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/SetVector.h"
41 #include "llvm/ADT/SmallMap.h"
42 #include "llvm/ADT/STLExtras.h"
43 #include "llvm/ADT/Statistic.h"
47 STATISTIC(NumChanged, "Number of insts reassociated");
48 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
49 STATISTIC(NumFactor , "Number of multiplies factored");
55 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {}
57 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) {
58 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start.
63 /// PrintOps - Print out the expression identified in the Ops list.
65 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
66 Module *M = I->getParent()->getParent()->getParent();
67 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
68 << *Ops[0].Op->getType() << '\t';
69 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
71 WriteAsOperand(dbgs(), Ops[i].Op, false, M);
72 dbgs() << ", #" << Ops[i].Rank << "] ";
78 /// \brief Utility class representing a base and exponent pair which form one
79 /// factor of some product.
84 Factor(Value *Base, unsigned Power) : Base(Base), Power(Power) {}
86 /// \brief Sort factors by their Base.
88 bool operator()(const Factor &LHS, const Factor &RHS) {
89 return LHS.Base < RHS.Base;
93 /// \brief Compare factors for equal bases.
95 bool operator()(const Factor &LHS, const Factor &RHS) {
96 return LHS.Base == RHS.Base;
100 /// \brief Sort factors in descending order by their power.
101 struct PowerDescendingSorter {
102 bool operator()(const Factor &LHS, const Factor &RHS) {
103 return LHS.Power > RHS.Power;
107 /// \brief Compare factors for equal powers.
109 bool operator()(const Factor &LHS, const Factor &RHS) {
110 return LHS.Power == RHS.Power;
117 class Reassociate : public FunctionPass {
118 DenseMap<BasicBlock*, unsigned> RankMap;
119 DenseMap<AssertingVH<Value>, unsigned> ValueRankMap;
120 SetVector<AssertingVH<Instruction> > RedoInsts;
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 OptimizeInst(Instruction *I);
152 char Reassociate::ID = 0;
153 INITIALIZE_PASS(Reassociate, "reassociate",
154 "Reassociate expressions", false, false)
156 // Public interface to the Reassociate pass
157 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
159 /// isReassociableOp - Return true if V is an instruction of the specified
160 /// opcode and if it only has one use.
161 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
162 if (V->hasOneUse() && isa<Instruction>(V) &&
163 cast<Instruction>(V)->getOpcode() == Opcode)
164 return cast<BinaryOperator>(V);
168 static bool isUnmovableInstruction(Instruction *I) {
169 if (I->getOpcode() == Instruction::PHI ||
170 I->getOpcode() == Instruction::LandingPad ||
171 I->getOpcode() == Instruction::Alloca ||
172 I->getOpcode() == Instruction::Load ||
173 I->getOpcode() == Instruction::Invoke ||
174 (I->getOpcode() == Instruction::Call &&
175 !isa<DbgInfoIntrinsic>(I)) ||
176 I->getOpcode() == Instruction::UDiv ||
177 I->getOpcode() == Instruction::SDiv ||
178 I->getOpcode() == Instruction::FDiv ||
179 I->getOpcode() == Instruction::URem ||
180 I->getOpcode() == Instruction::SRem ||
181 I->getOpcode() == Instruction::FRem)
186 void Reassociate::BuildRankMap(Function &F) {
189 // Assign distinct ranks to function arguments
190 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
191 ValueRankMap[&*I] = ++i;
193 ReversePostOrderTraversal<Function*> RPOT(&F);
194 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
195 E = RPOT.end(); I != E; ++I) {
197 unsigned BBRank = RankMap[BB] = ++i << 16;
199 // Walk the basic block, adding precomputed ranks for any instructions that
200 // we cannot move. This ensures that the ranks for these instructions are
201 // all different in the block.
202 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
203 if (isUnmovableInstruction(I))
204 ValueRankMap[&*I] = ++BBRank;
208 unsigned Reassociate::getRank(Value *V) {
209 Instruction *I = dyn_cast<Instruction>(V);
211 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
212 return 0; // Otherwise it's a global or constant, rank 0.
215 if (unsigned Rank = ValueRankMap[I])
216 return Rank; // Rank already known?
218 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
219 // we can reassociate expressions for code motion! Since we do not recurse
220 // for PHI nodes, we cannot have infinite recursion here, because there
221 // cannot be loops in the value graph that do not go through PHI nodes.
222 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
223 for (unsigned i = 0, e = I->getNumOperands();
224 i != e && Rank != MaxRank; ++i)
225 Rank = std::max(Rank, getRank(I->getOperand(i)));
227 // If this is a not or neg instruction, do not count it for rank. This
228 // assures us that X and ~X will have the same rank.
229 if (!I->getType()->isIntegerTy() ||
230 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
233 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = "
236 return ValueRankMap[I] = Rank;
239 /// LowerNegateToMultiply - Replace 0-X with X*-1.
241 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
242 Constant *Cst = Constant::getAllOnesValue(Neg->getType());
244 BinaryOperator *Res =
245 BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
247 Neg->replaceAllUsesWith(Res);
248 Res->setDebugLoc(Neg->getDebugLoc());
252 /// LinearizeExprTree - Given an associative binary expression, return the leaf
253 /// nodes in Ops. The original expression is the same as Ops[0] op ... Ops[N].
