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...
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/Constants.h"
26 #include "llvm/DerivedTypes.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Assembly/Writer.h"
32 #include "llvm/Support/CFG.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ValueHandle.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/ADT/PostOrderIterator.h"
37 #include "llvm/ADT/Statistic.h"
38 #include "llvm/ADT/DenseMap.h"
43 STATISTIC(NumLinear , "Number of insts linearized");
44 STATISTIC(NumChanged, "Number of insts reassociated");
45 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
46 STATISTIC(NumFactor , "Number of multiplies factored");
52 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {}
54 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) {
55 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start.
60 /// PrintOps - Print out the expression identified in the Ops list.
62 static void PrintOps(Instruction *I, const std::vector<ValueEntry> &Ops) {
63 Module *M = I->getParent()->getParent()->getParent();
64 errs() << Instruction::getOpcodeName(I->getOpcode()) << " "
65 << *Ops[0].Op->getType() << '\t';
66 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
68 WriteAsOperand(errs(), Ops[i].Op, false, M);
69 errs() << ", #" << Ops[i].Rank << "] ";
75 class Reassociate : public FunctionPass {
76 std::map<BasicBlock*, unsigned> RankMap;
77 std::map<AssertingVH<>, unsigned> ValueRankMap;
80 static char ID; // Pass identification, replacement for typeid
81 Reassociate() : FunctionPass(&ID) {}
83 bool runOnFunction(Function &F);
85 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
89 void BuildRankMap(Function &F);
90 unsigned getRank(Value *V);
91 void ReassociateExpression(BinaryOperator *I);
92 void RewriteExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops,
94 Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops);
95 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops);
96 void LinearizeExpr(BinaryOperator *I);
97 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
98 void ReassociateBB(BasicBlock *BB);
100 void RemoveDeadBinaryOp(Value *V);
104 char Reassociate::ID = 0;
105 static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions");
107 // Public interface to the Reassociate pass
108 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
110 void Reassociate::RemoveDeadBinaryOp(Value *V) {
111 Instruction *Op = dyn_cast<Instruction>(V);
112 if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty())
115 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1);
116 RemoveDeadBinaryOp(LHS);
117 RemoveDeadBinaryOp(RHS);
121 static bool isUnmovableInstruction(Instruction *I) {
122 if (I->getOpcode() == Instruction::PHI ||
123 I->getOpcode() == Instruction::Alloca ||
124 I->getOpcode() == Instruction::Load ||
125 I->getOpcode() == Instruction::Invoke ||
126 (I->getOpcode() == Instruction::Call &&
127 !isa<DbgInfoIntrinsic>(I)) ||
128 I->getOpcode() == Instruction::UDiv ||
129 I->getOpcode() == Instruction::SDiv ||
130 I->getOpcode() == Instruction::FDiv ||
131 I->getOpcode() == Instruction::URem ||
132 I->getOpcode() == Instruction::SRem ||
133 I->getOpcode() == Instruction::FRem)
138 void Reassociate::BuildRankMap(Function &F) {
141 // Assign distinct ranks to function arguments
142 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
143 ValueRankMap[&*I] = ++i;
145 ReversePostOrderTraversal<Function*> RPOT(&F);
146 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
147 E = RPOT.end(); I != E; ++I) {
149 unsigned BBRank = RankMap[BB] = ++i << 16;
151 // Walk the basic block, adding precomputed ranks for any instructions that
152 // we cannot move. This ensures that the ranks for these instructions are
153 // all different in the block.
154 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
155 if (isUnmovableInstruction(I))
156 ValueRankMap[&*I] = ++BBRank;
160 unsigned Reassociate::getRank(Value *V) {
161 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument...
163 Instruction *I = dyn_cast<Instruction>(V);
164 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0.
166 unsigned &CachedRank = ValueRankMap[I];
167 if (CachedRank) return CachedRank; // Rank already known?
169 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
170 // we can reassociate expressions for code motion! Since we do not recurse
171 // for PHI nodes, we cannot have infinite recursion here, because there
172 // cannot be loops in the value graph that do not go through PHI nodes.
