAndInstr(I, X, C3);
1368 // Put the original operands in the Redo list; hope they will be deleted
1370 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1371 RedoInsts.insert(T);
1372 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1373 RedoInsts.insert(T);
1378 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1379 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1381 Value *Reassociate::OptimizeXor(Instruction *I,
1382 SmallVectorImpl<ValueEntry> &Ops) {
1383 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1386 if (Ops.size() == 1)
1389 SmallVector<XorOpnd, 8> Opnds;
1390 SmallVector<XorOpnd*, 8> OpndPtrs;
1391 Type *Ty = Ops[0].Op->getType();
1392 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1394 // Step 1: Convert ValueEntry to XorOpnd
1395 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1396 Value *V = Ops[i].Op;
1397 if (!isa<ConstantInt>(V)) {
1399 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1402 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1405 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1406 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1407 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1408 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1409 // when new elements are added to the vector.
1410 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1411 OpndPtrs.push_back(&Opnds[i]);
1413 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1414 // the same symbolic value cluster together. For instance, the input operand
1415 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1416 // ("x | 123", "x & 789", "y & 456").
1417 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor());
1419 // Step 3: Combine adjacent operands
1420 XorOpnd *PrevOpnd = nullptr;
1421 bool Changed = false;
1422 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1423 XorOpnd *CurrOpnd = OpndPtrs[i];
1424 // The combined value
1427 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1428 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1431 *CurrOpnd = XorOpnd(CV);
1433 CurrOpnd->Invalidate();
1438 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1439 PrevOpnd = CurrOpnd;
1443 // step 3.2: When previous and current operands share the same symbolic
1444 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1446 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1447 // Remove previous operand
1448 PrevOpnd->Invalidate();
1450 *CurrOpnd = XorOpnd(CV);
1451 PrevOpnd = CurrOpnd;
1453 CurrOpnd->Invalidate();
1460 // Step 4: Reassemble the Ops
1463 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1464 XorOpnd &O = Opnds[i];
1467 ValueEntry VE(getRank(O.getValue()), O.getValue());
1470 if (ConstOpnd != 0) {
1471 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1472 ValueEntry VE(getRank(C), C);
1475 int Sz = Ops.size();
1477 return Ops.back().Op;
1479 assert(ConstOpnd == 0);
1480 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1487 /// Optimize a series of operands to an 'add' instruction. This
1488 /// optimizes based on identities. If it can be reduced to a single Value, it
1489 /// is returned, otherwise the Ops list is mutated as necessary.
1490 Value *Reassociate::OptimizeAdd(Instruction *I,
1491 SmallVectorImpl<ValueEntry> &Ops) {
1492 // Scan the operand lists looking for X and -X pairs. If we find any, we
1493 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1495 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1497 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1498 Value *TheOp = Ops[i].Op;
1499 // Check to see if we've seen this operand before. If so, we factor all
1500 // instances of the operand together. Due to our sorting criteria, we know
1501 // that these need to be next to each other in the vector.
1502 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1503 // Rescan the list, remove all instances of this operand from the expr.
1504 unsigned NumFound = 0;
1506 Ops.erase(Ops.begin()+i);
1508 } while (i != Ops.size() && Ops[i].Op == TheOp);
1510 DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1513 // Insert a new multiply.
1514 Type *Ty = TheOp->getType();
1515 Constant *C = Ty->isIntOrIntVectorTy() ?
1516 ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1517 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1519 // Now that we have inserted a multiply, optimize it. This allows us to
1520 // handle cases that require multiple factoring steps, such as this:
1521 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1522 RedoInsts.insert(Mul);
1524 // If every add operand was a duplicate, return the multiply.
1528 // Otherwise, we had some input that didn't have the dupe, such as
1529 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1530 // things being added by this operation.
1531 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1538 // Check for X and -X or X and ~X in the operand list.
1539 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1540 !BinaryOperator::isNot(TheOp))
1544 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1545 X = BinaryOperator::getNegArgument(TheOp);
1546 else if (BinaryOperator::isNot(TheOp))
1547 X = BinaryOperator::getNotArgument(TheOp);
1549 unsigned FoundX = FindInOperandList(Ops, i, X);
1553 // Remove X and -X from the operand list.
