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, etc.
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"
42 STATISTIC(NumLinear , "Number of insts linearized");
43 STATISTIC(NumChanged, "Number of insts reassociated");
44 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
45 STATISTIC(NumFactor , "Number of multiplies factored");
51 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {}
53 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) {
54 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start.
59 /// PrintOps - Print out the expression identified in the Ops list.
61 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
62 Module *M = I->getParent()->getParent()->getParent();
63 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
64 << *Ops[0].Op->getType() << '\t';
65 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
67 WriteAsOperand(dbgs(), Ops[i].Op, false, M);
68 dbgs() << ", #" << Ops[i].Rank << "] ";
74 class Reassociate : public FunctionPass {
75 DenseMap<BasicBlock*, unsigned> RankMap;
76 DenseMap<AssertingVH<>, unsigned> ValueRankMap;
79 static char ID; // Pass identification, replacement for typeid
80 Reassociate() : FunctionPass(ID) {
81 initializeReassociatePass(*PassRegistry::getPassRegistry());
84 bool runOnFunction(Function &F);
86 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
90 void BuildRankMap(Function &F);
91 unsigned getRank(Value *V);
92 Value *ReassociateExpression(BinaryOperator *I);
93 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops,
95 Value *OptimizeExpression(BinaryOperator *I,
96 SmallVectorImpl<ValueEntry> &Ops);
97 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
98 void LinearizeExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
99 void LinearizeExpr(BinaryOperator *I);
100 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
101 void ReassociateBB(BasicBlock *BB);
103 void RemoveDeadBinaryOp(Value *V);
107 char Reassociate::ID = 0;
108 INITIALIZE_PASS(Reassociate, "reassociate",
109 "Reassociate expressions", false, false)
111 // Public interface to the Reassociate pass
112 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
114 void Reassociate::RemoveDeadBinaryOp(Value *V) {
115 Instruction *Op = dyn_cast<Instruction>(V);
116 if (!Op || !isa<BinaryOperator>(Op) || !Op->use_empty())
119 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1);
121 ValueRankMap.erase(Op);
122 Op->eraseFromParent();
123 RemoveDeadBinaryOp(LHS);
124 RemoveDeadBinaryOp(RHS);
128 static bool isUnmovableInstruction(Instruction *I) {
129 if (I->getOpcode() == Instruction::PHI ||
130 I->getOpcode() == Instruction::Alloca ||
131 I->getOpcode() == Instruction::Load ||
132 I->getOpcode() == Instruction::Invoke ||
133 (I->getOpcode() == Instruction::Call &&
134 !isa<DbgInfoIntrinsic>(I)) ||
135 I->getOpcode() == Instruction::UDiv ||
136 I->getOpcode() == Instruction::SDiv ||
137 I->getOpcode() == Instruction::FDiv ||
138 I->getOpcode() == Instruction::URem ||
139 I->getOpcode() == Instruction::SRem ||
140 I->getOpcode() == Instruction::FRem)
145 void Reassociate::BuildRankMap(Function &F) {
148 // Assign distinct ranks to function arguments
149 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
150 ValueRankMap[&*I] = ++i;
152 ReversePostOrderTraversal<Function*> RPOT(&F);
153 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
154 E = RPOT.end(); I != E; ++I) {
156 unsigned BBRank = RankMap[BB] = ++i << 16;
158 // Walk the basic block, adding precomputed ranks for any instructions that
159 // we cannot move. This ensures that the ranks for these instructions are
160 // all different in the block.
161 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
162 if (isUnmovableInstruction(I))
163 ValueRankMap[&*I] = ++BBRank;
167 unsigned Reassociate::getRank(Value *V) {
168 Instruction *I = dyn_cast<Instruction>(V);
170 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
171 return 0; // Otherwise it's a global or constant, rank 0.
174 if (unsigned Rank = ValueRankMap[I])
175 return Rank; // Rank already known?
177 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
178 // we can reassociate expressions for code motion! Since we do not recurse
179 // for PHI nodes, we cannot have infinite recursion here, because there
180 // cannot be loops in the value graph that do not go through PHI nodes.
