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(
149 Function &F, 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 for (BasicBlock *BB : RPOT) {
159 unsigned BBRank = RankMap[BB] = ++i << 16;
161 // Walk the basic block, adding precomputed ranks for any instructions that
162 // we cannot move. This ensures that the ranks for these instructions are
163 // all different in the block.
164 for (Instruction &I : *BB)
165 if (mayBeMemoryDependent(I))
166 ValueRankMap[&I] = ++BBRank;
170 unsigned ReassociatePass::getRank(Value *V) {
171 Instruction *I = dyn_cast<Instruction>(V);
173 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument.
174 return 0; // Otherwise it's a global or constant, rank 0.
177 if (unsigned Rank = ValueRankMap[I])
178 return Rank; // Rank already known?
180 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
181 // we can reassociate expressions for code motion! Since we do not recurse
182 // for PHI nodes, we cannot have infinite recursion here, because there
183 // cannot be loops in the value graph that do not go through PHI nodes.
184 unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
185 for (unsigned i = 0, e = I->getNumOperands();
186 i != e && Rank != MaxRank; ++i)
187 Rank = std::max(Rank, getRank(I->getOperand(i)));
189 // If this is a not or neg instruction, do not count it for rank. This
190 // assures us that X and ~X will have the same rank.
191 if (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
192 !BinaryOperator::isFNeg(I))
195 DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
197 return ValueRankMap[I] = Rank;
200 // Canonicalize constants to RHS. Otherwise, sort the operands by rank.
201 void ReassociatePass::canonicalizeOperands(Instruction *I) {
202 assert(isa<BinaryOperator>(I) && "Expected binary operator.");
203 assert(I->isCommutative() && "Expected commutative operator.");
205 Value *LHS = I->getOperand(0);
206 Value *RHS = I->getOperand(1);
207 unsigned LHSRank = getRank(LHS);
208 unsigned RHSRank = getRank(RHS);
210 if (isa<Constant>(RHS))
213 if (isa<Constant>(LHS) || RHSRank < LHSRank)
214 cast<BinaryOperator>(I)->swapOperands();
217 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
218 Instruction *InsertBefore, Value *FlagsOp) {
219 if (S1->getType()->isIntOrIntVectorTy())
220 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
222 BinaryOperator *Res =
223 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
224 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
229 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
230 Instruction *InsertBefore, Value *FlagsOp) {
231 if (S1->getType()->isIntOrIntVectorTy())
232 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
234 BinaryOperator *Res =
235 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
236 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
241 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
242 Instruction *InsertBefore, Value *FlagsOp) {
243 if (S1->getType()->isIntOrIntVectorTy())
244 return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
246 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
247 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
252 /// Replace 0-X with X*-1.
253 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
254 Type *Ty = Neg->getType();
255 Constant *NegOne = Ty->isIntOrIntVectorTy() ?
256 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
258 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
259 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
261 Neg->replaceAllUsesWith(Res);
262 Res->setDebugLoc(Neg->getDebugLoc());
266 /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael
267 /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for
268 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic.
269 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every
270 /// even x in Bitwidth-bit arithmetic.
271 static unsigned CarmichaelShift(unsigned Bitwidth) {
277 /// Add the extra weight 'RHS' to the existing weight 'LHS',
278 /// reducing the combined weight using any special properties of the operation.
279 /// The existing weight LHS represents the computation X op X op ... op X where
280 /// X occurs LHS times. The combined weight represents X op X op ... op X with
281 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined
282 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even;
283 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second.
284 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) {
285 // If we were working with infinite precision arithmetic then the combined
286 // weight would be LHS + RHS. But we are using finite precision arithmetic,
287 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct
288 // for nilpotent operations and addition, but not for idempotent operations
289 // and multiplication), so it is important to correctly reduce the combined
290 // weight back into range if wrapping would be wrong.
292 // If RHS is zero then the weight didn't change.
293 if (RHS.isMinValue())
295 // If LHS is zero then the combined weight is RHS.
296 if (LHS.isMinValue()) {
300 // From this point on we know that neither LHS nor RHS is zero.
