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/PostOrderIterator.h"
39 #include "llvm/ADT/STLExtras.h"
40 #include "llvm/ADT/Statistic.h"
41 #include "llvm/ADT/DenseMap.h"
45 STATISTIC(NumLinear , "Number of insts linearized");
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,
139 Value *OptimizeExpression(BinaryOperator *I,
140 SmallVectorImpl<ValueEntry> &Ops);
141 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
142 bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
143 SmallVectorImpl<Factor> &Factors);
144 Value *buildMinimalMultiplyDAG(IRBuilder<> &Builder,
145 SmallVectorImpl<Factor> &Factors);
146 Value *OptimizeMul(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
147 void LinearizeExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
148 void LinearizeExpr(BinaryOperator *I);
149 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
150 void ReassociateInst(BasicBlock::iterator &BBI);
152 void RemoveDeadBinaryOp(Value *V);
156 char Reassociate::ID = 0;
157 INITIALIZE_PASS(Reassociate, "reassociate",
158 "Reassociate expressions", false, false)
160 // Public interface to the Reassociate pass
161 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
163 void Reassociate::RemoveDeadBinaryOp(Value *V) {
164 Instruction *Op = dyn_cast<Instruction>(V);
165 if (!Op || !isa<BinaryOperator>(Op))
168 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1);
170 ValueRankMap.erase(Op);
171 DeadInsts.push_back(Op);
172 RemoveDeadBinaryOp(LHS);
173 RemoveDeadBinaryOp(RHS);
176 static bool isUnmovableInstruction(Instruction *I) {
177 if (I->getOpcode() == Instruction::PHI ||
178 I->getOpcode() == Instruction::Alloca ||
179 I->getOpcode() == Instruction::Load ||
180 I->getOpcode() == Instruction::Invoke ||
181 (I->getOpcode() == Instruction::Call &&
182 !isa<DbgInfoIntrinsic>(I)) ||
183 I->getOpcode() == Instruction::UDiv ||
184 I->getOpcode() == Instruction::SDiv ||
185 I->getOpcode() == Instruction::FDiv ||
186 I->getOpcode() == Instruction::URem ||
187 I->getOpcode() == Instruction::SRem ||
188 I->getOpcode() == Instruction::FRem)
193 void Reassociate::BuildRankMap(Function &F) {
196 // Assign distinct ranks to function arguments
197 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
198 ValueRankMap[&*I] = ++i;
200 ReversePostOrderTraversal<Function*> RPOT(&F);
201 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
202 E = RPOT.end(); I != E; ++I) {
204 unsigned BBRank = RankMap[BB] = ++i << 16;
206 // Walk the basic block, adding precomputed ranks for any instructions that
207 // we cannot move. This ensures that the ranks for these instructions are
208 // all different in the block.
209 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
210 if (isUnmovableInstruction(I))
211 ValueRankMap[&*I] = ++BBRank;
215 unsigned Reassociate::getRank(Value *V) {
216 Instruction *I = dyn_cast<Instruction>(V);
218 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
219 return 0; // Otherwise it's a global or constant, rank 0.
222 if (unsigned Rank = ValueRankMap[I])
223 return Rank; // Rank already known?
225 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
226 // we can reassociate expressions for code motion! Since we do not recurse
227 // for PHI nodes, we cannot have infinite recursion here, because there
228 // cannot be loops in the value graph that do not go through PHI nodes.
229 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
230 for (unsigned i = 0, e = I->getNumOperands();
231 i != e && Rank != MaxRank; ++i)
232 Rank = std::max(Rank, getRank(I->getOperand(i)));
234 // If this is a not or neg instruction, do not count it for rank. This
235 // assures us that X and ~X will have the same rank.
236 if (!I->getType()->isIntegerTy() ||
237 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
240 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = "
243 return ValueRankMap[I] = Rank;
246 /// isReassociableOp - Return true if V is an instruction of the specified
247 /// opcode and if it only has one use.
248 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
249 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) &&
250 cast<Instruction>(V)->getOpcode() == Opcode)
251 return cast<BinaryOperator>(V);
255 /// LowerNegateToMultiply - Replace 0-X with X*-1.
257 static Instruction *LowerNegateToMultiply(Instruction *Neg,
258 DenseMap<AssertingVH<Value>, unsigned> &ValueRankMap) {
259 Constant *Cst = Constant::getAllOnesValue(Neg->getType());
261 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
262 ValueRankMap.erase(Neg);
264 Neg->replaceAllUsesWith(Res);
265 Res->setDebugLoc(Neg->getDebugLoc());
266 Neg->eraseFromParent();
270 // Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'.