173 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
174 for (unsigned i = 0, e = I->getNumOperands();
175 i != e && Rank != MaxRank; ++i)
176 Rank = std::max(Rank, getRank(I->getOperand(i)));
178 // If this is a not or neg instruction, do not count it for rank. This
179 // assures us that X and ~X will have the same rank.
180 if (!I->getType()->isInteger() ||
181 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
184 //DEBUG(errs() << "Calculated Rank[" << V->getName() << "] = "
187 return CachedRank = Rank;
190 /// isReassociableOp - Return true if V is an instruction of the specified
191 /// opcode and if it only has one use.
192 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
193 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) &&
194 cast<Instruction>(V)->getOpcode() == Opcode)
195 return cast<BinaryOperator>(V);
199 /// LowerNegateToMultiply - Replace 0-X with X*-1.
201 static Instruction *LowerNegateToMultiply(Instruction *Neg,
202 std::map<AssertingVH<>, unsigned> &ValueRankMap) {
203 Constant *Cst = Constant::getAllOnesValue(Neg->getType());
205 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
206 ValueRankMap.erase(Neg);
208 Neg->replaceAllUsesWith(Res);
209 Neg->eraseFromParent();
213 // Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'.
214 // Note that if D is also part of the expression tree that we recurse to
215 // linearize it as well. Besides that case, this does not recurse into A,B, or
217 void Reassociate::LinearizeExpr(BinaryOperator *I) {
218 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
219 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1));
220 assert(isReassociableOp(LHS, I->getOpcode()) &&
221 isReassociableOp(RHS, I->getOpcode()) &&
222 "Not an expression that needs linearization?");
224 DEBUG(errs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n');
226 // Move the RHS instruction to live immediately before I, avoiding breaking
227 // dominator properties.
230 // Move operands around to do the linearization.
231 I->setOperand(1, RHS->getOperand(0));
232 RHS->setOperand(0, LHS);
233 I->setOperand(0, RHS);
237 DEBUG(errs() << "Linearized: " << *I << '\n');
239 // If D is part of this expression tree, tail recurse.
240 if (isReassociableOp(I->getOperand(1), I->getOpcode()))
245 /// LinearizeExprTree - Given an associative binary expression tree, traverse
246 /// all of the uses putting it into canonical form. This forces a left-linear
247 /// form of the the expression (((a+b)+c)+d), and collects information about the
248 /// rank of the non-tree operands.
250 /// NOTE: These intentionally destroys the expression tree operands (turning
251 /// them into undef values) to reduce #uses of the values. This means that the
252 /// caller MUST use something like RewriteExprTree to put the values back in.
254 void Reassociate::LinearizeExprTree(BinaryOperator *I,
255 std::vector<ValueEntry> &Ops) {
256 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
257 unsigned Opcode = I->getOpcode();
259 // First step, linearize the expression if it is in ((A+B)+(C+D)) form.
260 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode);
261 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode);
263 // If this is a multiply expression tree and it contains internal negations,
264 // transform them into multiplies by -1 so they can be reassociated.
265 if (I->getOpcode() == Instruction::Mul) {
266 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) {
267 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap);
268 LHSBO = isReassociableOp(LHS, Opcode);
270 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) {
271 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap);
272 RHSBO = isReassociableOp(RHS, Opcode);
278 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As
279 // such, just remember these operands and their rank.
280 Ops.push_back(ValueEntry(getRank(LHS), LHS));
281 Ops.push_back(ValueEntry(getRank(RHS), RHS));
283 // Clear the leaves out.
284 I->setOperand(0, UndefValue::get(I->getType()));
285 I->setOperand(1, UndefValue::get(I->getType()));
288 // Turn X+(Y+Z) -> (Y+Z)+X
289 std::swap(LHSBO, RHSBO);
291 bool Success = !I->swapOperands();
292 assert(Success && "swapOperands failed");
297 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not
298 // part of the expression tree.
300 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0));
301 RHS = I->getOperand(1);
305 // Okay, now we know that the LHS is a nested expression and that the RHS is
306 // not. Perform reassociation.
