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/ValueHandle.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Transforms/Scalar.h"
43 #include "llvm/Transforms/Utils/Local.h"
46 using namespace reassociate;
48 #define DEBUG_TYPE "reassociate"
50 STATISTIC(NumChanged, "Number of insts reassociated");
51 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
52 STATISTIC(NumFactor , "Number of multiplies factored");
55 /// Print out the expression identified in the Ops list.
57 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
58 Module *M = I->getModule();
59 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
60 << *Ops[0].Op->getType() << '\t';
61 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
63 Ops[i].Op->printAsOperand(dbgs(), false, M);
64 dbgs() << ", #" << Ops[i].Rank << "] ";
69 /// Utility class representing a non-constant Xor-operand. We classify
70 /// non-constant Xor-Operands into two categories:
71 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0
73 /// C2.1) The operand is in the form of "X | C", where C is a non-zero
75 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
76 /// operand as "E | 0"
77 class llvm::reassociate::XorOpnd {
81 bool isInvalid() const { return SymbolicPart == nullptr; }
82 bool isOrExpr() const { return isOr; }
83 Value *getValue() const { return OrigVal; }
84 Value *getSymbolicPart() const { return SymbolicPart; }
85 unsigned getSymbolicRank() const { return SymbolicRank; }
86 const APInt &getConstPart() const { return ConstPart; }
88 void Invalidate() { SymbolicPart = OrigVal = nullptr; }
89 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
95 unsigned SymbolicRank;
99 XorOpnd::XorOpnd(Value *V) {
100 assert(!isa<ConstantInt>(V) && "No ConstantInt");
102 Instruction *I = dyn_cast<Instruction>(V);
105 if (I && (I->getOpcode() == Instruction::Or ||
106 I->getOpcode() == Instruction::And)) {
107 Value *V0 = I->getOperand(0);
108 Value *V1 = I->getOperand(1);
109 if (isa<ConstantInt>(V0))
112 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) {
113 ConstPart = C->getValue();
115 isOr = (I->getOpcode() == Instruction::Or);
120 // view the operand as "V | 0"
122 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth());
126 /// Return true if V is an instruction of the specified opcode and if it
127 /// only has one use.
128 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
129 if (V->hasOneUse() && isa<Instruction>(V) &&
130 cast<Instruction>(V)->getOpcode() == Opcode &&
131 (!isa<FPMathOperator>(V) ||
132 cast<Instruction>(V)->hasUnsafeAlgebra()))
133 return cast<BinaryOperator>(V);
137 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
139 if (V->hasOneUse() && isa<Instruction>(V) &&
140 (cast<Instruction>(V)->getOpcode() == Opcode1 ||
141 cast<Instruction>(V)->getOpcode() == Opcode2) &&
142 (!isa<FPMathOperator>(V) ||
143 cast<Instruction>(V)->hasUnsafeAlgebra()))
144 return cast<BinaryOperator>(V);
148 void ReassociatePass::BuildRankMap(Function &F,
149 ReversePostOrderTraversal<Function*> &RPOT) {
152 // Assign distinct ranks to function arguments.
153 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
154 ValueRankMap[&*I] = ++i;
155 DEBUG(dbgs() << "Calculated Rank[" << I->getName() << "] = " << i << "\n");
158 // Traverse basic blocks in ReversePostOrder
159 for (BasicBlock *BB : RPOT) {
160 unsigned BBRank = RankMap[BB] = ++i << 16;
162 // Walk the basic block, adding precomputed ranks for any instructions that
163 // we cannot move. This ensures that the ranks for these instructions are
164 // all different in the block.
165 for (Instruction &I : *BB)
166 if (mayBeMemoryDependent(I))
167 ValueRankMap[&I] = ++BBRank;
171 unsigned ReassociatePass::getRank(Value *V) {
172 Instruction *I = dyn_cast<Instruction>(V);
174 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
175 return 0; // Otherwise it's a global or constant, rank 0.
178 if (unsigned Rank = ValueRankMap[I])
179 return Rank; // Rank already known?
181 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
182 // we can reassociate expressions for code motion! Since we do not recurse
183 // for PHI nodes, we cannot have infinite recursion here, because there
184 // cannot be loops in the value graph that do not go through PHI nodes.
185 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
186 for (unsigned i = 0, e = I->getNumOperands();
187 i != e && Rank != MaxRank; ++i)
188 Rank = std::max(Rank, getRank(I->getOperand(i)));
190 // If this is a not or neg instruction, do not count it for rank. This
191 // assures us that X and ~X will have the same rank.
192 if (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
193 !BinaryOperator::isFNeg(I))
196 DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
198 return ValueRankMap[I] = Rank;
201 // Canonicalize constants to RHS. Otherwise, sort the operands by rank.
202 void ReassociatePass::canonicalizeOperands(Instruction *I) {
203 assert(isa<BinaryOperator>(I) && "Expected binary operator.");
204 assert(I->isCommutative() && "Expected commutative operator.");
206 Value *LHS = I->getOperand(0);
207 Value *RHS = I->getOperand(1);
208 unsigned LHSRank = getRank(LHS);
209 unsigned RHSRank = getRank(RHS);
211 if (isa<Constant>(RHS))
214 if (isa<Constant>(LHS) || RHSRank < LHSRank)
215 cast<BinaryOperator>(I)->swapOperands();
218 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
219 Instruction *InsertBefore, Value *FlagsOp) {
220 if (S1->getType()->isIntOrIntVectorTy())
221 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
223 BinaryOperator *Res =
224 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
225 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
230 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
231 Instruction *InsertBefore, Value *FlagsOp) {
232 if (S1->getType()->isIntOrIntVectorTy())
233 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
235 BinaryOperator *Res =
236 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
237 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
242 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
243 Instruction *InsertBefore, Value *FlagsOp) {
244 if (S1->getType()->isIntOrIntVectorTy())
245 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
247 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
248 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
253 /// Replace 0-X with X*-1.
254 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
255 Type *Ty = Neg->getType();
256 Constant *NegOne = Ty->isIntOrIntVectorTy() ?
257 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
259 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
260 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
262 Neg->replaceAllUsesWith(Res);
263 Res->setDebugLoc(Neg->getDebugLoc());
267 /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
268 /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for
269 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
270 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
271 /// even x in Bitwidth-bit arithmetic.
272 static unsigned CarmichaelShift(unsigned Bitwidth) {
278 /// Add the extra weight 'RHS' to the existing weight 'LHS',
279 /// reducing the combined weight using any special properties of the operation.
280 /// The existing weight LHS represents the computation X op X op ... op X where
281 /// X occurs LHS times. The combined weight represents X op X op ... op X with
282 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
283 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
284 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
285 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
286 // If we were working with infinite precision arithmetic then the combined
287 // weight would be LHS + RHS. But we are using finite precision arithmetic,
288 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
289 // for nilpotent operations and addition, but not for idempotent operations
290 // and multiplication), so it is important to correctly reduce the combined
291 // weight back into range if wrapping would be wrong.
293 // If RHS is zero then the weight didn't change.
294 if (RHS.isMinValue())
296 // If LHS is zero then the combined weight is RHS.
297 if (LHS.isMinValue()) {
301 // From this point on we know that neither LHS nor RHS is zero.
303 if (Instruction::isIdempotent(Opcode)) {
304 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
305 // weight of 1. Keeping weights at zero or one also means that wrapping is
307 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
308 return; // Return a weight of 1.
310 if (Instruction::isNilpotent(Opcode)) {
311 // Nilpotent means X op X === 0, so reduce weights modulo 2.
312 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
313 LHS = 0; // 1 + 1 === 0 modulo 2.
316 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
317 // TODO: Reduce the weight by exploiting nsw/nuw?
322 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
323 "Unknown associative operation!");
324 unsigned Bitwidth = LHS.getBitWidth();
325 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
326 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
327 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
328 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
329 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
330 // which by a happy accident means that they can always be represented using
332 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
333 // the Carmichael number).