302 if (Instruction::isIdempotent(Opcode)) {
303 // Idempotent means X op X === X, so any non-zero weight is equivalent to a
304 // weight of 1. Keeping weights at zero or one also means that wrapping is
306 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
307 return; // Return a weight of 1.
309 if (Instruction::isNilpotent(Opcode)) {
310 // Nilpotent means X op X === 0, so reduce weights modulo 2.
311 assert(LHS == 1 && RHS == 1 && "Weights not reduced!");
312 LHS = 0; // 1 + 1 === 0 modulo 2.
315 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) {
316 // TODO: Reduce the weight by exploiting nsw/nuw?
321 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&
322 "Unknown associative operation!");
323 unsigned Bitwidth = LHS.getBitWidth();
324 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth
325 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth
326 // bit number x, since either x is odd in which case x^CM = 1, or x is even in
327 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples
328 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth)
329 // which by a happy accident means that they can always be represented using
331 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than
332 // the Carmichael number).
334 /// CM - The value of Carmichael's lambda function.
335 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth));
336 // Any weight W >= Threshold can be replaced with W - CM.
337 APInt Threshold = CM + Bitwidth;
338 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!");
339 // For Bitwidth 4 or more the following sum does not overflow.
341 while (LHS.uge(Threshold))
344 // To avoid problems with overflow do everything the same as above but using
346 unsigned CM = 1U << CarmichaelShift(Bitwidth);
347 unsigned Threshold = CM + Bitwidth;
348 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&
349 "Weights not reduced!");
350 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue();
351 while (Total >= Threshold)
357 typedef std::pair<Value*, APInt> RepeatedValue;
359 /// Given an associative binary expression, return the leaf
360 /// nodes in Ops along with their weights (how many times the leaf occurs). The
361 /// original expression is the same as
362 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
364 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
368 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
370 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
372 /// This routine may modify the function, in which case it returns 'true'. The
373 /// changes it makes may well be destructive, changing the value computed by 'I'
374 /// to something completely different. Thus if the routine returns 'true' then
375 /// you MUST either replace I with a new expression computed from the Ops array,
376 /// or use RewriteExprTree to put the values back in.
378 /// A leaf node is either not a binary operation of the same kind as the root
379 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
380 /// opcode), or is the same kind of binary operator but has a use which either
381 /// does not belong to the expression, or does belong to the expression but is
382 /// a leaf node. Every leaf node has at least one use that is a non-leaf node
383 /// of the expression, while for non-leaf nodes (except for the root 'I') every
384 /// use is a non-leaf node of the expression.
387 /// expression graph node names
397 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
398 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
400 /// The expression is maximal: if some instruction is a binary operator of the
401 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
402 /// then the instruction also belongs to the expression, is not a leaf node of
403 /// it, and its operands also belong to the expression (but may be leaf nodes).
405 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
406 /// order to ensure that every non-root node in the expression has *exactly one*
407 /// use by a non-leaf node of the expression. This destruction means that the
408 /// caller MUST either replace 'I' with a new expression or use something like
409 /// RewriteExprTree to put the values back in if the routine indicates that it
410 /// made a change by returning 'true'.
412 /// In the above example either the right operand of A or the left operand of B
413 /// will be replaced by undef. If it is B's operand then this gives:
417 /// + + | A, B - operand of B replaced with undef
423 /// Note that such undef operands can only be reached by passing through 'I'.
424 /// For example, if you visit operands recursively starting from a leaf node
425 /// then you will never see such an undef operand unless you get back to 'I',
426 /// which requires passing through a phi node.
428 /// Note that this routine may also mutate binary operators of the wrong type
429 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
430 /// of the expression) if it can turn them into binary operators of the right
431 /// type and thus make the expression bigger.
433 static bool LinearizeExprTree(BinaryOperator *I,
434 SmallVectorImpl<RepeatedValue> &Ops) {
435 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
436 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits();
437 unsigned Opcode = I->getOpcode();
438 assert(I->isAssociative() && I->isCommutative() &&
439 "Expected an associative and commutative operation!");
441 // Visit all operands of the expression, keeping track of their weight (the
442 // number of paths from the expression root to the operand, or if you like
443 // the number of times that operand occurs in the linearized expression).
