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 errs() << Instruction::getOpcodeName(I->getOpcode()) << " "
64 << *Ops[0].Op->getType() << '\t';
65 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
67 WriteAsOperand(errs(), Ops[i].Op, false, M);
68 errs() << ", #" << 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) {}
82 bool runOnFunction(Function &F);
84 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
88 void BuildRankMap(Function &F);
89 unsigned getRank(Value *V);
90 Value *ReassociateExpression(BinaryOperator *I);
91 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops,
93 Value *OptimizeExpression(BinaryOperator *I,
94 SmallVectorImpl<ValueEntry> &Ops);
95 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
96 void LinearizeExprTree(BinaryOperator *I, SmallVectorImpl<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) || !Op->use_empty())
116 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1);
118 ValueRankMap.erase(Op);
119 Op->eraseFromParent();
120 RemoveDeadBinaryOp(LHS);
121 RemoveDeadBinaryOp(RHS);
125 static bool isUnmovableInstruction(Instruction *I) {
126 if (I->getOpcode() == Instruction::PHI ||
127 I->getOpcode() == Instruction::Alloca ||
128 I->getOpcode() == Instruction::Load ||
129 I->getOpcode() == Instruction::Invoke ||
130 (I->getOpcode() == Instruction::Call &&
131 !isa<DbgInfoIntrinsic>(I)) ||
132 I->getOpcode() == Instruction::UDiv ||
133 I->getOpcode() == Instruction::SDiv ||
134 I->getOpcode() == Instruction::FDiv ||
135 I->getOpcode() == Instruction::URem ||
136 I->getOpcode() == Instruction::SRem ||
137 I->getOpcode() == Instruction::FRem)
142 void Reassociate::BuildRankMap(Function &F) {
145 // Assign distinct ranks to function arguments
146 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
147 ValueRankMap[&*I] = ++i;
149 ReversePostOrderTraversal<Function*> RPOT(&F);
150 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
151 E = RPOT.end(); I != E; ++I) {
153 unsigned BBRank = RankMap[BB] = ++i << 16;
155 // Walk the basic block, adding precomputed ranks for any instructions that
156 // we cannot move. This ensures that the ranks for these instructions are
157 // all different in the block.
158 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
159 if (isUnmovableInstruction(I))
160 ValueRankMap[&*I] = ++BBRank;
164 unsigned Reassociate::getRank(Value *V) {
165 Instruction *I = dyn_cast<Instruction>(V);
167 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
168 return 0; // Otherwise it's a global or constant, rank 0.
171 if (unsigned Rank = ValueRankMap[I])
172 return Rank; // 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 ValueRankMap[I] = 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 DenseMap<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 if (Constant *C = dyn_cast<Constant>(V))
380 return ConstantExpr::getNeg(C);
382 // We are trying to expose opportunity for reassociation. One of the things
383 // that we want to do to achieve this is to push a negation as deep into an
384 // expression chain as possible, to expose the add instructions. In practice,
385 // this means that we turn this:
386 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
387 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
388 // the constants. We assume that instcombine will clean up the mess later if
389 // we introduce tons of unnecessary negation instructions.
391 if (Instruction *I = dyn_cast<Instruction>(V))
392 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
393 // Push the negates through the add.
394 I->setOperand(0, NegateValue(I->getOperand(0), BI));
395 I->setOperand(1, NegateValue(I->getOperand(1), BI));
397 // We must move the add instruction here, because the neg instructions do
398 // not dominate the old add instruction in general. By moving it, we are
399 // assured that the neg instructions we just inserted dominate the
400 // instruction we are about to insert after them.
403 I->setName(I->getName()+".neg");
407 // Okay, we need to materialize a negated version of V with an instruction.
408 // Scan the use lists of V to see if we have one already.
409 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
410 if (!BinaryOperator::isNeg(*UI)) continue;
412 // We found one! Now we have to make sure that the definition dominates
413 // this use. We do this by moving it to the entry block (if it is a
414 // non-instruction value) or right after the definition. These negates will
415 // be zapped by reassociate later, so we don't need much finesse here.
