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/Transforms/Utils/Local.h"
26 #include "llvm/Constants.h"
27 #include "llvm/DerivedTypes.h"
28 #include "llvm/Function.h"
29 #include "llvm/Instructions.h"
30 #include "llvm/IntrinsicInst.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Assembly/Writer.h"
33 #include "llvm/Support/CFG.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ValueHandle.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/ADT/PostOrderIterator.h"
38 #include "llvm/ADT/Statistic.h"
39 #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 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
65 << *Ops[0].Op->getType() << '\t';
66 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
68 WriteAsOperand(dbgs(), Ops[i].Op, false, M);
69 dbgs() << ", #" << Ops[i].Rank << "] ";
75 class Reassociate : public FunctionPass {
76 DenseMap<BasicBlock*, unsigned> RankMap;
77 DenseMap<AssertingVH<>, unsigned> ValueRankMap;
78 DenseMap<Value *, DbgValueInst *> DbgValues;
79 SmallVector<WeakVH, 8> RedoInsts;
80 SmallVector<WeakVH, 8> DeadInsts;
83 static char ID; // Pass identification, replacement for typeid
84 Reassociate() : FunctionPass(ID) {
85 initializeReassociatePass(*PassRegistry::getPassRegistry());
88 bool runOnFunction(Function &F);
90 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
94 void BuildRankMap(Function &F);
95 unsigned getRank(Value *V);
96 Value *ReassociateExpression(BinaryOperator *I);
97 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops,
99 Value *OptimizeExpression(BinaryOperator *I,
100 SmallVectorImpl<ValueEntry> &Ops);
101 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops);
102 void LinearizeExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops);
103 void LinearizeExpr(BinaryOperator *I);
104 Value *RemoveFactorFromExpression(Value *V, Value *Factor);
105 void ReassociateInst(BasicBlock::iterator &BBI);
107 void RemoveDeadBinaryOp(Value *V);
109 /// collectDbgValues - Collect all llvm.dbg.value intrinsics.
110 void collectDbgValues(Function &F);
114 char Reassociate::ID = 0;
115 INITIALIZE_PASS(Reassociate, "reassociate",
116 "Reassociate expressions", false, false)
118 // Public interface to the Reassociate pass
119 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); }
121 void Reassociate::RemoveDeadBinaryOp(Value *V) {
122 Instruction *Op = dyn_cast<Instruction>(V);
123 if (!Op || !isa<BinaryOperator>(Op))
126 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1);
128 ValueRankMap.erase(Op);
129 DeadInsts.push_back(Op);
130 RemoveDeadBinaryOp(LHS);
131 RemoveDeadBinaryOp(RHS);
135 static bool isUnmovableInstruction(Instruction *I) {
136 if (I->getOpcode() == Instruction::PHI ||
137 I->getOpcode() == Instruction::Alloca ||
138 I->getOpcode() == Instruction::Load ||
139 I->getOpcode() == Instruction::Invoke ||
140 (I->getOpcode() == Instruction::Call &&
141 !isa<DbgInfoIntrinsic>(I)) ||
142 I->getOpcode() == Instruction::UDiv ||
143 I->getOpcode() == Instruction::SDiv ||
144 I->getOpcode() == Instruction::FDiv ||
145 I->getOpcode() == Instruction::URem ||
146 I->getOpcode() == Instruction::SRem ||
147 I->getOpcode() == Instruction::FRem)
152 void Reassociate::BuildRankMap(Function &F) {
155 // Assign distinct ranks to function arguments
156 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
157 ValueRankMap[&*I] = ++i;
159 ReversePostOrderTraversal<Function*> RPOT(&F);
160 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
161 E = RPOT.end(); I != E; ++I) {
163 unsigned BBRank = RankMap[BB] = ++i << 16;
165 // Walk the basic block, adding precomputed ranks for any instructions that
166 // we cannot move. This ensures that the ranks for these instructions are
167 // all different in the block.
