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 cerr << "," << Ops[i].Rank;
69 class VISIBILITY_HIDDEN Reassociate : public FunctionPass {
70 std::map<BasicBlock*, unsigned> RankMap;
71 std::map<Value*, unsigned> ValueRankMap;
74 static char ID; // Pass identification, replacement for typeid
75 Reassociate() : FunctionPass(&ID) {}
77 bool runOnFunction(Function &F);
79 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
83 void BuildRankMap(Function &F);
84 unsigned getRank(Value *V);
85 void ReassociateExpression(BinaryOperator *I);
86 void RewriteExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops,
88 Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops);
89 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops);
90 void LinearizeExpr(BinaryOperator *I);
91 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
92 void ReassociateBB(BasicBlock *BB);
94 void RemoveDeadBinaryOp(Value *V);
98 char Reassociate::ID = 0;
99 static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions");
101 // Public interface to the Reassociate pass
102 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
104 void Reassociate::RemoveDeadBinaryOp(Value *V) {
105 Instruction *Op = dyn_cast<Instruction>(V);
106 if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty())
109 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1);
110 RemoveDeadBinaryOp(LHS);
111 RemoveDeadBinaryOp(RHS);
115 static bool isUnmovableInstruction(Instruction *I) {
116 if (I->getOpcode() == Instruction::PHI ||
117 I->getOpcode() == Instruction::Alloca ||
118 I->getOpcode() == Instruction::Load ||
119 I->getOpcode() == Instruction::Malloc ||
120 I->getOpcode() == Instruction::Invoke ||
121 I->getOpcode() == Instruction::Call ||
122 I->getOpcode() == Instruction::UDiv ||
123 I->getOpcode() == Instruction::SDiv ||
124 I->getOpcode() == Instruction::FDiv ||
125 I->getOpcode() == Instruction::URem ||
126 I->getOpcode() == Instruction::SRem ||
127 I->getOpcode() == Instruction::FRem)
132 void Reassociate::BuildRankMap(Function &F) {
135 // Assign distinct ranks to function arguments
136 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
137 ValueRankMap[I] = ++i;
139 ReversePostOrderTraversal<Function*> RPOT(&F);
140 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
141 E = RPOT.end(); I != E; ++I) {
143 unsigned BBRank = RankMap[BB] = ++i << 16;
145 // Walk the basic block, adding precomputed ranks for any instructions that
146 // we cannot move. This ensures that the ranks for these instructions are
147 // all different in the block.
148 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
149 if (isUnmovableInstruction(I))
150 ValueRankMap[I] = ++BBRank;
154 unsigned Reassociate::getRank(Value *V) {
155 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument...
157 Instruction *I = dyn_cast<Instruction>(V);
158 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0.
160 unsigned &CachedRank = ValueRankMap[I];
161 if (CachedRank) return CachedRank; // Rank already known?
163 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
164 // we can reassociate expressions for code motion! Since we do not recurse
165 // for PHI nodes, we cannot have infinite recursion here, because there
166 // cannot be loops in the value graph that do not go through PHI nodes.
167 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
168 for (unsigned i = 0, e = I->getNumOperands();
169 i != e && Rank != MaxRank; ++i)
170 Rank = std::max(Rank, getRank(I->getOperand(i)));
172 // If this is a not or neg instruction, do not count it for rank. This
173 // assures us that X and ~X will have the same rank.
174 if (!I->getType()->isInteger() ||
175 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
178 //DOUT << "Calculated Rank[" << V->getName() << "] = "
181 return CachedRank = Rank;
184 /// isReassociableOp - Return true if V is an instruction of the specified
185 /// opcode and if it only has one use.
186 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
187 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) &&
188 cast<Instruction>(V)->getOpcode() == Opcode)
189 return cast<BinaryOperator>(V);
193 /// LowerNegateToMultiply - Replace 0-X with X*-1.
195 static Instruction *LowerNegateToMultiply(Instruction *Neg) {
196 Constant *Cst = ConstantInt::getAllOnesValue(Neg->getType());
198 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
200 Neg->replaceAllUsesWith(Res);
201 Neg->eraseFromParent();
205 // Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'.
206 // Note that if D is also part of the expression tree that we recurse to
207 // linearize it as well. Besides that case, this does not recurse into A,B, or
209 void Reassociate::LinearizeExpr(BinaryOperator *I) {
210 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
211 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1));
212 assert(isReassociableOp(LHS, I->getOpcode()) &&
213 isReassociableOp(RHS, I->getOpcode()) &&
214 "Not an expression that needs linearization?");
216 DOUT << "Linear" << *LHS << *RHS << *I;
218 // Move the RHS instruction to live immediately before I, avoiding breaking
219 // dominator properties.