254 /// Note that a node may occur multiple times in Ops, but if so all occurrences
255 /// are consecutive in the vector.
257 /// A leaf node is either not a binary operation of the same kind as the root
258 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
259 /// opcode), or is the same kind of binary operator but has a use which either
260 /// does not belong to the expression, or does belong to the expression but is
261 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
262 /// of the expression, while for non-leaf nodes (except for the root 'I') every
263 /// use is a non-leaf node of the expression.
266 /// expression graph node names
276 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
277 /// that order) C, E, F, F, G, G.
279 /// The expression is maximal: if some instruction is a binary operator of the
280 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
281 /// then the instruction also belongs to the expression, is not a leaf node of
282 /// it, and its operands also belong to the expression (but may be leaf nodes).
284 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
285 /// order to ensure that every non-root node in the expression has *exactly one*
286 /// use by a non-leaf node of the expression. This destruction means that the
287 /// caller MUST either replace 'I' with a new expression or use something like
288 /// RewriteExprTree to put the values back in.
290 /// In the above example either the right operand of A or the left operand of B
291 /// will be replaced by undef. If it is B's operand then this gives:
295 /// + + | A, B - operand of B replaced with undef
301 /// Note that such undef operands can only be reached by passing through 'I'.
302 /// For example, if you visit operands recursively starting from a leaf node
303 /// then you will never see such an undef operand unless you get back to 'I',
304 /// which requires passing through a phi node.
306 /// Note that this routine may also mutate binary operators of the wrong type
307 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
308 /// of the expression) if it can turn them into binary operators of the right
309 /// type and thus make the expression bigger.
311 void Reassociate::LinearizeExprTree(BinaryOperator *I,
312 SmallVectorImpl<ValueEntry> &Ops) {
313 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
315 // Visit all operands of the expression, keeping track of their weight (the
316 // number of paths from the expression root to the operand, or if you like
317 // the number of times that operand occurs in the linearized expression).
318 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
319 // while A has weight two.
321 // Worklist of non-leaf nodes (their operands are in the expression too) along
322 // with their weights, representing a certain number of paths to the operator.
323 // If an operator occurs in the worklist multiple times then we found multiple
324 // ways to get to it.
325 SmallVector<std::pair<BinaryOperator*, unsigned>, 8> Worklist; // (Op, Weight)
326 Worklist.push_back(std::make_pair(I, 1));
327 unsigned Opcode = I->getOpcode();
329 // Leaves of the expression are values that either aren't the right kind of
330 // operation (eg: a constant, or a multiply in an add tree), or are, but have
331 // some uses that are not inside the expression. For example, in I = X + X,
332 // X = A + B, the value X has two uses (by I) that are in the expression. If
333 // X has any other uses, for example in a return instruction, then we consider
334 // X to be a leaf, and won't analyze it further. When we first visit a value,
335 // if it has more than one use then at first we conservatively consider it to
336 // be a leaf. Later, as the expression is explored, we may discover some more
337 // uses of the value from inside the expression. If all uses turn out to be
338 // from within the expression (and the value is a binary operator of the right
339 // kind) then the value is no longer considered to be a leaf, and its operands
342 // Leaves - Keeps track of the set of putative leaves as well as the number of
343 // paths to each leaf seen so far.
344 typedef SmallMap<Value*, unsigned, 8> LeafMap;
345 LeafMap Leaves; // Leaf -> Total weight so far.
346 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
349 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
351 while (!Worklist.empty()) {
352 std::pair<BinaryOperator*, unsigned> P = Worklist.pop_back_val();
353 I = P.first; // We examine the operands of this binary operator.
354 assert(P.second >= 1 && "No paths to here, so how did we get here?!");
356 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
357 Value *Op = I->getOperand(OpIdx);
358 unsigned Weight = P.second; // Number of paths to this operand.
359 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
360 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
362 // If this is a binary operation of the right kind with only one use then
363 // add its operands to the expression.
364 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
365 assert(Visited.insert(Op) && "Not first visit!");
366 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
367 Worklist.push_back(std::make_pair(BO, Weight));
371 // Appears to be a leaf. Is the operand already in the set of leaves?
372 LeafMap::iterator It = Leaves.find(Op);
373 if (It == Leaves.end()) {
374 // Not in the leaf map. Must be the first time we saw this operand.
375 assert(Visited.insert(Op) && "Not first visit!");
376 if (!Op->hasOneUse()) {
377 // This value has uses not accounted for by the expression, so it is
378 // not safe to modify. Mark it as being a leaf.
379 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
380 LeafOrder.push_back(Op);
384 // No uses outside the expression, try morphing it.
385 } else if (It != Leaves.end()) {
386 // Already in the leaf map.
387 assert(Visited.count(Op) && "In leaf map but not visited!");
389 // Update the number of paths to the leaf.
390 It->second += Weight;
392 // The leaf already has one use from inside the expression. As we want
393 // exactly one such use, drop this new use of the leaf.
394 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
395 I->setOperand(OpIdx, UndefValue::get(I->getType()));
398 // If the leaf is a binary operation of the right kind and we now see
399 // that its multiple original uses were in fact all by nodes belonging
400 // to the expression, then no longer consider it to be a leaf and add
401 // its operands to the expression.
402 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
403 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
404 Worklist.push_back(std::make_pair(BO, It->second));
409 // If we still have uses that are not accounted for by the expression
410 // then it is not safe to modify the value.
