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/Transforms/Utils/Local.h"
26 #include "llvm/Constants.h"
27 #include "llvm/DerivedTypes.h"
28 #include "llvm/Function.h"
29 #include "llvm/Instructions.h"
30 #include "llvm/IntrinsicInst.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Assembly/Writer.h"
33 #include "llvm/Support/CFG.h"
34 #include "llvm/Support/IRBuilder.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ValueHandle.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/ADT/DenseMap.h"
39 #include "llvm/ADT/PostOrderIterator.h"
40 #include "llvm/ADT/SmallMap.h"
41 #include "llvm/ADT/STLExtras.h"
42 #include "llvm/ADT/Statistic.h"
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 WriteAsOperand(dbgs(), Ops[i].Op, 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;
116 class Reassociate : public FunctionPass {
117 DenseMap<BasicBlock*, unsigned> RankMap;
118 DenseMap<AssertingVH<Value>, unsigned> ValueRankMap;
119 SmallVector<WeakVH, 8> RedoInsts;
120 SmallVector<WeakVH, 8> DeadInsts;
123 static char ID; // Pass identification, replacement for typeid
124 Reassociate() : FunctionPass(ID) {
125 initializeReassociatePass(*PassRegistry::getPassRegistry());
128 bool runOnFunction(Function &F);
130 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
131 AU.setPreservesCFG();
134 void BuildRankMap(Function &F);
135 unsigned getRank(Value *V);
136 Value *ReassociateExpression(BinaryOperator *I);
137 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
138 Value *OptimizeExpression(BinaryOperator *I,
139 SmallVectorImpl<ValueEntry> &Ops);
140 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
141 bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
142 SmallVectorImpl<Factor> &Factors);
143 Value *buildMinimalMultiplyDAG(IRBuilder<> &Builder,
144 SmallVectorImpl<Factor> &Factors);
145 Value *OptimizeMul(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
146 void LinearizeExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
147 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
148 void ReassociateInst(BasicBlock::iterator &BBI);
150 void RemoveDeadBinaryOp(Value *V);
154 char Reassociate::ID = 0;
155 INITIALIZE_PASS(Reassociate, "reassociate",
156 "Reassociate expressions", false, false)
158 // Public interface to the Reassociate pass
159 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
161 /// isReassociableOp - Return true if V is an instruction of the specified
162 /// opcode and if it only has one use.
163 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
164 if (V->hasOneUse() && isa<Instruction>(V) &&
165 cast<Instruction>(V)->getOpcode() == Opcode)
166 return cast<BinaryOperator>(V);
170 void Reassociate::RemoveDeadBinaryOp(Value *V) {
171 BinaryOperator *Op = dyn_cast<BinaryOperator>(V);
175 ValueRankMap.erase(Op);
176 DeadInsts.push_back(Op);
178 BinaryOperator *LHS = isReassociableOp(Op->getOperand(0), Op->getOpcode());
179 BinaryOperator *RHS = isReassociableOp(Op->getOperand(1), Op->getOpcode());
180 Op->setOperand(0, UndefValue::get(Op->getType()));
181 Op->setOperand(1, UndefValue::get(Op->getType()));
184 RemoveDeadBinaryOp(LHS);
186 RemoveDeadBinaryOp(RHS);
189 static bool isUnmovableInstruction(Instruction *I) {
190 if (I->getOpcode() == Instruction::PHI ||
191 I->getOpcode() == Instruction::LandingPad ||
192 I->getOpcode() == Instruction::Alloca ||
193 I->getOpcode() == Instruction::Load ||
194 I->getOpcode() == Instruction::Invoke ||
195 (I->getOpcode() == Instruction::Call &&
196 !isa<DbgInfoIntrinsic>(I)) ||
197 I->getOpcode() == Instruction::UDiv ||
198 I->getOpcode() == Instruction::SDiv ||
199 I->getOpcode() == Instruction::FDiv ||
200 I->getOpcode() == Instruction::URem ||
201 I->getOpcode() == Instruction::SRem ||
202 I->getOpcode() == Instruction::FRem)
207 void Reassociate::BuildRankMap(Function &F) {
210 // Assign distinct ranks to function arguments
211 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
212 ValueRankMap[&*I] = ++i;
214 ReversePostOrderTraversal<Function*> RPOT(&F);
215 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
216 E = RPOT.end(); I != E; ++I) {
218 unsigned BBRank = RankMap[BB] = ++i << 16;
220 // Walk the basic block, adding precomputed ranks for any instructions that
221 // we cannot move. This ensures that the ranks for these instructions are
222 // all different in the block.
223 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
224 if (isUnmovableInstruction(I))
225 ValueRankMap[&*I] = ++BBRank;
229 unsigned Reassociate::getRank(Value *V) {
230 Instruction *I = dyn_cast<Instruction>(V);
232 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
233 return 0; // Otherwise it's a global or constant, rank 0.
236 if (unsigned Rank = ValueRankMap[I])
237 return Rank; // Rank already known?
239 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
240 // we can reassociate expressions for code motion! Since we do not recurse
241 // for PHI nodes, we cannot have infinite recursion here, because there
242 // cannot be loops in the value graph that do not go through PHI nodes.
243 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
244 for (unsigned i = 0, e = I->getNumOperands();
245 i != e && Rank != MaxRank; ++i)
246 Rank = std::max(Rank, getRank(I->getOperand(i)));
248 // If this is a not or neg instruction, do not count it for rank. This
249 // assures us that X and ~X will have the same rank.
