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/APFloat.h"
25 #include "llvm/ADT/APInt.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/PostOrderIterator.h"
28 #include "llvm/ADT/SetVector.h"
29 #include "llvm/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Analysis/GlobalsModRef.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/IR/Argument.h"
35 #include "llvm/IR/BasicBlock.h"
36 #include "llvm/IR/CFG.h"
37 #include "llvm/IR/Constant.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/IRBuilder.h"
41 #include "llvm/IR/InstrTypes.h"
42 #include "llvm/IR/Instruction.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/Operator.h"
45 #include "llvm/IR/PassManager.h"
46 #include "llvm/IR/PatternMatch.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/IR/User.h"
49 #include "llvm/IR/Value.h"
50 #include "llvm/IR/ValueHandle.h"
51 #include "llvm/Pass.h"
52 #include "llvm/Support/Casting.h"
53 #include "llvm/Support/Debug.h"
54 #include "llvm/Support/ErrorHandling.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Scalar.h"
57 #include "llvm/Transforms/Utils/Local.h"
63 using namespace reassociate;
65 #define DEBUG_TYPE "reassociate"
67 STATISTIC(NumChanged, "Number of insts reassociated");
68 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
69 STATISTIC(NumFactor , "Number of multiplies factored");
72 /// Print out the expression identified in the Ops list.
73 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
74 Module *M = I->getModule();
75 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
76 << *Ops[0].Op->getType() << '\t';
77 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
79 Ops[i].Op->printAsOperand(dbgs(), false, M);
80 dbgs() << ", #" << Ops[i].Rank << "] ";
85 /// Utility class representing a non-constant Xor-operand. We classify
86 /// non-constant Xor-Operands into two categories:
87 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0
89 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
91 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
92 /// operand as "E | 0"
93 class llvm::reassociate::XorOpnd {
97 bool isInvalid() const { return SymbolicPart == nullptr; }
98 bool isOrExpr() const { return isOr; }
99 Value *getValue() const { return OrigVal; }
100 Value *getSymbolicPart() const { return SymbolicPart; }
101 unsigned getSymbolicRank() const { return SymbolicRank; }
102 const APInt &getConstPart() const { return ConstPart; }
104 void Invalidate() { SymbolicPart = OrigVal = nullptr; }
105 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
111 unsigned SymbolicRank;
115 XorOpnd::XorOpnd(Value *V) {
116 assert(!isa<ConstantInt>(V) && "No ConstantInt");
118 Instruction *I = dyn_cast<Instruction>(V);
121 if (I && (I->getOpcode() == Instruction::Or ||
122 I->getOpcode() == Instruction::And)) {
123 Value *V0 = I->getOperand(0);
124 Value *V1 = I->getOperand(1);
126 if (match(V0, PatternMatch::m_APInt(C)))
129 if (match(V1, PatternMatch::m_APInt(C))) {
132 isOr = (I->getOpcode() == Instruction::Or);
137 // view the operand as "V | 0"
139 ConstPart = APInt::getNullValue(V->getType()->getScalarSizeInBits());
143 /// Return true if V is an instruction of the specified opcode and if it
144 /// only has one use.
145 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
146 if (V->hasOneUse() && isa<Instruction>(V) &&
147 cast<Instruction>(V)->getOpcode() == Opcode &&
148 (!isa<FPMathOperator>(V) ||
149 cast<Instruction>(V)->hasUnsafeAlgebra()))
150 return cast<BinaryOperator>(V);
154 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
156 if (V->hasOneUse() && isa<Instruction>(V) &&
157 (cast<Instruction>(V)->getOpcode() == Opcode1 ||
158 cast<Instruction>(V)->getOpcode() == Opcode2) &&
159 (!isa<FPMathOperator>(V) ||
160 cast<Instruction>(V)->hasUnsafeAlgebra()))
161 return cast<BinaryOperator>(V);
165 void ReassociatePass::BuildRankMap(Function &F,
166 ReversePostOrderTraversal<Function*> &RPOT) {
169 // Assign distinct ranks to function arguments.
170 for (auto &Arg : F.args()) {
171 ValueRankMap[&Arg] = ++Rank;
172 DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank
176 // Traverse basic blocks in ReversePostOrder
177 for (BasicBlock *BB : RPOT) {
178 unsigned BBRank = RankMap[BB] = ++Rank << 16;
180 // Walk the basic block, adding precomputed ranks for any instructions that
181 // we cannot move. This ensures that the ranks for these instructions are
182 // all different in the block.
183 for (Instruction &I : *BB)
184 if (mayBeMemoryDependent(I))
185 ValueRankMap[&I] = ++BBRank;
189 unsigned ReassociatePass::getRank(Value *V) {
190 Instruction *I = dyn_cast<Instruction>(V);
192 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
193 return 0; // Otherwise it's a global or constant, rank 0.
196 if (unsigned Rank = ValueRankMap[I])
197 return Rank; // Rank already known?
199 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
200 // we can reassociate expressions for code motion! Since we do not recurse
201 // for PHI nodes, we cannot have infinite recursion here, because there
202 // cannot be loops in the value graph that do not go through PHI nodes.
203 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
204 for (unsigned i = 0, e = I->getNumOperands();
205 i != e && Rank != MaxRank; ++i)
206 Rank = std::max(Rank, getRank(I->getOperand(i)));
208 // If this is a not or neg instruction, do not count it for rank. This
209 // assures us that X and ~X will have the same rank.
210 if (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
211 !BinaryOperator::isFNeg(I))
214 DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
216 return ValueRankMap[I] = Rank;
219 // Canonicalize constants to RHS. Otherwise, sort the operands by rank.
220 void ReassociatePass::canonicalizeOperands(Instruction *I) {
221 assert(isa<BinaryOperator>(I) && "Expected binary operator.");
222 assert(I->isCommutative() && "Expected commutative operator.");
224 Value *LHS = I->getOperand(0);
225 Value *RHS = I->getOperand(1);
226 if (LHS == RHS || isa<Constant>(RHS))
228 if (isa<Constant>(LHS) || getRank(RHS) < getRank(LHS))
229 cast<BinaryOperator>(I)->swapOperands();
232 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
233 Instruction *InsertBefore, Value *FlagsOp) {
234 if (S1->getType()->isIntOrIntVectorTy())
235 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
237 BinaryOperator *Res =
238 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
239 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
244 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
245 Instruction *InsertBefore, Value *FlagsOp) {
246 if (S1->getType()->isIntOrIntVectorTy())
247 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
249 BinaryOperator *Res =
250 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
251 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
256 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
257 Instruction *InsertBefore, Value *FlagsOp) {
258 if (S1->getType()->isIntOrIntVectorTy())
259 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
261 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
262 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
267 /// Replace 0-X with X*-1.
268 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
269 Type *Ty = Neg->getType();
270 Constant *NegOne = Ty->isIntOrIntVectorTy() ?
271 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
273 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
274 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
276 Neg->replaceAllUsesWith(Res);
277 Res->setDebugLoc(Neg->getDebugLoc());
281 /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
282 /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for
283 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
284 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
285 /// even x in Bitwidth-bit arithmetic.
286 static unsigned CarmichaelShift(unsigned Bitwidth) {
292 /// Add the extra weight 'RHS' to the existing weight 'LHS',
293 /// reducing the combined weight using any special properties of the operation.
294 /// The existing weight LHS represents the computation X op X op ... op X where
295 /// X occurs LHS times. The combined weight represents X op X op ... op X with
296 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
297 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
298 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
299 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
300 // If we were working with infinite precision arithmetic then the combined
301 // weight would be LHS + RHS. But we are using finite precision arithmetic,
302 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
303 // for nilpotent operations and addition, but not for idempotent operations
304 // and multiplication), so it is important to correctly reduce the combined
305 // weight back into range if wrapping would be wrong.
307 // If RHS is zero then the weight didn't change.
308 if (RHS.isMinValue())
310 // If LHS is zero then the combined weight is RHS.
311 if (LHS.isMinValue()) {
315 // From this point on we know that neither LHS nor RHS is zero.
