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 #include "llvm/Transforms/Scalar/Reassociate.h"
24 #include "llvm/ADT/DenseMap.h"
25 #include "llvm/ADT/PostOrderIterator.h"
26 #include "llvm/ADT/STLExtras.h"
27 #include "llvm/ADT/SetVector.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/Analysis/GlobalsModRef.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/CFG.h"
32 #include "llvm/IR/Constants.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/IR/IRBuilder.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/PatternMatch.h"
39 #include "llvm/IR/ValueHandle.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Transforms/Scalar.h"
44 #include "llvm/Transforms/Utils/Local.h"
47 using namespace reassociate;
49 #define DEBUG_TYPE "reassociate"
51 STATISTIC(NumChanged, "Number of insts reassociated");
52 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
53 STATISTIC(NumFactor , "Number of multiplies factored");
56 /// Print out the expression identified in the Ops list.
58 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
59 Module *M = I->getModule();
60 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
61 << *Ops[0].Op->getType() << '\t';
62 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
64 Ops[i].Op->printAsOperand(dbgs(), false, M);
65 dbgs() << ", #" << Ops[i].Rank << "] ";
70 /// Utility class representing a non-constant Xor-operand. We classify
71 /// non-constant Xor-Operands into two categories:
72 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0
74 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
76 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
77 /// operand as "E | 0"
78 class llvm::reassociate::XorOpnd {
82 bool isInvalid() const { return SymbolicPart == nullptr; }
83 bool isOrExpr() const { return isOr; }
84 Value *getValue() const { return OrigVal; }
85 Value *getSymbolicPart() const { return SymbolicPart; }
86 unsigned getSymbolicRank() const { return SymbolicRank; }
87 const APInt &getConstPart() const { return ConstPart; }
89 void Invalidate() { SymbolicPart = OrigVal = nullptr; }
90 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
96 unsigned SymbolicRank;
100 XorOpnd::XorOpnd(Value *V) {
101 assert(!isa<ConstantInt>(V) && "No ConstantInt");
103 Instruction *I = dyn_cast<Instruction>(V);
106 if (I && (I->getOpcode() == Instruction::Or ||
107 I->getOpcode() == Instruction::And)) {
108 Value *V0 = I->getOperand(0);
109 Value *V1 = I->getOperand(1);
111 if (match(V0, PatternMatch::m_APInt(C)))
114 if (match(V1, PatternMatch::m_APInt(C))) {
117 isOr = (I->getOpcode() == Instruction::Or);
122 // view the operand as "V | 0"
124 ConstPart = APInt::getNullValue(V->getType()->getScalarSizeInBits());
128 /// Return true if V is an instruction of the specified opcode and if it
129 /// only has one use.
130 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
131 if (V->hasOneUse() && isa<Instruction>(V) &&
132 cast<Instruction>(V)->getOpcode() == Opcode &&
133 (!isa<FPMathOperator>(V) ||
134 cast<Instruction>(V)->hasUnsafeAlgebra()))
135 return cast<BinaryOperator>(V);
139 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
141 if (V->hasOneUse() && isa<Instruction>(V) &&
142 (cast<Instruction>(V)->getOpcode() == Opcode1 ||
143 cast<Instruction>(V)->getOpcode() == Opcode2) &&
144 (!isa<FPMathOperator>(V) ||
145 cast<Instruction>(V)->hasUnsafeAlgebra()))
146 return cast<BinaryOperator>(V);
150 void ReassociatePass::BuildRankMap(Function &F,
151 ReversePostOrderTraversal<Function*> &RPOT) {
154 // Assign distinct ranks to function arguments.
155 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
156 ValueRankMap[&*I] = ++i;
157 DEBUG(dbgs() << "Calculated Rank[" << I->getName() << "] = " << i << "\n");
160 // Traverse basic blocks in ReversePostOrder
161 for (BasicBlock *BB : RPOT) {
162 unsigned BBRank = RankMap[BB] = ++i << 16;
164 // Walk the basic block, adding precomputed ranks for any instructions that
165 // we cannot move. This ensures that the ranks for these instructions are
166 // all different in the block.
167 for (Instruction &I : *BB)
168 if (mayBeMemoryDependent(I))
169 ValueRankMap[&I] = ++BBRank;
173 unsigned ReassociatePass::getRank(Value *V) {
174 Instruction *I = dyn_cast<Instruction>(V);
176 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
177 return 0; // Otherwise it's a global or constant, rank 0.
180 if (unsigned Rank = ValueRankMap[I])
181 return Rank; // Rank already known?
183 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
184 // we can reassociate expressions for code motion! Since we do not recurse
185 // for PHI nodes, we cannot have infinite recursion here, because there
186 // cannot be loops in the value graph that do not go through PHI nodes.
187 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
188 for (unsigned i = 0, e = I->getNumOperands();
189 i != e && Rank != MaxRank; ++i)
190 Rank = std::max(Rank, getRank(I->getOperand(i)));
192 // If this is a not or neg instruction, do not count it for rank. This
193 // assures us that X and ~X will have the same rank.
194 if (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
195 !BinaryOperator::isFNeg(I))
198 DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
200 return ValueRankMap[I] = Rank;
203 // Canonicalize constants to RHS. Otherwise, sort the operands by rank.
204 void ReassociatePass::canonicalizeOperands(Instruction *I) {
205 assert(isa<BinaryOperator>(I) && "Expected binary operator.");
206 assert(I->isCommutative() && "Expected commutative operator.");
208 Value *LHS = I->getOperand(0);
209 Value *RHS = I->getOperand(1);
210 unsigned LHSRank = getRank(LHS);
211 unsigned RHSRank = getRank(RHS);
213 if (isa<Constant>(RHS))
216 if (isa<Constant>(LHS) || RHSRank < LHSRank)
217 cast<BinaryOperator>(I)->swapOperands();
220 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
221 Instruction *InsertBefore, Value *FlagsOp) {
222 if (S1->getType()->isIntOrIntVectorTy())
223 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
225 BinaryOperator *Res =
226 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
227 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
232 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
233 Instruction *InsertBefore, Value *FlagsOp) {
234 if (S1->getType()->isIntOrIntVectorTy())
235 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
237 BinaryOperator *Res =
238 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
239 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
244 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
245 Instruction *InsertBefore, Value *FlagsOp) {
246 if (S1->getType()->isIntOrIntVectorTy())
247 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
249 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
250 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
255 /// Replace 0-X with X*-1.
256 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
257 Type *Ty = Neg->getType();
258 Constant *NegOne = Ty->isIntOrIntVectorTy() ?
259 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
261 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
262 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
264 Neg->replaceAllUsesWith(Res);
265 Res->setDebugLoc(Neg->getDebugLoc());
269 /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
270 /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for
271 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
272 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
273 /// even x in Bitwidth-bit arithmetic.
274 static unsigned CarmichaelShift(unsigned Bitwidth) {
280 /// Add the extra weight 'RHS' to the existing weight 'LHS',
281 /// reducing the combined weight using any special properties of the operation.
282 /// The existing weight LHS represents the computation X op X op ... op X where
283 /// X occurs LHS times. The combined weight represents X op X op ... op X with
284 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
285 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
286 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
287 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
288 // If we were working with infinite precision arithmetic then the combined
289 // weight would be LHS + RHS. But we are using finite precision arithmetic,
290 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
291 // for nilpotent operations and addition, but not for idempotent operations
292 // and multiplication), so it is important to correctly reduce the combined
293 // weight back into range if wrapping would be wrong.
295 // If RHS is zero then the weight didn't change.
296 if (RHS.isMinValue())
298 // If LHS is zero then the combined weight is RHS.
299 if (LHS.isMinValue()) {
303 // From this point on we know that neither LHS nor RHS is zero.
305 if (Instruction::isIdempotent(Opcode)) {
306 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
307 // weight of 1. Keeping weights at zero or one also means that wrapping is
309 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
310 return; // Return a weight of 1.