1554 if (Ops.size() == 2 &&
1555 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1556 return Constant::getNullValue(X->getType());
1558 // Remove X and ~X from the operand list.
1559 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1560 return Constant::getAllOnesValue(X->getType());
1562 Ops.erase(Ops.begin()+i);
1566 --i; // Need to back up an extra one.
1567 Ops.erase(Ops.begin()+FoundX);
1569 --i; // Revisit element.
1570 e -= 2; // Removed two elements.
1572 // if X and ~X we append -1 to the operand list.
1573 if (BinaryOperator::isNot(TheOp)) {
1574 Value *V = Constant::getAllOnesValue(X->getType());
1575 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1580 // Scan the operand list, checking to see if there are any common factors
1581 // between operands. Consider something like A*A+A*B*C+D. We would like to
1582 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1583 // To efficiently find this, we count the number of times a factor occurs
1584 // for any ADD operands that are MULs.
1585 DenseMap<Value*, unsigned> FactorOccurrences;
1587 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1588 // where they are actually the same multiply.
1589 unsigned MaxOcc = 0;
1590 Value *MaxOccVal = nullptr;
1591 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1592 BinaryOperator *BOp =
1593 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1597 // Compute all of the factors of this added value.
1598 SmallVector<Value*, 8> Factors;
1599 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1600 assert(Factors.size() > 1 && "Bad linearize!");
1602 // Add one to FactorOccurrences for each unique factor in this op.
1603 SmallPtrSet<Value*, 8> Duplicates;
1604 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1605 Value *Factor = Factors[i];
1606 if (!Duplicates.insert(Factor).second)
1609 unsigned Occ = ++FactorOccurrences[Factor];
1615 // If Factor is a negative constant, add the negated value as a factor
1616 // because we can percolate the negate out. Watch for minint, which
1617 // cannot be positivified.
1618 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1619 if (CI->isNegative() && !CI->isMinValue(true)) {
1620 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1621 assert(!Duplicates.count(Factor) &&
1622 "Shouldn't have two constant factors, missed a canonicalize");
1623 unsigned Occ = ++FactorOccurrences[Factor];
1629 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1630 if (CF->isNegative()) {
1631 APFloat F(CF->getValueAPF());
1633 Factor = ConstantFP::get(CF->getContext(), F);
1634 assert(!Duplicates.count(Factor) &&
1635 "Shouldn't have two constant factors, missed a canonicalize");
1636 unsigned Occ = ++FactorOccurrences[Factor];
1646 // If any factor occurred more than one time, we can pull it out.
1648 DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1651 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1652 // this, we could otherwise run into situations where removing a factor
1653 // from an expression will drop a use of maxocc, and this can cause
1654 // RemoveFactorFromExpression on successive values to behave differently.
1655 Instruction *DummyInst =
1656 I->getType()->isIntOrIntVectorTy()
1657 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1658 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1660 SmallVector<WeakVH, 4> NewMulOps;
1661 for (unsigned i = 0; i != Ops.size(); ++i) {
1662 // Only try to remove factors from expressions we're allowed to.
1663 BinaryOperator *BOp =
1664 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1668 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1669 // The factorized operand may occur several times. Convert them all in
1671 for (unsigned j = Ops.size(); j != i;) {
1673 if (Ops[j].Op == Ops[i].Op) {
1674 NewMulOps.push_back(V);
1675 Ops.erase(Ops.begin()+j);
1682 // No need for extra uses anymore.
1685 unsigned NumAddedValues = NewMulOps.size();
1686 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1688 // Now that we have inserted the add tree, optimize it. This allows us to
1689 // handle cases that require multiple factoring steps, such as this:
1690 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1691 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1692 (void)NumAddedValues;
1693 if (Instruction *VI = dyn_cast<Instruction>(V))
1694 RedoInsts.insert(VI);
1696 // Create the multiply.
1697 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1699 // Rerun associate on the multiply in case the inner expression turned into
1700 // a multiply. We want to make sure that we keep things in canonical form.
1701 RedoInsts.insert(V2);
1703 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1704 // entire result expression is just the multiply "A*(B+C)".
1708 // Otherwise, we had some input that didn't have the factor, such as
1709 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1710 // things being added by this operation.