411 if (!Op->hasOneUse())
414 // No uses outside the expression, try morphing it.
416 Leaves.erase(It); // Since the value may be morphed below.
419 // At this point we have a value which, first of all, is not a binary
420 // expression of the right kind, and secondly, is only used inside the
421 // expression. This means that it can safely be modified. See if we
422 // can usefully morph it into an expression of the right kind.
423 assert((!isa<Instruction>(Op) ||
424 cast<Instruction>(Op)->getOpcode() != Opcode) &&
425 "Should have been handled above!");
426 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
428 // If this is a multiply expression, turn any internal negations into
429 // multiplies by -1 so they can be reassociated.
430 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op);
431 if (Opcode == Instruction::Mul && BO && BinaryOperator::isNeg(BO)) {
432 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
433 BO = LowerNegateToMultiply(BO);
434 DEBUG(dbgs() << *BO << 'n');
435 Worklist.push_back(std::make_pair(BO, Weight));
440 // Failed to morph into an expression of the right type. This really is
442 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
443 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
444 LeafOrder.push_back(Op);
449 // The leaves, repeated according to their weights, represent the linearized
450 // form of the expression.
451 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
452 Value *V = LeafOrder[i];
453 LeafMap::iterator It = Leaves.find(V);
454 if (It == Leaves.end())
455 // Leaf already output, or node initially thought to be a leaf wasn't.
457 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
458 unsigned Weight = It->second;
459 assert(Weight > 0 && "No paths to this value!");
460 // FIXME: Rather than repeating values Weight times, use a vector of
461 // (ValueEntry, multiplicity) pairs.
462 Ops.append(Weight, ValueEntry(getRank(V), V));
463 // Ensure the leaf is only output once.
468 // RewriteExprTree - Now that the operands for this expression tree are
469 // linearized and optimized, emit them in-order.
470 void Reassociate::RewriteExprTree(BinaryOperator *I,
471 SmallVectorImpl<ValueEntry> &Ops) {
472 assert(Ops.size() > 1 && "Single values should be used directly!");
474 // Since our optimizations never increase the number of operations, the new
475 // expression can always be written by reusing the existing binary operators
476 // from the original expression tree, without creating any new instructions,
477 // though the rewritten expression may have a completely different topology.
478 // We take care to not change anything if the new expression will be the same
479 // as the original. If more than trivial changes (like commuting operands)
480 // were made then we are obliged to clear out any optional subclass data like
483 /// NodesToRewrite - Nodes from the original expression available for writing
484 /// the new expression into.
485 SmallVector<BinaryOperator*, 8> NodesToRewrite;
486 unsigned Opcode = I->getOpcode();
487 NodesToRewrite.push_back(I);
489 // ExpressionChanged - Non-null if the rewritten expression differs from the
490 // original in some non-trivial way, requiring the clearing of optional flags.
491 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
492 BinaryOperator *ExpressionChanged = 0;
493 BinaryOperator *Previous;
494 BinaryOperator *Op = 0;
495 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
496 assert(!NodesToRewrite.empty() &&
497 "Optimized expressions has more nodes than original!");
498 Previous = Op; Op = NodesToRewrite.pop_back_val();
499 if (ExpressionChanged)
500 // Compactify the tree instructions together with each other to guarantee
501 // that the expression tree is dominated by all of Ops.
502 Op->moveBefore(Previous);
504 // The last operation (which comes earliest in the IR) is special as both
505 // operands will come from Ops, rather than just one with the other being
507 if (i+2 == Ops.size()) {
508 Value *NewLHS = Ops[i].Op;
509 Value *NewRHS = Ops[i+1].Op;
510 Value *OldLHS = Op->getOperand(0);
511 Value *OldRHS = Op->getOperand(1);
513 if (NewLHS == OldLHS && NewRHS == OldRHS)
514 // Nothing changed, leave it alone.
517 if (NewLHS == OldRHS && NewRHS == OldLHS) {
518 // The order of the operands was reversed. Swap them.
519 DEBUG(dbgs() << "RA: " << *Op << '\n');
521 DEBUG(dbgs() << "TO: " << *Op << '\n');
527 // The new operation differs non-trivially from the original. Overwrite
528 // the old operands with the new ones.
529 DEBUG(dbgs() << "RA: " << *Op << '\n');
530 if (NewLHS != OldLHS) {
531 if (BinaryOperator *BO = isReassociableOp(OldLHS, Opcode))
532 NodesToRewrite.push_back(BO);
533 Op->setOperand(0, NewLHS);
535 if (NewRHS != OldRHS) {
536 if (BinaryOperator *BO = isReassociableOp(OldRHS, Opcode))
537 NodesToRewrite.push_back(BO);
538 Op->setOperand(1, NewRHS);
540 DEBUG(dbgs() << "TO: " << *Op << '\n');
542 ExpressionChanged = Op;
549 // Not the last operation. The left-hand side will be a sub-expression
550 // while the right-hand side will be the current element of Ops.
551 Value *NewRHS = Ops[i].Op;
552 if (NewRHS != Op->getOperand(1)) {
553 DEBUG(dbgs() << "RA: " << *Op << '\n');
554 if (NewRHS == Op->getOperand(0)) {
555 // The new right-hand side was already present as the left operand. If
556 // we are lucky then swapping the operands will sort out both of them.