250 if (!I->getType()->isIntegerTy() ||
251 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
254 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = "
257 return ValueRankMap[I] = Rank;
260 /// LowerNegateToMultiply - Replace 0-X with X*-1.
262 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg,
263 DenseMap<AssertingVH<Value>, unsigned> &ValueRankMap) {
264 Constant *Cst = Constant::getAllOnesValue(Neg->getType());
266 BinaryOperator *Res =
267 BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
268 ValueRankMap.erase(Neg);
270 Neg->replaceAllUsesWith(Res);
271 Res->setDebugLoc(Neg->getDebugLoc());
272 Neg->eraseFromParent();
276 /// LinearizeExprTree - Given an associative binary expression, return the leaf
277 /// nodes in Ops. The original expression is the same as Ops[0] op ... Ops[N].
278 /// Note that a node may occur multiple times in Ops, but if so all occurrences
279 /// are consecutive in the vector.
281 /// A leaf node is either not a binary operation of the same kind as the root
282 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
283 /// opcode), or is the same kind of binary operator but has a use which either
284 /// does not belong to the expression, or does belong to the expression but is
285 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
286 /// of the expression, while for non-leaf nodes (except for the root 'I') every
287 /// use is a non-leaf node of the expression.
290 /// expression graph node names
300 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
301 /// that order) C, E, F, F, G, G.
303 /// The expression is maximal: if some instruction is a binary operator of the
304 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
305 /// then the instruction also belongs to the expression, is not a leaf node of
306 /// it, and its operands also belong to the expression (but may be leaf nodes).
308 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
309 /// order to ensure that every non-root node in the expression has *exactly one*
310 /// use by a non-leaf node of the expression. This destruction means that the
311 /// caller MUST either replace 'I' with a new expression or use something like
312 /// RewriteExprTree to put the values back in.
314 /// In the above example either the right operand of A or the left operand of B
315 /// will be replaced by undef. If it is B's operand then this gives:
319 /// + + | A, B - operand of B replaced with undef
325 /// Note that such undef operands can only be reached by passing through 'I'.
326 /// For example, if you visit operands recursively starting from a leaf node
327 /// then you will never see such an undef operand unless you get back to 'I',
328 /// which requires passing through a phi node.
330 /// Note that this routine may also mutate binary operators of the wrong type
331 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
332 /// of the expression) if it can turn them into binary operators of the right
333 /// type and thus make the expression bigger.
335 void Reassociate::LinearizeExprTree(BinaryOperator *I,
336 SmallVectorImpl<ValueEntry> &Ops) {
337 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
339 // Visit all operands of the expression, keeping track of their weight (the
340 // number of paths from the expression root to the operand, or if you like
341 // the number of times that operand occurs in the linearized expression).
342 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
343 // while A has weight two.
345 // Worklist of non-leaf nodes (their operands are in the expression too) along
346 // with their weights, representing a certain number of paths to the operator.
347 // If an operator occurs in the worklist multiple times then we found multiple
348 // ways to get to it.
349 SmallVector<std::pair<BinaryOperator*, unsigned>, 8> Worklist; // (Op, Weight)
350 Worklist.push_back(std::make_pair(I, 1));
351 unsigned Opcode = I->getOpcode();
353 // Leaves of the expression are values that either aren't the right kind of
354 // operation (eg: a constant, or a multiply in an add tree), or are, but have
355 // some uses that are not inside the expression. For example, in I = X + X,
356 // X = A + B, the value X has two uses (by I) that are in the expression. If
357 // X has any other uses, for example in a return instruction, then we consider
358 // X to be a leaf, and won't analyze it further. When we first visit a value,
359 // if it has more than one use then at first we conservatively consider it to
360 // be a leaf. Later, as the expression is explored, we may discover some more
361 // uses of the value from inside the expression. If all uses turn out to be
362 // from within the expression (and the value is a binary operator of the right
363 // kind) then the value is no longer considered to be a leaf, and its operands
366 // Leaves - Keeps track of the set of putative leaves as well as the number of
367 // paths to each leaf seen so far.
368 typedef SmallMap<Value*, unsigned, 8> LeafMap;
369 LeafMap Leaves; // Leaf -> Total weight so far.
370 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
373 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
375 while (!Worklist.empty()) {
376 std::pair<BinaryOperator*, unsigned> P = Worklist.pop_back_val();
377 I = P.first; // We examine the operands of this binary operator.
378 assert(P.second >= 1 && "No paths to here, so how did we get here?!");
380 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
381 Value *Op = I->getOperand(OpIdx);
382 unsigned Weight = P.second; // Number of paths to this operand.
383 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
384 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
386 // If this is a binary operation of the right kind with only one use then
387 // add its operands to the expression.
388 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
389 assert(Visited.insert(Op) && "Not first visit!");
390 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
391 Worklist.push_back(std::make_pair(BO, Weight));
395 // Appears to be a leaf. Is the operand already in the set of leaves?
396 LeafMap::iterator It = Leaves.find(Op);
397 if (It == Leaves.end()) {
398 // Not in the leaf map. Must be the first time we saw this operand.
399 assert(Visited.insert(Op) && "Not first visit!");
400 if (!Op->hasOneUse()) {
401 // This value has uses not accounted for by the expression, so it is
402 // not safe to modify. Mark it as being a leaf.
403 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
404 LeafOrder.push_back(Op);
408 // No uses outside the expression, try morphing it.