307 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!");
309 // Move LHS right before I to make sure that the tree expression dominates all
311 LHSBO->moveBefore(I);
313 // Linearize the expression tree on the LHS.
314 LinearizeExprTree(LHSBO, Ops);
316 // Remember the RHS operand and its rank.
317 Ops.push_back(ValueEntry(getRank(RHS), RHS));
319 // Clear the RHS leaf out.
320 I->setOperand(1, UndefValue::get(I->getType()));
323 // RewriteExprTree - Now that the operands for this expression tree are
324 // linearized and optimized, emit them in-order. This function is written to be
326 void Reassociate::RewriteExprTree(BinaryOperator *I,
327 std::vector<ValueEntry> &Ops,
329 if (i+2 == Ops.size()) {
330 if (I->getOperand(0) != Ops[i].Op ||
331 I->getOperand(1) != Ops[i+1].Op) {
332 Value *OldLHS = I->getOperand(0);
333 DEBUG(errs() << "RA: " << *I << '\n');
334 I->setOperand(0, Ops[i].Op);
335 I->setOperand(1, Ops[i+1].Op);
336 DEBUG(errs() << "TO: " << *I << '\n');
340 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3)
341 // delete the extra, now dead, nodes.
342 RemoveDeadBinaryOp(OldLHS);
346 assert(i+2 < Ops.size() && "Ops index out of range!");
348 if (I->getOperand(1) != Ops[i].Op) {
349 DEBUG(errs() << "RA: " << *I << '\n');
350 I->setOperand(1, Ops[i].Op);
351 DEBUG(errs() << "TO: " << *I << '\n');
356 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
357 assert(LHS->getOpcode() == I->getOpcode() &&
358 "Improper expression tree!");
360 // Compactify the tree instructions together with each other to guarantee
361 // that the expression tree is dominated by all of Ops.
363 RewriteExprTree(LHS, Ops, i+1);
368 // NegateValue - Insert instructions before the instruction pointed to by BI,
369 // that computes the negative version of the value specified. The negative
370 // version of the value is returned, and BI is left pointing at the instruction
371 // that should be processed next by the reassociation pass.
373 static Value *NegateValue(Value *V, Instruction *BI) {
374 // We are trying to expose opportunity for reassociation. One of the things
375 // that we want to do to achieve this is to push a negation as deep into an
376 // expression chain as possible, to expose the add instructions. In practice,
377 // this means that we turn this:
378 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
379 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
380 // the constants. We assume that instcombine will clean up the mess later if
381 // we introduce tons of unnecessary negation instructions...
383 if (Instruction *I = dyn_cast<Instruction>(V))
384 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
385 // Push the negates through the add.
386 I->setOperand(0, NegateValue(I->getOperand(0), BI));
387 I->setOperand(1, NegateValue(I->getOperand(1), BI));
389 // We must move the add instruction here, because the neg instructions do
390 // not dominate the old add instruction in general. By moving it, we are
391 // assured that the neg instructions we just inserted dominate the
392 // instruction we are about to insert after them.
395 I->setName(I->getName()+".neg");
399 // Insert a 'neg' instruction that subtracts the value from zero to get the
402 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
405 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
406 /// X-Y into (X + -Y).
407 static bool ShouldBreakUpSubtract(Instruction *Sub) {
408 // If this is a negation, we can't split it up!
409 if (BinaryOperator::isNeg(Sub))
412 // Don't bother to break this up unless either the LHS is an associable add or
413 // subtract or if this is only used by one.
414 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
415 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
417 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
418 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
420 if (Sub->hasOneUse() &&
421 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
422 isReassociableOp(Sub->use_back(), Instruction::Sub)))
428 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
429 /// only used by an add, transform this into (X+(0-Y)) to promote better
431 static Instruction *BreakUpSubtract(Instruction *Sub,
432 std::map<AssertingVH<>, unsigned> &ValueRankMap) {
433 // Convert a subtract into an add and a neg instruction... so that sub
434 // instructions can be commuted with other add instructions...
436 // Calculate the negative value of Operand 1 of the sub instruction...
