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/Pass.h"
30 #include "llvm/Assembly/Writer.h"
31 #include "llvm/Support/CFG.h"
32 #include "llvm/Support/Compiler.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/ADT/PostOrderIterator.h"
35 #include "llvm/ADT/Statistic.h"
40 STATISTIC(NumLinear , "Number of insts linearized");
41 STATISTIC(NumChanged, "Number of insts reassociated");
42 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
43 STATISTIC(NumFactor , "Number of multiplies factored");
46 struct VISIBILITY_HIDDEN ValueEntry {
49 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {}
51 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) {
52 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start.
56 /// PrintOps - Print out the expression identified in the Ops list.
58 static void PrintOps(Instruction *I, const std::vector<ValueEntry> &Ops) {
59 Module *M = I->getParent()->getParent()->getParent();
60 cerr << Instruction::getOpcodeName(I->getOpcode()) << " "
61 << *Ops[0].Op->getType();
62 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
63 WriteAsOperand(*cerr.stream() << " ", Ops[i].Op, false, M)
64 << "," << Ops[i].Rank;
68 class VISIBILITY_HIDDEN Reassociate : public FunctionPass {
69 std::map<BasicBlock*, unsigned> RankMap;
70 std::map<Value*, unsigned> ValueRankMap;
73 static char ID; // Pass identification, replacement for typeid
74 Reassociate() : FunctionPass((intptr_t)&ID) {}
76 bool runOnFunction(Function &F);
78 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
82 void BuildRankMap(Function &F);
83 unsigned getRank(Value *V);
84 void ReassociateExpression(BinaryOperator *I);
85 void RewriteExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops,
87 Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops);
88 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops);
89 void LinearizeExpr(BinaryOperator *I);
90 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
91 void ReassociateBB(BasicBlock *BB);
93 void RemoveDeadBinaryOp(Value *V);
97 char Reassociate::ID = 0;
98 static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions");
100 // Public interface to the Reassociate pass
101 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
103 void Reassociate::RemoveDeadBinaryOp(Value *V) {
104 Instruction *Op = dyn_cast<Instruction>(V);
105 if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty())
108 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1);
109 RemoveDeadBinaryOp(LHS);
110 RemoveDeadBinaryOp(RHS);
114 static bool isUnmovableInstruction(Instruction *I) {
115 if (I->getOpcode() == Instruction::PHI ||
116 I->getOpcode() == Instruction::Alloca ||
117 I->getOpcode() == Instruction::Load ||
118 I->getOpcode() == Instruction::Malloc ||
119 I->getOpcode() == Instruction::Invoke ||
120 I->getOpcode() == Instruction::Call ||
121 I->getOpcode() == Instruction::UDiv ||
122 I->getOpcode() == Instruction::SDiv ||
123 I->getOpcode() == Instruction::FDiv ||
124 I->getOpcode() == Instruction::URem ||
125 I->getOpcode() == Instruction::SRem ||
126 I->getOpcode() == Instruction::FRem)
131 void Reassociate::BuildRankMap(Function &F) {
134 // Assign distinct ranks to function arguments
135 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
136 ValueRankMap[I] = ++i;
138 ReversePostOrderTraversal<Function*> RPOT(&F);
139 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
140 E = RPOT.end(); I != E; ++I) {
142 unsigned BBRank = RankMap[BB] = ++i << 16;
144 // Walk the basic block, adding precomputed ranks for any instructions that
145 // we cannot move. This ensures that the ranks for these instructions are
146 // all different in the block.
147 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
148 if (isUnmovableInstruction(I))
149 ValueRankMap[I] = ++BBRank;
153 unsigned Reassociate::getRank(Value *V) {
154 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument...
156 Instruction *I = dyn_cast<Instruction>(V);
157 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0.
159 unsigned &CachedRank = ValueRankMap[I];
160 if (CachedRank) return CachedRank; // Rank already known?