409 } else if (It != Leaves.end()) {
410 // Already in the leaf map.
411 assert(Visited.count(Op) && "In leaf map but not visited!");
413 // Update the number of paths to the leaf.
414 It->second += Weight;
416 // The leaf already has one use from inside the expression. As we want
417 // exactly one such use, drop this new use of the leaf.
418 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
419 I->setOperand(OpIdx, UndefValue::get(I->getType()));
422 // If the leaf is a binary operation of the right kind and we now see
423 // that its multiple original uses were in fact all by nodes belonging
424 // to the expression, then no longer consider it to be a leaf and add
425 // its operands to the expression.
426 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
427 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
428 Worklist.push_back(std::make_pair(BO, It->second));
433 // If we still have uses that are not accounted for by the expression
434 // then it is not safe to modify the value.
435 if (!Op->hasOneUse())
438 // No uses outside the expression, try morphing it.
440 Leaves.erase(It); // Since the value may be morphed below.
443 // At this point we have a value which, first of all, is not a binary
444 // expression of the right kind, and secondly, is only used inside the
445 // expression. This means that it can safely be modified. See if we
446 // can usefully morph it into an expression of the right kind.
447 assert((!isa<Instruction>(Op) ||
448 cast<Instruction>(Op)->getOpcode() != Opcode) &&
449 "Should have been handled above!");
450 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
452 // If this is a multiply expression, turn any internal negations into
453 // multiplies by -1 so they can be reassociated.
454 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op);
455 if (Opcode == Instruction::Mul && BO && BinaryOperator::isNeg(BO)) {
456 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
457 BO = LowerNegateToMultiply(BO, ValueRankMap);
458 DEBUG(dbgs() << *BO << 'n');
459 Worklist.push_back(std::make_pair(BO, Weight));
464 // Failed to morph into an expression of the right type. This really is
466 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
467 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
468 LeafOrder.push_back(Op);
473 // The leaves, repeated according to their weights, represent the linearized
474 // form of the expression.
475 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
476 Value *V = LeafOrder[i];
477 LeafMap::iterator It = Leaves.find(V);
478 if (It == Leaves.end())
479 // Leaf already output, or node initially thought to be a leaf wasn't.
481 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
482 unsigned Weight = It->second;
483 assert(Weight > 0 && "No paths to this value!");
484 // FIXME: Rather than repeating values Weight times, use a vector of
485 // (ValueEntry, multiplicity) pairs.
486 Ops.append(Weight, ValueEntry(getRank(V), V));
487 // Ensure the leaf is only output once.
492 // RewriteExprTree - Now that the operands for this expression tree are
493 // linearized and optimized, emit them in-order.
494 void Reassociate::RewriteExprTree(BinaryOperator *I,
495 SmallVectorImpl<ValueEntry> &Ops) {
496 assert(Ops.size() > 1 && "Single values should be used directly!");
498 // Since our optimizations never increase the number of operations, the new
499 // expression can always be written by reusing the existing binary operators
500 // from the original expression tree, without creating any new instructions,
501 // though the rewritten expression may have a completely different topology.
502 // We take care to not change anything if the new expression will be the same
503 // as the original. If more than trivial changes (like commuting operands)
504 // were made then we are obliged to clear out any optional subclass data like
507 /// NodesToRewrite - Nodes from the original expression available for writing
508 /// the new expression into.
509 SmallVector<BinaryOperator*, 8> NodesToRewrite;
510 unsigned Opcode = I->getOpcode();
511 NodesToRewrite.push_back(I);
513 // ExpressionChanged - Non-null if the rewritten expression differs from the
514 // original in some non-trivial way, requiring the clearing of optional flags.
515 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
516 BinaryOperator *ExpressionChanged = 0;
517 BinaryOperator *Previous;
518 BinaryOperator *Op = 0;
519 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
520 assert(!NodesToRewrite.empty() &&
521 "Optimized expressions has more nodes than original!");
522 Previous = Op; Op = NodesToRewrite.pop_back_val();
523 if (ExpressionChanged)
524 // Compactify the tree instructions together with each other to guarantee
525 // that the expression tree is dominated by all of Ops.
526 Op->moveBefore(Previous);
528 // The last operation (which comes earliest in the IR) is special as both
529 // operands will come from Ops, rather than just one with the other being
531 if (i+2 == Ops.size()) {
532 Value *NewLHS = Ops[i].Op;
533 Value *NewRHS = Ops[i+1].Op;
534 Value *OldLHS = Op->getOperand(0);
535 Value *OldRHS = Op->getOperand(1);
537 if (NewLHS == OldLHS && NewRHS == OldRHS)
538 // Nothing changed, leave it alone.
541 if (NewLHS == OldRHS && NewRHS == OldLHS) {
542 // The order of the operands was reversed. Swap them.
543 DEBUG(dbgs() << "RA: " << *Op << '\n');
545 DEBUG(dbgs() << "TO: " << *Op << '\n');
551 // The new operation differs non-trivially from the original. Overwrite
552 // the old operands with the new ones.
553 DEBUG(dbgs() << "RA: " << *Op << '\n');
554 if (NewLHS != OldLHS) {
555 if (BinaryOperator *BO = isReassociableOp(OldLHS, Opcode))
556 NodesToRewrite.push_back(BO);
557 Op->setOperand(0, NewLHS);
559 if (NewRHS != OldRHS) {
560 if (BinaryOperator *BO = isReassociableOp(OldRHS, Opcode))
561 NodesToRewrite.push_back(BO);
562 Op->setOperand(1, NewRHS);
564 DEBUG(dbgs() << "TO: " << *Op << '\n');
566 ExpressionChanged = Op;
573 // Not the last operation. The left-hand side will be a sub-expression
574 // while the right-hand side will be the current element of Ops.