312 if (Instruction::isNilpotent(Opcode)) {
313 // Nilpotent means X op X === 0, so reduce weights modulo 2.
314 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
315 LHS = 0; // 1 + 1 === 0 modulo 2.
318 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
319 // TODO: Reduce the weight by exploiting nsw/nuw?
324 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
325 "Unknown associative operation!");
326 unsigned Bitwidth = LHS.getBitWidth();
327 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
328 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
329 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
330 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
331 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
332 // which by a happy accident means that they can always be represented using
334 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
335 // the Carmichael number).
337 /// CM - The value of Carmichael's lambda function.
338 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
339 // Any weight W >= Threshold can be replaced with W - CM.
340 APInt Threshold = CM + Bitwidth;
341 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
342 // For Bitwidth 4 or more the following sum does not overflow.
344 while (LHS.uge(Threshold))
347 // To avoid problems with overflow do everything the same as above but using
349 unsigned CM = 1U << CarmichaelShift(Bitwidth);
350 unsigned Threshold = CM + Bitwidth;
351 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
352 "Weights not reduced!");
353 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
354 while (Total >= Threshold)
360 typedef std::pair<Value*, APInt> RepeatedValue;
362 /// Given an associative binary expression, return the leaf
363 /// nodes in Ops along with their weights (how many times the leaf occurs). The
364 /// original expression is the same as
365 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
367 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
371 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
373 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
375 /// This routine may modify the function, in which case it returns 'true'. The
376 /// changes it makes may well be destructive, changing the value computed by 'I'
377 /// to something completely different. Thus if the routine returns 'true' then
378 /// you MUST either replace I with a new expression computed from the Ops array,
379 /// or use RewriteExprTree to put the values back in.
381 /// A leaf node is either not a binary operation of the same kind as the root
382 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
383 /// opcode), or is the same kind of binary operator but has a use which either
384 /// does not belong to the expression, or does belong to the expression but is
385 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
386 /// of the expression, while for non-leaf nodes (except for the root 'I') every
387 /// use is a non-leaf node of the expression.
390 /// expression graph node names
400 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
401 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
403 /// The expression is maximal: if some instruction is a binary operator of the
404 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
405 /// then the instruction also belongs to the expression, is not a leaf node of
406 /// it, and its operands also belong to the expression (but may be leaf nodes).
408 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
409 /// order to ensure that every non-root node in the expression has *exactly one*
410 /// use by a non-leaf node of the expression. This destruction means that the
411 /// caller MUST either replace 'I' with a new expression or use something like
412 /// RewriteExprTree to put the values back in if the routine indicates that it
413 /// made a change by returning 'true'.
415 /// In the above example either the right operand of A or the left operand of B
416 /// will be replaced by undef. If it is B's operand then this gives:
420 /// + + | A, B - operand of B replaced with undef
426 /// Note that such undef operands can only be reached by passing through 'I'.
427 /// For example, if you visit operands recursively starting from a leaf node
428 /// then you will never see such an undef operand unless you get back to 'I',
429 /// which requires passing through a phi node.
431 /// Note that this routine may also mutate binary operators of the wrong type
432 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
433 /// of the expression) if it can turn them into binary operators of the right
434 /// type and thus make the expression bigger.
436 static bool LinearizeExprTree(BinaryOperator *I,
437 SmallVectorImpl<RepeatedValue> &Ops) {
438 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
439 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
440 unsigned Opcode = I->getOpcode();
441 assert(I->isAssociative() && I->isCommutative() &&
442 "Expected an associative and commutative operation!");
444 // Visit all operands of the expression, keeping track of their weight (the
445 // number of paths from the expression root to the operand, or if you like
446 // the number of times that operand occurs in the linearized expression).
447 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
448 // while A has weight two.
450 // Worklist of non-leaf nodes (their operands are in the expression too) along
451 // with their weights, representing a certain number of paths to the operator.
452 // If an operator occurs in the worklist multiple times then we found multiple
453 // ways to get to it.
454 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
455 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
456 bool Changed = false;
458 // Leaves of the expression are values that either aren't the right kind of
459 // operation (eg: a constant, or a multiply in an add tree), or are, but have
460 // some uses that are not inside the expression. For example, in I = X + X,
461 // X = A + B, the value X has two uses (by I) that are in the expression. If
462 // X has any other uses, for example in a return instruction, then we consider
463 // X to be a leaf, and won't analyze it further. When we first visit a value,
464 // if it has more than one use then at first we conservatively consider it to
465 // be a leaf. Later, as the expression is explored, we may discover some more
466 // uses of the value from inside the expression. If all uses turn out to be
467 // from within the expression (and the value is a binary operator of the right
468 // kind) then the value is no longer considered to be a leaf, and its operands
471 // Leaves - Keeps track of the set of putative leaves as well as the number of
472 // paths to each leaf seen so far.
473 typedef DenseMap<Value*, APInt> LeafMap;
474 LeafMap Leaves; // Leaf -> Total weight so far.
475 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
478 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
480 while (!Worklist.empty()) {
481 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
482 I = P.first; // We examine the operands of this binary operator.
484 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
485 Value *Op = I->getOperand(OpIdx);
486 APInt Weight = P.second; // Number of paths to this operand.
487 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
488 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
490 // If this is a binary operation of the right kind with only one use then
491 // add its operands to the expression.
492 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
493 assert(Visited.insert(Op).second && "Not first visit!");
494 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
495 Worklist.push_back(std::make_pair(BO, Weight));
499 // Appears to be a leaf. Is the operand already in the set of leaves?
500 LeafMap::iterator It = Leaves.find(Op);
501 if (It == Leaves.end()) {
502 // Not in the leaf map. Must be the first time we saw this operand.
503 assert(Visited.insert(Op).second && "Not first visit!");
504 if (!Op->hasOneUse()) {
505 // This value has uses not accounted for by the expression, so it is
506 // not safe to modify. Mark it as being a leaf.
507 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
508 LeafOrder.push_back(Op);
512 // No uses outside the expression, try morphing it.
514 // Already in the leaf map.
515 assert(It != Leaves.end() && Visited.count(Op) &&
516 "In leaf map but not visited!");
518 // Update the number of paths to the leaf.
519 IncorporateWeight(It->second, Weight, Opcode);
521 #if 0 // TODO: Re-enable once PR13021 is fixed.
522 // The leaf already has one use from inside the expression. As we want
523 // exactly one such use, drop this new use of the leaf.
524 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
525 I->setOperand(OpIdx, UndefValue::get(I->getType()));
528 // If the leaf is a binary operation of the right kind and we now see
529 // that its multiple original uses were in fact all by nodes belonging
530 // to the expression, then no longer consider it to be a leaf and add
531 // its operands to the expression.
532 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
533 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
534 Worklist.push_back(std::make_pair(BO, It->second));
540 // If we still have uses that are not accounted for by the expression
541 // then it is not safe to modify the value.
542 if (!Op->hasOneUse())
545 // No uses outside the expression, try morphing it.
547 Leaves.erase(It); // Since the value may be morphed below.
550 // At this point we have a value which, first of all, is not a binary
551 // expression of the right kind, and secondly, is only used inside the
552 // expression. This means that it can safely be modified. See if we
553 // can usefully morph it into an expression of the right kind.
554 assert((!isa<Instruction>(Op) ||
555 cast<Instruction>(Op)->getOpcode() != Opcode
556 || (isa<FPMathOperator>(Op) &&
557 !cast<Instruction>(Op)->hasUnsafeAlgebra())) &&
558 "Should have been handled above!");
559 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
561 // If this is a multiply expression, turn any internal negations into
562 // multiplies by -1 so they can be reassociated.
563 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
564 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
565 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
566 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
567 BO = LowerNegateToMultiply(BO);
568 DEBUG(dbgs() << *BO << '\n');
569 Worklist.push_back(std::make_pair(BO, Weight));
574 // Failed to morph into an expression of the right type. This really is
576 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
577 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
578 LeafOrder.push_back(Op);
583 // The leaves, repeated according to their weights, represent the linearized
584 // form of the expression.
