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/Compiler.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/ADT/PostOrderIterator.h"
36 #include "llvm/ADT/Statistic.h"
41 STATISTIC(NumLinear , "Number of insts linearized");
42 STATISTIC(NumChanged, "Number of insts reassociated");
43 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
44 STATISTIC(NumFactor , "Number of multiplies factored");
47 struct VISIBILITY_HIDDEN ValueEntry {
50 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {}
52 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) {
53 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start.
58 /// PrintOps - Print out the expression identified in the Ops list.
60 static void PrintOps(Instruction *I, const std::vector<ValueEntry> &Ops) {
61 Module *M = I->getParent()->getParent()->getParent();
62 cerr << Instruction::getOpcodeName(I->getOpcode()) << " "
63 << *Ops[0].Op->getType();
64 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
65 WriteAsOperand(*cerr.stream() << " ", Ops[i].Op, false, M);
66 cerr << "," << Ops[i].Rank;
72 class VISIBILITY_HIDDEN Reassociate : public FunctionPass {
73 std::map<BasicBlock*, unsigned> RankMap;
74 std::map<Value*, unsigned> ValueRankMap;
77 static char ID; // Pass identification, replacement for typeid
78 Reassociate() : FunctionPass(&ID) {}
80 bool runOnFunction(Function &F);
82 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
86 void BuildRankMap(Function &F);
87 unsigned getRank(Value *V);
88 void ReassociateExpression(BinaryOperator *I);
89 void RewriteExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops,
91 Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops);
92 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops);
93 void LinearizeExpr(BinaryOperator *I);
94 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
95 void ReassociateBB(BasicBlock *BB);
97 void RemoveDeadBinaryOp(Value *V);
101 char Reassociate::ID = 0;
102 static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions");
104 // Public interface to the Reassociate pass
105 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
107 void Reassociate::RemoveDeadBinaryOp(Value *V) {
108 Instruction *Op = dyn_cast<Instruction>(V);
109 if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty())
112 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1);
113 RemoveDeadBinaryOp(LHS);
114 RemoveDeadBinaryOp(RHS);
118 static bool isUnmovableInstruction(Instruction *I) {
119 if (I->getOpcode() == Instruction::PHI ||
120 I->getOpcode() == Instruction::Alloca ||
121 I->getOpcode() == Instruction::Load ||
122 I->getOpcode() == Instruction::Malloc ||
123 I->getOpcode() == Instruction::Invoke ||
124 (I->getOpcode() == Instruction::Call &&
125 !isa<DbgInfoIntrinsic>(I)) ||
126 I->getOpcode() == Instruction::UDiv ||
127 I->getOpcode() == Instruction::SDiv ||
128 I->getOpcode() == Instruction::FDiv ||
129 I->getOpcode() == Instruction::URem ||
130 I->getOpcode() == Instruction::SRem ||
131 I->getOpcode() == Instruction::FRem)
136 void Reassociate::BuildRankMap(Function &F) {
139 // Assign distinct ranks to function arguments
140 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
141 ValueRankMap[I] = ++i;
143 ReversePostOrderTraversal<Function*> RPOT(&F);
144 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
145 E = RPOT.end(); I != E; ++I) {
147 unsigned BBRank = RankMap[BB] = ++i << 16;
149 // Walk the basic block, adding precomputed ranks for any instructions that
150 // we cannot move. This ensures that the ranks for these instructions are
151 // all different in the block.
152 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
153 if (isUnmovableInstruction(I))
154 ValueRankMap[I] = ++BBRank;
158 unsigned Reassociate::getRank(Value *V) {
159 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument...
161 Instruction *I = dyn_cast<Instruction>(V);
162 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0.
164 unsigned &CachedRank = ValueRankMap[I];
165 if (CachedRank) return CachedRank; // Rank already known?
167 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
168 // we can reassociate expressions for code motion! Since we do not recurse
169 // for PHI nodes, we cannot have infinite recursion here, because there
170 // cannot be loops in the value graph that do not go through PHI nodes.