437 // and set it as the RHS of the add instruction we just made...
439 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
441 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
444 // Everyone now refers to the add instruction.
445 ValueRankMap.erase(Sub);
446 Sub->replaceAllUsesWith(New);
447 Sub->eraseFromParent();
449 DEBUG(errs() << "Negated: " << *New << '\n');
453 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
454 /// by one, change this into a multiply by a constant to assist with further
456 static Instruction *ConvertShiftToMul(Instruction *Shl,
457 std::map<AssertingVH<>, unsigned> &ValueRankMap) {
458 // If an operand of this shift is a reassociable multiply, or if the shift
459 // is used by a reassociable multiply or add, turn into a multiply.
460 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) ||
462 (isReassociableOp(Shl->use_back(), Instruction::Mul) ||
463 isReassociableOp(Shl->use_back(), Instruction::Add)))) {
464 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
466 ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
468 Instruction *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst,
470 ValueRankMap.erase(Shl);
472 Shl->replaceAllUsesWith(Mul);
473 Shl->eraseFromParent();
479 // Scan backwards and forwards among values with the same rank as element i to
480 // see if X exists. If X does not exist, return i.
481 static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i,
483 unsigned XRank = Ops[i].Rank;
484 unsigned e = Ops.size();
485 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
489 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
495 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
496 /// and returning the result. Insert the tree before I.
497 static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl<Value*> &Ops){
498 if (Ops.size() == 1) return Ops.back();
500 Value *V1 = Ops.back();
502 Value *V2 = EmitAddTreeOfValues(I, Ops);
503 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
506 /// RemoveFactorFromExpression - If V is an expression tree that is a
507 /// multiplication sequence, and if this sequence contains a multiply by Factor,
508 /// remove Factor from the tree and return the new tree.
509 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
510 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
513 std::vector<ValueEntry> Factors;
514 LinearizeExprTree(BO, Factors);
516 bool FoundFactor = false;
517 for (unsigned i = 0, e = Factors.size(); i != e; ++i)
518 if (Factors[i].Op == Factor) {
520 Factors.erase(Factors.begin()+i);
524 // Make sure to restore the operands to the expression tree.
525 RewriteExprTree(BO, Factors);
529 if (Factors.size() == 1) return Factors[0].Op;
531 RewriteExprTree(BO, Factors);
535 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
536 /// add its operands as factors, otherwise add V to the list of factors.
537 static void FindSingleUseMultiplyFactors(Value *V,
538 SmallVectorImpl<Value*> &Factors) {
540 if ((!V->hasOneUse() && !V->use_empty()) ||
541 !(BO = dyn_cast<BinaryOperator>(V)) ||
542 BO->getOpcode() != Instruction::Mul) {
543 Factors.push_back(V);
547 // Otherwise, add the LHS and RHS to the list of factors.
548 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
549 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
554 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
555 std::vector<ValueEntry> &Ops) {
556 // Now that we have the linearized expression tree, try to optimize it.
557 // Start by folding any constants that we found.
558 bool IterateOptimization = false;
559 if (Ops.size() == 1) return Ops[0].Op;
561 unsigned Opcode = I->getOpcode();
563 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
564 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
566 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
567 return OptimizeExpression(I, Ops);
570 // Check for destructive annihilation due to a constant being used.
571 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
574 case Instruction::And:
575 if (CstVal->isZero()) { // ... & 0 -> 0
579 if (CstVal->isAllOnesValue()) // ... & -1 -> ...
582 case Instruction::Mul:
583 if (CstVal->isZero()) { // ... * 0 -> 0
588 if (cast<ConstantInt>(CstVal)->isOne())
589 Ops.pop_back(); // ... * 1 -> ...
591 case Instruction::Or:
592 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1
597 case Instruction::Add:
598 case Instruction::Xor:
599 if (CstVal->isZero()) // ... [|^+] 0 -> ...
603 if (Ops.size() == 1) return Ops[0].Op;
605 // Handle destructive annihilation due to identities between elements in the
606 // argument list here.