585 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
586 Value *V = LeafOrder[i];
587 LeafMap::iterator It = Leaves.find(V);
588 if (It == Leaves.end())
589 // Node initially thought to be a leaf wasn't.
591 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
592 APInt Weight = It->second;
593 if (Weight.isMinValue())
594 // Leaf already output or weight reduction eliminated it.
596 // Ensure the leaf is only output once.
598 Ops.push_back(std::make_pair(V, Weight));
601 // For nilpotent operations or addition there may be no operands, for example
602 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
603 // in both cases the weight reduces to 0 causing the value to be skipped.
605 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
606 assert(Identity && "Associative operation without identity!");
607 Ops.emplace_back(Identity, APInt(Bitwidth, 1));
613 /// Now that the operands for this expression tree are
614 /// linearized and optimized, emit them in-order.
615 void ReassociatePass::RewriteExprTree(BinaryOperator *I,
616 SmallVectorImpl<ValueEntry> &Ops) {
617 assert(Ops.size() > 1 && "Single values should be used directly!");
619 // Since our optimizations should never increase the number of operations, the
620 // new expression can usually be written reusing the existing binary operators
621 // from the original expression tree, without creating any new instructions,
622 // though the rewritten expression may have a completely different topology.
623 // We take care to not change anything if the new expression will be the same
624 // as the original. If more than trivial changes (like commuting operands)
625 // were made then we are obliged to clear out any optional subclass data like
628 /// NodesToRewrite - Nodes from the original expression available for writing
629 /// the new expression into.
630 SmallVector<BinaryOperator*, 8> NodesToRewrite;
631 unsigned Opcode = I->getOpcode();
632 BinaryOperator *Op = I;
634 /// NotRewritable - The operands being written will be the leaves of the new
635 /// expression and must not be used as inner nodes (via NodesToRewrite) by
636 /// mistake. Inner nodes are always reassociable, and usually leaves are not
637 /// (if they were they would have been incorporated into the expression and so
638 /// would not be leaves), so most of the time there is no danger of this. But
639 /// in rare cases a leaf may become reassociable if an optimization kills uses
640 /// of it, or it may momentarily become reassociable during rewriting (below)
641 /// due it being removed as an operand of one of its uses. Ensure that misuse
642 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
643 /// leaves and refusing to reuse any of them as inner nodes.
644 SmallPtrSet<Value*, 8> NotRewritable;
645 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
646 NotRewritable.insert(Ops[i].Op);
648 // ExpressionChanged - Non-null if the rewritten expression differs from the
649 // original in some non-trivial way, requiring the clearing of optional flags.
650 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
651 BinaryOperator *ExpressionChanged = nullptr;
652 for (unsigned i = 0; ; ++i) {
653 // The last operation (which comes earliest in the IR) is special as both
654 // operands will come from Ops, rather than just one with the other being
656 if (i+2 == Ops.size()) {
657 Value *NewLHS = Ops[i].Op;
658 Value *NewRHS = Ops[i+1].Op;
659 Value *OldLHS = Op->getOperand(0);
660 Value *OldRHS = Op->getOperand(1);
662 if (NewLHS == OldLHS && NewRHS == OldRHS)
663 // Nothing changed, leave it alone.
666 if (NewLHS == OldRHS && NewRHS == OldLHS) {
667 // The order of the operands was reversed. Swap them.
668 DEBUG(dbgs() << "RA: " << *Op << '\n');
670 DEBUG(dbgs() << "TO: " << *Op << '\n');
676 // The new operation differs non-trivially from the original. Overwrite
677 // the old operands with the new ones.
678 DEBUG(dbgs() << "RA: " << *Op << '\n');
679 if (NewLHS != OldLHS) {
680 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
681 if (BO && !NotRewritable.count(BO))
682 NodesToRewrite.push_back(BO);
683 Op->setOperand(0, NewLHS);
685 if (NewRHS != OldRHS) {
686 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
687 if (BO && !NotRewritable.count(BO))
688 NodesToRewrite.push_back(BO);
689 Op->setOperand(1, NewRHS);
691 DEBUG(dbgs() << "TO: " << *Op << '\n');
693 ExpressionChanged = Op;
700 // Not the last operation. The left-hand side will be a sub-expression
701 // while the right-hand side will be the current element of Ops.
702 Value *NewRHS = Ops[i].Op;
703 if (NewRHS != Op->getOperand(1)) {
704 DEBUG(dbgs() << "RA: " << *Op << '\n');
705 if (NewRHS == Op->getOperand(0)) {
706 // The new right-hand side was already present as the left operand. If
707 // we are lucky then swapping the operands will sort out both of them.
710 // Overwrite with the new right-hand side.
711 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
712 if (BO && !NotRewritable.count(BO))
713 NodesToRewrite.push_back(BO);
714 Op->setOperand(1, NewRHS);
715 ExpressionChanged = Op;
717 DEBUG(dbgs() << "TO: " << *Op << '\n');
722 // Now deal with the left-hand side. If this is already an operation node
723 // from the original expression then just rewrite the rest of the expression
725 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
726 if (BO && !NotRewritable.count(BO)) {
731 // Otherwise, grab a spare node from the original expression and use that as
732 // the left-hand side. If there are no nodes left then the optimizers made
733 // an expression with more nodes than the original! This usually means that
734 // they did something stupid but it might mean that the problem was just too
735 // hard (finding the mimimal number of multiplications needed to realize a
736 // multiplication expression is NP-complete). Whatever the reason, smart or
737 // stupid, create a new node if there are none left.
738 BinaryOperator *NewOp;
739 if (NodesToRewrite.empty()) {
740 Constant *Undef = UndefValue::get(I->getType());
741 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
742 Undef, Undef, "", I);
743 if (NewOp->getType()->isFPOrFPVectorTy())
744 NewOp->setFastMathFlags(I->getFastMathFlags());
746 NewOp = NodesToRewrite.pop_back_val();
749 DEBUG(dbgs() << "RA: " << *Op << '\n');
750 Op->setOperand(0, NewOp);
751 DEBUG(dbgs() << "TO: " << *Op << '\n');
752 ExpressionChanged = Op;
758 // If the expression changed non-trivially then clear out all subclass data
759 // starting from the operator specified in ExpressionChanged, and compactify
760 // the operators to just before the expression root to guarantee that the
761 // expression tree is dominated by all of Ops.
762 if (ExpressionChanged)
764 // Preserve FastMathFlags.
765 if (isa<FPMathOperator>(I)) {
766 FastMathFlags Flags = I->getFastMathFlags();
767 ExpressionChanged->clearSubclassOptionalData();
768 ExpressionChanged->setFastMathFlags(Flags);
770 ExpressionChanged->clearSubclassOptionalData();
772 if (ExpressionChanged == I)
774 ExpressionChanged->moveBefore(I);
775 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
778 // Throw away any left over nodes from the original expression.
779 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
780 RedoInsts.insert(NodesToRewrite[i]);
783 /// Insert instructions before the instruction pointed to by BI,
784 /// that computes the negative version of the value specified. The negative
785 /// version of the value is returned, and BI is left pointing at the instruction
786 /// that should be processed next by the reassociation pass.
787 /// Also add intermediate instructions to the redo list that are modified while
788 /// pushing the negates through adds. These will be revisited to see if
789 /// additional opportunities have been exposed.
790 static Value *NegateValue(Value *V, Instruction *BI,
791 SetVector<AssertingVH<Instruction>> &ToRedo) {
792 if (Constant *C = dyn_cast<Constant>(V)) {
793 if (C->getType()->isFPOrFPVectorTy()) {
794 return ConstantExpr::getFNeg(C);
796 return ConstantExpr::getNeg(C);
800 // We are trying to expose opportunity for reassociation. One of the things
801 // that we want to do to achieve this is to push a negation as deep into an
802 // expression chain as possible, to expose the add instructions. In practice,
803 // this means that we turn this:
804 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
805 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
806 // the constants. We assume that instcombine will clean up the mess later if
807 // we introduce tons of unnecessary negation instructions.