171 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
172 for (unsigned i = 0, e = I->getNumOperands();
173 i != e && Rank != MaxRank; ++i)
174 Rank = std::max(Rank, getRank(I->getOperand(i)));
176 // If this is a not or neg instruction, do not count it for rank. This
177 // assures us that X and ~X will have the same rank.
178 if (!I->getType()->isInteger() ||
179 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
182 //DOUT << "Calculated Rank[" << V->getName() << "] = "
185 return CachedRank = Rank;
188 /// isReassociableOp - Return true if V is an instruction of the specified
189 /// opcode and if it only has one use.
190 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
191 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) &&
192 cast<Instruction>(V)->getOpcode() == Opcode)
193 return cast<BinaryOperator>(V);
197 /// LowerNegateToMultiply - Replace 0-X with X*-1.
199 static Instruction *LowerNegateToMultiply(Instruction *Neg) {
200 Constant *Cst = ConstantInt::getAllOnesValue(Neg->getType());
202 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
204 Neg->replaceAllUsesWith(Res);
205 Neg->eraseFromParent();
209 // Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'.
210 // Note that if D is also part of the expression tree that we recurse to
211 // linearize it as well. Besides that case, this does not recurse into A,B, or
213 void Reassociate::LinearizeExpr(BinaryOperator *I) {
214 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
215 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1));
216 assert(isReassociableOp(LHS, I->getOpcode()) &&
217 isReassociableOp(RHS, I->getOpcode()) &&
218 "Not an expression that needs linearization?");
220 DOUT << "Linear" << *LHS << *RHS << *I;
222 // Move the RHS instruction to live immediately before I, avoiding breaking
223 // dominator properties.
226 // Move operands around to do the linearization.
227 I->setOperand(1, RHS->getOperand(0));
228 RHS->setOperand(0, LHS);
229 I->setOperand(0, RHS);
233 DOUT << "Linearized: " << *I;
235 // If D is part of this expression tree, tail recurse.
236 if (isReassociableOp(I->getOperand(1), I->getOpcode()))
241 /// LinearizeExprTree - Given an associative binary expression tree, traverse
242 /// all of the uses putting it into canonical form. This forces a left-linear
243 /// form of the the expression (((a+b)+c)+d), and collects information about the
244 /// rank of the non-tree operands.
246 /// NOTE: These intentionally destroys the expression tree operands (turning
247 /// them into undef values) to reduce #uses of the values. This means that the
248 /// caller MUST use something like RewriteExprTree to put the values back in.
250 void Reassociate::LinearizeExprTree(BinaryOperator *I,
251 std::vector<ValueEntry> &Ops) {
252 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
253 unsigned Opcode = I->getOpcode();
255 // First step, linearize the expression if it is in ((A+B)+(C+D)) form.
256 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode);
257 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode);
259 // If this is a multiply expression tree and it contains internal negations,
260 // transform them into multiplies by -1 so they can be reassociated.
261 if (I->getOpcode() == Instruction::Mul) {
262 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) {
263 LHS = LowerNegateToMultiply(cast<Instruction>(LHS));
264 LHSBO = isReassociableOp(LHS, Opcode);
266 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) {
267 RHS = LowerNegateToMultiply(cast<Instruction>(RHS));
268 RHSBO = isReassociableOp(RHS, Opcode);
274 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As
275 // such, just remember these operands and their rank.
276 Ops.push_back(ValueEntry(getRank(LHS), LHS));
277 Ops.push_back(ValueEntry(getRank(RHS), RHS));
279 // Clear the leaves out.
280 I->setOperand(0, UndefValue::get(I->getType()));
281 I->setOperand(1, UndefValue::get(I->getType()));
284 // Turn X+(Y+Z) -> (Y+Z)+X
285 std::swap(LHSBO, RHSBO);
287 bool Success = !I->swapOperands();
288 assert(Success && "swapOperands failed");
293 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not
294 // part of the expression tree.