335 /// CM - The value of Carmichael's lambda function.
336 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
337 // Any weight W >= Threshold can be replaced with W - CM.
338 APInt Threshold = CM + Bitwidth;
339 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
340 // For Bitwidth 4 or more the following sum does not overflow.
342 while (LHS.uge(Threshold))
345 // To avoid problems with overflow do everything the same as above but using
347 unsigned CM = 1U << CarmichaelShift(Bitwidth);
348 unsigned Threshold = CM + Bitwidth;
349 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
350 "Weights not reduced!");
351 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
352 while (Total >= Threshold)
358 typedef std::pair<Value*, APInt> RepeatedValue;
360 /// Given an associative binary expression, return the leaf
361 /// nodes in Ops along with their weights (how many times the leaf occurs). The
362 /// original expression is the same as
363 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
365 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
369 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
371 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
373 /// This routine may modify the function, in which case it returns 'true'. The
374 /// changes it makes may well be destructive, changing the value computed by 'I'
375 /// to something completely different. Thus if the routine returns 'true' then
376 /// you MUST either replace I with a new expression computed from the Ops array,
377 /// or use RewriteExprTree to put the values back in.
379 /// A leaf node is either not a binary operation of the same kind as the root
380 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
381 /// opcode), or is the same kind of binary operator but has a use which either
382 /// does not belong to the expression, or does belong to the expression but is
383 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
384 /// of the expression, while for non-leaf nodes (except for the root 'I') every
385 /// use is a non-leaf node of the expression.
388 /// expression graph node names
398 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
399 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
401 /// The expression is maximal: if some instruction is a binary operator of the
402 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
403 /// then the instruction also belongs to the expression, is not a leaf node of
404 /// it, and its operands also belong to the expression (but may be leaf nodes).
406 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
407 /// order to ensure that every non-root node in the expression has *exactly one*
408 /// use by a non-leaf node of the expression. This destruction means that the
409 /// caller MUST either replace 'I' with a new expression or use something like
410 /// RewriteExprTree to put the values back in if the routine indicates that it
411 /// made a change by returning 'true'.
413 /// In the above example either the right operand of A or the left operand of B
414 /// will be replaced by undef. If it is B's operand then this gives:
418 /// + + | A, B - operand of B replaced with undef
424 /// Note that such undef operands can only be reached by passing through 'I'.
425 /// For example, if you visit operands recursively starting from a leaf node
426 /// then you will never see such an undef operand unless you get back to 'I',
427 /// which requires passing through a phi node.
429 /// Note that this routine may also mutate binary operators of the wrong type
430 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
431 /// of the expression) if it can turn them into binary operators of the right
432 /// type and thus make the expression bigger.
434 static bool LinearizeExprTree(BinaryOperator *I,
435 SmallVectorImpl<RepeatedValue> &Ops) {
436 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
437 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
438 unsigned Opcode = I->getOpcode();
439 assert(I->isAssociative() && I->isCommutative() &&
440 "Expected an associative and commutative operation!");
442 // Visit all operands of the expression, keeping track of their weight (the
443 // number of paths from the expression root to the operand, or if you like
444 // the number of times that operand occurs in the linearized expression).
445 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
446 // while A has weight two.
448 // Worklist of non-leaf nodes (their operands are in the expression too) along
449 // with their weights, representing a certain number of paths to the operator.
450 // If an operator occurs in the worklist multiple times then we found multiple
451 // ways to get to it.
452 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
453 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
454 bool Changed = false;
456 // Leaves of the expression are values that either aren't the right kind of
457 // operation (eg: a constant, or a multiply in an add tree), or are, but have
458 // some uses that are not inside the expression. For example, in I = X + X,
459 // X = A + B, the value X has two uses (by I) that are in the expression. If
460 // X has any other uses, for example in a return instruction, then we consider
461 // X to be a leaf, and won't analyze it further. When we first visit a value,
462 // if it has more than one use then at first we conservatively consider it to
463 // be a leaf. Later, as the expression is explored, we may discover some more
464 // uses of the value from inside the expression. If all uses turn out to be
465 // from within the expression (and the value is a binary operator of the right
466 // kind) then the value is no longer considered to be a leaf, and its operands
469 // Leaves - Keeps track of the set of putative leaves as well as the number of
470 // paths to each leaf seen so far.
471 typedef DenseMap<Value*, APInt> LeafMap;
472 LeafMap Leaves; // Leaf -> Total weight so far.
473 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
476 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
478 while (!Worklist.empty()) {
479 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
480 I = P.first; // We examine the operands of this binary operator.
482 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
483 Value *Op = I->getOperand(OpIdx);
484 APInt Weight = P.second; // Number of paths to this operand.
485 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
486 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
488 // If this is a binary operation of the right kind with only one use then
489 // add its operands to the expression.
490 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
491 assert(Visited.insert(Op).second && "Not first visit!");
492 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
493 Worklist.push_back(std::make_pair(BO, Weight));
497 // Appears to be a leaf. Is the operand already in the set of leaves?
498 LeafMap::iterator It = Leaves.find(Op);
499 if (It == Leaves.end()) {
500 // Not in the leaf map. Must be the first time we saw this operand.
501 assert(Visited.insert(Op).second && "Not first visit!");
502 if (!Op->hasOneUse()) {
503 // This value has uses not accounted for by the expression, so it is
504 // not safe to modify. Mark it as being a leaf.
505 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
506 LeafOrder.push_back(Op);
510 // No uses outside the expression, try morphing it.
512 // Already in the leaf map.
513 assert(It != Leaves.end() && Visited.count(Op) &&
514 "In leaf map but not visited!");
516 // Update the number of paths to the leaf.
517 IncorporateWeight(It->second, Weight, Opcode);
519 #if 0 // TODO: Re-enable once PR13021 is fixed.
520 // The leaf already has one use from inside the expression. As we want
521 // exactly one such use, drop this new use of the leaf.
522 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
523 I->setOperand(OpIdx, UndefValue::get(I->getType()));
526 // If the leaf is a binary operation of the right kind and we now see
527 // that its multiple original uses were in fact all by nodes belonging
528 // to the expression, then no longer consider it to be a leaf and add
529 // its operands to the expression.
530 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
531 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
532 Worklist.push_back(std::make_pair(BO, It->second));
538 // If we still have uses that are not accounted for by the expression
539 // then it is not safe to modify the value.
540 if (!Op->hasOneUse())
543 // No uses outside the expression, try morphing it.
545 Leaves.erase(It); // Since the value may be morphed below.
548 // At this point we have a value which, first of all, is not a binary
549 // expression of the right kind, and secondly, is only used inside the
550 // expression. This means that it can safely be modified. See if we
551 // can usefully morph it into an expression of the right kind.
552 assert((!isa<Instruction>(Op) ||
553 cast<Instruction>(Op)->getOpcode() != Opcode
554 || (isa<FPMathOperator>(Op) &&
555 !cast<Instruction>(Op)->hasUnsafeAlgebra())) &&
556 "Should have been handled above!");
557 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
559 // If this is a multiply expression, turn any internal negations into
560 // multiplies by -1 so they can be reassociated.
561 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
562 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
563 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
564 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
565 BO = LowerNegateToMultiply(BO);
566 DEBUG(dbgs() << *BO << '\n');
567 Worklist.push_back(std::make_pair(BO, Weight));
572 // Failed to morph into an expression of the right type. This really is
574 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
575 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
576 LeafOrder.push_back(Op);
581 // The leaves, repeated according to their weights, represent the linearized
582 // form of the expression.
583 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
584 Value *V = LeafOrder[i];
585 LeafMap::iterator It = Leaves.find(V);
586 if (It == Leaves.end())
587 // Node initially thought to be a leaf wasn't.
589 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
590 APInt Weight = It->second;
591 if (Weight.isMinValue())
592 // Leaf already output or weight reduction eliminated it.
