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 SmallVectorImpl<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 Value *ReassociateExpression(BinaryOperator *I);
92 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops,
94 Value *OptimizeExpression(BinaryOperator *I,
95 SmallVectorImpl<ValueEntry> &Ops);
96 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
97 void LinearizeExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
98 void LinearizeExpr(BinaryOperator *I);
99 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
100 void ReassociateBB(BasicBlock *BB);
102 void RemoveDeadBinaryOp(Value *V);
106 char Reassociate::ID = 0;
107 static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions");
109 // Public interface to the Reassociate pass
110 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
112 void Reassociate::RemoveDeadBinaryOp(Value *V) {
113 Instruction *Op = dyn_cast<Instruction>(V);
114 if (!Op || !isa<BinaryOperator>(Op) || !Op->use_empty())
117 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1);
119 ValueRankMap.erase(Op);
120 Op->eraseFromParent();
121 RemoveDeadBinaryOp(LHS);
122 RemoveDeadBinaryOp(RHS);
126 static bool isUnmovableInstruction(Instruction *I) {
127 if (I->getOpcode() == Instruction::PHI ||
128 I->getOpcode() == Instruction::Alloca ||
129 I->getOpcode() == Instruction::Load ||
130 I->getOpcode() == Instruction::Invoke ||
131 (I->getOpcode() == Instruction::Call &&
132 !isa<DbgInfoIntrinsic>(I)) ||
133 I->getOpcode() == Instruction::UDiv ||
134 I->getOpcode() == Instruction::SDiv ||
135 I->getOpcode() == Instruction::FDiv ||
136 I->getOpcode() == Instruction::URem ||
137 I->getOpcode() == Instruction::SRem ||
138 I->getOpcode() == Instruction::FRem)
143 void Reassociate::BuildRankMap(Function &F) {
146 // Assign distinct ranks to function arguments
147 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
148 ValueRankMap[&*I] = ++i;
150 ReversePostOrderTraversal<Function*> RPOT(&F);
151 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
152 E = RPOT.end(); I != E; ++I) {
154 unsigned BBRank = RankMap[BB] = ++i << 16;
156 // Walk the basic block, adding precomputed ranks for any instructions that
157 // we cannot move. This ensures that the ranks for these instructions are
158 // all different in the block.
159 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
160 if (isUnmovableInstruction(I))
161 ValueRankMap[&*I] = ++BBRank;
165 unsigned Reassociate::getRank(Value *V) {
166 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument...
168 Instruction *I = dyn_cast<Instruction>(V);
169 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0.
171 unsigned &CachedRank = ValueRankMap[I];
172 if (CachedRank) return CachedRank; // Rank already known?
174 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
175 // we can reassociate expressions for code motion! Since we do not recurse
176 // for PHI nodes, we cannot have infinite recursion here, because there
177 // cannot be loops in the value graph that do not go through PHI nodes.
178 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
179 for (unsigned i = 0, e = I->getNumOperands();
180 i != e && Rank != MaxRank; ++i)
181 Rank = std::max(Rank, getRank(I->getOperand(i)));
183 // If this is a not or neg instruction, do not count it for rank. This
184 // assures us that X and ~X will have the same rank.
185 if (!I->getType()->isInteger() ||
186 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
189 //DEBUG(errs() << "Calculated Rank[" << V->getName() << "] = "
192 return CachedRank = Rank;
195 /// isReassociableOp - Return true if V is an instruction of the specified
196 /// opcode and if it only has one use.
197 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
198 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) &&
199 cast<Instruction>(V)->getOpcode() == Opcode)
200 return cast<BinaryOperator>(V);
204 /// LowerNegateToMultiply - Replace 0-X with X*-1.
206 static Instruction *LowerNegateToMultiply(Instruction *Neg,
207 std::map<AssertingVH<>, unsigned> &ValueRankMap) {
208 Constant *Cst = Constant::getAllOnesValue(Neg->getType());
210 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
211 ValueRankMap.erase(Neg);
213 Neg->replaceAllUsesWith(Res);
214 Neg->eraseFromParent();
218 // Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'.