271 // Note that if D is also part of the expression tree that we recurse to
272 // linearize it as well. Besides that case, this does not recurse into A,B, or
274 void Reassociate::LinearizeExpr(BinaryOperator *I) {
275 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
276 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1));
277 assert(isReassociableOp(LHS, I->getOpcode()) &&
278 isReassociableOp(RHS, I->getOpcode()) &&
279 "Not an expression that needs linearization?");
281 DEBUG(dbgs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n');
283 // Move the RHS instruction to live immediately before I, avoiding breaking
284 // dominator properties.
287 // Move operands around to do the linearization.
288 I->setOperand(1, RHS->getOperand(0));
289 RHS->setOperand(0, LHS);
290 I->setOperand(0, RHS);
292 // Conservatively clear all the optional flags, which may not hold
293 // after the reassociation.
294 I->clearSubclassOptionalData();
295 LHS->clearSubclassOptionalData();
296 RHS->clearSubclassOptionalData();
300 DEBUG(dbgs() << "Linearized: " << *I << '\n');
302 // If D is part of this expression tree, tail recurse.
303 if (isReassociableOp(I->getOperand(1), I->getOpcode()))
307 /// LinearizeExprTree - Given an associative binary expression tree, traverse
308 /// all of the uses putting it into canonical form. This forces a left-linear
309 /// form of the expression (((a+b)+c)+d), and collects information about the
310 /// rank of the non-tree operands.
312 /// NOTE: These intentionally destroys the expression tree operands (turning
313 /// them into undef values) to reduce #uses of the values. This means that the
314 /// caller MUST use something like RewriteExprTree to put the values back in.
316 void Reassociate::LinearizeExprTree(BinaryOperator *I,
317 SmallVectorImpl<ValueEntry> &Ops) {
318 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
319 unsigned Opcode = I->getOpcode();
321 // First step, linearize the expression if it is in ((A+B)+(C+D)) form.
322 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode);
323 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode);
325 // If this is a multiply expression tree and it contains internal negations,
326 // transform them into multiplies by -1 so they can be reassociated.
327 if (I->getOpcode() == Instruction::Mul) {
328 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) {
329 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap);
330 LHSBO = isReassociableOp(LHS, Opcode);
332 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) {
333 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap);
334 RHSBO = isReassociableOp(RHS, Opcode);
340 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As
341 // such, just remember these operands and their rank.
342 Ops.push_back(ValueEntry(getRank(LHS), LHS));
343 Ops.push_back(ValueEntry(getRank(RHS), RHS));
345 // Clear the leaves out.
346 I->setOperand(0, UndefValue::get(I->getType()));
347 I->setOperand(1, UndefValue::get(I->getType()));
351 // Turn X+(Y+Z) -> (Y+Z)+X
352 std::swap(LHSBO, RHSBO);
354 bool Success = !I->swapOperands();
355 assert(Success && "swapOperands failed");
359 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the RHS is not
360 // part of the expression tree.
362 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0));
363 RHS = I->getOperand(1);
367 // Okay, now we know that the LHS is a nested expression and that the RHS is
368 // not. Perform reassociation.
369 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!");
371 // Move LHS right before I to make sure that the tree expression dominates all
373 LHSBO->moveBefore(I);
375 // Linearize the expression tree on the LHS.
376 LinearizeExprTree(LHSBO, Ops);
378 // Remember the RHS operand and its rank.
379 Ops.push_back(ValueEntry(getRank(RHS), RHS));
381 // Clear the RHS leaf out.
382 I->setOperand(1, UndefValue::get(I->getType()));
385 // RewriteExprTree - Now that the operands for this expression tree are
386 // linearized and optimized, emit them in-order. This function is written to be
388 void Reassociate::RewriteExprTree(BinaryOperator *I,
389 SmallVectorImpl<ValueEntry> &Ops,
391 if (i+2 == Ops.size()) {
392 if (I->getOperand(0) != Ops[i].Op ||
393 I->getOperand(1) != Ops[i+1].Op) {
394 Value *OldLHS = I->getOperand(0);
395 DEBUG(dbgs() << "RA: " << *I << '\n');
396 I->setOperand(0, Ops[i].Op);
397 I->setOperand(1, Ops[i+1].Op);
399 // Clear all the optional flags, which may not hold after the
400 // reassociation if the expression involved more than just this operation.
402 I->clearSubclassOptionalData();
404 DEBUG(dbgs() << "TO: " << *I << '\n');
408 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3)
409 // delete the extra, now dead, nodes.