416 BinaryOperator *TheNeg = cast<BinaryOperator>(*UI);
418 // Verify that the negate is in this function, V might be a constant expr.
419 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
422 BasicBlock::iterator InsertPt;
423 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
424 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
425 InsertPt = II->getNormalDest()->begin();
427 InsertPt = InstInput;
430 while (isa<PHINode>(InsertPt)) ++InsertPt;
432 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
434 TheNeg->moveBefore(InsertPt);
438 // Insert a 'neg' instruction that subtracts the value from zero to get the
440 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
443 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
444 /// X-Y into (X + -Y).
445 static bool ShouldBreakUpSubtract(Instruction *Sub) {
446 // If this is a negation, we can't split it up!
447 if (BinaryOperator::isNeg(Sub))
450 // Don't bother to break this up unless either the LHS is an associable add or
451 // subtract or if this is only used by one.
452 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
453 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
455 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
456 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
458 if (Sub->hasOneUse() &&
459 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
460 isReassociableOp(Sub->use_back(), Instruction::Sub)))
466 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
467 /// only used by an add, transform this into (X+(0-Y)) to promote better
469 static Instruction *BreakUpSubtract(Instruction *Sub,
470 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) {
471 // Convert a subtract into an add and a neg instruction. This allows sub
472 // instructions to be commuted with other add instructions.
474 // Calculate the negative value of Operand 1 of the sub instruction,
475 // and set it as the RHS of the add instruction we just made.
477 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
479 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
482 // Everyone now refers to the add instruction.
483 ValueRankMap.erase(Sub);
484 Sub->replaceAllUsesWith(New);
485 Sub->eraseFromParent();
487 DEBUG(errs() << "Negated: " << *New << '\n');
491 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
492 /// by one, change this into a multiply by a constant to assist with further
494 static Instruction *ConvertShiftToMul(Instruction *Shl,
495 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) {
496 // If an operand of this shift is a reassociable multiply, or if the shift
497 // is used by a reassociable multiply or add, turn into a multiply.
498 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) ||
500 (isReassociableOp(Shl->use_back(), Instruction::Mul) ||
501 isReassociableOp(Shl->use_back(), Instruction::Add)))) {
502 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
503 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
506 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
507 ValueRankMap.erase(Shl);
509 Shl->replaceAllUsesWith(Mul);
510 Shl->eraseFromParent();
516 // Scan backwards and forwards among values with the same rank as element i to
517 // see if X exists. If X does not exist, return i. This is useful when
518 // scanning for 'x' when we see '-x' because they both get the same rank.
519 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
521 unsigned XRank = Ops[i].Rank;
522 unsigned e = Ops.size();
523 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
527 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
533 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
534 /// and returning the result. Insert the tree before I.
535 static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl<Value*> &Ops){
536 if (Ops.size() == 1) return Ops.back();
538 Value *V1 = Ops.back();
540 Value *V2 = EmitAddTreeOfValues(I, Ops);
541 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
544 /// RemoveFactorFromExpression - If V is an expression tree that is a
545 /// multiplication sequence, and if this sequence contains a multiply by Factor,
546 /// remove Factor from the tree and return the new tree.
547 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
548 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
551 SmallVector<ValueEntry, 8> Factors;
552 LinearizeExprTree(BO, Factors);
554 bool FoundFactor = false;
555 bool NeedsNegate = false;
556 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
557 if (Factors[i].Op == Factor) {
559 Factors.erase(Factors.begin()+i);
563 // If this is a negative version of this factor, remove it.
564 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor))
565 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
566 if (FC1->getValue() == -FC2->getValue()) {
567 FoundFactor = NeedsNegate = true;
568 Factors.erase(Factors.begin()+i);
574 // Make sure to restore the operands to the expression tree.
575 RewriteExprTree(BO, Factors);
579 BasicBlock::iterator InsertPt = BO; ++InsertPt;
581 // If this was just a single multiply, remove the multiply and return the only
582 // remaining operand.
583 if (Factors.size() == 1) {
584 ValueRankMap.erase(BO);
585 BO->eraseFromParent();
588 RewriteExprTree(BO, Factors);
593 V = BinaryOperator::CreateNeg(V, "neg", InsertPt);
598 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
599 /// add its operands as factors, otherwise add V to the list of factors.