181 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
182 for (unsigned i = 0, e = I->getNumOperands();
183 i != e && Rank != MaxRank; ++i)
184 Rank = std::max(Rank, getRank(I->getOperand(i)));
186 // If this is a not or neg instruction, do not count it for rank. This
187 // assures us that X and ~X will have the same rank.
188 if (!I->getType()->isIntegerTy() ||
189 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
192 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = "
195 return ValueRankMap[I] = Rank;
198 /// isReassociableOp - Return true if V is an instruction of the specified
199 /// opcode and if it only has one use.
200 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
201 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) &&
202 cast<Instruction>(V)->getOpcode() == Opcode)
203 return cast<BinaryOperator>(V);
207 /// LowerNegateToMultiply - Replace 0-X with X*-1.
209 static Instruction *LowerNegateToMultiply(Instruction *Neg,
210 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) {
211 Constant *Cst = Constant::getAllOnesValue(Neg->getType());
213 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
214 ValueRankMap.erase(Neg);
216 Neg->replaceAllUsesWith(Res);
217 Neg->eraseFromParent();
221 // Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'.
222 // Note that if D is also part of the expression tree that we recurse to
223 // linearize it as well. Besides that case, this does not recurse into A,B, or
225 void Reassociate::LinearizeExpr(BinaryOperator *I) {
226 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
227 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1));
228 assert(isReassociableOp(LHS, I->getOpcode()) &&
229 isReassociableOp(RHS, I->getOpcode()) &&
230 "Not an expression that needs linearization?");
232 DEBUG(dbgs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n');
234 // Move the RHS instruction to live immediately before I, avoiding breaking
235 // dominator properties.
238 // Move operands around to do the linearization.
239 I->setOperand(1, RHS->getOperand(0));
240 RHS->setOperand(0, LHS);
241 I->setOperand(0, RHS);
245 DEBUG(dbgs() << "Linearized: " << *I << '\n');
247 // If D is part of this expression tree, tail recurse.
248 if (isReassociableOp(I->getOperand(1), I->getOpcode()))
253 /// LinearizeExprTree - Given an associative binary expression tree, traverse
254 /// all of the uses putting it into canonical form. This forces a left-linear
255 /// form of the expression (((a+b)+c)+d), and collects information about the
256 /// rank of the non-tree operands.
258 /// NOTE: These intentionally destroys the expression tree operands (turning
259 /// them into undef values) to reduce #uses of the values. This means that the
260 /// caller MUST use something like RewriteExprTree to put the values back in.
262 void Reassociate::LinearizeExprTree(BinaryOperator *I,
263 SmallVectorImpl<ValueEntry> &Ops) {
264 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
265 unsigned Opcode = I->getOpcode();
267 // First step, linearize the expression if it is in ((A+B)+(C+D)) form.
268 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode);
269 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode);
271 // If this is a multiply expression tree and it contains internal negations,
272 // transform them into multiplies by -1 so they can be reassociated.
273 if (I->getOpcode() == Instruction::Mul) {
274 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) {
275 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap);
276 LHSBO = isReassociableOp(LHS, Opcode);
278 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) {
279 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap);
280 RHSBO = isReassociableOp(RHS, Opcode);
286 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As
287 // such, just remember these operands and their rank.
288 Ops.push_back(ValueEntry(getRank(LHS), LHS));
289 Ops.push_back(ValueEntry(getRank(RHS), RHS));
291 // Clear the leaves out.
292 I->setOperand(0, UndefValue::get(I->getType()));
293 I->setOperand(1, UndefValue::get(I->getType()));
297 // Turn X+(Y+Z) -> (Y+Z)+X
298 std::swap(LHSBO, RHSBO);
300 bool Success = !I->swapOperands();
301 assert(Success && "swapOperands failed");
305 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the RHS is not
306 // part of the expression tree.
308 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0));
309 RHS = I->getOperand(1);
313 // Okay, now we know that the LHS is a nested expression and that the RHS is
314 // not. Perform reassociation.
315 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!");
317 // Move LHS right before I to make sure that the tree expression dominates all
319 LHSBO->moveBefore(I);
321 // Linearize the expression tree on the LHS.