162 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
163 // we can reassociate expressions for code motion! Since we do not recurse
164 // for PHI nodes, we cannot have infinite recursion here, because there
165 // cannot be loops in the value graph that do not go through PHI nodes.
166 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
167 for (unsigned i = 0, e = I->getNumOperands();
168 i != e && Rank != MaxRank; ++i)
169 Rank = std::max(Rank, getRank(I->getOperand(i)));
171 // If this is a not or neg instruction, do not count it for rank. This
172 // assures us that X and ~X will have the same rank.
173 if (!I->getType()->isInteger() ||
174 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
177 //DOUT << "Calculated Rank[" << V->getName() << "] = "
180 return CachedRank = Rank;
183 /// isReassociableOp - Return true if V is an instruction of the specified
184 /// opcode and if it only has one use.
185 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
186 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) &&
187 cast<Instruction>(V)->getOpcode() == Opcode)
188 return cast<BinaryOperator>(V);
192 /// LowerNegateToMultiply - Replace 0-X with X*-1.
194 static Instruction *LowerNegateToMultiply(Instruction *Neg) {
195 Constant *Cst = ConstantInt::getAllOnesValue(Neg->getType());
197 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
199 Neg->replaceAllUsesWith(Res);
200 Neg->eraseFromParent();
204 // Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'.
205 // Note that if D is also part of the expression tree that we recurse to
206 // linearize it as well. Besides that case, this does not recurse into A,B, or
208 void Reassociate::LinearizeExpr(BinaryOperator *I) {
209 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
210 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1));
211 assert(isReassociableOp(LHS, I->getOpcode()) &&
212 isReassociableOp(RHS, I->getOpcode()) &&
213 "Not an expression that needs linearization?");
215 DOUT << "Linear" << *LHS << *RHS << *I;
217 // Move the RHS instruction to live immediately before I, avoiding breaking
218 // dominator properties.
221 // Move operands around to do the linearization.
222 I->setOperand(1, RHS->getOperand(0));
223 RHS->setOperand(0, LHS);
224 I->setOperand(0, RHS);
228 DOUT << "Linearized: " << *I;
230 // If D is part of this expression tree, tail recurse.
231 if (isReassociableOp(I->getOperand(1), I->getOpcode()))
236 /// LinearizeExprTree - Given an associative binary expression tree, traverse
237 /// all of the uses putting it into canonical form. This forces a left-linear
238 /// form of the the expression (((a+b)+c)+d), and collects information about the
239 /// rank of the non-tree operands.
241 /// NOTE: These intentionally destroys the expression tree operands (turning
242 /// them into undef values) to reduce #uses of the values. This means that the
243 /// caller MUST use something like RewriteExprTree to put the values back in.
245 void Reassociate::LinearizeExprTree(BinaryOperator *I,
246 std::vector<ValueEntry> &Ops) {
247 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
248 unsigned Opcode = I->getOpcode();
250 // First step, linearize the expression if it is in ((A+B)+(C+D)) form.
251 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode);
252 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode);
254 // If this is a multiply expression tree and it contains internal negations,
255 // transform them into multiplies by -1 so they can be reassociated.
256 if (I->getOpcode() == Instruction::Mul) {
257 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) {
258 LHS = LowerNegateToMultiply(cast<Instruction>(LHS));
259 LHSBO = isReassociableOp(LHS, Opcode);
261 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) {
262 RHS = LowerNegateToMultiply(cast<Instruction>(RHS));
263 RHSBO = isReassociableOp(RHS, Opcode);
269 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As
270 // such, just remember these operands and their rank.
271 Ops.push_back(ValueEntry(getRank(LHS), LHS));
272 Ops.push_back(ValueEntry(getRank(RHS), RHS));
274 // Clear the leaves out.
275 I->setOperand(0, UndefValue::get(I->getType()));
276 I->setOperand(1, UndefValue::get(I->getType()));
279 // Turn X+(Y+Z) -> (Y+Z)+X
280 std::swap(LHSBO, RHSBO);
282 bool Success = !I->swapOperands();
283 assert(Success && "swapOperands failed");
287 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not
288 // part of the expression tree.