600 static void FindSingleUseMultiplyFactors(Value *V,
601 SmallVectorImpl<Value*> &Factors) {
603 if ((!V->hasOneUse() && !V->use_empty()) ||
604 !(BO = dyn_cast<BinaryOperator>(V)) ||
605 BO->getOpcode() != Instruction::Mul) {
606 Factors.push_back(V);
610 // Otherwise, add the LHS and RHS to the list of factors.
611 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
612 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
615 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
616 /// instruction. This optimizes based on identities. If it can be reduced to
617 /// a single Value, it is returned, otherwise the Ops list is mutated as
619 static Value *OptimizeAndOrXor(unsigned Opcode,
620 SmallVectorImpl<ValueEntry> &Ops) {
621 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
622 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
623 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
624 // First, check for X and ~X in the operand list.
625 assert(i < Ops.size());
626 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
627 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
628 unsigned FoundX = FindInOperandList(Ops, i, X);
630 if (Opcode == Instruction::And) // ...&X&~X = 0
631 return Constant::getNullValue(X->getType());
633 if (Opcode == Instruction::Or) // ...|X|~X = -1
634 return Constant::getAllOnesValue(X->getType());
638 // Next, check for duplicate pairs of values, which we assume are next to
639 // each other, due to our sorting criteria.
640 assert(i < Ops.size());
641 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
642 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
643 // Drop duplicate values for And and Or.
644 Ops.erase(Ops.begin()+i);
650 // Drop pairs of values for Xor.
651 assert(Opcode == Instruction::Xor);
653 return Constant::getNullValue(Ops[0].Op->getType());
656 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
664 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
665 /// optimizes based on identities. If it can be reduced to a single Value, it
666 /// is returned, otherwise the Ops list is mutated as necessary.
667 Value *Reassociate::OptimizeAdd(Instruction *I,
668 SmallVectorImpl<ValueEntry> &Ops) {
669 // Scan the operand lists looking for X and -X pairs. If we find any, we
670 // can simplify the expression. X+-X == 0. While we're at it, scan for any
671 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
673 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1".
675 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
676 Value *TheOp = Ops[i].Op;
677 // Check to see if we've seen this operand before. If so, we factor all
678 // instances of the operand together. Due to our sorting criteria, we know
679 // that these need to be next to each other in the vector.
680 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
681 // Rescan the list, remove all instances of this operand from the expr.
682 unsigned NumFound = 0;
684 Ops.erase(Ops.begin()+i);
686 } while (i != Ops.size() && Ops[i].Op == TheOp);
688 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
691 // Insert a new multiply.
692 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound);
693 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I);
695 // Now that we have inserted a multiply, optimize it. This allows us to
696 // handle cases that require multiple factoring steps, such as this:
697 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
698 Mul = ReassociateExpression(cast<BinaryOperator>(Mul));
700 // If every add operand was a duplicate, return the multiply.
704 // Otherwise, we had some input that didn't have the dupe, such as
705 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
706 // things being added by this operation.
707 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
714 // Check for X and -X in the operand list.
715 if (!BinaryOperator::isNeg(TheOp))
718 Value *X = BinaryOperator::getNegArgument(TheOp);
719 unsigned FoundX = FindInOperandList(Ops, i, X);
723 // Remove X and -X from the operand list.
725 return Constant::getNullValue(X->getType());
727 Ops.erase(Ops.begin()+i);
731 --i; // Need to back up an extra one.
732 Ops.erase(Ops.begin()+FoundX);
734 --i; // Revisit element.
735 e -= 2; // Removed two elements.
738 // Scan the operand list, checking to see if there are any common factors
739 // between operands. Consider something like A*A+A*B*C+D. We would like to
740 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
741 // To efficiently find this, we count the number of times a factor occurs
742 // for any ADD operands that are MULs.
743 DenseMap<Value*, unsigned> FactorOccurrences;
745 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
746 // where they are actually the same multiply.