296 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0));
297 RHS = I->getOperand(1);
301 // Okay, now we know that the LHS is a nested expression and that the RHS is
302 // not. Perform reassociation.
303 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!");
305 // Move LHS right before I to make sure that the tree expression dominates all
307 LHSBO->moveBefore(I);
309 // Linearize the expression tree on the LHS.
310 LinearizeExprTree(LHSBO, Ops);
312 // Remember the RHS operand and its rank.
313 Ops.push_back(ValueEntry(getRank(RHS), RHS));
315 // Clear the RHS leaf out.
316 I->setOperand(1, UndefValue::get(I->getType()));
319 // RewriteExprTree - Now that the operands for this expression tree are
320 // linearized and optimized, emit them in-order. This function is written to be
322 void Reassociate::RewriteExprTree(BinaryOperator *I,
323 std::vector<ValueEntry> &Ops,
325 if (i+2 == Ops.size()) {
326 if (I->getOperand(0) != Ops[i].Op ||
327 I->getOperand(1) != Ops[i+1].Op) {
328 Value *OldLHS = I->getOperand(0);
329 DOUT << "RA: " << *I;
330 I->setOperand(0, Ops[i].Op);
331 I->setOperand(1, Ops[i+1].Op);
332 DOUT << "TO: " << *I;
336 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3)
337 // delete the extra, now dead, nodes.
338 RemoveDeadBinaryOp(OldLHS);
342 assert(i+2 < Ops.size() && "Ops index out of range!");
344 if (I->getOperand(1) != Ops[i].Op) {
345 DOUT << "RA: " << *I;
346 I->setOperand(1, Ops[i].Op);
347 DOUT << "TO: " << *I;
352 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
353 assert(LHS->getOpcode() == I->getOpcode() &&
354 "Improper expression tree!");
356 // Compactify the tree instructions together with each other to guarantee
357 // that the expression tree is dominated by all of Ops.
359 RewriteExprTree(LHS, Ops, i+1);
364 // NegateValue - Insert instructions before the instruction pointed to by BI,
365 // that computes the negative version of the value specified. The negative
366 // version of the value is returned, and BI is left pointing at the instruction
367 // that should be processed next by the reassociation pass.
369 static Value *NegateValue(Value *V, Instruction *BI) {
370 // We are trying to expose opportunity for reassociation. One of the things
371 // that we want to do to achieve this is to push a negation as deep into an
372 // expression chain as possible, to expose the add instructions. In practice,
373 // this means that we turn this:
374 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
375 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
376 // the constants. We assume that instcombine will clean up the mess later if
377 // we introduce tons of unnecessary negation instructions...
379 if (Instruction *I = dyn_cast<Instruction>(V))
380 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
381 // Push the negates through the add.
382 I->setOperand(0, NegateValue(I->getOperand(0), BI));
383 I->setOperand(1, NegateValue(I->getOperand(1), BI));
385 // We must move the add instruction here, because the neg instructions do
386 // not dominate the old add instruction in general. By moving it, we are
387 // assured that the neg instructions we just inserted dominate the
388 // instruction we are about to insert after them.
391 I->setName(I->getName()+".neg");
395 // Insert a 'neg' instruction that subtracts the value from zero to get the
398 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
401 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
402 /// X-Y into (X + -Y).
403 static bool ShouldBreakUpSubtract(Instruction *Sub) {
404 // If this is a negation, we can't split it up!
405 if (BinaryOperator::isNeg(Sub))
408 // Don't bother to break this up unless either the LHS is an associable add or
409 // subtract or if this is only used by one.
410 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
411 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
413 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
414 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
416 if (Sub->hasOneUse() &&
417 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
418 isReassociableOp(Sub->use_back(), Instruction::Sub)))
424 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
425 /// only used by an add, transform this into (X+(0-Y)) to promote better
427 static Instruction *BreakUpSubtract(Instruction *Sub) {
428 // Convert a subtract into an add and a neg instruction... so that sub
429 // instructions can be commuted with other add instructions...