168 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
169 if (isUnmovableInstruction(I))
170 ValueRankMap[&*I] = ++BBRank;
174 unsigned Reassociate::getRank(Value *V) {
175 Instruction *I = dyn_cast<Instruction>(V);
177 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
178 return 0; // Otherwise it's a global or constant, rank 0.
181 if (unsigned Rank = ValueRankMap[I])
182 return Rank; // Rank already known?
184 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
185 // we can reassociate expressions for code motion! Since we do not recurse
186 // for PHI nodes, we cannot have infinite recursion here, because there
187 // cannot be loops in the value graph that do not go through PHI nodes.
188 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
189 for (unsigned i = 0, e = I->getNumOperands();
190 i != e && Rank != MaxRank; ++i)
191 Rank = std::max(Rank, getRank(I->getOperand(i)));
193 // If this is a not or neg instruction, do not count it for rank. This
194 // assures us that X and ~X will have the same rank.
195 if (!I->getType()->isIntegerTy() ||
196 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I)))
199 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = "
202 return ValueRankMap[I] = Rank;
205 /// isReassociableOp - Return true if V is an instruction of the specified
206 /// opcode and if it only has one use.
207 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
208 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) &&
209 cast<Instruction>(V)->getOpcode() == Opcode)
210 return cast<BinaryOperator>(V);
214 /// LowerNegateToMultiply - Replace 0-X with X*-1.
216 static Instruction *LowerNegateToMultiply(Instruction *Neg,
217 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) {
218 Constant *Cst = Constant::getAllOnesValue(Neg->getType());
220 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg);
221 ValueRankMap.erase(Neg);
223 Neg->replaceAllUsesWith(Res);
224 Res->setDebugLoc(Neg->getDebugLoc());
225 Neg->eraseFromParent();
229 // Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'.
230 // Note that if D is also part of the expression tree that we recurse to
231 // linearize it as well. Besides that case, this does not recurse into A,B, or
233 void Reassociate::LinearizeExpr(BinaryOperator *I) {
234 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
235 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1));
236 assert(isReassociableOp(LHS, I->getOpcode()) &&
237 isReassociableOp(RHS, I->getOpcode()) &&
238 "Not an expression that needs linearization?");
240 DEBUG(dbgs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n');
242 // Move the RHS instruction to live immediately before I, avoiding breaking
243 // dominator properties.
246 // Move operands around to do the linearization.
247 I->setOperand(1, RHS->getOperand(0));
248 RHS->setOperand(0, LHS);
249 I->setOperand(0, RHS);
251 // Conservatively clear all the optional flags, which may not hold
252 // after the reassociation.
253 I->clearSubclassOptionalData();
254 LHS->clearSubclassOptionalData();
255 RHS->clearSubclassOptionalData();
259 DEBUG(dbgs() << "Linearized: " << *I << '\n');
261 // If D is part of this expression tree, tail recurse.
262 if (isReassociableOp(I->getOperand(1), I->getOpcode()))
267 /// LinearizeExprTree - Given an associative binary expression tree, traverse
268 /// all of the uses putting it into canonical form. This forces a left-linear
269 /// form of the expression (((a+b)+c)+d), and collects information about the
270 /// rank of the non-tree operands.
272 /// NOTE: These intentionally destroys the expression tree operands (turning
273 /// them into undef values) to reduce #uses of the values. This means that the
274 /// caller MUST use something like RewriteExprTree to put the values back in.
276 void Reassociate::LinearizeExprTree(BinaryOperator *I,
277 SmallVectorImpl<ValueEntry> &Ops) {
278 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
279 unsigned Opcode = I->getOpcode();
281 // First step, linearize the expression if it is in ((A+B)+(C+D)) form.
282 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode);
283 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode);
285 // If this is a multiply expression tree and it contains internal negations,
286 // transform them into multiplies by -1 so they can be reassociated.
287 if (I->getOpcode() == Instruction::Mul) {
288 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) {
289 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap);
290 LHSBO = isReassociableOp(LHS, Opcode);
292 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) {
293 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap);
294 RHSBO = isReassociableOp(RHS, Opcode);
300 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As
301 // such, just remember these operands and their rank.