748 Value *MaxOccVal = 0;
749 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
750 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op);
751 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty())
754 // Compute all of the factors of this added value.
755 SmallVector<Value*, 8> Factors;
756 FindSingleUseMultiplyFactors(BOp, Factors);
757 assert(Factors.size() > 1 && "Bad linearize!");
759 // Add one to FactorOccurrences for each unique factor in this op.
760 SmallPtrSet<Value*, 8> Duplicates;
761 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
762 Value *Factor = Factors[i];
763 if (!Duplicates.insert(Factor)) continue;
765 unsigned Occ = ++FactorOccurrences[Factor];
766 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
768 // If Factor is a negative constant, add the negated value as a factor
769 // because we can percolate the negate out. Watch for minint, which
770 // cannot be positivified.
771 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor))
772 if (CI->getValue().isNegative() && !CI->getValue().isMinSignedValue()) {
773 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
774 assert(!Duplicates.count(Factor) &&
775 "Shouldn't have two constant factors, missed a canonicalize");
777 unsigned Occ = ++FactorOccurrences[Factor];
778 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
783 // If any factor occurred more than one time, we can pull it out.
785 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
788 // Create a new instruction that uses the MaxOccVal twice. If we don't do
789 // this, we could otherwise run into situations where removing a factor
790 // from an expression will drop a use of maxocc, and this can cause
791 // RemoveFactorFromExpression on successive values to behave differently.
792 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
793 SmallVector<Value*, 4> NewMulOps;
794 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
795 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
796 NewMulOps.push_back(V);
797 Ops.erase(Ops.begin()+i);
802 // No need for extra uses anymore.
805 unsigned NumAddedValues = NewMulOps.size();
806 Value *V = EmitAddTreeOfValues(I, NewMulOps);
808 // Now that we have inserted the add tree, optimize it. This allows us to
809 // handle cases that require multiple factoring steps, such as this:
810 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
811 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
812 V = ReassociateExpression(cast<BinaryOperator>(V));
814 // Create the multiply.
815 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
817 // Rerun associate on the multiply in case the inner expression turned into
818 // a multiply. We want to make sure that we keep things in canonical form.
819 V2 = ReassociateExpression(cast<BinaryOperator>(V2));
821 // If every add operand included the factor (e.g. "A*B + A*C"), then the
822 // entire result expression is just the multiply "A*(B+C)".
826 // Otherwise, we had some input that didn't have the factor, such as
827 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
828 // things being added by this operation.
829 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
835 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
836 SmallVectorImpl<ValueEntry> &Ops) {
837 // Now that we have the linearized expression tree, try to optimize it.
838 // Start by folding any constants that we found.
839 bool IterateOptimization = false;
840 if (Ops.size() == 1) return Ops[0].Op;
842 unsigned Opcode = I->getOpcode();
844 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
845 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
847 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
848 return OptimizeExpression(I, Ops);
851 // Check for destructive annihilation due to a constant being used.
852 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
855 case Instruction::And:
856 if (CstVal->isZero()) // X & 0 -> 0
858 if (CstVal->isAllOnesValue()) // X & -1 -> X
861 case Instruction::Mul:
862 if (CstVal->isZero()) { // X * 0 -> 0
867 if (cast<ConstantInt>(CstVal)->isOne())
868 Ops.pop_back(); // X * 1 -> X
870 case Instruction::Or:
871 if (CstVal->isAllOnesValue()) // X | -1 -> -1
874 case Instruction::Add:
875 case Instruction::Xor:
876 if (CstVal->isZero()) // X [|^+] 0 -> X
880 if (Ops.size() == 1) return Ops[0].Op;
882 // Handle destructive annihilation due to identities between elements in the
883 // argument list here.
886 case Instruction::And:
887 case Instruction::Or:
888 case Instruction::Xor: {
889 unsigned NumOps = Ops.size();
890 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
892 IterateOptimization |= Ops.size() != NumOps;
896 case Instruction::Add: {
897 unsigned NumOps = Ops.size();
898 if (Value *Result = OptimizeAdd(I, Ops))
900 IterateOptimization |= Ops.size() != NumOps;
904 //case Instruction::Mul:
907 if (IterateOptimization)
908 return OptimizeExpression(I, Ops);
913 /// ReassociateBB - Inspect all of the instructions in this basic block,
914 /// reassociating them as we go.