609 case Instruction::And:
610 case Instruction::Or:
611 case Instruction::Xor:
612 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
613 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
614 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
615 // First, check for X and ~X in the operand list.
616 assert(i < Ops.size());
617 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
618 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
619 unsigned FoundX = FindInOperandList(Ops, i, X);
621 if (Opcode == Instruction::And) { // ...&X&~X = 0
623 return Constant::getNullValue(X->getType());
626 if (Opcode == Instruction::Or) { // ...|X|~X = -1
628 return Constant::getAllOnesValue(X->getType());
633 // Next, check for duplicate pairs of values, which we assume are next to
634 // each other, due to our sorting criteria.
635 assert(i < Ops.size());
636 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
637 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
638 // Drop duplicate values.
639 Ops.erase(Ops.begin()+i);
641 IterateOptimization = true;
644 assert(Opcode == Instruction::Xor);
647 return Constant::getNullValue(Ops[0].Op->getType());
650 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
652 IterateOptimization = true;
659 case Instruction::Add:
660 // Scan the operand lists looking for X and -X pairs. If we find any, we
661 // can simplify the expression. X+-X == 0.
662 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
663 assert(i < Ops.size());
664 // Check for X and -X in the operand list.
665 if (!BinaryOperator::isNeg(Ops[i].Op))
668 Value *X = BinaryOperator::getNegArgument(Ops[i].Op);
669 unsigned FoundX = FindInOperandList(Ops, i, X);
673 // Remove X and -X from the operand list.
674 if (Ops.size() == 2) {
676 return Constant::getNullValue(X->getType());
678 Ops.erase(Ops.begin()+i);
682 --i; // Need to back up an extra one.
683 Ops.erase(Ops.begin()+FoundX);
684 IterateOptimization = true;
686 --i; // Revisit element.
687 e -= 2; // Removed two elements.
690 // Scan the operand list, checking to see if there are any common factors
691 // between operands. Consider something like A*A+A*B*C+D. We would like to
692 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
693 // To efficiently find this, we count the number of times a factor occurs
694 // for any ADD operands that are MULs.
695 DenseMap<Value*, unsigned> FactorOccurrences;
697 Value *MaxOccVal = 0;
698 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
699 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op);
700 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty())
703 // Compute all of the factors of this added value.
704 SmallVector<Value*, 8> Factors;
705 FindSingleUseMultiplyFactors(BOp, Factors);
706 assert(Factors.size() > 1 && "Bad linearize!");
708 // Add one to FactorOccurrences for each unique factor in this op.
709 if (Factors.size() == 2) {
710 unsigned Occ = ++FactorOccurrences[Factors[0]];
711 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; }
712 if (Factors[0] != Factors[1]) { // Don't double count A*A.
713 Occ = ++FactorOccurrences[Factors[1]];
714 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; }
717 SmallPtrSet<Value*, 4> Duplicates;
718 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
719 if (!Duplicates.insert(Factors[i])) continue;
721 unsigned Occ = ++FactorOccurrences[Factors[i]];
722 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; }
727 // If any factor occurred more than one time, we can pull it out.
729 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n");
731 // Create a new instruction that uses the MaxOccVal twice. If we don't do
732 // this, we could otherwise run into situations where removing a factor
733 // from an expression will drop a use of maxocc, and this can cause
734 // RemoveFactorFromExpression on successive values to behave differently.
735 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
736 SmallVector<Value*, 4> NewMulOps;
737 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
738 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
739 NewMulOps.push_back(V);
740 Ops.erase(Ops.begin()+i);
745 // No need for extra uses anymore.
748 unsigned NumAddedValues = NewMulOps.size();
749 Value *V = EmitAddTreeOfValues(I, NewMulOps);
750 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
752 // Now that we have inserted V and its sole use, optimize it. This allows
753 // us to handle cases that require multiple factoring steps, such as this:
754 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
755 if (NumAddedValues > 1)
756 ReassociateExpression(cast<BinaryOperator>(V));
763 // Add the new value to the list of things being added.
764 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
766 // Rewrite the tree so that there is now a use of V.