302 Ops.push_back(ValueEntry(getRank(LHS), LHS));
303 Ops.push_back(ValueEntry(getRank(RHS), RHS));
305 // Clear the leaves out.
306 I->setOperand(0, UndefValue::get(I->getType()));
307 I->setOperand(1, UndefValue::get(I->getType()));
311 // Turn X+(Y+Z) -> (Y+Z)+X
312 std::swap(LHSBO, RHSBO);
314 bool Success = !I->swapOperands();
315 assert(Success && "swapOperands failed");
319 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the RHS is not
320 // part of the expression tree.
322 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0));
323 RHS = I->getOperand(1);
327 // Okay, now we know that the LHS is a nested expression and that the RHS is
328 // not. Perform reassociation.
329 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!");
331 // Move LHS right before I to make sure that the tree expression dominates all
333 LHSBO->moveBefore(I);
335 // Linearize the expression tree on the LHS.
336 LinearizeExprTree(LHSBO, Ops);
338 // Remember the RHS operand and its rank.
339 Ops.push_back(ValueEntry(getRank(RHS), RHS));
341 // Clear the RHS leaf out.
342 I->setOperand(1, UndefValue::get(I->getType()));
345 // RewriteExprTree - Now that the operands for this expression tree are
346 // linearized and optimized, emit them in-order. This function is written to be
348 void Reassociate::RewriteExprTree(BinaryOperator *I,
349 SmallVectorImpl<ValueEntry> &Ops,
351 // If this operation was representing debug info of a value then it
352 // is no longer true, so remove the dbg.value instrinsic.
353 if (DbgValueInst *DVI = DbgValues.lookup(I))
354 DeadInsts.push_back(DVI);
356 if (i+2 == Ops.size()) {
357 if (I->getOperand(0) != Ops[i].Op ||
358 I->getOperand(1) != Ops[i+1].Op) {
359 Value *OldLHS = I->getOperand(0);
360 DEBUG(dbgs() << "RA: " << *I << '\n');
361 I->setOperand(0, Ops[i].Op);
362 I->setOperand(1, Ops[i+1].Op);
364 // Clear all the optional flags, which may not hold after the
365 // reassociation if the expression involved more than just this operation.
367 I->clearSubclassOptionalData();
369 DEBUG(dbgs() << "TO: " << *I << '\n');
373 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3)
374 // delete the extra, now dead, nodes.
375 RemoveDeadBinaryOp(OldLHS);
379 assert(i+2 < Ops.size() && "Ops index out of range!");
381 if (I->getOperand(1) != Ops[i].Op) {
382 DEBUG(dbgs() << "RA: " << *I << '\n');
383 I->setOperand(1, Ops[i].Op);
385 // Conservatively clear all the optional flags, which may not hold
386 // after the reassociation.
387 I->clearSubclassOptionalData();
389 DEBUG(dbgs() << "TO: " << *I << '\n');
394 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
395 assert(LHS->getOpcode() == I->getOpcode() &&
396 "Improper expression tree!");
398 // Compactify the tree instructions together with each other to guarantee
399 // that the expression tree is dominated by all of Ops.
401 RewriteExprTree(LHS, Ops, i+1);
406 // NegateValue - Insert instructions before the instruction pointed to by BI,
407 // that computes the negative version of the value specified. The negative
408 // version of the value is returned, and BI is left pointing at the instruction
409 // that should be processed next by the reassociation pass.
411 static Value *NegateValue(Value *V, Instruction *BI) {
412 if (Constant *C = dyn_cast<Constant>(V))
413 return ConstantExpr::getNeg(C);
415 // We are trying to expose opportunity for reassociation. One of the things
416 // that we want to do to achieve this is to push a negation as deep into an
417 // expression chain as possible, to expose the add instructions. In practice,
418 // this means that we turn this:
419 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
420 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
421 // the constants. We assume that instcombine will clean up the mess later if
422 // we introduce tons of unnecessary negation instructions.
424 if (Instruction *I = dyn_cast<Instruction>(V))
425 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
426 // Push the negates through the add.