915 void Reassociate::ReassociateBB(BasicBlock *BB) {
916 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) {
917 Instruction *BI = BBI++;
918 if (BI->getOpcode() == Instruction::Shl &&
919 isa<ConstantInt>(BI->getOperand(1)))
920 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) {
925 // Reject cases where it is pointless to do this.
926 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() ||
927 isa<VectorType>(BI->getType()))
928 continue; // Floating point ops are not associative.
930 // If this is a subtract instruction which is not already in negate form,
931 // see if we can convert it to X+-Y.
932 if (BI->getOpcode() == Instruction::Sub) {
933 if (ShouldBreakUpSubtract(BI)) {
934 BI = BreakUpSubtract(BI, ValueRankMap);
936 } else if (BinaryOperator::isNeg(BI)) {
937 // Otherwise, this is a negation. See if the operand is a multiply tree
938 // and if this is not an inner node of a multiply tree.
939 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) &&
941 !isReassociableOp(BI->use_back(), Instruction::Mul))) {
942 BI = LowerNegateToMultiply(BI, ValueRankMap);
948 // If this instruction is a commutative binary operator, process it.
949 if (!BI->isAssociative()) continue;
950 BinaryOperator *I = cast<BinaryOperator>(BI);
952 // If this is an interior node of a reassociable tree, ignore it until we
953 // get to the root of the tree, to avoid N^2 analysis.
954 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode()))
957 // If this is an add tree that is used by a sub instruction, ignore it
958 // until we process the subtract.
959 if (I->hasOneUse() && I->getOpcode() == Instruction::Add &&
960 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub)
963 ReassociateExpression(I);
967 Value *Reassociate::ReassociateExpression(BinaryOperator *I) {
969 // First, walk the expression tree, linearizing the tree, collecting the
970 // operand information.
971 SmallVector<ValueEntry, 8> Ops;
972 LinearizeExprTree(I, Ops);
974 DEBUG(errs() << "RAIn:\t"; PrintOps(I, Ops); errs() << '\n');
976 // Now that we have linearized the tree to a list and have gathered all of
977 // the operands and their ranks, sort the operands by their rank. Use a
978 // stable_sort so that values with equal ranks will have their relative
979 // positions maintained (and so the compiler is deterministic). Note that
980 // this sorts so that the highest ranking values end up at the beginning of
982 std::stable_sort(Ops.begin(), Ops.end());
984 // OptimizeExpression - Now that we have the expression tree in a convenient
985 // sorted form, optimize it globally if possible.
986 if (Value *V = OptimizeExpression(I, Ops)) {
987 // This expression tree simplified to something that isn't a tree,
989 DEBUG(errs() << "Reassoc to scalar: " << *V << '\n');
990 I->replaceAllUsesWith(V);
991 RemoveDeadBinaryOp(I);
996 // We want to sink immediates as deeply as possible except in the case where
997 // this is a multiply tree used only by an add, and the immediate is a -1.
998 // In this case we reassociate to put the negation on the outside so that we
999 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
1000 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
1001 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
1002 isa<ConstantInt>(Ops.back().Op) &&
1003 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
1004 ValueEntry Tmp = Ops.pop_back_val();
1005 Ops.insert(Ops.begin(), Tmp);
1008 DEBUG(errs() << "RAOut:\t"; PrintOps(I, Ops); errs() << '\n');
1010 if (Ops.size() == 1) {
1011 // This expression tree simplified to something that isn't a tree,
1013 I->replaceAllUsesWith(Ops[0].Op);
1014 RemoveDeadBinaryOp(I);
1018 // Now that we ordered and optimized the expressions, splat them back into
1019 // the expression tree, removing any unneeded nodes.
1020 RewriteExprTree(I, Ops);
1025 bool Reassociate::runOnFunction(Function &F) {
1026 // Recalculate the rank map for F
1030 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
1033 // We are done with the rank map.
1035 ValueRankMap.clear();