1711 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1717 /// \brief Build up a vector of value/power pairs factoring a product.
1719 /// Given a series of multiplication operands, build a vector of factors and
1720 /// the powers each is raised to when forming the final product. Sort them in
1721 /// the order of descending power.
1723 /// (x*x) -> [(x, 2)]
1724 /// ((x*x)*x) -> [(x, 3)]
1725 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1727 /// \returns Whether any factors have a power greater than one.
1728 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1729 SmallVectorImpl<Factor> &Factors) {
1730 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1731 // Compute the sum of powers of simplifiable factors.
1732 unsigned FactorPowerSum = 0;
1733 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1734 Value *Op = Ops[Idx-1].Op;
1736 // Count the number of occurrences of this value.
1738 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1740 // Track for simplification all factors which occur 2 or more times.
1742 FactorPowerSum += Count;
1745 // We can only simplify factors if the sum of the powers of our simplifiable
1746 // factors is 4 or higher. When that is the case, we will *always* have
1747 // a simplification. This is an important invariant to prevent cyclicly
1748 // trying to simplify already minimal formations.
1749 if (FactorPowerSum < 4)
1752 // Now gather the simplifiable factors, removing them from Ops.
1754 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1755 Value *Op = Ops[Idx-1].Op;
1757 // Count the number of occurrences of this value.
1759 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1763 // Move an even number of occurrences to Factors.
1766 FactorPowerSum += Count;
1767 Factors.push_back(Factor(Op, Count));
1768 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1771 // None of the adjustments above should have reduced the sum of factor powers
1772 // below our mininum of '4'.
1773 assert(FactorPowerSum >= 4);
1775 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1779 /// \brief Build a tree of multiplies, computing the product of Ops.
1780 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1781 SmallVectorImpl<Value*> &Ops) {
1782 if (Ops.size() == 1)
1785 Value *LHS = Ops.pop_back_val();
1787 if (LHS->getType()->isIntOrIntVectorTy())
1788 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1790 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1791 } while (!Ops.empty());
1796 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1798 /// Given a vector of values raised to various powers, where no two values are
1799 /// equal and the powers are sorted in decreasing order, compute the minimal
1800 /// DAG of multiplies to compute the final product, and return that product
1802 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1803 SmallVectorImpl<Factor> &Factors) {
1804 assert(Factors[0].Power);
1805 SmallVector<Value *, 4> OuterProduct;
1806 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1807 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1808 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1813 // We want to multiply across all the factors with the same power so that
1814 // we can raise them to that power as a single entity. Build a mini tree
1816 SmallVector<Value *, 4> InnerProduct;
1817 InnerProduct.push_back(Factors[LastIdx].Base);
1819 InnerProduct.push_back(Factors[Idx].Base);
1821 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1823 // Reset the base value of the first factor to the new expression tree.
1824 // We'll remove all the factors with the same power in a second pass.
1825 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1826 if (Instruction *MI = dyn_cast<Instruction>(M))
1827 RedoInsts.insert(MI);
1831 // Unique factors with equal powers -- we've folded them into the first one's
1833 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1834 Factor::PowerEqual()),
1837 // Iteratively collect the base of each factor with an add power into the
1838 // outer product, and halve each power in preparation for squaring the
1840 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1841 if (Factors[Idx].Power & 1)
1842 OuterProduct.push_back(Factors[Idx].Base);
1843 Factors[Idx].Power >>= 1;
1845 if (Factors[0].Power) {
1846 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1847 OuterProduct.push_back(SquareRoot);
1848 OuterProduct.push_back(SquareRoot);
1850 if (OuterProduct.size() == 1)
1851 return OuterProduct.front();
1853 Value *V = buildMultiplyTree(Builder, OuterProduct);
1857 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1858 SmallVectorImpl<ValueEntry> &Ops) {
1859 // We can only optimize the multiplies when there is a chain of more than
1860 // three, such that a balanced tree might require fewer total multiplies.
1864 // Try to turn linear trees of multiplies without other uses of the
1865 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1867 SmallVector<Factor, 4> Factors;
1868 if (!collectMultiplyFactors(Ops, Factors))
1869 return nullptr; // All distinct factors, so nothing left for us to do.