410 RemoveDeadBinaryOp(OldLHS);
414 assert(i+2 < Ops.size() && "Ops index out of range!");
416 if (I->getOperand(1) != Ops[i].Op) {
417 DEBUG(dbgs() << "RA: " << *I << '\n');
418 I->setOperand(1, Ops[i].Op);
420 // Conservatively clear all the optional flags, which may not hold
421 // after the reassociation.
422 I->clearSubclassOptionalData();
424 DEBUG(dbgs() << "TO: " << *I << '\n');
429 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
430 assert(LHS->getOpcode() == I->getOpcode() &&
431 "Improper expression tree!");
433 // Compactify the tree instructions together with each other to guarantee
434 // that the expression tree is dominated by all of Ops.
436 RewriteExprTree(LHS, Ops, i+1);
439 /// NegateValue - Insert instructions before the instruction pointed to by BI,
440 /// that computes the negative version of the value specified. The negative
441 /// version of the value is returned, and BI is left pointing at the instruction
442 /// that should be processed next by the reassociation pass.
443 static Value *NegateValue(Value *V, Instruction *BI) {
444 if (Constant *C = dyn_cast<Constant>(V))
445 return ConstantExpr::getNeg(C);
447 // We are trying to expose opportunity for reassociation. One of the things
448 // that we want to do to achieve this is to push a negation as deep into an
449 // expression chain as possible, to expose the add instructions. In practice,
450 // this means that we turn this:
451 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
452 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
453 // the constants. We assume that instcombine will clean up the mess later if
454 // we introduce tons of unnecessary negation instructions.
456 if (Instruction *I = dyn_cast<Instruction>(V))
457 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
458 // Push the negates through the add.
459 I->setOperand(0, NegateValue(I->getOperand(0), BI));
460 I->setOperand(1, NegateValue(I->getOperand(1), BI));
462 // We must move the add instruction here, because the neg instructions do
463 // not dominate the old add instruction in general. By moving it, we are
464 // assured that the neg instructions we just inserted dominate the
465 // instruction we are about to insert after them.
468 I->setName(I->getName()+".neg");
472 // Okay, we need to materialize a negated version of V with an instruction.
473 // Scan the use lists of V to see if we have one already.
474 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
476 if (!BinaryOperator::isNeg(U)) continue;
478 // We found one! Now we have to make sure that the definition dominates
479 // this use. We do this by moving it to the entry block (if it is a
480 // non-instruction value) or right after the definition. These negates will
481 // be zapped by reassociate later, so we don't need much finesse here.
482 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
484 // Verify that the negate is in this function, V might be a constant expr.
485 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
488 BasicBlock::iterator InsertPt;
489 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
490 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
491 InsertPt = II->getNormalDest()->begin();
493 InsertPt = InstInput;
496 while (isa<PHINode>(InsertPt)) ++InsertPt;
498 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
500 TheNeg->moveBefore(InsertPt);
504 // Insert a 'neg' instruction that subtracts the value from zero to get the
506 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
509 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
510 /// X-Y into (X + -Y).
511 static bool ShouldBreakUpSubtract(Instruction *Sub) {
512 // If this is a negation, we can't split it up!
513 if (BinaryOperator::isNeg(Sub))
516 // Don't bother to break this up unless either the LHS is an associable add or
517 // subtract or if this is only used by one.
518 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
519 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
521 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
522 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
524 if (Sub->hasOneUse() &&
525 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
526 isReassociableOp(Sub->use_back(), Instruction::Sub)))
532 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
533 /// only used by an add, transform this into (X+(0-Y)) to promote better
535 static Instruction *BreakUpSubtract(Instruction *Sub,
536 DenseMap<AssertingVH<Value>, unsigned> &ValueRankMap) {
537 // Convert a subtract into an add and a neg instruction. This allows sub
538 // instructions to be commuted with other add instructions.
540 // Calculate the negative value of Operand 1 of the sub instruction,
541 // and set it as the RHS of the add instruction we just made.
543 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
545 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
548 // Everyone now refers to the add instruction.
549 ValueRankMap.erase(Sub);
550 Sub->replaceAllUsesWith(New);
551 New->setDebugLoc(Sub->getDebugLoc());
552 Sub->eraseFromParent();
554 DEBUG(dbgs() << "Negated: " << *New << '\n');
558 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
559 /// by one, change this into a multiply by a constant to assist with further
561 static Instruction *ConvertShiftToMul(Instruction *Shl,
562 DenseMap<AssertingVH<Value>, unsigned> &ValueRankMap) {
563 // If an operand of this shift is a reassociable multiply, or if the shift
564 // is used by a reassociable multiply or add, turn into a multiply.