219 // Note that if D is also part of the expression tree that we recurse to
220 // linearize it as well. Besides that case, this does not recurse into A,B, or
222 void Reassociate::LinearizeExpr(BinaryOperator *I) {
223 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
224 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1));
225 assert(isReassociableOp(LHS, I->getOpcode()) &&
226 isReassociableOp(RHS, I->getOpcode()) &&
227 "Not an expression that needs linearization?");
229 DEBUG(errs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n');
231 // Move the RHS instruction to live immediately before I, avoiding breaking
232 // dominator properties.
235 // Move operands around to do the linearization.
236 I->setOperand(1, RHS->getOperand(0));
237 RHS->setOperand(0, LHS);
238 I->setOperand(0, RHS);
242 DEBUG(errs() << "Linearized: " << *I << '\n');
244 // If D is part of this expression tree, tail recurse.
245 if (isReassociableOp(I->getOperand(1), I->getOpcode()))
250 /// LinearizeExprTree - Given an associative binary expression tree, traverse
251 /// all of the uses putting it into canonical form. This forces a left-linear
252 /// form of the the expression (((a+b)+c)+d), and collects information about the
253 /// rank of the non-tree operands.
255 /// NOTE: These intentionally destroys the expression tree operands (turning
256 /// them into undef values) to reduce #uses of the values. This means that the
257 /// caller MUST use something like RewriteExprTree to put the values back in.
259 void Reassociate::LinearizeExprTree(BinaryOperator *I,
260 SmallVectorImpl<ValueEntry> &Ops) {
261 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
262 unsigned Opcode = I->getOpcode();
264 // First step, linearize the expression if it is in ((A+B)+(C+D)) form.
265 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode);
266 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode);
268 // If this is a multiply expression tree and it contains internal negations,
269 // transform them into multiplies by -1 so they can be reassociated.
270 if (I->getOpcode() == Instruction::Mul) {
271 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) {
272 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap);
273 LHSBO = isReassociableOp(LHS, Opcode);
275 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) {
276 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap);
277 RHSBO = isReassociableOp(RHS, Opcode);
283 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As
284 // such, just remember these operands and their rank.
285 Ops.push_back(ValueEntry(getRank(LHS), LHS));
286 Ops.push_back(ValueEntry(getRank(RHS), RHS));
288 // Clear the leaves out.
289 I->setOperand(0, UndefValue::get(I->getType()));
290 I->setOperand(1, UndefValue::get(I->getType()));
294 // Turn X+(Y+Z) -> (Y+Z)+X
295 std::swap(LHSBO, RHSBO);
297 bool Success = !I->swapOperands();
298 assert(Success && "swapOperands failed");
302 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not
303 // part of the expression tree.
305 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0));
306 RHS = I->getOperand(1);
310 // Okay, now we know that the LHS is a nested expression and that the RHS is
311 // not. Perform reassociation.
312 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!");
314 // Move LHS right before I to make sure that the tree expression dominates all
316 LHSBO->moveBefore(I);
318 // Linearize the expression tree on the LHS.
319 LinearizeExprTree(LHSBO, Ops);
321 // Remember the RHS operand and its rank.
322 Ops.push_back(ValueEntry(getRank(RHS), RHS));
324 // Clear the RHS leaf out.
325 I->setOperand(1, UndefValue::get(I->getType()));
328 // RewriteExprTree - Now that the operands for this expression tree are
329 // linearized and optimized, emit them in-order. This function is written to be
331 void Reassociate::RewriteExprTree(BinaryOperator *I,
332 SmallVectorImpl<ValueEntry> &Ops,
334 if (i+2 == Ops.size()) {
335 if (I->getOperand(0) != Ops[i].Op ||
336 I->getOperand(1) != Ops[i+1].Op) {
337 Value *OldLHS = I->getOperand(0);
338 DEBUG(errs() << "RA: " << *I << '\n');
339 I->setOperand(0, Ops[i].Op);
340 I->setOperand(1, Ops[i+1].Op);
341 DEBUG(errs() << "TO: " << *I << '\n');
345 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3)
346 // delete the extra, now dead, nodes.