290 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0));
291 RHS = I->getOperand(1);
295 // Okay, now we know that the LHS is a nested expression and that the RHS is
296 // not. Perform reassociation.
297 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!");
299 // Move LHS right before I to make sure that the tree expression dominates all
301 LHSBO->moveBefore(I);
303 // Linearize the expression tree on the LHS.
304 LinearizeExprTree(LHSBO, Ops);
306 // Remember the RHS operand and its rank.
307 Ops.push_back(ValueEntry(getRank(RHS), RHS));
309 // Clear the RHS leaf out.
310 I->setOperand(1, UndefValue::get(I->getType()));
313 // RewriteExprTree - Now that the operands for this expression tree are
314 // linearized and optimized, emit them in-order. This function is written to be
316 void Reassociate::RewriteExprTree(BinaryOperator *I,
317 std::vector<ValueEntry> &Ops,
319 if (i+2 == Ops.size()) {
320 if (I->getOperand(0) != Ops[i].Op ||
321 I->getOperand(1) != Ops[i+1].Op) {
322 Value *OldLHS = I->getOperand(0);
323 DOUT << "RA: " << *I;
324 I->setOperand(0, Ops[i].Op);
325 I->setOperand(1, Ops[i+1].Op);
326 DOUT << "TO: " << *I;
330 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3)
331 // delete the extra, now dead, nodes.
332 RemoveDeadBinaryOp(OldLHS);
336 assert(i+2 < Ops.size() && "Ops index out of range!");
338 if (I->getOperand(1) != Ops[i].Op) {
339 DOUT << "RA: " << *I;
340 I->setOperand(1, Ops[i].Op);
341 DOUT << "TO: " << *I;
346 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
347 assert(LHS->getOpcode() == I->getOpcode() &&
348 "Improper expression tree!");
350 // Compactify the tree instructions together with each other to guarantee
351 // that the expression tree is dominated by all of Ops.
353 RewriteExprTree(LHS, Ops, i+1);
358 // NegateValue - Insert instructions before the instruction pointed to by BI,
359 // that computes the negative version of the value specified. The negative
360 // version of the value is returned, and BI is left pointing at the instruction
361 // that should be processed next by the reassociation pass.
363 static Value *NegateValue(Value *V, Instruction *BI) {
364 // We are trying to expose opportunity for reassociation. One of the things
365 // that we want to do to achieve this is to push a negation as deep into an
366 // expression chain as possible, to expose the add instructions. In practice,
367 // this means that we turn this:
368 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
369 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
370 // the constants. We assume that instcombine will clean up the mess later if
371 // we introduce tons of unnecessary negation instructions...
373 if (Instruction *I = dyn_cast<Instruction>(V))
374 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
375 // Push the negates through the add.
376 I->setOperand(0, NegateValue(I->getOperand(0), BI));
377 I->setOperand(1, NegateValue(I->getOperand(1), BI));
379 // We must move the add instruction here, because the neg instructions do
380 // not dominate the old add instruction in general. By moving it, we are
381 // assured that the neg instructions we just inserted dominate the
382 // instruction we are about to insert after them.
385 I->setName(I->getName()+".neg");
389 // Insert a 'neg' instruction that subtracts the value from zero to get the
392 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
395 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
396 /// X-Y into (X + -Y).
397 static bool ShouldBreakUpSubtract(Instruction *Sub) {
398 // If this is a negation, we can't split it up!
399 if (BinaryOperator::isNeg(Sub))
402 // Don't bother to break this up unless either the LHS is an associable add or
403 // subtract or if this is only used by one.
404 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
405 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
407 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
408 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
410 if (Sub->hasOneUse() &&
411 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
412 isReassociableOp(Sub->use_back(), Instruction::Sub)))
418 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
419 /// only used by an add, transform this into (X+(0-Y)) to promote better
421 static Instruction *BreakUpSubtract(Instruction *Sub) {
422 // Convert a subtract into an add and a neg instruction... so that sub
423 // instructions can be commuted with other add instructions...