222 // Move operands around to do the linearization.
223 I->setOperand(1, RHS->getOperand(0));
224 RHS->setOperand(0, LHS);
225 I->setOperand(0, RHS);
229 DOUT << "Linearized: " << *I;
231 // If D is part of this expression tree, tail recurse.
232 if (isReassociableOp(I->getOperand(1), I->getOpcode()))
237 /// LinearizeExprTree - Given an associative binary expression tree, traverse
238 /// all of the uses putting it into canonical form. This forces a left-linear
239 /// form of the the expression (((a+b)+c)+d), and collects information about the
240 /// rank of the non-tree operands.
242 /// NOTE: These intentionally destroys the expression tree operands (turning
243 /// them into undef values) to reduce #uses of the values. This means that the
244 /// caller MUST use something like RewriteExprTree to put the values back in.
246 void Reassociate::LinearizeExprTree(BinaryOperator *I,
247 std::vector<ValueEntry> &Ops) {
248 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
249 unsigned Opcode = I->getOpcode();
251 // First step, linearize the expression if it is in ((A+B)+(C+D)) form.
252 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode);
253 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode);
255 // If this is a multiply expression tree and it contains internal negations,
256 // transform them into multiplies by -1 so they can be reassociated.
257 if (I->getOpcode() == Instruction::Mul) {
258 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) {
259 LHS = LowerNegateToMultiply(cast<Instruction>(LHS));
260 LHSBO = isReassociableOp(LHS, Opcode);
262 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) {
263 RHS = LowerNegateToMultiply(cast<Instruction>(RHS));
264 RHSBO = isReassociableOp(RHS, Opcode);
270 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As
271 // such, just remember these operands and their rank.
272 Ops.push_back(ValueEntry(getRank(LHS), LHS));
273 Ops.push_back(ValueEntry(getRank(RHS), RHS));
275 // Clear the leaves out.
276 I->setOperand(0, UndefValue::get(I->getType()));
277 I->setOperand(1, UndefValue::get(I->getType()));
280 // Turn X+(Y+Z) -> (Y+Z)+X
281 std::swap(LHSBO, RHSBO);
283 bool Success = !I->swapOperands();
284 assert(Success && "swapOperands failed");
288 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not
289 // part of the expression tree.
291 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0));
292 RHS = I->getOperand(1);
296 // Okay, now we know that the LHS is a nested expression and that the RHS is
297 // not. Perform reassociation.
298 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!");
300 // Move LHS right before I to make sure that the tree expression dominates all
302 LHSBO->moveBefore(I);
304 // Linearize the expression tree on the LHS.
305 LinearizeExprTree(LHSBO, Ops);
307 // Remember the RHS operand and its rank.
308 Ops.push_back(ValueEntry(getRank(RHS), RHS));
310 // Clear the RHS leaf out.
311 I->setOperand(1, UndefValue::get(I->getType()));
314 // RewriteExprTree - Now that the operands for this expression tree are
315 // linearized and optimized, emit them in-order. This function is written to be
317 void Reassociate::RewriteExprTree(BinaryOperator *I,
318 std::vector<ValueEntry> &Ops,
320 if (i+2 == Ops.size()) {
321 if (I->getOperand(0) != Ops[i].Op ||
322 I->getOperand(1) != Ops[i+1].Op) {
323 Value *OldLHS = I->getOperand(0);
324 DOUT << "RA: " << *I;
325 I->setOperand(0, Ops[i].Op);
326 I->setOperand(1, Ops[i+1].Op);
327 DOUT << "TO: " << *I;
331 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3)
332 // delete the extra, now dead, nodes.
333 RemoveDeadBinaryOp(OldLHS);
337 assert(i+2 < Ops.size() && "Ops index out of range!");
339 if (I->getOperand(1) != Ops[i].Op) {
340 DOUT << "RA: " << *I;
341 I->setOperand(1, Ops[i].Op);
342 DOUT << "TO: " << *I;
347 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
348 assert(LHS->getOpcode() == I->getOpcode() &&
349 "Improper expression tree!");
351 // Compactify the tree instructions together with each other to guarantee
352 // that the expression tree is dominated by all of Ops.
354 RewriteExprTree(LHS, Ops, i+1);
359 // NegateValue - Insert instructions before the instruction pointed to by BI,
360 // that computes the negative version of the value specified. The negative
361 // version of the value is returned, and BI is left pointing at the instruction
362 // that should be processed next by the reassociation pass.
