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 #define DEBUG_TYPE "reassociate"
24 #include "llvm/Transforms/Scalar.h"
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/PostOrderIterator.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SetVector.h"
29 #include "llvm/ADT/Statistic.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/Pass.h"
37 #include "llvm/Support/CFG.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ValueHandle.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Transforms/Utils/Local.h"
45 STATISTIC(NumChanged, "Number of insts reassociated");
46 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
47 STATISTIC(NumFactor , "Number of multiplies factored");
53 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {}
55 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) {
56 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start.
61 /// PrintOps - Print out the expression identified in the Ops list.
63 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
64 Module *M = I->getParent()->getParent()->getParent();
65 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
66 << *Ops[0].Op->getType() << '\t';
67 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
69 Ops[i].Op->printAsOperand(dbgs(), false, M);
70 dbgs() << ", #" << Ops[i].Rank << "] ";
76 /// \brief Utility class representing a base and exponent pair which form one
77 /// factor of some product.
82 Factor(Value *Base, unsigned Power) : Base(Base), Power(Power) {}
84 /// \brief Sort factors by their Base.
86 bool operator()(const Factor &LHS, const Factor &RHS) {
87 return LHS.Base < RHS.Base;
91 /// \brief Compare factors for equal bases.
93 bool operator()(const Factor &LHS, const Factor &RHS) {
94 return LHS.Base == RHS.Base;
98 /// \brief Sort factors in descending order by their power.
99 struct PowerDescendingSorter {
100 bool operator()(const Factor &LHS, const Factor &RHS) {
101 return LHS.Power > RHS.Power;
105 /// \brief Compare factors for equal powers.
107 bool operator()(const Factor &LHS, const Factor &RHS) {
108 return LHS.Power == RHS.Power;
113 /// Utility class representing a non-constant Xor-operand. We classify
114 /// non-constant Xor-Operands into two categories:
115 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0
117 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
119 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
120 /// operand as "E | 0"
125 bool isInvalid() const { return SymbolicPart == 0; }
126 bool isOrExpr() const { return isOr; }
127 Value *getValue() const { return OrigVal; }
128 Value *getSymbolicPart() const { return SymbolicPart; }
129 unsigned getSymbolicRank() const { return SymbolicRank; }
130 const APInt &getConstPart() const { return ConstPart; }
132 void Invalidate() { SymbolicPart = OrigVal = 0; }
133 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
135 // Sort the XorOpnd-Pointer in ascending order of symbolic-value-rank.
136 // The purpose is twofold:
137 // 1) Cluster together the operands sharing the same symbolic-value.
138 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
139 // could potentially shorten crital path, and expose more loop-invariants.
140 // Note that values' rank are basically defined in RPO order (FIXME).
141 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
142 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
143 // "z" in the order of X-Y-Z is better than any other orders.
144 struct PtrSortFunctor {
145 bool operator()(XorOpnd * const &LHS, XorOpnd * const &RHS) {
146 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
153 unsigned SymbolicRank;
159 class Reassociate : public FunctionPass {
160 DenseMap<BasicBlock*, unsigned> RankMap;
161 DenseMap<AssertingVH<Value>, unsigned> ValueRankMap;
162 SetVector<AssertingVH<Instruction> > RedoInsts;
165 static char ID; // Pass identification, replacement for typeid
166 Reassociate() : FunctionPass(ID) {
167 initializeReassociatePass(*PassRegistry::getPassRegistry());
170 bool runOnFunction(Function &F);
172 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
173 AU.setPreservesCFG();
176 void BuildRankMap(Function &F);
177 unsigned getRank(Value *V);
178 void ReassociateExpression(BinaryOperator *I);
179 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
180 Value *OptimizeExpression(BinaryOperator *I,
181 SmallVectorImpl<ValueEntry> &Ops);
182 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
183 Value *OptimizeXor(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
184 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, APInt &ConstOpnd,
186 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
187 APInt &ConstOpnd, Value *&Res);
188 bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
189 SmallVectorImpl<Factor> &Factors);
190 Value *buildMinimalMultiplyDAG(IRBuilder<> &Builder,
191 SmallVectorImpl<Factor> &Factors);
192 Value *OptimizeMul(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
193 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
194 void EraseInst(Instruction *I);
195 void OptimizeInst(Instruction *I);
199 XorOpnd::XorOpnd(Value *V) {
200 assert(!isa<ConstantInt>(V) && "No ConstantInt");
202 Instruction *I = dyn_cast<Instruction>(V);
205 if (I && (I->getOpcode() == Instruction::Or ||
206 I->getOpcode() == Instruction::And)) {
207 Value *V0 = I->getOperand(0);
208 Value *V1 = I->getOperand(1);
209 if (isa<ConstantInt>(V0))
212 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) {
213 ConstPart = C->getValue();
215 isOr = (I->getOpcode() == Instruction::Or);
220 // view the operand as "V | 0"
222 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth());
226 char Reassociate::ID = 0;
227 INITIALIZE_PASS(Reassociate, "reassociate",
228 "Reassociate expressions", false, false)
230 // Public interface to the Reassociate pass
231 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
233 /// isReassociableOp - Return true if V is an instruction of the specified
234 /// opcode and if it only has one use.
235 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
236 if (V->hasOneUse() && isa<Instruction>(V) &&
237 cast<Instruction>(V)->getOpcode() == Opcode)
238 return cast<BinaryOperator>(V);
242 static bool isUnmovableInstruction(Instruction *I) {
243 switch (I->getOpcode()) {
244 case Instruction::PHI:
245 case Instruction::LandingPad:
246 case Instruction::Alloca:
247 case Instruction::Load:
248 case Instruction::Invoke:
249 case Instruction::UDiv:
250 case Instruction::SDiv:
251 case Instruction::FDiv:
252 case Instruction::URem:
253 case Instruction::SRem:
254 case Instruction::FRem:
256 case Instruction::Call:
257 return !isa<DbgInfoIntrinsic>(I);
263 void Reassociate::BuildRankMap(Function &F) {
266 // Assign distinct ranks to function arguments
267 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
268 ValueRankMap[&*I] = ++i;
270 ReversePostOrderTraversal<Function*> RPOT(&F);
271 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
272 E = RPOT.end(); I != E; ++I) {
274 unsigned BBRank = RankMap[BB] = ++i << 16;
276 // Walk the basic block, adding precomputed ranks for any instructions that
277 // we cannot move. This ensures that the ranks for these instructions are
278 // all different in the block.
279 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
280 if (isUnmovableInstruction(I))
281 ValueRankMap[&*I] = ++BBRank;
285 unsigned Reassociate::getRank(Value *V) {
286 Instruction *I = dyn_cast<Instruction>(V);
288 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
289 return 0; // Otherwise it's a global or constant, rank 0.
292 if (unsigned Rank = ValueRankMap[I])
293 return Rank; // Rank already known?