559 // Overwrite with the new right-hand side.
560 if (BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode))
561 NodesToRewrite.push_back(BO);
562 Op->setOperand(1, NewRHS);
563 ExpressionChanged = Op;
565 DEBUG(dbgs() << "TO: " << *Op << '\n');
570 // Now deal with the left-hand side. If this is already an operation node
571 // from the original expression then just rewrite the rest of the expression
573 if (BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode)) {
574 NodesToRewrite.push_back(BO);
578 // Otherwise, grab a spare node from the original expression and use that as
579 // the left-hand side.
580 assert(!NodesToRewrite.empty() &&
581 "Optimized expressions has more nodes than original!");
582 DEBUG(dbgs() << "RA: " << *Op << '\n');
583 Op->setOperand(0, NodesToRewrite.back());
584 DEBUG(dbgs() << "TO: " << *Op << '\n');
585 ExpressionChanged = Op;
590 // If the expression changed non-trivially then clear out all subclass data
591 // starting from the operator specified in ExpressionChanged.
592 if (ExpressionChanged) {
594 ExpressionChanged->clearSubclassOptionalData();
595 if (ExpressionChanged == I)
597 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->use_begin());
601 // Throw away any left over nodes from the original expression.
602 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
603 RedoInsts.insert(NodesToRewrite[i]);
606 /// NegateValue - Insert instructions before the instruction pointed to by BI,
607 /// that computes the negative version of the value specified. The negative
608 /// version of the value is returned, and BI is left pointing at the instruction
609 /// that should be processed next by the reassociation pass.
610 static Value *NegateValue(Value *V, Instruction *BI) {
611 if (Constant *C = dyn_cast<Constant>(V))
612 return ConstantExpr::getNeg(C);
614 // We are trying to expose opportunity for reassociation. One of the things
615 // that we want to do to achieve this is to push a negation as deep into an
616 // expression chain as possible, to expose the add instructions. In practice,
617 // this means that we turn this:
618 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
619 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
620 // the constants. We assume that instcombine will clean up the mess later if
621 // we introduce tons of unnecessary negation instructions.
623 if (BinaryOperator *I = isReassociableOp(V, Instruction::Add)) {
624 // Push the negates through the add.
625 I->setOperand(0, NegateValue(I->getOperand(0), BI));
626 I->setOperand(1, NegateValue(I->getOperand(1), BI));
628 // We must move the add instruction here, because the neg instructions do
629 // not dominate the old add instruction in general. By moving it, we are
630 // assured that the neg instructions we just inserted dominate the
631 // instruction we are about to insert after them.
634 I->setName(I->getName()+".neg");
638 // Okay, we need to materialize a negated version of V with an instruction.
639 // Scan the use lists of V to see if we have one already.
640 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
642 if (!BinaryOperator::isNeg(U)) continue;
644 // We found one! Now we have to make sure that the definition dominates
645 // this use. We do this by moving it to the entry block (if it is a
646 // non-instruction value) or right after the definition. These negates will
647 // be zapped by reassociate later, so we don't need much finesse here.
648 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
650 // Verify that the negate is in this function, V might be a constant expr.
651 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
654 BasicBlock::iterator InsertPt;
655 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
656 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
657 InsertPt = II->getNormalDest()->begin();
659 InsertPt = InstInput;
662 while (isa<PHINode>(InsertPt)) ++InsertPt;
664 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
666 TheNeg->moveBefore(InsertPt);
670 // Insert a 'neg' instruction that subtracts the value from zero to get the
672 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
675 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
676 /// X-Y into (X + -Y).
677 static bool ShouldBreakUpSubtract(Instruction *Sub) {
678 // If this is a negation, we can't split it up!
679 if (BinaryOperator::isNeg(Sub))
682 // Don't bother to break this up unless either the LHS is an associable add or
683 // subtract or if this is only used by one.
684 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
685 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
687 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
688 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
690 if (Sub->hasOneUse() &&
691 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
692 isReassociableOp(Sub->use_back(), Instruction::Sub)))
698 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
699 /// only used by an add, transform this into (X+(0-Y)) to promote better
701 static BinaryOperator *BreakUpSubtract(Instruction *Sub) {
702 // Convert a subtract into an add and a neg instruction. This allows sub
703 // instructions to be commuted with other add instructions.
705 // Calculate the negative value of Operand 1 of the sub instruction,
706 // and set it as the RHS of the add instruction we just made.
708 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
709 BinaryOperator *New =
710 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
713 // Everyone now refers to the add instruction.
714 Sub->replaceAllUsesWith(New);
715 New->setDebugLoc(Sub->getDebugLoc());
717 DEBUG(dbgs() << "Negated: " << *New << '\n');
721 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
722 /// by one, change this into a multiply by a constant to assist with further
724 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
725 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
726 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
728 BinaryOperator *Mul =
729 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
731 Shl->replaceAllUsesWith(Mul);
732 Mul->setDebugLoc(Shl->getDebugLoc());
736 /// FindInOperandList - Scan backwards and forwards among values with the same
737 /// rank as element i to see if X exists. If X does not exist, return i. This
738 /// is useful when scanning for 'x' when we see '-x' because they both get the
740 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
742 unsigned XRank = Ops[i].Rank;
743 unsigned e = Ops.size();
744 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
748 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
754 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
755 /// and returning the result. Insert the tree before I.