594 // Ensure the leaf is only output once.
596 Ops.push_back(std::make_pair(V, Weight));
599 // For nilpotent operations or addition there may be no operands, for example
600 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
601 // in both cases the weight reduces to 0 causing the value to be skipped.
603 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
604 assert(Identity && "Associative operation without identity!");
605 Ops.emplace_back(Identity, APInt(Bitwidth, 1));
611 /// Now that the operands for this expression tree are
612 /// linearized and optimized, emit them in-order.
613 void ReassociatePass::RewriteExprTree(BinaryOperator *I,
614 SmallVectorImpl<ValueEntry> &Ops) {
615 assert(Ops.size() > 1 && "Single values should be used directly!");
617 // Since our optimizations should never increase the number of operations, the
618 // new expression can usually be written reusing the existing binary operators
619 // from the original expression tree, without creating any new instructions,
620 // though the rewritten expression may have a completely different topology.
621 // We take care to not change anything if the new expression will be the same
622 // as the original. If more than trivial changes (like commuting operands)
623 // were made then we are obliged to clear out any optional subclass data like
626 /// NodesToRewrite - Nodes from the original expression available for writing
627 /// the new expression into.
628 SmallVector<BinaryOperator*, 8> NodesToRewrite;
629 unsigned Opcode = I->getOpcode();
630 BinaryOperator *Op = I;
632 /// NotRewritable - The operands being written will be the leaves of the new
633 /// expression and must not be used as inner nodes (via NodesToRewrite) by
634 /// mistake. Inner nodes are always reassociable, and usually leaves are not
635 /// (if they were they would have been incorporated into the expression and so
636 /// would not be leaves), so most of the time there is no danger of this. But
637 /// in rare cases a leaf may become reassociable if an optimization kills uses
638 /// of it, or it may momentarily become reassociable during rewriting (below)
639 /// due it being removed as an operand of one of its uses. Ensure that misuse
640 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
641 /// leaves and refusing to reuse any of them as inner nodes.
642 SmallPtrSet<Value*, 8> NotRewritable;
643 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
644 NotRewritable.insert(Ops[i].Op);
646 // ExpressionChanged - Non-null if the rewritten expression differs from the
647 // original in some non-trivial way, requiring the clearing of optional flags.
648 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
649 BinaryOperator *ExpressionChanged = nullptr;
650 for (unsigned i = 0; ; ++i) {
651 // The last operation (which comes earliest in the IR) is special as both
652 // operands will come from Ops, rather than just one with the other being
654 if (i+2 == Ops.size()) {
655 Value *NewLHS = Ops[i].Op;
656 Value *NewRHS = Ops[i+1].Op;
657 Value *OldLHS = Op->getOperand(0);
658 Value *OldRHS = Op->getOperand(1);
660 if (NewLHS == OldLHS && NewRHS == OldRHS)
661 // Nothing changed, leave it alone.
664 if (NewLHS == OldRHS && NewRHS == OldLHS) {
665 // The order of the operands was reversed. Swap them.
666 DEBUG(dbgs() << "RA: " << *Op << '\n');
668 DEBUG(dbgs() << "TO: " << *Op << '\n');
674 // The new operation differs non-trivially from the original. Overwrite
675 // the old operands with the new ones.
676 DEBUG(dbgs() << "RA: " << *Op << '\n');
677 if (NewLHS != OldLHS) {
678 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
679 if (BO && !NotRewritable.count(BO))
680 NodesToRewrite.push_back(BO);
681 Op->setOperand(0, NewLHS);
683 if (NewRHS != OldRHS) {
684 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
685 if (BO && !NotRewritable.count(BO))
686 NodesToRewrite.push_back(BO);
687 Op->setOperand(1, NewRHS);
689 DEBUG(dbgs() << "TO: " << *Op << '\n');
691 ExpressionChanged = Op;
698 // Not the last operation. The left-hand side will be a sub-expression
699 // while the right-hand side will be the current element of Ops.
700 Value *NewRHS = Ops[i].Op;
701 if (NewRHS != Op->getOperand(1)) {
702 DEBUG(dbgs() << "RA: " << *Op << '\n');
703 if (NewRHS == Op->getOperand(0)) {
704 // The new right-hand side was already present as the left operand. If
705 // we are lucky then swapping the operands will sort out both of them.
708 // Overwrite with the new right-hand side.
709 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
710 if (BO && !NotRewritable.count(BO))
711 NodesToRewrite.push_back(BO);
712 Op->setOperand(1, NewRHS);
713 ExpressionChanged = Op;
715 DEBUG(dbgs() << "TO: " << *Op << '\n');
720 // Now deal with the left-hand side. If this is already an operation node
721 // from the original expression then just rewrite the rest of the expression
723 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
724 if (BO && !NotRewritable.count(BO)) {
729 // Otherwise, grab a spare node from the original expression and use that as
730 // the left-hand side. If there are no nodes left then the optimizers made
731 // an expression with more nodes than the original! This usually means that
732 // they did something stupid but it might mean that the problem was just too
733 // hard (finding the mimimal number of multiplications needed to realize a
734 // multiplication expression is NP-complete). Whatever the reason, smart or
735 // stupid, create a new node if there are none left.
736 BinaryOperator *NewOp;
737 if (NodesToRewrite.empty()) {
738 Constant *Undef = UndefValue::get(I->getType());
739 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
740 Undef, Undef, "", I);
741 if (NewOp->getType()->isFPOrFPVectorTy())
742 NewOp->setFastMathFlags(I->getFastMathFlags());
744 NewOp = NodesToRewrite.pop_back_val();
747 DEBUG(dbgs() << "RA: " << *Op << '\n');
748 Op->setOperand(0, NewOp);
749 DEBUG(dbgs() << "TO: " << *Op << '\n');
750 ExpressionChanged = Op;
756 // If the expression changed non-trivially then clear out all subclass data
757 // starting from the operator specified in ExpressionChanged, and compactify
758 // the operators to just before the expression root to guarantee that the
759 // expression tree is dominated by all of Ops.
760 if (ExpressionChanged)
762 // Preserve FastMathFlags.
763 if (isa<FPMathOperator>(I)) {
764 FastMathFlags Flags = I->getFastMathFlags();
765 ExpressionChanged->clearSubclassOptionalData();
766 ExpressionChanged->setFastMathFlags(Flags);
768 ExpressionChanged->clearSubclassOptionalData();
770 if (ExpressionChanged == I)
772 ExpressionChanged->moveBefore(I);
773 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
776 // Throw away any left over nodes from the original expression.
777 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
778 RedoInsts.insert(NodesToRewrite[i]);
781 /// Insert instructions before the instruction pointed to by BI,
782 /// that computes the negative version of the value specified. The negative
783 /// version of the value is returned, and BI is left pointing at the instruction
784 /// that should be processed next by the reassociation pass.
785 /// Also add intermediate instructions to the redo list that are modified while
786 /// pushing the negates through adds. These will be revisited to see if
787 /// additional opportunities have been exposed.
788 static Value *NegateValue(Value *V, Instruction *BI,
789 SetVector<AssertingVH<Instruction>> &ToRedo) {
790 if (Constant *C = dyn_cast<Constant>(V)) {
791 if (C->getType()->isFPOrFPVectorTy()) {
792 return ConstantExpr::getFNeg(C);
794 return ConstantExpr::getNeg(C);
798 // We are trying to expose opportunity for reassociation. One of the things
799 // that we want to do to achieve this is to push a negation as deep into an
800 // expression chain as possible, to expose the add instructions. In practice,
801 // this means that we turn this:
802 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
803 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
804 // the constants. We assume that instcombine will clean up the mess later if
805 // we introduce tons of unnecessary negation instructions.
807 if (BinaryOperator *I =
808 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
809 // Push the negates through the add.