575 Value *NewRHS = Ops[i].Op;
576 if (NewRHS != Op->getOperand(1)) {
577 DEBUG(dbgs() << "RA: " << *Op << '\n');
578 if (NewRHS == Op->getOperand(0)) {
579 // The new right-hand side was already present as the left operand. If
580 // we are lucky then swapping the operands will sort out both of them.
583 // Overwrite with the new right-hand side.
584 if (BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode))
585 NodesToRewrite.push_back(BO);
586 Op->setOperand(1, NewRHS);
587 ExpressionChanged = Op;
589 DEBUG(dbgs() << "TO: " << *Op << '\n');
594 // Now deal with the left-hand side. If this is already an operation node
595 // from the original expression then just rewrite the rest of the expression
597 if (BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode)) {
598 NodesToRewrite.push_back(BO);
602 // Otherwise, grab a spare node from the original expression and use that as
603 // the left-hand side.
604 assert(!NodesToRewrite.empty() &&
605 "Optimized expressions has more nodes than original!");
606 DEBUG(dbgs() << "RA: " << *Op << '\n');
607 Op->setOperand(0, NodesToRewrite.back());
608 DEBUG(dbgs() << "TO: " << *Op << '\n');
609 ExpressionChanged = Op;
614 // If the expression changed non-trivially then clear out all subclass data
615 // starting from the operator specified in ExpressionChanged.
616 if (ExpressionChanged) {
618 ExpressionChanged->clearSubclassOptionalData();
619 if (ExpressionChanged == I)
621 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->use_begin());
625 // Throw away any left over nodes from the original expression.
626 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
627 RemoveDeadBinaryOp(NodesToRewrite[i]);
630 /// NegateValue - Insert instructions before the instruction pointed to by BI,
631 /// that computes the negative version of the value specified. The negative
632 /// version of the value is returned, and BI is left pointing at the instruction
633 /// that should be processed next by the reassociation pass.
634 static Value *NegateValue(Value *V, Instruction *BI) {
635 if (Constant *C = dyn_cast<Constant>(V))
636 return ConstantExpr::getNeg(C);
638 // We are trying to expose opportunity for reassociation. One of the things
639 // that we want to do to achieve this is to push a negation as deep into an
640 // expression chain as possible, to expose the add instructions. In practice,
641 // this means that we turn this:
642 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
643 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
644 // the constants. We assume that instcombine will clean up the mess later if
645 // we introduce tons of unnecessary negation instructions.
647 if (BinaryOperator *I = isReassociableOp(V, Instruction::Add)) {
648 // Push the negates through the add.
649 I->setOperand(0, NegateValue(I->getOperand(0), BI));
650 I->setOperand(1, NegateValue(I->getOperand(1), BI));
652 // We must move the add instruction here, because the neg instructions do
653 // not dominate the old add instruction in general. By moving it, we are
654 // assured that the neg instructions we just inserted dominate the
655 // instruction we are about to insert after them.
658 I->setName(I->getName()+".neg");
662 // Okay, we need to materialize a negated version of V with an instruction.
663 // Scan the use lists of V to see if we have one already.
664 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
666 if (!BinaryOperator::isNeg(U)) continue;
668 // We found one! Now we have to make sure that the definition dominates
669 // this use. We do this by moving it to the entry block (if it is a
670 // non-instruction value) or right after the definition. These negates will
671 // be zapped by reassociate later, so we don't need much finesse here.
672 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
674 // Verify that the negate is in this function, V might be a constant expr.
675 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
678 BasicBlock::iterator InsertPt;
679 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
680 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
681 InsertPt = II->getNormalDest()->begin();
683 InsertPt = InstInput;
686 while (isa<PHINode>(InsertPt)) ++InsertPt;
688 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
690 TheNeg->moveBefore(InsertPt);
694 // Insert a 'neg' instruction that subtracts the value from zero to get the
696 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
699 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
700 /// X-Y into (X + -Y).
701 static bool ShouldBreakUpSubtract(Instruction *Sub) {
702 // If this is a negation, we can't split it up!
703 if (BinaryOperator::isNeg(Sub))
706 // Don't bother to break this up unless either the LHS is an associable add or
707 // subtract or if this is only used by one.
708 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
709 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
711 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
712 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
714 if (Sub->hasOneUse() &&
715 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
716 isReassociableOp(Sub->use_back(), Instruction::Sub)))
722 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
723 /// only used by an add, transform this into (X+(0-Y)) to promote better
725 static Instruction *BreakUpSubtract(Instruction *Sub,
726 DenseMap<AssertingVH<Value>, unsigned> &ValueRankMap) {
727 // Convert a subtract into an add and a neg instruction. This allows sub
728 // instructions to be commuted with other add instructions.
730 // Calculate the negative value of Operand 1 of the sub instruction,
731 // and set it as the RHS of the add instruction we just made.
733 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
735 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
738 // Everyone now refers to the add instruction.