347 RemoveDeadBinaryOp(OldLHS);
351 assert(i+2 < Ops.size() && "Ops index out of range!");
353 if (I->getOperand(1) != Ops[i].Op) {
354 DEBUG(errs() << "RA: " << *I << '\n');
355 I->setOperand(1, Ops[i].Op);
356 DEBUG(errs() << "TO: " << *I << '\n');
361 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
362 assert(LHS->getOpcode() == I->getOpcode() &&
363 "Improper expression tree!");
365 // Compactify the tree instructions together with each other to guarantee
366 // that the expression tree is dominated by all of Ops.
368 RewriteExprTree(LHS, Ops, i+1);
373 // NegateValue - Insert instructions before the instruction pointed to by BI,
374 // that computes the negative version of the value specified. The negative
375 // version of the value is returned, and BI is left pointing at the instruction
376 // that should be processed next by the reassociation pass.
378 static Value *NegateValue(Value *V, Instruction *BI) {
379 // We are trying to expose opportunity for reassociation. One of the things
380 // that we want to do to achieve this is to push a negation as deep into an
381 // expression chain as possible, to expose the add instructions. In practice,
382 // this means that we turn this:
383 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
384 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
385 // the constants. We assume that instcombine will clean up the mess later if
386 // we introduce tons of unnecessary negation instructions...
388 if (Instruction *I = dyn_cast<Instruction>(V))
389 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
390 // Push the negates through the add.
391 I->setOperand(0, NegateValue(I->getOperand(0), BI));
392 I->setOperand(1, NegateValue(I->getOperand(1), BI));
394 // We must move the add instruction here, because the neg instructions do
395 // not dominate the old add instruction in general. By moving it, we are
396 // assured that the neg instructions we just inserted dominate the
397 // instruction we are about to insert after them.
400 I->setName(I->getName()+".neg");
404 // Insert a 'neg' instruction that subtracts the value from zero to get the
407 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
410 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
411 /// X-Y into (X + -Y).
412 static bool ShouldBreakUpSubtract(Instruction *Sub) {
413 // If this is a negation, we can't split it up!
414 if (BinaryOperator::isNeg(Sub))
417 // Don't bother to break this up unless either the LHS is an associable add or
418 // subtract or if this is only used by one.
419 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
420 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
422 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
423 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
425 if (Sub->hasOneUse() &&
426 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
427 isReassociableOp(Sub->use_back(), Instruction::Sub)))
433 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
434 /// only used by an add, transform this into (X+(0-Y)) to promote better
436 static Instruction *BreakUpSubtract(Instruction *Sub,
437 std::map<AssertingVH<>, unsigned> &ValueRankMap) {
438 // Convert a subtract into an add and a neg instruction... so that sub
439 // instructions can be commuted with other add instructions...
441 // Calculate the negative value of Operand 1 of the sub instruction...
442 // and set it as the RHS of the add instruction we just made...
444 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
446 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
449 // Everyone now refers to the add instruction.
450 ValueRankMap.erase(Sub);
451 Sub->replaceAllUsesWith(New);
452 Sub->eraseFromParent();
454 DEBUG(errs() << "Negated: " << *New << '\n');
458 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
459 /// by one, change this into a multiply by a constant to assist with further
461 static Instruction *ConvertShiftToMul(Instruction *Shl,
462 std::map<AssertingVH<>, unsigned> &ValueRankMap) {
463 // If an operand of this shift is a reassociable multiply, or if the shift
464 // is used by a reassociable multiply or add, turn into a multiply.
465 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) ||
467 (isReassociableOp(Shl->use_back(), Instruction::Mul) ||
468 isReassociableOp(Shl->use_back(), Instruction::Add)))) {
469 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
470 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
473 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
474 ValueRankMap.erase(Shl);
476 Shl->replaceAllUsesWith(Mul);
477 Shl->eraseFromParent();
483 // Scan backwards and forwards among values with the same rank as element i to
484 // see if X exists. If X does not exist, return i.
485 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
487 unsigned XRank = Ops[i].Rank;
488 unsigned e = Ops.size();
489 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
493 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
499 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
500 /// and returning the result. Insert the tree before I.
501 static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl<Value*> &Ops){
502 if (Ops.size() == 1) return Ops.back();
504 Value *V1 = Ops.back();
506 Value *V2 = EmitAddTreeOfValues(I, Ops);
507 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
510 /// RemoveFactorFromExpression - If V is an expression tree that is a
511 /// multiplication sequence, and if this sequence contains a multiply by Factor,
512 /// remove Factor from the tree and return the new tree.