295 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
296 // we can reassociate expressions for code motion! Since we do not recurse
297 // for PHI nodes, we cannot have infinite recursion here, because there
298 // cannot be loops in the value graph that do not go through PHI nodes.
299 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
300 for (unsigned i = 0, e = I->getNumOperands();
301 i != e && Rank != MaxRank; ++i)
302 Rank = std::max(Rank, getRank(I->getOperand(i)));
304 // If this is a not or neg instruction, do not count it for rank. This
305 // assures us that X and ~X will have the same rank.
306 if (!I->getType()->isIntegerTy() ||
307 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
310 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = "
313 return ValueRankMap[I] = Rank;
316 /// LowerNegateToMultiply - Replace 0-X with X*-1.
318 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
319 Constant *Cst = Constant::getAllOnesValue(Neg->getType());
321 BinaryOperator *Res =
322 BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
323 Neg->setOperand(1, Constant::getNullValue(Neg->getType())); // Drop use of op.
325 Neg->replaceAllUsesWith(Res);
326 Res->setDebugLoc(Neg->getDebugLoc());
330 /// CarmichaelShift - Returns k such that lambda(2^Bitwidth) = 2^k, where lambda
331 /// is the Carmichael function. This means that x^(2^k) === 1 mod 2^Bitwidth for
332 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
333 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
334 /// even x in Bitwidth-bit arithmetic.
335 static unsigned CarmichaelShift(unsigned Bitwidth) {
341 /// IncorporateWeight - Add the extra weight 'RHS' to the existing weight 'LHS',
342 /// reducing the combined weight using any special properties of the operation.
343 /// The existing weight LHS represents the computation X op X op ... op X where
344 /// X occurs LHS times. The combined weight represents X op X op ... op X with
345 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
346 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
347 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
348 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
349 // If we were working with infinite precision arithmetic then the combined
350 // weight would be LHS + RHS. But we are using finite precision arithmetic,
351 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
352 // for nilpotent operations and addition, but not for idempotent operations
353 // and multiplication), so it is important to correctly reduce the combined
354 // weight back into range if wrapping would be wrong.
356 // If RHS is zero then the weight didn't change.
357 if (RHS.isMinValue())
359 // If LHS is zero then the combined weight is RHS.
360 if (LHS.isMinValue()) {
364 // From this point on we know that neither LHS nor RHS is zero.
366 if (Instruction::isIdempotent(Opcode)) {
367 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
368 // weight of 1. Keeping weights at zero or one also means that wrapping is
370 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
371 return; // Return a weight of 1.
373 if (Instruction::isNilpotent(Opcode)) {
374 // Nilpotent means X op X === 0, so reduce weights modulo 2.
375 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
376 LHS = 0; // 1 + 1 === 0 modulo 2.
379 if (Opcode == Instruction::Add) {
380 // TODO: Reduce the weight by exploiting nsw/nuw?
385 assert(Opcode == Instruction::Mul && "Unknown associative operation!");
386 unsigned Bitwidth = LHS.getBitWidth();
387 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
388 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
389 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
390 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
391 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
392 // which by a happy accident means that they can always be represented using
394 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
395 // the Carmichael number).
397 /// CM - The value of Carmichael's lambda function.
398 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
399 // Any weight W >= Threshold can be replaced with W - CM.
400 APInt Threshold = CM + Bitwidth;
401 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
402 // For Bitwidth 4 or more the following sum does not overflow.
404 while (LHS.uge(Threshold))
407 // To avoid problems with overflow do everything the same as above but using
409 unsigned CM = 1U << CarmichaelShift(Bitwidth);
410 unsigned Threshold = CM + Bitwidth;
411 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
412 "Weights not reduced!");
413 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
414 while (Total >= Threshold)
420 typedef std::pair<Value*, APInt> RepeatedValue;
422 /// LinearizeExprTree - Given an associative binary expression, return the leaf
423 /// nodes in Ops along with their weights (how many times the leaf occurs). The
424 /// original expression is the same as
425 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
427 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
431 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
433 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
435 /// This routine may modify the function, in which case it returns 'true'. The
436 /// changes it makes may well be destructive, changing the value computed by 'I'
437 /// to something completely different. Thus if the routine returns 'true' then
438 /// you MUST either replace I with a new expression computed from the Ops array,
439 /// or use RewriteExprTree to put the values back in.
441 /// A leaf node is either not a binary operation of the same kind as the root
442 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
443 /// opcode), or is the same kind of binary operator but has a use which either
444 /// does not belong to the expression, or does belong to the expression but is
445 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
446 /// of the expression, while for non-leaf nodes (except for the root 'I') every
447 /// use is a non-leaf node of the expression.
450 /// expression graph node names
460 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
461 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
463 /// The expression is maximal: if some instruction is a binary operator of the
464 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
465 /// then the instruction also belongs to the expression, is not a leaf node of
466 /// it, and its operands also belong to the expression (but may be leaf nodes).
468 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
469 /// order to ensure that every non-root node in the expression has *exactly one*
470 /// use by a non-leaf node of the expression. This destruction means that the
471 /// caller MUST either replace 'I' with a new expression or use something like
472 /// RewriteExprTree to put the values back in if the routine indicates that it
473 /// made a change by returning 'true'.
475 /// In the above example either the right operand of A or the left operand of B
476 /// will be replaced by undef. If it is B's operand then this gives:
480 /// + + | A, B - operand of B replaced with undef
486 /// Note that such undef operands can only be reached by passing through 'I'.
487 /// For example, if you visit operands recursively starting from a leaf node
488 /// then you will never see such an undef operand unless you get back to 'I',
489 /// which requires passing through a phi node.
491 /// Note that this routine may also mutate binary operators of the wrong type
492 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
493 /// of the expression) if it can turn them into binary operators of the right
494 /// type and thus make the expression bigger.
496 static bool LinearizeExprTree(BinaryOperator *I,
497 SmallVectorImpl<RepeatedValue> &Ops) {
498 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
499 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
500 unsigned Opcode = I->getOpcode();
501 assert(Instruction::isAssociative(Opcode) &&
502 Instruction::isCommutative(Opcode) &&
503 "Expected an associative and commutative operation!");
505 // Visit all operands of the expression, keeping track of their weight (the
506 // number of paths from the expression root to the operand, or if you like
507 // the number of times that operand occurs in the linearized expression).
508 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
509 // while A has weight two.
511 // Worklist of non-leaf nodes (their operands are in the expression too) along
512 // with their weights, representing a certain number of paths to the operator.
513 // If an operator occurs in the worklist multiple times then we found multiple
514 // ways to get to it.