565 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) ||
567 (isReassociableOp(Shl->use_back(), Instruction::Mul) ||
568 isReassociableOp(Shl->use_back(), Instruction::Add)))) {
569 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
570 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
573 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
574 ValueRankMap.erase(Shl);
576 Shl->replaceAllUsesWith(Mul);
577 Mul->setDebugLoc(Shl->getDebugLoc());
578 Shl->eraseFromParent();
584 /// FindInOperandList - Scan backwards and forwards among values with the same
585 /// rank as element i to see if X exists. If X does not exist, return i. This
586 /// is useful when scanning for 'x' when we see '-x' because they both get the
588 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
590 unsigned XRank = Ops[i].Rank;
591 unsigned e = Ops.size();
592 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
596 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
602 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
603 /// and returning the result. Insert the tree before I.
604 static Value *EmitAddTreeOfValues(Instruction *I,
605 SmallVectorImpl<WeakVH> &Ops){
606 if (Ops.size() == 1) return Ops.back();
608 Value *V1 = Ops.back();
610 Value *V2 = EmitAddTreeOfValues(I, Ops);
611 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
614 /// RemoveFactorFromExpression - If V is an expression tree that is a
615 /// multiplication sequence, and if this sequence contains a multiply by Factor,
616 /// remove Factor from the tree and return the new tree.
617 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
618 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
621 SmallVector<ValueEntry, 8> Factors;
622 LinearizeExprTree(BO, Factors);
624 bool FoundFactor = false;
625 bool NeedsNegate = false;
626 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
627 if (Factors[i].Op == Factor) {
629 Factors.erase(Factors.begin()+i);
633 // If this is a negative version of this factor, remove it.
634 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor))
635 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
636 if (FC1->getValue() == -FC2->getValue()) {
637 FoundFactor = NeedsNegate = true;
638 Factors.erase(Factors.begin()+i);
644 // Make sure to restore the operands to the expression tree.
645 RewriteExprTree(BO, Factors);
649 BasicBlock::iterator InsertPt = BO; ++InsertPt;
651 // If this was just a single multiply, remove the multiply and return the only
652 // remaining operand.
653 if (Factors.size() == 1) {
654 ValueRankMap.erase(BO);
655 DeadInsts.push_back(BO);
658 RewriteExprTree(BO, Factors);
663 V = BinaryOperator::CreateNeg(V, "neg", InsertPt);
668 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
669 /// add its operands as factors, otherwise add V to the list of factors.
671 /// Ops is the top-level list of add operands we're trying to factor.
672 static void FindSingleUseMultiplyFactors(Value *V,
673 SmallVectorImpl<Value*> &Factors,
674 const SmallVectorImpl<ValueEntry> &Ops,
677 if (!(V->hasOneUse() || V->use_empty()) || // More than one use.
678 !(BO = dyn_cast<BinaryOperator>(V)) ||
679 BO->getOpcode() != Instruction::Mul) {
680 Factors.push_back(V);
684 // If this value has a single use because it is another input to the add
685 // tree we're reassociating and we dropped its use, it actually has two
686 // uses and we can't factor it.
688 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
689 if (Ops[i].Op == V) {
690 Factors.push_back(V);
696 // Otherwise, add the LHS and RHS to the list of factors.
697 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops, false);
698 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops, false);
701 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
702 /// instruction. This optimizes based on identities. If it can be reduced to
703 /// a single Value, it is returned, otherwise the Ops list is mutated as
705 static Value *OptimizeAndOrXor(unsigned Opcode,
706 SmallVectorImpl<ValueEntry> &Ops) {
707 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
708 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
709 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
710 // First, check for X and ~X in the operand list.
711 assert(i < Ops.size());
712 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
713 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
714 unsigned FoundX = FindInOperandList(Ops, i, X);
716 if (Opcode == Instruction::And) // ...&X&~X = 0
717 return Constant::getNullValue(X->getType());
719 if (Opcode == Instruction::Or) // ...|X|~X = -1
720 return Constant::getAllOnesValue(X->getType());
724 // Next, check for duplicate pairs of values, which we assume are next to
725 // each other, due to our sorting criteria.
726 assert(i < Ops.size());
727 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
728 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
729 // Drop duplicate values for And and Or.
730 Ops.erase(Ops.begin()+i);
736 // Drop pairs of values for Xor.
737 assert(Opcode == Instruction::Xor);
739 return Constant::getNullValue(Ops[0].Op->getType());
742 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
750 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
751 /// optimizes based on identities. If it can be reduced to a single Value, it
752 /// is returned, otherwise the Ops list is mutated as necessary.