756 static Value *EmitAddTreeOfValues(Instruction *I,
757 SmallVectorImpl<WeakVH> &Ops){
758 if (Ops.size() == 1) return Ops.back();
760 Value *V1 = Ops.back();
762 Value *V2 = EmitAddTreeOfValues(I, Ops);
763 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
766 /// RemoveFactorFromExpression - If V is an expression tree that is a
767 /// multiplication sequence, and if this sequence contains a multiply by Factor,
768 /// remove Factor from the tree and return the new tree.
769 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
770 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
773 SmallVector<ValueEntry, 8> Factors;
774 LinearizeExprTree(BO, Factors);
776 bool FoundFactor = false;
777 bool NeedsNegate = false;
778 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
779 if (Factors[i].Op == Factor) {
781 Factors.erase(Factors.begin()+i);
785 // If this is a negative version of this factor, remove it.
786 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor))
787 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
788 if (FC1->getValue() == -FC2->getValue()) {
789 FoundFactor = NeedsNegate = true;
790 Factors.erase(Factors.begin()+i);
796 // Make sure to restore the operands to the expression tree.
797 RewriteExprTree(BO, Factors);
801 BasicBlock::iterator InsertPt = BO; ++InsertPt;
803 // If this was just a single multiply, remove the multiply and return the only
804 // remaining operand.
805 if (Factors.size() == 1) {
806 RedoInsts.insert(BO);
809 RewriteExprTree(BO, Factors);
814 V = BinaryOperator::CreateNeg(V, "neg", InsertPt);
819 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
820 /// add its operands as factors, otherwise add V to the list of factors.
822 /// Ops is the top-level list of add operands we're trying to factor.
823 static void FindSingleUseMultiplyFactors(Value *V,
824 SmallVectorImpl<Value*> &Factors,
825 const SmallVectorImpl<ValueEntry> &Ops) {
826 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
828 Factors.push_back(V);
832 // Otherwise, add the LHS and RHS to the list of factors.
833 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
834 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
837 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
838 /// instruction. This optimizes based on identities. If it can be reduced to
839 /// a single Value, it is returned, otherwise the Ops list is mutated as
841 static Value *OptimizeAndOrXor(unsigned Opcode,
842 SmallVectorImpl<ValueEntry> &Ops) {
843 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
844 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
845 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
846 // First, check for X and ~X in the operand list.
847 assert(i < Ops.size());
848 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
849 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
850 unsigned FoundX = FindInOperandList(Ops, i, X);
852 if (Opcode == Instruction::And) // ...&X&~X = 0
853 return Constant::getNullValue(X->getType());
855 if (Opcode == Instruction::Or) // ...|X|~X = -1
856 return Constant::getAllOnesValue(X->getType());
860 // Next, check for duplicate pairs of values, which we assume are next to
861 // each other, due to our sorting criteria.
862 assert(i < Ops.size());
863 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
864 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
865 // Drop duplicate values for And and Or.
866 Ops.erase(Ops.begin()+i);
872 // Drop pairs of values for Xor.
873 assert(Opcode == Instruction::Xor);
875 return Constant::getNullValue(Ops[0].Op->getType());
878 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
886 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
887 /// optimizes based on identities. If it can be reduced to a single Value, it
888 /// is returned, otherwise the Ops list is mutated as necessary.
889 Value *Reassociate::OptimizeAdd(Instruction *I,
890 SmallVectorImpl<ValueEntry> &Ops) {
891 // Scan the operand lists looking for X and -X pairs. If we find any, we
892 // can simplify the expression. X+-X == 0. While we're at it, scan for any
893 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
895 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1".
897 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
898 Value *TheOp = Ops[i].Op;
899 // Check to see if we've seen this operand before. If so, we factor all
900 // instances of the operand together. Due to our sorting criteria, we know
901 // that these need to be next to each other in the vector.
902 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
903 // Rescan the list, remove all instances of this operand from the expr.
904 unsigned NumFound = 0;
906 Ops.erase(Ops.begin()+i);
908 } while (i != Ops.size() && Ops[i].Op == TheOp);
910 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
913 // Insert a new multiply.
914 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound);
915 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I);
917 // Now that we have inserted a multiply, optimize it. This allows us to
918 // handle cases that require multiple factoring steps, such as this:
919 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
920 RedoInsts.insert(cast<Instruction>(Mul));
922 // If every add operand was a duplicate, return the multiply.
926 // Otherwise, we had some input that didn't have the dupe, such as
927 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
928 // things being added by this operation.
929 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
936 // Check for X and -X in the operand list.
937 if (!BinaryOperator::isNeg(TheOp))
940 Value *X = BinaryOperator::getNegArgument(TheOp);
941 unsigned FoundX = FindInOperandList(Ops, i, X);
945 // Remove X and -X from the operand list.
947 return Constant::getNullValue(X->getType());
949 Ops.erase(Ops.begin()+i);
953 --i; // Need to back up an extra one.
954 Ops.erase(Ops.begin()+FoundX);
956 --i; // Revisit element.
957 e -= 2; // Removed two elements.
960 // Scan the operand list, checking to see if there are any common factors
961 // between operands. Consider something like A*A+A*B*C+D. We would like to
962 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
963 // To efficiently find this, we count the number of times a factor occurs
964 // for any ADD operands that are MULs.