515 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
516 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
517 bool MadeChange = false;
519 // Leaves of the expression are values that either aren't the right kind of
520 // operation (eg: a constant, or a multiply in an add tree), or are, but have
521 // some uses that are not inside the expression. For example, in I = X + X,
522 // X = A + B, the value X has two uses (by I) that are in the expression. If
523 // X has any other uses, for example in a return instruction, then we consider
524 // X to be a leaf, and won't analyze it further. When we first visit a value,
525 // if it has more than one use then at first we conservatively consider it to
526 // be a leaf. Later, as the expression is explored, we may discover some more
527 // uses of the value from inside the expression. If all uses turn out to be
528 // from within the expression (and the value is a binary operator of the right
529 // kind) then the value is no longer considered to be a leaf, and its operands
532 // Leaves - Keeps track of the set of putative leaves as well as the number of
533 // paths to each leaf seen so far.
534 typedef DenseMap<Value*, APInt> LeafMap;
535 LeafMap Leaves; // Leaf -> Total weight so far.
536 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
539 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
541 while (!Worklist.empty()) {
542 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
543 I = P.first; // We examine the operands of this binary operator.
545 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
546 Value *Op = I->getOperand(OpIdx);
547 APInt Weight = P.second; // Number of paths to this operand.
548 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
549 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
551 // If this is a binary operation of the right kind with only one use then
552 // add its operands to the expression.
553 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
554 assert(Visited.insert(Op) && "Not first visit!");
555 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
556 Worklist.push_back(std::make_pair(BO, Weight));
560 // Appears to be a leaf. Is the operand already in the set of leaves?
561 LeafMap::iterator It = Leaves.find(Op);
562 if (It == Leaves.end()) {
563 // Not in the leaf map. Must be the first time we saw this operand.
564 assert(Visited.insert(Op) && "Not first visit!");
565 if (!Op->hasOneUse()) {
566 // This value has uses not accounted for by the expression, so it is
567 // not safe to modify. Mark it as being a leaf.
568 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
569 LeafOrder.push_back(Op);
573 // No uses outside the expression, try morphing it.
574 } else if (It != Leaves.end()) {
575 // Already in the leaf map.
576 assert(Visited.count(Op) && "In leaf map but not visited!");
578 // Update the number of paths to the leaf.
579 IncorporateWeight(It->second, Weight, Opcode);
581 #if 0 // TODO: Re-enable once PR13021 is fixed.
582 // The leaf already has one use from inside the expression. As we want
583 // exactly one such use, drop this new use of the leaf.
584 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
585 I->setOperand(OpIdx, UndefValue::get(I->getType()));
588 // If the leaf is a binary operation of the right kind and we now see
589 // that its multiple original uses were in fact all by nodes belonging
590 // to the expression, then no longer consider it to be a leaf and add
591 // its operands to the expression.
592 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
593 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
594 Worklist.push_back(std::make_pair(BO, It->second));
600 // If we still have uses that are not accounted for by the expression
601 // then it is not safe to modify the value.
602 if (!Op->hasOneUse())
605 // No uses outside the expression, try morphing it.
607 Leaves.erase(It); // Since the value may be morphed below.
610 // At this point we have a value which, first of all, is not a binary
611 // expression of the right kind, and secondly, is only used inside the
612 // expression. This means that it can safely be modified. See if we
613 // can usefully morph it into an expression of the right kind.
614 assert((!isa<Instruction>(Op) ||
615 cast<Instruction>(Op)->getOpcode() != Opcode) &&
616 "Should have been handled above!");
617 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
619 // If this is a multiply expression, turn any internal negations into
620 // multiplies by -1 so they can be reassociated.
621 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op);
622 if (Opcode == Instruction::Mul && BO && BinaryOperator::isNeg(BO)) {
623 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
624 BO = LowerNegateToMultiply(BO);
625 DEBUG(dbgs() << *BO << 'n');
626 Worklist.push_back(std::make_pair(BO, Weight));
631 // Failed to morph into an expression of the right type. This really is
633 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
634 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
635 LeafOrder.push_back(Op);
640 // The leaves, repeated according to their weights, represent the linearized
641 // form of the expression.
642 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
643 Value *V = LeafOrder[i];
644 LeafMap::iterator It = Leaves.find(V);
645 if (It == Leaves.end())
646 // Node initially thought to be a leaf wasn't.
648 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
649 APInt Weight = It->second;
650 if (Weight.isMinValue())
651 // Leaf already output or weight reduction eliminated it.
653 // Ensure the leaf is only output once.
655 Ops.push_back(std::make_pair(V, Weight));
658 // For nilpotent operations or addition there may be no operands, for example
659 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
660 // in both cases the weight reduces to 0 causing the value to be skipped.
662 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
663 assert(Identity && "Associative operation without identity!");
664 Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1)));
670 // RewriteExprTree - Now that the operands for this expression tree are
671 // linearized and optimized, emit them in-order.
672 void Reassociate::RewriteExprTree(BinaryOperator *I,
673 SmallVectorImpl<ValueEntry> &Ops) {
674 assert(Ops.size() > 1 && "Single values should be used directly!");
676 // Since our optimizations should never increase the number of operations, the
677 // new expression can usually be written reusing the existing binary operators
678 // from the original expression tree, without creating any new instructions,
679 // though the rewritten expression may have a completely different topology.
680 // We take care to not change anything if the new expression will be the same
681 // as the original. If more than trivial changes (like commuting operands)
682 // were made then we are obliged to clear out any optional subclass data like
685 /// NodesToRewrite - Nodes from the original expression available for writing
686 /// the new expression into.
687 SmallVector<BinaryOperator*, 8> NodesToRewrite;
688 unsigned Opcode = I->getOpcode();
689 BinaryOperator *Op = I;
691 /// NotRewritable - The operands being written will be the leaves of the new
692 /// expression and must not be used as inner nodes (via NodesToRewrite) by
693 /// mistake. Inner nodes are always reassociable, and usually leaves are not
694 /// (if they were they would have been incorporated into the expression and so
695 /// would not be leaves), so most of the time there is no danger of this. But
696 /// in rare cases a leaf may become reassociable if an optimization kills uses
697 /// of it, or it may momentarily become reassociable during rewriting (below)
698 /// due it being removed as an operand of one of its uses. Ensure that misuse
699 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
700 /// leaves and refusing to reuse any of them as inner nodes.
701 SmallPtrSet<Value*, 8> NotRewritable;
702 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
703 NotRewritable.insert(Ops[i].Op);
705 // ExpressionChanged - Non-null if the rewritten expression differs from the
706 // original in some non-trivial way, requiring the clearing of optional flags.
707 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
708 BinaryOperator *ExpressionChanged = 0;
709 for (unsigned i = 0; ; ++i) {
710 // The last operation (which comes earliest in the IR) is special as both
711 // operands will come from Ops, rather than just one with the other being
713 if (i+2 == Ops.size()) {
714 Value *NewLHS = Ops[i].Op;
715 Value *NewRHS = Ops[i+1].Op;
716 Value *OldLHS = Op->getOperand(0);
717 Value *OldRHS = Op->getOperand(1);
719 if (NewLHS == OldLHS && NewRHS == OldRHS)
720 // Nothing changed, leave it alone.