753 Value *Reassociate::OptimizeAdd(Instruction *I,
754 SmallVectorImpl<ValueEntry> &Ops) {
755 // Scan the operand lists looking for X and -X pairs. If we find any, we
756 // can simplify the expression. X+-X == 0. While we're at it, scan for any
757 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
759 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1".
761 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
762 Value *TheOp = Ops[i].Op;
763 // Check to see if we've seen this operand before. If so, we factor all
764 // instances of the operand together. Due to our sorting criteria, we know
765 // that these need to be next to each other in the vector.
766 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
767 // Rescan the list, remove all instances of this operand from the expr.
768 unsigned NumFound = 0;
770 Ops.erase(Ops.begin()+i);
772 } while (i != Ops.size() && Ops[i].Op == TheOp);
774 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
777 // Insert a new multiply.
778 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound);
779 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I);
781 // Now that we have inserted a multiply, optimize it. This allows us to
782 // handle cases that require multiple factoring steps, such as this:
783 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
784 RedoInsts.push_back(Mul);
786 // If every add operand was a duplicate, return the multiply.
790 // Otherwise, we had some input that didn't have the dupe, such as
791 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
792 // things being added by this operation.
793 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
800 // Check for X and -X in the operand list.
801 if (!BinaryOperator::isNeg(TheOp))
804 Value *X = BinaryOperator::getNegArgument(TheOp);
805 unsigned FoundX = FindInOperandList(Ops, i, X);
809 // Remove X and -X from the operand list.
811 return Constant::getNullValue(X->getType());
813 Ops.erase(Ops.begin()+i);
817 --i; // Need to back up an extra one.
818 Ops.erase(Ops.begin()+FoundX);
820 --i; // Revisit element.
821 e -= 2; // Removed two elements.
824 // Scan the operand list, checking to see if there are any common factors
825 // between operands. Consider something like A*A+A*B*C+D. We would like to
826 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
827 // To efficiently find this, we count the number of times a factor occurs
828 // for any ADD operands that are MULs.
829 DenseMap<Value*, unsigned> FactorOccurrences;
831 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
832 // where they are actually the same multiply.
834 Value *MaxOccVal = 0;
835 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
836 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op);
837 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty())
840 // Compute all of the factors of this added value.
841 SmallVector<Value*, 8> Factors;
842 FindSingleUseMultiplyFactors(BOp, Factors, Ops, true);
843 assert(Factors.size() > 1 && "Bad linearize!");
845 // Add one to FactorOccurrences for each unique factor in this op.
846 SmallPtrSet<Value*, 8> Duplicates;
847 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
848 Value *Factor = Factors[i];
849 if (!Duplicates.insert(Factor)) continue;
851 unsigned Occ = ++FactorOccurrences[Factor];
852 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
854 // If Factor is a negative constant, add the negated value as a factor
855 // because we can percolate the negate out. Watch for minint, which
856 // cannot be positivified.
857 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor))
858 if (CI->isNegative() && !CI->isMinValue(true)) {
859 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
860 assert(!Duplicates.count(Factor) &&
861 "Shouldn't have two constant factors, missed a canonicalize");
863 unsigned Occ = ++FactorOccurrences[Factor];
864 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
869 // If any factor occurred more than one time, we can pull it out.
871 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
874 // Create a new instruction that uses the MaxOccVal twice. If we don't do
875 // this, we could otherwise run into situations where removing a factor
876 // from an expression will drop a use of maxocc, and this can cause
877 // RemoveFactorFromExpression on successive values to behave differently.
878 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
879 SmallVector<WeakVH, 4> NewMulOps;
880 for (unsigned i = 0; i != Ops.size(); ++i) {
881 // Only try to remove factors from expressions we're allowed to.
882 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op);
883 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty())
886 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
887 // The factorized operand may occur several times. Convert them all in
889 for (unsigned j = Ops.size(); j != i;) {
891 if (Ops[j].Op == Ops[i].Op) {
892 NewMulOps.push_back(V);
893 Ops.erase(Ops.begin()+j);
900 // No need for extra uses anymore.
903 unsigned NumAddedValues = NewMulOps.size();
904 Value *V = EmitAddTreeOfValues(I, NewMulOps);
906 // Now that we have inserted the add tree, optimize it. This allows us to
907 // handle cases that require multiple factoring steps, such as this:
908 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
909 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
910 (void)NumAddedValues;
911 V = ReassociateExpression(cast<BinaryOperator>(V));
913 // Create the multiply.
