1 //===- Reassociate.cpp - Reassociate binary expressions -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass reassociates commutative expressions in an order that is designed
11 // to promote better constant propagation, GCSE, LICM, PRE, etc.
13 // For example: 4 + (x + 5) -> x + (4 + 5)
15 // In the implementation of this algorithm, constants are assigned rank = 0,
16 // function arguments are rank = 1, and other values are assigned ranks
17 // corresponding to the reverse post order traversal of current function
18 // (starting at 2), which effectively gives values in deep loops higher rank
19 // than values not in loops.
21 //===----------------------------------------------------------------------===//
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/ADT/DenseMap.h"
25 #include "llvm/ADT/PostOrderIterator.h"
26 #include "llvm/ADT/STLExtras.h"
27 #include "llvm/ADT/SetVector.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/IR/CFG.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/ValueHandle.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Transforms/Utils/Local.h"
44 #define DEBUG_TYPE "reassociate"
46 STATISTIC(NumChanged, "Number of insts reassociated");
47 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
48 STATISTIC(NumFactor , "Number of multiplies factored");
54 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {}
56 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) {
57 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start.
62 /// PrintOps - Print out the expression identified in the Ops list.
64 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
65 Module *M = I->getParent()->getParent()->getParent();
66 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
67 << *Ops[0].Op->getType() << '\t';
68 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
70 Ops[i].Op->printAsOperand(dbgs(), false, M);
71 dbgs() << ", #" << Ops[i].Rank << "] ";
77 /// \brief Utility class representing a base and exponent pair which form one
78 /// factor of some product.
83 Factor(Value *Base, unsigned Power) : Base(Base), Power(Power) {}
85 /// \brief Sort factors by their Base.
87 bool operator()(const Factor &LHS, const Factor &RHS) {
88 return LHS.Base < RHS.Base;
92 /// \brief Compare factors for equal bases.
94 bool operator()(const Factor &LHS, const Factor &RHS) {
95 return LHS.Base == RHS.Base;
99 /// \brief Sort factors in descending order by their power.
100 struct PowerDescendingSorter {
101 bool operator()(const Factor &LHS, const Factor &RHS) {
102 return LHS.Power > RHS.Power;
106 /// \brief Compare factors for equal powers.
108 bool operator()(const Factor &LHS, const Factor &RHS) {
109 return LHS.Power == RHS.Power;
114 /// Utility class representing a non-constant Xor-operand. We classify
115 /// non-constant Xor-Operands into two categories:
116 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0
118 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
120 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
121 /// operand as "E | 0"
126 bool isInvalid() const { return SymbolicPart == nullptr; }
127 bool isOrExpr() const { return isOr; }
128 Value *getValue() const { return OrigVal; }
129 Value *getSymbolicPart() const { return SymbolicPart; }
130 unsigned getSymbolicRank() const { return SymbolicRank; }
131 const APInt &getConstPart() const { return ConstPart; }
133 void Invalidate() { SymbolicPart = OrigVal = nullptr; }
134 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
136 // Sort the XorOpnd-Pointer in ascending order of symbolic-value-rank.
137 // The purpose is twofold:
138 // 1) Cluster together the operands sharing the same symbolic-value.
139 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
140 // could potentially shorten crital path, and expose more loop-invariants.
141 // Note that values' rank are basically defined in RPO order (FIXME).
142 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
143 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
144 // "z" in the order of X-Y-Z is better than any other orders.
145 struct PtrSortFunctor {
146 bool operator()(XorOpnd * const &LHS, XorOpnd * const &RHS) {
147 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
154 unsigned SymbolicRank;
160 class Reassociate : public FunctionPass {
161 DenseMap<BasicBlock*, unsigned> RankMap;
162 DenseMap<AssertingVH<Value>, unsigned> ValueRankMap;
163 SetVector<AssertingVH<Instruction> > RedoInsts;
166 static char ID; // Pass identification, replacement for typeid
167 Reassociate() : FunctionPass(ID) {
168 initializeReassociatePass(*PassRegistry::getPassRegistry());
171 bool runOnFunction(Function &F) override;
173 void getAnalysisUsage(AnalysisUsage &AU) const override {
174 AU.setPreservesCFG();
177 void BuildRankMap(Function &F);
178 unsigned getRank(Value *V);
179 void 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 /// isReassociableOp - Return true if V is an instruction of the specified
237 /// opcode and if it only has one use.
238 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
239 if (V->hasOneUse() && isa<Instruction>(V) &&
240 cast<Instruction>(V)->getOpcode() == Opcode &&
241 (!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 Type *Ty = V->getType();
325 if ((!Ty->isIntegerTy() && !Ty->isFloatingPointTy()) ||
326 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
327 !BinaryOperator::isFNeg(I)))
330 DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
332 return ValueRankMap[I] = Rank;
335 // Canonicalize constants to RHS. Otherwise, sort the operands by rank.
336 void Reassociate::canonicalizeOperands(Instruction *I) {
337 assert(isa<BinaryOperator>(I) && "Expected binary operator.");
338 assert(I->isCommutative() && "Expected commutative operator.");
340 Value *LHS = I->getOperand(0);
341 Value *RHS = I->getOperand(1);
342 unsigned LHSRank = getRank(LHS);
343 unsigned RHSRank = getRank(RHS);
345 if (isa<Constant>(RHS))
348 if (isa<Constant>(LHS) || RHSRank < LHSRank)
349 cast<BinaryOperator>(I)->swapOperands();
352 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
353 Instruction *InsertBefore, Value *FlagsOp) {
354 if (S1->getType()->isIntegerTy())
355 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
357 BinaryOperator *Res =
358 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
359 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
364 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
365 Instruction *InsertBefore, Value *FlagsOp) {
366 if (S1->getType()->isIntegerTy())
367 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
369 BinaryOperator *Res =
370 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
371 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
376 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
377 Instruction *InsertBefore, Value *FlagsOp) {
378 if (S1->getType()->isIntegerTy())
379 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
381 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
382 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
387 /// LowerNegateToMultiply - Replace 0-X with X*-1.
389 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
390 Type *Ty = Neg->getType();
391 Constant *NegOne = Ty->isIntegerTy() ? ConstantInt::getAllOnesValue(Ty)
392 : ConstantFP::get(Ty, -1.0);
394 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
395 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
397 if (Ty->isIntegerTy()) {
398 bool NSW = cast<BinaryOperator>(Neg)->hasNoSignedWrap();
399 bool NUW = cast<BinaryOperator>(Neg)->hasNoUnsignedWrap();
401 Res->setHasNoSignedWrap(true);
402 Res->setHasNoUnsignedWrap(NUW);
404 Neg->replaceAllUsesWith(Res);
405 Res->setDebugLoc(Neg->getDebugLoc());
409 /// CarmichaelShift - Returns k such that lambda(2^Bitwidth) = 2^k, where lambda
410 /// is the Carmichael function. This means that x^(2^k) === 1 mod 2^Bitwidth for
411 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
412 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
413 /// even x in Bitwidth-bit arithmetic.
414 static unsigned CarmichaelShift(unsigned Bitwidth) {
420 /// IncorporateWeight - Add the extra weight 'RHS' to the existing weight 'LHS',
421 /// reducing the combined weight using any special properties of the operation.
422 /// The existing weight LHS represents the computation X op X op ... op X where
423 /// X occurs LHS times. The combined weight represents X op X op ... op X with
424 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
425 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
426 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
427 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
428 // If we were working with infinite precision arithmetic then the combined
429 // weight would be LHS + RHS. But we are using finite precision arithmetic,
430 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
431 // for nilpotent operations and addition, but not for idempotent operations
432 // and multiplication), so it is important to correctly reduce the combined
433 // weight back into range if wrapping would be wrong.