513 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
514 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
517 SmallVector<ValueEntry, 8> Factors;
518 LinearizeExprTree(BO, Factors);
520 bool FoundFactor = false;
521 for (unsigned i = 0, e = Factors.size(); i != e; ++i)
522 if (Factors[i].Op == Factor) {
524 Factors.erase(Factors.begin()+i);
528 // Make sure to restore the operands to the expression tree.
529 RewriteExprTree(BO, Factors);
533 // If this was just a single multiply, remove the multiply and return the only
534 // remaining operand.
535 if (Factors.size() == 1) {
536 ValueRankMap.erase(BO);
537 BO->eraseFromParent();
538 return Factors[0].Op;
541 RewriteExprTree(BO, Factors);
545 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
546 /// add its operands as factors, otherwise add V to the list of factors.
547 static void FindSingleUseMultiplyFactors(Value *V,
548 SmallVectorImpl<Value*> &Factors) {
550 if ((!V->hasOneUse() && !V->use_empty()) ||
551 !(BO = dyn_cast<BinaryOperator>(V)) ||
552 BO->getOpcode() != Instruction::Mul) {
553 Factors.push_back(V);
557 // Otherwise, add the LHS and RHS to the list of factors.
558 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
559 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
562 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
563 /// instruction. This optimizes based on identities. If it can be reduced to
564 /// a single Value, it is returned, otherwise the Ops list is mutated as
566 static Value *OptimizeAndOrXor(unsigned Opcode,
567 SmallVectorImpl<ValueEntry> &Ops) {
568 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
569 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
570 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
571 // First, check for X and ~X in the operand list.
572 assert(i < Ops.size());
573 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
574 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
575 unsigned FoundX = FindInOperandList(Ops, i, X);
577 if (Opcode == Instruction::And) // ...&X&~X = 0
578 return Constant::getNullValue(X->getType());
580 if (Opcode == Instruction::Or) // ...|X|~X = -1
581 return Constant::getAllOnesValue(X->getType());
585 // Next, check for duplicate pairs of values, which we assume are next to
586 // each other, due to our sorting criteria.
587 assert(i < Ops.size());
588 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
589 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
590 // Drop duplicate values for And and Or.
591 Ops.erase(Ops.begin()+i);
597 // Drop pairs of values for Xor.
598 assert(Opcode == Instruction::Xor);
600 return Constant::getNullValue(Ops[0].Op->getType());
603 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
611 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
612 /// optimizes based on identities. If it can be reduced to a single Value, it
613 /// is returned, otherwise the Ops list is mutated as necessary.
614 Value *Reassociate::OptimizeAdd(Instruction *I,
615 SmallVectorImpl<ValueEntry> &Ops) {
616 // Scan the operand lists looking for X and -X pairs. If we find any, we
617 // can simplify the expression. X+-X == 0. While we're at it, scan for any
618 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
619 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
620 Value *TheOp = Ops[i].Op;
621 // Check to see if we've seen this operand before. If so, we factor all
622 // instances of the operand together. Due to our sorting criteria, we know
623 // that these need to be next to each other in the vector.
624 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
625 // Rescan the list, remove all instances of this operand from the expr.
626 unsigned NumFound = 0;
628 Ops.erase(Ops.begin()+i);
630 } while (i != Ops.size() && Ops[i].Op == TheOp);
632 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
635 // Insert a new multiply.
636 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound);
637 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I);
639 // Now that we have inserted a multiply, optimize it. This allows us to
640 // handle cases that require multiple factoring steps, such as this:
641 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
642 Mul = ReassociateExpression(cast<BinaryOperator>(Mul));
644 // If every add operand was a duplicate, return the multiply.
648 // Otherwise, we had some input that didn't have the dupe, such as
649 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
650 // things being added by this operation.
651 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
658 // Check for X and -X in the operand list.
659 if (!BinaryOperator::isNeg(TheOp))
662 Value *X = BinaryOperator::getNegArgument(TheOp);
663 unsigned FoundX = FindInOperandList(Ops, i, X);
667 // Remove X and -X from the operand list.
669 return Constant::getNullValue(X->getType());
671 Ops.erase(Ops.begin()+i);
675 --i; // Need to back up an extra one.