809 if (BinaryOperator *I =
810 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
811 // Push the negates through the add.
812 I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo));
813 I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo));
814 if (I->getOpcode() == Instruction::Add) {
815 I->setHasNoUnsignedWrap(false);
816 I->setHasNoSignedWrap(false);
819 // We must move the add instruction here, because the neg instructions do
820 // not dominate the old add instruction in general. By moving it, we are
821 // assured that the neg instructions we just inserted dominate the
822 // instruction we are about to insert after them.
825 I->setName(I->getName()+".neg");
827 // Add the intermediate negates to the redo list as processing them later
828 // could expose more reassociating opportunities.
833 // Okay, we need to materialize a negated version of V with an instruction.
834 // Scan the use lists of V to see if we have one already.
835 for (User *U : V->users()) {
836 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
839 // We found one! Now we have to make sure that the definition dominates
840 // this use. We do this by moving it to the entry block (if it is a
841 // non-instruction value) or right after the definition. These negates will
842 // be zapped by reassociate later, so we don't need much finesse here.
843 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
845 // Verify that the negate is in this function, V might be a constant expr.
846 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
849 BasicBlock::iterator InsertPt;
850 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
851 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
852 InsertPt = II->getNormalDest()->begin();
854 InsertPt = ++InstInput->getIterator();
856 while (isa<PHINode>(InsertPt)) ++InsertPt;
858 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
860 TheNeg->moveBefore(&*InsertPt);
861 if (TheNeg->getOpcode() == Instruction::Sub) {
862 TheNeg->setHasNoUnsignedWrap(false);
863 TheNeg->setHasNoSignedWrap(false);
865 TheNeg->andIRFlags(BI);
867 ToRedo.insert(TheNeg);
871 // Insert a 'neg' instruction that subtracts the value from zero to get the
873 BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI);
874 ToRedo.insert(NewNeg);
878 /// Return true if we should break up this subtract of X-Y into (X + -Y).
879 static bool ShouldBreakUpSubtract(Instruction *Sub) {
880 // If this is a negation, we can't split it up!
881 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
884 // Don't breakup X - undef.
885 if (isa<UndefValue>(Sub->getOperand(1)))
888 // Don't bother to break this up unless either the LHS is an associable add or
889 // subtract or if this is only used by one.
890 Value *V0 = Sub->getOperand(0);
891 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
892 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
894 Value *V1 = Sub->getOperand(1);
895 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
896 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
898 Value *VB = Sub->user_back();
899 if (Sub->hasOneUse() &&
900 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
901 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
907 /// If we have (X-Y), and if either X is an add, or if this is only used by an
908 /// add, transform this into (X+(0-Y)) to promote better reassociation.
909 static BinaryOperator *
910 BreakUpSubtract(Instruction *Sub, SetVector<AssertingVH<Instruction>> &ToRedo) {
911 // Convert a subtract into an add and a neg instruction. This allows sub
912 // instructions to be commuted with other add instructions.
914 // Calculate the negative value of Operand 1 of the sub instruction,
915 // and set it as the RHS of the add instruction we just made.
917 Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo);
918 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
919 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
920 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
923 // Everyone now refers to the add instruction.
924 Sub->replaceAllUsesWith(New);
925 New->setDebugLoc(Sub->getDebugLoc());
927 DEBUG(dbgs() << "Negated: " << *New << '\n');
931 /// If this is a shift of a reassociable multiply or is used by one, change
932 /// this into a multiply by a constant to assist with further reassociation.
933 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
934 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
935 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
937 BinaryOperator *Mul =
938 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
939 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
942 // Everyone now refers to the mul instruction.
943 Shl->replaceAllUsesWith(Mul);
944 Mul->setDebugLoc(Shl->getDebugLoc());
946 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
947 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
949 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
950 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
952 Mul->setHasNoSignedWrap(true);
953 Mul->setHasNoUnsignedWrap(NUW);
957 /// Scan backwards and forwards among values with the same rank as element i
958 /// to see if X exists. If X does not exist, return i. This is useful when
959 /// scanning for 'x' when we see '-x' because they both get the same rank.
960 static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops,
961 unsigned i, Value *X) {
962 unsigned XRank = Ops[i].Rank;
963 unsigned e = Ops.size();
964 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
967 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
968 if (Instruction *I2 = dyn_cast<Instruction>(X))
969 if (I1->isIdenticalTo(I2))
973 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
976 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
977 if (Instruction *I2 = dyn_cast<Instruction>(X))
978 if (I1->isIdenticalTo(I2))
984 /// Emit a tree of add instructions, summing Ops together
985 /// and returning the result. Insert the tree before I.
986 static Value *EmitAddTreeOfValues(Instruction *I,
987 SmallVectorImpl<WeakTrackingVH> &Ops) {
988 if (Ops.size() == 1) return Ops.back();
990 Value *V1 = Ops.back();
992 Value *V2 = EmitAddTreeOfValues(I, Ops);
993 return CreateAdd(V2, V1, "tmp", I, I);
996 /// If V is an expression tree that is a multiplication sequence,
997 /// and if this sequence contains a multiply by Factor,
998 /// remove Factor from the tree and return the new tree.
999 Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) {
1000 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1004 SmallVector<RepeatedValue, 8> Tree;
1005 MadeChange |= LinearizeExprTree(BO, Tree);
1006 SmallVector<ValueEntry, 8> Factors;
1007 Factors.reserve(Tree.size());
1008 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1009 RepeatedValue E = Tree[i];
1010 Factors.append(E.second.getZExtValue(),
1011 ValueEntry(getRank(E.first), E.first));
1014 bool FoundFactor = false;
1015 bool NeedsNegate = false;
1016 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1017 if (Factors[i].Op == Factor) {
1019 Factors.erase(Factors.begin()+i);
1023 // If this is a negative version of this factor, remove it.
1024 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1025 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1026 if (FC1->getValue() == -FC2->getValue()) {
1027 FoundFactor = NeedsNegate = true;
1028 Factors.erase(Factors.begin()+i);
1031 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1032 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1033 const APFloat &F1 = FC1->getValueAPF();
1034 APFloat F2(FC2->getValueAPF());
1036 if (F1.compare(F2) == APFloat::cmpEqual) {
1037 FoundFactor = NeedsNegate = true;
1038 Factors.erase(Factors.begin() + i);
1046 // Make sure to restore the operands to the expression tree.
1047 RewriteExprTree(BO, Factors);
1051 BasicBlock::iterator InsertPt = ++BO->getIterator();
1053 // If this was just a single multiply, remove the multiply and return the only
1054 // remaining operand.
1055 if (Factors.size() == 1) {
1056 RedoInsts.insert(BO);
1059 RewriteExprTree(BO, Factors);
1064 V = CreateNeg(V, "neg", &*InsertPt, BO);
1069 /// If V is a single-use multiply, recursively add its operands as factors,
1070 /// otherwise add V to the list of factors.
1072 /// Ops is the top-level list of add operands we're trying to factor.
1073 static void FindSingleUseMultiplyFactors(Value *V,
1074 SmallVectorImpl<Value*> &Factors) {
1075 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1077 Factors.push_back(V);
1081 // Otherwise, add the LHS and RHS to the list of factors.
1082 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
1083 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
1086 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1087 /// This optimizes based on identities. If it can be reduced to a single Value,
1088 /// it is returned, otherwise the Ops list is mutated as necessary.
1089 static Value *OptimizeAndOrXor(unsigned Opcode,
1090 SmallVectorImpl<ValueEntry> &Ops) {
1091 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1092 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1093 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1094 // First, check for X and ~X in the operand list.
1095 assert(i < Ops.size());
1096 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1097 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1098 unsigned FoundX = FindInOperandList(Ops, i, X);
1100 if (Opcode == Instruction::And) // ...&X&~X = 0
1101 return Constant::getNullValue(X->getType());
1103 if (Opcode == Instruction::Or) // ...|X|~X = -1
1104 return Constant::getAllOnesValue(X->getType());
1108 // Next, check for duplicate pairs of values, which we assume are next to
1109 // each other, due to our sorting criteria.