435 // If RHS is zero then the weight didn't change.
436 if (RHS.isMinValue())
438 // If LHS is zero then the combined weight is RHS.
439 if (LHS.isMinValue()) {
443 // From this point on we know that neither LHS nor RHS is zero.
445 if (Instruction::isIdempotent(Opcode)) {
446 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
447 // weight of 1. Keeping weights at zero or one also means that wrapping is
449 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
450 return; // Return a weight of 1.
452 if (Instruction::isNilpotent(Opcode)) {
453 // Nilpotent means X op X === 0, so reduce weights modulo 2.
454 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
455 LHS = 0; // 1 + 1 === 0 modulo 2.
458 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
459 // TODO: Reduce the weight by exploiting nsw/nuw?
464 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
465 "Unknown associative operation!");
466 unsigned Bitwidth = LHS.getBitWidth();
467 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
468 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
469 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
470 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
471 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
472 // which by a happy accident means that they can always be represented using
474 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
475 // the Carmichael number).
477 /// CM - The value of Carmichael's lambda function.
478 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
479 // Any weight W >= Threshold can be replaced with W - CM.
480 APInt Threshold = CM + Bitwidth;
481 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
482 // For Bitwidth 4 or more the following sum does not overflow.
484 while (LHS.uge(Threshold))
487 // To avoid problems with overflow do everything the same as above but using
489 unsigned CM = 1U << CarmichaelShift(Bitwidth);
490 unsigned Threshold = CM + Bitwidth;
491 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
492 "Weights not reduced!");
493 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
494 while (Total >= Threshold)
500 typedef std::pair<Value*, APInt> RepeatedValue;
502 /// LinearizeExprTree - Given an associative binary expression, return the leaf
503 /// nodes in Ops along with their weights (how many times the leaf occurs). The
504 /// original expression is the same as
505 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
507 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
511 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
513 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
515 /// This routine may modify the function, in which case it returns 'true'. The
516 /// changes it makes may well be destructive, changing the value computed by 'I'
517 /// to something completely different. Thus if the routine returns 'true' then
518 /// you MUST either replace I with a new expression computed from the Ops array,
519 /// or use RewriteExprTree to put the values back in.
521 /// A leaf node is either not a binary operation of the same kind as the root
522 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
523 /// opcode), or is the same kind of binary operator but has a use which either
524 /// does not belong to the expression, or does belong to the expression but is
525 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
526 /// of the expression, while for non-leaf nodes (except for the root 'I') every
527 /// use is a non-leaf node of the expression.
530 /// expression graph node names
540 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
541 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
543 /// The expression is maximal: if some instruction is a binary operator of the
544 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
545 /// then the instruction also belongs to the expression, is not a leaf node of
546 /// it, and its operands also belong to the expression (but may be leaf nodes).
548 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
549 /// order to ensure that every non-root node in the expression has *exactly one*
550 /// use by a non-leaf node of the expression. This destruction means that the
551 /// caller MUST either replace 'I' with a new expression or use something like
552 /// RewriteExprTree to put the values back in if the routine indicates that it
553 /// made a change by returning 'true'.
555 /// In the above example either the right operand of A or the left operand of B
556 /// will be replaced by undef. If it is B's operand then this gives:
560 /// + + | A, B - operand of B replaced with undef
566 /// Note that such undef operands can only be reached by passing through 'I'.
567 /// For example, if you visit operands recursively starting from a leaf node
568 /// then you will never see such an undef operand unless you get back to 'I',
569 /// which requires passing through a phi node.
571 /// Note that this routine may also mutate binary operators of the wrong type
572 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
573 /// of the expression) if it can turn them into binary operators of the right
574 /// type and thus make the expression bigger.
576 static bool LinearizeExprTree(BinaryOperator *I,
577 SmallVectorImpl<RepeatedValue> &Ops) {
578 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
579 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
580 unsigned Opcode = I->getOpcode();
581 assert(I->isAssociative() && I->isCommutative() &&
582 "Expected an associative and commutative operation!");
584 // Visit all operands of the expression, keeping track of their weight (the
585 // number of paths from the expression root to the operand, or if you like
586 // the number of times that operand occurs in the linearized expression).
587 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
588 // while A has weight two.
590 // Worklist of non-leaf nodes (their operands are in the expression too) along
591 // with their weights, representing a certain number of paths to the operator.
592 // If an operator occurs in the worklist multiple times then we found multiple
593 // ways to get to it.
594 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
595 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
596 bool Changed = false;
598 // Leaves of the expression are values that either aren't the right kind of
599 // operation (eg: a constant, or a multiply in an add tree), or are, but have
600 // some uses that are not inside the expression. For example, in I = X + X,
601 // X = A + B, the value X has two uses (by I) that are in the expression. If
602 // X has any other uses, for example in a return instruction, then we consider
603 // X to be a leaf, and won't analyze it further. When we first visit a value,
604 // if it has more than one use then at first we conservatively consider it to
605 // be a leaf. Later, as the expression is explored, we may discover some more
606 // uses of the value from inside the expression. If all uses turn out to be
607 // from within the expression (and the value is a binary operator of the right
608 // kind) then the value is no longer considered to be a leaf, and its operands
611 // Leaves - Keeps track of the set of putative leaves as well as the number of
612 // paths to each leaf seen so far.
613 typedef DenseMap<Value*, APInt> LeafMap;
614 LeafMap Leaves; // Leaf -> Total weight so far.
615 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
618 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
620 while (!Worklist.empty()) {
621 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
622 I = P.first; // We examine the operands of this binary operator.
624 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
625 Value *Op = I->getOperand(OpIdx);
626 APInt Weight = P.second; // Number of paths to this operand.
627 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
628 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
630 // If this is a binary operation of the right kind with only one use then
631 // add its operands to the expression.
632 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
633 assert(Visited.insert(Op).second && "Not first visit!");
634 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
635 Worklist.push_back(std::make_pair(BO, Weight));
639 // Appears to be a leaf. Is the operand already in the set of leaves?
640 LeafMap::iterator It = Leaves.find(Op);
641 if (It == Leaves.end()) {
642 // Not in the leaf map. Must be the first time we saw this operand.
643 assert(Visited.insert(Op).second && "Not first visit!");
644 if (!Op->hasOneUse()) {
645 // This value has uses not accounted for by the expression, so it is
646 // not safe to modify. Mark it as being a leaf.
647 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
648 LeafOrder.push_back(Op);
652 // No uses outside the expression, try morphing it.
653 } else if (It != Leaves.end()) {
654 // Already in the leaf map.
655 assert(Visited.count(Op) && "In leaf map but not visited!");
657 // Update the number of paths to the leaf.
658 IncorporateWeight(It->second, Weight, Opcode);
660 #if 0 // TODO: Re-enable once PR13021 is fixed.
661 // The leaf already has one use from inside the expression. As we want
662 // exactly one such use, drop this new use of the leaf.
663 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
664 I->setOperand(OpIdx, UndefValue::get(I->getType()));
667 // If the leaf is a binary operation of the right kind and we now see
668 // that its multiple original uses were in fact all by nodes belonging
669 // to the expression, then no longer consider it to be a leaf and add
670 // its operands to the expression.