364 static Value *NegateValue(Value *V, Instruction *BI) {
365 // We are trying to expose opportunity for reassociation. One of the things
366 // that we want to do to achieve this is to push a negation as deep into an
367 // expression chain as possible, to expose the add instructions. In practice,
368 // this means that we turn this:
369 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
370 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
371 // the constants. We assume that instcombine will clean up the mess later if
372 // we introduce tons of unnecessary negation instructions...
374 if (Instruction *I = dyn_cast<Instruction>(V))
375 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
376 // Push the negates through the add.
377 I->setOperand(0, NegateValue(I->getOperand(0), BI));
378 I->setOperand(1, NegateValue(I->getOperand(1), BI));
380 // We must move the add instruction here, because the neg instructions do
381 // not dominate the old add instruction in general. By moving it, we are
382 // assured that the neg instructions we just inserted dominate the
383 // instruction we are about to insert after them.
386 I->setName(I->getName()+".neg");
390 // Insert a 'neg' instruction that subtracts the value from zero to get the
393 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
396 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
397 /// X-Y into (X + -Y).
398 static bool ShouldBreakUpSubtract(Instruction *Sub) {
399 // If this is a negation, we can't split it up!
400 if (BinaryOperator::isNeg(Sub))
403 // Don't bother to break this up unless either the LHS is an associable add or
404 // subtract or if this is only used by one.
405 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
406 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
408 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
409 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
411 if (Sub->hasOneUse() &&
412 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
413 isReassociableOp(Sub->use_back(), Instruction::Sub)))
419 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
420 /// only used by an add, transform this into (X+(0-Y)) to promote better
422 static Instruction *BreakUpSubtract(Instruction *Sub) {
423 // Convert a subtract into an add and a neg instruction... so that sub
424 // instructions can be commuted with other add instructions...
426 // Calculate the negative value of Operand 1 of the sub instruction...
427 // and set it as the RHS of the add instruction we just made...
429 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
431 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
434 // Everyone now refers to the add instruction.
435 Sub->replaceAllUsesWith(New);
436 Sub->eraseFromParent();
438 DOUT << "Negated: " << *New;
442 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
443 /// by one, change this into a multiply by a constant to assist with further
445 static Instruction *ConvertShiftToMul(Instruction *Shl) {
446 // If an operand of this shift is a reassociable multiply, or if the shift
447 // is used by a reassociable multiply or add, turn into a multiply.
448 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) ||
450 (isReassociableOp(Shl->use_back(), Instruction::Mul) ||
451 isReassociableOp(Shl->use_back(), Instruction::Add)))) {
452 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
453 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
455 Instruction *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst,
458 Shl->replaceAllUsesWith(Mul);
459 Shl->eraseFromParent();
465 // Scan backwards and forwards among values with the same rank as element i to
466 // see if X exists. If X does not exist, return i.
467 static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i,
469 unsigned XRank = Ops[i].Rank;
470 unsigned e = Ops.size();
471 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
475 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
481 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
482 /// and returning the result. Insert the tree before I.
483 static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) {
484 if (Ops.size() == 1) return Ops.back();
486 Value *V1 = Ops.back();
488 Value *V2 = EmitAddTreeOfValues(I, Ops);
489 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
492 /// RemoveFactorFromExpression - If V is an expression tree that is a
493 /// multiplication sequence, and if this sequence contains a multiply by Factor,
494 /// remove Factor from the tree and return the new tree.
495 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
496 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
499 std::vector<ValueEntry> Factors;
500 LinearizeExprTree(BO, Factors);
502 bool FoundFactor = false;
503 for (unsigned i = 0, e = Factors.size(); i != e; ++i)
504 if (Factors[i].Op == Factor) {
506 Factors.erase(Factors.begin()+i);
510 // Make sure to restore the operands to the expression tree.
511 RewriteExprTree(BO, Factors);
515 if (Factors.size() == 1) return Factors[0].Op;
517 RewriteExprTree(BO, Factors);
521 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
522 /// add its operands as factors, otherwise add V to the list of factors.
523 static void FindSingleUseMultiplyFactors(Value *V,
524 std::vector<Value*> &Factors) {
526 if ((!V->hasOneUse() && !V->use_empty()) ||
527 !(BO = dyn_cast<BinaryOperator>(V)) ||
528 BO->getOpcode() != Instruction::Mul) {
529 Factors.push_back(V);
533 // Otherwise, add the LHS and RHS to the list of factors.