723 if (NewLHS == OldRHS && NewRHS == OldLHS) {
724 // The order of the operands was reversed. Swap them.
725 DEBUG(dbgs() << "RA: " << *Op << '\n');
727 DEBUG(dbgs() << "TO: " << *Op << '\n');
733 // The new operation differs non-trivially from the original. Overwrite
734 // the old operands with the new ones.
735 DEBUG(dbgs() << "RA: " << *Op << '\n');
736 if (NewLHS != OldLHS) {
737 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
738 if (BO && !NotRewritable.count(BO))
739 NodesToRewrite.push_back(BO);
740 Op->setOperand(0, NewLHS);
742 if (NewRHS != OldRHS) {
743 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
744 if (BO && !NotRewritable.count(BO))
745 NodesToRewrite.push_back(BO);
746 Op->setOperand(1, NewRHS);
748 DEBUG(dbgs() << "TO: " << *Op << '\n');
750 ExpressionChanged = Op;
757 // Not the last operation. The left-hand side will be a sub-expression
758 // while the right-hand side will be the current element of Ops.
759 Value *NewRHS = Ops[i].Op;
760 if (NewRHS != Op->getOperand(1)) {
761 DEBUG(dbgs() << "RA: " << *Op << '\n');
762 if (NewRHS == Op->getOperand(0)) {
763 // The new right-hand side was already present as the left operand. If
764 // we are lucky then swapping the operands will sort out both of them.
767 // Overwrite with the new right-hand side.
768 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
769 if (BO && !NotRewritable.count(BO))
770 NodesToRewrite.push_back(BO);
771 Op->setOperand(1, NewRHS);
772 ExpressionChanged = Op;
774 DEBUG(dbgs() << "TO: " << *Op << '\n');
779 // Now deal with the left-hand side. If this is already an operation node
780 // from the original expression then just rewrite the rest of the expression
782 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
783 if (BO && !NotRewritable.count(BO)) {
788 // Otherwise, grab a spare node from the original expression and use that as
789 // the left-hand side. If there are no nodes left then the optimizers made
790 // an expression with more nodes than the original! This usually means that
791 // they did something stupid but it might mean that the problem was just too
792 // hard (finding the mimimal number of multiplications needed to realize a
793 // multiplication expression is NP-complete). Whatever the reason, smart or
794 // stupid, create a new node if there are none left.
795 BinaryOperator *NewOp;
796 if (NodesToRewrite.empty()) {
797 Constant *Undef = UndefValue::get(I->getType());
798 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
799 Undef, Undef, "", I);
801 NewOp = NodesToRewrite.pop_back_val();
804 DEBUG(dbgs() << "RA: " << *Op << '\n');
805 Op->setOperand(0, NewOp);
806 DEBUG(dbgs() << "TO: " << *Op << '\n');
807 ExpressionChanged = Op;
813 // If the expression changed non-trivially then clear out all subclass data
814 // starting from the operator specified in ExpressionChanged, and compactify
815 // the operators to just before the expression root to guarantee that the
816 // expression tree is dominated by all of Ops.
817 if (ExpressionChanged)
819 ExpressionChanged->clearSubclassOptionalData();
820 if (ExpressionChanged == I)
822 ExpressionChanged->moveBefore(I);
823 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->use_begin());
826 // Throw away any left over nodes from the original expression.
827 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
828 RedoInsts.insert(NodesToRewrite[i]);
831 /// NegateValue - Insert instructions before the instruction pointed to by BI,
832 /// that computes the negative version of the value specified. The negative
833 /// version of the value is returned, and BI is left pointing at the instruction
834 /// that should be processed next by the reassociation pass.
835 static Value *NegateValue(Value *V, Instruction *BI) {
836 if (Constant *C = dyn_cast<Constant>(V))
837 return ConstantExpr::getNeg(C);
839 // We are trying to expose opportunity for reassociation. One of the things
840 // that we want to do to achieve this is to push a negation as deep into an
841 // expression chain as possible, to expose the add instructions. In practice,
842 // this means that we turn this:
843 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
844 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
845 // the constants. We assume that instcombine will clean up the mess later if
846 // we introduce tons of unnecessary negation instructions.
848 if (BinaryOperator *I = isReassociableOp(V, Instruction::Add)) {
849 // Push the negates through the add.
850 I->setOperand(0, NegateValue(I->getOperand(0), BI));
851 I->setOperand(1, NegateValue(I->getOperand(1), BI));
853 // We must move the add instruction here, because the neg instructions do
854 // not dominate the old add instruction in general. By moving it, we are
855 // assured that the neg instructions we just inserted dominate the
856 // instruction we are about to insert after them.
859 I->setName(I->getName()+".neg");
863 // Okay, we need to materialize a negated version of V with an instruction.
864 // Scan the use lists of V to see if we have one already.
865 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
867 if (!BinaryOperator::isNeg(U)) continue;
869 // We found one! Now we have to make sure that the definition dominates
870 // this use. We do this by moving it to the entry block (if it is a
871 // non-instruction value) or right after the definition. These negates will
872 // be zapped by reassociate later, so we don't need much finesse here.
873 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
875 // Verify that the negate is in this function, V might be a constant expr.
876 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
879 BasicBlock::iterator InsertPt;
880 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
881 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
882 InsertPt = II->getNormalDest()->begin();
884 InsertPt = InstInput;
887 while (isa<PHINode>(InsertPt)) ++InsertPt;
889 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
891 TheNeg->moveBefore(InsertPt);
895 // Insert a 'neg' instruction that subtracts the value from zero to get the
897 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
900 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
901 /// X-Y into (X + -Y).
902 static bool ShouldBreakUpSubtract(Instruction *Sub) {
903 // If this is a negation, we can't split it up!
904 if (BinaryOperator::isNeg(Sub))
907 // Don't bother to break this up unless either the LHS is an associable add or
908 // subtract or if this is only used by one.
909 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
910 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
912 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
913 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
915 if (Sub->hasOneUse() &&
916 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
917 isReassociableOp(Sub->use_back(), Instruction::Sub)))
923 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
924 /// only used by an add, transform this into (X+(0-Y)) to promote better
926 static BinaryOperator *BreakUpSubtract(Instruction *Sub) {
927 // Convert a subtract into an add and a neg instruction. This allows sub
928 // instructions to be commuted with other add instructions.
930 // Calculate the negative value of Operand 1 of the sub instruction,
931 // and set it as the RHS of the add instruction we just made.
933 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
934 BinaryOperator *New =
935 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
936 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
937 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
940 // Everyone now refers to the add instruction.