810 I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo));
811 I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo));
812 if (I->getOpcode() == Instruction::Add) {
813 I->setHasNoUnsignedWrap(false);
814 I->setHasNoSignedWrap(false);
817 // We must move the add instruction here, because the neg instructions do
818 // not dominate the old add instruction in general. By moving it, we are
819 // assured that the neg instructions we just inserted dominate the
820 // instruction we are about to insert after them.
823 I->setName(I->getName()+".neg");
825 // Add the intermediate negates to the redo list as processing them later
826 // could expose more reassociating opportunities.
831 // Okay, we need to materialize a negated version of V with an instruction.
832 // Scan the use lists of V to see if we have one already.
833 for (User *U : V->users()) {
834 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
837 // We found one! Now we have to make sure that the definition dominates
838 // this use. We do this by moving it to the entry block (if it is a
839 // non-instruction value) or right after the definition. These negates will
840 // be zapped by reassociate later, so we don't need much finesse here.
841 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
843 // Verify that the negate is in this function, V might be a constant expr.
844 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
847 BasicBlock::iterator InsertPt;
848 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
849 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
850 InsertPt = II->getNormalDest()->begin();
852 InsertPt = ++InstInput->getIterator();
854 while (isa<PHINode>(InsertPt)) ++InsertPt;
856 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
858 TheNeg->moveBefore(&*InsertPt);
859 if (TheNeg->getOpcode() == Instruction::Sub) {
860 TheNeg->setHasNoUnsignedWrap(false);
861 TheNeg->setHasNoSignedWrap(false);
863 TheNeg->andIRFlags(BI);
865 ToRedo.insert(TheNeg);
869 // Insert a 'neg' instruction that subtracts the value from zero to get the
871 BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI);
872 ToRedo.insert(NewNeg);
876 /// Return true if we should break up this subtract of X-Y into (X + -Y).
877 static bool ShouldBreakUpSubtract(Instruction *Sub) {
878 // If this is a negation, we can't split it up!
879 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
882 // Don't breakup X - undef.
883 if (isa<UndefValue>(Sub->getOperand(1)))
886 // Don't bother to break this up unless either the LHS is an associable add or
887 // subtract or if this is only used by one.
888 Value *V0 = Sub->getOperand(0);
889 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
890 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
892 Value *V1 = Sub->getOperand(1);
893 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
894 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
896 Value *VB = Sub->user_back();
897 if (Sub->hasOneUse() &&
898 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
899 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
905 /// If we have (X-Y), and if either X is an add, or if this is only used by an
906 /// add, transform this into (X+(0-Y)) to promote better reassociation.
907 static BinaryOperator *
908 BreakUpSubtract(Instruction *Sub, SetVector<AssertingVH<Instruction>> &ToRedo) {
909 // Convert a subtract into an add and a neg instruction. This allows sub
910 // instructions to be commuted with other add instructions.
912 // Calculate the negative value of Operand 1 of the sub instruction,
913 // and set it as the RHS of the add instruction we just made.
915 Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo);
916 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
917 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
918 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
921 // Everyone now refers to the add instruction.
922 Sub->replaceAllUsesWith(New);
923 New->setDebugLoc(Sub->getDebugLoc());
925 DEBUG(dbgs() << "Negated: " << *New << '\n');
929 /// If this is a shift of a reassociable multiply or is used by one, change
930 /// this into a multiply by a constant to assist with further reassociation.
931 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
932 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
933 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
935 BinaryOperator *Mul =
936 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
937 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
940 // Everyone now refers to the mul instruction.
941 Shl->replaceAllUsesWith(Mul);
942 Mul->setDebugLoc(Shl->getDebugLoc());
944 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
945 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
947 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
948 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
950 Mul->setHasNoSignedWrap(true);
951 Mul->setHasNoUnsignedWrap(NUW);
955 /// Scan backwards and forwards among values with the same rank as element i
956 /// to see if X exists. If X does not exist, return i. This is useful when
957 /// scanning for 'x' when we see '-x' because they both get the same rank.
958 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
960 unsigned XRank = Ops[i].Rank;
961 unsigned e = Ops.size();
962 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
965 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
966 if (Instruction *I2 = dyn_cast<Instruction>(X))
967 if (I1->isIdenticalTo(I2))
971 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
974 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
975 if (Instruction *I2 = dyn_cast<Instruction>(X))
976 if (I1->isIdenticalTo(I2))
982 /// Emit a tree of add instructions, summing Ops together
983 /// and returning the result. Insert the tree before I.
984 static Value *EmitAddTreeOfValues(Instruction *I,
985 SmallVectorImpl<WeakTrackingVH> &Ops) {
986 if (Ops.size() == 1) return Ops.back();
988 Value *V1 = Ops.back();
990 Value *V2 = EmitAddTreeOfValues(I, Ops);
991 return CreateAdd(V2, V1, "tmp", I, I);
994 /// If V is an expression tree that is a multiplication sequence,
995 /// and if this sequence contains a multiply by Factor,
996 /// remove Factor from the tree and return the new tree.
997 Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) {
998 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1002 SmallVector<RepeatedValue, 8> Tree;
1003 MadeChange |= LinearizeExprTree(BO, Tree);
1004 SmallVector<ValueEntry, 8> Factors;
1005 Factors.reserve(Tree.size());
1006 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1007 RepeatedValue E = Tree[i];
1008 Factors.append(E.second.getZExtValue(),
1009 ValueEntry(getRank(E.first), E.first));
1012 bool FoundFactor = false;
1013 bool NeedsNegate = false;
1014 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1015 if (Factors[i].Op == Factor) {
1017 Factors.erase(Factors.begin()+i);
1021 // If this is a negative version of this factor, remove it.
1022 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1023 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1024 if (FC1->getValue() == -FC2->getValue()) {
1025 FoundFactor = NeedsNegate = true;
1026 Factors.erase(Factors.begin()+i);
1029 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1030 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1031 const APFloat &F1 = FC1->getValueAPF();
1032 APFloat F2(FC2->getValueAPF());
1034 if (F1.compare(F2) == APFloat::cmpEqual) {
1035 FoundFactor = NeedsNegate = true;
1036 Factors.erase(Factors.begin() + i);
1044 // Make sure to restore the operands to the expression tree.
1045 RewriteExprTree(BO, Factors);
1049 BasicBlock::iterator InsertPt = ++BO->getIterator();
1051 // If this was just a single multiply, remove the multiply and return the only
1052 // remaining operand.
1053 if (Factors.size() == 1) {
1054 RedoInsts.insert(BO);
1057 RewriteExprTree(BO, Factors);
1062 V = CreateNeg(V, "neg", &*InsertPt, BO);
1067 /// If V is a single-use multiply, recursively add its operands as factors,
1068 /// otherwise add V to the list of factors.
1070 /// Ops is the top-level list of add operands we're trying to factor.
1071 static void FindSingleUseMultiplyFactors(Value *V,
1072 SmallVectorImpl<Value*> &Factors) {
1073 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1075 Factors.push_back(V);
1079 // Otherwise, add the LHS and RHS to the list of factors.
1080 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
1081 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
1084 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1085 /// This optimizes based on identities. If it can be reduced to a single Value,
1086 /// it is returned, otherwise the Ops list is mutated as necessary.
1087 static Value *OptimizeAndOrXor(unsigned Opcode,
1088 SmallVectorImpl<ValueEntry> &Ops) {
1089 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1090 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1091 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1092 // First, check for X and ~X in the operand list.
1093 assert(i < Ops.size());
1094 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1095 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1096 unsigned FoundX = FindInOperandList(Ops, i, X);
1098 if (Opcode == Instruction::And) // ...&X&~X = 0
1099 return Constant::getNullValue(X->getType());
1101 if (Opcode == Instruction::Or) // ...|X|~X = -1
1102 return Constant::getAllOnesValue(X->getType());
1106 // Next, check for duplicate pairs of values, which we assume are next to
1107 // each other, due to our sorting criteria.
1108 assert(i < Ops.size());
1109 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1110 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1111 // Drop duplicate values for And and Or.