671 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
672 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
673 Worklist.push_back(std::make_pair(BO, It->second));
679 // If we still have uses that are not accounted for by the expression
680 // then it is not safe to modify the value.
681 if (!Op->hasOneUse())
684 // No uses outside the expression, try morphing it.
686 Leaves.erase(It); // Since the value may be morphed below.
689 // At this point we have a value which, first of all, is not a binary
690 // expression of the right kind, and secondly, is only used inside the
691 // expression. This means that it can safely be modified. See if we
692 // can usefully morph it into an expression of the right kind.
693 assert((!isa<Instruction>(Op) ||
694 cast<Instruction>(Op)->getOpcode() != Opcode
695 || (isa<FPMathOperator>(Op) &&
696 !cast<Instruction>(Op)->hasUnsafeAlgebra())) &&
697 "Should have been handled above!");
698 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
700 // If this is a multiply expression, turn any internal negations into
701 // multiplies by -1 so they can be reassociated.
702 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
703 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
704 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
705 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
706 BO = LowerNegateToMultiply(BO);
707 DEBUG(dbgs() << *BO << '\n');
708 Worklist.push_back(std::make_pair(BO, Weight));
713 // Failed to morph into an expression of the right type. This really is
715 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
716 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
717 LeafOrder.push_back(Op);
722 // The leaves, repeated according to their weights, represent the linearized
723 // form of the expression.
724 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
725 Value *V = LeafOrder[i];
726 LeafMap::iterator It = Leaves.find(V);
727 if (It == Leaves.end())
728 // Node initially thought to be a leaf wasn't.
730 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
731 APInt Weight = It->second;
732 if (Weight.isMinValue())
733 // Leaf already output or weight reduction eliminated it.
735 // Ensure the leaf is only output once.
737 Ops.push_back(std::make_pair(V, Weight));
740 // For nilpotent operations or addition there may be no operands, for example
741 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
742 // in both cases the weight reduces to 0 causing the value to be skipped.
744 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
745 assert(Identity && "Associative operation without identity!");
746 Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1)));
752 // RewriteExprTree - Now that the operands for this expression tree are
753 // linearized and optimized, emit them in-order.
754 void Reassociate::RewriteExprTree(BinaryOperator *I,
755 SmallVectorImpl<ValueEntry> &Ops) {
756 assert(Ops.size() > 1 && "Single values should be used directly!");
758 // Since our optimizations should never increase the number of operations, the
759 // new expression can usually be written reusing the existing binary operators
760 // from the original expression tree, without creating any new instructions,
761 // though the rewritten expression may have a completely different topology.
762 // We take care to not change anything if the new expression will be the same
763 // as the original. If more than trivial changes (like commuting operands)
764 // were made then we are obliged to clear out any optional subclass data like
767 /// NodesToRewrite - Nodes from the original expression available for writing
768 /// the new expression into.
769 SmallVector<BinaryOperator*, 8> NodesToRewrite;
770 unsigned Opcode = I->getOpcode();
771 BinaryOperator *Op = I;
773 /// NotRewritable - The operands being written will be the leaves of the new
774 /// expression and must not be used as inner nodes (via NodesToRewrite) by
775 /// mistake. Inner nodes are always reassociable, and usually leaves are not
776 /// (if they were they would have been incorporated into the expression and so
777 /// would not be leaves), so most of the time there is no danger of this. But
778 /// in rare cases a leaf may become reassociable if an optimization kills uses
779 /// of it, or it may momentarily become reassociable during rewriting (below)
780 /// due it being removed as an operand of one of its uses. Ensure that misuse
781 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
782 /// leaves and refusing to reuse any of them as inner nodes.
783 SmallPtrSet<Value*, 8> NotRewritable;
784 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
785 NotRewritable.insert(Ops[i].Op);
787 // ExpressionChanged - Non-null if the rewritten expression differs from the
788 // original in some non-trivial way, requiring the clearing of optional flags.
789 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
790 BinaryOperator *ExpressionChanged = nullptr;
791 for (unsigned i = 0; ; ++i) {
792 // The last operation (which comes earliest in the IR) is special as both
793 // operands will come from Ops, rather than just one with the other being
795 if (i+2 == Ops.size()) {
796 Value *NewLHS = Ops[i].Op;
797 Value *NewRHS = Ops[i+1].Op;
798 Value *OldLHS = Op->getOperand(0);
799 Value *OldRHS = Op->getOperand(1);
801 if (NewLHS == OldLHS && NewRHS == OldRHS)
802 // Nothing changed, leave it alone.
805 if (NewLHS == OldRHS && NewRHS == OldLHS) {
806 // The order of the operands was reversed. Swap them.
807 DEBUG(dbgs() << "RA: " << *Op << '\n');
809 DEBUG(dbgs() << "TO: " << *Op << '\n');
815 // The new operation differs non-trivially from the original. Overwrite
816 // the old operands with the new ones.
817 DEBUG(dbgs() << "RA: " << *Op << '\n');
818 if (NewLHS != OldLHS) {
819 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
820 if (BO && !NotRewritable.count(BO))
821 NodesToRewrite.push_back(BO);
822 Op->setOperand(0, NewLHS);
824 if (NewRHS != OldRHS) {
825 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
826 if (BO && !NotRewritable.count(BO))
827 NodesToRewrite.push_back(BO);
828 Op->setOperand(1, NewRHS);
830 DEBUG(dbgs() << "TO: " << *Op << '\n');
832 ExpressionChanged = Op;
839 // Not the last operation. The left-hand side will be a sub-expression
840 // while the right-hand side will be the current element of Ops.
841 Value *NewRHS = Ops[i].Op;
842 if (NewRHS != Op->getOperand(1)) {
843 DEBUG(dbgs() << "RA: " << *Op << '\n');
844 if (NewRHS == Op->getOperand(0)) {
845 // The new right-hand side was already present as the left operand. If
846 // we are lucky then swapping the operands will sort out both of them.
849 // Overwrite with the new right-hand side.
850 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
851 if (BO && !NotRewritable.count(BO))
852 NodesToRewrite.push_back(BO);
853 Op->setOperand(1, NewRHS);
854 ExpressionChanged = Op;
856 DEBUG(dbgs() << "TO: " << *Op << '\n');
861 // Now deal with the left-hand side. If this is already an operation node
862 // from the original expression then just rewrite the rest of the expression
864 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
865 if (BO && !NotRewritable.count(BO)) {
870 // Otherwise, grab a spare node from the original expression and use that as
871 // the left-hand side. If there are no nodes left then the optimizers made
872 // an expression with more nodes than the original! This usually means that
873 // they did something stupid but it might mean that the problem was just too
874 // hard (finding the mimimal number of multiplications needed to realize a
875 // multiplication expression is NP-complete). Whatever the reason, smart or
876 // stupid, create a new node if there are none left.
877 BinaryOperator *NewOp;
878 if (NodesToRewrite.empty()) {
879 Constant *Undef = UndefValue::get(I->getType());
880 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
881 Undef, Undef, "", I);
882 if (NewOp->getType()->isFloatingPointTy())
883 NewOp->setFastMathFlags(I->getFastMathFlags());
885 NewOp = NodesToRewrite.pop_back_val();
888 DEBUG(dbgs() << "RA: " << *Op << '\n');
889 Op->setOperand(0, NewOp);
890 DEBUG(dbgs() << "TO: " << *Op << '\n');
891 ExpressionChanged = Op;
897 // If the expression changed non-trivially then clear out all subclass data
898 // starting from the operator specified in ExpressionChanged, and compactify
899 // the operators to just before the expression root to guarantee that the
900 // expression tree is dominated by all of Ops.