425 // Calculate the negative value of Operand 1 of the sub instruction...
426 // and set it as the RHS of the add instruction we just made...
428 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
430 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
433 // Everyone now refers to the add instruction.
434 Sub->replaceAllUsesWith(New);
435 Sub->eraseFromParent();
437 DOUT << "Negated: " << *New;
441 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
442 /// by one, change this into a multiply by a constant to assist with further
444 static Instruction *ConvertShiftToMul(Instruction *Shl) {
445 // If an operand of this shift is a reassociable multiply, or if the shift
446 // is used by a reassociable multiply or add, turn into a multiply.
447 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) ||
449 (isReassociableOp(Shl->use_back(), Instruction::Mul) ||
450 isReassociableOp(Shl->use_back(), Instruction::Add)))) {
451 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
452 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
454 Instruction *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst,
457 Shl->replaceAllUsesWith(Mul);
458 Shl->eraseFromParent();
464 // Scan backwards and forwards among values with the same rank as element i to
465 // see if X exists. If X does not exist, return i.
466 static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i,
468 unsigned XRank = Ops[i].Rank;
469 unsigned e = Ops.size();
470 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
474 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
480 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
481 /// and returning the result. Insert the tree before I.
482 static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) {
483 if (Ops.size() == 1) return Ops.back();
485 Value *V1 = Ops.back();
487 Value *V2 = EmitAddTreeOfValues(I, Ops);
488 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
491 /// RemoveFactorFromExpression - If V is an expression tree that is a
492 /// multiplication sequence, and if this sequence contains a multiply by Factor,
493 /// remove Factor from the tree and return the new tree.
494 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
495 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
498 std::vector<ValueEntry> Factors;
499 LinearizeExprTree(BO, Factors);
501 bool FoundFactor = false;
502 for (unsigned i = 0, e = Factors.size(); i != e; ++i)
503 if (Factors[i].Op == Factor) {
505 Factors.erase(Factors.begin()+i);
509 // Make sure to restore the operands to the expression tree.
510 RewriteExprTree(BO, Factors);
514 if (Factors.size() == 1) return Factors[0].Op;
516 RewriteExprTree(BO, Factors);
520 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
521 /// add its operands as factors, otherwise add V to the list of factors.
522 static void FindSingleUseMultiplyFactors(Value *V,
523 std::vector<Value*> &Factors) {
525 if ((!V->hasOneUse() && !V->use_empty()) ||
526 !(BO = dyn_cast<BinaryOperator>(V)) ||
527 BO->getOpcode() != Instruction::Mul) {
528 Factors.push_back(V);
532 // Otherwise, add the LHS and RHS to the list of factors.
533 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
534 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
539 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
540 std::vector<ValueEntry> &Ops) {
541 // Now that we have the linearized expression tree, try to optimize it.
542 // Start by folding any constants that we found.
543 bool IterateOptimization = false;
544 if (Ops.size() == 1) return Ops[0].Op;
546 unsigned Opcode = I->getOpcode();
548 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
549 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
551 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
552 return OptimizeExpression(I, Ops);
555 // Check for destructive annihilation due to a constant being used.
556 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
559 case Instruction::And:
560 if (CstVal->isZero()) { // ... & 0 -> 0
563 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ...
567 case Instruction::Mul:
568 if (CstVal->isZero()) { // ... * 0 -> 0
571 } else if (cast<ConstantInt>(CstVal)->isOne()) {
572 Ops.pop_back(); // ... * 1 -> ...
575 case Instruction::Or:
576 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1
581 case Instruction::Add:
582 case Instruction::Xor:
583 if (CstVal->isZero()) // ... [|^+] 0 -> ...
587 if (Ops.size() == 1) return Ops[0].Op;
589 // Handle destructive annihilation do to identities between elements in the
590 // argument list here.