317 if (Instruction::isIdempotent(Opcode)) {
318 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
319 // weight of 1. Keeping weights at zero or one also means that wrapping is
321 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
322 return; // Return a weight of 1.
324 if (Instruction::isNilpotent(Opcode)) {
325 // Nilpotent means X op X === 0, so reduce weights modulo 2.
326 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
327 LHS = 0; // 1 + 1 === 0 modulo 2.
330 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
331 // TODO: Reduce the weight by exploiting nsw/nuw?
336 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
337 "Unknown associative operation!");
338 unsigned Bitwidth = LHS.getBitWidth();
339 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
340 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
341 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
342 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
343 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
344 // which by a happy accident means that they can always be represented using
346 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
347 // the Carmichael number).
349 /// CM - The value of Carmichael's lambda function.
350 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
351 // Any weight W >= Threshold can be replaced with W - CM.
352 APInt Threshold = CM + Bitwidth;
353 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
354 // For Bitwidth 4 or more the following sum does not overflow.
356 while (LHS.uge(Threshold))
359 // To avoid problems with overflow do everything the same as above but using
361 unsigned CM = 1U << CarmichaelShift(Bitwidth);
362 unsigned Threshold = CM + Bitwidth;
363 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
364 "Weights not reduced!");
365 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
366 while (Total >= Threshold)
372 using RepeatedValue = std::pair<Value*, APInt>;
374 /// Given an associative binary expression, return the leaf
375 /// nodes in Ops along with their weights (how many times the leaf occurs). The
376 /// original expression is the same as
377 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
379 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
383 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
385 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
387 /// This routine may modify the function, in which case it returns 'true'. The
388 /// changes it makes may well be destructive, changing the value computed by 'I'
389 /// to something completely different. Thus if the routine returns 'true' then
390 /// you MUST either replace I with a new expression computed from the Ops array,
391 /// or use RewriteExprTree to put the values back in.
393 /// A leaf node is either not a binary operation of the same kind as the root
394 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
395 /// opcode), or is the same kind of binary operator but has a use which either
396 /// does not belong to the expression, or does belong to the expression but is
397 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
398 /// of the expression, while for non-leaf nodes (except for the root 'I') every
399 /// use is a non-leaf node of the expression.
402 /// expression graph node names
412 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
413 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
415 /// The expression is maximal: if some instruction is a binary operator of the
416 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
417 /// then the instruction also belongs to the expression, is not a leaf node of
418 /// it, and its operands also belong to the expression (but may be leaf nodes).
420 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
421 /// order to ensure that every non-root node in the expression has *exactly one*
422 /// use by a non-leaf node of the expression. This destruction means that the
423 /// caller MUST either replace 'I' with a new expression or use something like
424 /// RewriteExprTree to put the values back in if the routine indicates that it
425 /// made a change by returning 'true'.
427 /// In the above example either the right operand of A or the left operand of B
428 /// will be replaced by undef. If it is B's operand then this gives:
432 /// + + | A, B - operand of B replaced with undef
438 /// Note that such undef operands can only be reached by passing through 'I'.
439 /// For example, if you visit operands recursively starting from a leaf node
440 /// then you will never see such an undef operand unless you get back to 'I',
441 /// which requires passing through a phi node.
443 /// Note that this routine may also mutate binary operators of the wrong type
444 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
445 /// of the expression) if it can turn them into binary operators of the right
446 /// type and thus make the expression bigger.
447 static bool LinearizeExprTree(BinaryOperator *I,
448 SmallVectorImpl<RepeatedValue> &Ops) {
449 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
450 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
451 unsigned Opcode = I->getOpcode();
452 assert(I->isAssociative() && I->isCommutative() &&
453 "Expected an associative and commutative operation!");
455 // Visit all operands of the expression, keeping track of their weight (the
456 // number of paths from the expression root to the operand, or if you like
457 // the number of times that operand occurs in the linearized expression).
458 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
459 // while A has weight two.
461 // Worklist of non-leaf nodes (their operands are in the expression too) along
462 // with their weights, representing a certain number of paths to the operator.
463 // If an operator occurs in the worklist multiple times then we found multiple
464 // ways to get to it.
465 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
466 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
467 bool Changed = false;
469 // Leaves of the expression are values that either aren't the right kind of
470 // operation (eg: a constant, or a multiply in an add tree), or are, but have
471 // some uses that are not inside the expression. For example, in I = X + X,
472 // X = A + B, the value X has two uses (by I) that are in the expression. If
473 // X has any other uses, for example in a return instruction, then we consider
474 // X to be a leaf, and won't analyze it further. When we first visit a value,
475 // if it has more than one use then at first we conservatively consider it to
476 // be a leaf. Later, as the expression is explored, we may discover some more
477 // uses of the value from inside the expression. If all uses turn out to be
478 // from within the expression (and the value is a binary operator of the right
479 // kind) then the value is no longer considered to be a leaf, and its operands
482 // Leaves - Keeps track of the set of putative leaves as well as the number of
483 // paths to each leaf seen so far.
484 using LeafMap = DenseMap<Value *, APInt>;
485 LeafMap Leaves; // Leaf -> Total weight so far.
486 SmallVector<Value *, 8> LeafOrder; // Ensure deterministic leaf output order.
489 SmallPtrSet<Value *, 8> Visited; // For sanity checking the iteration scheme.
491 while (!Worklist.empty()) {
492 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
493 I = P.first; // We examine the operands of this binary operator.
495 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
496 Value *Op = I->getOperand(OpIdx);
497 APInt Weight = P.second; // Number of paths to this operand.
498 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
499 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
501 // If this is a binary operation of the right kind with only one use then
502 // add its operands to the expression.
503 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
504 assert(Visited.insert(Op).second && "Not first visit!");
505 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
506 Worklist.push_back(std::make_pair(BO, Weight));
510 // Appears to be a leaf. Is the operand already in the set of leaves?
511 LeafMap::iterator It = Leaves.find(Op);
512 if (It == Leaves.end()) {
513 // Not in the leaf map. Must be the first time we saw this operand.
514 assert(Visited.insert(Op).second && "Not first visit!");
515 if (!Op->hasOneUse()) {
516 // This value has uses not accounted for by the expression, so it is
517 // not safe to modify. Mark it as being a leaf.
518 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
519 LeafOrder.push_back(Op);
523 // No uses outside the expression, try morphing it.
525 // Already in the leaf map.
526 assert(It != Leaves.end() && Visited.count(Op) &&
527 "In leaf map but not visited!");
529 // Update the number of paths to the leaf.
530 IncorporateWeight(It->second, Weight, Opcode);
532 #if 0 // TODO: Re-enable once PR13021 is fixed.
533 // The leaf already has one use from inside the expression. As we want
534 // exactly one such use, drop this new use of the leaf.
535 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
536 I->setOperand(OpIdx, UndefValue::get(I->getType()));
539 // If the leaf is a binary operation of the right kind and we now see
540 // that its multiple original uses were in fact all by nodes belonging
541 // to the expression, then no longer consider it to be a leaf and add
542 // its operands to the expression.
543 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
544 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
545 Worklist.push_back(std::make_pair(BO, It->second));
551 // If we still have uses that are not accounted for by the expression
552 // then it is not safe to modify the value.
553 if (!Op->hasOneUse())
556 // No uses outside the expression, try morphing it.
558 Leaves.erase(It); // Since the value may be morphed below.
561 // At this point we have a value which, first of all, is not a binary
562 // expression of the right kind, and secondly, is only used inside the
563 // expression. This means that it can safely be modified. See if we
564 // can usefully morph it into an expression of the right kind.
565 assert((!isa<Instruction>(Op) ||
566 cast<Instruction>(Op)->getOpcode() != Opcode
567 || (isa<FPMathOperator>(Op) &&
568 !cast<Instruction>(Op)->hasUnsafeAlgebra())) &&
569 "Should have been handled above!");
570 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
572 // If this is a multiply expression, turn any internal negations into
573 // multiplies by -1 so they can be reassociated.