965 DenseMap<Value*, unsigned> FactorOccurrences;
967 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
968 // where they are actually the same multiply.
970 Value *MaxOccVal = 0;
971 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
972 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul);
976 // Compute all of the factors of this added value.
977 SmallVector<Value*, 8> Factors;
978 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
979 assert(Factors.size() > 1 && "Bad linearize!");
981 // Add one to FactorOccurrences for each unique factor in this op.
982 SmallPtrSet<Value*, 8> Duplicates;
983 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
984 Value *Factor = Factors[i];
985 if (!Duplicates.insert(Factor)) continue;
987 unsigned Occ = ++FactorOccurrences[Factor];
988 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
990 // If Factor is a negative constant, add the negated value as a factor
991 // because we can percolate the negate out. Watch for minint, which
992 // cannot be positivified.
993 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor))
994 if (CI->isNegative() && !CI->isMinValue(true)) {
995 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
996 assert(!Duplicates.count(Factor) &&
997 "Shouldn't have two constant factors, missed a canonicalize");
999 unsigned Occ = ++FactorOccurrences[Factor];
1000 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
1005 // If any factor occurred more than one time, we can pull it out.
1007 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1010 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1011 // this, we could otherwise run into situations where removing a factor
1012 // from an expression will drop a use of maxocc, and this can cause
1013 // RemoveFactorFromExpression on successive values to behave differently.
1014 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
1015 SmallVector<WeakVH, 4> NewMulOps;
1016 for (unsigned i = 0; i != Ops.size(); ++i) {
1017 // Only try to remove factors from expressions we're allowed to.
1018 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul);
1022 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1023 // The factorized operand may occur several times. Convert them all in
1025 for (unsigned j = Ops.size(); j != i;) {
1027 if (Ops[j].Op == Ops[i].Op) {
1028 NewMulOps.push_back(V);
1029 Ops.erase(Ops.begin()+j);
1036 // No need for extra uses anymore.
1039 unsigned NumAddedValues = NewMulOps.size();
1040 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1042 // Now that we have inserted the add tree, optimize it. This allows us to
1043 // handle cases that require multiple factoring steps, such as this:
1044 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1045 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1046 (void)NumAddedValues;
1047 if (Instruction *VI = dyn_cast<Instruction>(V))
1048 RedoInsts.insert(VI);
1050 // Create the multiply.
1051 Instruction *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
1053 // Rerun associate on the multiply in case the inner expression turned into
1054 // a multiply. We want to make sure that we keep things in canonical form.
1055 RedoInsts.insert(V2);
1057 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1058 // entire result expression is just the multiply "A*(B+C)".
1062 // Otherwise, we had some input that didn't have the factor, such as
1063 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1064 // things being added by this operation.
1065 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1072 /// \brief Predicate tests whether a ValueEntry's op is in a map.
1073 struct IsValueInMap {
1074 const DenseMap<Value *, unsigned> ⤅
1076 IsValueInMap(const DenseMap<Value *, unsigned> &Map) : Map(Map) {}
1078 bool operator()(const ValueEntry &Entry) {
1079 return Map.find(Entry.Op) != Map.end();
1084 /// \brief Build up a vector of value/power pairs factoring a product.
1086 /// Given a series of multiplication operands, build a vector of factors and
1087 /// the powers each is raised to when forming the final product. Sort them in
1088 /// the order of descending power.
1090 /// (x*x) -> [(x, 2)]
1091 /// ((x*x)*x) -> [(x, 3)]
1092 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1094 /// \returns Whether any factors have a power greater than one.
1095 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1096 SmallVectorImpl<Factor> &Factors) {
1097 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1098 // Compute the sum of powers of simplifiable factors.
1099 unsigned FactorPowerSum = 0;
1100 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1101 Value *Op = Ops[Idx-1].Op;
1103 // Count the number of occurrences of this value.
1105 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1107 // Track for simplification all factors which occur 2 or more times.
1109 FactorPowerSum += Count;
1112 // We can only simplify factors if the sum of the powers of our simplifiable
1113 // factors is 4 or higher. When that is the case, we will *always* have
1114 // a simplification. This is an important invariant to prevent cyclicly
1115 // trying to simplify already minimal formations.
1116 if (FactorPowerSum < 4)
1119 // Now gather the simplifiable factors, removing them from Ops.
1121 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1122 Value *Op = Ops[Idx-1].Op;
1124 // Count the number of occurrences of this value.
1126 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1130 // Move an even number of occurrences to Factors.
1133 FactorPowerSum += Count;
1134 Factors.push_back(Factor(Op, Count));
1135 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1138 // None of the adjustments above should have reduced the sum of factor powers
1139 // below our mininum of '4'.
1140 assert(FactorPowerSum >= 4);
1142 std::sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1146 /// \brief Build a tree of multiplies, computing the product of Ops.