941 Sub->replaceAllUsesWith(New);
942 New->setDebugLoc(Sub->getDebugLoc());
944 DEBUG(dbgs() << "Negated: " << *New << '\n');
948 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
949 /// by one, change this into a multiply by a constant to assist with further
951 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
952 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
953 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
955 BinaryOperator *Mul =
956 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
957 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
959 Shl->replaceAllUsesWith(Mul);
960 Mul->setDebugLoc(Shl->getDebugLoc());
964 /// FindInOperandList - Scan backwards and forwards among values with the same
965 /// rank as element i to see if X exists. If X does not exist, return i. This
966 /// is useful when scanning for 'x' when we see '-x' because they both get the
968 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
970 unsigned XRank = Ops[i].Rank;
971 unsigned e = Ops.size();
972 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
976 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
982 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
983 /// and returning the result. Insert the tree before I.
984 static Value *EmitAddTreeOfValues(Instruction *I,
985 SmallVectorImpl<WeakVH> &Ops){
986 if (Ops.size() == 1) return Ops.back();
988 Value *V1 = Ops.back();
990 Value *V2 = EmitAddTreeOfValues(I, Ops);
991 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
994 /// RemoveFactorFromExpression - If V is an expression tree that is a
995 /// multiplication sequence, and if this sequence contains a multiply by Factor,
996 /// remove Factor from the tree and return the new tree.
997 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
998 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
1001 SmallVector<RepeatedValue, 8> Tree;
1002 MadeChange |= LinearizeExprTree(BO, Tree);
1003 SmallVector<ValueEntry, 8> Factors;
1004 Factors.reserve(Tree.size());
1005 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1006 RepeatedValue E = Tree[i];
1007 Factors.append(E.second.getZExtValue(),
1008 ValueEntry(getRank(E.first), E.first));
1011 bool FoundFactor = false;
1012 bool NeedsNegate = false;
1013 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1014 if (Factors[i].Op == Factor) {
1016 Factors.erase(Factors.begin()+i);
1020 // If this is a negative version of this factor, remove it.
1021 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor))
1022 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1023 if (FC1->getValue() == -FC2->getValue()) {
1024 FoundFactor = NeedsNegate = true;
1025 Factors.erase(Factors.begin()+i);
1031 // Make sure to restore the operands to the expression tree.
1032 RewriteExprTree(BO, Factors);
1036 BasicBlock::iterator InsertPt = BO; ++InsertPt;
1038 // If this was just a single multiply, remove the multiply and return the only
1039 // remaining operand.
1040 if (Factors.size() == 1) {
1041 RedoInsts.insert(BO);
1044 RewriteExprTree(BO, Factors);
1049 V = BinaryOperator::CreateNeg(V, "neg", InsertPt);
1054 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
1055 /// add its operands as factors, otherwise add V to the list of factors.
1057 /// Ops is the top-level list of add operands we're trying to factor.
1058 static void FindSingleUseMultiplyFactors(Value *V,
1059 SmallVectorImpl<Value*> &Factors,
1060 const SmallVectorImpl<ValueEntry> &Ops) {
1061 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
1063 Factors.push_back(V);
1067 // Otherwise, add the LHS and RHS to the list of factors.
1068 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
1069 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
1072 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
1073 /// instruction. This optimizes based on identities. If it can be reduced to
1074 /// a single Value, it is returned, otherwise the Ops list is mutated as
1076 static Value *OptimizeAndOrXor(unsigned Opcode,
1077 SmallVectorImpl<ValueEntry> &Ops) {
1078 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1079 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1080 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1081 // First, check for X and ~X in the operand list.
1082 assert(i < Ops.size());
1083 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1084 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1085 unsigned FoundX = FindInOperandList(Ops, i, X);
1087 if (Opcode == Instruction::And) // ...&X&~X = 0
1088 return Constant::getNullValue(X->getType());
1090 if (Opcode == Instruction::Or) // ...|X|~X = -1
1091 return Constant::getAllOnesValue(X->getType());
1095 // Next, check for duplicate pairs of values, which we assume are next to
1096 // each other, due to our sorting criteria.
1097 assert(i < Ops.size());
1098 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1099 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1100 // Drop duplicate values for And and Or.
1101 Ops.erase(Ops.begin()+i);
1107 // Drop pairs of values for Xor.
1108 assert(Opcode == Instruction::Xor);
1110 return Constant::getNullValue(Ops[0].Op->getType());
1113 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1121 /// Helper funciton of CombineXorOpnd(). It creates a bitwise-and
1122 /// instruction with the given two operands, and return the resulting
1123 /// instruction. There are two special cases: 1) if the constant operand is 0,
1124 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1126 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1127 const APInt &ConstOpnd) {
1128 if (ConstOpnd != 0) {
1129 if (!ConstOpnd.isAllOnesValue()) {
1130 LLVMContext &Ctx = Opnd->getType()->getContext();
1132 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1133 "and.ra", InsertBefore);
1134 I->setDebugLoc(InsertBefore->getDebugLoc());
1142 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1143 // into "R ^ C", where C would be 0, and R is a symbolic value.
1145 // If it was successful, true is returned, and the "R" and "C" is returned
1146 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1147 // and both "Res" and "ConstOpnd" remain unchanged.
1149 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1150 APInt &ConstOpnd, Value *&Res) {
1151 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1152 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1153 // = (x & ~c1) ^ (c1 ^ c2)
1154 // It is useful only when c1 == c2.
1155 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) {
1156 if (!Opnd1->getValue()->hasOneUse())
1159 const APInt &C1 = Opnd1->getConstPart();
1160 if (C1 != ConstOpnd)
1163 Value *X = Opnd1->getSymbolicPart();
1164 Res = createAndInstr(I, X, ~C1);
1165 // ConstOpnd was C2, now C1 ^ C2.
1168 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1169 RedoInsts.insert(T);
1176 // Helper function of OptimizeXor(). It tries to simplify
1177 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1180 // If it was successful, true is returned, and the "R" and "C" is returned
1181 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1182 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1183 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1184 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2,
1185 APInt &ConstOpnd, Value *&Res) {
1186 Value *X = Opnd1->getSymbolicPart();
1187 if (X != Opnd2->getSymbolicPart())
1190 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1191 int DeadInstNum = 1;
1192 if (Opnd1->getValue()->hasOneUse())
1194 if (Opnd2->getValue()->hasOneUse())
1198 // (x | c1) ^ (x & c2)
1199 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1200 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1201 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1203 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1204 if (Opnd2->isOrExpr())
1205 std::swap(Opnd1, Opnd2);
1207 const APInt &C1 = Opnd1->getConstPart();
1208 const APInt &C2 = Opnd2->getConstPart();
1209 APInt C3((~C1) ^ C2);
1211 // Do not increase code size!