593 case Instruction::And:
594 case Instruction::Or:
595 case Instruction::Xor:
596 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
597 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
598 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
599 // First, check for X and ~X in the operand list.
600 assert(i < Ops.size());
601 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
602 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
603 unsigned FoundX = FindInOperandList(Ops, i, X);
605 if (Opcode == Instruction::And) { // ...&X&~X = 0
607 return Constant::getNullValue(X->getType());
608 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1
610 return ConstantInt::getAllOnesValue(X->getType());
615 // Next, check for duplicate pairs of values, which we assume are next to
616 // each other, due to our sorting criteria.
617 assert(i < Ops.size());
618 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
619 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
620 // Drop duplicate values.
621 Ops.erase(Ops.begin()+i);
623 IterateOptimization = true;
626 assert(Opcode == Instruction::Xor);
629 return Constant::getNullValue(Ops[0].Op->getType());
632 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
634 IterateOptimization = true;
641 case Instruction::Add:
642 // Scan the operand lists looking for X and -X pairs. If we find any, we
643 // can simplify the expression. X+-X == 0.
644 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
645 assert(i < Ops.size());
646 // Check for X and -X in the operand list.
647 if (BinaryOperator::isNeg(Ops[i].Op)) {
648 Value *X = BinaryOperator::getNegArgument(Ops[i].Op);
649 unsigned FoundX = FindInOperandList(Ops, i, X);
651 // Remove X and -X from the operand list.
652 if (Ops.size() == 2) {
654 return Constant::getNullValue(X->getType());
656 Ops.erase(Ops.begin()+i);
660 --i; // Need to back up an extra one.
661 Ops.erase(Ops.begin()+FoundX);
662 IterateOptimization = true;
664 --i; // Revisit element.
665 e -= 2; // Removed two elements.
672 // Scan the operand list, checking to see if there are any common factors
673 // between operands. Consider something like A*A+A*B*C+D. We would like to
674 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
675 // To efficiently find this, we count the number of times a factor occurs
676 // for any ADD operands that are MULs.
677 std::map<Value*, unsigned> FactorOccurrences;
679 Value *MaxOccVal = 0;
680 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
681 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) {
682 if (BOp->getOpcode() == Instruction::Mul && BOp->use_empty()) {
683 // Compute all of the factors of this added value.
684 std::vector<Value*> Factors;
685 FindSingleUseMultiplyFactors(BOp, Factors);
686 assert(Factors.size() > 1 && "Bad linearize!");
688 // Add one to FactorOccurrences for each unique factor in this op.
689 if (Factors.size() == 2) {
690 unsigned Occ = ++FactorOccurrences[Factors[0]];
691 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; }
692 if (Factors[0] != Factors[1]) { // Don't double count A*A.
693 Occ = ++FactorOccurrences[Factors[1]];
694 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; }
697 std::set<Value*> Duplicates;
698 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
699 if (Duplicates.insert(Factors[i]).second) {
700 unsigned Occ = ++FactorOccurrences[Factors[i]];
701 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; }
709 // If any factor occurred more than one time, we can pull it out.
711 DOUT << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n";
713 // Create a new instruction that uses the MaxOccVal twice. If we don't do
714 // this, we could otherwise run into situations where removing a factor
715 // from an expression will drop a use of maxocc, and this can cause
716 // RemoveFactorFromExpression on successive values to behave differently.
717 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
718 std::vector<Value*> NewMulOps;
719 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
720 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
721 NewMulOps.push_back(V);
722 Ops.erase(Ops.begin()+i);
727 // No need for extra uses anymore.
730 unsigned NumAddedValues = NewMulOps.size();
731 Value *V = EmitAddTreeOfValues(I, NewMulOps);
732 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
734 // Now that we have inserted V and its sole use, optimize it. This allows
735 // us to handle cases that require multiple factoring steps, such as this:
736 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
737 if (NumAddedValues > 1)
738 ReassociateExpression(cast<BinaryOperator>(V));
745 // Add the new value to the list of things being added.
746 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
748 // Rewrite the tree so that there is now a use of V.