574 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
575 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
576 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
577 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
578 BO = LowerNegateToMultiply(BO);
579 DEBUG(dbgs() << *BO << '\n');
580 Worklist.push_back(std::make_pair(BO, Weight));
585 // Failed to morph into an expression of the right type. This really is
587 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
588 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
589 LeafOrder.push_back(Op);
594 // The leaves, repeated according to their weights, represent the linearized
595 // form of the expression.
596 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
597 Value *V = LeafOrder[i];
598 LeafMap::iterator It = Leaves.find(V);
599 if (It == Leaves.end())
600 // Node initially thought to be a leaf wasn't.
602 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
603 APInt Weight = It->second;
604 if (Weight.isMinValue())
605 // Leaf already output or weight reduction eliminated it.
607 // Ensure the leaf is only output once.
609 Ops.push_back(std::make_pair(V, Weight));
612 // For nilpotent operations or addition there may be no operands, for example
613 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
614 // in both cases the weight reduces to 0 causing the value to be skipped.
616 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
617 assert(Identity && "Associative operation without identity!");
618 Ops.emplace_back(Identity, APInt(Bitwidth, 1));
624 /// Now that the operands for this expression tree are
625 /// linearized and optimized, emit them in-order.
626 void ReassociatePass::RewriteExprTree(BinaryOperator *I,
627 SmallVectorImpl<ValueEntry> &Ops) {
628 assert(Ops.size() > 1 && "Single values should be used directly!");
630 // Since our optimizations should never increase the number of operations, the
631 // new expression can usually be written reusing the existing binary operators
632 // from the original expression tree, without creating any new instructions,
633 // though the rewritten expression may have a completely different topology.
634 // We take care to not change anything if the new expression will be the same
635 // as the original. If more than trivial changes (like commuting operands)
636 // were made then we are obliged to clear out any optional subclass data like
639 /// NodesToRewrite - Nodes from the original expression available for writing
640 /// the new expression into.
641 SmallVector<BinaryOperator*, 8> NodesToRewrite;
642 unsigned Opcode = I->getOpcode();
643 BinaryOperator *Op = I;
645 /// NotRewritable - The operands being written will be the leaves of the new
646 /// expression and must not be used as inner nodes (via NodesToRewrite) by
647 /// mistake. Inner nodes are always reassociable, and usually leaves are not
648 /// (if they were they would have been incorporated into the expression and so
649 /// would not be leaves), so most of the time there is no danger of this. But
650 /// in rare cases a leaf may become reassociable if an optimization kills uses
651 /// of it, or it may momentarily become reassociable during rewriting (below)
652 /// due it being removed as an operand of one of its uses. Ensure that misuse
653 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
654 /// leaves and refusing to reuse any of them as inner nodes.
655 SmallPtrSet<Value*, 8> NotRewritable;
656 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
657 NotRewritable.insert(Ops[i].Op);
659 // ExpressionChanged - Non-null if the rewritten expression differs from the
660 // original in some non-trivial way, requiring the clearing of optional flags.
661 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
662 BinaryOperator *ExpressionChanged = nullptr;
663 for (unsigned i = 0; ; ++i) {
664 // The last operation (which comes earliest in the IR) is special as both
665 // operands will come from Ops, rather than just one with the other being
667 if (i+2 == Ops.size()) {
668 Value *NewLHS = Ops[i].Op;
669 Value *NewRHS = Ops[i+1].Op;
670 Value *OldLHS = Op->getOperand(0);
671 Value *OldRHS = Op->getOperand(1);
673 if (NewLHS == OldLHS && NewRHS == OldRHS)
674 // Nothing changed, leave it alone.
677 if (NewLHS == OldRHS && NewRHS == OldLHS) {
678 // The order of the operands was reversed. Swap them.
679 DEBUG(dbgs() << "RA: " << *Op << '\n');
681 DEBUG(dbgs() << "TO: " << *Op << '\n');
687 // The new operation differs non-trivially from the original. Overwrite
688 // the old operands with the new ones.
689 DEBUG(dbgs() << "RA: " << *Op << '\n');
690 if (NewLHS != OldLHS) {
691 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
692 if (BO && !NotRewritable.count(BO))
693 NodesToRewrite.push_back(BO);
694 Op->setOperand(0, NewLHS);
696 if (NewRHS != OldRHS) {
697 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
698 if (BO && !NotRewritable.count(BO))
699 NodesToRewrite.push_back(BO);
700 Op->setOperand(1, NewRHS);
702 DEBUG(dbgs() << "TO: " << *Op << '\n');
704 ExpressionChanged = Op;
711 // Not the last operation. The left-hand side will be a sub-expression
712 // while the right-hand side will be the current element of Ops.
713 Value *NewRHS = Ops[i].Op;
714 if (NewRHS != Op->getOperand(1)) {
715 DEBUG(dbgs() << "RA: " << *Op << '\n');
716 if (NewRHS == Op->getOperand(0)) {
717 // The new right-hand side was already present as the left operand. If
718 // we are lucky then swapping the operands will sort out both of them.
721 // Overwrite with the new right-hand side.
722 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
723 if (BO && !NotRewritable.count(BO))
724 NodesToRewrite.push_back(BO);
725 Op->setOperand(1, NewRHS);
726 ExpressionChanged = Op;
728 DEBUG(dbgs() << "TO: " << *Op << '\n');
733 // Now deal with the left-hand side. If this is already an operation node
734 // from the original expression then just rewrite the rest of the expression
736 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
737 if (BO && !NotRewritable.count(BO)) {
742 // Otherwise, grab a spare node from the original expression and use that as
743 // the left-hand side. If there are no nodes left then the optimizers made
744 // an expression with more nodes than the original! This usually means that
745 // they did something stupid but it might mean that the problem was just too
746 // hard (finding the mimimal number of multiplications needed to realize a
747 // multiplication expression is NP-complete). Whatever the reason, smart or
748 // stupid, create a new node if there are none left.
749 BinaryOperator *NewOp;
750 if (NodesToRewrite.empty()) {
751 Constant *Undef = UndefValue::get(I->getType());
752 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
753 Undef, Undef, "", I);
754 if (NewOp->getType()->isFPOrFPVectorTy())
755 NewOp->setFastMathFlags(I->getFastMathFlags());
757 NewOp = NodesToRewrite.pop_back_val();
760 DEBUG(dbgs() << "RA: " << *Op << '\n');
761 Op->setOperand(0, NewOp);
762 DEBUG(dbgs() << "TO: " << *Op << '\n');
763 ExpressionChanged = Op;
769 // If the expression changed non-trivially then clear out all subclass data
770 // starting from the operator specified in ExpressionChanged, and compactify
771 // the operators to just before the expression root to guarantee that the
772 // expression tree is dominated by all of Ops.
773 if (ExpressionChanged)
775 // Preserve FastMathFlags.
776 if (isa<FPMathOperator>(I)) {
777 FastMathFlags Flags = I->getFastMathFlags();
778 ExpressionChanged->clearSubclassOptionalData();
779 ExpressionChanged->setFastMathFlags(Flags);
781 ExpressionChanged->clearSubclassOptionalData();
783 if (ExpressionChanged == I)
785 ExpressionChanged->moveBefore(I);
786 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
789 // Throw away any left over nodes from the original expression.
790 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
791 RedoInsts.insert(NodesToRewrite[i]);
794 /// Insert instructions before the instruction pointed to by BI,
795 /// that computes the negative version of the value specified. The negative
796 /// version of the value is returned, and BI is left pointing at the instruction
797 /// that should be processed next by the reassociation pass.
798 /// Also add intermediate instructions to the redo list that are modified while
799 /// pushing the negates through adds. These will be revisited to see if
800 /// additional opportunities have been exposed.
801 static Value *NegateValue(Value *V, Instruction *BI,
802 SetVector<AssertingVH<Instruction>> &ToRedo) {
803 if (Constant *C = dyn_cast<Constant>(V)) {
804 if (C->getType()->isFPOrFPVectorTy()) {
805 return ConstantExpr::getFNeg(C);
807 return ConstantExpr::getNeg(C);
810 // We are trying to expose opportunity for reassociation. One of the things
811 // that we want to do to achieve this is to push a negation as deep into an
812 // expression chain as possible, to expose the add instructions. In practice,
813 // this means that we turn this:
814 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
815 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
816 // the constants. We assume that instcombine will clean up the mess later if
817 // we introduce tons of unnecessary negation instructions.