1212 if (C3 != 0 && !C3.isAllOnesValue()) {
1213 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1214 if (NewInstNum > DeadInstNum)
1218 Res = createAndInstr(I, X, C3);
1221 } else if (Opnd1->isOrExpr()) {
1222 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1224 const APInt &C1 = Opnd1->getConstPart();
1225 const APInt &C2 = Opnd2->getConstPart();
1228 // Do not increase code size
1229 if (C3 != 0 && !C3.isAllOnesValue()) {
1230 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1231 if (NewInstNum > DeadInstNum)
1235 Res = createAndInstr(I, X, C3);
1238 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1240 const APInt &C1 = Opnd1->getConstPart();
1241 const APInt &C2 = Opnd2->getConstPart();
1243 Res = createAndInstr(I, X, C3);
1246 // Put the original operands in the Redo list; hope they will be deleted
1248 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1249 RedoInsts.insert(T);
1250 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1251 RedoInsts.insert(T);
1256 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1257 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1259 Value *Reassociate::OptimizeXor(Instruction *I,
1260 SmallVectorImpl<ValueEntry> &Ops) {
1261 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1264 if (Ops.size() == 1)
1267 SmallVector<XorOpnd, 8> Opnds;
1268 SmallVector<XorOpnd*, 8> OpndPtrs;
1269 Type *Ty = Ops[0].Op->getType();
1270 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1272 // Step 1: Convert ValueEntry to XorOpnd
1273 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1274 Value *V = Ops[i].Op;
1275 if (!isa<ConstantInt>(V)) {
1277 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1280 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1283 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1284 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1285 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1286 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1287 // when new elements are added to the vector.
1288 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1289 OpndPtrs.push_back(&Opnds[i]);
1291 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1292 // the same symbolic value cluster together. For instance, the input operand
1293 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1294 // ("x | 123", "x & 789", "y & 456").
1295 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor());
1297 // Step 3: Combine adjacent operands
1298 XorOpnd *PrevOpnd = 0;
1299 bool Changed = false;
1300 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1301 XorOpnd *CurrOpnd = OpndPtrs[i];
1302 // The combined value
1305 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1306 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1309 *CurrOpnd = XorOpnd(CV);
1311 CurrOpnd->Invalidate();
1316 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1317 PrevOpnd = CurrOpnd;
1321 // step 3.2: When previous and current operands share the same symbolic
1322 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1324 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1325 // Remove previous operand
1326 PrevOpnd->Invalidate();
1328 *CurrOpnd = XorOpnd(CV);
1329 PrevOpnd = CurrOpnd;
1331 CurrOpnd->Invalidate();
1338 // Step 4: Reassemble the Ops
1341 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1342 XorOpnd &O = Opnds[i];
1345 ValueEntry VE(getRank(O.getValue()), O.getValue());
1348 if (ConstOpnd != 0) {
1349 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1350 ValueEntry VE(getRank(C), C);
1353 int Sz = Ops.size();
1355 return Ops.back().Op;
1357 assert(ConstOpnd == 0);
1358 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1365 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
1366 /// optimizes based on identities. If it can be reduced to a single Value, it
1367 /// is returned, otherwise the Ops list is mutated as necessary.
1368 Value *Reassociate::OptimizeAdd(Instruction *I,
1369 SmallVectorImpl<ValueEntry> &Ops) {
1370 // Scan the operand lists looking for X and -X pairs. If we find any, we
1371 // can simplify the expression. X+-X == 0. While we're at it, scan for any
1372 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1374 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1".
1376 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1377 Value *TheOp = Ops[i].Op;
1378 // Check to see if we've seen this operand before. If so, we factor all
1379 // instances of the operand together. Due to our sorting criteria, we know
1380 // that these need to be next to each other in the vector.
1381 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1382 // Rescan the list, remove all instances of this operand from the expr.
1383 unsigned NumFound = 0;
1385 Ops.erase(Ops.begin()+i);
1387 } while (i != Ops.size() && Ops[i].Op == TheOp);
1389 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1392 // Insert a new multiply.
1393 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound);
1394 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I);
1396 // Now that we have inserted a multiply, optimize it. This allows us to
1397 // handle cases that require multiple factoring steps, such as this:
1398 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1399 RedoInsts.insert(cast<Instruction>(Mul));
1401 // If every add operand was a duplicate, return the multiply.
1405 // Otherwise, we had some input that didn't have the dupe, such as
1406 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1407 // things being added by this operation.
1408 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1415 // Check for X and -X in the operand list.
1416 if (!BinaryOperator::isNeg(TheOp))
1419 Value *X = BinaryOperator::getNegArgument(TheOp);
1420 unsigned FoundX = FindInOperandList(Ops, i, X);
1424 // Remove X and -X from the operand list.
1425 if (Ops.size() == 2)
1426 return Constant::getNullValue(X->getType());
1428 Ops.erase(Ops.begin()+i);
1432 --i; // Need to back up an extra one.
1433 Ops.erase(Ops.begin()+FoundX);
1435 --i; // Revisit element.
1436 e -= 2; // Removed two elements.
1439 // Scan the operand list, checking to see if there are any common factors
1440 // between operands. Consider something like A*A+A*B*C+D. We would like to
1441 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1442 // To efficiently find this, we count the number of times a factor occurs
1443 // for any ADD operands that are MULs.
1444 DenseMap<Value*, unsigned> FactorOccurrences;
1446 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1447 // where they are actually the same multiply.
1448 unsigned MaxOcc = 0;
1449 Value *MaxOccVal = 0;
1450 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1451 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul);
1455 // Compute all of the factors of this added value.
1456 SmallVector<Value*, 8> Factors;
1457 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1458 assert(Factors.size() > 1 && "Bad linearize!");
1460 // Add one to FactorOccurrences for each unique factor in this op.
1461 SmallPtrSet<Value*, 8> Duplicates;
1462 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1463 Value *Factor = Factors[i];
1464 if (!Duplicates.insert(Factor)) continue;
1466 unsigned Occ = ++FactorOccurrences[Factor];
1467 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
1469 // If Factor is a negative constant, add the negated value as a factor
1470 // because we can percolate the negate out. Watch for minint, which
1471 // cannot be positivified.
1472 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor))
1473 if (CI->isNegative() && !CI->isMinValue(true)) {
1474 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1475 assert(!Duplicates.count(Factor) &&
1476 "Shouldn't have two constant factors, missed a canonicalize");
1478 unsigned Occ = ++FactorOccurrences[Factor];
1479 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
1484 // If any factor occurred more than one time, we can pull it out.
1486 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1489 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1490 // this, we could otherwise run into situations where removing a factor
1491 // from an expression will drop a use of maxocc, and this can cause
1492 // RemoveFactorFromExpression on successive values to behave differently.
1493 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
1494 SmallVector<WeakVH, 4> NewMulOps;
1495 for (unsigned i = 0; i != Ops.size(); ++i) {
1496 // Only try to remove factors from expressions we're allowed to.
