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 Value *OptimizeAdd(std::vector<ValueEntry> &Ops);
96 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops);
97 void LinearizeExpr(BinaryOperator *I);
98 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
99 void ReassociateBB(BasicBlock *BB);
101 void RemoveDeadBinaryOp(Value *V);
105 char Reassociate::ID = 0;
106 static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions");
108 // Public interface to the Reassociate pass
109 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
111 void Reassociate::RemoveDeadBinaryOp(Value *V) {
112 Instruction *Op = dyn_cast<Instruction>(V);
113 if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty())
116 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1);
117 RemoveDeadBinaryOp(LHS);
118 RemoveDeadBinaryOp(RHS);
122 static bool isUnmovableInstruction(Instruction *I) {
123 if (I->getOpcode() == Instruction::PHI ||
124 I->getOpcode() == Instruction::Alloca ||
125 I->getOpcode() == Instruction::Load ||
126 I->getOpcode() == Instruction::Invoke ||
127 (I->getOpcode() == Instruction::Call &&
128 !isa<DbgInfoIntrinsic>(I)) ||
129 I->getOpcode() == Instruction::UDiv ||
130 I->getOpcode() == Instruction::SDiv ||
131 I->getOpcode() == Instruction::FDiv ||
132 I->getOpcode() == Instruction::URem ||
133 I->getOpcode() == Instruction::SRem ||
134 I->getOpcode() == Instruction::FRem)
139 void Reassociate::BuildRankMap(Function &F) {
142 // Assign distinct ranks to function arguments
143 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
144 ValueRankMap[&*I] = ++i;
146 ReversePostOrderTraversal<Function*> RPOT(&F);
147 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
148 E = RPOT.end(); I != E; ++I) {
150 unsigned BBRank = RankMap[BB] = ++i << 16;
152 // Walk the basic block, adding precomputed ranks for any instructions that
153 // we cannot move. This ensures that the ranks for these instructions are
154 // all different in the block.
155 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
156 if (isUnmovableInstruction(I))
157 ValueRankMap[&*I] = ++BBRank;
161 unsigned Reassociate::getRank(Value *V) {
162 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument...
164 Instruction *I = dyn_cast<Instruction>(V);
165 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0.
167 unsigned &CachedRank = ValueRankMap[I];
168 if (CachedRank) return CachedRank; // Rank already known?
170 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
171 // we can reassociate expressions for code motion! Since we do not recurse
172 // for PHI nodes, we cannot have infinite recursion here, because there
173 // cannot be loops in the value graph that do not go through PHI nodes.
174 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
175 for (unsigned i = 0, e = I->getNumOperands();
176 i != e && Rank != MaxRank; ++i)
177 Rank = std::max(Rank, getRank(I->getOperand(i)));
179 // If this is a not or neg instruction, do not count it for rank. This
180 // assures us that X and ~X will have the same rank.
181 if (!I->getType()->isInteger() ||
182 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
185 //DEBUG(errs() << "Calculated Rank[" << V->getName() << "] = "
188 return CachedRank = Rank;
191 /// isReassociableOp - Return true if V is an instruction of the specified
192 /// opcode and if it only has one use.
193 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
194 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) &&
195 cast<Instruction>(V)->getOpcode() == Opcode)
196 return cast<BinaryOperator>(V);
200 /// LowerNegateToMultiply - Replace 0-X with X*-1.
202 static Instruction *LowerNegateToMultiply(Instruction *Neg,
203 std::map<AssertingVH<>, unsigned> &ValueRankMap) {
204 Constant *Cst = Constant::getAllOnesValue(Neg->getType());
206 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
207 ValueRankMap.erase(Neg);
209 Neg->replaceAllUsesWith(Res);
210 Neg->eraseFromParent();
214 // Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'.
215 // Note that if D is also part of the expression tree that we recurse to
216 // linearize it as well. Besides that case, this does not recurse into A,B, or
218 void Reassociate::LinearizeExpr(BinaryOperator *I) {
219 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
220 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1));
221 assert(isReassociableOp(LHS, I->getOpcode()) &&
222 isReassociableOp(RHS, I->getOpcode()) &&
223 "Not an expression that needs linearization?");
225 DEBUG(errs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n');
227 // Move the RHS instruction to live immediately before I, avoiding breaking
228 // dominator properties.