676 Ops.erase(Ops.begin()+FoundX);
678 --i; // Revisit element.
679 e -= 2; // Removed two elements.
682 // Scan the operand list, checking to see if there are any common factors
683 // between operands. Consider something like A*A+A*B*C+D. We would like to
684 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
685 // To efficiently find this, we count the number of times a factor occurs
686 // for any ADD operands that are MULs.
687 DenseMap<Value*, unsigned> FactorOccurrences;
689 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
690 // where they are actually the same multiply.
692 Value *MaxOccVal = 0;
693 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
694 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op);
695 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty())
698 // Compute all of the factors of this added value.
699 SmallVector<Value*, 8> Factors;
700 FindSingleUseMultiplyFactors(BOp, Factors);
701 assert(Factors.size() > 1 && "Bad linearize!");
703 // Add one to FactorOccurrences for each unique factor in this op.
704 if (Factors.size() == 2) {
705 unsigned Occ = ++FactorOccurrences[Factors[0]];
706 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; }
707 if (Factors[0] != Factors[1]) { // Don't double count A*A.
708 Occ = ++FactorOccurrences[Factors[1]];
709 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; }
712 SmallPtrSet<Value*, 4> Duplicates;
713 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
714 if (!Duplicates.insert(Factors[i])) continue;
716 unsigned Occ = ++FactorOccurrences[Factors[i]];
717 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; }
722 // If any factor occurred more than one time, we can pull it out.
724 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
727 // Create a new instruction that uses the MaxOccVal twice. If we don't do
728 // this, we could otherwise run into situations where removing a factor
729 // from an expression will drop a use of maxocc, and this can cause
730 // RemoveFactorFromExpression on successive values to behave differently.
731 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
732 SmallVector<Value*, 4> NewMulOps;
733 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
734 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
735 NewMulOps.push_back(V);
736 Ops.erase(Ops.begin()+i);
741 // No need for extra uses anymore.
744 unsigned NumAddedValues = NewMulOps.size();
745 Value *V = EmitAddTreeOfValues(I, NewMulOps);
747 // Now that we have inserted the add tree, optimize it. This allows us to
748 // handle cases that require multiple factoring steps, such as this:
749 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
750 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
751 V = ReassociateExpression(cast<BinaryOperator>(V));
753 // Create the multiply.
754 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
756 // Rerun associate on the multiply in case the inner expression turned into
757 // a multiply. We want to make sure that we keep things in canonical form.
758 V2 = ReassociateExpression(cast<BinaryOperator>(V2));
760 // If every add operand included the factor (e.g. "A*B + A*C"), then the
761 // entire result expression is just the multiply "A*(B+C)".
765 // Otherwise, we had some input that didn't have the factor, such as
766 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
767 // things being added by this operation.
768 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
774 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
775 SmallVectorImpl<ValueEntry> &Ops) {
776 // Now that we have the linearized expression tree, try to optimize it.
777 // Start by folding any constants that we found.
778 bool IterateOptimization = false;
779 if (Ops.size() == 1) return Ops[0].Op;
781 unsigned Opcode = I->getOpcode();
783 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
784 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
786 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
787 return OptimizeExpression(I, Ops);
790 // Check for destructive annihilation due to a constant being used.
791 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
794 case Instruction::And:
795 if (CstVal->isZero()) // ... & 0 -> 0
797 if (CstVal->isAllOnesValue()) // ... & -1 -> ...
800 case Instruction::Mul:
801 if (CstVal->isZero()) { // ... * 0 -> 0
806 if (cast<ConstantInt>(CstVal)->isOne())
807 Ops.pop_back(); // ... * 1 -> ...
809 case Instruction::Or:
810 if (CstVal->isAllOnesValue()) // ... | -1 -> -1
813 case Instruction::Add:
814 case Instruction::Xor:
815 if (CstVal->isZero()) // ... [|^+] 0 -> ...
819 if (Ops.size() == 1) return Ops[0].Op;
821 // Handle destructive annihilation due to identities between elements in the
822 // argument list here.