1497 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul);
1501 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1502 // The factorized operand may occur several times. Convert them all in
1504 for (unsigned j = Ops.size(); j != i;) {
1506 if (Ops[j].Op == Ops[i].Op) {
1507 NewMulOps.push_back(V);
1508 Ops.erase(Ops.begin()+j);
1515 // No need for extra uses anymore.
1518 unsigned NumAddedValues = NewMulOps.size();
1519 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1521 // Now that we have inserted the add tree, optimize it. This allows us to
1522 // handle cases that require multiple factoring steps, such as this:
1523 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1524 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1525 (void)NumAddedValues;
1526 if (Instruction *VI = dyn_cast<Instruction>(V))
1527 RedoInsts.insert(VI);
1529 // Create the multiply.
1530 Instruction *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
1532 // Rerun associate on the multiply in case the inner expression turned into
1533 // a multiply. We want to make sure that we keep things in canonical form.
1534 RedoInsts.insert(V2);
1536 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1537 // entire result expression is just the multiply "A*(B+C)".
1541 // Otherwise, we had some input that didn't have the factor, such as
1542 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1543 // things being added by this operation.
1544 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1550 /// \brief Build up a vector of value/power pairs factoring a product.
1552 /// Given a series of multiplication operands, build a vector of factors and
1553 /// the powers each is raised to when forming the final product. Sort them in
1554 /// the order of descending power.
1556 /// (x*x) -> [(x, 2)]
1557 /// ((x*x)*x) -> [(x, 3)]
1558 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1560 /// \returns Whether any factors have a power greater than one.
1561 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1562 SmallVectorImpl<Factor> &Factors) {
1563 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1564 // Compute the sum of powers of simplifiable factors.
1565 unsigned FactorPowerSum = 0;
1566 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1567 Value *Op = Ops[Idx-1].Op;
1569 // Count the number of occurrences of this value.
1571 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1573 // Track for simplification all factors which occur 2 or more times.
1575 FactorPowerSum += Count;
1578 // We can only simplify factors if the sum of the powers of our simplifiable
1579 // factors is 4 or higher. When that is the case, we will *always* have
1580 // a simplification. This is an important invariant to prevent cyclicly
1581 // trying to simplify already minimal formations.
1582 if (FactorPowerSum < 4)
1585 // Now gather the simplifiable factors, removing them from Ops.
1587 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1588 Value *Op = Ops[Idx-1].Op;
1590 // Count the number of occurrences of this value.
1592 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1596 // Move an even number of occurrences to Factors.
1599 FactorPowerSum += Count;
1600 Factors.push_back(Factor(Op, Count));
1601 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1604 // None of the adjustments above should have reduced the sum of factor powers
1605 // below our mininum of '4'.
1606 assert(FactorPowerSum >= 4);
1608 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1612 /// \brief Build a tree of multiplies, computing the product of Ops.
1613 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1614 SmallVectorImpl<Value*> &Ops) {
1615 if (Ops.size() == 1)
1618 Value *LHS = Ops.pop_back_val();
1620 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1621 } while (!Ops.empty());
1626 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1628 /// Given a vector of values raised to various powers, where no two values are
1629 /// equal and the powers are sorted in decreasing order, compute the minimal
1630 /// DAG of multiplies to compute the final product, and return that product
1632 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1633 SmallVectorImpl<Factor> &Factors) {
1634 assert(Factors[0].Power);
1635 SmallVector<Value *, 4> OuterProduct;
1636 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1637 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1638 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1643 // We want to multiply across all the factors with the same power so that
1644 // we can raise them to that power as a single entity. Build a mini tree
1646 SmallVector<Value *, 4> InnerProduct;
1647 InnerProduct.push_back(Factors[LastIdx].Base);
1649 InnerProduct.push_back(Factors[Idx].Base);
1651 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1653 // Reset the base value of the first factor to the new expression tree.
1654 // We'll remove all the factors with the same power in a second pass.
1655 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1656 if (Instruction *MI = dyn_cast<Instruction>(M))
1657 RedoInsts.insert(MI);
1661 // Unique factors with equal powers -- we've folded them into the first one's
1663 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1664 Factor::PowerEqual()),
1667 // Iteratively collect the base of each factor with an add power into the
1668 // outer product, and halve each power in preparation for squaring the
1670 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1671 if (Factors[Idx].Power & 1)
1672 OuterProduct.push_back(Factors[Idx].Base);
1673 Factors[Idx].Power >>= 1;
1675 if (Factors[0].Power) {
1676 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1677 OuterProduct.push_back(SquareRoot);
1678 OuterProduct.push_back(SquareRoot);
1680 if (OuterProduct.size() == 1)
1681 return OuterProduct.front();
1683 Value *V = buildMultiplyTree(Builder, OuterProduct);
1687 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1688 SmallVectorImpl<ValueEntry> &Ops) {
1689 // We can only optimize the multiplies when there is a chain of more than
1690 // three, such that a balanced tree might require fewer total multiplies.
1694 // Try to turn linear trees of multiplies without other uses of the
1695 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1697 SmallVector<Factor, 4> Factors;
1698 if (!collectMultiplyFactors(Ops, Factors))
1699 return 0; // All distinct factors, so nothing left for us to do.
1701 IRBuilder<> Builder(I);
1702 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1706 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1707 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1711 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1712 SmallVectorImpl<ValueEntry> &Ops) {
1713 // Now that we have the linearized expression tree, try to optimize it.
1714 // Start by folding any constants that we found.
1716 unsigned Opcode = I->getOpcode();
1717 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1718 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1719 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1721 // If there was nothing but constants then we are done.
1725 // Put the combined constant back at the end of the operand list, except if
1726 // there is no point. For example, an add of 0 gets dropped here, while a
1727 // multiplication by zero turns the whole expression into zero.
1728 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1729 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1731 Ops.push_back(ValueEntry(0, Cst));
1734 if (Ops.size() == 1) return Ops[0].Op;
1736 // Handle destructive annihilation due to identities between elements in the
1737 // argument list here.
1738 unsigned NumOps = Ops.size();
1741 case Instruction::And:
1742 case Instruction::Or:
1743 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1747 case Instruction::Xor:
1748 if (Value *Result = OptimizeXor(I, Ops))
1752 case Instruction::Add:
1753 if (Value *Result = OptimizeAdd(I, Ops))
1757 case Instruction::Mul:
1758 if (Value *Result = OptimizeMul(I, Ops))
1763 if (Ops.size() != NumOps)
1764 return OptimizeExpression(I, Ops);
1768 /// EraseInst - Zap the given instruction, adding interesting operands to the
1770 void Reassociate::EraseInst(Instruction *I) {
1771 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1772 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1773 // Erase the dead instruction.
1774 ValueRankMap.erase(I);
1775 RedoInsts.remove(I);
1776 I->eraseFromParent();
1777 // Optimize its operands.
1778 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1779 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1780 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1781 // If this is a node in an expression tree, climb to the expression root
1782 // and add that since that's where optimization actually happens.