901 if (ExpressionChanged)
903 // Preserve FastMathFlags.
904 if (isa<FPMathOperator>(I)) {
905 FastMathFlags Flags = I->getFastMathFlags();
906 ExpressionChanged->clearSubclassOptionalData();
907 ExpressionChanged->setFastMathFlags(Flags);
909 ExpressionChanged->clearSubclassOptionalData();
911 if (ExpressionChanged == I)
913 ExpressionChanged->moveBefore(I);
914 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
917 // Throw away any left over nodes from the original expression.
918 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
919 RedoInsts.insert(NodesToRewrite[i]);
922 /// NegateValue - Insert instructions before the instruction pointed to by BI,
923 /// that computes the negative version of the value specified. The negative
924 /// version of the value is returned, and BI is left pointing at the instruction
925 /// that should be processed next by the reassociation pass.
926 static Value *NegateValue(Value *V, Instruction *BI) {
927 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
928 return ConstantExpr::getFNeg(C);
929 if (Constant *C = dyn_cast<Constant>(V))
930 return ConstantExpr::getNeg(C);
932 // We are trying to expose opportunity for reassociation. One of the things
933 // that we want to do to achieve this is to push a negation as deep into an
934 // expression chain as possible, to expose the add instructions. In practice,
935 // this means that we turn this:
936 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
937 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
938 // the constants. We assume that instcombine will clean up the mess later if
939 // we introduce tons of unnecessary negation instructions.
941 if (BinaryOperator *I =
942 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
943 // Push the negates through the add.
944 I->setOperand(0, NegateValue(I->getOperand(0), BI));
945 I->setOperand(1, NegateValue(I->getOperand(1), BI));
947 // We must move the add instruction here, because the neg instructions do
948 // not dominate the old add instruction in general. By moving it, we are
949 // assured that the neg instructions we just inserted dominate the
950 // instruction we are about to insert after them.
953 I->setName(I->getName()+".neg");
957 // Okay, we need to materialize a negated version of V with an instruction.
958 // Scan the use lists of V to see if we have one already.
959 for (User *U : V->users()) {
960 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
963 // We found one! Now we have to make sure that the definition dominates
964 // this use. We do this by moving it to the entry block (if it is a
965 // non-instruction value) or right after the definition. These negates will
966 // be zapped by reassociate later, so we don't need much finesse here.
967 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
969 // Verify that the negate is in this function, V might be a constant expr.
970 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
973 BasicBlock::iterator InsertPt;
974 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
975 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
976 InsertPt = II->getNormalDest()->begin();
978 InsertPt = InstInput;
981 while (isa<PHINode>(InsertPt)) ++InsertPt;
983 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
985 TheNeg->moveBefore(InsertPt);
989 // Insert a 'neg' instruction that subtracts the value from zero to get the
991 return CreateNeg(V, V->getName() + ".neg", BI, BI);
994 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
995 /// X-Y into (X + -Y).
996 static bool ShouldBreakUpSubtract(Instruction *Sub) {
997 // If this is a negation, we can't split it up!
998 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
1001 // Don't breakup X - undef.
1002 if (isa<UndefValue>(Sub->getOperand(1)))
1005 // Don't bother to break this up unless either the LHS is an associable add or
1006 // subtract or if this is only used by one.
1007 Value *V0 = Sub->getOperand(0);
1008 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
1009 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
1011 Value *V1 = Sub->getOperand(1);
1012 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
1013 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
1015 Value *VB = Sub->user_back();
1016 if (Sub->hasOneUse() &&
1017 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
1018 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
1024 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
1025 /// only used by an add, transform this into (X+(0-Y)) to promote better
1027 static BinaryOperator *BreakUpSubtract(Instruction *Sub) {
1028 // Convert a subtract into an add and a neg instruction. This allows sub
1029 // instructions to be commuted with other add instructions.
1031 // Calculate the negative value of Operand 1 of the sub instruction,
1032 // and set it as the RHS of the add instruction we just made.
1034 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
1035 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
1036 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
1037 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
1040 // Everyone now refers to the add instruction.
1041 Sub->replaceAllUsesWith(New);
1042 New->setDebugLoc(Sub->getDebugLoc());
1044 DEBUG(dbgs() << "Negated: " << *New << '\n');
1048 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
1049 /// by one, change this into a multiply by a constant to assist with further
1051 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1052 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
1053 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
1055 BinaryOperator *Mul =
1056 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
1057 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
1060 // Everyone now refers to the mul instruction.
1061 Shl->replaceAllUsesWith(Mul);
1062 Mul->setDebugLoc(Shl->getDebugLoc());
1064 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
1065 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
1067 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
1068 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
1070 Mul->setHasNoSignedWrap(true);
1071 Mul->setHasNoUnsignedWrap(NUW);
1075 /// FindInOperandList - Scan backwards and forwards among values with the same
1076 /// rank as element i to see if X exists. If X does not exist, return i. This
1077 /// is useful when scanning for 'x' when we see '-x' because they both get the
1079 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
1081 unsigned XRank = Ops[i].Rank;
1082 unsigned e = Ops.size();
1083 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
1086 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1087 if (Instruction *I2 = dyn_cast<Instruction>(X))
1088 if (I1->isIdenticalTo(I2))
1092 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
1095 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1096 if (Instruction *I2 = dyn_cast<Instruction>(X))
1097 if (I1->isIdenticalTo(I2))
1103 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
1104 /// and returning the result. Insert the tree before I.
1105 static Value *EmitAddTreeOfValues(Instruction *I,
1106 SmallVectorImpl<WeakVH> &Ops){
1107 if (Ops.size() == 1) return Ops.back();
1109 Value *V1 = Ops.back();
1111 Value *V2 = EmitAddTreeOfValues(I, Ops);
1112 return CreateAdd(V2, V1, "tmp", I, I);
1115 /// RemoveFactorFromExpression - If V is an expression tree that is a
1116 /// multiplication sequence, and if this sequence contains a multiply by Factor,
1117 /// remove Factor from the tree and return the new tree.
1118 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
1119 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1123 SmallVector<RepeatedValue, 8> Tree;
1124 MadeChange |= LinearizeExprTree(BO, Tree);
1125 SmallVector<ValueEntry, 8> Factors;
1126 Factors.reserve(Tree.size());
1127 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1128 RepeatedValue E = Tree[i];
1129 Factors.append(E.second.getZExtValue(),
1130 ValueEntry(getRank(E.first), E.first));
1133 bool FoundFactor = false;
1134 bool NeedsNegate = false;
1135 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1136 if (Factors[i].Op == Factor) {
1138 Factors.erase(Factors.begin()+i);
1142 // If this is a negative version of this factor, remove it.
1143 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1144 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1145 if (FC1->getValue() == -FC2->getValue()) {
1146 FoundFactor = NeedsNegate = true;
1147 Factors.erase(Factors.begin()+i);
1150 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1151 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1152 APFloat F1(FC1->getValueAPF());
1153 APFloat F2(FC2->getValueAPF());
1155 if (F1.compare(F2) == APFloat::cmpEqual) {
1156 FoundFactor = NeedsNegate = true;
1157 Factors.erase(Factors.begin() + i);
1165 // Make sure to restore the operands to the expression tree.
1166 RewriteExprTree(BO, Factors);
1170 BasicBlock::iterator InsertPt = BO; ++InsertPt;
1172 // If this was just a single multiply, remove the multiply and return the only
1173 // remaining operand.