534 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
535 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
540 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
541 std::vector<ValueEntry> &Ops) {
542 // Now that we have the linearized expression tree, try to optimize it.
543 // Start by folding any constants that we found.
544 bool IterateOptimization = false;
545 if (Ops.size() == 1) return Ops[0].Op;
547 unsigned Opcode = I->getOpcode();
549 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
550 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
552 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
553 return OptimizeExpression(I, Ops);
556 // Check for destructive annihilation due to a constant being used.
557 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
560 case Instruction::And:
561 if (CstVal->isZero()) { // ... & 0 -> 0
564 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ...
568 case Instruction::Mul:
569 if (CstVal->isZero()) { // ... * 0 -> 0
572 } else if (cast<ConstantInt>(CstVal)->isOne()) {
573 Ops.pop_back(); // ... * 1 -> ...
576 case Instruction::Or:
577 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1
582 case Instruction::Add:
583 case Instruction::Xor:
584 if (CstVal->isZero()) // ... [|^+] 0 -> ...
588 if (Ops.size() == 1) return Ops[0].Op;
590 // Handle destructive annihilation do to identities between elements in the
591 // argument list here.
594 case Instruction::And:
595 case Instruction::Or:
596 case Instruction::Xor:
597 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
598 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
599 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
600 // First, check for X and ~X in the operand list.
601 assert(i < Ops.size());
602 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
603 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
604 unsigned FoundX = FindInOperandList(Ops, i, X);
606 if (Opcode == Instruction::And) { // ...&X&~X = 0
608 return Constant::getNullValue(X->getType());
609 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1
611 return ConstantInt::getAllOnesValue(X->getType());
616 // Next, check for duplicate pairs of values, which we assume are next to
617 // each other, due to our sorting criteria.
618 assert(i < Ops.size());
619 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
620 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
621 // Drop duplicate values.
622 Ops.erase(Ops.begin()+i);
624 IterateOptimization = true;
627 assert(Opcode == Instruction::Xor);
630 return Constant::getNullValue(Ops[0].Op->getType());
633 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
635 IterateOptimization = true;
642 case Instruction::Add:
643 // Scan the operand lists looking for X and -X pairs. If we find any, we
644 // can simplify the expression. X+-X == 0.
645 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
646 assert(i < Ops.size());
647 // Check for X and -X in the operand list.
648 if (BinaryOperator::isNeg(Ops[i].Op)) {
649 Value *X = BinaryOperator::getNegArgument(Ops[i].Op);
650 unsigned FoundX = FindInOperandList(Ops, i, X);
652 // Remove X and -X from the operand list.
653 if (Ops.size() == 2) {
655 return Constant::getNullValue(X->getType());
657 Ops.erase(Ops.begin()+i);
661 --i; // Need to back up an extra one.
662 Ops.erase(Ops.begin()+FoundX);
663 IterateOptimization = true;
665 --i; // Revisit element.
666 e -= 2; // Removed two elements.
673 // Scan the operand list, checking to see if there are any common factors
674 // between operands. Consider something like A*A+A*B*C+D. We would like to
675 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
676 // To efficiently find this, we count the number of times a factor occurs
677 // for any ADD operands that are MULs.
678 std::map<Value*, unsigned> FactorOccurrences;
680 Value *MaxOccVal = 0;
681 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
682 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) {
683 if (BOp->getOpcode() == Instruction::Mul && BOp->use_empty()) {
684 // Compute all of the factors of this added value.
685 std::vector<Value*> Factors;
686 FindSingleUseMultiplyFactors(BOp, Factors);
687 assert(Factors.size() > 1 && "Bad linearize!");
689 // Add one to FactorOccurrences for each unique factor in this op.
690 if (Factors.size() == 2) {
691 unsigned Occ = ++FactorOccurrences[Factors[0]];
692 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; }
693 if (Factors[0] != Factors[1]) { // Don't double count A*A.
694 Occ = ++FactorOccurrences[Factors[1]];
695 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; }
698 std::set<Value*> Duplicates;
699 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
700 if (Duplicates.insert(Factors[i]).second) {
701 unsigned Occ = ++FactorOccurrences[Factors[i]];
702 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; }
710 // If any factor occurred more than one time, we can pull it out.
712 DOUT << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n";
714 // Create a new instruction that uses the MaxOccVal twice. If we don't do
715 // this, we could otherwise run into situations where removing a factor
716 // from an expression will drop a use of maxocc, and this can cause
717 // RemoveFactorFromExpression on successive values to behave differently.
718 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
719 std::vector<Value*> NewMulOps;
720 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
721 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
722 NewMulOps.push_back(V);
723 Ops.erase(Ops.begin()+i);
728 // No need for extra uses anymore.