322 LinearizeExprTree(LHSBO, Ops);
324 // Remember the RHS operand and its rank.
325 Ops.push_back(ValueEntry(getRank(RHS), RHS));
327 // Clear the RHS leaf out.
328 I->setOperand(1, UndefValue::get(I->getType()));
331 // RewriteExprTree - Now that the operands for this expression tree are
332 // linearized and optimized, emit them in-order. This function is written to be
334 void Reassociate::RewriteExprTree(BinaryOperator *I,
335 SmallVectorImpl<ValueEntry> &Ops,
337 if (i+2 == Ops.size()) {
338 if (I->getOperand(0) != Ops[i].Op ||
339 I->getOperand(1) != Ops[i+1].Op) {
340 Value *OldLHS = I->getOperand(0);
341 DEBUG(dbgs() << "RA: " << *I << '\n');
342 I->setOperand(0, Ops[i].Op);
343 I->setOperand(1, Ops[i+1].Op);
344 DEBUG(dbgs() << "TO: " << *I << '\n');
348 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3)
349 // delete the extra, now dead, nodes.
350 RemoveDeadBinaryOp(OldLHS);
354 assert(i+2 < Ops.size() && "Ops index out of range!");
356 if (I->getOperand(1) != Ops[i].Op) {
357 DEBUG(dbgs() << "RA: " << *I << '\n');
358 I->setOperand(1, Ops[i].Op);
359 DEBUG(dbgs() << "TO: " << *I << '\n');
364 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
365 assert(LHS->getOpcode() == I->getOpcode() &&
366 "Improper expression tree!");
368 // Compactify the tree instructions together with each other to guarantee
369 // that the expression tree is dominated by all of Ops.
371 RewriteExprTree(LHS, Ops, i+1);
376 // NegateValue - Insert instructions before the instruction pointed to by BI,
377 // that computes the negative version of the value specified. The negative
378 // version of the value is returned, and BI is left pointing at the instruction
379 // that should be processed next by the reassociation pass.
381 static Value *NegateValue(Value *V, Instruction *BI) {
382 if (Constant *C = dyn_cast<Constant>(V))
383 return ConstantExpr::getNeg(C);
385 // We are trying to expose opportunity for reassociation. One of the things
386 // that we want to do to achieve this is to push a negation as deep into an
387 // expression chain as possible, to expose the add instructions. In practice,
388 // this means that we turn this:
389 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
390 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
391 // the constants. We assume that instcombine will clean up the mess later if
392 // we introduce tons of unnecessary negation instructions.
394 if (Instruction *I = dyn_cast<Instruction>(V))
395 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
396 // Push the negates through the add.
397 I->setOperand(0, NegateValue(I->getOperand(0), BI));
398 I->setOperand(1, NegateValue(I->getOperand(1), BI));
400 // We must move the add instruction here, because the neg instructions do
401 // not dominate the old add instruction in general. By moving it, we are
402 // assured that the neg instructions we just inserted dominate the
403 // instruction we are about to insert after them.
406 I->setName(I->getName()+".neg");
410 // Okay, we need to materialize a negated version of V with an instruction.
411 // Scan the use lists of V to see if we have one already.
412 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
414 if (!BinaryOperator::isNeg(U)) continue;
416 // We found one! Now we have to make sure that the definition dominates
417 // this use. We do this by moving it to the entry block (if it is a
418 // non-instruction value) or right after the definition. These negates will
419 // be zapped by reassociate later, so we don't need much finesse here.
420 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
422 // Verify that the negate is in this function, V might be a constant expr.
423 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
426 BasicBlock::iterator InsertPt;
427 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
428 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
429 InsertPt = II->getNormalDest()->begin();
431 InsertPt = InstInput;
434 while (isa<PHINode>(InsertPt)) ++InsertPt;
436 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
438 TheNeg->moveBefore(InsertPt);
442 // Insert a 'neg' instruction that subtracts the value from zero to get the
444 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
447 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
448 /// X-Y into (X + -Y).
449 static bool ShouldBreakUpSubtract(Instruction *Sub) {
450 // If this is a negation, we can't split it up!