427 I->setOperand(0, NegateValue(I->getOperand(0), BI));
428 I->setOperand(1, NegateValue(I->getOperand(1), BI));
430 // We must move the add instruction here, because the neg instructions do
431 // not dominate the old add instruction in general. By moving it, we are
432 // assured that the neg instructions we just inserted dominate the
433 // instruction we are about to insert after them.
436 I->setName(I->getName()+".neg");
440 // Okay, we need to materialize a negated version of V with an instruction.
441 // Scan the use lists of V to see if we have one already.
442 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
444 if (!BinaryOperator::isNeg(U)) continue;
446 // We found one! Now we have to make sure that the definition dominates
447 // this use. We do this by moving it to the entry block (if it is a
448 // non-instruction value) or right after the definition. These negates will
449 // be zapped by reassociate later, so we don't need much finesse here.
450 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
452 // Verify that the negate is in this function, V might be a constant expr.
453 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
456 BasicBlock::iterator InsertPt;
457 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
458 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
459 InsertPt = II->getNormalDest()->begin();
461 InsertPt = InstInput;
464 while (isa<PHINode>(InsertPt)) ++InsertPt;
466 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
468 TheNeg->moveBefore(InsertPt);
472 // Insert a 'neg' instruction that subtracts the value from zero to get the
474 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
477 /// ShouldBreakUpSubtract - Return true if we should break up this subtract of
478 /// X-Y into (X + -Y).
479 static bool ShouldBreakUpSubtract(Instruction *Sub) {
480 // If this is a negation, we can't split it up!
481 if (BinaryOperator::isNeg(Sub))
484 // Don't bother to break this up unless either the LHS is an associable add or
485 // subtract or if this is only used by one.
486 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) ||
487 isReassociableOp(Sub->getOperand(0), Instruction::Sub))
489 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) ||
490 isReassociableOp(Sub->getOperand(1), Instruction::Sub))
492 if (Sub->hasOneUse() &&
493 (isReassociableOp(Sub->use_back(), Instruction::Add) ||
494 isReassociableOp(Sub->use_back(), Instruction::Sub)))
500 /// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is
501 /// only used by an add, transform this into (X+(0-Y)) to promote better
503 static Instruction *BreakUpSubtract(Instruction *Sub,
504 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) {
505 // Convert a subtract into an add and a neg instruction. This allows sub
506 // instructions to be commuted with other add instructions.
508 // Calculate the negative value of Operand 1 of the sub instruction,
509 // and set it as the RHS of the add instruction we just made.
511 Value *NegVal = NegateValue(Sub->getOperand(1), Sub);
513 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub);
516 // Everyone now refers to the add instruction.
517 ValueRankMap.erase(Sub);
518 Sub->replaceAllUsesWith(New);
519 New->setDebugLoc(Sub->getDebugLoc());
520 Sub->eraseFromParent();
522 DEBUG(dbgs() << "Negated: " << *New << '\n');
526 /// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used
527 /// by one, change this into a multiply by a constant to assist with further
529 static Instruction *ConvertShiftToMul(Instruction *Shl,
530 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) {
531 // If an operand of this shift is a reassociable multiply, or if the shift
532 // is used by a reassociable multiply or add, turn into a multiply.
533 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) ||
535 (isReassociableOp(Shl->use_back(), Instruction::Mul) ||
536 isReassociableOp(Shl->use_back(), Instruction::Add)))) {
537 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
538 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
541 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
542 ValueRankMap.erase(Shl);
544 Shl->replaceAllUsesWith(Mul);
545 Mul->setDebugLoc(Shl->getDebugLoc());
546 Shl->eraseFromParent();
552 // Scan backwards and forwards among values with the same rank as element i to
553 // see if X exists. If X does not exist, return i. This is useful when
554 // scanning for 'x' when we see '-x' because they both get the same rank.
555 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
557 unsigned XRank = Ops[i].Rank;
558 unsigned e = Ops.size();
559 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j)
563 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j)
569 /// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together
570 /// and returning the result. Insert the tree before I.