739 ValueRankMap.erase(Sub);
740 Sub->replaceAllUsesWith(New);
741 New->setDebugLoc(Sub->getDebugLoc());
742 Sub->eraseFromParent();
744 DEBUG(dbgs() << "Negated: " << *New << '\n');
748 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
749 /// by one, change this into a multiply by a constant to assist with further
751 static Instruction *ConvertShiftToMul(Instruction *Shl,
752 DenseMap<AssertingVH<Value>, unsigned> &ValueRankMap) {
753 // If an operand of this shift is a reassociable multiply, or if the shift
754 // is used by a reassociable multiply or add, turn into a multiply.
755 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) ||
757 (isReassociableOp(Shl->use_back(), Instruction::Mul) ||
758 isReassociableOp(Shl->use_back(), Instruction::Add)))) {
759 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
760 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
763 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
764 ValueRankMap.erase(Shl);
766 Shl->replaceAllUsesWith(Mul);
767 Mul->setDebugLoc(Shl->getDebugLoc());
768 Shl->eraseFromParent();
774 /// FindInOperandList - Scan backwards and forwards among values with the same
775 /// rank as element i to see if X exists. If X does not exist, return i. This
776 /// is useful when scanning for 'x' when we see '-x' because they both get the
778 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
780 unsigned XRank = Ops[i].Rank;
781 unsigned e = Ops.size();
782 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
786 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
792 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
793 /// and returning the result. Insert the tree before I.
794 static Value *EmitAddTreeOfValues(Instruction *I,
795 SmallVectorImpl<WeakVH> &Ops){
796 if (Ops.size() == 1) return Ops.back();
798 Value *V1 = Ops.back();
800 Value *V2 = EmitAddTreeOfValues(I, Ops);
801 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
804 /// RemoveFactorFromExpression - If V is an expression tree that is a
805 /// multiplication sequence, and if this sequence contains a multiply by Factor,
806 /// remove Factor from the tree and return the new tree.
807 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
808 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
811 SmallVector<ValueEntry, 8> Factors;
812 LinearizeExprTree(BO, Factors);
814 bool FoundFactor = false;
815 bool NeedsNegate = false;
816 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
817 if (Factors[i].Op == Factor) {
819 Factors.erase(Factors.begin()+i);
823 // If this is a negative version of this factor, remove it.
824 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor))
825 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
826 if (FC1->getValue() == -FC2->getValue()) {
827 FoundFactor = NeedsNegate = true;
828 Factors.erase(Factors.begin()+i);
834 // Make sure to restore the operands to the expression tree.
835 RewriteExprTree(BO, Factors);
839 BasicBlock::iterator InsertPt = BO; ++InsertPt;
841 // If this was just a single multiply, remove the multiply and return the only
842 // remaining operand.
843 if (Factors.size() == 1) {
844 RemoveDeadBinaryOp(BO);
847 RewriteExprTree(BO, Factors);
852 V = BinaryOperator::CreateNeg(V, "neg", InsertPt);
857 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
858 /// add its operands as factors, otherwise add V to the list of factors.
860 /// Ops is the top-level list of add operands we're trying to factor.
861 static void FindSingleUseMultiplyFactors(Value *V,
862 SmallVectorImpl<Value*> &Factors,
863 const SmallVectorImpl<ValueEntry> &Ops) {
864 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
866 Factors.push_back(V);
870 // Otherwise, add the LHS and RHS to the list of factors.
871 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
872 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
875 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
876 /// instruction. This optimizes based on identities. If it can be reduced to
877 /// a single Value, it is returned, otherwise the Ops list is mutated as
879 static Value *OptimizeAndOrXor(unsigned Opcode,
880 SmallVectorImpl<ValueEntry> &Ops) {
881 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
882 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
883 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
884 // First, check for X and ~X in the operand list.
885 assert(i < Ops.size());
886 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
887 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
888 unsigned FoundX = FindInOperandList(Ops, i, X);
890 if (Opcode == Instruction::And) // ...&X&~X = 0
891 return Constant::getNullValue(X->getType());
893 if (Opcode == Instruction::Or) // ...|X|~X = -1
894 return Constant::getAllOnesValue(X->getType());
898 // Next, check for duplicate pairs of values, which we assume are next to
899 // each other, due to our sorting criteria.
900 assert(i < Ops.size());
901 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
902 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
903 // Drop duplicate values for And and Or.
904 Ops.erase(Ops.begin()+i);
910 // Drop pairs of values for Xor.
911 assert(Opcode == Instruction::Xor);
913 return Constant::getNullValue(Ops[0].Op->getType());
916 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
924 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
925 /// optimizes based on identities. If it can be reduced to a single Value, it
926 /// is returned, otherwise the Ops list is mutated as necessary.
927 Value *Reassociate::OptimizeAdd(Instruction *I,
928 SmallVectorImpl<ValueEntry> &Ops) {
929 // Scan the operand lists looking for X and -X pairs. If we find any, we
930 // can simplify the expression. X+-X == 0. While we're at it, scan for any
931 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
933 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1".
935 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
936 Value *TheOp = Ops[i].Op;
937 // Check to see if we've seen this operand before. If so, we factor all
938 // instances of the operand together. Due to our sorting criteria, we know
939 // that these need to be next to each other in the vector.
940 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
941 // Rescan the list, remove all instances of this operand from the expr.
942 unsigned NumFound = 0;
944 Ops.erase(Ops.begin()+i);
946 } while (i != Ops.size() && Ops[i].Op == TheOp);
948 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
951 // Insert a new multiply.