819 if (BinaryOperator *I =
820 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
821 // Push the negates through the add.
822 I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo));
823 I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo));
824 if (I->getOpcode() == Instruction::Add) {
825 I->setHasNoUnsignedWrap(false);
826 I->setHasNoSignedWrap(false);
829 // We must move the add instruction here, because the neg instructions do
830 // not dominate the old add instruction in general. By moving it, we are
831 // assured that the neg instructions we just inserted dominate the
832 // instruction we are about to insert after them.
835 I->setName(I->getName()+".neg");
837 // Add the intermediate negates to the redo list as processing them later
838 // could expose more reassociating opportunities.
843 // Okay, we need to materialize a negated version of V with an instruction.
844 // Scan the use lists of V to see if we have one already.
845 for (User *U : V->users()) {
846 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
849 // We found one! Now we have to make sure that the definition dominates
850 // this use. We do this by moving it to the entry block (if it is a
851 // non-instruction value) or right after the definition. These negates will
852 // be zapped by reassociate later, so we don't need much finesse here.
853 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
855 // Verify that the negate is in this function, V might be a constant expr.
856 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
859 BasicBlock::iterator InsertPt;
860 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
861 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
862 InsertPt = II->getNormalDest()->begin();
864 InsertPt = ++InstInput->getIterator();
866 while (isa<PHINode>(InsertPt)) ++InsertPt;
868 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
870 TheNeg->moveBefore(&*InsertPt);
871 if (TheNeg->getOpcode() == Instruction::Sub) {
872 TheNeg->setHasNoUnsignedWrap(false);
873 TheNeg->setHasNoSignedWrap(false);
875 TheNeg->andIRFlags(BI);
877 ToRedo.insert(TheNeg);
881 // Insert a 'neg' instruction that subtracts the value from zero to get the
883 BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI);
884 ToRedo.insert(NewNeg);
888 /// Return true if we should break up this subtract of X-Y into (X + -Y).
889 static bool ShouldBreakUpSubtract(Instruction *Sub) {
890 // If this is a negation, we can't split it up!
891 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
894 // Don't breakup X - undef.
895 if (isa<UndefValue>(Sub->getOperand(1)))
898 // Don't bother to break this up unless either the LHS is an associable add or
899 // subtract or if this is only used by one.
900 Value *V0 = Sub->getOperand(0);
901 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
902 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
904 Value *V1 = Sub->getOperand(1);
905 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
906 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
908 Value *VB = Sub->user_back();
909 if (Sub->hasOneUse() &&
910 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
911 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
917 /// If we have (X-Y), and if either X is an add, or if this is only used by an
918 /// add, transform this into (X+(0-Y)) to promote better reassociation.
919 static BinaryOperator *
920 BreakUpSubtract(Instruction *Sub, SetVector<AssertingVH<Instruction>> &ToRedo) {
921 // Convert a subtract into an add and a neg instruction. This allows sub
922 // instructions to be commuted with other add instructions.
924 // Calculate the negative value of Operand 1 of the sub instruction,
925 // and set it as the RHS of the add instruction we just made.
926 Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo);
927 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
928 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
929 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
932 // Everyone now refers to the add instruction.
933 Sub->replaceAllUsesWith(New);
934 New->setDebugLoc(Sub->getDebugLoc());
936 DEBUG(dbgs() << "Negated: " << *New << '\n');
940 /// If this is a shift of a reassociable multiply or is used by one, change
941 /// this into a multiply by a constant to assist with further reassociation.
942 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
943 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
944 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
946 BinaryOperator *Mul =
947 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
948 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
951 // Everyone now refers to the mul instruction.
952 Shl->replaceAllUsesWith(Mul);
953 Mul->setDebugLoc(Shl->getDebugLoc());
955 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
956 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
958 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
959 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
961 Mul->setHasNoSignedWrap(true);
962 Mul->setHasNoUnsignedWrap(NUW);
966 /// Scan backwards and forwards among values with the same rank as element i
967 /// to see if X exists. If X does not exist, return i. This is useful when
968 /// scanning for 'x' when we see '-x' because they both get the same rank.
969 static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops,
970 unsigned i, Value *X) {
971 unsigned XRank = Ops[i].Rank;
972 unsigned e = Ops.size();
973 for (unsigned j = i+1; j != e && 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))
982 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
985 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
986 if (Instruction *I2 = dyn_cast<Instruction>(X))
987 if (I1->isIdenticalTo(I2))
993 /// Emit a tree of add instructions, summing Ops together
994 /// and returning the result. Insert the tree before I.
995 static Value *EmitAddTreeOfValues(Instruction *I,
996 SmallVectorImpl<WeakTrackingVH> &Ops) {
997 if (Ops.size() == 1) return Ops.back();
999 Value *V1 = Ops.back();
1001 Value *V2 = EmitAddTreeOfValues(I, Ops);
1002 return CreateAdd(V2, V1, "tmp", I, I);
1005 /// If V is an expression tree that is a multiplication sequence,
1006 /// and if this sequence contains a multiply by Factor,
1007 /// remove Factor from the tree and return the new tree.
1008 Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) {
1009 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1013 SmallVector<RepeatedValue, 8> Tree;
1014 MadeChange |= LinearizeExprTree(BO, Tree);
1015 SmallVector<ValueEntry, 8> Factors;
1016 Factors.reserve(Tree.size());
1017 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1018 RepeatedValue E = Tree[i];
1019 Factors.append(E.second.getZExtValue(),
1020 ValueEntry(getRank(E.first), E.first));
1023 bool FoundFactor = false;
1024 bool NeedsNegate = false;
1025 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1026 if (Factors[i].Op == Factor) {
1028 Factors.erase(Factors.begin()+i);
1032 // If this is a negative version of this factor, remove it.
1033 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1034 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1035 if (FC1->getValue() == -FC2->getValue()) {
1036 FoundFactor = NeedsNegate = true;
1037 Factors.erase(Factors.begin()+i);
1040 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1041 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1042 const APFloat &F1 = FC1->getValueAPF();
1043 APFloat F2(FC2->getValueAPF());
1045 if (F1.compare(F2) == APFloat::cmpEqual) {
1046 FoundFactor = NeedsNegate = true;
1047 Factors.erase(Factors.begin() + i);
1055 // Make sure to restore the operands to the expression tree.
1056 RewriteExprTree(BO, Factors);
1060 BasicBlock::iterator InsertPt = ++BO->getIterator();
1062 // If this was just a single multiply, remove the multiply and return the only
1063 // remaining operand.
1064 if (Factors.size() == 1) {
1065 RedoInsts.insert(BO);
1068 RewriteExprTree(BO, Factors);
1073 V = CreateNeg(V, "neg", &*InsertPt, BO);
1078 /// If V is a single-use multiply, recursively add its operands as factors,
1079 /// otherwise add V to the list of factors.
1081 /// Ops is the top-level list of add operands we're trying to factor.
1082 static void FindSingleUseMultiplyFactors(Value *V,
1083 SmallVectorImpl<Value*> &Factors) {
1084 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1086 Factors.push_back(V);
1090 // Otherwise, add the LHS and RHS to the list of factors.
1091 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
1092 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
1095 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1096 /// This optimizes based on identities. If it can be reduced to a single Value,
1097 /// it is returned, otherwise the Ops list is mutated as necessary.
1098 static Value *OptimizeAndOrXor(unsigned Opcode,
1099 SmallVectorImpl<ValueEntry> &Ops) {
1100 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1101 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1102 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1103 // First, check for X and ~X in the operand list.
1104 assert(i < Ops.size());
1105 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1106 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1107 unsigned FoundX = FindInOperandList(Ops, i, X);
1109 if (Opcode == Instruction::And) // ...&X&~X = 0
1110 return Constant::getNullValue(X->getType());
1112 if (Opcode == Instruction::Or) // ...|X|~X = -1
1113 return Constant::getAllOnesValue(X->getType());
1117 // Next, check for duplicate pairs of values, which we assume are next to
1118 // each other, due to our sorting criteria.