767 RewriteExprTree(I, Ops);
768 return OptimizeExpression(I, Ops);
771 //case Instruction::Mul:
774 if (IterateOptimization)
775 return OptimizeExpression(I, Ops);
780 /// ReassociateBB - Inspect all of the instructions in this basic block,
781 /// reassociating them as we go.
782 void Reassociate::ReassociateBB(BasicBlock *BB) {
783 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) {
784 Instruction *BI = BBI++;
785 if (BI->getOpcode() == Instruction::Shl &&
786 isa<ConstantInt>(BI->getOperand(1)))
787 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) {
792 // Reject cases where it is pointless to do this.
793 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() ||
794 isa<VectorType>(BI->getType()))
795 continue; // Floating point ops are not associative.
797 // If this is a subtract instruction which is not already in negate form,
798 // see if we can convert it to X+-Y.
799 if (BI->getOpcode() == Instruction::Sub) {
800 if (ShouldBreakUpSubtract(BI)) {
801 BI = BreakUpSubtract(BI, ValueRankMap);
803 } else if (BinaryOperator::isNeg(BI)) {
804 // Otherwise, this is a negation. See if the operand is a multiply tree
805 // and if this is not an inner node of a multiply tree.
806 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) &&
808 !isReassociableOp(BI->use_back(), Instruction::Mul))) {
809 BI = LowerNegateToMultiply(BI, ValueRankMap);
815 // If this instruction is a commutative binary operator, process it.
816 if (!BI->isAssociative()) continue;
817 BinaryOperator *I = cast<BinaryOperator>(BI);
819 // If this is an interior node of a reassociable tree, ignore it until we
820 // get to the root of the tree, to avoid N^2 analysis.
821 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode()))
824 // If this is an add tree that is used by a sub instruction, ignore it
825 // until we process the subtract.
826 if (I->hasOneUse() && I->getOpcode() == Instruction::Add &&
827 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub)
830 ReassociateExpression(I);
834 void Reassociate::ReassociateExpression(BinaryOperator *I) {
836 // First, walk the expression tree, linearizing the tree, collecting
837 std::vector<ValueEntry> Ops;
838 LinearizeExprTree(I, Ops);
840 DEBUG(errs() << "RAIn:\t"; PrintOps(I, Ops); errs() << "\n");
842 // Now that we have linearized the tree to a list and have gathered all of
843 // the operands and their ranks, sort the operands by their rank. Use a
844 // stable_sort so that values with equal ranks will have their relative
845 // positions maintained (and so the compiler is deterministic). Note that
846 // this sorts so that the highest ranking values end up at the beginning of
848 std::stable_sort(Ops.begin(), Ops.end());
850 // OptimizeExpression - Now that we have the expression tree in a convenient
851 // sorted form, optimize it globally if possible.
852 if (Value *V = OptimizeExpression(I, Ops)) {
853 // This expression tree simplified to something that isn't a tree,
855 DEBUG(errs() << "Reassoc to scalar: " << *V << "\n");
856 I->replaceAllUsesWith(V);
857 RemoveDeadBinaryOp(I);
861 // We want to sink immediates as deeply as possible except in the case where
862 // this is a multiply tree used only by an add, and the immediate is a -1.
863 // In this case we reassociate to put the negation on the outside so that we
864 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
865 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
866 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
867 isa<ConstantInt>(Ops.back().Op) &&
868 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
869 Ops.insert(Ops.begin(), Ops.back());
873 DEBUG(errs() << "RAOut:\t"; PrintOps(I, Ops); errs() << "\n");
875 if (Ops.size() == 1) {
876 // This expression tree simplified to something that isn't a tree,
878 I->replaceAllUsesWith(Ops[0].Op);
879 RemoveDeadBinaryOp(I);
881 // Now that we ordered and optimized the expressions, splat them back into
882 // the expression tree, removing any unneeded nodes.
883 RewriteExprTree(I, Ops);
888 bool Reassociate::runOnFunction(Function &F) {
889 // Recalculate the rank map for F
893 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
896 // We are done with the rank map...
898 ValueRankMap.clear();