451 if (BinaryOperator::isNeg(Sub))
454 // Don't bother to break this up unless either the LHS is an associable add or
455 // subtract or if this is only used by one.
456 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
457 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
459 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
460 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
462 if (Sub->hasOneUse() &&
463 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
464 isReassociableOp(Sub->use_back(), Instruction::Sub)))
470 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
471 /// only used by an add, transform this into (X+(0-Y)) to promote better
473 static Instruction *BreakUpSubtract(Instruction *Sub,
474 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) {
475 // Convert a subtract into an add and a neg instruction. This allows sub
476 // instructions to be commuted with other add instructions.
478 // Calculate the negative value of Operand 1 of the sub instruction,
479 // and set it as the RHS of the add instruction we just made.
481 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
483 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
486 // Everyone now refers to the add instruction.
487 ValueRankMap.erase(Sub);
488 Sub->replaceAllUsesWith(New);
489 Sub->eraseFromParent();
491 DEBUG(dbgs() << "Negated: " << *New << '\n');
495 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
496 /// by one, change this into a multiply by a constant to assist with further
498 static Instruction *ConvertShiftToMul(Instruction *Shl,
499 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) {
500 // If an operand of this shift is a reassociable multiply, or if the shift
501 // is used by a reassociable multiply or add, turn into a multiply.
502 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) ||
504 (isReassociableOp(Shl->use_back(), Instruction::Mul) ||
505 isReassociableOp(Shl->use_back(), Instruction::Add)))) {
506 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
507 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
510 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
511 ValueRankMap.erase(Shl);
513 Shl->replaceAllUsesWith(Mul);
514 Shl->eraseFromParent();
520 // Scan backwards and forwards among values with the same rank as element i to
521 // see if X exists. If X does not exist, return i. This is useful when
522 // scanning for 'x' when we see '-x' because they both get the same rank.
523 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
525 unsigned XRank = Ops[i].Rank;
526 unsigned e = Ops.size();
527 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
531 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
537 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
538 /// and returning the result. Insert the tree before I.
539 static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl<Value*> &Ops){
540 if (Ops.size() == 1) return Ops.back();
542 Value *V1 = Ops.back();
544 Value *V2 = EmitAddTreeOfValues(I, Ops);
545 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
548 /// RemoveFactorFromExpression - If V is an expression tree that is a
549 /// multiplication sequence, and if this sequence contains a multiply by Factor,
550 /// remove Factor from the tree and return the new tree.
551 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
552 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
555 SmallVector<ValueEntry, 8> Factors;
556 LinearizeExprTree(BO, Factors);
558 bool FoundFactor = false;
559 bool NeedsNegate = false;
560 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
561 if (Factors[i].Op == Factor) {
563 Factors.erase(Factors.begin()+i);
567 // If this is a negative version of this factor, remove it.
568 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor))
569 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
570 if (FC1->getValue() == -FC2->getValue()) {
571 FoundFactor = NeedsNegate = true;
572 Factors.erase(Factors.begin()+i);
578 // Make sure to restore the operands to the expression tree.
579 RewriteExprTree(BO, Factors);
583 BasicBlock::iterator InsertPt = BO; ++InsertPt;
585 // If this was just a single multiply, remove the multiply and return the only
586 // remaining operand.
587 if (Factors.size() == 1) {
588 ValueRankMap.erase(BO);
589 BO->eraseFromParent();
592 RewriteExprTree(BO, Factors);
597 V = BinaryOperator::CreateNeg(V, "neg", InsertPt);
602 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
603 /// add its operands as factors, otherwise add V to the list of factors.
605 /// Ops is the top-level list of add operands we're trying to factor.
606 static void FindSingleUseMultiplyFactors(Value *V,
607 SmallVectorImpl<Value*> &Factors,
608 const SmallVectorImpl<ValueEntry> &Ops,
611 if (!(V->hasOneUse() || V->use_empty()) || // More than one use.
612 !(BO = dyn_cast<BinaryOperator>(V)) ||
613 BO->getOpcode() != Instruction::Mul) {
614 Factors.push_back(V);
618 // If this value has a single use because it is another input to the add
619 // tree we're reassociating and we dropped its use, it actually has two
620 // uses and we can't factor it.