1119 assert(i < Ops.size());
1120 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1121 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1122 // Drop duplicate values for And and Or.
1123 Ops.erase(Ops.begin()+i);
1129 // Drop pairs of values for Xor.
1130 assert(Opcode == Instruction::Xor);
1132 return Constant::getNullValue(Ops[0].Op->getType());
1135 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1143 /// Helper function of CombineXorOpnd(). It creates a bitwise-and
1144 /// instruction with the given two operands, and return the resulting
1145 /// instruction. There are two special cases: 1) if the constant operand is 0,
1146 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1148 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1149 const APInt &ConstOpnd) {
1150 if (ConstOpnd.isNullValue())
1153 if (ConstOpnd.isAllOnesValue())
1156 Instruction *I = BinaryOperator::CreateAnd(
1157 Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra",
1159 I->setDebugLoc(InsertBefore->getDebugLoc());
1163 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1164 // into "R ^ C", where C would be 0, and R is a symbolic value.
1166 // If it was successful, true is returned, and the "R" and "C" is returned
1167 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1168 // and both "Res" and "ConstOpnd" remain unchanged.
1169 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1170 APInt &ConstOpnd, Value *&Res) {
1171 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1172 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1173 // = (x & ~c1) ^ (c1 ^ c2)
1174 // It is useful only when c1 == c2.
1175 if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isNullValue())
1178 if (!Opnd1->getValue()->hasOneUse())
1181 const APInt &C1 = Opnd1->getConstPart();
1182 if (C1 != ConstOpnd)
1185 Value *X = Opnd1->getSymbolicPart();
1186 Res = createAndInstr(I, X, ~C1);
1187 // ConstOpnd was C2, now C1 ^ C2.
1190 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1191 RedoInsts.insert(T);
1195 // Helper function of OptimizeXor(). It tries to simplify
1196 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1199 // If it was successful, true is returned, and the "R" and "C" is returned
1200 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1201 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1202 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1203 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1204 XorOpnd *Opnd2, APInt &ConstOpnd,
1206 Value *X = Opnd1->getSymbolicPart();
1207 if (X != Opnd2->getSymbolicPart())
1210 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1211 int DeadInstNum = 1;
1212 if (Opnd1->getValue()->hasOneUse())
1214 if (Opnd2->getValue()->hasOneUse())
1218 // (x | c1) ^ (x & c2)
1219 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1220 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1221 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1223 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1224 if (Opnd2->isOrExpr())
1225 std::swap(Opnd1, Opnd2);
1227 const APInt &C1 = Opnd1->getConstPart();
1228 const APInt &C2 = Opnd2->getConstPart();
1229 APInt C3((~C1) ^ C2);
1231 // Do not increase code size!
1232 if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1233 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1234 if (NewInstNum > DeadInstNum)
1238 Res = createAndInstr(I, X, C3);
1240 } else if (Opnd1->isOrExpr()) {
1241 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1243 const APInt &C1 = Opnd1->getConstPart();
1244 const APInt &C2 = Opnd2->getConstPart();
1247 // Do not increase code size
1248 if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1249 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1250 if (NewInstNum > DeadInstNum)
1254 Res = createAndInstr(I, X, C3);
1257 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1259 const APInt &C1 = Opnd1->getConstPart();
1260 const APInt &C2 = Opnd2->getConstPart();
1262 Res = createAndInstr(I, X, C3);
1265 // Put the original operands in the Redo list; hope they will be deleted
1267 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1268 RedoInsts.insert(T);
1269 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1270 RedoInsts.insert(T);
1275 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1276 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1278 Value *ReassociatePass::OptimizeXor(Instruction *I,
1279 SmallVectorImpl<ValueEntry> &Ops) {
1280 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1283 if (Ops.size() == 1)
1286 SmallVector<XorOpnd, 8> Opnds;
1287 SmallVector<XorOpnd*, 8> OpndPtrs;
1288 Type *Ty = Ops[0].Op->getType();
1289 APInt ConstOpnd(Ty->getScalarSizeInBits(), 0);
1291 // Step 1: Convert ValueEntry to XorOpnd
1292 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1293 Value *V = Ops[i].Op;
1295 // TODO: Support non-splat vectors.
1296 if (match(V, PatternMatch::m_APInt(C))) {
1300 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1305 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1306 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1307 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1308 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1309 // when new elements are added to the vector.
1310 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1311 OpndPtrs.push_back(&Opnds[i]);
1313 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1314 // the same symbolic value cluster together. For instance, the input operand
1315 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1316 // ("x | 123", "x & 789", "y & 456").
1318 // The purpose is twofold:
1319 // 1) Cluster together the operands sharing the same symbolic-value.
1320 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
1321 // could potentially shorten crital path, and expose more loop-invariants.
1322 // Note that values' rank are basically defined in RPO order (FIXME).
1323 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1324 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1325 // "z" in the order of X-Y-Z is better than any other orders.
1326 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(),
1327 [](XorOpnd *LHS, XorOpnd *RHS) {
1328 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1331 // Step 3: Combine adjacent operands
1332 XorOpnd *PrevOpnd = nullptr;
1333 bool Changed = false;
1334 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1335 XorOpnd *CurrOpnd = OpndPtrs[i];
1336 // The combined value
1339 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1340 if (!ConstOpnd.isNullValue() &&
1341 CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1344 *CurrOpnd = XorOpnd(CV);
1346 CurrOpnd->Invalidate();
1351 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1352 PrevOpnd = CurrOpnd;
1356 // step 3.2: When previous and current operands share the same symbolic
1357 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1358 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1359 // Remove previous operand
1360 PrevOpnd->Invalidate();
1362 *CurrOpnd = XorOpnd(CV);
1363 PrevOpnd = CurrOpnd;
1365 CurrOpnd->Invalidate();
1372 // Step 4: Reassemble the Ops
1375 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1376 XorOpnd &O = Opnds[i];
1379 ValueEntry VE(getRank(O.getValue()), O.getValue());
1382 if (!ConstOpnd.isNullValue()) {
1383 Value *C = ConstantInt::get(Ty, ConstOpnd);
1384 ValueEntry VE(getRank(C), C);
1387 unsigned Sz = Ops.size();
1389 return Ops.back().Op;
1391 assert(ConstOpnd.isNullValue());
1392 return ConstantInt::get(Ty, ConstOpnd);
1399 /// Optimize a series of operands to an 'add' instruction. This
1400 /// optimizes based on identities. If it can be reduced to a single Value, it
1401 /// is returned, otherwise the Ops list is mutated as necessary.
1402 Value *ReassociatePass::OptimizeAdd(Instruction *I,
1403 SmallVectorImpl<ValueEntry> &Ops) {
1404 // Scan the operand lists looking for X and -X pairs. If we find any, we
1405 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1407 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1409 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1410 Value *TheOp = Ops[i].Op;
1411 // Check to see if we've seen this operand before. If so, we factor all
1412 // instances of the operand together. Due to our sorting criteria, we know
1413 // that these need to be next to each other in the vector.
1414 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1415 // Rescan the list, remove all instances of this operand from the expr.
1416 unsigned NumFound = 0;
1418 Ops.erase(Ops.begin()+i);
1420 } while (i != Ops.size() && Ops[i].Op == TheOp);
1422 DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1425 // Insert a new multiply.
1426 Type *Ty = TheOp->getType();
1427 Constant *C = Ty->isIntOrIntVectorTy() ?
1428 ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1429 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1431 // Now that we have inserted a multiply, optimize it. This allows us to
1432 // handle cases that require multiple factoring steps, such as this:
1433 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1434 RedoInsts.insert(Mul);
1436 // If every add operand was a duplicate, return the multiply.
1440 // Otherwise, we had some input that didn't have the dupe, such as
1441 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1442 // things being added by this operation.
1443 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1450 // Check for X and -X or X and ~X in the operand list.