444 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
445 // while A has weight two.
447 // Worklist of non-leaf nodes (their operands are in the expression too) along
448 // with their weights, representing a certain number of paths to the operator.
449 // If an operator occurs in the worklist multiple times then we found multiple
450 // ways to get to it.
451 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight)
452 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1)));
453 bool Changed = false;
455 // Leaves of the expression are values that either aren't the right kind of
456 // operation (eg: a constant, or a multiply in an add tree), or are, but have
457 // some uses that are not inside the expression. For example, in I = X + X,
458 // X = A + B, the value X has two uses (by I) that are in the expression. If
459 // X has any other uses, for example in a return instruction, then we consider
460 // X to be a leaf, and won't analyze it further. When we first visit a value,
461 // if it has more than one use then at first we conservatively consider it to
462 // be a leaf. Later, as the expression is explored, we may discover some more
463 // uses of the value from inside the expression. If all uses turn out to be
464 // from within the expression (and the value is a binary operator of the right
465 // kind) then the value is no longer considered to be a leaf, and its operands
468 // Leaves - Keeps track of the set of putative leaves as well as the number of
469 // paths to each leaf seen so far.
470 typedef DenseMap<Value*, APInt> LeafMap;
471 LeafMap Leaves; // Leaf -> Total weight so far.
472 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order.
475 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme.
477 while (!Worklist.empty()) {
478 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val();
479 I = P.first; // We examine the operands of this binary operator.
481 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands.
482 Value *Op = I->getOperand(OpIdx);
483 APInt Weight = P.second; // Number of paths to this operand.
484 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
485 assert(!Op->use_empty() && "No uses, so how did we get to it?!");
487 // If this is a binary operation of the right kind with only one use then
488 // add its operands to the expression.
489 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
490 assert(Visited.insert(Op).second && "Not first visit!");
491 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
492 Worklist.push_back(std::make_pair(BO, Weight));
496 // Appears to be a leaf. Is the operand already in the set of leaves?
497 LeafMap::iterator It = Leaves.find(Op);
498 if (It == Leaves.end()) {
499 // Not in the leaf map. Must be the first time we saw this operand.
500 assert(Visited.insert(Op).second && "Not first visit!");
501 if (!Op->hasOneUse()) {
502 // This value has uses not accounted for by the expression, so it is
503 // not safe to modify. Mark it as being a leaf.
504 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
505 LeafOrder.push_back(Op);
509 // No uses outside the expression, try morphing it.
510 } else if (It != Leaves.end()) {
511 // Already in the leaf map.
512 assert(Visited.count(Op) && "In leaf map but not visited!");
514 // Update the number of paths to the leaf.
515 IncorporateWeight(It->second, Weight, Opcode);
517 #if 0 // TODO: Re-enable once PR13021 is fixed.
518 // The leaf already has one use from inside the expression. As we want
519 // exactly one such use, drop this new use of the leaf.
520 assert(!Op->hasOneUse() && "Only one use, but we got here twice!");
521 I->setOperand(OpIdx, UndefValue::get(I->getType()));
524 // If the leaf is a binary operation of the right kind and we now see
525 // that its multiple original uses were in fact all by nodes belonging
526 // to the expression, then no longer consider it to be a leaf and add
527 // its operands to the expression.
528 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
529 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n");
530 Worklist.push_back(std::make_pair(BO, It->second));
536 // If we still have uses that are not accounted for by the expression
537 // then it is not safe to modify the value.
538 if (!Op->hasOneUse())
541 // No uses outside the expression, try morphing it.
543 Leaves.erase(It); // Since the value may be morphed below.
546 // At this point we have a value which, first of all, is not a binary
547 // expression of the right kind, and secondly, is only used inside the
548 // expression. This means that it can safely be modified. See if we
549 // can usefully morph it into an expression of the right kind.