431 // Calculate the negative value of Operand 1 of the sub instruction...
432 // and set it as the RHS of the add instruction we just made...
434 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
436 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
439 // Everyone now refers to the add instruction.
440 Sub->replaceAllUsesWith(New);
441 Sub->eraseFromParent();
443 DOUT << "Negated: " << *New;
447 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
448 /// by one, change this into a multiply by a constant to assist with further
450 static Instruction *ConvertShiftToMul(Instruction *Shl) {
451 // If an operand of this shift is a reassociable multiply, or if the shift
452 // is used by a reassociable multiply or add, turn into a multiply.
453 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) ||
455 (isReassociableOp(Shl->use_back(), Instruction::Mul) ||
456 isReassociableOp(Shl->use_back(), Instruction::Add)))) {
457 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
458 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
460 Instruction *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst,
463 Shl->replaceAllUsesWith(Mul);
464 Shl->eraseFromParent();
470 // Scan backwards and forwards among values with the same rank as element i to
471 // see if X exists. If X does not exist, return i.
472 static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i,
474 unsigned XRank = Ops[i].Rank;
475 unsigned e = Ops.size();
476 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
480 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
486 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
487 /// and returning the result. Insert the tree before I.
488 static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) {
489 if (Ops.size() == 1) return Ops.back();
491 Value *V1 = Ops.back();
493 Value *V2 = EmitAddTreeOfValues(I, Ops);
494 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
497 /// RemoveFactorFromExpression - If V is an expression tree that is a
498 /// multiplication sequence, and if this sequence contains a multiply by Factor,
499 /// remove Factor from the tree and return the new tree.
500 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
501 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
504 std::vector<ValueEntry> Factors;
505 LinearizeExprTree(BO, Factors);
507 bool FoundFactor = false;
508 for (unsigned i = 0, e = Factors.size(); i != e; ++i)
509 if (Factors[i].Op == Factor) {
511 Factors.erase(Factors.begin()+i);
515 // Make sure to restore the operands to the expression tree.
516 RewriteExprTree(BO, Factors);
520 if (Factors.size() == 1) return Factors[0].Op;
522 RewriteExprTree(BO, Factors);
526 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
527 /// add its operands as factors, otherwise add V to the list of factors.
528 static void FindSingleUseMultiplyFactors(Value *V,
529 std::vector<Value*> &Factors) {
531 if ((!V->hasOneUse() && !V->use_empty()) ||
532 !(BO = dyn_cast<BinaryOperator>(V)) ||
533 BO->getOpcode() != Instruction::Mul) {
534 Factors.push_back(V);
538 // Otherwise, add the LHS and RHS to the list of factors.
539 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
540 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
545 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
546 std::vector<ValueEntry> &Ops) {
547 // Now that we have the linearized expression tree, try to optimize it.
548 // Start by folding any constants that we found.
549 bool IterateOptimization = false;
550 if (Ops.size() == 1) return Ops[0].Op;
552 unsigned Opcode = I->getOpcode();
554 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
555 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
557 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
558 return OptimizeExpression(I, Ops);
561 // Check for destructive annihilation due to a constant being used.
562 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
565 case Instruction::And:
566 if (CstVal->isZero()) { // ... & 0 -> 0
569 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ...
573 case Instruction::Mul:
574 if (CstVal->isZero()) { // ... * 0 -> 0
577 } else if (cast<ConstantInt>(CstVal)->isOne()) {
578 Ops.pop_back(); // ... * 1 -> ...
581 case Instruction::Or:
582 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1
587 case Instruction::Add:
588 case Instruction::Xor:
589 if (CstVal->isZero()) // ... [|^+] 0 -> ...
593 if (Ops.size() == 1) return Ops[0].Op;
595 // Handle destructive annihilation do to identities between elements in the
596 // argument list here.
599 case Instruction::And:
600 case Instruction::Or:
601 case Instruction::Xor:
602 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
603 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
604 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
605 // First, check for X and ~X in the operand list.