1110 assert(i < Ops.size());
1111 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1112 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1113 // Drop duplicate values for And and Or.
1114 Ops.erase(Ops.begin()+i);
1120 // Drop pairs of values for Xor.
1121 assert(Opcode == Instruction::Xor);
1123 return Constant::getNullValue(Ops[0].Op->getType());
1126 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1134 /// Helper function of CombineXorOpnd(). It creates a bitwise-and
1135 /// instruction with the given two operands, and return the resulting
1136 /// instruction. There are two special cases: 1) if the constant operand is 0,
1137 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1139 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1140 const APInt &ConstOpnd) {
1141 if (ConstOpnd.isNullValue())
1144 if (ConstOpnd.isAllOnesValue())
1147 Instruction *I = BinaryOperator::CreateAnd(
1148 Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra",
1150 I->setDebugLoc(InsertBefore->getDebugLoc());
1154 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1155 // into "R ^ C", where C would be 0, and R is a symbolic value.
1157 // If it was successful, true is returned, and the "R" and "C" is returned
1158 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1159 // and both "Res" and "ConstOpnd" remain unchanged.
1161 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1162 APInt &ConstOpnd, Value *&Res) {
1163 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1164 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1165 // = (x & ~c1) ^ (c1 ^ c2)
1166 // It is useful only when c1 == c2.
1167 if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isNullValue())
1170 if (!Opnd1->getValue()->hasOneUse())
1173 const APInt &C1 = Opnd1->getConstPart();
1174 if (C1 != ConstOpnd)
1177 Value *X = Opnd1->getSymbolicPart();
1178 Res = createAndInstr(I, X, ~C1);
1179 // ConstOpnd was C2, now C1 ^ C2.
1182 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1183 RedoInsts.insert(T);
1188 // Helper function of OptimizeXor(). It tries to simplify
1189 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1192 // If it was successful, true is returned, and the "R" and "C" is returned
1193 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1194 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1195 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1196 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1197 XorOpnd *Opnd2, APInt &ConstOpnd,
1199 Value *X = Opnd1->getSymbolicPart();
1200 if (X != Opnd2->getSymbolicPart())
1203 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1204 int DeadInstNum = 1;
1205 if (Opnd1->getValue()->hasOneUse())
1207 if (Opnd2->getValue()->hasOneUse())
1211 // (x | c1) ^ (x & c2)
1212 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1213 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1214 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1216 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1217 if (Opnd2->isOrExpr())
1218 std::swap(Opnd1, Opnd2);
1220 const APInt &C1 = Opnd1->getConstPart();
1221 const APInt &C2 = Opnd2->getConstPart();
1222 APInt C3((~C1) ^ C2);
1224 // Do not increase code size!
1225 if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1226 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1227 if (NewInstNum > DeadInstNum)
1231 Res = createAndInstr(I, X, C3);
1234 } else if (Opnd1->isOrExpr()) {
1235 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1237 const APInt &C1 = Opnd1->getConstPart();
1238 const APInt &C2 = Opnd2->getConstPart();
1241 // Do not increase code size
1242 if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1243 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1244 if (NewInstNum > DeadInstNum)
1248 Res = createAndInstr(I, X, C3);
1251 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1253 const APInt &C1 = Opnd1->getConstPart();
1254 const APInt &C2 = Opnd2->getConstPart();
1256 Res = createAndInstr(I, X, C3);
1259 // Put the original operands in the Redo list; hope they will be deleted
1261 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1262 RedoInsts.insert(T);
1263 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1264 RedoInsts.insert(T);
1269 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1270 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1272 Value *ReassociatePass::OptimizeXor(Instruction *I,
1273 SmallVectorImpl<ValueEntry> &Ops) {
1274 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1277 if (Ops.size() == 1)
1280 SmallVector<XorOpnd, 8> Opnds;
1281 SmallVector<XorOpnd*, 8> OpndPtrs;
1282 Type *Ty = Ops[0].Op->getType();
1283 APInt ConstOpnd(Ty->getScalarSizeInBits(), 0);
1285 // Step 1: Convert ValueEntry to XorOpnd
1286 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1287 Value *V = Ops[i].Op;
1289 // TODO: Support non-splat vectors.
1290 if (match(V, PatternMatch::m_APInt(C))) {
1294 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1299 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1300 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1301 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1302 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1303 // when new elements are added to the vector.
1304 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1305 OpndPtrs.push_back(&Opnds[i]);
1307 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1308 // the same symbolic value cluster together. For instance, the input operand
1309 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1310 // ("x | 123", "x & 789", "y & 456").
1312 // The purpose is twofold:
1313 // 1) Cluster together the operands sharing the same symbolic-value.
1314 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
1315 // could potentially shorten crital path, and expose more loop-invariants.
1316 // Note that values' rank are basically defined in RPO order (FIXME).
1317 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1318 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1319 // "z" in the order of X-Y-Z is better than any other orders.
1320 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(),
1321 [](XorOpnd *LHS, XorOpnd *RHS) {
1322 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1325 // Step 3: Combine adjacent operands
1326 XorOpnd *PrevOpnd = nullptr;
1327 bool Changed = false;
1328 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1329 XorOpnd *CurrOpnd = OpndPtrs[i];
1330 // The combined value
1333 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1334 if (!ConstOpnd.isNullValue() &&
1335 CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1338 *CurrOpnd = XorOpnd(CV);
1340 CurrOpnd->Invalidate();
1345 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1346 PrevOpnd = CurrOpnd;
1350 // step 3.2: When previous and current operands share the same symbolic
1351 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1353 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1354 // Remove previous operand
1355 PrevOpnd->Invalidate();
1357 *CurrOpnd = XorOpnd(CV);
1358 PrevOpnd = CurrOpnd;
1360 CurrOpnd->Invalidate();
1367 // Step 4: Reassemble the Ops
1370 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1371 XorOpnd &O = Opnds[i];
1374 ValueEntry VE(getRank(O.getValue()), O.getValue());
1377 if (!ConstOpnd.isNullValue()) {
1378 Value *C = ConstantInt::get(Ty, ConstOpnd);
1379 ValueEntry VE(getRank(C), C);
1382 unsigned Sz = Ops.size();
1384 return Ops.back().Op;
1386 assert(ConstOpnd.isNullValue());
1387 return ConstantInt::get(Ty, ConstOpnd);
1394 /// Optimize a series of operands to an 'add' instruction. This
1395 /// optimizes based on identities. If it can be reduced to a single Value, it
1396 /// is returned, otherwise the Ops list is mutated as necessary.
1397 Value *ReassociatePass::OptimizeAdd(Instruction *I,
1398 SmallVectorImpl<ValueEntry> &Ops) {
1399 // Scan the operand lists looking for X and -X pairs. If we find any, we
1400 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1402 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1404 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1405 Value *TheOp = Ops[i].Op;
1406 // Check to see if we've seen this operand before. If so, we factor all
1407 // instances of the operand together. Due to our sorting criteria, we know
1408 // that these need to be next to each other in the vector.
1409 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1410 // Rescan the list, remove all instances of this operand from the expr.
1411 unsigned NumFound = 0;
1413 Ops.erase(Ops.begin()+i);
1415 } while (i != Ops.size() && Ops[i].Op == TheOp);
1417 DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1420 // Insert a new multiply.
1421 Type *Ty = TheOp->getType();
1422 Constant *C = Ty->isIntOrIntVectorTy() ?
1423 ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1424 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1426 // Now that we have inserted a multiply, optimize it. This allows us to
1427 // handle cases that require multiple factoring steps, such as this:
1428 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1429 RedoInsts.insert(Mul);
1431 // If every add operand was a duplicate, return the multiply.
1435 // Otherwise, we had some input that didn't have the dupe, such as
1436 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1437 // things being added by this operation.