1174 if (Factors.size() == 1) {
1175 RedoInsts.insert(BO);
1178 RewriteExprTree(BO, Factors);
1183 V = CreateNeg(V, "neg", InsertPt, BO);
1188 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
1189 /// add its operands as factors, otherwise add V to the list of factors.
1191 /// Ops is the top-level list of add operands we're trying to factor.
1192 static void FindSingleUseMultiplyFactors(Value *V,
1193 SmallVectorImpl<Value*> &Factors,
1194 const SmallVectorImpl<ValueEntry> &Ops) {
1195 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1197 Factors.push_back(V);
1201 // Otherwise, add the LHS and RHS to the list of factors.
1202 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
1203 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
1206 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
1207 /// instruction. This optimizes based on identities. If it can be reduced to
1208 /// a single Value, it is returned, otherwise the Ops list is mutated as
1210 static Value *OptimizeAndOrXor(unsigned Opcode,
1211 SmallVectorImpl<ValueEntry> &Ops) {
1212 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1213 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1214 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1215 // First, check for X and ~X in the operand list.
1216 assert(i < Ops.size());
1217 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1218 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1219 unsigned FoundX = FindInOperandList(Ops, i, X);
1221 if (Opcode == Instruction::And) // ...&X&~X = 0
1222 return Constant::getNullValue(X->getType());
1224 if (Opcode == Instruction::Or) // ...|X|~X = -1
1225 return Constant::getAllOnesValue(X->getType());
1229 // Next, check for duplicate pairs of values, which we assume are next to
1230 // each other, due to our sorting criteria.
1231 assert(i < Ops.size());
1232 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1233 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1234 // Drop duplicate values for And and Or.
1235 Ops.erase(Ops.begin()+i);
1241 // Drop pairs of values for Xor.
1242 assert(Opcode == Instruction::Xor);
1244 return Constant::getNullValue(Ops[0].Op->getType());
1247 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1255 /// Helper funciton of CombineXorOpnd(). It creates a bitwise-and
1256 /// instruction with the given two operands, and return the resulting
1257 /// instruction. There are two special cases: 1) if the constant operand is 0,
1258 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1260 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1261 const APInt &ConstOpnd) {
1262 if (ConstOpnd != 0) {
1263 if (!ConstOpnd.isAllOnesValue()) {
1264 LLVMContext &Ctx = Opnd->getType()->getContext();
1266 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1267 "and.ra", InsertBefore);
1268 I->setDebugLoc(InsertBefore->getDebugLoc());
1276 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1277 // into "R ^ C", where C would be 0, and R is a symbolic value.
1279 // If it was successful, true is returned, and the "R" and "C" is returned
1280 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1281 // and both "Res" and "ConstOpnd" remain unchanged.
1283 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1284 APInt &ConstOpnd, Value *&Res) {
1285 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1286 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1287 // = (x & ~c1) ^ (c1 ^ c2)
1288 // It is useful only when c1 == c2.
1289 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) {
1290 if (!Opnd1->getValue()->hasOneUse())
1293 const APInt &C1 = Opnd1->getConstPart();
1294 if (C1 != ConstOpnd)
1297 Value *X = Opnd1->getSymbolicPart();
1298 Res = createAndInstr(I, X, ~C1);
1299 // ConstOpnd was C2, now C1 ^ C2.
1302 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1303 RedoInsts.insert(T);
1310 // Helper function of OptimizeXor(). It tries to simplify
1311 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1314 // If it was successful, true is returned, and the "R" and "C" is returned
1315 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1316 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1317 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1318 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
1319 APInt &ConstOpnd, Value *&Res) {
1320 Value *X = Opnd1->getSymbolicPart();
1321 if (X != Opnd2->getSymbolicPart())
1324 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1325 int DeadInstNum = 1;
1326 if (Opnd1->getValue()->hasOneUse())
1328 if (Opnd2->getValue()->hasOneUse())
1332 // (x | c1) ^ (x & c2)
1333 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1334 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1335 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1337 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1338 if (Opnd2->isOrExpr())
1339 std::swap(Opnd1, Opnd2);
1341 const APInt &C1 = Opnd1->getConstPart();
1342 const APInt &C2 = Opnd2->getConstPart();
1343 APInt C3((~C1) ^ C2);
1345 // Do not increase code size!
1346 if (C3 != 0 && !C3.isAllOnesValue()) {
1347 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1348 if (NewInstNum > DeadInstNum)
1352 Res = createAndInstr(I, X, C3);
1355 } else if (Opnd1->isOrExpr()) {
1356 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1358 const APInt &C1 = Opnd1->getConstPart();
1359 const APInt &C2 = Opnd2->getConstPart();
1362 // Do not increase code size
1363 if (C3 != 0 && !C3.isAllOnesValue()) {
1364 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1365 if (NewInstNum > DeadInstNum)
1369 Res = createAndInstr(I, X, C3);
1372 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1374 const APInt &C1 = Opnd1->getConstPart();
1375 const APInt &C2 = Opnd2->getConstPart();
1377 Res = createAndInstr(I, X, C3);
1380 // Put the original operands in the Redo list; hope they will be deleted
1382 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1383 RedoInsts.insert(T);
1384 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1385 RedoInsts.insert(T);
1390 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1391 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1393 Value *Reassociate::OptimizeXor(Instruction *I,
1394 SmallVectorImpl<ValueEntry> &Ops) {
1395 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1398 if (Ops.size() == 1)
1401 SmallVector<XorOpnd, 8> Opnds;
1402 SmallVector<XorOpnd*, 8> OpndPtrs;
1403 Type *Ty = Ops[0].Op->getType();
1404 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1406 // Step 1: Convert ValueEntry to XorOpnd
1407 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1408 Value *V = Ops[i].Op;
1409 if (!isa<ConstantInt>(V)) {
1411 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1414 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1417 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1418 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1419 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1420 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1421 // when new elements are added to the vector.
1422 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1423 OpndPtrs.push_back(&Opnds[i]);
1425 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1426 // the same symbolic value cluster together. For instance, the input operand
1427 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1428 // ("x | 123", "x & 789", "y & 456").
1429 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor());
1431 // Step 3: Combine adjacent operands
1432 XorOpnd *PrevOpnd = nullptr;
1433 bool Changed = false;
1434 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1435 XorOpnd *CurrOpnd = OpndPtrs[i];
1436 // The combined value
1439 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1440 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1443 *CurrOpnd = XorOpnd(CV);
1445 CurrOpnd->Invalidate();
1450 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1451 PrevOpnd = CurrOpnd;
1455 // step 3.2: When previous and current operands share the same symbolic
1456 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1458 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1459 // Remove previous operand
1460 PrevOpnd->Invalidate();
1462 *CurrOpnd = XorOpnd(CV);
1463 PrevOpnd = CurrOpnd;
1465 CurrOpnd->Invalidate();
1472 // Step 4: Reassemble the Ops
1475 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1476 XorOpnd &O = Opnds[i];
1479 ValueEntry VE(getRank(O.getValue()), O.getValue());
1482 if (ConstOpnd != 0) {
1483 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1484 ValueEntry VE(getRank(C), C);
1487 int Sz = Ops.size();
1489 return Ops.back().Op;
1491 assert(ConstOpnd == 0);
1492 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1499 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
1500 /// optimizes based on identities. If it can be reduced to a single Value, it
1501 /// is returned, otherwise the Ops list is mutated as necessary.