1112 Ops.erase(Ops.begin()+i);
1118 // Drop pairs of values for Xor.
1119 assert(Opcode == Instruction::Xor);
1121 return Constant::getNullValue(Ops[0].Op->getType());
1124 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1132 /// Helper function of CombineXorOpnd(). It creates a bitwise-and
1133 /// instruction with the given two operands, and return the resulting
1134 /// instruction. There are two special cases: 1) if the constant operand is 0,
1135 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1137 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1138 const APInt &ConstOpnd) {
1139 if (!ConstOpnd.isNullValue()) {
1140 if (!ConstOpnd.isAllOnesValue()) {
1141 LLVMContext &Ctx = Opnd->getType()->getContext();
1143 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1144 "and.ra", InsertBefore);
1145 I->setDebugLoc(InsertBefore->getDebugLoc());
1153 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1154 // into "R ^ C", where C would be 0, and R is a symbolic value.
1156 // If it was successful, true is returned, and the "R" and "C" is returned
1157 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1158 // and both "Res" and "ConstOpnd" remain unchanged.
1160 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1161 APInt &ConstOpnd, Value *&Res) {
1162 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1163 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1164 // = (x & ~c1) ^ (c1 ^ c2)
1165 // It is useful only when c1 == c2.
1166 if (Opnd1->isOrExpr() && !Opnd1->getConstPart().isNullValue()) {
1167 if (!Opnd1->getValue()->hasOneUse())
1170 const APInt &C1 = Opnd1->getConstPart();
1171 if (C1 != ConstOpnd)
1174 Value *X = Opnd1->getSymbolicPart();
1175 Res = createAndInstr(I, X, ~C1);
1176 // ConstOpnd was C2, now C1 ^ C2.
1179 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1180 RedoInsts.insert(T);
1187 // Helper function of OptimizeXor(). It tries to simplify
1188 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1191 // If it was successful, true is returned, and the "R" and "C" is returned
1192 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1193 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1194 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1195 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1196 XorOpnd *Opnd2, APInt &ConstOpnd,
1198 Value *X = Opnd1->getSymbolicPart();
1199 if (X != Opnd2->getSymbolicPart())
1202 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1203 int DeadInstNum = 1;
1204 if (Opnd1->getValue()->hasOneUse())
1206 if (Opnd2->getValue()->hasOneUse())
1210 // (x | c1) ^ (x & c2)
1211 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1212 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1213 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1215 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1216 if (Opnd2->isOrExpr())
1217 std::swap(Opnd1, Opnd2);
1219 const APInt &C1 = Opnd1->getConstPart();
1220 const APInt &C2 = Opnd2->getConstPart();
1221 APInt C3((~C1) ^ C2);
1223 // Do not increase code size!
1224 if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1225 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1226 if (NewInstNum > DeadInstNum)
1230 Res = createAndInstr(I, X, C3);
1233 } else if (Opnd1->isOrExpr()) {
1234 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1236 const APInt &C1 = Opnd1->getConstPart();
1237 const APInt &C2 = Opnd2->getConstPart();
1240 // Do not increase code size
1241 if (!C3.isNullValue() && !C3.isAllOnesValue()) {
1242 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1243 if (NewInstNum > DeadInstNum)
1247 Res = createAndInstr(I, X, C3);
1250 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1252 const APInt &C1 = Opnd1->getConstPart();
1253 const APInt &C2 = Opnd2->getConstPart();
1255 Res = createAndInstr(I, X, C3);
1258 // Put the original operands in the Redo list; hope they will be deleted
1260 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1261 RedoInsts.insert(T);
1262 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1263 RedoInsts.insert(T);
1268 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1269 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1271 Value *ReassociatePass::OptimizeXor(Instruction *I,
1272 SmallVectorImpl<ValueEntry> &Ops) {
1273 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1276 if (Ops.size() == 1)
1279 Type *Ty = Ops[0].Op->getType();
1281 // TODO: We should optimize vector Xor instructions, but they are
1282 // currently unsupported.
1283 if (Ty->isVectorTy())
1286 SmallVector<XorOpnd, 8> Opnds;
1287 SmallVector<XorOpnd*, 8> OpndPtrs;
1288 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1290 // Step 1: Convert ValueEntry to XorOpnd
1291 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1292 Value *V = Ops[i].Op;
1293 if (!isa<ConstantInt>(V)) {
1295 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1298 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1301 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1302 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1303 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1304 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1305 // when new elements are added to the vector.
1306 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1307 OpndPtrs.push_back(&Opnds[i]);
1309 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1310 // the same symbolic value cluster together. For instance, the input operand
1311 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1312 // ("x | 123", "x & 789", "y & 456").
1314 // The purpose is twofold:
1315 // 1) Cluster together the operands sharing the same symbolic-value.
1316 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
1317 // could potentially shorten crital path, and expose more loop-invariants.
1318 // Note that values' rank are basically defined in RPO order (FIXME).
1319 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1320 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1321 // "z" in the order of X-Y-Z is better than any other orders.
1322 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(),
1323 [](XorOpnd *LHS, XorOpnd *RHS) {
1324 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1327 // Step 3: Combine adjacent operands
1328 XorOpnd *PrevOpnd = nullptr;
1329 bool Changed = false;
1330 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1331 XorOpnd *CurrOpnd = OpndPtrs[i];
1332 // The combined value
1335 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1336 if (!ConstOpnd.isNullValue() &&
1337 CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1340 *CurrOpnd = XorOpnd(CV);
1342 CurrOpnd->Invalidate();
1347 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1348 PrevOpnd = CurrOpnd;
1352 // step 3.2: When previous and current operands share the same symbolic
1353 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1355 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1356 // Remove previous operand
1357 PrevOpnd->Invalidate();
1359 *CurrOpnd = XorOpnd(CV);
1360 PrevOpnd = CurrOpnd;
1362 CurrOpnd->Invalidate();
1369 // Step 4: Reassemble the Ops
1372 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1373 XorOpnd &O = Opnds[i];
1376 ValueEntry VE(getRank(O.getValue()), O.getValue());
1379 if (!ConstOpnd.isNullValue()) {
1380 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1381 ValueEntry VE(getRank(C), C);
1384 int Sz = Ops.size();
1386 return Ops.back().Op;
1388 assert(ConstOpnd == 0);
1389 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1396 /// Optimize a series of operands to an 'add' instruction. This
1397 /// optimizes based on identities. If it can be reduced to a single Value, it
1398 /// is returned, otherwise the Ops list is mutated as necessary.
1399 Value *ReassociatePass::OptimizeAdd(Instruction *I,
1400 SmallVectorImpl<ValueEntry> &Ops) {
1401 // Scan the operand lists looking for X and -X pairs. If we find any, we
1402 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1404 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1406 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1407 Value *TheOp = Ops[i].Op;
1408 // Check to see if we've seen this operand before. If so, we factor all
1409 // instances of the operand together. Due to our sorting criteria, we know
1410 // that these need to be next to each other in the vector.
1411 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1412 // Rescan the list, remove all instances of this operand from the expr.
1413 unsigned NumFound = 0;
1415 Ops.erase(Ops.begin()+i);
1417 } while (i != Ops.size() && Ops[i].Op == TheOp);
1419 DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1422 // Insert a new multiply.
1423 Type *Ty = TheOp->getType();
1424 Constant *C = Ty->isIntOrIntVectorTy() ?
1425 ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1426 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1428 // Now that we have inserted a multiply, optimize it. This allows us to
1429 // handle cases that require multiple factoring steps, such as this:
1430 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1431 RedoInsts.insert(Mul);
1433 // If every add operand was a duplicate, return the multiply.
1437 // Otherwise, we had some input that didn't have the dupe, such as
1438 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1439 // things being added by this operation.
1440 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1447 // Check for X and -X or X and ~X in the operand list.