1871 IRBuilder<> Builder(I);
1872 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1876 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1877 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1881 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1882 SmallVectorImpl<ValueEntry> &Ops) {
1883 // Now that we have the linearized expression tree, try to optimize it.
1884 // Start by folding any constants that we found.
1885 Constant *Cst = nullptr;
1886 unsigned Opcode = I->getOpcode();
1887 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1888 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1889 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1891 // If there was nothing but constants then we are done.
1895 // Put the combined constant back at the end of the operand list, except if
1896 // there is no point. For example, an add of 0 gets dropped here, while a
1897 // multiplication by zero turns the whole expression into zero.
1898 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1899 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1901 Ops.push_back(ValueEntry(0, Cst));
1904 if (Ops.size() == 1) return Ops[0].Op;
1906 // Handle destructive annihilation due to identities between elements in the
1907 // argument list here.
1908 unsigned NumOps = Ops.size();
1911 case Instruction::And:
1912 case Instruction::Or:
1913 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1917 case Instruction::Xor:
1918 if (Value *Result = OptimizeXor(I, Ops))
1922 case Instruction::Add:
1923 case Instruction::FAdd:
1924 if (Value *Result = OptimizeAdd(I, Ops))
1928 case Instruction::Mul:
1929 case Instruction::FMul:
1930 if (Value *Result = OptimizeMul(I, Ops))
1935 if (Ops.size() != NumOps)
1936 return OptimizeExpression(I, Ops);
1940 /// Zap the given instruction, adding interesting operands to the work list.
1941 void Reassociate::EraseInst(Instruction *I) {
1942 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1943 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1944 // Erase the dead instruction.
1945 ValueRankMap.erase(I);
1946 RedoInsts.remove(I);
1947 I->eraseFromParent();
1948 // Optimize its operands.
1949 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1950 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1951 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1952 // If this is a node in an expression tree, climb to the expression root
1953 // and add that since that's where optimization actually happens.
1954 unsigned Opcode = Op->getOpcode();
1955 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1956 Visited.insert(Op).second)
1957 Op = Op->user_back();
1958 RedoInsts.insert(Op);
1962 // Canonicalize expressions of the following form:
1963 // x + (-Constant * y) -> x - (Constant * y)
1964 // x - (-Constant * y) -> x + (Constant * y)
1965 Instruction *Reassociate::canonicalizeNegConstExpr(Instruction *I) {
1966 if (!I->hasOneUse() || I->getType()->isVectorTy())
1969 // Must be a mul, fmul, or fdiv instruction.
1970 unsigned Opcode = I->getOpcode();
1971 if (Opcode != Instruction::Mul && Opcode != Instruction::FMul &&
1972 Opcode != Instruction::FDiv)
1975 // Must have at least one constant operand.
1976 Constant *C0 = dyn_cast<Constant>(I->getOperand(0));
1977 Constant *C1 = dyn_cast<Constant>(I->getOperand(1));
1981 // Must be a negative ConstantInt or ConstantFP.
1982 Constant *C = C0 ? C0 : C1;
1983 unsigned ConstIdx = C0 ? 0 : 1;
1984 if (auto *CI = dyn_cast<ConstantInt>(C)) {
1985 if (!CI->isNegative() || CI->isMinValue(true))
1987 } else if (auto *CF = dyn_cast<ConstantFP>(C)) {
1988 if (!CF->isNegative())
1993 // User must be a binary operator with one or more uses.
1994 Instruction *User = I->user_back();
1995 if (!isa<BinaryOperator>(User) || !User->getNumUses())
1998 unsigned UserOpcode = User->getOpcode();
1999 if (UserOpcode != Instruction::Add && UserOpcode != Instruction::FAdd &&
2000 UserOpcode != Instruction::Sub && UserOpcode != Instruction::FSub)
2003 // Subtraction is not commutative. Explicitly, the following transform is
2004 // not valid: (-Constant * y) - x -> x + (Constant * y)
2005 if (!User->isCommutative() && User->getOperand(1) != I)
2008 // Change the sign of the constant.