231 // Move operands around to do the linearization.
232 I->setOperand(1, RHS->getOperand(0));
233 RHS->setOperand(0, LHS);
234 I->setOperand(0, RHS);
238 DEBUG(errs() << "Linearized: " << *I << '\n');
240 // If D is part of this expression tree, tail recurse.
241 if (isReassociableOp(I->getOperand(1), I->getOpcode()))
246 /// LinearizeExprTree - Given an associative binary expression tree, traverse
247 /// all of the uses putting it into canonical form. This forces a left-linear
248 /// form of the the expression (((a+b)+c)+d), and collects information about the
249 /// rank of the non-tree operands.
251 /// NOTE: These intentionally destroys the expression tree operands (turning
252 /// them into undef values) to reduce #uses of the values. This means that the
253 /// caller MUST use something like RewriteExprTree to put the values back in.
255 void Reassociate::LinearizeExprTree(BinaryOperator *I,
256 std::vector<ValueEntry> &Ops) {
257 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
258 unsigned Opcode = I->getOpcode();
260 // First step, linearize the expression if it is in ((A+B)+(C+D)) form.
261 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode);
262 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode);
264 // If this is a multiply expression tree and it contains internal negations,
265 // transform them into multiplies by -1 so they can be reassociated.
266 if (I->getOpcode() == Instruction::Mul) {
267 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) {
268 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap);
269 LHSBO = isReassociableOp(LHS, Opcode);
271 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) {
272 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap);
273 RHSBO = isReassociableOp(RHS, Opcode);
279 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As
280 // such, just remember these operands and their rank.
281 Ops.push_back(ValueEntry(getRank(LHS), LHS));
282 Ops.push_back(ValueEntry(getRank(RHS), RHS));
284 // Clear the leaves out.
285 I->setOperand(0, UndefValue::get(I->getType()));
286 I->setOperand(1, UndefValue::get(I->getType()));
290 // Turn X+(Y+Z) -> (Y+Z)+X
291 std::swap(LHSBO, RHSBO);
293 bool Success = !I->swapOperands();
294 assert(Success && "swapOperands failed");
298 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not
299 // part of the expression tree.
301 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0));
302 RHS = I->getOperand(1);
306 // Okay, now we know that the LHS is a nested expression and that the RHS is
307 // not. Perform reassociation.
308 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!");
310 // Move LHS right before I to make sure that the tree expression dominates all
312 LHSBO->moveBefore(I);
314 // Linearize the expression tree on the LHS.
315 LinearizeExprTree(LHSBO, Ops);
317 // Remember the RHS operand and its rank.
318 Ops.push_back(ValueEntry(getRank(RHS), RHS));
320 // Clear the RHS leaf out.
321 I->setOperand(1, UndefValue::get(I->getType()));
324 // RewriteExprTree - Now that the operands for this expression tree are
325 // linearized and optimized, emit them in-order. This function is written to be
327 void Reassociate::RewriteExprTree(BinaryOperator *I,
328 std::vector<ValueEntry> &Ops,
330 if (i+2 == Ops.size()) {
331 if (I->getOperand(0) != Ops[i].Op ||
332 I->getOperand(1) != Ops[i+1].Op) {
333 Value *OldLHS = I->getOperand(0);
334 DEBUG(errs() << "RA: " << *I << '\n');
335 I->setOperand(0, Ops[i].Op);
336 I->setOperand(1, Ops[i+1].Op);
337 DEBUG(errs() << "TO: " << *I << '\n');
341 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3)
342 // delete the extra, now dead, nodes.
343 RemoveDeadBinaryOp(OldLHS);
347 assert(i+2 < Ops.size() && "Ops index out of range!");
349 if (I->getOperand(1) != Ops[i].Op) {
350 DEBUG(errs() << "RA: " << *I << '\n');
351 I->setOperand(1, Ops[i].Op);
352 DEBUG(errs() << "TO: " << *I << '\n');
357 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
358 assert(LHS->getOpcode() == I->getOpcode() &&
359 "Improper expression tree!");
361 // Compactify the tree instructions together with each other to guarantee
362 // that the expression tree is dominated by all of Ops.
364 RewriteExprTree(LHS, Ops, i+1);
369 // NegateValue - Insert instructions before the instruction pointed to by BI,
370 // that computes the negative version of the value specified. The negative
371 // version of the value is returned, and BI is left pointing at the instruction
372 // that should be processed next by the reassociation pass.