952 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound);
953 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I);
955 // Now that we have inserted a multiply, optimize it. This allows us to
956 // handle cases that require multiple factoring steps, such as this:
957 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
958 RedoInsts.push_back(Mul);
960 // If every add operand was a duplicate, return the multiply.
964 // Otherwise, we had some input that didn't have the dupe, such as
965 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
966 // things being added by this operation.
967 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
974 // Check for X and -X in the operand list.
975 if (!BinaryOperator::isNeg(TheOp))
978 Value *X = BinaryOperator::getNegArgument(TheOp);
979 unsigned FoundX = FindInOperandList(Ops, i, X);
983 // Remove X and -X from the operand list.
985 return Constant::getNullValue(X->getType());
987 Ops.erase(Ops.begin()+i);
991 --i; // Need to back up an extra one.
992 Ops.erase(Ops.begin()+FoundX);
994 --i; // Revisit element.
995 e -= 2; // Removed two elements.
998 // Scan the operand list, checking to see if there are any common factors
999 // between operands. Consider something like A*A+A*B*C+D. We would like to
1000 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1001 // To efficiently find this, we count the number of times a factor occurs
1002 // for any ADD operands that are MULs.
1003 DenseMap<Value*, unsigned> FactorOccurrences;
1005 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1006 // where they are actually the same multiply.
1007 unsigned MaxOcc = 0;
1008 Value *MaxOccVal = 0;
1009 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1010 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul);
1014 // Compute all of the factors of this added value.
1015 SmallVector<Value*, 8> Factors;
1016 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1017 assert(Factors.size() > 1 && "Bad linearize!");
1019 // Add one to FactorOccurrences for each unique factor in this op.
1020 SmallPtrSet<Value*, 8> Duplicates;
1021 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1022 Value *Factor = Factors[i];
1023 if (!Duplicates.insert(Factor)) continue;
1025 unsigned Occ = ++FactorOccurrences[Factor];
1026 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
1028 // If Factor is a negative constant, add the negated value as a factor
1029 // because we can percolate the negate out. Watch for minint, which
1030 // cannot be positivified.
1031 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor))
1032 if (CI->isNegative() && !CI->isMinValue(true)) {
1033 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1034 assert(!Duplicates.count(Factor) &&
1035 "Shouldn't have two constant factors, missed a canonicalize");
1037 unsigned Occ = ++FactorOccurrences[Factor];
1038 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
1043 // If any factor occurred more than one time, we can pull it out.
1045 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1048 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1049 // this, we could otherwise run into situations where removing a factor
1050 // from an expression will drop a use of maxocc, and this can cause
1051 // RemoveFactorFromExpression on successive values to behave differently.
1052 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
1053 SmallVector<WeakVH, 4> NewMulOps;
1054 for (unsigned i = 0; i != Ops.size(); ++i) {
1055 // Only try to remove factors from expressions we're allowed to.
1056 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul);
1060 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1061 // The factorized operand may occur several times. Convert them all in
1063 for (unsigned j = Ops.size(); j != i;) {
1065 if (Ops[j].Op == Ops[i].Op) {
1066 NewMulOps.push_back(V);
1067 Ops.erase(Ops.begin()+j);
1074 // No need for extra uses anymore.
1077 unsigned NumAddedValues = NewMulOps.size();
1078 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1080 // Now that we have inserted the add tree, optimize it. This allows us to
1081 // handle cases that require multiple factoring steps, such as this:
1082 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1083 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1084 (void)NumAddedValues;
1085 RedoInsts.push_back(V);
1087 // Create the multiply.
1088 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
1090 // Rerun associate on the multiply in case the inner expression turned into
1091 // a multiply. We want to make sure that we keep things in canonical form.
1092 RedoInsts.push_back(V2);
1094 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1095 // entire result expression is just the multiply "A*(B+C)".
1099 // Otherwise, we had some input that didn't have the factor, such as
1100 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1101 // things being added by this operation.
1102 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1109 /// \brief Predicate tests whether a ValueEntry's op is in a map.
1110 struct IsValueInMap {
1111 const DenseMap<Value *, unsigned> ⤅
1113 IsValueInMap(const DenseMap<Value *, unsigned> &Map) : Map(Map) {}
1115 bool operator()(const ValueEntry &Entry) {
1116 return Map.find(Entry.Op) != Map.end();
1121 /// \brief Build up a vector of value/power pairs factoring a product.
1123 /// Given a series of multiplication operands, build a vector of factors and
1124 /// the powers each is raised to when forming the final product. Sort them in
1125 /// the order of descending power.
1127 /// (x*x) -> [(x, 2)]
1128 /// ((x*x)*x) -> [(x, 3)]
1129 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1131 /// \returns Whether any factors have a power greater than one.
1132 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1133 SmallVectorImpl<Factor> &Factors) {
1134 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1135 // Compute the sum of powers of simplifiable factors.
1136 unsigned FactorPowerSum = 0;
1137 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1138 Value *Op = Ops[Idx-1].Op;
1140 // Count the number of occurrences of this value.
1142 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1144 // Track for simplification all factors which occur 2 or more times.
1146 FactorPowerSum += Count;
1149 // We can only simplify factors if the sum of the powers of our simplifiable
1150 // factors is 4 or higher. When that is the case, we will *always* have
1151 // a simplification. This is an important invariant to prevent cyclicly
1152 // trying to simplify already minimal formations.
1153 if (FactorPowerSum < 4)
1156 // Now gather the simplifiable factors, removing them from Ops.