622 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
623 if (Ops[i].Op == V) {
624 Factors.push_back(V);
630 // Otherwise, add the LHS and RHS to the list of factors.
631 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops, false);
632 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops, false);
635 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
636 /// instruction. This optimizes based on identities. If it can be reduced to
637 /// a single Value, it is returned, otherwise the Ops list is mutated as
639 static Value *OptimizeAndOrXor(unsigned Opcode,
640 SmallVectorImpl<ValueEntry> &Ops) {
641 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
642 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
643 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
644 // First, check for X and ~X in the operand list.
645 assert(i < Ops.size());
646 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
647 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
648 unsigned FoundX = FindInOperandList(Ops, i, X);
650 if (Opcode == Instruction::And) // ...&X&~X = 0
651 return Constant::getNullValue(X->getType());
653 if (Opcode == Instruction::Or) // ...|X|~X = -1
654 return Constant::getAllOnesValue(X->getType());
658 // Next, check for duplicate pairs of values, which we assume are next to
659 // each other, due to our sorting criteria.
660 assert(i < Ops.size());
661 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
662 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
663 // Drop duplicate values for And and Or.
664 Ops.erase(Ops.begin()+i);
670 // Drop pairs of values for Xor.
671 assert(Opcode == Instruction::Xor);
673 return Constant::getNullValue(Ops[0].Op->getType());
676 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
684 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
685 /// optimizes based on identities. If it can be reduced to a single Value, it
686 /// is returned, otherwise the Ops list is mutated as necessary.
687 Value *Reassociate::OptimizeAdd(Instruction *I,
688 SmallVectorImpl<ValueEntry> &Ops) {
689 // Scan the operand lists looking for X and -X pairs. If we find any, we
690 // can simplify the expression. X+-X == 0. While we're at it, scan for any
691 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
693 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1".
695 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
696 Value *TheOp = Ops[i].Op;
697 // Check to see if we've seen this operand before. If so, we factor all
698 // instances of the operand together. Due to our sorting criteria, we know
699 // that these need to be next to each other in the vector.
700 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
701 // Rescan the list, remove all instances of this operand from the expr.
702 unsigned NumFound = 0;
704 Ops.erase(Ops.begin()+i);
706 } while (i != Ops.size() && Ops[i].Op == TheOp);
708 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
711 // Insert a new multiply.
712 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound);
713 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I);
715 // Now that we have inserted a multiply, optimize it. This allows us to
716 // handle cases that require multiple factoring steps, such as this:
717 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
718 Mul = ReassociateExpression(cast<BinaryOperator>(Mul));
720 // If every add operand was a duplicate, return the multiply.
724 // Otherwise, we had some input that didn't have the dupe, such as
725 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
726 // things being added by this operation.
727 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
734 // Check for X and -X in the operand list.
735 if (!BinaryOperator::isNeg(TheOp))
738 Value *X = BinaryOperator::getNegArgument(TheOp);
739 unsigned FoundX = FindInOperandList(Ops, i, X);
743 // Remove X and -X from the operand list.
745 return Constant::getNullValue(X->getType());
747 Ops.erase(Ops.begin()+i);
751 --i; // Need to back up an extra one.
752 Ops.erase(Ops.begin()+FoundX);
754 --i; // Revisit element.
755 e -= 2; // Removed two elements.
758 // Scan the operand list, checking to see if there are any common factors
759 // between operands. Consider something like A*A+A*B*C+D. We would like to
760 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
761 // To efficiently find this, we count the number of times a factor occurs
762 // for any ADD operands that are MULs.
763 DenseMap<Value*, unsigned> FactorOccurrences;
765 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
766 // where they are actually the same multiply.
768 Value *MaxOccVal = 0;
769 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
770 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op);
771 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty())
774 // Compute all of the factors of this added value.
775 SmallVector<Value*, 8> Factors;
776 FindSingleUseMultiplyFactors(BOp, Factors, Ops, true);
777 assert(Factors.size() > 1 && "Bad linearize!");
779 // Add one to FactorOccurrences for each unique factor in this op.