1451 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1452 !BinaryOperator::isNot(TheOp))
1456 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1457 X = BinaryOperator::getNegArgument(TheOp);
1458 else if (BinaryOperator::isNot(TheOp))
1459 X = BinaryOperator::getNotArgument(TheOp);
1461 unsigned FoundX = FindInOperandList(Ops, i, X);
1465 // Remove X and -X from the operand list.
1466 if (Ops.size() == 2 &&
1467 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1468 return Constant::getNullValue(X->getType());
1470 // Remove X and ~X from the operand list.
1471 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1472 return Constant::getAllOnesValue(X->getType());
1474 Ops.erase(Ops.begin()+i);
1478 --i; // Need to back up an extra one.
1479 Ops.erase(Ops.begin()+FoundX);
1481 --i; // Revisit element.
1482 e -= 2; // Removed two elements.
1484 // if X and ~X we append -1 to the operand list.
1485 if (BinaryOperator::isNot(TheOp)) {
1486 Value *V = Constant::getAllOnesValue(X->getType());
1487 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1492 // Scan the operand list, checking to see if there are any common factors
1493 // between operands. Consider something like A*A+A*B*C+D. We would like to
1494 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1495 // To efficiently find this, we count the number of times a factor occurs
1496 // for any ADD operands that are MULs.
1497 DenseMap<Value*, unsigned> FactorOccurrences;
1499 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1500 // where they are actually the same multiply.
1501 unsigned MaxOcc = 0;
1502 Value *MaxOccVal = nullptr;
1503 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1504 BinaryOperator *BOp =
1505 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1509 // Compute all of the factors of this added value.
1510 SmallVector<Value*, 8> Factors;
1511 FindSingleUseMultiplyFactors(BOp, Factors);
1512 assert(Factors.size() > 1 && "Bad linearize!");
1514 // Add one to FactorOccurrences for each unique factor in this op.
1515 SmallPtrSet<Value*, 8> Duplicates;
1516 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1517 Value *Factor = Factors[i];
1518 if (!Duplicates.insert(Factor).second)
1521 unsigned Occ = ++FactorOccurrences[Factor];
1527 // If Factor is a negative constant, add the negated value as a factor
1528 // because we can percolate the negate out. Watch for minint, which
1529 // cannot be positivified.
1530 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1531 if (CI->isNegative() && !CI->isMinValue(true)) {
1532 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1533 if (!Duplicates.insert(Factor).second)
1535 unsigned Occ = ++FactorOccurrences[Factor];
1541 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1542 if (CF->isNegative()) {
1543 APFloat F(CF->getValueAPF());
1545 Factor = ConstantFP::get(CF->getContext(), F);
1546 if (!Duplicates.insert(Factor).second)
1548 unsigned Occ = ++FactorOccurrences[Factor];
1558 // If any factor occurred more than one time, we can pull it out.
1560 DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1563 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1564 // this, we could otherwise run into situations where removing a factor
1565 // from an expression will drop a use of maxocc, and this can cause
1566 // RemoveFactorFromExpression on successive values to behave differently.
1567 Instruction *DummyInst =
1568 I->getType()->isIntOrIntVectorTy()
1569 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1570 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1572 SmallVector<WeakTrackingVH, 4> NewMulOps;
1573 for (unsigned i = 0; i != Ops.size(); ++i) {
1574 // Only try to remove factors from expressions we're allowed to.
1575 BinaryOperator *BOp =
1576 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1580 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1581 // The factorized operand may occur several times. Convert them all in
1583 for (unsigned j = Ops.size(); j != i;) {
1585 if (Ops[j].Op == Ops[i].Op) {
1586 NewMulOps.push_back(V);
1587 Ops.erase(Ops.begin()+j);
1594 // No need for extra uses anymore.
1595 DummyInst->deleteValue();
1597 unsigned NumAddedValues = NewMulOps.size();
1598 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1600 // Now that we have inserted the add tree, optimize it. This allows us to
1601 // handle cases that require multiple factoring steps, such as this:
1602 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1603 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1604 (void)NumAddedValues;
1605 if (Instruction *VI = dyn_cast<Instruction>(V))
1606 RedoInsts.insert(VI);
1608 // Create the multiply.
1609 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1611 // Rerun associate on the multiply in case the inner expression turned into
1612 // a multiply. We want to make sure that we keep things in canonical form.
1613 RedoInsts.insert(V2);
1615 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1616 // entire result expression is just the multiply "A*(B+C)".
1620 // Otherwise, we had some input that didn't have the factor, such as
1621 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1622 // things being added by this operation.
1623 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1629 /// \brief Build up a vector of value/power pairs factoring a product.
1631 /// Given a series of multiplication operands, build a vector of factors and
1632 /// the powers each is raised to when forming the final product. Sort them in
1633 /// the order of descending power.
1635 /// (x*x) -> [(x, 2)]
1636 /// ((x*x)*x) -> [(x, 3)]
1637 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1639 /// \returns Whether any factors have a power greater than one.
1640 static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1641 SmallVectorImpl<Factor> &Factors) {
1642 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1643 // Compute the sum of powers of simplifiable factors.
1644 unsigned FactorPowerSum = 0;
1645 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1646 Value *Op = Ops[Idx-1].Op;
1648 // Count the number of occurrences of this value.
1650 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1652 // Track for simplification all factors which occur 2 or more times.
1654 FactorPowerSum += Count;
1657 // We can only simplify factors if the sum of the powers of our simplifiable
1658 // factors is 4 or higher. When that is the case, we will *always* have
1659 // a simplification. This is an important invariant to prevent cyclicly
1660 // trying to simplify already minimal formations.
1661 if (FactorPowerSum < 4)
1664 // Now gather the simplifiable factors, removing them from Ops.
1666 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1667 Value *Op = Ops[Idx-1].Op;
1669 // Count the number of occurrences of this value.
1671 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1675 // Move an even number of occurrences to Factors.
1678 FactorPowerSum += Count;
1679 Factors.push_back(Factor(Op, Count));
1680 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1683 // None of the adjustments above should have reduced the sum of factor powers
1684 // below our mininum of '4'.
1685 assert(FactorPowerSum >= 4);
1687 std::stable_sort(Factors.begin(), Factors.end(),
1688 [](const Factor &LHS, const Factor &RHS) {
1689 return LHS.Power > RHS.Power;
1694 /// \brief Build a tree of multiplies, computing the product of Ops.
1695 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1696 SmallVectorImpl<Value*> &Ops) {
1697 if (Ops.size() == 1)
1700 Value *LHS = Ops.pop_back_val();
1702 if (LHS->getType()->isIntOrIntVectorTy())
1703 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1705 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1706 } while (!Ops.empty());
1711 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1713 /// Given a vector of values raised to various powers, where no two values are
1714 /// equal and the powers are sorted in decreasing order, compute the minimal
1715 /// DAG of multiplies to compute the final product, and return that product
1718 ReassociatePass::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1719 SmallVectorImpl<Factor> &Factors) {
1720 assert(Factors[0].Power);
1721 SmallVector<Value *, 4> OuterProduct;
1722 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1723 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1724 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1729 // We want to multiply across all the factors with the same power so that
1730 // we can raise them to that power as a single entity. Build a mini tree
1732 SmallVector<Value *, 4> InnerProduct;
1733 InnerProduct.push_back(Factors[LastIdx].Base);
1735 InnerProduct.push_back(Factors[Idx].Base);
1737 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1739 // Reset the base value of the first factor to the new expression tree.
1740 // We'll remove all the factors with the same power in a second pass.