1158 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1159 Value *Op = Ops[Idx-1].Op;
1161 // Count the number of occurrences of this value.
1163 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1167 // Move an even number of occurrences to Factors.
1170 FactorPowerSum += Count;
1171 Factors.push_back(Factor(Op, Count));
1172 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1175 // None of the adjustments above should have reduced the sum of factor powers
1176 // below our mininum of '4'.
1177 assert(FactorPowerSum >= 4);
1179 std::sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1183 /// \brief Build a tree of multiplies, computing the product of Ops.
1184 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1185 SmallVectorImpl<Value*> &Ops) {
1186 if (Ops.size() == 1)
1189 Value *LHS = Ops.pop_back_val();
1191 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1192 } while (!Ops.empty());
1197 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1199 /// Given a vector of values raised to various powers, where no two values are
1200 /// equal and the powers are sorted in decreasing order, compute the minimal
1201 /// DAG of multiplies to compute the final product, and return that product
1203 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1204 SmallVectorImpl<Factor> &Factors) {
1205 assert(Factors[0].Power);
1206 SmallVector<Value *, 4> OuterProduct;
1207 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1208 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1209 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1214 // We want to multiply across all the factors with the same power so that
1215 // we can raise them to that power as a single entity. Build a mini tree
1217 SmallVector<Value *, 4> InnerProduct;
1218 InnerProduct.push_back(Factors[LastIdx].Base);
1220 InnerProduct.push_back(Factors[Idx].Base);
1222 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1224 // Reset the base value of the first factor to the new expression tree.
1225 // We'll remove all the factors with the same power in a second pass.
1226 Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1227 RedoInsts.push_back(Factors[LastIdx].Base);
1231 // Unique factors with equal powers -- we've folded them into the first one's
1233 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1234 Factor::PowerEqual()),
1237 // Iteratively collect the base of each factor with an add power into the
1238 // outer product, and halve each power in preparation for squaring the
1240 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1241 if (Factors[Idx].Power & 1)
1242 OuterProduct.push_back(Factors[Idx].Base);
1243 Factors[Idx].Power >>= 1;
1245 if (Factors[0].Power) {
1246 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1247 OuterProduct.push_back(SquareRoot);
1248 OuterProduct.push_back(SquareRoot);
1250 if (OuterProduct.size() == 1)
1251 return OuterProduct.front();
1253 Value *V = buildMultiplyTree(Builder, OuterProduct);
1257 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1258 SmallVectorImpl<ValueEntry> &Ops) {
1259 // We can only optimize the multiplies when there is a chain of more than
1260 // three, such that a balanced tree might require fewer total multiplies.
1264 // Try to turn linear trees of multiplies without other uses of the
1265 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1267 SmallVector<Factor, 4> Factors;
1268 if (!collectMultiplyFactors(Ops, Factors))
1269 return 0; // All distinct factors, so nothing left for us to do.
1271 IRBuilder<> Builder(I);
1272 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1276 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1277 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1281 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1282 SmallVectorImpl<ValueEntry> &Ops) {
1283 // Now that we have the linearized expression tree, try to optimize it.
1284 // Start by folding any constants that we found.
1285 bool IterateOptimization = false;
1286 if (Ops.size() == 1) return Ops[0].Op;
1288 unsigned Opcode = I->getOpcode();
1290 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
1291 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
1293 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
1294 return OptimizeExpression(I, Ops);
1297 // Check for destructive annihilation due to a constant being used.
1298 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
1301 case Instruction::And:
1302 if (CstVal->isZero()) // X & 0 -> 0
1304 if (CstVal->isAllOnesValue()) // X & -1 -> X
1307 case Instruction::Mul:
1308 if (CstVal->isZero()) { // X * 0 -> 0
1313 if (cast<ConstantInt>(CstVal)->isOne())
1314 Ops.pop_back(); // X * 1 -> X
1316 case Instruction::Or:
1317 if (CstVal->isAllOnesValue()) // X | -1 -> -1
1320 case Instruction::Add:
1321 case Instruction::Xor:
1322 if (CstVal->isZero()) // X [|^+] 0 -> X
1326 if (Ops.size() == 1) return Ops[0].Op;
1328 // Handle destructive annihilation due to identities between elements in the
1329 // argument list here.
1330 unsigned NumOps = Ops.size();
1333 case Instruction::And:
1334 case Instruction::Or:
1335 case Instruction::Xor:
1336 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1340 case Instruction::Add:
1341 if (Value *Result = OptimizeAdd(I, Ops))
1345 case Instruction::Mul:
1346 if (Value *Result = OptimizeMul(I, Ops))
1351 if (IterateOptimization || Ops.size() != NumOps)
1352 return OptimizeExpression(I, Ops);
1356 /// ReassociateInst - Inspect and reassociate the instruction at the
1357 /// given position, post-incrementing the position.
1358 void Reassociate::ReassociateInst(BasicBlock::iterator &BBI) {
1359 Instruction *BI = BBI++;
1360 if (BI->getOpcode() == Instruction::Shl &&
1361 isa<ConstantInt>(BI->getOperand(1)))
1362 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) {
1367 // Floating point binary operators are not associative, but we can still
1368 // commute (some) of them, to canonicalize the order of their operands.
1369 // This can potentially expose more CSE opportunities, and makes writing
1370 // other transformations simpler.
1371 if (isa<BinaryOperator>(BI) &&
1372 (BI->getType()->isFloatingPointTy() || BI->getType()->isVectorTy())) {
1373 // FAdd and FMul can be commuted.