1502 Value *Reassociate::OptimizeAdd(Instruction *I,
1503 SmallVectorImpl<ValueEntry> &Ops) {
1504 // Scan the operand lists looking for X and -X pairs. If we find any, we
1505 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1507 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1509 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1510 Value *TheOp = Ops[i].Op;
1511 // Check to see if we've seen this operand before. If so, we factor all
1512 // instances of the operand together. Due to our sorting criteria, we know
1513 // that these need to be next to each other in the vector.
1514 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1515 // Rescan the list, remove all instances of this operand from the expr.
1516 unsigned NumFound = 0;
1518 Ops.erase(Ops.begin()+i);
1520 } while (i != Ops.size() && Ops[i].Op == TheOp);
1522 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1525 // Insert a new multiply.
1526 Type *Ty = TheOp->getType();
1527 Constant *C = Ty->isIntegerTy() ? ConstantInt::get(Ty, NumFound)
1528 : ConstantFP::get(Ty, NumFound);
1529 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1531 // Now that we have inserted a multiply, optimize it. This allows us to
1532 // handle cases that require multiple factoring steps, such as this:
1533 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1534 RedoInsts.insert(Mul);
1536 // If every add operand was a duplicate, return the multiply.
1540 // Otherwise, we had some input that didn't have the dupe, such as
1541 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1542 // things being added by this operation.
1543 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1550 // Check for X and -X or X and ~X in the operand list.
1551 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1552 !BinaryOperator::isNot(TheOp))
1556 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1557 X = BinaryOperator::getNegArgument(TheOp);
1558 else if (BinaryOperator::isNot(TheOp))
1559 X = BinaryOperator::getNotArgument(TheOp);
1561 unsigned FoundX = FindInOperandList(Ops, i, X);
1565 // Remove X and -X from the operand list.
1566 if (Ops.size() == 2 &&
1567 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1568 return Constant::getNullValue(X->getType());
1570 // Remove X and ~X from the operand list.
1571 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1572 return Constant::getAllOnesValue(X->getType());
1574 Ops.erase(Ops.begin()+i);
1578 --i; // Need to back up an extra one.
1579 Ops.erase(Ops.begin()+FoundX);
1581 --i; // Revisit element.
1582 e -= 2; // Removed two elements.
1584 // if X and ~X we append -1 to the operand list.
1585 if (BinaryOperator::isNot(TheOp)) {
1586 Value *V = Constant::getAllOnesValue(X->getType());
1587 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1592 // Scan the operand list, checking to see if there are any common factors
1593 // between operands. Consider something like A*A+A*B*C+D. We would like to
1594 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1595 // To efficiently find this, we count the number of times a factor occurs
1596 // for any ADD operands that are MULs.
1597 DenseMap<Value*, unsigned> FactorOccurrences;
1599 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1600 // where they are actually the same multiply.
1601 unsigned MaxOcc = 0;
1602 Value *MaxOccVal = nullptr;
1603 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1604 BinaryOperator *BOp =
1605 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1609 // Compute all of the factors of this added value.
1610 SmallVector<Value*, 8> Factors;
1611 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1612 assert(Factors.size() > 1 && "Bad linearize!");
1614 // Add one to FactorOccurrences for each unique factor in this op.
1615 SmallPtrSet<Value*, 8> Duplicates;
1616 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1617 Value *Factor = Factors[i];
1618 if (!Duplicates.insert(Factor).second)
1621 unsigned Occ = ++FactorOccurrences[Factor];
1627 // If Factor is a negative constant, add the negated value as a factor
1628 // because we can percolate the negate out. Watch for minint, which
1629 // cannot be positivified.
1630 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1631 if (CI->isNegative() && !CI->isMinValue(true)) {
1632 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1633 assert(!Duplicates.count(Factor) &&
1634 "Shouldn't have two constant factors, missed a canonicalize");
1635 unsigned Occ = ++FactorOccurrences[Factor];
1641 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1642 if (CF->isNegative()) {
1643 APFloat F(CF->getValueAPF());
1645 Factor = ConstantFP::get(CF->getContext(), F);
1646 assert(!Duplicates.count(Factor) &&
1647 "Shouldn't have two constant factors, missed a canonicalize");
1648 unsigned Occ = ++FactorOccurrences[Factor];
1658 // If any factor occurred more than one time, we can pull it out.
1660 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1663 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1664 // this, we could otherwise run into situations where removing a factor
1665 // from an expression will drop a use of maxocc, and this can cause
1666 // RemoveFactorFromExpression on successive values to behave differently.
1667 Instruction *DummyInst =
1668 I->getType()->isIntegerTy()
1669 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1670 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1672 SmallVector<WeakVH, 4> NewMulOps;
1673 for (unsigned i = 0; i != Ops.size(); ++i) {
1674 // Only try to remove factors from expressions we're allowed to.
1675 BinaryOperator *BOp =
1676 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1680 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1681 // The factorized operand may occur several times. Convert them all in
1683 for (unsigned j = Ops.size(); j != i;) {
1685 if (Ops[j].Op == Ops[i].Op) {
1686 NewMulOps.push_back(V);
1687 Ops.erase(Ops.begin()+j);
1694 // No need for extra uses anymore.
1697 unsigned NumAddedValues = NewMulOps.size();
1698 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1700 // Now that we have inserted the add tree, optimize it. This allows us to
1701 // handle cases that require multiple factoring steps, such as this:
1702 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1703 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1704 (void)NumAddedValues;
1705 if (Instruction *VI = dyn_cast<Instruction>(V))
1706 RedoInsts.insert(VI);
1708 // Create the multiply.
1709 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1711 // Rerun associate on the multiply in case the inner expression turned into
1712 // a multiply. We want to make sure that we keep things in canonical form.
1713 RedoInsts.insert(V2);
1715 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1716 // entire result expression is just the multiply "A*(B+C)".
1720 // Otherwise, we had some input that didn't have the factor, such as
1721 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1722 // things being added by this operation.
1723 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1729 /// \brief Build up a vector of value/power pairs factoring a product.
1731 /// Given a series of multiplication operands, build a vector of factors and
1732 /// the powers each is raised to when forming the final product. Sort them in
1733 /// the order of descending power.
1735 /// (x*x) -> [(x, 2)]
1736 /// ((x*x)*x) -> [(x, 3)]
1737 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1739 /// \returns Whether any factors have a power greater than one.
1740 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1741 SmallVectorImpl<Factor> &Factors) {
1742 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1743 // Compute the sum of powers of simplifiable factors.
1744 unsigned FactorPowerSum = 0;
1745 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1746 Value *Op = Ops[Idx-1].Op;
1748 // Count the number of occurrences of this value.
1750 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1752 // Track for simplification all factors which occur 2 or more times.
1754 FactorPowerSum += Count;
1757 // We can only simplify factors if the sum of the powers of our simplifiable
1758 // factors is 4 or higher. When that is the case, we will *always* have
1759 // a simplification. This is an important invariant to prevent cyclicly
1760 // trying to simplify already minimal formations.
1761 if (FactorPowerSum < 4)
1764 // Now gather the simplifiable factors, removing them from Ops.
1766 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1767 Value *Op = Ops[Idx-1].Op;
1769 // Count the number of occurrences of this value.
1771 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1775 // Move an even number of occurrences to Factors.
1778 FactorPowerSum += Count;
1779 Factors.push_back(Factor(Op, Count));
1780 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1783 // None of the adjustments above should have reduced the sum of factor powers
1784 // below our mininum of '4'.
1785 assert(FactorPowerSum >= 4);
1787 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1791 /// \brief Build a tree of multiplies, computing the product of Ops.