1438 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1445 // Check for X and -X or X and ~X in the operand list.
1446 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1447 !BinaryOperator::isNot(TheOp))
1451 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1452 X = BinaryOperator::getNegArgument(TheOp);
1453 else if (BinaryOperator::isNot(TheOp))
1454 X = BinaryOperator::getNotArgument(TheOp);
1456 unsigned FoundX = FindInOperandList(Ops, i, X);
1460 // Remove X and -X from the operand list.
1461 if (Ops.size() == 2 &&
1462 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1463 return Constant::getNullValue(X->getType());
1465 // Remove X and ~X from the operand list.
1466 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1467 return Constant::getAllOnesValue(X->getType());
1469 Ops.erase(Ops.begin()+i);
1473 --i; // Need to back up an extra one.
1474 Ops.erase(Ops.begin()+FoundX);
1476 --i; // Revisit element.
1477 e -= 2; // Removed two elements.
1479 // if X and ~X we append -1 to the operand list.
1480 if (BinaryOperator::isNot(TheOp)) {
1481 Value *V = Constant::getAllOnesValue(X->getType());
1482 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1487 // Scan the operand list, checking to see if there are any common factors
1488 // between operands. Consider something like A*A+A*B*C+D. We would like to
1489 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1490 // To efficiently find this, we count the number of times a factor occurs
1491 // for any ADD operands that are MULs.
1492 DenseMap<Value*, unsigned> FactorOccurrences;
1494 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1495 // where they are actually the same multiply.
1496 unsigned MaxOcc = 0;
1497 Value *MaxOccVal = nullptr;
1498 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1499 BinaryOperator *BOp =
1500 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1504 // Compute all of the factors of this added value.
1505 SmallVector<Value*, 8> Factors;
1506 FindSingleUseMultiplyFactors(BOp, Factors);
1507 assert(Factors.size() > 1 && "Bad linearize!");
1509 // Add one to FactorOccurrences for each unique factor in this op.
1510 SmallPtrSet<Value*, 8> Duplicates;
1511 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1512 Value *Factor = Factors[i];
1513 if (!Duplicates.insert(Factor).second)
1516 unsigned Occ = ++FactorOccurrences[Factor];
1522 // If Factor is a negative constant, add the negated value as a factor
1523 // because we can percolate the negate out. Watch for minint, which
1524 // cannot be positivified.
1525 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1526 if (CI->isNegative() && !CI->isMinValue(true)) {
1527 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1528 if (!Duplicates.insert(Factor).second)
1530 unsigned Occ = ++FactorOccurrences[Factor];
1536 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1537 if (CF->isNegative()) {
1538 APFloat F(CF->getValueAPF());
1540 Factor = ConstantFP::get(CF->getContext(), F);
1541 if (!Duplicates.insert(Factor).second)
1543 unsigned Occ = ++FactorOccurrences[Factor];
1553 // If any factor occurred more than one time, we can pull it out.
1555 DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1558 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1559 // this, we could otherwise run into situations where removing a factor
1560 // from an expression will drop a use of maxocc, and this can cause
1561 // RemoveFactorFromExpression on successive values to behave differently.
1562 Instruction *DummyInst =
1563 I->getType()->isIntOrIntVectorTy()
1564 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1565 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1567 SmallVector<WeakTrackingVH, 4> NewMulOps;
1568 for (unsigned i = 0; i != Ops.size(); ++i) {
1569 // Only try to remove factors from expressions we're allowed to.
1570 BinaryOperator *BOp =
1571 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1575 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1576 // The factorized operand may occur several times. Convert them all in
1578 for (unsigned j = Ops.size(); j != i;) {
1580 if (Ops[j].Op == Ops[i].Op) {
1581 NewMulOps.push_back(V);
1582 Ops.erase(Ops.begin()+j);
1589 // No need for extra uses anymore.
1590 DummyInst->deleteValue();
1592 unsigned NumAddedValues = NewMulOps.size();
1593 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1595 // Now that we have inserted the add tree, optimize it. This allows us to
1596 // handle cases that require multiple factoring steps, such as this:
1597 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1598 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1599 (void)NumAddedValues;
1600 if (Instruction *VI = dyn_cast<Instruction>(V))
1601 RedoInsts.insert(VI);
1603 // Create the multiply.
1604 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1606 // Rerun associate on the multiply in case the inner expression turned into
1607 // a multiply. We want to make sure that we keep things in canonical form.
1608 RedoInsts.insert(V2);
1610 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1611 // entire result expression is just the multiply "A*(B+C)".
1615 // Otherwise, we had some input that didn't have the factor, such as
1616 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1617 // things being added by this operation.
1618 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1624 /// \brief Build up a vector of value/power pairs factoring a product.
1626 /// Given a series of multiplication operands, build a vector of factors and
1627 /// the powers each is raised to when forming the final product. Sort them in
1628 /// the order of descending power.
1630 /// (x*x) -> [(x, 2)]
1631 /// ((x*x)*x) -> [(x, 3)]
1632 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1634 /// \returns Whether any factors have a power greater than one.
1635 static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1636 SmallVectorImpl<Factor> &Factors) {
1637 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1638 // Compute the sum of powers of simplifiable factors.
1639 unsigned FactorPowerSum = 0;
1640 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1641 Value *Op = Ops[Idx-1].Op;
1643 // Count the number of occurrences of this value.
1645 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1647 // Track for simplification all factors which occur 2 or more times.
1649 FactorPowerSum += Count;
1652 // We can only simplify factors if the sum of the powers of our simplifiable
1653 // factors is 4 or higher. When that is the case, we will *always* have
1654 // a simplification. This is an important invariant to prevent cyclicly
1655 // trying to simplify already minimal formations.
1656 if (FactorPowerSum < 4)
1659 // Now gather the simplifiable factors, removing them from Ops.
1661 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1662 Value *Op = Ops[Idx-1].Op;
1664 // Count the number of occurrences of this value.
1666 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1670 // Move an even number of occurrences to Factors.
1673 FactorPowerSum += Count;
1674 Factors.push_back(Factor(Op, Count));
1675 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1678 // None of the adjustments above should have reduced the sum of factor powers
1679 // below our mininum of '4'.
1680 assert(FactorPowerSum >= 4);
1682 std::stable_sort(Factors.begin(), Factors.end(),
1683 [](const Factor &LHS, const Factor &RHS) {
1684 return LHS.Power > RHS.Power;
1689 /// \brief Build a tree of multiplies, computing the product of Ops.
1690 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1691 SmallVectorImpl<Value*> &Ops) {
1692 if (Ops.size() == 1)
1695 Value *LHS = Ops.pop_back_val();
1697 if (LHS->getType()->isIntOrIntVectorTy())
1698 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1700 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1701 } while (!Ops.empty());
1706 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1708 /// Given a vector of values raised to various powers, where no two values are
1709 /// equal and the powers are sorted in decreasing order, compute the minimal
1710 /// DAG of multiplies to compute the final product, and return that product
1713 ReassociatePass::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1714 SmallVectorImpl<Factor> &Factors) {
1715 assert(Factors[0].Power);
1716 SmallVector<Value *, 4> OuterProduct;
1717 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1718 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1719 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1724 // We want to multiply across all the factors with the same power so that
1725 // we can raise them to that power as a single entity. Build a mini tree
1727 SmallVector<Value *, 4> InnerProduct;
1728 InnerProduct.push_back(Factors[LastIdx].Base);
1730 InnerProduct.push_back(Factors[Idx].Base);
1732 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1734 // Reset the base value of the first factor to the new expression tree.
1735 // We'll remove all the factors with the same power in a second pass.