1374 if (BI->getOpcode() != Instruction::FMul &&
1375 BI->getOpcode() != Instruction::FAdd)
1378 Value *LHS = BI->getOperand(0);
1379 Value *RHS = BI->getOperand(1);
1380 unsigned LHSRank = getRank(LHS);
1381 unsigned RHSRank = getRank(RHS);
1383 // Sort the operands by rank.
1384 if (RHSRank < LHSRank) {
1385 BI->setOperand(0, RHS);
1386 BI->setOperand(1, LHS);
1392 // Do not reassociate operations that we do not understand.
1393 if (!isa<BinaryOperator>(BI))
1396 // Do not reassociate boolean (i1) expressions. We want to preserve the
1397 // original order of evaluation for short-circuited comparisons that
1398 // SimplifyCFG has folded to AND/OR expressions. If the expression
1399 // is not further optimized, it is likely to be transformed back to a
1400 // short-circuited form for code gen, and the source order may have been
1401 // optimized for the most likely conditions.
1402 if (BI->getType()->isIntegerTy(1))
1405 // If this is a subtract instruction which is not already in negate form,
1406 // see if we can convert it to X+-Y.
1407 if (BI->getOpcode() == Instruction::Sub) {
1408 if (ShouldBreakUpSubtract(BI)) {
1409 BI = BreakUpSubtract(BI, ValueRankMap);
1410 // Reset the BBI iterator in case BreakUpSubtract changed the
1411 // instruction it points to.
1415 } else if (BinaryOperator::isNeg(BI)) {
1416 // Otherwise, this is a negation. See if the operand is a multiply tree
1417 // and if this is not an inner node of a multiply tree.
1418 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) &&
1419 (!BI->hasOneUse() ||
1420 !isReassociableOp(BI->use_back(), Instruction::Mul))) {
1421 BI = LowerNegateToMultiply(BI, ValueRankMap);
1427 // If this instruction is a commutative binary operator, process it.
1428 if (!BI->isAssociative()) return;
1429 BinaryOperator *I = cast<BinaryOperator>(BI);
1431 // If this is an interior node of a reassociable tree, ignore it until we
1432 // get to the root of the tree, to avoid N^2 analysis.
1433 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode()))
1436 // If this is an add tree that is used by a sub instruction, ignore it
1437 // until we process the subtract.
1438 if (I->hasOneUse() && I->getOpcode() == Instruction::Add &&
1439 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub)
1442 ReassociateExpression(I);
1445 Value *Reassociate::ReassociateExpression(BinaryOperator *I) {
1447 // First, walk the expression tree, linearizing the tree, collecting the
1448 // operand information.
1449 SmallVector<ValueEntry, 8> Ops;
1450 LinearizeExprTree(I, Ops);
1452 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
1454 // Now that we have linearized the tree to a list and have gathered all of
1455 // the operands and their ranks, sort the operands by their rank. Use a
1456 // stable_sort so that values with equal ranks will have their relative
1457 // positions maintained (and so the compiler is deterministic). Note that
1458 // this sorts so that the highest ranking values end up at the beginning of
1460 std::stable_sort(Ops.begin(), Ops.end());
1462 // OptimizeExpression - Now that we have the expression tree in a convenient
1463 // sorted form, optimize it globally if possible.
1464 if (Value *V = OptimizeExpression(I, Ops)) {
1465 // This expression tree simplified to something that isn't a tree,
1467 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
1468 I->replaceAllUsesWith(V);
1469 if (Instruction *VI = dyn_cast<Instruction>(V))
1470 VI->setDebugLoc(I->getDebugLoc());
1471 RemoveDeadBinaryOp(I);
1476 // We want to sink immediates as deeply as possible except in the case where
1477 // this is a multiply tree used only by an add, and the immediate is a -1.
1478 // In this case we reassociate to put the negation on the outside so that we
1479 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
1480 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
1481 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
1482 isa<ConstantInt>(Ops.back().Op) &&
1483 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
1484 ValueEntry Tmp = Ops.pop_back_val();
1485 Ops.insert(Ops.begin(), Tmp);
1488 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
1490 if (Ops.size() == 1) {
1491 // This expression tree simplified to something that isn't a tree,
1493 I->replaceAllUsesWith(Ops[0].Op);
1494 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
1495 OI->setDebugLoc(I->getDebugLoc());
1496 RemoveDeadBinaryOp(I);
1500 // Now that we ordered and optimized the expressions, splat them back into
1501 // the expression tree, removing any unneeded nodes.
1502 RewriteExprTree(I, Ops);
1506 bool Reassociate::runOnFunction(Function &F) {
1507 // Recalculate the rank map for F
1511 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
1512 for (BasicBlock::iterator BBI = FI->begin(); BBI != FI->end(); )
1513 ReassociateInst(BBI);
1515 // Now that we're done, revisit any instructions which are likely to
1516 // have secondary reassociation opportunities.
1517 while (!RedoInsts.empty())
1518 if (Value *V = RedoInsts.pop_back_val()) {
1519 BasicBlock::iterator BBI = cast<Instruction>(V);
1520 ReassociateInst(BBI);
1523 // We are done with the rank map.
1525 ValueRankMap.clear();
1527 // Now that we're done, delete any instructions which are no longer used.
1528 while (!DeadInsts.empty())
1529 if (Value *V = DeadInsts.pop_back_val())
1530 RecursivelyDeleteTriviallyDeadInstructions(V);