571 static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl<Value*> &Ops){
572 if (Ops.size() == 1) return Ops.back();
574 Value *V1 = Ops.back();
576 Value *V2 = EmitAddTreeOfValues(I, Ops);
577 return BinaryOperator::CreateAdd(V2, V1, "tmp", I);
580 /// RemoveFactorFromExpression - If V is an expression tree that is a
581 /// multiplication sequence, and if this sequence contains a multiply by Factor,
582 /// remove Factor from the tree and return the new tree.
583 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) {
584 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul);
587 SmallVector<ValueEntry, 8> Factors;
588 LinearizeExprTree(BO, Factors);
590 bool FoundFactor = false;
591 bool NeedsNegate = false;
592 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
593 if (Factors[i].Op == Factor) {
595 Factors.erase(Factors.begin()+i);
599 // If this is a negative version of this factor, remove it.
600 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor))
601 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
602 if (FC1->getValue() == -FC2->getValue()) {
603 FoundFactor = NeedsNegate = true;
604 Factors.erase(Factors.begin()+i);
610 // Make sure to restore the operands to the expression tree.
611 RewriteExprTree(BO, Factors);
615 BasicBlock::iterator InsertPt = BO; ++InsertPt;
617 // If this was just a single multiply, remove the multiply and return the only
618 // remaining operand.
619 if (Factors.size() == 1) {
620 ValueRankMap.erase(BO);
621 DeadInsts.push_back(BO);
624 RewriteExprTree(BO, Factors);
629 V = BinaryOperator::CreateNeg(V, "neg", InsertPt);
634 /// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively
635 /// add its operands as factors, otherwise add V to the list of factors.
637 /// Ops is the top-level list of add operands we're trying to factor.
638 static void FindSingleUseMultiplyFactors(Value *V,
639 SmallVectorImpl<Value*> &Factors,
640 const SmallVectorImpl<ValueEntry> &Ops,
643 if (!(V->hasOneUse() || V->use_empty()) || // More than one use.
644 !(BO = dyn_cast<BinaryOperator>(V)) ||
645 BO->getOpcode() != Instruction::Mul) {
646 Factors.push_back(V);
650 // If this value has a single use because it is another input to the add
651 // tree we're reassociating and we dropped its use, it actually has two
652 // uses and we can't factor it.
654 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
655 if (Ops[i].Op == V) {
656 Factors.push_back(V);
662 // Otherwise, add the LHS and RHS to the list of factors.
663 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops, false);
664 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops, false);
667 /// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor'
668 /// instruction. This optimizes based on identities. If it can be reduced to
669 /// a single Value, it is returned, otherwise the Ops list is mutated as
671 static Value *OptimizeAndOrXor(unsigned Opcode,
672 SmallVectorImpl<ValueEntry> &Ops) {
673 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
674 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
675 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
676 // First, check for X and ~X in the operand list.
677 assert(i < Ops.size());
678 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
679 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
680 unsigned FoundX = FindInOperandList(Ops, i, X);
682 if (Opcode == Instruction::And) // ...&X&~X = 0
683 return Constant::getNullValue(X->getType());
685 if (Opcode == Instruction::Or) // ...|X|~X = -1
686 return Constant::getAllOnesValue(X->getType());
690 // Next, check for duplicate pairs of values, which we assume are next to
691 // each other, due to our sorting criteria.
692 assert(i < Ops.size());
693 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
694 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
695 // Drop duplicate values for And and Or.
696 Ops.erase(Ops.begin()+i);
702 // Drop pairs of values for Xor.
703 assert(Opcode == Instruction::Xor);
705 return Constant::getNullValue(Ops[0].Op->getType());
708 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
716 /// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This
717 /// optimizes based on identities. If it can be reduced to a single Value, it
718 /// is returned, otherwise the Ops list is mutated as necessary.
719 Value *Reassociate::OptimizeAdd(Instruction *I,
720 SmallVectorImpl<ValueEntry> &Ops) {
721 // Scan the operand lists looking for X and -X pairs. If we find any, we
722 // can simplify the expression. X+-X == 0. While we're at it, scan for any
723 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
725 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1".