1741 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1742 if (Instruction *MI = dyn_cast<Instruction>(M))
1743 RedoInsts.insert(MI);
1747 // Unique factors with equal powers -- we've folded them into the first one's
1749 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1750 [](const Factor &LHS, const Factor &RHS) {
1751 return LHS.Power == RHS.Power;
1755 // Iteratively collect the base of each factor with an add power into the
1756 // outer product, and halve each power in preparation for squaring the
1758 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1759 if (Factors[Idx].Power & 1)
1760 OuterProduct.push_back(Factors[Idx].Base);
1761 Factors[Idx].Power >>= 1;
1763 if (Factors[0].Power) {
1764 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1765 OuterProduct.push_back(SquareRoot);
1766 OuterProduct.push_back(SquareRoot);
1768 if (OuterProduct.size() == 1)
1769 return OuterProduct.front();
1771 Value *V = buildMultiplyTree(Builder, OuterProduct);
1775 Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
1776 SmallVectorImpl<ValueEntry> &Ops) {
1777 // We can only optimize the multiplies when there is a chain of more than
1778 // three, such that a balanced tree might require fewer total multiplies.
1782 // Try to turn linear trees of multiplies without other uses of the
1783 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1785 SmallVector<Factor, 4> Factors;
1786 if (!collectMultiplyFactors(Ops, Factors))
1787 return nullptr; // All distinct factors, so nothing left for us to do.
1789 IRBuilder<> Builder(I);
1790 // The reassociate transformation for FP operations is performed only
1791 // if unsafe algebra is permitted by FastMathFlags. Propagate those flags
1792 // to the newly generated operations.
1793 if (auto FPI = dyn_cast<FPMathOperator>(I))
1794 Builder.setFastMathFlags(FPI->getFastMathFlags());
1796 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1800 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1801 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1805 Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
1806 SmallVectorImpl<ValueEntry> &Ops) {
1807 // Now that we have the linearized expression tree, try to optimize it.
1808 // Start by folding any constants that we found.
1809 Constant *Cst = nullptr;
1810 unsigned Opcode = I->getOpcode();
1811 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1812 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1813 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1815 // If there was nothing but constants then we are done.
1819 // Put the combined constant back at the end of the operand list, except if
1820 // there is no point. For example, an add of 0 gets dropped here, while a
1821 // multiplication by zero turns the whole expression into zero.
1822 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1823 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1825 Ops.push_back(ValueEntry(0, Cst));
1828 if (Ops.size() == 1) return Ops[0].Op;
1830 // Handle destructive annihilation due to identities between elements in the
1831 // argument list here.
1832 unsigned NumOps = Ops.size();
1835 case Instruction::And:
1836 case Instruction::Or:
1837 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1841 case Instruction::Xor:
1842 if (Value *Result = OptimizeXor(I, Ops))
1846 case Instruction::Add:
1847 case Instruction::FAdd:
1848 if (Value *Result = OptimizeAdd(I, Ops))
1852 case Instruction::Mul:
1853 case Instruction::FMul:
1854 if (Value *Result = OptimizeMul(I, Ops))
1859 if (Ops.size() != NumOps)
1860 return OptimizeExpression(I, Ops);
1864 // Remove dead instructions and if any operands are trivially dead add them to
1865 // Insts so they will be removed as well.
1866 void ReassociatePass::RecursivelyEraseDeadInsts(
1867 Instruction *I, SetVector<AssertingVH<Instruction>> &Insts) {
1868 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1869 SmallVector<Value *, 4> Ops(I->op_begin(), I->op_end());
1870 ValueRankMap.erase(I);
1872 RedoInsts.remove(I);
1873 I->eraseFromParent();
1875 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1876 if (OpInst->use_empty())
1877 Insts.insert(OpInst);
1880 /// Zap the given instruction, adding interesting operands to the work list.
1881 void ReassociatePass::EraseInst(Instruction *I) {
1882 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1883 DEBUG(dbgs() << "Erasing dead inst: "; I->dump());
1885 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1886 // Erase the dead instruction.
1887 ValueRankMap.erase(I);
1888 RedoInsts.remove(I);
1889 I->eraseFromParent();
1890 // Optimize its operands.
1891 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1892 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1893 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1894 // If this is a node in an expression tree, climb to the expression root
1895 // and add that since that's where optimization actually happens.
1896 unsigned Opcode = Op->getOpcode();
1897 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1898 Visited.insert(Op).second)
1899 Op = Op->user_back();
1900 RedoInsts.insert(Op);
1906 // Canonicalize expressions of the following form:
1907 // x + (-Constant * y) -> x - (Constant * y)
1908 // x - (-Constant * y) -> x + (Constant * y)
1909 Instruction *ReassociatePass::canonicalizeNegConstExpr(Instruction *I) {
1910 if (!I->hasOneUse() || I->getType()->isVectorTy())
1913 // Must be a fmul or fdiv instruction.
1914 unsigned Opcode = I->getOpcode();
1915 if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv)
1918 auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0));
1919 auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1));
1921 // Both operands are constant, let it get constant folded away.
1925 ConstantFP *CF = C0 ? C0 : C1;
1927 // Must have one constant operand.
1931 // Must be a negative ConstantFP.
1932 if (!CF->isNegative())
1935 // User must be a binary operator with one or more uses.
1936 Instruction *User = I->user_back();
1937 if (!isa<BinaryOperator>(User) || User->use_empty())
1940 unsigned UserOpcode = User->getOpcode();
1941 if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub)
1944 // Subtraction is not commutative. Explicitly, the following transform is
1945 // not valid: (-Constant * y) - x -> x + (Constant * y)
1946 if (!User->isCommutative() && User->getOperand(1) != I)
1949 // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the
1950 // resulting subtract will be broken up later. This can get us into an
1951 // infinite loop during reassociation.
1952 if (UserOpcode == Instruction::FAdd && ShouldBreakUpSubtract(User))
1955 // Change the sign of the constant.
1956 APFloat Val = CF->getValueAPF();
1958 I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val));
1960 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
1961 // ((-Const*y) + x) -> (x + (-Const*y)).
1962 if (User->getOperand(0) == I && User->isCommutative())
1963 cast<BinaryOperator>(User)->swapOperands();
1965 Value *Op0 = User->getOperand(0);
1966 Value *Op1 = User->getOperand(1);
1968 switch (UserOpcode) {
1970 llvm_unreachable("Unexpected Opcode!");
1971 case Instruction::FAdd:
1972 NI = BinaryOperator::CreateFSub(Op0, Op1);
1973 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1975 case Instruction::FSub:
1976 NI = BinaryOperator::CreateFAdd(Op0, Op1);
1977 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1981 NI->insertBefore(User);
1982 NI->setName(User->getName());
1983 User->replaceAllUsesWith(NI);
1984 NI->setDebugLoc(I->getDebugLoc());
1985 RedoInsts.insert(I);
1990 /// Inspect and optimize the given instruction. Note that erasing
1991 /// instructions is not allowed.
1992 void ReassociatePass::OptimizeInst(Instruction *I) {
1993 // Only consider operations that we understand.
1994 if (!isa<BinaryOperator>(I))
1997 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
1998 // If an operand of this shift is a reassociable multiply, or if the shift
1999 // is used by a reassociable multiply or add, turn into a multiply.
2000 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
2002 (isReassociableOp(I->user_back(), Instruction::Mul) ||
2003 isReassociableOp(I->user_back(), Instruction::Add)))) {
2004 Instruction *NI = ConvertShiftToMul(I);
2005 RedoInsts.insert(I);
2010 // Canonicalize negative constants out of expressions.
2011 if (Instruction *Res = canonicalizeNegConstExpr(I))
2014 // Commute binary operators, to canonicalize the order of their operands.
2015 // This can potentially expose more CSE opportunities, and makes writing other
2016 // transformations simpler.
2017 if (I->isCommutative())
2018 canonicalizeOperands(I);
2020 // Don't optimize floating point instructions that don't have unsafe algebra.
2021 if (I->getType()->isFPOrFPVectorTy() && !I->hasUnsafeAlgebra())
2024 // Do not reassociate boolean (i1) expressions. We want to preserve the
2025 // original order of evaluation for short-circuited comparisons that
2026 // SimplifyCFG has folded to AND/OR expressions. If the expression
2027 // is not further optimized, it is likely to be transformed back to a
2028 // short-circuited form for code gen, and the source order may have been
2029 // optimized for the most likely conditions.
2030 if (I->getType()->isIntegerTy(1))
2033 // If this is a subtract instruction which is not already in negate form,
2034 // see if we can convert it to X+-Y.