914 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
916 // Rerun associate on the multiply in case the inner expression turned into
917 // a multiply. We want to make sure that we keep things in canonical form.
918 V2 = ReassociateExpression(cast<BinaryOperator>(V2));
920 // If every add operand included the factor (e.g. "A*B + A*C"), then the
921 // entire result expression is just the multiply "A*(B+C)".
925 // Otherwise, we had some input that didn't have the factor, such as
926 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
927 // things being added by this operation.
928 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
935 /// \brief Predicate tests whether a ValueEntry's op is in a map.
936 struct IsValueInMap {
937 const DenseMap<Value *, unsigned> ⤅
939 IsValueInMap(const DenseMap<Value *, unsigned> &Map) : Map(Map) {}
941 bool operator()(const ValueEntry &Entry) {
942 return Map.find(Entry.Op) != Map.end();
947 /// \brief Build up a vector of value/power pairs factoring a product.
949 /// Given a series of multiplication operands, build a vector of factors and
950 /// the powers each is raised to when forming the final product. Sort them in
951 /// the order of descending power.
953 /// (x*x) -> [(x, 2)]
954 /// ((x*x)*x) -> [(x, 3)]
955 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
957 /// \returns Whether any factors have a power greater than one.
958 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
959 SmallVectorImpl<Factor> &Factors) {
960 unsigned FactorPowerSum = 0;
961 DenseMap<Value *, unsigned> FactorCounts;
962 for (unsigned LastIdx = 0, Idx = 0, Size = Ops.size(); Idx < Size; ++Idx) {
963 // Note that 'use_empty' uses means the only use is in the linearized tree
964 // represented by Ops -- we remove the values from the actual operations to
965 // reduce their use count.
966 if (!Ops[Idx].Op->use_empty()) {
971 if (LastIdx == Idx || Ops[LastIdx].Op != Ops[Idx].Op) {
975 // Track for simplification all factors which occur 2 or more times.
976 DenseMap<Value *, unsigned>::iterator CountIt;
978 llvm::tie(CountIt, Inserted)
979 = FactorCounts.insert(std::make_pair(Ops[Idx].Op, 2));
982 Factors.push_back(Factor(Ops[Idx].Op, 2));
988 // We can only simplify factors if the sum of the powers of our simplifiable
989 // factors is 4 or higher. When that is the case, we will *always* have
990 // a simplification. This is an important invariant to prevent cyclicly
991 // trying to simplify already minimal formations.
992 if (FactorPowerSum < 4)
995 // Remove all the operands which are in the map.
996 Ops.erase(std::remove_if(Ops.begin(), Ops.end(), IsValueInMap(FactorCounts)),
999 // Record the adjusted power for the simplification factors. We add back into
1000 // the Ops list any values with an odd power, and make the power even. This
1001 // allows the outer-most multiplication tree to remain in tact during
1003 unsigned OldOpsSize = Ops.size();
1004 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1005 Factors[Idx].Power = FactorCounts[Factors[Idx].Base];
1006 if (Factors[Idx].Power & 1) {
1007 Ops.push_back(ValueEntry(getRank(Factors[Idx].Base), Factors[Idx].Base));
1008 --Factors[Idx].Power;
1012 // None of the adjustments above should have reduced the sum of factor powers
1013 // below our mininum of '4'.
1014 assert(FactorPowerSum >= 4);
1016 // Patch up the sort of the ops vector by sorting the factors we added back
1017 // onto the back, and merging the two sequences.
1018 if (OldOpsSize != Ops.size()) {
1019 SmallVectorImpl<ValueEntry>::iterator MiddleIt = Ops.begin() + OldOpsSize;
1020 std::sort(MiddleIt, Ops.end());
1021 std::inplace_merge(Ops.begin(), MiddleIt, Ops.end());
1024 std::sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter());
1028 /// \brief Build a tree of multiplies, computing the product of Ops.
1029 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1030 SmallVectorImpl<Value*> &Ops) {
1031 if (Ops.size() == 1)
1034 Value *LHS = Ops.pop_back_val();
1036 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1037 } while (!Ops.empty());
1042 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1044 /// Given a vector of values raised to various powers, where no two values are
1045 /// equal and the powers are sorted in decreasing order, compute the minimal
1046 /// DAG of multiplies to compute the final product, and return that product
1048 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1049 SmallVectorImpl<Factor> &Factors) {
1050 assert(Factors[0].Power);
1051 SmallVector<Value *, 4> OuterProduct;
1052 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1053 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1054 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1059 // We want to multiply across all the factors with the same power so that
1060 // we can raise them to that power as a single entity. Build a mini tree
1062 SmallVector<Value *, 4> InnerProduct;
1063 InnerProduct.push_back(Factors[LastIdx].Base);
1065 InnerProduct.push_back(Factors[Idx].Base);
1067 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1069 // Reset the base value of the first factor to the new expression tree.