606 assert(i < Ops.size());
607 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
608 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
609 unsigned FoundX = FindInOperandList(Ops, i, X);
611 if (Opcode == Instruction::And) { // ...&X&~X = 0
613 return Constant::getNullValue(X->getType());
614 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1
616 return ConstantInt::getAllOnesValue(X->getType());
621 // Next, check for duplicate pairs of values, which we assume are next to
622 // each other, due to our sorting criteria.
623 assert(i < Ops.size());
624 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
625 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
626 // Drop duplicate values.
627 Ops.erase(Ops.begin()+i);
629 IterateOptimization = true;
632 assert(Opcode == Instruction::Xor);
635 return Constant::getNullValue(Ops[0].Op->getType());
638 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
640 IterateOptimization = true;
647 case Instruction::Add:
648 // Scan the operand lists looking for X and -X pairs. If we find any, we
649 // can simplify the expression. X+-X == 0.
650 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
651 assert(i < Ops.size());
652 // Check for X and -X in the operand list.
653 if (BinaryOperator::isNeg(Ops[i].Op)) {
654 Value *X = BinaryOperator::getNegArgument(Ops[i].Op);
655 unsigned FoundX = FindInOperandList(Ops, i, X);
657 // Remove X and -X from the operand list.
658 if (Ops.size() == 2) {
660 return Constant::getNullValue(X->getType());
662 Ops.erase(Ops.begin()+i);
666 --i; // Need to back up an extra one.
667 Ops.erase(Ops.begin()+FoundX);
668 IterateOptimization = true;
670 --i; // Revisit element.
671 e -= 2; // Removed two elements.
678 // Scan the operand list, checking to see if there are any common factors
679 // between operands. Consider something like A*A+A*B*C+D. We would like to
680 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
681 // To efficiently find this, we count the number of times a factor occurs
682 // for any ADD operands that are MULs.
683 std::map<Value*, unsigned> FactorOccurrences;
685 Value *MaxOccVal = 0;
686 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
687 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) {
688 if (BOp->getOpcode() == Instruction::Mul && BOp->use_empty()) {
689 // Compute all of the factors of this added value.
690 std::vector<Value*> Factors;
691 FindSingleUseMultiplyFactors(BOp, Factors);
692 assert(Factors.size() > 1 && "Bad linearize!");
694 // Add one to FactorOccurrences for each unique factor in this op.
695 if (Factors.size() == 2) {
696 unsigned Occ = ++FactorOccurrences[Factors[0]];
697 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; }
698 if (Factors[0] != Factors[1]) { // Don't double count A*A.
699 Occ = ++FactorOccurrences[Factors[1]];
700 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; }
703 std::set<Value*> Duplicates;
704 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
705 if (Duplicates.insert(Factors[i]).second) {
706 unsigned Occ = ++FactorOccurrences[Factors[i]];
707 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; }
715 // If any factor occurred more than one time, we can pull it out.
717 DOUT << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n";
719 // Create a new instruction that uses the MaxOccVal twice. If we don't do
720 // this, we could otherwise run into situations where removing a factor
721 // from an expression will drop a use of maxocc, and this can cause
722 // RemoveFactorFromExpression on successive values to behave differently.
723 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
724 std::vector<Value*> NewMulOps;
725 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
726 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
727 NewMulOps.push_back(V);
728 Ops.erase(Ops.begin()+i);
733 // No need for extra uses anymore.
736 unsigned NumAddedValues = NewMulOps.size();
737 Value *V = EmitAddTreeOfValues(I, NewMulOps);
738 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
740 // Now that we have inserted V and its sole use, optimize it. This allows
741 // us to handle cases that require multiple factoring steps, such as this:
742 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
743 if (NumAddedValues > 1)
744 ReassociateExpression(cast<BinaryOperator>(V));
751 // Add the new value to the list of things being added.
752 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
754 // Rewrite the tree so that there is now a use of V.