1448 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1449 !BinaryOperator::isNot(TheOp))
1453 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1454 X = BinaryOperator::getNegArgument(TheOp);
1455 else if (BinaryOperator::isNot(TheOp))
1456 X = BinaryOperator::getNotArgument(TheOp);
1458 unsigned FoundX = FindInOperandList(Ops, i, X);
1462 // Remove X and -X from the operand list.
1463 if (Ops.size() == 2 &&
1464 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1465 return Constant::getNullValue(X->getType());
1467 // Remove X and ~X from the operand list.
1468 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1469 return Constant::getAllOnesValue(X->getType());
1471 Ops.erase(Ops.begin()+i);
1475 --i; // Need to back up an extra one.
1476 Ops.erase(Ops.begin()+FoundX);
1478 --i; // Revisit element.
1479 e -= 2; // Removed two elements.
1481 // if X and ~X we append -1 to the operand list.
1482 if (BinaryOperator::isNot(TheOp)) {
1483 Value *V = Constant::getAllOnesValue(X->getType());
1484 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1489 // Scan the operand list, checking to see if there are any common factors
1490 // between operands. Consider something like A*A+A*B*C+D. We would like to
1491 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1492 // To efficiently find this, we count the number of times a factor occurs
1493 // for any ADD operands that are MULs.
1494 DenseMap<Value*, unsigned> FactorOccurrences;
1496 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1497 // where they are actually the same multiply.
1498 unsigned MaxOcc = 0;
1499 Value *MaxOccVal = nullptr;
1500 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1501 BinaryOperator *BOp =
1502 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1506 // Compute all of the factors of this added value.
1507 SmallVector<Value*, 8> Factors;
1508 FindSingleUseMultiplyFactors(BOp, Factors);
1509 assert(Factors.size() > 1 && "Bad linearize!");
1511 // Add one to FactorOccurrences for each unique factor in this op.
1512 SmallPtrSet<Value*, 8> Duplicates;
1513 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1514 Value *Factor = Factors[i];
1515 if (!Duplicates.insert(Factor).second)
1518 unsigned Occ = ++FactorOccurrences[Factor];
1524 // If Factor is a negative constant, add the negated value as a factor
1525 // because we can percolate the negate out. Watch for minint, which
1526 // cannot be positivified.
1527 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1528 if (CI->isNegative() && !CI->isMinValue(true)) {
1529 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1530 if (!Duplicates.insert(Factor).second)
1532 unsigned Occ = ++FactorOccurrences[Factor];
1538 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1539 if (CF->isNegative()) {
1540 APFloat F(CF->getValueAPF());
1542 Factor = ConstantFP::get(CF->getContext(), F);
1543 if (!Duplicates.insert(Factor).second)
1545 unsigned Occ = ++FactorOccurrences[Factor];
1555 // If any factor occurred more than one time, we can pull it out.
1557 DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1560 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1561 // this, we could otherwise run into situations where removing a factor
1562 // from an expression will drop a use of maxocc, and this can cause
1563 // RemoveFactorFromExpression on successive values to behave differently.
1564 Instruction *DummyInst =
1565 I->getType()->isIntOrIntVectorTy()
1566 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1567 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1569 SmallVector<WeakTrackingVH, 4> NewMulOps;
1570 for (unsigned i = 0; i != Ops.size(); ++i) {
1571 // Only try to remove factors from expressions we're allowed to.
1572 BinaryOperator *BOp =
1573 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1577 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1578 // The factorized operand may occur several times. Convert them all in
1580 for (unsigned j = Ops.size(); j != i;) {
1582 if (Ops[j].Op == Ops[i].Op) {
1583 NewMulOps.push_back(V);
1584 Ops.erase(Ops.begin()+j);
1591 // No need for extra uses anymore.
1592 DummyInst->deleteValue();
1594 unsigned NumAddedValues = NewMulOps.size();
1595 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1597 // Now that we have inserted the add tree, optimize it. This allows us to
1598 // handle cases that require multiple factoring steps, such as this:
1599 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1600 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1601 (void)NumAddedValues;
1602 if (Instruction *VI = dyn_cast<Instruction>(V))
1603 RedoInsts.insert(VI);
1605 // Create the multiply.
1606 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1608 // Rerun associate on the multiply in case the inner expression turned into
1609 // a multiply. We want to make sure that we keep things in canonical form.
1610 RedoInsts.insert(V2);
1612 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1613 // entire result expression is just the multiply "A*(B+C)".
1617 // Otherwise, we had some input that didn't have the factor, such as
1618 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1619 // things being added by this operation.
1620 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1626 /// \brief Build up a vector of value/power pairs factoring a product.
1628 /// Given a series of multiplication operands, build a vector of factors and
1629 /// the powers each is raised to when forming the final product. Sort them in
1630 /// the order of descending power.
1632 /// (x*x) -> [(x, 2)]
1633 /// ((x*x)*x) -> [(x, 3)]
1634 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1636 /// \returns Whether any factors have a power greater than one.
1637 static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1638 SmallVectorImpl<Factor> &Factors) {
1639 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1640 // Compute the sum of powers of simplifiable factors.
1641 unsigned FactorPowerSum = 0;
1642 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1643 Value *Op = Ops[Idx-1].Op;
1645 // Count the number of occurrences of this value.
1647 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1649 // Track for simplification all factors which occur 2 or more times.
1651 FactorPowerSum += Count;
1654 // We can only simplify factors if the sum of the powers of our simplifiable
1655 // factors is 4 or higher. When that is the case, we will *always* have
1656 // a simplification. This is an important invariant to prevent cyclicly
1657 // trying to simplify already minimal formations.
1658 if (FactorPowerSum < 4)
1661 // Now gather the simplifiable factors, removing them from Ops.
1663 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1664 Value *Op = Ops[Idx-1].Op;
1666 // Count the number of occurrences of this value.
1668 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1672 // Move an even number of occurrences to Factors.
1675 FactorPowerSum += Count;
1676 Factors.push_back(Factor(Op, Count));
1677 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1680 // None of the adjustments above should have reduced the sum of factor powers
1681 // below our mininum of '4'.
1682 assert(FactorPowerSum >= 4);
1684 std::stable_sort(Factors.begin(), Factors.end(),
1685 [](const Factor &LHS, const Factor &RHS) {
1686 return LHS.Power > RHS.Power;
1691 /// \brief Build a tree of multiplies, computing the product of Ops.
1692 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1693 SmallVectorImpl<Value*> &Ops) {
1694 if (Ops.size() == 1)
1697 Value *LHS = Ops.pop_back_val();
1699 if (LHS->getType()->isIntOrIntVectorTy())
1700 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1702 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1703 } while (!Ops.empty());
1708 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1710 /// Given a vector of values raised to various powers, where no two values are
1711 /// equal and the powers are sorted in decreasing order, compute the minimal
1712 /// DAG of multiplies to compute the final product, and return that product
1715 ReassociatePass::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1716 SmallVectorImpl<Factor> &Factors) {
1717 assert(Factors[0].Power);
1718 SmallVector<Value *, 4> OuterProduct;
1719 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1720 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1721 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1726 // We want to multiply across all the factors with the same power so that
1727 // we can raise them to that power as a single entity. Build a mini tree
1729 SmallVector<Value *, 4> InnerProduct;
1730 InnerProduct.push_back(Factors[LastIdx].Base);
1732 InnerProduct.push_back(Factors[Idx].Base);
1734 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1736 // Reset the base value of the first factor to the new expression tree.
1737 // We'll remove all the factors with the same power in a second pass.