374 static Value *NegateValue(Value *V, Instruction *BI) {
375 // We are trying to expose opportunity for reassociation. One of the things
376 // that we want to do to achieve this is to push a negation as deep into an
377 // expression chain as possible, to expose the add instructions. In practice,
378 // this means that we turn this:
379 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
380 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
381 // the constants. We assume that instcombine will clean up the mess later if
382 // we introduce tons of unnecessary negation instructions...
384 if (Instruction *I = dyn_cast<Instruction>(V))
385 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
386 // Push the negates through the add.
387 I->setOperand(0, NegateValue(I->getOperand(0), BI));
388 I->setOperand(1, NegateValue(I->getOperand(1), BI));
390 // We must move the add instruction here, because the neg instructions do
391 // not dominate the old add instruction in general. By moving it, we are
392 // assured that the neg instructions we just inserted dominate the
393 // instruction we are about to insert after them.
396 I->setName(I->getName()+".neg");
400 // Insert a 'neg' instruction that subtracts the value from zero to get the
403 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
406 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
407 /// X-Y into (X + -Y).
408 static bool ShouldBreakUpSubtract(Instruction *Sub) {
409 // If this is a negation, we can't split it up!
410 if (BinaryOperator::isNeg(Sub))
413 // Don't bother to break this up unless either the LHS is an associable add or
414 // subtract or if this is only used by one.
415 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
416 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
418 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
419 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
421 if (Sub->hasOneUse() &&
422 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
423 isReassociableOp(Sub->use_back(), Instruction::Sub)))
429 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
430 /// only used by an add, transform this into (X+(0-Y)) to promote better
432 static Instruction *BreakUpSubtract(Instruction *Sub,
433 std::map<AssertingVH<>, unsigned> &ValueRankMap) {
434 // Convert a subtract into an add and a neg instruction... so that sub
435 // instructions can be commuted with other add instructions...
437 // Calculate the negative value of Operand 1 of the sub instruction...
438 // and set it as the RHS of the add instruction we just made...
440 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
442 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
445 // Everyone now refers to the add instruction.
446 ValueRankMap.erase(Sub);
447 Sub->replaceAllUsesWith(New);
448 Sub->eraseFromParent();
450 DEBUG(errs() << "Negated: " << *New << '\n');
454 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
455 /// by one, change this into a multiply by a constant to assist with further
457 static Instruction *ConvertShiftToMul(Instruction *Shl,
458 std::map<AssertingVH<>, unsigned> &ValueRankMap) {
459 // If an operand of this shift is a reassociable multiply, or if the shift
460 // is used by a reassociable multiply or add, turn into a multiply.
461 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) ||
463 (isReassociableOp(Shl->use_back(), Instruction::Mul) ||
464 isReassociableOp(Shl->use_back(), Instruction::Add)))) {
465 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
466 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
469 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
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);
552 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
553 /// instruction. This optimizes based on identities. If it can be reduced to
554 /// a single Value, it is returned, otherwise the Ops list is mutated as
556 static Value *OptimizeAndOrXor(unsigned Opcode, std::vector<ValueEntry> &Ops) {
557 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
558 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
559 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
560 // First, check for X and ~X in the operand list.
561 assert(i < Ops.size());
562 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
563 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
564 unsigned FoundX = FindInOperandList(Ops, i, X);
566 if (Opcode == Instruction::And) { // ...&X&~X = 0
568 return Constant::getNullValue(X->getType());
571 if (Opcode == Instruction::Or) { // ...|X|~X = -1
573 return Constant::getAllOnesValue(X->getType());
578 // Next, check for duplicate pairs of values, which we assume are next to
579 // each other, due to our sorting criteria.
580 assert(i < Ops.size());
581 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
582 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
583 // Drop duplicate values.
584 Ops.erase(Ops.begin()+i);
588 assert(Opcode == Instruction::Xor);
591 return Constant::getNullValue(Ops[0].Op->getType());
594 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
603 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
604 /// optimizes based on identities. If it can be reduced to a single Value, it
605 /// is returned, otherwise the Ops list is mutated as necessary.
606 Value *Reassociate::OptimizeAdd(std::vector<ValueEntry> &Ops) {
607 // Scan the operand lists looking for X and -X pairs. If we find any, we
608 // can simplify the expression. X+-X == 0.