2009 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
2010 I->setOperand(ConstIdx, ConstantInt::get(CI->getContext(), -CI->getValue()));
2012 ConstantFP *CF = cast<ConstantFP>(C);
2013 APFloat Val = CF->getValueAPF();
2015 I->setOperand(ConstIdx, ConstantFP::get(CF->getContext(), Val));
2018 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
2019 // ((-Const*y) + x) -> (x + (-Const*y)).
2020 if (User->getOperand(0) == I && User->isCommutative())
2021 cast<BinaryOperator>(User)->swapOperands();
2023 Value *Op0 = User->getOperand(0);
2024 Value *Op1 = User->getOperand(1);
2026 switch(UserOpcode) {
2028 llvm_unreachable("Unexpected Opcode!");
2029 case Instruction::Add:
2030 NI = BinaryOperator::CreateSub(Op0, Op1);
2032 case Instruction::Sub:
2033 NI = BinaryOperator::CreateAdd(Op0, Op1);
2035 case Instruction::FAdd:
2036 NI = BinaryOperator::CreateFSub(Op0, Op1);
2037 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2039 case Instruction::FSub:
2040 NI = BinaryOperator::CreateFAdd(Op0, Op1);
2041 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2045 NI->insertBefore(User);
2046 NI->setName(User->getName());
2047 User->replaceAllUsesWith(NI);
2048 NI->setDebugLoc(I->getDebugLoc());
2049 RedoInsts.insert(I);
2054 /// Inspect and optimize the given instruction. Note that erasing
2055 /// instructions is not allowed.
2056 void Reassociate::OptimizeInst(Instruction *I) {
2057 // Only consider operations that we understand.
2058 if (!isa<BinaryOperator>(I))
2061 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
2062 // If an operand of this shift is a reassociable multiply, or if the shift
2063 // is used by a reassociable multiply or add, turn into a multiply.
2064 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
2066 (isReassociableOp(I->user_back(), Instruction::Mul) ||
2067 isReassociableOp(I->user_back(), Instruction::Add)))) {
2068 Instruction *NI = ConvertShiftToMul(I);
2069 RedoInsts.insert(I);
2074 // Canonicalize negative constants out of expressions.
2075 if (Instruction *Res = canonicalizeNegConstExpr(I))
2078 // Commute binary operators, to canonicalize the order of their operands.
2079 // This can potentially expose more CSE opportunities, and makes writing other
2080 // transformations simpler.
2081 if (I->isCommutative())
2082 canonicalizeOperands(I);
2084 // TODO: We should optimize vector Xor instructions, but they are
2085 // currently unsupported.
2086 if (I->getType()->isVectorTy() && I->getOpcode() == Instruction::Xor)
2089 // Don't optimize floating point instructions that don't have unsafe algebra.
2090 if (I->getType()->isFloatingPointTy() && !I->hasUnsafeAlgebra())
2093 // Do not reassociate boolean (i1) expressions. We want to preserve the
2094 // original order of evaluation for short-circuited comparisons that
2095 // SimplifyCFG has folded to AND/OR expressions. If the expression
2096 // is not further optimized, it is likely to be transformed back to a
2097 // short-circuited form for code gen, and the source order may have been
2098 // optimized for the most likely conditions.
2099 if (I->getType()->isIntegerTy(1))
2102 // If this is a subtract instruction which is not already in negate form,
2103 // see if we can convert it to X+-Y.
2104 if (I->getOpcode() == Instruction::Sub) {
2105 if (ShouldBreakUpSubtract(I)) {
2106 Instruction *NI = BreakUpSubtract(I);
2107 RedoInsts.insert(I);
2110 } else if (BinaryOperator::isNeg(I)) {
2111 // Otherwise, this is a negation. See if the operand is a multiply tree
2112 // and if this is not an inner node of a multiply tree.
2113 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2115 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2116 Instruction *NI = LowerNegateToMultiply(I);
2117 RedoInsts.insert(I);
2122 } else if (I->getOpcode() == Instruction::FSub) {
2123 if (ShouldBreakUpSubtract(I)) {
2124 Instruction *NI = BreakUpSubtract(I);
2125 RedoInsts.insert(I);
2128 } else if (BinaryOperator::isFNeg(I)) {
2129 // Otherwise, this is a negation. See if the operand is a multiply tree
2130 // and if this is not an inner node of a multiply tree.