1070 // We'll remove all the factors with the same power in a second pass.
1071 Factors[LastIdx].Base
1072 = ReassociateExpression(
1073 cast<BinaryOperator>(buildMultiplyTree(Builder, InnerProduct)));
1077 // Unique factors with equal powers -- we've folded them into the first one's
1079 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1080 Factor::PowerEqual()),
1083 // Iteratively collect the base of each factor with an add power into the
1084 // outer product, and halve each power in preparation for squaring the
1086 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1087 if (Factors[Idx].Power & 1)
1088 OuterProduct.push_back(Factors[Idx].Base);
1089 Factors[Idx].Power >>= 1;
1091 if (Factors[0].Power) {
1092 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1093 OuterProduct.push_back(SquareRoot);
1094 OuterProduct.push_back(SquareRoot);
1096 if (OuterProduct.size() == 1)
1097 return OuterProduct.front();
1099 return ReassociateExpression(
1100 cast<BinaryOperator>(buildMultiplyTree(Builder, OuterProduct)));
1103 Value *Reassociate::OptimizeMul(BinaryOperator *I,
1104 SmallVectorImpl<ValueEntry> &Ops) {
1105 // We can only optimize the multiplies when there is a chain of more than
1106 // three, such that a balanced tree might require fewer total multiplies.
1110 // Try to turn linear trees of multiplies without other uses of the
1111 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1113 SmallVector<Factor, 4> Factors;
1114 if (!collectMultiplyFactors(Ops, Factors))
1115 return 0; // All distinct factors, so nothing left for us to do.
1117 IRBuilder<> Builder(I);
1118 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1122 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1123 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1127 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
1128 SmallVectorImpl<ValueEntry> &Ops) {
1129 // Now that we have the linearized expression tree, try to optimize it.
1130 // Start by folding any constants that we found.
1131 bool IterateOptimization = false;
1132 if (Ops.size() == 1) return Ops[0].Op;
1134 unsigned Opcode = I->getOpcode();
1136 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
1137 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
1139 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
1140 return OptimizeExpression(I, Ops);
1143 // Check for destructive annihilation due to a constant being used.
1144 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
1147 case Instruction::And:
1148 if (CstVal->isZero()) // X & 0 -> 0
1150 if (CstVal->isAllOnesValue()) // X & -1 -> X
1153 case Instruction::Mul:
1154 if (CstVal->isZero()) { // X * 0 -> 0
1159 if (cast<ConstantInt>(CstVal)->isOne())
1160 Ops.pop_back(); // X * 1 -> X
1162 case Instruction::Or:
1163 if (CstVal->isAllOnesValue()) // X | -1 -> -1
1166 case Instruction::Add:
1167 case Instruction::Xor:
1168 if (CstVal->isZero()) // X [|^+] 0 -> X
1172 if (Ops.size() == 1) return Ops[0].Op;
1174 // Handle destructive annihilation due to identities between elements in the
1175 // argument list here.
1176 unsigned NumOps = Ops.size();
1179 case Instruction::And:
1180 case Instruction::Or:
1181 case Instruction::Xor:
1182 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1186 case Instruction::Add:
1187 if (Value *Result = OptimizeAdd(I, Ops))
1191 case Instruction::Mul:
1192 if (Value *Result = OptimizeMul(I, Ops))
1197 if (IterateOptimization || Ops.size() != NumOps)
1198 return OptimizeExpression(I, Ops);
1202 /// ReassociateInst - Inspect and reassociate the instruction at the
1203 /// given position, post-incrementing the position.
1204 void Reassociate::ReassociateInst(BasicBlock::iterator &BBI) {
1205 Instruction *BI = BBI++;
1206 if (BI->getOpcode() == Instruction::Shl &&
1207 isa<ConstantInt>(BI->getOperand(1)))
1208 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) {
1213 // Reject cases where it is pointless to do this.
1214 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPointTy() ||
1215 BI->getType()->isVectorTy())
1216 return; // Floating point ops are not associative.
1218 // Do not reassociate boolean (i1) expressions. We want to preserve the
1219 // original order of evaluation for short-circuited comparisons that
1220 // SimplifyCFG has folded to AND/OR expressions. If the expression
1221 // is not further optimized, it is likely to be transformed back to a
1222 // short-circuited form for code gen, and the source order may have been
1223 // optimized for the most likely conditions.