1736 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1737 if (Instruction *MI = dyn_cast<Instruction>(M))
1738 RedoInsts.insert(MI);
1742 // Unique factors with equal powers -- we've folded them into the first one's
1744 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1745 [](const Factor &LHS, const Factor &RHS) {
1746 return LHS.Power == RHS.Power;
1750 // Iteratively collect the base of each factor with an add power into the
1751 // outer product, and halve each power in preparation for squaring the
1753 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1754 if (Factors[Idx].Power & 1)
1755 OuterProduct.push_back(Factors[Idx].Base);
1756 Factors[Idx].Power >>= 1;
1758 if (Factors[0].Power) {
1759 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1760 OuterProduct.push_back(SquareRoot);
1761 OuterProduct.push_back(SquareRoot);
1763 if (OuterProduct.size() == 1)
1764 return OuterProduct.front();
1766 Value *V = buildMultiplyTree(Builder, OuterProduct);
1770 Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
1771 SmallVectorImpl<ValueEntry> &Ops) {
1772 // We can only optimize the multiplies when there is a chain of more than
1773 // three, such that a balanced tree might require fewer total multiplies.
1777 // Try to turn linear trees of multiplies without other uses of the
1778 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1780 SmallVector<Factor, 4> Factors;
1781 if (!collectMultiplyFactors(Ops, Factors))
1782 return nullptr; // All distinct factors, so nothing left for us to do.
1784 IRBuilder<> Builder(I);
1785 // The reassociate transformation for FP operations is performed only
1786 // if unsafe algebra is permitted by FastMathFlags. Propagate those flags
1787 // to the newly generated operations.
1788 if (auto FPI = dyn_cast<FPMathOperator>(I))
1789 Builder.setFastMathFlags(FPI->getFastMathFlags());
1791 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1795 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1796 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1800 Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
1801 SmallVectorImpl<ValueEntry> &Ops) {
1802 // Now that we have the linearized expression tree, try to optimize it.
1803 // Start by folding any constants that we found.
1804 Constant *Cst = nullptr;
1805 unsigned Opcode = I->getOpcode();
1806 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1807 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1808 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1810 // If there was nothing but constants then we are done.
1814 // Put the combined constant back at the end of the operand list, except if
1815 // there is no point. For example, an add of 0 gets dropped here, while a
1816 // multiplication by zero turns the whole expression into zero.
1817 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1818 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1820 Ops.push_back(ValueEntry(0, Cst));
1823 if (Ops.size() == 1) return Ops[0].Op;
1825 // Handle destructive annihilation due to identities between elements in the
1826 // argument list here.
1827 unsigned NumOps = Ops.size();
1830 case Instruction::And:
1831 case Instruction::Or:
1832 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1836 case Instruction::Xor:
1837 if (Value *Result = OptimizeXor(I, Ops))
1841 case Instruction::Add:
1842 case Instruction::FAdd:
1843 if (Value *Result = OptimizeAdd(I, Ops))
1847 case Instruction::Mul:
1848 case Instruction::FMul:
1849 if (Value *Result = OptimizeMul(I, Ops))
1854 if (Ops.size() != NumOps)
1855 return OptimizeExpression(I, Ops);
1859 // Remove dead instructions and if any operands are trivially dead add them to
1860 // Insts so they will be removed as well.
1861 void ReassociatePass::RecursivelyEraseDeadInsts(
1862 Instruction *I, SetVector<AssertingVH<Instruction>> &Insts) {
1863 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1864 SmallVector<Value *, 4> Ops(I->op_begin(), I->op_end());
1865 ValueRankMap.erase(I);
1867 RedoInsts.remove(I);
1868 I->eraseFromParent();
1870 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1871 if (OpInst->use_empty())
1872 Insts.insert(OpInst);
1875 /// Zap the given instruction, adding interesting operands to the work list.
1876 void ReassociatePass::EraseInst(Instruction *I) {
1877 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1878 DEBUG(dbgs() << "Erasing dead inst: "; I->dump());
1880 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1881 // Erase the dead instruction.
1882 ValueRankMap.erase(I);
1883 RedoInsts.remove(I);
1884 I->eraseFromParent();
1885 // Optimize its operands.
1886 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1887 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1888 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1889 // If this is a node in an expression tree, climb to the expression root
1890 // and add that since that's where optimization actually happens.
1891 unsigned Opcode = Op->getOpcode();
1892 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1893 Visited.insert(Op).second)
1894 Op = Op->user_back();
1895 RedoInsts.insert(Op);
1899 // Canonicalize expressions of the following form:
1900 // x + (-Constant * y) -> x - (Constant * y)
1901 // x - (-Constant * y) -> x + (Constant * y)
1902 Instruction *ReassociatePass::canonicalizeNegConstExpr(Instruction *I) {
1903 if (!I->hasOneUse() || I->getType()->isVectorTy())
1906 // Must be a fmul or fdiv instruction.
1907 unsigned Opcode = I->getOpcode();
1908 if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv)
1911 auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0));
1912 auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1));
1914 // Both operands are constant, let it get constant folded away.
1918 ConstantFP *CF = C0 ? C0 : C1;
1920 // Must have one constant operand.
1924 // Must be a negative ConstantFP.
1925 if (!CF->isNegative())
1928 // User must be a binary operator with one or more uses.
1929 Instruction *User = I->user_back();
1930 if (!isa<BinaryOperator>(User) || User->use_empty())
1933 unsigned UserOpcode = User->getOpcode();
1934 if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub)
1937 // Subtraction is not commutative. Explicitly, the following transform is
1938 // not valid: (-Constant * y) - x -> x + (Constant * y)
1939 if (!User->isCommutative() && User->getOperand(1) != I)
1942 // Change the sign of the constant.
1943 APFloat Val = CF->getValueAPF();
1945 I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val));
1947 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
1948 // ((-Const*y) + x) -> (x + (-Const*y)).
1949 if (User->getOperand(0) == I && User->isCommutative())
1950 cast<BinaryOperator>(User)->swapOperands();
1952 Value *Op0 = User->getOperand(0);
1953 Value *Op1 = User->getOperand(1);
1955 switch (UserOpcode) {
1957 llvm_unreachable("Unexpected Opcode!");
1958 case Instruction::FAdd:
1959 NI = BinaryOperator::CreateFSub(Op0, Op1);
1960 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1962 case Instruction::FSub:
1963 NI = BinaryOperator::CreateFAdd(Op0, Op1);
1964 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1968 NI->insertBefore(User);
1969 NI->setName(User->getName());
1970 User->replaceAllUsesWith(NI);
1971 NI->setDebugLoc(I->getDebugLoc());
1972 RedoInsts.insert(I);
1977 /// Inspect and optimize the given instruction. Note that erasing
1978 /// instructions is not allowed.
1979 void ReassociatePass::OptimizeInst(Instruction *I) {
1980 // Only consider operations that we understand.
1981 if (!isa<BinaryOperator>(I))
1984 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
1985 // If an operand of this shift is a reassociable multiply, or if the shift
1986 // is used by a reassociable multiply or add, turn into a multiply.
1987 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
1989 (isReassociableOp(I->user_back(), Instruction::Mul) ||
1990 isReassociableOp(I->user_back(), Instruction::Add)))) {
1991 Instruction *NI = ConvertShiftToMul(I);
1992 RedoInsts.insert(I);
1997 // Canonicalize negative constants out of expressions.
1998 if (Instruction *Res = canonicalizeNegConstExpr(I))
2001 // Commute binary operators, to canonicalize the order of their operands.
2002 // This can potentially expose more CSE opportunities, and makes writing other
2003 // transformations simpler.
2004 if (I->isCommutative())
2005 canonicalizeOperands(I);
2007 // Don't optimize floating point instructions that don't have unsafe algebra.
2008 if (I->getType()->isFPOrFPVectorTy() && !I->hasUnsafeAlgebra())
2011 // Do not reassociate boolean (i1) expressions. We want to preserve the
2012 // original order of evaluation for short-circuited comparisons that
2013 // SimplifyCFG has folded to AND/OR expressions. If the expression
2014 // is not further optimized, it is likely to be transformed back to a
2015 // short-circuited form for code gen, and the source order may have been
2016 // optimized for the most likely conditions.
2017 if (I->getType()->isIntegerTy(1))
2020 // If this is a subtract instruction which is not already in negate form,
2021 // see if we can convert it to X+-Y.