2131 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2133 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2134 Instruction *NI = LowerNegateToMultiply(I);
2135 RedoInsts.insert(I);
2142 // If this instruction is an associative binary operator, process it.
2143 if (!I->isAssociative()) return;
2144 BinaryOperator *BO = cast<BinaryOperator>(I);
2146 // If this is an interior node of a reassociable tree, ignore it until we
2147 // get to the root of the tree, to avoid N^2 analysis.
2148 unsigned Opcode = BO->getOpcode();
2149 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode)
2152 // If this is an add tree that is used by a sub instruction, ignore it
2153 // until we process the subtract.
2154 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2155 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2157 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2158 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2161 ReassociateExpression(BO);
2164 void Reassociate::ReassociateExpression(BinaryOperator *I) {
2165 // First, walk the expression tree, linearizing the tree, collecting the
2166 // operand information.
2167 SmallVector<RepeatedValue, 8> Tree;
2168 MadeChange |= LinearizeExprTree(I, Tree);
2169 SmallVector<ValueEntry, 8> Ops;
2170 Ops.reserve(Tree.size());
2171 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2172 RepeatedValue E = Tree[i];
2173 Ops.append(E.second.getZExtValue(),
2174 ValueEntry(getRank(E.first), E.first));
2177 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2179 // Now that we have linearized the tree to a list and have gathered all of
2180 // the operands and their ranks, sort the operands by their rank. Use a
2181 // stable_sort so that values with equal ranks will have their relative
2182 // positions maintained (and so the compiler is deterministic). Note that
2183 // this sorts so that the highest ranking values end up at the beginning of
2185 std::stable_sort(Ops.begin(), Ops.end());
2187 // Now that we have the expression tree in a convenient
2188 // sorted form, optimize it globally if possible.
2189 if (Value *V = OptimizeExpression(I, Ops)) {
2191 // Self-referential expression in unreachable code.
2193 // This expression tree simplified to something that isn't a tree,
2195 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2196 I->replaceAllUsesWith(V);
2197 if (Instruction *VI = dyn_cast<Instruction>(V))
2198 VI->setDebugLoc(I->getDebugLoc());
2199 RedoInsts.insert(I);
2204 // We want to sink immediates as deeply as possible except in the case where
2205 // this is a multiply tree used only by an add, and the immediate is a -1.
2206 // In this case we reassociate to put the negation on the outside so that we
2207 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2208 if (I->hasOneUse()) {
2209 if (I->getOpcode() == Instruction::Mul &&
2210 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2211 isa<ConstantInt>(Ops.back().Op) &&
2212 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
2213 ValueEntry Tmp = Ops.pop_back_val();
2214 Ops.insert(Ops.begin(), Tmp);
2215 } else if (I->getOpcode() == Instruction::FMul &&
2216 cast<Instruction>(I->user_back())->getOpcode() ==
2217 Instruction::FAdd &&
2218 isa<ConstantFP>(Ops.back().Op) &&
2219 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2220 ValueEntry Tmp = Ops.pop_back_val();
2221 Ops.insert(Ops.begin(), Tmp);
2225 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2227 if (Ops.size() == 1) {
2229 // Self-referential expression in unreachable code.
2232 // This expression tree simplified to something that isn't a tree,
2234 I->replaceAllUsesWith(Ops[0].Op);
2235 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2236 OI->setDebugLoc(I->getDebugLoc());
2237 RedoInsts.insert(I);
2241 // Now that we ordered and optimized the expressions, splat them back into
2242 // the expression tree, removing any unneeded nodes.
2243 RewriteExprTree(I, Ops);
2246 bool Reassociate::runOnFunction(Function &F) {
2247 if (skipOptnoneFunction(F))
2250 // Calculate the rank map for F
2254 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
2255 // Optimize every instruction in the basic block.
2256 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; )
2257 if (isInstructionTriviallyDead(II)) {
2261 assert(II->getParent() == BI && "Moved to a different block!");
2265 // If this produced extra instructions to optimize, handle them now.
2266 while (!RedoInsts.empty()) {
2267 Instruction *I = RedoInsts.pop_back_val();
2268 if (isInstructionTriviallyDead(I))
2275 // We are done with the rank map.
2277 ValueRankMap.clear();