1147 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1148 SmallVectorImpl<Value*> &Ops) {
1149 if (Ops.size() == 1)
1152 Value *LHS = Ops.pop_back_val();
1154 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1155 } while (!Ops.empty());
1160 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1162 /// Given a vector of values raised to various powers, where no two values are
1163 /// equal and the powers are sorted in decreasing order, compute the minimal
1164 /// DAG of multiplies to compute the final product, and return that product
1166 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1167 SmallVectorImpl<Factor> &Factors) {
1168 assert(Factors[0].Power);
1169 SmallVector<Value *, 4> OuterProduct;
1170 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1171 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1172 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1177 // We want to multiply across all the factors with the same power so that
1178 // we can raise them to that power as a single entity. Build a mini tree
1180 SmallVector<Value *, 4> InnerProduct;
1181 InnerProduct.push_back(Factors[LastIdx].Base);
1183 InnerProduct.push_back(Factors[Idx].Base);
1185 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1187 // Reset the base value of the first factor to the new expression tree.
1188 // We'll remove all the factors with the same power in a second pass.
1189 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1190 if (Instruction *MI = dyn_cast<Instruction>(M))
1191 RedoInsts.insert(MI);
1195 // Unique factors with equal powers -- we've folded them into the first one's
1197 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1198 Factor::PowerEqual()),
1201 // Iteratively collect the base of each factor with an add power into the
1202 // outer product, and halve each power in preparation for squaring the
1204 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1205 if (Factors[Idx].Power & 1)
1206 OuterProduct.push_back(Factors[Idx].Base);
1207 Factors[Idx].Power >>= 1;
1209 if (Factors[0].Power) {
1210 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1211 OuterProduct.push_back(SquareRoot);
1212 OuterProduct.push_back(SquareRoot);
1214 if (OuterProduct.size() == 1)
1215 return OuterProduct.front();
1217 Value *V = buildMultiplyTree(Builder, OuterProduct);
1221 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1222 SmallVectorImpl<ValueEntry> &Ops) {
1223 // We can only optimize the multiplies when there is a chain of more than
1224 // three, such that a balanced tree might require fewer total multiplies.
1228 // Try to turn linear trees of multiplies without other uses of the
1229 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1231 SmallVector<Factor, 4> Factors;
1232 if (!collectMultiplyFactors(Ops, Factors))
1233 return 0; // All distinct factors, so nothing left for us to do.
1235 IRBuilder<> Builder(I);
1236 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1240 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1241 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1245 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1246 SmallVectorImpl<ValueEntry> &Ops) {
1247 // Now that we have the linearized expression tree, try to optimize it.
1248 // Start by folding any constants that we found.
1249 if (Ops.size() == 1) return Ops[0].Op;
1251 unsigned Opcode = I->getOpcode();
1253 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
1254 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
1256 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
1257 return OptimizeExpression(I, Ops);
1260 // Check for destructive annihilation due to a constant being used.
1261 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
1264 case Instruction::And:
1265 if (CstVal->isZero()) // X & 0 -> 0
1267 if (CstVal->isAllOnesValue()) // X & -1 -> X
1270 case Instruction::Mul:
1271 if (CstVal->isZero()) { // X * 0 -> 0
1276 if (cast<ConstantInt>(CstVal)->isOne())
1277 Ops.pop_back(); // X * 1 -> X
1279 case Instruction::Or:
1280 if (CstVal->isAllOnesValue()) // X | -1 -> -1
1283 case Instruction::Add:
1284 case Instruction::Xor:
1285 if (CstVal->isZero()) // X [|^+] 0 -> X
1289 if (Ops.size() == 1) return Ops[0].Op;
1291 // Handle destructive annihilation due to identities between elements in the
1292 // argument list here.
1293 unsigned NumOps = Ops.size();
1296 case Instruction::And:
1297 case Instruction::Or:
1298 case Instruction::Xor:
1299 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1303 case Instruction::Add:
1304 if (Value *Result = OptimizeAdd(I, Ops))
1308 case Instruction::Mul:
1309 if (Value *Result = OptimizeMul(I, Ops))
1314 if (Ops.size() != NumOps)
1315 return OptimizeExpression(I, Ops);
1319 /// OptimizeInst - Inspect and optimize the given instruction, possibly erasing
1321 void Reassociate::OptimizeInst(Instruction *I) {
1322 // Reassociation can expose instructions as dead. Erasing them, removing uses,
1323 // can free up their operands for reassociation.
1324 if (isInstructionTriviallyDead(I)) {
1325 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1326 // Erase the dead instruction.
1327 ValueRankMap.erase(I);
1328 I->eraseFromParent();
1329 // Optimize its operands.
1330 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1331 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1332 // If this is a node in an expression tree, climb to the expression root
1333 // and add that since that's where optimization actually happens.
1334 unsigned Opcode = Op->getOpcode();
1335 while (Op->hasOneUse() && Op->use_back()->getOpcode() == Opcode)
1336 Op = Op->use_back();
1337 RedoInsts.insert(Op);
1342 // Only consider operations that we understand.
1343 if (!isa<BinaryOperator>(I))
1346 if (I->getOpcode() == Instruction::Shl &&
1347 isa<ConstantInt>(I->getOperand(1)))
1348 // If an operand of this shift is a reassociable multiply, or if the shift
1349 // is used by a reassociable multiply or add, turn into a multiply.
1350 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
1352 (isReassociableOp(I->use_back(), Instruction::Mul) ||
1353 isReassociableOp(I->use_back(), Instruction::Add)))) {
1354 Instruction *NI = ConvertShiftToMul(I);
1355 ValueRankMap.erase(I);
1356 I->eraseFromParent();
1361 // Floating point binary operators are not associative, but we can still
1362 // commute (some) of them, to canonicalize the order of their operands.