1224 if (BI->getType()->isIntegerTy(1))
1227 // If this is a subtract instruction which is not already in negate form,
1228 // see if we can convert it to X+-Y.
1229 if (BI->getOpcode() == Instruction::Sub) {
1230 if (ShouldBreakUpSubtract(BI)) {
1231 BI = BreakUpSubtract(BI, ValueRankMap);
1232 // Reset the BBI iterator in case BreakUpSubtract changed the
1233 // instruction it points to.
1237 } else if (BinaryOperator::isNeg(BI)) {
1238 // Otherwise, this is a negation. See if the operand is a multiply tree
1239 // and if this is not an inner node of a multiply tree.
1240 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) &&
1241 (!BI->hasOneUse() ||
1242 !isReassociableOp(BI->use_back(), Instruction::Mul))) {
1243 BI = LowerNegateToMultiply(BI, ValueRankMap);
1249 // If this instruction is a commutative binary operator, process it.
1250 if (!BI->isAssociative()) return;
1251 BinaryOperator *I = cast<BinaryOperator>(BI);
1253 // If this is an interior node of a reassociable tree, ignore it until we
1254 // get to the root of the tree, to avoid N^2 analysis.
1255 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode()))
1258 // If this is an add tree that is used by a sub instruction, ignore it
1259 // until we process the subtract.
1260 if (I->hasOneUse() && I->getOpcode() == Instruction::Add &&
1261 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub)
1264 ReassociateExpression(I);
1267 Value *Reassociate::ReassociateExpression(BinaryOperator *I) {
1269 // First, walk the expression tree, linearizing the tree, collecting the
1270 // operand information.
1271 SmallVector<ValueEntry, 8> Ops;
1272 LinearizeExprTree(I, Ops);
1274 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
1276 // Now that we have linearized the tree to a list and have gathered all of
1277 // the operands and their ranks, sort the operands by their rank. Use a
1278 // stable_sort so that values with equal ranks will have their relative
1279 // positions maintained (and so the compiler is deterministic). Note that
1280 // this sorts so that the highest ranking values end up at the beginning of
1282 std::stable_sort(Ops.begin(), Ops.end());
1284 // OptimizeExpression - Now that we have the expression tree in a convenient
1285 // sorted form, optimize it globally if possible.
1286 if (Value *V = OptimizeExpression(I, Ops)) {
1287 // This expression tree simplified to something that isn't a tree,
1289 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
1290 I->replaceAllUsesWith(V);
1291 if (Instruction *VI = dyn_cast<Instruction>(V))
1292 VI->setDebugLoc(I->getDebugLoc());
1293 RemoveDeadBinaryOp(I);
1298 // We want to sink immediates as deeply as possible except in the case where
1299 // this is a multiply tree used only by an add, and the immediate is a -1.
1300 // In this case we reassociate to put the negation on the outside so that we
1301 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
1302 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
1303 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
1304 isa<ConstantInt>(Ops.back().Op) &&
1305 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
1306 ValueEntry Tmp = Ops.pop_back_val();
1307 Ops.insert(Ops.begin(), Tmp);
1310 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
1312 if (Ops.size() == 1) {
1313 // This expression tree simplified to something that isn't a tree,
1315 I->replaceAllUsesWith(Ops[0].Op);
1316 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
1317 OI->setDebugLoc(I->getDebugLoc());
1318 RemoveDeadBinaryOp(I);
1322 // Now that we ordered and optimized the expressions, splat them back into
1323 // the expression tree, removing any unneeded nodes.
1324 RewriteExprTree(I, Ops);
1328 bool Reassociate::runOnFunction(Function &F) {
1329 // Recalculate the rank map for F
1333 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
1334 for (BasicBlock::iterator BBI = FI->begin(); BBI != FI->end(); )
1335 ReassociateInst(BBI);
1337 // Now that we're done, revisit any instructions which are likely to
1338 // have secondary reassociation opportunities.
1339 while (!RedoInsts.empty())
1340 if (Value *V = RedoInsts.pop_back_val()) {
1341 BasicBlock::iterator BBI = cast<Instruction>(V);
1342 ReassociateInst(BBI);
1345 // Now that we're done, delete any instructions which are no longer used.
1346 while (!DeadInsts.empty())
1347 if (Value *V = DeadInsts.pop_back_val())
1348 RecursivelyDeleteTriviallyDeadInstructions(V);
1350 // We are done with the rank map.
1352 ValueRankMap.clear();