825 case Instruction::And:
826 case Instruction::Or:
827 case Instruction::Xor: {
828 unsigned NumOps = Ops.size();
829 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
831 IterateOptimization |= Ops.size() != NumOps;
835 case Instruction::Add: {
836 unsigned NumOps = Ops.size();
837 if (Value *Result = OptimizeAdd(I, Ops))
839 IterateOptimization |= Ops.size() != NumOps;
843 //case Instruction::Mul:
846 if (IterateOptimization)
847 return OptimizeExpression(I, Ops);
852 /// ReassociateBB - Inspect all of the instructions in this basic block,
853 /// reassociating them as we go.
854 void Reassociate::ReassociateBB(BasicBlock *BB) {
855 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) {
856 Instruction *BI = BBI++;
857 if (BI->getOpcode() == Instruction::Shl &&
858 isa<ConstantInt>(BI->getOperand(1)))
859 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) {
864 // Reject cases where it is pointless to do this.
865 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() ||
866 isa<VectorType>(BI->getType()))
867 continue; // Floating point ops are not associative.
869 // If this is a subtract instruction which is not already in negate form,
870 // see if we can convert it to X+-Y.
871 if (BI->getOpcode() == Instruction::Sub) {
872 if (ShouldBreakUpSubtract(BI)) {
873 BI = BreakUpSubtract(BI, ValueRankMap);
875 } else if (BinaryOperator::isNeg(BI)) {
876 // Otherwise, this is a negation. See if the operand is a multiply tree
877 // and if this is not an inner node of a multiply tree.
878 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) &&
880 !isReassociableOp(BI->use_back(), Instruction::Mul))) {
881 BI = LowerNegateToMultiply(BI, ValueRankMap);
887 // If this instruction is a commutative binary operator, process it.
888 if (!BI->isAssociative()) continue;
889 BinaryOperator *I = cast<BinaryOperator>(BI);
891 // If this is an interior node of a reassociable tree, ignore it until we
892 // get to the root of the tree, to avoid N^2 analysis.
893 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode()))
896 // If this is an add tree that is used by a sub instruction, ignore it
897 // until we process the subtract.
898 if (I->hasOneUse() && I->getOpcode() == Instruction::Add &&
899 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub)
902 ReassociateExpression(I);
906 Value *Reassociate::ReassociateExpression(BinaryOperator *I) {
908 // First, walk the expression tree, linearizing the tree, collecting the
909 // operand information.
910 SmallVector<ValueEntry, 8> Ops;
911 LinearizeExprTree(I, Ops);
913 DEBUG(errs() << "RAIn:\t"; PrintOps(I, Ops); errs() << '\n');
915 // Now that we have linearized the tree to a list and have gathered all of
916 // the operands and their ranks, sort the operands by their rank. Use a
917 // stable_sort so that values with equal ranks will have their relative
918 // positions maintained (and so the compiler is deterministic). Note that
919 // this sorts so that the highest ranking values end up at the beginning of
921 std::stable_sort(Ops.begin(), Ops.end());
923 // OptimizeExpression - Now that we have the expression tree in a convenient
924 // sorted form, optimize it globally if possible.
925 if (Value *V = OptimizeExpression(I, Ops)) {
926 // This expression tree simplified to something that isn't a tree,
928 DEBUG(errs() << "Reassoc to scalar: " << *V << '\n');
929 I->replaceAllUsesWith(V);
930 RemoveDeadBinaryOp(I);
935 // We want to sink immediates as deeply as possible except in the case where
936 // this is a multiply tree used only by an add, and the immediate is a -1.
937 // In this case we reassociate to put the negation on the outside so that we
938 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
939 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
940 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
941 isa<ConstantInt>(Ops.back().Op) &&
942 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
943 ValueEntry Tmp = Ops.pop_back_val();
944 Ops.insert(Ops.begin(), Tmp);
947 DEBUG(errs() << "RAOut:\t"; PrintOps(I, Ops); errs() << '\n');
949 if (Ops.size() == 1) {
950 // This expression tree simplified to something that isn't a tree,
952 I->replaceAllUsesWith(Ops[0].Op);
953 RemoveDeadBinaryOp(I);
957 // Now that we ordered and optimized the expressions, splat them back into
958 // the expression tree, removing any unneeded nodes.
959 RewriteExprTree(I, Ops);
964 bool Reassociate::runOnFunction(Function &F) {
965 // Recalculate the rank map for F
969 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
972 // We are done with the rank map...
974 ValueRankMap.clear();