780 SmallPtrSet<Value*, 8> Duplicates;
781 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
782 Value *Factor = Factors[i];
783 if (!Duplicates.insert(Factor)) continue;
785 unsigned Occ = ++FactorOccurrences[Factor];
786 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
788 // If Factor is a negative constant, add the negated value as a factor
789 // because we can percolate the negate out. Watch for minint, which
790 // cannot be positivified.
791 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor))
792 if (CI->getValue().isNegative() && !CI->getValue().isMinSignedValue()) {
793 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
794 assert(!Duplicates.count(Factor) &&
795 "Shouldn't have two constant factors, missed a canonicalize");
797 unsigned Occ = ++FactorOccurrences[Factor];
798 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
803 // If any factor occurred more than one time, we can pull it out.
805 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
808 // Create a new instruction that uses the MaxOccVal twice. If we don't do
809 // this, we could otherwise run into situations where removing a factor
810 // from an expression will drop a use of maxocc, and this can cause
811 // RemoveFactorFromExpression on successive values to behave differently.
812 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
813 SmallVector<Value*, 4> NewMulOps;
814 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
815 // Only try to remove factors from expressions we're allowed to.
816 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op);
817 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty())
820 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
821 NewMulOps.push_back(V);
822 Ops.erase(Ops.begin()+i);
827 // No need for extra uses anymore.
830 unsigned NumAddedValues = NewMulOps.size();
831 Value *V = EmitAddTreeOfValues(I, NewMulOps);
833 // Now that we have inserted the add tree, optimize it. This allows us to
834 // handle cases that require multiple factoring steps, such as this:
835 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
836 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
837 (void)NumAddedValues;
838 V = ReassociateExpression(cast<BinaryOperator>(V));
840 // Create the multiply.
841 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
843 // Rerun associate on the multiply in case the inner expression turned into
844 // a multiply. We want to make sure that we keep things in canonical form.
845 V2 = ReassociateExpression(cast<BinaryOperator>(V2));
847 // If every add operand included the factor (e.g. "A*B + A*C"), then the
848 // entire result expression is just the multiply "A*(B+C)".
852 // Otherwise, we had some input that didn't have the factor, such as
853 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
854 // things being added by this operation.
855 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
861 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
862 SmallVectorImpl<ValueEntry> &Ops) {
863 // Now that we have the linearized expression tree, try to optimize it.
864 // Start by folding any constants that we found.
865 bool IterateOptimization = false;
866 if (Ops.size() == 1) return Ops[0].Op;
868 unsigned Opcode = I->getOpcode();
870 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
871 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
873 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
874 return OptimizeExpression(I, Ops);
877 // Check for destructive annihilation due to a constant being used.
878 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
881 case Instruction::And:
882 if (CstVal->isZero()) // X & 0 -> 0
884 if (CstVal->isAllOnesValue()) // X & -1 -> X
887 case Instruction::Mul:
888 if (CstVal->isZero()) { // X * 0 -> 0
893 if (cast<ConstantInt>(CstVal)->isOne())
894 Ops.pop_back(); // X * 1 -> X
896 case Instruction::Or:
897 if (CstVal->isAllOnesValue()) // X | -1 -> -1
900 case Instruction::Add:
901 case Instruction::Xor:
902 if (CstVal->isZero()) // X [|^+] 0 -> X
906 if (Ops.size() == 1) return Ops[0].Op;
908 // Handle destructive annihilation due to identities between elements in the
909 // argument list here.
912 case Instruction::And:
913 case Instruction::Or:
914 case Instruction::Xor: {
915 unsigned NumOps = Ops.size();
916 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
918 IterateOptimization |= Ops.size() != NumOps;
922 case Instruction::Add: {
923 unsigned NumOps = Ops.size();
924 if (Value *Result = OptimizeAdd(I, Ops))
926 IterateOptimization |= Ops.size() != NumOps;
930 //case Instruction::Mul:
933 if (IterateOptimization)
934 return OptimizeExpression(I, Ops);
939 /// ReassociateBB - Inspect all of the instructions in this basic block,
940 /// reassociating them as we go.