1738 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1739 if (Instruction *MI = dyn_cast<Instruction>(M))
1740 RedoInsts.insert(MI);
1744 // Unique factors with equal powers -- we've folded them into the first one's
1746 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1747 [](const Factor &LHS, const Factor &RHS) {
1748 return LHS.Power == RHS.Power;
1752 // Iteratively collect the base of each factor with an add power into the
1753 // outer product, and halve each power in preparation for squaring the
1755 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1756 if (Factors[Idx].Power & 1)
1757 OuterProduct.push_back(Factors[Idx].Base);
1758 Factors[Idx].Power >>= 1;
1760 if (Factors[0].Power) {
1761 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1762 OuterProduct.push_back(SquareRoot);
1763 OuterProduct.push_back(SquareRoot);
1765 if (OuterProduct.size() == 1)
1766 return OuterProduct.front();
1768 Value *V = buildMultiplyTree(Builder, OuterProduct);
1772 Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
1773 SmallVectorImpl<ValueEntry> &Ops) {
1774 // We can only optimize the multiplies when there is a chain of more than
1775 // three, such that a balanced tree might require fewer total multiplies.
1779 // Try to turn linear trees of multiplies without other uses of the
1780 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1782 SmallVector<Factor, 4> Factors;
1783 if (!collectMultiplyFactors(Ops, Factors))
1784 return nullptr; // All distinct factors, so nothing left for us to do.
1786 IRBuilder<> Builder(I);
1787 // The reassociate transformation for FP operations is performed only
1788 // if unsafe algebra is permitted by FastMathFlags. Propagate those flags
1789 // to the newly generated operations.
1790 if (auto FPI = dyn_cast<FPMathOperator>(I))
1791 Builder.setFastMathFlags(FPI->getFastMathFlags());
1793 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1797 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1798 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1802 Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
1803 SmallVectorImpl<ValueEntry> &Ops) {
1804 // Now that we have the linearized expression tree, try to optimize it.
1805 // Start by folding any constants that we found.
1806 Constant *Cst = nullptr;
1807 unsigned Opcode = I->getOpcode();
1808 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1809 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1810 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1812 // If there was nothing but constants then we are done.
1816 // Put the combined constant back at the end of the operand list, except if
1817 // there is no point. For example, an add of 0 gets dropped here, while a
1818 // multiplication by zero turns the whole expression into zero.
1819 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1820 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1822 Ops.push_back(ValueEntry(0, Cst));
1825 if (Ops.size() == 1) return Ops[0].Op;
1827 // Handle destructive annihilation due to identities between elements in the
1828 // argument list here.
1829 unsigned NumOps = Ops.size();
1832 case Instruction::And:
1833 case Instruction::Or:
1834 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1838 case Instruction::Xor:
1839 if (Value *Result = OptimizeXor(I, Ops))
1843 case Instruction::Add:
1844 case Instruction::FAdd:
1845 if (Value *Result = OptimizeAdd(I, Ops))
1849 case Instruction::Mul:
1850 case Instruction::FMul:
1851 if (Value *Result = OptimizeMul(I, Ops))
1856 if (Ops.size() != NumOps)
1857 return OptimizeExpression(I, Ops);
1861 // Remove dead instructions and if any operands are trivially dead add them to
1862 // Insts so they will be removed as well.
1863 void ReassociatePass::RecursivelyEraseDeadInsts(
1864 Instruction *I, SetVector<AssertingVH<Instruction>> &Insts) {
1865 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1866 SmallVector<Value *, 4> Ops(I->op_begin(), I->op_end());
1867 ValueRankMap.erase(I);
1869 RedoInsts.remove(I);
1870 I->eraseFromParent();
1872 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1873 if (OpInst->use_empty())
1874 Insts.insert(OpInst);
1877 /// Zap the given instruction, adding interesting operands to the work list.
1878 void ReassociatePass::EraseInst(Instruction *I) {
1879 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1880 DEBUG(dbgs() << "Erasing dead inst: "; I->dump());
1882 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1883 // Erase the dead instruction.
1884 ValueRankMap.erase(I);
1885 RedoInsts.remove(I);
1886 I->eraseFromParent();
1887 // Optimize its operands.
1888 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1889 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1890 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1891 // If this is a node in an expression tree, climb to the expression root
1892 // and add that since that's where optimization actually happens.
1893 unsigned Opcode = Op->getOpcode();
1894 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1895 Visited.insert(Op).second)
1896 Op = Op->user_back();
1897 RedoInsts.insert(Op);
1901 // Canonicalize expressions of the following form:
1902 // x + (-Constant * y) -> x - (Constant * y)
1903 // x - (-Constant * y) -> x + (Constant * y)
1904 Instruction *ReassociatePass::canonicalizeNegConstExpr(Instruction *I) {
1905 if (!I->hasOneUse() || I->getType()->isVectorTy())
1908 // Must be a fmul or fdiv instruction.
1909 unsigned Opcode = I->getOpcode();
1910 if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv)
1913 auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0));
1914 auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1));
1916 // Both operands are constant, let it get constant folded away.
1920 ConstantFP *CF = C0 ? C0 : C1;
1922 // Must have one constant operand.
1926 // Must be a negative ConstantFP.
1927 if (!CF->isNegative())
1930 // User must be a binary operator with one or more uses.
1931 Instruction *User = I->user_back();
1932 if (!isa<BinaryOperator>(User) || User->use_empty())
1935 unsigned UserOpcode = User->getOpcode();
1936 if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub)
1939 // Subtraction is not commutative. Explicitly, the following transform is
1940 // not valid: (-Constant * y) - x -> x + (Constant * y)
1941 if (!User->isCommutative() && User->getOperand(1) != I)
1944 // Change the sign of the constant.
1945 APFloat Val = CF->getValueAPF();
1947 I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val));
1949 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
1950 // ((-Const*y) + x) -> (x + (-Const*y)).
1951 if (User->getOperand(0) == I && User->isCommutative())
1952 cast<BinaryOperator>(User)->swapOperands();
1954 Value *Op0 = User->getOperand(0);
1955 Value *Op1 = User->getOperand(1);
1957 switch (UserOpcode) {
1959 llvm_unreachable("Unexpected Opcode!");
1960 case Instruction::FAdd:
1961 NI = BinaryOperator::CreateFSub(Op0, Op1);
1962 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1964 case Instruction::FSub:
1965 NI = BinaryOperator::CreateFAdd(Op0, Op1);
1966 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1970 NI->insertBefore(User);
1971 NI->setName(User->getName());
1972 User->replaceAllUsesWith(NI);
1973 NI->setDebugLoc(I->getDebugLoc());
1974 RedoInsts.insert(I);
1979 /// Inspect and optimize the given instruction. Note that erasing
1980 /// instructions is not allowed.
1981 void ReassociatePass::OptimizeInst(Instruction *I) {
1982 // Only consider operations that we understand.
1983 if (!isa<BinaryOperator>(I))
1986 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
1987 // If an operand of this shift is a reassociable multiply, or if the shift
1988 // is used by a reassociable multiply or add, turn into a multiply.
1989 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
1991 (isReassociableOp(I->user_back(), Instruction::Mul) ||
1992 isReassociableOp(I->user_back(), Instruction::Add)))) {
1993 Instruction *NI = ConvertShiftToMul(I);
1994 RedoInsts.insert(I);
1999 // Canonicalize negative constants out of expressions.
2000 if (Instruction *Res = canonicalizeNegConstExpr(I))
2003 // Commute binary operators, to canonicalize the order of their operands.
2004 // This can potentially expose more CSE opportunities, and makes writing other
2005 // transformations simpler.
2006 if (I->isCommutative())
2007 canonicalizeOperands(I);
2009 // Don't optimize floating point instructions that don't have unsafe algebra.
2010 if (I->getType()->isFPOrFPVectorTy() && !I->hasUnsafeAlgebra())
2013 // Do not reassociate boolean (i1) expressions. We want to preserve the
2014 // original order of evaluation for short-circuited comparisons that
2015 // SimplifyCFG has folded to AND/OR expressions. If the expression
2016 // is not further optimized, it is likely to be transformed back to a
2017 // short-circuited form for code gen, and the source order may have been
2018 // optimized for the most likely conditions.
2019 if (I->getType()->isIntegerTy(1))
2022 // If this is a subtract instruction which is not already in negate form,
2023 // see if we can convert it to X+-Y.