727 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
728 Value *TheOp = Ops[i].Op;
729 // Check to see if we've seen this operand before. If so, we factor all
730 // instances of the operand together. Due to our sorting criteria, we know
731 // that these need to be next to each other in the vector.
732 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
733 // Rescan the list, remove all instances of this operand from the expr.
734 unsigned NumFound = 0;
736 Ops.erase(Ops.begin()+i);
738 } while (i != Ops.size() && Ops[i].Op == TheOp);
740 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
743 // Insert a new multiply.
744 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound);
745 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I);
747 // Now that we have inserted a multiply, optimize it. This allows us to
748 // handle cases that require multiple factoring steps, such as this:
749 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
750 RedoInsts.push_back(Mul);
752 // If every add operand was a duplicate, return the multiply.
756 // Otherwise, we had some input that didn't have the dupe, such as
757 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
758 // things being added by this operation.
759 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
766 // Check for X and -X in the operand list.
767 if (!BinaryOperator::isNeg(TheOp))
770 Value *X = BinaryOperator::getNegArgument(TheOp);
771 unsigned FoundX = FindInOperandList(Ops, i, X);
775 // Remove X and -X from the operand list.
777 return Constant::getNullValue(X->getType());
779 Ops.erase(Ops.begin()+i);
783 --i; // Need to back up an extra one.
784 Ops.erase(Ops.begin()+FoundX);
786 --i; // Revisit element.
787 e -= 2; // Removed two elements.
790 // Scan the operand list, checking to see if there are any common factors
791 // between operands. Consider something like A*A+A*B*C+D. We would like to
792 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
793 // To efficiently find this, we count the number of times a factor occurs
794 // for any ADD operands that are MULs.
795 DenseMap<Value*, unsigned> FactorOccurrences;
797 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
798 // where they are actually the same multiply.
800 Value *MaxOccVal = 0;
801 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
802 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op);
803 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty())
806 // Compute all of the factors of this added value.
807 SmallVector<Value*, 8> Factors;
808 FindSingleUseMultiplyFactors(BOp, Factors, Ops, true);
809 assert(Factors.size() > 1 && "Bad linearize!");
811 // Add one to FactorOccurrences for each unique factor in this op.
812 SmallPtrSet<Value*, 8> Duplicates;
813 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
814 Value *Factor = Factors[i];
815 if (!Duplicates.insert(Factor)) continue;
817 unsigned Occ = ++FactorOccurrences[Factor];
818 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
820 // If Factor is a negative constant, add the negated value as a factor
821 // because we can percolate the negate out. Watch for minint, which
822 // cannot be positivified.
823 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor))
824 if (CI->isNegative() && !CI->isMinValue(true)) {
825 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
826 assert(!Duplicates.count(Factor) &&
827 "Shouldn't have two constant factors, missed a canonicalize");
829 unsigned Occ = ++FactorOccurrences[Factor];
830 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; }
835 // If any factor occurred more than one time, we can pull it out.
837 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
840 // Create a new instruction that uses the MaxOccVal twice. If we don't do
841 // this, we could otherwise run into situations where removing a factor
842 // from an expression will drop a use of maxocc, and this can cause
843 // RemoveFactorFromExpression on successive values to behave differently.
844 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal);
845 SmallVector<Value*, 4> NewMulOps;
846 for (unsigned i = 0; i != Ops.size(); ++i) {
847 // Only try to remove factors from expressions we're allowed to.
848 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op);
849 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty())
852 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
853 // The factorized operand may occur several times. Convert them all in
855 for (unsigned j = Ops.size(); j != i;) {
857 if (Ops[j].Op == Ops[i].Op) {
858 NewMulOps.push_back(V);
859 Ops.erase(Ops.begin()+j);
866 // No need for extra uses anymore.
869 unsigned NumAddedValues = NewMulOps.size();
870 Value *V = EmitAddTreeOfValues(I, NewMulOps);
872 // Now that we have inserted the add tree, optimize it. This allows us to
873 // handle cases that require multiple factoring steps, such as this:
874 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
875 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
876 (void)NumAddedValues;
877 V = ReassociateExpression(cast<BinaryOperator>(V));
879 // Create the multiply.