1792 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1793 SmallVectorImpl<Value*> &Ops) {
1794 if (Ops.size() == 1)
1797 Value *LHS = Ops.pop_back_val();
1799 if (LHS->getType()->isIntegerTy())
1800 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1802 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1803 } while (!Ops.empty());
1808 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1810 /// Given a vector of values raised to various powers, where no two values are
1811 /// equal and the powers are sorted in decreasing order, compute the minimal
1812 /// DAG of multiplies to compute the final product, and return that product
1814 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1815 SmallVectorImpl<Factor> &Factors) {
1816 assert(Factors[0].Power);
1817 SmallVector<Value *, 4> OuterProduct;
1818 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1819 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1820 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1825 // We want to multiply across all the factors with the same power so that
1826 // we can raise them to that power as a single entity. Build a mini tree
1828 SmallVector<Value *, 4> InnerProduct;
1829 InnerProduct.push_back(Factors[LastIdx].Base);
1831 InnerProduct.push_back(Factors[Idx].Base);
1833 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1835 // Reset the base value of the first factor to the new expression tree.
1836 // We'll remove all the factors with the same power in a second pass.
1837 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1838 if (Instruction *MI = dyn_cast<Instruction>(M))
1839 RedoInsts.insert(MI);
1843 // Unique factors with equal powers -- we've folded them into the first one's
1845 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1846 Factor::PowerEqual()),
1849 // Iteratively collect the base of each factor with an add power into the
1850 // outer product, and halve each power in preparation for squaring the
1852 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1853 if (Factors[Idx].Power & 1)
1854 OuterProduct.push_back(Factors[Idx].Base);
1855 Factors[Idx].Power >>= 1;
1857 if (Factors[0].Power) {
1858 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1859 OuterProduct.push_back(SquareRoot);
1860 OuterProduct.push_back(SquareRoot);
1862 if (OuterProduct.size() == 1)
1863 return OuterProduct.front();
1865 Value *V = buildMultiplyTree(Builder, OuterProduct);
1869 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1870 SmallVectorImpl<ValueEntry> &Ops) {
1871 // We can only optimize the multiplies when there is a chain of more than
1872 // three, such that a balanced tree might require fewer total multiplies.
1876 // Try to turn linear trees of multiplies without other uses of the
1877 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1879 SmallVector<Factor, 4> Factors;
1880 if (!collectMultiplyFactors(Ops, Factors))
1881 return nullptr; // All distinct factors, so nothing left for us to do.
1883 IRBuilder<> Builder(I);
1884 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1888 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1889 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1893 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1894 SmallVectorImpl<ValueEntry> &Ops) {
1895 // Now that we have the linearized expression tree, try to optimize it.
1896 // Start by folding any constants that we found.
1897 Constant *Cst = nullptr;
1898 unsigned Opcode = I->getOpcode();
1899 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1900 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1901 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1903 // If there was nothing but constants then we are done.
1907 // Put the combined constant back at the end of the operand list, except if
1908 // there is no point. For example, an add of 0 gets dropped here, while a
1909 // multiplication by zero turns the whole expression into zero.
1910 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1911 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1913 Ops.push_back(ValueEntry(0, Cst));
1916 if (Ops.size() == 1) return Ops[0].Op;
1918 // Handle destructive annihilation due to identities between elements in the
1919 // argument list here.
1920 unsigned NumOps = Ops.size();
1923 case Instruction::And:
1924 case Instruction::Or:
1925 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1929 case Instruction::Xor:
1930 if (Value *Result = OptimizeXor(I, Ops))
1934 case Instruction::Add:
1935 case Instruction::FAdd:
1936 if (Value *Result = OptimizeAdd(I, Ops))
1940 case Instruction::Mul:
1941 case Instruction::FMul:
1942 if (Value *Result = OptimizeMul(I, Ops))
1947 if (Ops.size() != NumOps)
1948 return OptimizeExpression(I, Ops);
1952 /// EraseInst - Zap the given instruction, adding interesting operands to the
1954 void Reassociate::EraseInst(Instruction *I) {
1955 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1956 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1957 // Erase the dead instruction.
1958 ValueRankMap.erase(I);
1959 RedoInsts.remove(I);
1960 I->eraseFromParent();
1961 // Optimize its operands.
1962 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1963 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1964 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1965 // If this is a node in an expression tree, climb to the expression root
1966 // and add that since that's where optimization actually happens.
1967 unsigned Opcode = Op->getOpcode();
1968 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1969 Visited.insert(Op).second)
1970 Op = Op->user_back();
1971 RedoInsts.insert(Op);
1975 // Canonicalize expressions of the following form:
1976 // x + (-Constant * y) -> x - (Constant * y)
1977 // x - (-Constant * y) -> x + (Constant * y)
1978 Instruction *Reassociate::canonicalizeNegConstExpr(Instruction *I) {
1979 if (!I->hasOneUse() || I->getType()->isVectorTy())
1982 // Must be a mul, fmul, or fdiv instruction.
1983 unsigned Opcode = I->getOpcode();
1984 if (Opcode != Instruction::Mul && Opcode != Instruction::FMul &&
1985 Opcode != Instruction::FDiv)
1988 // Must have at least one constant operand.
1989 Constant *C0 = dyn_cast<Constant>(I->getOperand(0));
1990 Constant *C1 = dyn_cast<Constant>(I->getOperand(1));
1994 // Must be a negative ConstantInt or ConstantFP.
1995 Constant *C = C0 ? C0 : C1;
1996 unsigned ConstIdx = C0 ? 0 : 1;
1997 if (auto *CI = dyn_cast<ConstantInt>(C)) {
1998 if (!CI->isNegative())
2000 } else if (auto *CF = dyn_cast<ConstantFP>(C)) {
2001 if (!CF->isNegative())
2006 // User must be a binary operator with one or more uses.
2007 Instruction *User = I->user_back();
2008 if (!isa<BinaryOperator>(User) || !User->getNumUses())
2011 unsigned UserOpcode = User->getOpcode();
2012 if (UserOpcode != Instruction::Add && UserOpcode != Instruction::FAdd &&
2013 UserOpcode != Instruction::Sub && UserOpcode != Instruction::FSub)
2016 // Subtraction is not commutative. Explicitly, the following transform is
2017 // not valid: (-Constant * y) - x -> x + (Constant * y)
2018 if (!User->isCommutative() && User->getOperand(1) != I)
2021 // Change the sign of the constant.
2022 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
2023 I->setOperand(ConstIdx, ConstantInt::get(CI->getContext(), -CI->getValue()));
2025 ConstantFP *CF = cast<ConstantFP>(C);
2026 APFloat Val = CF->getValueAPF();
2028 I->setOperand(ConstIdx, ConstantFP::get(CF->getContext(), Val));
2031 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
2032 // ((-Const*y) + x) -> (x + (-Const*y)).
2033 if (User->getOperand(0) == I && User->isCommutative())
2034 cast<BinaryOperator>(User)->swapOperands();
2036 Value *Op0 = User->getOperand(0);
2037 Value *Op1 = User->getOperand(1);
2039 switch(UserOpcode) {
2041 llvm_unreachable("Unexpected Opcode!");
2042 case Instruction::Add:
2043 NI = BinaryOperator::CreateSub(Op0, Op1);
2045 case Instruction::Sub:
2046 NI = BinaryOperator::CreateAdd(Op0, Op1);
2048 case Instruction::FAdd:
2049 NI = BinaryOperator::CreateFSub(Op0, Op1);
2050 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2052 case Instruction::FSub:
2053 NI = BinaryOperator::CreateFAdd(Op0, Op1);
2054 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
2058 NI->insertBefore(User);
2059 NI->setName(User->getName());
2060 User->replaceAllUsesWith(NI);
2061 NI->setDebugLoc(I->getDebugLoc());
2062 RedoInsts.insert(I);
2067 /// OptimizeInst - Inspect and optimize the given instruction. Note that erasing
2068 /// instructions is not allowed.