550 assert((!isa<Instruction>(Op) ||
551 cast<Instruction>(Op)->getOpcode() != Opcode
552 || (isa<FPMathOperator>(Op) &&
553 !cast<Instruction>(Op)->hasUnsafeAlgebra())) &&
554 "Should have been handled above!");
555 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
557 // If this is a multiply expression, turn any internal negations into
558 // multiplies by -1 so they can be reassociated.
559 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
560 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
561 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
562 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
563 BO = LowerNegateToMultiply(BO);
564 DEBUG(dbgs() << *BO << '\n');
565 Worklist.push_back(std::make_pair(BO, Weight));
570 // Failed to morph into an expression of the right type. This really is
572 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
573 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
574 LeafOrder.push_back(Op);
579 // The leaves, repeated according to their weights, represent the linearized
580 // form of the expression.
581 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) {
582 Value *V = LeafOrder[i];
583 LeafMap::iterator It = Leaves.find(V);
584 if (It == Leaves.end())
585 // Node initially thought to be a leaf wasn't.
587 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
588 APInt Weight = It->second;
589 if (Weight.isMinValue())
590 // Leaf already output or weight reduction eliminated it.
592 // Ensure the leaf is only output once.
594 Ops.push_back(std::make_pair(V, Weight));
597 // For nilpotent operations or addition there may be no operands, for example
598 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
599 // in both cases the weight reduces to 0 causing the value to be skipped.
601 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
602 assert(Identity && "Associative operation without identity!");
603 Ops.emplace_back(Identity, APInt(Bitwidth, 1));
609 /// Now that the operands for this expression tree are
610 /// linearized and optimized, emit them in-order.
611 void ReassociatePass::RewriteExprTree(BinaryOperator *I,
612 SmallVectorImpl<ValueEntry> &Ops) {
613 assert(Ops.size() > 1 && "Single values should be used directly!");
615 // Since our optimizations should never increase the number of operations, the
616 // new expression can usually be written reusing the existing binary operators
617 // from the original expression tree, without creating any new instructions,
618 // though the rewritten expression may have a completely different topology.
619 // We take care to not change anything if the new expression will be the same
620 // as the original. If more than trivial changes (like commuting operands)
621 // were made then we are obliged to clear out any optional subclass data like
624 /// NodesToRewrite - Nodes from the original expression available for writing
625 /// the new expression into.
626 SmallVector<BinaryOperator*, 8> NodesToRewrite;
627 unsigned Opcode = I->getOpcode();
628 BinaryOperator *Op = I;
630 /// NotRewritable - The operands being written will be the leaves of the new
631 /// expression and must not be used as inner nodes (via NodesToRewrite) by
632 /// mistake. Inner nodes are always reassociable, and usually leaves are not
633 /// (if they were they would have been incorporated into the expression and so
634 /// would not be leaves), so most of the time there is no danger of this. But
635 /// in rare cases a leaf may become reassociable if an optimization kills uses
636 /// of it, or it may momentarily become reassociable during rewriting (below)
637 /// due it being removed as an operand of one of its uses. Ensure that misuse
638 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
639 /// leaves and refusing to reuse any of them as inner nodes.
640 SmallPtrSet<Value*, 8> NotRewritable;
641 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
642 NotRewritable.insert(Ops[i].Op);
644 // ExpressionChanged - Non-null if the rewritten expression differs from the
645 // original in some non-trivial way, requiring the clearing of optional flags.
646 // Flags are cleared from the operator in ExpressionChanged up to I inclusive.
647 BinaryOperator *ExpressionChanged = nullptr;
648 for (unsigned i = 0; ; ++i) {
649 // The last operation (which comes earliest in the IR) is special as both
650 // operands will come from Ops, rather than just one with the other being
652 if (i+2 == Ops.size()) {
653 Value *NewLHS = Ops[i].Op;
654 Value *NewRHS = Ops[i+1].Op;
655 Value *OldLHS = Op->getOperand(0);
656 Value *OldRHS = Op->getOperand(1);
658 if (NewLHS == OldLHS && NewRHS == OldRHS)
659 // Nothing changed, leave it alone.
662 if (NewLHS == OldRHS && NewRHS == OldLHS) {
663 // The order of the operands was reversed. Swap them.