880 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I);
882 // Rerun associate on the multiply in case the inner expression turned into
883 // a multiply. We want to make sure that we keep things in canonical form.
884 V2 = ReassociateExpression(cast<BinaryOperator>(V2));
886 // If every add operand included the factor (e.g. "A*B + A*C"), then the
887 // entire result expression is just the multiply "A*(B+C)".
891 // Otherwise, we had some input that didn't have the factor, such as
892 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
893 // things being added by this operation.
894 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
900 Value *Reassociate::OptimizeExpression(BinaryOperator *I,
901 SmallVectorImpl<ValueEntry> &Ops) {
902 // Now that we have the linearized expression tree, try to optimize it.
903 // Start by folding any constants that we found.
904 bool IterateOptimization = false;
905 if (Ops.size() == 1) return Ops[0].Op;
907 unsigned Opcode = I->getOpcode();
909 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op))
910 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) {
912 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2);
913 return OptimizeExpression(I, Ops);
916 // Check for destructive annihilation due to a constant being used.
917 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op))
920 case Instruction::And:
921 if (CstVal->isZero()) // X & 0 -> 0
923 if (CstVal->isAllOnesValue()) // X & -1 -> X
926 case Instruction::Mul:
927 if (CstVal->isZero()) { // X * 0 -> 0
932 if (cast<ConstantInt>(CstVal)->isOne())
933 Ops.pop_back(); // X * 1 -> X
935 case Instruction::Or:
936 if (CstVal->isAllOnesValue()) // X | -1 -> -1
939 case Instruction::Add:
940 case Instruction::Xor:
941 if (CstVal->isZero()) // X [|^+] 0 -> X
945 if (Ops.size() == 1) return Ops[0].Op;
947 // Handle destructive annihilation due to identities between elements in the
948 // argument list here.
951 case Instruction::And:
952 case Instruction::Or:
953 case Instruction::Xor: {
954 unsigned NumOps = Ops.size();
955 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
957 IterateOptimization |= Ops.size() != NumOps;
961 case Instruction::Add: {
962 unsigned NumOps = Ops.size();
963 if (Value *Result = OptimizeAdd(I, Ops))
965 IterateOptimization |= Ops.size() != NumOps;
969 //case Instruction::Mul:
972 if (IterateOptimization)
973 return OptimizeExpression(I, Ops);
978 /// ReassociateInst - Inspect and reassociate the instruction at the
979 /// given position, post-incrementing the position.
980 void Reassociate::ReassociateInst(BasicBlock::iterator &BBI) {
981 Instruction *BI = BBI++;
982 if (BI->getOpcode() == Instruction::Shl &&
983 isa<ConstantInt>(BI->getOperand(1)))
984 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) {
989 // Reject cases where it is pointless to do this.
990 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPointTy() ||
991 BI->getType()->isVectorTy())
992 return; // Floating point ops are not associative.
994 // Do not reassociate boolean (i1) expressions. We want to preserve the
995 // original order of evaluation for short-circuited comparisons that
996 // SimplifyCFG has folded to AND/OR expressions. If the expression
997 // is not further optimized, it is likely to be transformed back to a
998 // short-circuited form for code gen, and the source order may have been
999 // optimized for the most likely conditions.
1000 if (BI->getType()->isIntegerTy(1))
1003 // If this is a subtract instruction which is not already in negate form,
1004 // see if we can convert it to X+-Y.
1005 if (BI->getOpcode() == Instruction::Sub) {
1006 if (ShouldBreakUpSubtract(BI)) {
1007 BI = BreakUpSubtract(BI, ValueRankMap);
1008 // Reset the BBI iterator in case BreakUpSubtract changed the
1009 // instruction it points to.
1013 } else if (BinaryOperator::isNeg(BI)) {
1014 // Otherwise, this is a negation. See if the operand is a multiply tree
1015 // and if this is not an inner node of a multiply tree.