609 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
610 assert(i < Ops.size());
611 // Check for X and -X in the operand list.
612 if (!BinaryOperator::isNeg(Ops[i].Op))
615 Value *X = BinaryOperator::getNegArgument(Ops[i].Op);
616 unsigned FoundX = FindInOperandList(Ops, i, X);
620 // Remove X and -X from the operand list.
621 if (Ops.size() == 2) {
623 return Constant::getNullValue(X->getType());
626 Ops.erase(Ops.begin()+i);
630 --i; // Need to back up an extra one.
631 Ops.erase(Ops.begin()+FoundX);
633 --i; // Revisit element.
634 e -= 2; // Removed two elements.
639 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
640 std::vector<ValueEntry> &Ops) {
641 // Now that we have the linearized expression tree, try to optimize it.
642 // Start by folding any constants that we found.
643 bool IterateOptimization = false;
644 if (Ops.size() == 1) return Ops[0].Op;
646 unsigned Opcode = I->getOpcode();
648 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
649 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
651 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
652 return OptimizeExpression(I, Ops);
655 // Check for destructive annihilation due to a constant being used.
656 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
659 case Instruction::And:
660 if (CstVal->isZero()) { // ... & 0 -> 0
664 if (CstVal->isAllOnesValue()) // ... & -1 -> ...
667 case Instruction::Mul:
668 if (CstVal->isZero()) { // ... * 0 -> 0
673 if (cast<ConstantInt>(CstVal)->isOne())
674 Ops.pop_back(); // ... * 1 -> ...
676 case Instruction::Or:
677 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1
682 case Instruction::Add:
683 case Instruction::Xor:
684 if (CstVal->isZero()) // ... [|^+] 0 -> ...
688 if (Ops.size() == 1) return Ops[0].Op;
690 // Handle destructive annihilation due to identities between elements in the
691 // argument list here.
694 case Instruction::And:
695 case Instruction::Or:
696 case Instruction::Xor: {
697 unsigned NumOps = Ops.size();
698 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
700 IterateOptimization |= Ops.size() != NumOps;
704 case Instruction::Add: {
705 unsigned NumOps = Ops.size();
706 if (Value *Result = OptimizeAdd(Ops))
708 IterateOptimization |= Ops.size() != NumOps;
711 // Scan the operand list, checking to see if there are any common factors
712 // between operands. Consider something like A*A+A*B*C+D. We would like to
713 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
714 // To efficiently find this, we count the number of times a factor occurs
715 // for any ADD operands that are MULs.
716 DenseMap<Value*, unsigned> FactorOccurrences;
718 Value *MaxOccVal = 0;
719 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
720 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op);
721 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty())
724 // Compute all of the factors of this added value.
725 SmallVector<Value*, 8> Factors;
726 FindSingleUseMultiplyFactors(BOp, Factors);
727 assert(Factors.size() > 1 && "Bad linearize!");
729 // Add one to FactorOccurrences for each unique factor in this op.
730 if (Factors.size() == 2) {
731 unsigned Occ = ++FactorOccurrences[Factors[0]];
732 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; }
733 if (Factors[0] != Factors[1]) { // Don't double count A*A.
734 Occ = ++FactorOccurrences[Factors[1]];
735 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; }
738 SmallPtrSet<Value*, 4> Duplicates;
739 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
740 if (!Duplicates.insert(Factors[i])) continue;
742 unsigned Occ = ++FactorOccurrences[Factors[i]];
743 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; }
748 // If any factor occurred more than one time, we can pull it out.
750 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n");
752 // Create a new instruction that uses the MaxOccVal twice. If we don't do
753 // this, we could otherwise run into situations where removing a factor
754 // from an expression will drop a use of maxocc, and this can cause
755 // RemoveFactorFromExpression on successive values to behave differently.
756 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
757 SmallVector<Value*, 4> NewMulOps;
758 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
759 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
760 NewMulOps.push_back(V);
761 Ops.erase(Ops.begin()+i);
766 // No need for extra uses anymore.
769 unsigned NumAddedValues = NewMulOps.size();
770 Value *V = EmitAddTreeOfValues(I, NewMulOps);
771 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
773 // Now that we have inserted V and its sole use, optimize it. This allows
774 // us to handle cases that require multiple factoring steps, such as this:
775 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
776 if (NumAddedValues > 1)
777 ReassociateExpression(cast<BinaryOperator>(V));
784 // Add the new value to the list of things being added.