2024 if (I->getOpcode() == Instruction::Sub) {
2025 if (ShouldBreakUpSubtract(I)) {
2026 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2027 RedoInsts.insert(I);
2030 } else if (BinaryOperator::isNeg(I)) {
2031 // Otherwise, this is a negation. See if the operand is a multiply tree
2032 // and if this is not an inner node of a multiply tree.
2033 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2035 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2036 Instruction *NI = LowerNegateToMultiply(I);
2037 // If the negate was simplified, revisit the users to see if we can
2038 // reassociate further.
2039 for (User *U : NI->users()) {
2040 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2041 RedoInsts.insert(Tmp);
2043 RedoInsts.insert(I);
2048 } else if (I->getOpcode() == Instruction::FSub) {
2049 if (ShouldBreakUpSubtract(I)) {
2050 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2051 RedoInsts.insert(I);
2054 } else if (BinaryOperator::isFNeg(I)) {
2055 // Otherwise, this is a negation. See if the operand is a multiply tree
2056 // and if this is not an inner node of a multiply tree.
2057 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2059 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2060 // If the negate was simplified, revisit the users to see if we can
2061 // reassociate further.
2062 Instruction *NI = LowerNegateToMultiply(I);
2063 for (User *U : NI->users()) {
2064 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2065 RedoInsts.insert(Tmp);
2067 RedoInsts.insert(I);
2074 // If this instruction is an associative binary operator, process it.
2075 if (!I->isAssociative()) return;
2076 BinaryOperator *BO = cast<BinaryOperator>(I);
2078 // If this is an interior node of a reassociable tree, ignore it until we
2079 // get to the root of the tree, to avoid N^2 analysis.
2080 unsigned Opcode = BO->getOpcode();
2081 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2082 // During the initial run we will get to the root of the tree.
2083 // But if we get here while we are redoing instructions, there is no
2084 // guarantee that the root will be visited. So Redo later
2085 if (BO->user_back() != BO &&
2086 BO->getParent() == BO->user_back()->getParent())
2087 RedoInsts.insert(BO->user_back());
2091 // If this is an add tree that is used by a sub instruction, ignore it
2092 // until we process the subtract.
2093 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2094 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2096 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2097 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2100 ReassociateExpression(BO);
2103 void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2104 // First, walk the expression tree, linearizing the tree, collecting the
2105 // operand information.
2106 SmallVector<RepeatedValue, 8> Tree;
2107 MadeChange |= LinearizeExprTree(I, Tree);
2108 SmallVector<ValueEntry, 8> Ops;
2109 Ops.reserve(Tree.size());
2110 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2111 RepeatedValue E = Tree[i];
2112 Ops.append(E.second.getZExtValue(),
2113 ValueEntry(getRank(E.first), E.first));
2116 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2118 // Now that we have linearized the tree to a list and have gathered all of
2119 // the operands and their ranks, sort the operands by their rank. Use a
2120 // stable_sort so that values with equal ranks will have their relative
2121 // positions maintained (and so the compiler is deterministic). Note that
2122 // this sorts so that the highest ranking values end up at the beginning of
2124 std::stable_sort(Ops.begin(), Ops.end());
2126 // Now that we have the expression tree in a convenient
2127 // sorted form, optimize it globally if possible.
2128 if (Value *V = OptimizeExpression(I, Ops)) {
2130 // Self-referential expression in unreachable code.
2132 // This expression tree simplified to something that isn't a tree,
2134 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2135 I->replaceAllUsesWith(V);
2136 if (Instruction *VI = dyn_cast<Instruction>(V))
2137 VI->setDebugLoc(I->getDebugLoc());
2138 RedoInsts.insert(I);
2143 // We want to sink immediates as deeply as possible except in the case where
2144 // this is a multiply tree used only by an add, and the immediate is a -1.
2145 // In this case we reassociate to put the negation on the outside so that we
2146 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2147 if (I->hasOneUse()) {
2148 if (I->getOpcode() == Instruction::Mul &&
2149 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2150 isa<ConstantInt>(Ops.back().Op) &&
2151 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
2152 ValueEntry Tmp = Ops.pop_back_val();
2153 Ops.insert(Ops.begin(), Tmp);
2154 } else if (I->getOpcode() == Instruction::FMul &&
2155 cast<Instruction>(I->user_back())->getOpcode() ==
2156 Instruction::FAdd &&
2157 isa<ConstantFP>(Ops.back().Op) &&
2158 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2159 ValueEntry Tmp = Ops.pop_back_val();
2160 Ops.insert(Ops.begin(), Tmp);
2164 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2166 if (Ops.size() == 1) {
2168 // Self-referential expression in unreachable code.
2171 // This expression tree simplified to something that isn't a tree,
2173 I->replaceAllUsesWith(Ops[0].Op);
2174 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2175 OI->setDebugLoc(I->getDebugLoc());
2176 RedoInsts.insert(I);
2180 // Now that we ordered and optimized the expressions, splat them back into
2181 // the expression tree, removing any unneeded nodes.
2182 RewriteExprTree(I, Ops);
2185 PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) {
2186 // Get the functions basic blocks in Reverse Post Order. This order is used by
2187 // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic
2188 // blocks (it has been seen that the analysis in this pass could hang when
2189 // analysing dead basic blocks).
2190 ReversePostOrderTraversal<Function *> RPOT(&F);
2192 // Calculate the rank map for F.
2193 BuildRankMap(F, RPOT);
2196 // Traverse the same blocks that was analysed by BuildRankMap.
2197 for (BasicBlock *BI : RPOT) {
2198 assert(RankMap.count(&*BI) && "BB should be ranked.");
2199 // Optimize every instruction in the basic block.
2200 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;)
2201 if (isInstructionTriviallyDead(&*II)) {
2205 assert(II->getParent() == &*BI && "Moved to a different block!");
2209 // Make a copy of all the instructions to be redone so we can remove dead
2211 SetVector<AssertingVH<Instruction>> ToRedo(RedoInsts);
2212 // Iterate over all instructions to be reevaluated and remove trivially dead
2213 // instructions. If any operand of the trivially dead instruction becomes
2214 // dead mark it for deletion as well. Continue this process until all
2215 // trivially dead instructions have been removed.
2216 while (!ToRedo.empty()) {
2217 Instruction *I = ToRedo.pop_back_val();
2218 if (isInstructionTriviallyDead(I)) {
2219 RecursivelyEraseDeadInsts(I, ToRedo);
2224 // Now that we have removed dead instructions, we can reoptimize the
2225 // remaining instructions.
2226 while (!RedoInsts.empty()) {
2227 Instruction *I = RedoInsts.pop_back_val();
2228 if (isInstructionTriviallyDead(I))
2235 // We are done with the rank map.
2237 ValueRankMap.clear();
2240 PreservedAnalyses PA;
2241 PA.preserveSet<CFGAnalyses>();
2242 PA.preserve<GlobalsAA>();
2246 return PreservedAnalyses::all();
2250 class ReassociateLegacyPass : public FunctionPass {
2251 ReassociatePass Impl;
2253 static char ID; // Pass identification, replacement for typeid
2254 ReassociateLegacyPass() : FunctionPass(ID) {
2255 initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry());
2258 bool runOnFunction(Function &F) override {
2259 if (skipFunction(F))
2262 FunctionAnalysisManager DummyFAM;
2263 auto PA = Impl.run(F, DummyFAM);
2264 return !PA.areAllPreserved();
2267 void getAnalysisUsage(AnalysisUsage &AU) const override {
2268 AU.setPreservesCFG();
2269 AU.addPreserved<GlobalsAAWrapperPass>();
2274 char ReassociateLegacyPass::ID = 0;
2275 INITIALIZE_PASS(ReassociateLegacyPass, "reassociate",
2276 "Reassociate expressions", false, false)
2278 // Public interface to the Reassociate pass
2279 FunctionPass *llvm::createReassociatePass() {
2280 return new ReassociateLegacyPass();