2069 void Reassociate::OptimizeInst(Instruction *I) {
2070 // Only consider operations that we understand.
2071 if (!isa<BinaryOperator>(I))
2074 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
2075 // If an operand of this shift is a reassociable multiply, or if the shift
2076 // is used by a reassociable multiply or add, turn into a multiply.
2077 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
2079 (isReassociableOp(I->user_back(), Instruction::Mul) ||
2080 isReassociableOp(I->user_back(), Instruction::Add)))) {
2081 Instruction *NI = ConvertShiftToMul(I);
2082 RedoInsts.insert(I);
2087 // Canonicalize negative constants out of expressions.
2088 if (Instruction *Res = canonicalizeNegConstExpr(I))
2091 // Commute binary operators, to canonicalize the order of their operands.
2092 // This can potentially expose more CSE opportunities, and makes writing other
2093 // transformations simpler.
2094 if (I->isCommutative())
2095 canonicalizeOperands(I);
2097 // Don't optimize vector instructions.
2098 if (I->getType()->isVectorTy())
2101 // Don't optimize floating point instructions that don't have unsafe algebra.
2102 if (I->getType()->isFloatingPointTy() && !I->hasUnsafeAlgebra())
2105 // Do not reassociate boolean (i1) expressions. We want to preserve the
2106 // original order of evaluation for short-circuited comparisons that
2107 // SimplifyCFG has folded to AND/OR expressions. If the expression
2108 // is not further optimized, it is likely to be transformed back to a
2109 // short-circuited form for code gen, and the source order may have been
2110 // optimized for the most likely conditions.
2111 if (I->getType()->isIntegerTy(1))
2114 // If this is a subtract instruction which is not already in negate form,
2115 // see if we can convert it to X+-Y.
2116 if (I->getOpcode() == Instruction::Sub) {
2117 if (ShouldBreakUpSubtract(I)) {
2118 Instruction *NI = BreakUpSubtract(I);
2119 RedoInsts.insert(I);
2122 } else if (BinaryOperator::isNeg(I)) {
2123 // Otherwise, this is a negation. See if the operand is a multiply tree
2124 // and if this is not an inner node of a multiply tree.
2125 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2127 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2128 Instruction *NI = LowerNegateToMultiply(I);
2129 RedoInsts.insert(I);
2134 } else if (I->getOpcode() == Instruction::FSub) {
2135 if (ShouldBreakUpSubtract(I)) {
2136 Instruction *NI = BreakUpSubtract(I);
2137 RedoInsts.insert(I);
2140 } else if (BinaryOperator::isFNeg(I)) {
2141 // Otherwise, this is a negation. See if the operand is a multiply tree
2142 // and if this is not an inner node of a multiply tree.
2143 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2145 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2146 Instruction *NI = LowerNegateToMultiply(I);
2147 RedoInsts.insert(I);
2154 // If this instruction is an associative binary operator, process it.
2155 if (!I->isAssociative()) return;
2156 BinaryOperator *BO = cast<BinaryOperator>(I);
2158 // If this is an interior node of a reassociable tree, ignore it until we
2159 // get to the root of the tree, to avoid N^2 analysis.
2160 unsigned Opcode = BO->getOpcode();
2161 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode)
2164 // If this is an add tree that is used by a sub instruction, ignore it
2165 // until we process the subtract.
2166 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2167 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2169 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2170 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2173 ReassociateExpression(BO);
2176 void Reassociate::ReassociateExpression(BinaryOperator *I) {
2177 assert(!I->getType()->isVectorTy() &&
2178 "Reassociation of vector instructions is not supported.");
2180 // First, walk the expression tree, linearizing the tree, collecting the
2181 // operand information.
2182 SmallVector<RepeatedValue, 8> Tree;
2183 MadeChange |= LinearizeExprTree(I, Tree);
2184 SmallVector<ValueEntry, 8> Ops;
2185 Ops.reserve(Tree.size());
2186 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2187 RepeatedValue E = Tree[i];
2188 Ops.append(E.second.getZExtValue(),
2189 ValueEntry(getRank(E.first), E.first));
2192 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2194 // Now that we have linearized the tree to a list and have gathered all of
2195 // the operands and their ranks, sort the operands by their rank. Use a
2196 // stable_sort so that values with equal ranks will have their relative
2197 // positions maintained (and so the compiler is deterministic). Note that
2198 // this sorts so that the highest ranking values end up at the beginning of
2200 std::stable_sort(Ops.begin(), Ops.end());
2202 // OptimizeExpression - Now that we have the expression tree in a convenient
2203 // sorted form, optimize it globally if possible.
2204 if (Value *V = OptimizeExpression(I, Ops)) {
2206 // Self-referential expression in unreachable code.
2208 // This expression tree simplified to something that isn't a tree,
2210 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2211 I->replaceAllUsesWith(V);
2212 if (Instruction *VI = dyn_cast<Instruction>(V))
2213 VI->setDebugLoc(I->getDebugLoc());
2214 RedoInsts.insert(I);
2219 // We want to sink immediates as deeply as possible except in the case where
2220 // this is a multiply tree used only by an add, and the immediate is a -1.
2221 // In this case we reassociate to put the negation on the outside so that we
2222 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2223 if (I->hasOneUse()) {
2224 if (I->getOpcode() == Instruction::Mul &&
2225 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2226 isa<ConstantInt>(Ops.back().Op) &&
2227 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
2228 ValueEntry Tmp = Ops.pop_back_val();
2229 Ops.insert(Ops.begin(), Tmp);
2230 } else if (I->getOpcode() == Instruction::FMul &&
2231 cast<Instruction>(I->user_back())->getOpcode() ==
2232 Instruction::FAdd &&
2233 isa<ConstantFP>(Ops.back().Op) &&
2234 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2235 ValueEntry Tmp = Ops.pop_back_val();
2236 Ops.insert(Ops.begin(), Tmp);
2240 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2242 if (Ops.size() == 1) {
2244 // Self-referential expression in unreachable code.
2247 // This expression tree simplified to something that isn't a tree,
2249 I->replaceAllUsesWith(Ops[0].Op);
2250 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2251 OI->setDebugLoc(I->getDebugLoc());
2252 RedoInsts.insert(I);
2256 // Now that we ordered and optimized the expressions, splat them back into
2257 // the expression tree, removing any unneeded nodes.
2258 RewriteExprTree(I, Ops);
2261 bool Reassociate::runOnFunction(Function &F) {
2262 if (skipOptnoneFunction(F))
2265 // Calculate the rank map for F
2269 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
2270 // Optimize every instruction in the basic block.
2271 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; )
2272 if (isInstructionTriviallyDead(II)) {
2276 assert(II->getParent() == BI && "Moved to a different block!");
2280 // If this produced extra instructions to optimize, handle them now.
2281 while (!RedoInsts.empty()) {
2282 Instruction *I = RedoInsts.pop_back_val();
2283 if (isInstructionTriviallyDead(I))
2290 // We are done with the rank map.
2292 ValueRankMap.clear();