1783 unsigned Opcode = Op->getOpcode();
1784 while (Op->hasOneUse() && Op->use_back()->getOpcode() == Opcode &&
1786 Op = Op->use_back();
1787 RedoInsts.insert(Op);
1791 /// OptimizeInst - Inspect and optimize the given instruction. Note that erasing
1792 /// instructions is not allowed.
1793 void Reassociate::OptimizeInst(Instruction *I) {
1794 // Only consider operations that we understand.
1795 if (!isa<BinaryOperator>(I))
1798 if (I->getOpcode() == Instruction::Shl &&
1799 isa<ConstantInt>(I->getOperand(1)))
1800 // If an operand of this shift is a reassociable multiply, or if the shift
1801 // is used by a reassociable multiply or add, turn into a multiply.
1802 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
1804 (isReassociableOp(I->use_back(), Instruction::Mul) ||
1805 isReassociableOp(I->use_back(), Instruction::Add)))) {
1806 Instruction *NI = ConvertShiftToMul(I);
1807 RedoInsts.insert(I);
1812 // Floating point binary operators are not associative, but we can still
1813 // commute (some) of them, to canonicalize the order of their operands.
1814 // This can potentially expose more CSE opportunities, and makes writing
1815 // other transformations simpler.
1816 if ((I->getType()->isFloatingPointTy() || I->getType()->isVectorTy())) {
1817 // FAdd and FMul can be commuted.
1818 if (I->getOpcode() != Instruction::FMul &&
1819 I->getOpcode() != Instruction::FAdd)
1822 Value *LHS = I->getOperand(0);
1823 Value *RHS = I->getOperand(1);
1824 unsigned LHSRank = getRank(LHS);
1825 unsigned RHSRank = getRank(RHS);
1827 // Sort the operands by rank.
1828 if (RHSRank < LHSRank) {
1829 I->setOperand(0, RHS);
1830 I->setOperand(1, LHS);
1836 // Do not reassociate boolean (i1) expressions. We want to preserve the
1837 // original order of evaluation for short-circuited comparisons that
1838 // SimplifyCFG has folded to AND/OR expressions. If the expression
1839 // is not further optimized, it is likely to be transformed back to a
1840 // short-circuited form for code gen, and the source order may have been
1841 // optimized for the most likely conditions.
1842 if (I->getType()->isIntegerTy(1))
1845 // If this is a subtract instruction which is not already in negate form,
1846 // see if we can convert it to X+-Y.
1847 if (I->getOpcode() == Instruction::Sub) {
1848 if (ShouldBreakUpSubtract(I)) {
1849 Instruction *NI = BreakUpSubtract(I);
1850 RedoInsts.insert(I);
1853 } else if (BinaryOperator::isNeg(I)) {
1854 // Otherwise, this is a negation. See if the operand is a multiply tree
1855 // and if this is not an inner node of a multiply tree.
1856 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
1858 !isReassociableOp(I->use_back(), Instruction::Mul))) {
1859 Instruction *NI = LowerNegateToMultiply(I);
1860 RedoInsts.insert(I);
1867 // If this instruction is an associative binary operator, process it.
1868 if (!I->isAssociative()) return;
1869 BinaryOperator *BO = cast<BinaryOperator>(I);
1871 // If this is an interior node of a reassociable tree, ignore it until we
1872 // get to the root of the tree, to avoid N^2 analysis.
1873 unsigned Opcode = BO->getOpcode();
1874 if (BO->hasOneUse() && BO->use_back()->getOpcode() == Opcode)
1877 // If this is an add tree that is used by a sub instruction, ignore it
1878 // until we process the subtract.
1879 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
1880 cast<Instruction>(BO->use_back())->getOpcode() == Instruction::Sub)
1883 ReassociateExpression(BO);
1886 void Reassociate::ReassociateExpression(BinaryOperator *I) {
1888 // First, walk the expression tree, linearizing the tree, collecting the
1889 // operand information.
1890 SmallVector<RepeatedValue, 8> Tree;
1891 MadeChange |= LinearizeExprTree(I, Tree);
1892 SmallVector<ValueEntry, 8> Ops;
1893 Ops.reserve(Tree.size());
1894 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1895 RepeatedValue E = Tree[i];
1896 Ops.append(E.second.getZExtValue(),
1897 ValueEntry(getRank(E.first), E.first));
1900 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
1902 // Now that we have linearized the tree to a list and have gathered all of
1903 // the operands and their ranks, sort the operands by their rank. Use a
1904 // stable_sort so that values with equal ranks will have their relative
1905 // positions maintained (and so the compiler is deterministic). Note that
1906 // this sorts so that the highest ranking values end up at the beginning of
1908 std::stable_sort(Ops.begin(), Ops.end());
1910 // OptimizeExpression - Now that we have the expression tree in a convenient
1911 // sorted form, optimize it globally if possible.
1912 if (Value *V = OptimizeExpression(I, Ops)) {
1914 // Self-referential expression in unreachable code.
1916 // This expression tree simplified to something that isn't a tree,
1918 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
1919 I->replaceAllUsesWith(V);
1920 if (Instruction *VI = dyn_cast<Instruction>(V))
1921 VI->setDebugLoc(I->getDebugLoc());
1922 RedoInsts.insert(I);
1927 // We want to sink immediates as deeply as possible except in the case where
1928 // this is a multiply tree used only by an add, and the immediate is a -1.
1929 // In this case we reassociate to put the negation on the outside so that we
1930 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
1931 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
1932 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
1933 isa<ConstantInt>(Ops.back().Op) &&
1934 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
1935 ValueEntry Tmp = Ops.pop_back_val();
1936 Ops.insert(Ops.begin(), Tmp);
1939 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
1941 if (Ops.size() == 1) {
1943 // Self-referential expression in unreachable code.
1946 // This expression tree simplified to something that isn't a tree,
1948 I->replaceAllUsesWith(Ops[0].Op);
1949 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
1950 OI->setDebugLoc(I->getDebugLoc());
1951 RedoInsts.insert(I);
1955 // Now that we ordered and optimized the expressions, splat them back into
1956 // the expression tree, removing any unneeded nodes.
1957 RewriteExprTree(I, Ops);
1960 bool Reassociate::runOnFunction(Function &F) {
1961 if (skipOptnoneFunction(F))
1964 // Calculate the rank map for F
1968 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
1969 // Optimize every instruction in the basic block.
1970 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; )
1971 if (isInstructionTriviallyDead(II)) {
1975 assert(II->getParent() == BI && "Moved to a different block!");
1979 // If this produced extra instructions to optimize, handle them now.
1980 while (!RedoInsts.empty()) {
1981 Instruction *I = RedoInsts.pop_back_val();
1982 if (isInstructionTriviallyDead(I))
1989 // We are done with the rank map.
1991 ValueRankMap.clear();