749 RewriteExprTree(I, Ops);
750 return OptimizeExpression(I, Ops);
753 //case Instruction::Mul:
756 if (IterateOptimization)
757 return OptimizeExpression(I, Ops);
762 /// ReassociateBB - Inspect all of the instructions in this basic block,
763 /// reassociating them as we go.
764 void Reassociate::ReassociateBB(BasicBlock *BB) {
765 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) {
766 Instruction *BI = BBI++;
767 if (BI->getOpcode() == Instruction::Shl &&
768 isa<ConstantInt>(BI->getOperand(1)))
769 if (Instruction *NI = ConvertShiftToMul(BI)) {
774 // Reject cases where it is pointless to do this.
775 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() ||
776 isa<VectorType>(BI->getType()))
777 continue; // Floating point ops are not associative.
779 // If this is a subtract instruction which is not already in negate form,
780 // see if we can convert it to X+-Y.
781 if (BI->getOpcode() == Instruction::Sub) {
782 if (ShouldBreakUpSubtract(BI)) {
783 BI = BreakUpSubtract(BI);
785 } else if (BinaryOperator::isNeg(BI)) {
786 // Otherwise, this is a negation. See if the operand is a multiply tree
787 // and if this is not an inner node of a multiply tree.
788 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) &&
790 !isReassociableOp(BI->use_back(), Instruction::Mul))) {
791 BI = LowerNegateToMultiply(BI);
797 // If this instruction is a commutative binary operator, process it.
798 if (!BI->isAssociative()) continue;
799 BinaryOperator *I = cast<BinaryOperator>(BI);
801 // If this is an interior node of a reassociable tree, ignore it until we
802 // get to the root of the tree, to avoid N^2 analysis.
803 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode()))
806 // If this is an add tree that is used by a sub instruction, ignore it
807 // until we process the subtract.
808 if (I->hasOneUse() && I->getOpcode() == Instruction::Add &&
809 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub)
812 ReassociateExpression(I);
816 void Reassociate::ReassociateExpression(BinaryOperator *I) {
818 // First, walk the expression tree, linearizing the tree, collecting
819 std::vector<ValueEntry> Ops;
820 LinearizeExprTree(I, Ops);
822 DOUT << "RAIn:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n";
824 // Now that we have linearized the tree to a list and have gathered all of
825 // the operands and their ranks, sort the operands by their rank. Use a
826 // stable_sort so that values with equal ranks will have their relative
827 // positions maintained (and so the compiler is deterministic). Note that
828 // this sorts so that the highest ranking values end up at the beginning of
830 std::stable_sort(Ops.begin(), Ops.end());
832 // OptimizeExpression - Now that we have the expression tree in a convenient
833 // sorted form, optimize it globally if possible.
834 if (Value *V = OptimizeExpression(I, Ops)) {
835 // This expression tree simplified to something that isn't a tree,
837 DOUT << "Reassoc to scalar: " << *V << "\n";
838 I->replaceAllUsesWith(V);
839 RemoveDeadBinaryOp(I);
843 // We want to sink immediates as deeply as possible except in the case where
844 // this is a multiply tree used only by an add, and the immediate is a -1.
845 // In this case we reassociate to put the negation on the outside so that we
846 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
847 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
848 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
849 isa<ConstantInt>(Ops.back().Op) &&
850 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
851 Ops.insert(Ops.begin(), Ops.back());
855 DOUT << "RAOut:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n";
857 if (Ops.size() == 1) {
858 // This expression tree simplified to something that isn't a tree,
860 I->replaceAllUsesWith(Ops[0].Op);
861 RemoveDeadBinaryOp(I);
863 // Now that we ordered and optimized the expressions, splat them back into
864 // the expression tree, removing any unneeded nodes.
865 RewriteExprTree(I, Ops);
870 bool Reassociate::runOnFunction(Function &F) {
871 // Recalculate the rank map for F
875 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
878 // We are done with the rank map...
880 ValueRankMap.clear();