941 void Reassociate::ReassociateBB(BasicBlock *BB) {
942 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) {
943 Instruction *BI = BBI++;
944 if (BI->getOpcode() == Instruction::Shl &&
945 isa<ConstantInt>(BI->getOperand(1)))
946 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) {
951 // Reject cases where it is pointless to do this.
952 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPointTy() ||
953 BI->getType()->isVectorTy())
954 continue; // Floating point ops are not associative.
956 // Do not reassociate boolean (i1) expressions. We want to preserve the
957 // original order of evaluation for short-circuited comparisons that
958 // SimplifyCFG has folded to AND/OR expressions. If the expression
959 // is not further optimized, it is likely to be transformed back to a
960 // short-circuited form for code gen, and the source order may have been
961 // optimized for the most likely conditions.
962 if (BI->getType()->isIntegerTy(1))
965 // If this is a subtract instruction which is not already in negate form,
966 // see if we can convert it to X+-Y.
967 if (BI->getOpcode() == Instruction::Sub) {
968 if (ShouldBreakUpSubtract(BI)) {
969 BI = BreakUpSubtract(BI, ValueRankMap);
970 // Reset the BBI iterator in case BreakUpSubtract changed the
971 // instruction it points to.
975 } else if (BinaryOperator::isNeg(BI)) {
976 // Otherwise, this is a negation. See if the operand is a multiply tree
977 // and if this is not an inner node of a multiply tree.
978 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) &&
980 !isReassociableOp(BI->use_back(), Instruction::Mul))) {
981 BI = LowerNegateToMultiply(BI, ValueRankMap);
987 // If this instruction is a commutative binary operator, process it.
988 if (!BI->isAssociative()) continue;
989 BinaryOperator *I = cast<BinaryOperator>(BI);
991 // If this is an interior node of a reassociable tree, ignore it until we
992 // get to the root of the tree, to avoid N^2 analysis.
993 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode()))
996 // If this is an add tree that is used by a sub instruction, ignore it
997 // until we process the subtract.
998 if (I->hasOneUse() && I->getOpcode() == Instruction::Add &&
999 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub)
1002 ReassociateExpression(I);
1006 Value *Reassociate::ReassociateExpression(BinaryOperator *I) {
1008 // First, walk the expression tree, linearizing the tree, collecting the
1009 // operand information.
1010 SmallVector<ValueEntry, 8> Ops;
1011 LinearizeExprTree(I, Ops);
1013 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
1015 // Now that we have linearized the tree to a list and have gathered all of
1016 // the operands and their ranks, sort the operands by their rank. Use a
1017 // stable_sort so that values with equal ranks will have their relative
1018 // positions maintained (and so the compiler is deterministic). Note that
1019 // this sorts so that the highest ranking values end up at the beginning of
1021 std::stable_sort(Ops.begin(), Ops.end());
1023 // OptimizeExpression - Now that we have the expression tree in a convenient
1024 // sorted form, optimize it globally if possible.
1025 if (Value *V = OptimizeExpression(I, Ops)) {
1026 // This expression tree simplified to something that isn't a tree,
1028 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
1029 I->replaceAllUsesWith(V);
1030 RemoveDeadBinaryOp(I);
1035 // We want to sink immediates as deeply as possible except in the case where
1036 // this is a multiply tree used only by an add, and the immediate is a -1.
1037 // In this case we reassociate to put the negation on the outside so that we
1038 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
1039 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
1040 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
1041 isa<ConstantInt>(Ops.back().Op) &&
1042 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
1043 ValueEntry Tmp = Ops.pop_back_val();
1044 Ops.insert(Ops.begin(), Tmp);
1047 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
1049 if (Ops.size() == 1) {
1050 // This expression tree simplified to something that isn't a tree,
1052 I->replaceAllUsesWith(Ops[0].Op);
1053 RemoveDeadBinaryOp(I);
1057 // Now that we ordered and optimized the expressions, splat them back into
1058 // the expression tree, removing any unneeded nodes.
1059 RewriteExprTree(I, Ops);
1064 bool Reassociate::runOnFunction(Function &F) {
1065 // Recalculate the rank map for F
1069 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
1072 // We are done with the rank map.
1074 ValueRankMap.clear();