785 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
787 // Rewrite the tree so that there is now a use of V.
788 RewriteExprTree(I, Ops);
789 return OptimizeExpression(I, Ops);
792 //case Instruction::Mul:
795 if (IterateOptimization)
796 return OptimizeExpression(I, Ops);
801 /// ReassociateBB - Inspect all of the instructions in this basic block,
802 /// reassociating them as we go.
803 void Reassociate::ReassociateBB(BasicBlock *BB) {
804 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) {
805 Instruction *BI = BBI++;
806 if (BI->getOpcode() == Instruction::Shl &&
807 isa<ConstantInt>(BI->getOperand(1)))
808 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) {
813 // Reject cases where it is pointless to do this.
814 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() ||
815 isa<VectorType>(BI->getType()))
816 continue; // Floating point ops are not associative.
818 // If this is a subtract instruction which is not already in negate form,
819 // see if we can convert it to X+-Y.
820 if (BI->getOpcode() == Instruction::Sub) {
821 if (ShouldBreakUpSubtract(BI)) {
822 BI = BreakUpSubtract(BI, ValueRankMap);
824 } else if (BinaryOperator::isNeg(BI)) {
825 // Otherwise, this is a negation. See if the operand is a multiply tree
826 // and if this is not an inner node of a multiply tree.
827 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) &&
829 !isReassociableOp(BI->use_back(), Instruction::Mul))) {
830 BI = LowerNegateToMultiply(BI, ValueRankMap);
836 // If this instruction is a commutative binary operator, process it.
837 if (!BI->isAssociative()) continue;
838 BinaryOperator *I = cast<BinaryOperator>(BI);
840 // If this is an interior node of a reassociable tree, ignore it until we
841 // get to the root of the tree, to avoid N^2 analysis.
842 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode()))
845 // If this is an add tree that is used by a sub instruction, ignore it
846 // until we process the subtract.
847 if (I->hasOneUse() && I->getOpcode() == Instruction::Add &&
848 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub)
851 ReassociateExpression(I);
855 void Reassociate::ReassociateExpression(BinaryOperator *I) {
857 // First, walk the expression tree, linearizing the tree, collecting
858 std::vector<ValueEntry> Ops;
859 LinearizeExprTree(I, Ops);
861 DEBUG(errs() << "RAIn:\t"; PrintOps(I, Ops); errs() << "\n");
863 // Now that we have linearized the tree to a list and have gathered all of
864 // the operands and their ranks, sort the operands by their rank. Use a
865 // stable_sort so that values with equal ranks will have their relative
866 // positions maintained (and so the compiler is deterministic). Note that
867 // this sorts so that the highest ranking values end up at the beginning of
869 std::stable_sort(Ops.begin(), Ops.end());
871 // OptimizeExpression - Now that we have the expression tree in a convenient
872 // sorted form, optimize it globally if possible.
873 if (Value *V = OptimizeExpression(I, Ops)) {
874 // This expression tree simplified to something that isn't a tree,
876 DEBUG(errs() << "Reassoc to scalar: " << *V << "\n");
877 I->replaceAllUsesWith(V);
878 RemoveDeadBinaryOp(I);
882 // We want to sink immediates as deeply as possible except in the case where
883 // this is a multiply tree used only by an add, and the immediate is a -1.
884 // In this case we reassociate to put the negation on the outside so that we
885 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
886 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
887 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
888 isa<ConstantInt>(Ops.back().Op) &&
889 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
890 Ops.insert(Ops.begin(), Ops.back());
894 DEBUG(errs() << "RAOut:\t"; PrintOps(I, Ops); errs() << "\n");
896 if (Ops.size() == 1) {
897 // This expression tree simplified to something that isn't a tree,
899 I->replaceAllUsesWith(Ops[0].Op);
900 RemoveDeadBinaryOp(I);
902 // Now that we ordered and optimized the expressions, splat them back into
903 // the expression tree, removing any unneeded nodes.
904 RewriteExprTree(I, Ops);
909 bool Reassociate::runOnFunction(Function &F) {
910 // Recalculate the rank map for F
914 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
917 // We are done with the rank map...
919 ValueRankMap.clear();