1016 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) &&
1017 (!BI->hasOneUse() ||
1018 !isReassociableOp(BI->use_back(), Instruction::Mul))) {
1019 BI = LowerNegateToMultiply(BI, ValueRankMap);
1025 // If this instruction is a commutative binary operator, process it.
1026 if (!BI->isAssociative()) return;
1027 BinaryOperator *I = cast<BinaryOperator>(BI);
1029 // If this is an interior node of a reassociable tree, ignore it until we
1030 // get to the root of the tree, to avoid N^2 analysis.
1031 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode()))
1034 // If this is an add tree that is used by a sub instruction, ignore it
1035 // until we process the subtract.
1036 if (I->hasOneUse() && I->getOpcode() == Instruction::Add &&
1037 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub)
1040 ReassociateExpression(I);
1043 Value *Reassociate::ReassociateExpression(BinaryOperator *I) {
1045 // First, walk the expression tree, linearizing the tree, collecting the
1046 // operand information.
1047 SmallVector<ValueEntry, 8> Ops;
1048 LinearizeExprTree(I, Ops);
1050 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
1052 // Now that we have linearized the tree to a list and have gathered all of
1053 // the operands and their ranks, sort the operands by their rank. Use a
1054 // stable_sort so that values with equal ranks will have their relative
1055 // positions maintained (and so the compiler is deterministic). Note that
1056 // this sorts so that the highest ranking values end up at the beginning of
1058 std::stable_sort(Ops.begin(), Ops.end());
1060 // OptimizeExpression - Now that we have the expression tree in a convenient
1061 // sorted form, optimize it globally if possible.
1062 if (Value *V = OptimizeExpression(I, Ops)) {
1063 // This expression tree simplified to something that isn't a tree,
1065 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
1066 I->replaceAllUsesWith(V);
1067 if (Instruction *VI = dyn_cast<Instruction>(V))
1068 VI->setDebugLoc(I->getDebugLoc());
1069 RemoveDeadBinaryOp(I);
1074 // We want to sink immediates as deeply as possible except in the case where
1075 // this is a multiply tree used only by an add, and the immediate is a -1.
1076 // In this case we reassociate to put the negation on the outside so that we
1077 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
1078 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() &&
1079 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add &&
1080 isa<ConstantInt>(Ops.back().Op) &&
1081 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
1082 ValueEntry Tmp = Ops.pop_back_val();
1083 Ops.insert(Ops.begin(), Tmp);
1086 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
1088 if (Ops.size() == 1) {
1089 // This expression tree simplified to something that isn't a tree,
1091 I->replaceAllUsesWith(Ops[0].Op);
1092 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
1093 OI->setDebugLoc(I->getDebugLoc());
1094 RemoveDeadBinaryOp(I);
1098 // Now that we ordered and optimized the expressions, splat them back into
1099 // the expression tree, removing any unneeded nodes.
1100 RewriteExprTree(I, Ops);
1105 bool Reassociate::runOnFunction(Function &F) {
1106 collectDbgValues(F);
1107 // Recalculate the rank map for F
1111 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
1112 for (BasicBlock::iterator BBI = FI->begin(); BBI != FI->end(); )
1113 ReassociateInst(BBI);
1115 // Now that we're done, revisit any instructions which are likely to
1116 // have secondary reassociation opportunities.
1117 while (!RedoInsts.empty())
1118 if (Value *V = RedoInsts.pop_back_val()) {
1119 BasicBlock::iterator BBI = cast<Instruction>(V);
1120 ReassociateInst(BBI);
1123 // Now that we're done, delete any instructions which are no longer used.
1124 while (!DeadInsts.empty())
1125 if (Value *V = DeadInsts.pop_back_val())
1126 if (!RecursivelyDeleteTriviallyDeadInstructions(V))
1127 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(V))
1128 DVI->eraseFromParent();
1130 // We are done with the rank map.
1132 ValueRankMap.clear();
1137 /// collectDbgValues - Collect all llvm.dbg.value intrinsics.
1138 void Reassociate::collectDbgValues(Function &F) {
1139 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
1140 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
1142 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI))
1143 DbgValues[DVI->getValue()] = DVI;