2035 if (I->getOpcode() == Instruction::Sub) {
2036 if (ShouldBreakUpSubtract(I)) {
2037 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2038 RedoInsts.insert(I);
2041 } else if (BinaryOperator::isNeg(I)) {
2042 // Otherwise, this is a negation. See if the operand is a multiply tree
2043 // and if this is not an inner node of a multiply tree.
2044 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2046 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2047 Instruction *NI = LowerNegateToMultiply(I);
2048 // If the negate was simplified, revisit the users to see if we can
2049 // reassociate further.
2050 for (User *U : NI->users()) {
2051 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2052 RedoInsts.insert(Tmp);
2054 RedoInsts.insert(I);
2059 } else if (I->getOpcode() == Instruction::FSub) {
2060 if (ShouldBreakUpSubtract(I)) {
2061 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2062 RedoInsts.insert(I);
2065 } else if (BinaryOperator::isFNeg(I)) {
2066 // Otherwise, this is a negation. See if the operand is a multiply tree
2067 // and if this is not an inner node of a multiply tree.
2068 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2070 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2071 // If the negate was simplified, revisit the users to see if we can
2072 // reassociate further.
2073 Instruction *NI = LowerNegateToMultiply(I);
2074 for (User *U : NI->users()) {
2075 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2076 RedoInsts.insert(Tmp);
2078 RedoInsts.insert(I);
2085 // If this instruction is an associative binary operator, process it.
2086 if (!I->isAssociative()) return;
2087 BinaryOperator *BO = cast<BinaryOperator>(I);
2089 // If this is an interior node of a reassociable tree, ignore it until we
2090 // get to the root of the tree, to avoid N^2 analysis.
2091 unsigned Opcode = BO->getOpcode();
2092 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2093 // During the initial run we will get to the root of the tree.
2094 // But if we get here while we are redoing instructions, there is no
2095 // guarantee that the root will be visited. So Redo later
2096 if (BO->user_back() != BO &&
2097 BO->getParent() == BO->user_back()->getParent())
2098 RedoInsts.insert(BO->user_back());
2102 // If this is an add tree that is used by a sub instruction, ignore it
2103 // until we process the subtract.
2104 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2105 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2107 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2108 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2111 ReassociateExpression(BO);
2114 void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2115 // First, walk the expression tree, linearizing the tree, collecting the
2116 // operand information.
2117 SmallVector<RepeatedValue, 8> Tree;
2118 MadeChange |= LinearizeExprTree(I, Tree);
2119 SmallVector<ValueEntry, 8> Ops;
2120 Ops.reserve(Tree.size());
2121 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2122 RepeatedValue E = Tree[i];
2123 Ops.append(E.second.getZExtValue(),
2124 ValueEntry(getRank(E.first), E.first));
2127 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2129 // Now that we have linearized the tree to a list and have gathered all of
2130 // the operands and their ranks, sort the operands by their rank. Use a
2131 // stable_sort so that values with equal ranks will have their relative
2132 // positions maintained (and so the compiler is deterministic). Note that
2133 // this sorts so that the highest ranking values end up at the beginning of
2135 std::stable_sort(Ops.begin(), Ops.end());
2137 // Now that we have the expression tree in a convenient
2138 // sorted form, optimize it globally if possible.
2139 if (Value *V = OptimizeExpression(I, Ops)) {
2141 // Self-referential expression in unreachable code.
2143 // This expression tree simplified to something that isn't a tree,
2145 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2146 I->replaceAllUsesWith(V);
2147 if (Instruction *VI = dyn_cast<Instruction>(V))
2148 if (I->getDebugLoc())
2149 VI->setDebugLoc(I->getDebugLoc());
2150 RedoInsts.insert(I);
2155 // We want to sink immediates as deeply as possible except in the case where
2156 // this is a multiply tree used only by an add, and the immediate is a -1.
2157 // In this case we reassociate to put the negation on the outside so that we
2158 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2159 if (I->hasOneUse()) {
2160 if (I->getOpcode() == Instruction::Mul &&
2161 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2162 isa<ConstantInt>(Ops.back().Op) &&
2163 cast<ConstantInt>(Ops.back().Op)->isMinusOne()) {
2164 ValueEntry Tmp = Ops.pop_back_val();
2165 Ops.insert(Ops.begin(), Tmp);
2166 } else if (I->getOpcode() == Instruction::FMul &&
2167 cast<Instruction>(I->user_back())->getOpcode() ==
2168 Instruction::FAdd &&
2169 isa<ConstantFP>(Ops.back().Op) &&
2170 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2171 ValueEntry Tmp = Ops.pop_back_val();
2172 Ops.insert(Ops.begin(), Tmp);
2176 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2178 if (Ops.size() == 1) {
2180 // Self-referential expression in unreachable code.
2183 // This expression tree simplified to something that isn't a tree,
2185 I->replaceAllUsesWith(Ops[0].Op);
2186 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2187 OI->setDebugLoc(I->getDebugLoc());
2188 RedoInsts.insert(I);
2192 // Now that we ordered and optimized the expressions, splat them back into
2193 // the expression tree, removing any unneeded nodes.
2194 RewriteExprTree(I, Ops);
2197 PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) {
2198 // Get the functions basic blocks in Reverse Post Order. This order is used by
2199 // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic
2200 // blocks (it has been seen that the analysis in this pass could hang when
2201 // analysing dead basic blocks).
2202 ReversePostOrderTraversal<Function *> RPOT(&F);
2204 // Calculate the rank map for F.
2205 BuildRankMap(F, RPOT);
2208 // Traverse the same blocks that was analysed by BuildRankMap.
2209 for (BasicBlock *BI : RPOT) {
2210 assert(RankMap.count(&*BI) && "BB should be ranked.");
2211 // Optimize every instruction in the basic block.
2212 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;)
2213 if (isInstructionTriviallyDead(&*II)) {
2217 assert(II->getParent() == &*BI && "Moved to a different block!");
2221 // Make a copy of all the instructions to be redone so we can remove dead
2223 SetVector<AssertingVH<Instruction>> ToRedo(RedoInsts);
2224 // Iterate over all instructions to be reevaluated and remove trivially dead
2225 // instructions. If any operand of the trivially dead instruction becomes
2226 // dead mark it for deletion as well. Continue this process until all
2227 // trivially dead instructions have been removed.
2228 while (!ToRedo.empty()) {
2229 Instruction *I = ToRedo.pop_back_val();
2230 if (isInstructionTriviallyDead(I)) {
2231 RecursivelyEraseDeadInsts(I, ToRedo);
2236 // Now that we have removed dead instructions, we can reoptimize the
2237 // remaining instructions.
2238 while (!RedoInsts.empty()) {
2239 Instruction *I = RedoInsts.pop_back_val();
2240 if (isInstructionTriviallyDead(I))
2247 // We are done with the rank map.
2249 ValueRankMap.clear();
2252 PreservedAnalyses PA;
2253 PA.preserveSet<CFGAnalyses>();
2254 PA.preserve<GlobalsAA>();
2258 return PreservedAnalyses::all();
2263 class ReassociateLegacyPass : public FunctionPass {
2264 ReassociatePass Impl;
2267 static char ID; // Pass identification, replacement for typeid
2269 ReassociateLegacyPass() : FunctionPass(ID) {
2270 initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry());
2273 bool runOnFunction(Function &F) override {
2274 if (skipFunction(F))
2277 FunctionAnalysisManager DummyFAM;
2278 auto PA = Impl.run(F, DummyFAM);
2279 return !PA.areAllPreserved();
2282 void getAnalysisUsage(AnalysisUsage &AU) const override {
2283 AU.setPreservesCFG();
2284 AU.addPreserved<GlobalsAAWrapperPass>();
2288 } // end anonymous namespace
2290 char ReassociateLegacyPass::ID = 0;
2292 INITIALIZE_PASS(ReassociateLegacyPass, "reassociate",
2293 "Reassociate expressions", false, false)
2295 // Public interface to the Reassociate pass
2296 FunctionPass *llvm::createReassociatePass() {
2297 return new ReassociateLegacyPass();