755 RewriteExprTree(I, Ops);
756 return OptimizeExpression(I, Ops);
759 //case Instruction::Mul:
762 if (IterateOptimization)
763 return OptimizeExpression(I, Ops);
768 /// ReassociateBB - Inspect all of the instructions in this basic block,
769 /// reassociating them as we go.
770 void Reassociate::ReassociateBB(BasicBlock *BB) {
771 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) {
772 Instruction *BI = BBI++;
773 if (BI->getOpcode() == Instruction::Shl &&
774 isa<ConstantInt>(BI->getOperand(1)))
775 if (Instruction *NI = ConvertShiftToMul(BI)) {
780 // Reject cases where it is pointless to do this.
781 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() ||
782 isa<VectorType>(BI->getType()))
783 continue; // Floating point ops are not associative.
785 // If this is a subtract instruction which is not already in negate form,
786 // see if we can convert it to X+-Y.
787 if (BI->getOpcode() == Instruction::Sub) {
788 if (ShouldBreakUpSubtract(BI)) {
789 BI = BreakUpSubtract(BI);
791 } else if (BinaryOperator::isNeg(BI)) {
792 // Otherwise, this is a negation. See if the operand is a multiply tree
793 // and if this is not an inner node of a multiply tree.
794 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) &&
796 !isReassociableOp(BI->use_back(), Instruction::Mul))) {
797 BI = LowerNegateToMultiply(BI);
803 // If this instruction is a commutative binary operator, process it.
804 if (!BI->isAssociative()) continue;
805 BinaryOperator *I = cast<BinaryOperator>(BI);
807 // If this is an interior node of a reassociable tree, ignore it until we
808 // get to the root of the tree, to avoid N^2 analysis.
809 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode()))
812 // If this is an add tree that is used by a sub instruction, ignore it
813 // until we process the subtract.
814 if (I->hasOneUse() && I->getOpcode() == Instruction::Add &&
815 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub)
818 ReassociateExpression(I);
822 void Reassociate::ReassociateExpression(BinaryOperator *I) {
824 // First, walk the expression tree, linearizing the tree, collecting
825 std::vector<ValueEntry> Ops;
826 LinearizeExprTree(I, Ops);
828 DOUT << "RAIn:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n";
830 // Now that we have linearized the tree to a list and have gathered all of
831 // the operands and their ranks, sort the operands by their rank. Use a
832 // stable_sort so that values with equal ranks will have their relative
833 // positions maintained (and so the compiler is deterministic). Note that
834 // this sorts so that the highest ranking values end up at the beginning of
836 std::stable_sort(Ops.begin(), Ops.end());
838 // OptimizeExpression - Now that we have the expression tree in a convenient
839 // sorted form, optimize it globally if possible.
840 if (Value *V = OptimizeExpression(I, Ops)) {
841 // This expression tree simplified to something that isn't a tree,
843 DOUT << "Reassoc to scalar: " << *V << "\n";
844 I->replaceAllUsesWith(V);
845 RemoveDeadBinaryOp(I);
849 // We want to sink immediates as deeply as possible except in the case where
850 // this is a multiply tree used only by an add, and the immediate is a -1.
851 // In this case we reassociate to put the negation on the outside so that we
852 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
853 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
854 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
855 isa<ConstantInt>(Ops.back().Op) &&
856 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
857 Ops.insert(Ops.begin(), Ops.back());
861 DOUT << "RAOut:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n";
863 if (Ops.size() == 1) {
864 // This expression tree simplified to something that isn't a tree,
866 I->replaceAllUsesWith(Ops[0].Op);
867 RemoveDeadBinaryOp(I);
869 // Now that we ordered and optimized the expressions, splat them back into
870 // the expression tree, removing any unneeded nodes.
871 RewriteExprTree(I, Ops);
876 bool Reassociate::runOnFunction(Function &F) {
877 // Recalculate the rank map for F
881 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
884 // We are done with the rank map...
886 ValueRankMap.clear();