2022 if (I->getOpcode() == Instruction::Sub) {
2023 if (ShouldBreakUpSubtract(I)) {
2024 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2025 RedoInsts.insert(I);
2028 } else if (BinaryOperator::isNeg(I)) {
2029 // Otherwise, this is a negation. See if the operand is a multiply tree
2030 // and if this is not an inner node of a multiply tree.
2031 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2033 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2034 Instruction *NI = LowerNegateToMultiply(I);
2035 // If the negate was simplified, revisit the users to see if we can
2036 // reassociate further.
2037 for (User *U : NI->users()) {
2038 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2039 RedoInsts.insert(Tmp);
2041 RedoInsts.insert(I);
2046 } else if (I->getOpcode() == Instruction::FSub) {
2047 if (ShouldBreakUpSubtract(I)) {
2048 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2049 RedoInsts.insert(I);
2052 } else if (BinaryOperator::isFNeg(I)) {
2053 // Otherwise, this is a negation. See if the operand is a multiply tree
2054 // and if this is not an inner node of a multiply tree.
2055 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2057 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2058 // If the negate was simplified, revisit the users to see if we can
2059 // reassociate further.
2060 Instruction *NI = LowerNegateToMultiply(I);
2061 for (User *U : NI->users()) {
2062 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2063 RedoInsts.insert(Tmp);
2065 RedoInsts.insert(I);
2072 // If this instruction is an associative binary operator, process it.
2073 if (!I->isAssociative()) return;
2074 BinaryOperator *BO = cast<BinaryOperator>(I);
2076 // If this is an interior node of a reassociable tree, ignore it until we
2077 // get to the root of the tree, to avoid N^2 analysis.
2078 unsigned Opcode = BO->getOpcode();
2079 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2080 // During the initial run we will get to the root of the tree.
2081 // But if we get here while we are redoing instructions, there is no
2082 // guarantee that the root will be visited. So Redo later
2083 if (BO->user_back() != BO &&
2084 BO->getParent() == BO->user_back()->getParent())
2085 RedoInsts.insert(BO->user_back());
2089 // If this is an add tree that is used by a sub instruction, ignore it
2090 // until we process the subtract.
2091 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2092 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2094 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2095 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2098 ReassociateExpression(BO);
2101 void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2102 // First, walk the expression tree, linearizing the tree, collecting the
2103 // operand information.
2104 SmallVector<RepeatedValue, 8> Tree;
2105 MadeChange |= LinearizeExprTree(I, Tree);
2106 SmallVector<ValueEntry, 8> Ops;
2107 Ops.reserve(Tree.size());
2108 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2109 RepeatedValue E = Tree[i];
2110 Ops.append(E.second.getZExtValue(),
2111 ValueEntry(getRank(E.first), E.first));
2114 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2116 // Now that we have linearized the tree to a list and have gathered all of
2117 // the operands and their ranks, sort the operands by their rank. Use a
2118 // stable_sort so that values with equal ranks will have their relative
2119 // positions maintained (and so the compiler is deterministic). Note that
2120 // this sorts so that the highest ranking values end up at the beginning of
2122 std::stable_sort(Ops.begin(), Ops.end());
2124 // Now that we have the expression tree in a convenient
2125 // sorted form, optimize it globally if possible.
2126 if (Value *V = OptimizeExpression(I, Ops)) {
2128 // Self-referential expression in unreachable code.
2130 // This expression tree simplified to something that isn't a tree,
2132 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2133 I->replaceAllUsesWith(V);
2134 if (Instruction *VI = dyn_cast<Instruction>(V))
2135 VI->setDebugLoc(I->getDebugLoc());
2136 RedoInsts.insert(I);
2141 // We want to sink immediates as deeply as possible except in the case where
2142 // this is a multiply tree used only by an add, and the immediate is a -1.
2143 // In this case we reassociate to put the negation on the outside so that we
2144 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2145 if (I->hasOneUse()) {
2146 if (I->getOpcode() == Instruction::Mul &&
2147 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2148 isa<ConstantInt>(Ops.back().Op) &&
2149 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
2150 ValueEntry Tmp = Ops.pop_back_val();
2151 Ops.insert(Ops.begin(), Tmp);
2152 } else if (I->getOpcode() == Instruction::FMul &&
2153 cast<Instruction>(I->user_back())->getOpcode() ==
2154 Instruction::FAdd &&
2155 isa<ConstantFP>(Ops.back().Op) &&
2156 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2157 ValueEntry Tmp = Ops.pop_back_val();
2158 Ops.insert(Ops.begin(), Tmp);
2162 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2164 if (Ops.size() == 1) {
2166 // Self-referential expression in unreachable code.
2169 // This expression tree simplified to something that isn't a tree,
2171 I->replaceAllUsesWith(Ops[0].Op);
2172 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2173 OI->setDebugLoc(I->getDebugLoc());
2174 RedoInsts.insert(I);
2178 // Now that we ordered and optimized the expressions, splat them back into
2179 // the expression tree, removing any unneeded nodes.
2180 RewriteExprTree(I, Ops);
2183 PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) {
2184 // Get the functions basic blocks in Reverse Post Order. This order is used by
2185 // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic
2186 // blocks (it has been seen that the analysis in this pass could hang when
2187 // analysing dead basic blocks).
2188 ReversePostOrderTraversal<Function *> RPOT(&F);
2190 // Calculate the rank map for F.
2191 BuildRankMap(F, RPOT);
2194 // Traverse the same blocks that was analysed by BuildRankMap.
2195 for (BasicBlock *BI : RPOT) {
2196 assert(RankMap.count(&*BI) && "BB should be ranked.");
2197 // Optimize every instruction in the basic block.
2198 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;)
2199 if (isInstructionTriviallyDead(&*II)) {
2203 assert(II->getParent() == &*BI && "Moved to a different block!");
2207 // Make a copy of all the instructions to be redone so we can remove dead
2209 SetVector<AssertingVH<Instruction>> ToRedo(RedoInsts);
2210 // Iterate over all instructions to be reevaluated and remove trivially dead
2211 // instructions. If any operand of the trivially dead instruction becomes
2212 // dead mark it for deletion as well. Continue this process until all
2213 // trivially dead instructions have been removed.
2214 while (!ToRedo.empty()) {
2215 Instruction *I = ToRedo.pop_back_val();
2216 if (isInstructionTriviallyDead(I)) {
2217 RecursivelyEraseDeadInsts(I, ToRedo);
2222 // Now that we have removed dead instructions, we can reoptimize the
2223 // remaining instructions.
2224 while (!RedoInsts.empty()) {
2225 Instruction *I = RedoInsts.pop_back_val();
2226 if (isInstructionTriviallyDead(I))
2233 // We are done with the rank map.
2235 ValueRankMap.clear();
2238 PreservedAnalyses PA;
2239 PA.preserveSet<CFGAnalyses>();
2240 PA.preserve<GlobalsAA>();
2244 return PreservedAnalyses::all();
2248 class ReassociateLegacyPass : public FunctionPass {
2249 ReassociatePass Impl;
2251 static char ID; // Pass identification, replacement for typeid
2252 ReassociateLegacyPass() : FunctionPass(ID) {
2253 initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry());
2256 bool runOnFunction(Function &F) override {
2257 if (skipFunction(F))
2260 FunctionAnalysisManager DummyFAM;
2261 auto PA = Impl.run(F, DummyFAM);
2262 return !PA.areAllPreserved();
2265 void getAnalysisUsage(AnalysisUsage &AU) const override {
2266 AU.setPreservesCFG();
2267 AU.addPreserved<GlobalsAAWrapperPass>();
2272 char ReassociateLegacyPass::ID = 0;
2273 INITIALIZE_PASS(ReassociateLegacyPass, "reassociate",
2274 "Reassociate expressions", false, false)
2276 // Public interface to the Reassociate pass
2277 FunctionPass *llvm::createReassociatePass() {
2278 return new ReassociateLegacyPass();