731 unsigned NumAddedValues = NewMulOps.size();
732 Value *V = EmitAddTreeOfValues(I, NewMulOps);
733 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
735 // Now that we have inserted V and its sole use, optimize it. This allows
736 // us to handle cases that require multiple factoring steps, such as this:
737 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
738 if (NumAddedValues > 1)
739 ReassociateExpression(cast<BinaryOperator>(V));
746 // Add the new value to the list of things being added.
747 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
749 // Rewrite the tree so that there is now a use of V.
750 RewriteExprTree(I, Ops);
751 return OptimizeExpression(I, Ops);
754 //case Instruction::Mul:
757 if (IterateOptimization)
758 return OptimizeExpression(I, Ops);
763 /// ReassociateBB - Inspect all of the instructions in this basic block,
764 /// reassociating them as we go.
765 void Reassociate::ReassociateBB(BasicBlock *BB) {
766 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) {
767 Instruction *BI = BBI++;
768 if (BI->getOpcode() == Instruction::Shl &&
769 isa<ConstantInt>(BI->getOperand(1)))
770 if (Instruction *NI = ConvertShiftToMul(BI)) {
775 // Reject cases where it is pointless to do this.
776 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() ||
777 isa<VectorType>(BI->getType()))
778 continue; // Floating point ops are not associative.
780 // If this is a subtract instruction which is not already in negate form,
781 // see if we can convert it to X+-Y.
782 if (BI->getOpcode() == Instruction::Sub) {
783 if (ShouldBreakUpSubtract(BI)) {
784 BI = BreakUpSubtract(BI);
786 } else if (BinaryOperator::isNeg(BI)) {
787 // Otherwise, this is a negation. See if the operand is a multiply tree
788 // and if this is not an inner node of a multiply tree.
789 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) &&
791 !isReassociableOp(BI->use_back(), Instruction::Mul))) {
792 BI = LowerNegateToMultiply(BI);
798 // If this instruction is a commutative binary operator, process it.
799 if (!BI->isAssociative()) continue;
800 BinaryOperator *I = cast<BinaryOperator>(BI);
802 // If this is an interior node of a reassociable tree, ignore it until we
803 // get to the root of the tree, to avoid N^2 analysis.
804 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode()))
807 // If this is an add tree that is used by a sub instruction, ignore it
808 // until we process the subtract.
809 if (I->hasOneUse() && I->getOpcode() == Instruction::Add &&
810 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub)
813 ReassociateExpression(I);
817 void Reassociate::ReassociateExpression(BinaryOperator *I) {
819 // First, walk the expression tree, linearizing the tree, collecting
820 std::vector<ValueEntry> Ops;
821 LinearizeExprTree(I, Ops);
823 DOUT << "RAIn:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n";
825 // Now that we have linearized the tree to a list and have gathered all of
826 // the operands and their ranks, sort the operands by their rank. Use a
827 // stable_sort so that values with equal ranks will have their relative
828 // positions maintained (and so the compiler is deterministic). Note that
829 // this sorts so that the highest ranking values end up at the beginning of
831 std::stable_sort(Ops.begin(), Ops.end());
833 // OptimizeExpression - Now that we have the expression tree in a convenient
834 // sorted form, optimize it globally if possible.
835 if (Value *V = OptimizeExpression(I, Ops)) {
836 // This expression tree simplified to something that isn't a tree,
838 DOUT << "Reassoc to scalar: " << *V << "\n";
839 I->replaceAllUsesWith(V);
840 RemoveDeadBinaryOp(I);
844 // We want to sink immediates as deeply as possible except in the case where
845 // this is a multiply tree used only by an add, and the immediate is a -1.
846 // In this case we reassociate to put the negation on the outside so that we
847 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
848 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
849 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
850 isa<ConstantInt>(Ops.back().Op) &&
851 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
852 Ops.insert(Ops.begin(), Ops.back());
856 DOUT << "RAOut:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n";
858 if (Ops.size() == 1) {
859 // This expression tree simplified to something that isn't a tree,
861 I->replaceAllUsesWith(Ops[0].Op);
862 RemoveDeadBinaryOp(I);
864 // Now that we ordered and optimized the expressions, splat them back into
865 // the expression tree, removing any unneeded nodes.
866 RewriteExprTree(I, Ops);
871 bool Reassociate::runOnFunction(Function &F) {
872 // Recalculate the rank map for F
876 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
879 // We are done with the rank map...
881 ValueRankMap.clear();