1363 // This can potentially expose more CSE opportunities, and makes writing
1364 // other transformations simpler.
1365 if ((I->getType()->isFloatingPointTy() || I->getType()->isVectorTy())) {
1366 // FAdd and FMul can be commuted.
1367 if (I->getOpcode() != Instruction::FMul &&
1368 I->getOpcode() != Instruction::FAdd)
1371 Value *LHS = I->getOperand(0);
1372 Value *RHS = I->getOperand(1);
1373 unsigned LHSRank = getRank(LHS);
1374 unsigned RHSRank = getRank(RHS);
1376 // Sort the operands by rank.
1377 if (RHSRank < LHSRank) {
1378 I->setOperand(0, RHS);
1379 I->setOperand(1, LHS);
1385 // Do not reassociate boolean (i1) expressions. We want to preserve the
1386 // original order of evaluation for short-circuited comparisons that
1387 // SimplifyCFG has folded to AND/OR expressions. If the expression
1388 // is not further optimized, it is likely to be transformed back to a
1389 // short-circuited form for code gen, and the source order may have been
1390 // optimized for the most likely conditions.
1391 if (I->getType()->isIntegerTy(1))
1394 // If this is a subtract instruction which is not already in negate form,
1395 // see if we can convert it to X+-Y.
1396 if (I->getOpcode() == Instruction::Sub) {
1397 if (ShouldBreakUpSubtract(I)) {
1398 Instruction *NI = BreakUpSubtract(I);
1399 ValueRankMap.erase(I);
1400 I->eraseFromParent();
1403 } else if (BinaryOperator::isNeg(I)) {
1404 // Otherwise, this is a negation. See if the operand is a multiply tree
1405 // and if this is not an inner node of a multiply tree.
1406 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
1408 !isReassociableOp(I->use_back(), Instruction::Mul))) {
1409 Instruction *NI = LowerNegateToMultiply(I);
1410 ValueRankMap.erase(I);
1411 I->eraseFromParent();
1418 // If this instruction is an associative binary operator, process it.
1419 if (!I->isAssociative()) return;
1420 BinaryOperator *BO = cast<BinaryOperator>(I);
1422 // If this is an interior node of a reassociable tree, ignore it until we
1423 // get to the root of the tree, to avoid N^2 analysis.
1424 if (BO->hasOneUse() && BO->use_back()->getOpcode() == BO->getOpcode())
1427 // If this is an add tree that is used by a sub instruction, ignore it
1428 // until we process the subtract.
1429 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
1430 cast<Instruction>(BO->use_back())->getOpcode() == Instruction::Sub)
1433 ReassociateExpression(BO);
1436 Value *Reassociate::ReassociateExpression(BinaryOperator *I) {
1438 // First, walk the expression tree, linearizing the tree, collecting the
1439 // operand information.
1440 SmallVector<ValueEntry, 8> Ops;
1441 LinearizeExprTree(I, Ops);
1443 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
1445 // Now that we have linearized the tree to a list and have gathered all of
1446 // the operands and their ranks, sort the operands by their rank. Use a
1447 // stable_sort so that values with equal ranks will have their relative
1448 // positions maintained (and so the compiler is deterministic). Note that
1449 // this sorts so that the highest ranking values end up at the beginning of
1451 std::stable_sort(Ops.begin(), Ops.end());
1453 // OptimizeExpression - Now that we have the expression tree in a convenient
1454 // sorted form, optimize it globally if possible.
1455 if (Value *V = OptimizeExpression(I, Ops)) {
1456 // This expression tree simplified to something that isn't a tree,
1458 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
1459 I->replaceAllUsesWith(V);
1460 if (Instruction *VI = dyn_cast<Instruction>(V))
1461 VI->setDebugLoc(I->getDebugLoc());
1462 RedoInsts.insert(I);
1467 // We want to sink immediates as deeply as possible except in the case where
1468 // this is a multiply tree used only by an add, and the immediate is a -1.
1469 // In this case we reassociate to put the negation on the outside so that we
1470 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
1471 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
1472 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
1473 isa<ConstantInt>(Ops.back().Op) &&
1474 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
1475 ValueEntry Tmp = Ops.pop_back_val();
1476 Ops.insert(Ops.begin(), Tmp);
1479 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
1481 if (Ops.size() == 1) {
1482 // This expression tree simplified to something that isn't a tree,
1484 I->replaceAllUsesWith(Ops[0].Op);
1485 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
1486 OI->setDebugLoc(I->getDebugLoc());
1487 RedoInsts.insert(I);
1491 // Now that we ordered and optimized the expressions, splat them back into
1492 // the expression tree, removing any unneeded nodes.
1493 RewriteExprTree(I, Ops);
1497 bool Reassociate::runOnFunction(Function &F) {
1498 // Calculate the rank map for F
1502 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI)
1503 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; ) {
1504 // Optimize the current instruction, possibly erasing it. If this creates
1505 // new instructions that need optimizing then optimize all such too before
1506 // moving on to the next instruction.
1507 RedoInsts.insert(AssertingVH<Instruction>(II));
1508 while (!RedoInsts.empty()) {
1509 Instruction *I = RedoInsts.pop_back_val();
1510 if ((Instruction*)II == I)
1516 // We are done with the rank map.
1518 ValueRankMap.clear();