664 DEBUG(dbgs() << "RA: " << *Op << '\n');
666 DEBUG(dbgs() << "TO: " << *Op << '\n');
672 // The new operation differs non-trivially from the original. Overwrite
673 // the old operands with the new ones.
674 DEBUG(dbgs() << "RA: " << *Op << '\n');
675 if (NewLHS != OldLHS) {
676 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
677 if (BO && !NotRewritable.count(BO))
678 NodesToRewrite.push_back(BO);
679 Op->setOperand(0, NewLHS);
681 if (NewRHS != OldRHS) {
682 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
683 if (BO && !NotRewritable.count(BO))
684 NodesToRewrite.push_back(BO);
685 Op->setOperand(1, NewRHS);
687 DEBUG(dbgs() << "TO: " << *Op << '\n');
689 ExpressionChanged = Op;
696 // Not the last operation. The left-hand side will be a sub-expression
697 // while the right-hand side will be the current element of Ops.
698 Value *NewRHS = Ops[i].Op;
699 if (NewRHS != Op->getOperand(1)) {
700 DEBUG(dbgs() << "RA: " << *Op << '\n');
701 if (NewRHS == Op->getOperand(0)) {
702 // The new right-hand side was already present as the left operand. If
703 // we are lucky then swapping the operands will sort out both of them.
706 // Overwrite with the new right-hand side.
707 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
708 if (BO && !NotRewritable.count(BO))
709 NodesToRewrite.push_back(BO);
710 Op->setOperand(1, NewRHS);
711 ExpressionChanged = Op;
713 DEBUG(dbgs() << "TO: " << *Op << '\n');
718 // Now deal with the left-hand side. If this is already an operation node
719 // from the original expression then just rewrite the rest of the expression
721 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
722 if (BO && !NotRewritable.count(BO)) {
727 // Otherwise, grab a spare node from the original expression and use that as
728 // the left-hand side. If there are no nodes left then the optimizers made
729 // an expression with more nodes than the original! This usually means that
730 // they did something stupid but it might mean that the problem was just too
731 // hard (finding the mimimal number of multiplications needed to realize a
732 // multiplication expression is NP-complete). Whatever the reason, smart or
733 // stupid, create a new node if there are none left.
734 BinaryOperator *NewOp;
735 if (NodesToRewrite.empty()) {
736 Constant *Undef = UndefValue::get(I->getType());
737 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
738 Undef, Undef, "", I);
739 if (NewOp->getType()->isFPOrFPVectorTy())
740 NewOp->setFastMathFlags(I->getFastMathFlags());
742 NewOp = NodesToRewrite.pop_back_val();
745 DEBUG(dbgs() << "RA: " << *Op << '\n');
746 Op->setOperand(0, NewOp);
747 DEBUG(dbgs() << "TO: " << *Op << '\n');
748 ExpressionChanged = Op;
754 // If the expression changed non-trivially then clear out all subclass data
755 // starting from the operator specified in ExpressionChanged, and compactify
756 // the operators to just before the expression root to guarantee that the
757 // expression tree is dominated by all of Ops.
758 if (ExpressionChanged)
760 // Preserve FastMathFlags.
761 if (isa<FPMathOperator>(I)) {
762 FastMathFlags Flags = I->getFastMathFlags();
763 ExpressionChanged->clearSubclassOptionalData();
764 ExpressionChanged->setFastMathFlags(Flags);
766 ExpressionChanged->clearSubclassOptionalData();
768 if (ExpressionChanged == I)
770 ExpressionChanged->moveBefore(I);
771 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin());
774 // Throw away any left over nodes from the original expression.
775 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i)
776 RedoInsts.insert(NodesToRewrite[i]);
779 /// Insert instructions before the instruction pointed to by BI,
780 /// that computes the negative version of the value specified. The negative
781 /// version of the value is returned, and BI is left pointing at the instruction
782 /// that should be processed next by the reassociation pass.
783 /// Also add intermediate instructions to the redo list that are modified while
784 /// pushing the negates through adds. These will be revisited to see if
785 /// additional opportunities have been exposed.
786 static Value *NegateValue(Value *V, Instruction *BI,
787 SetVector<AssertingVH<Instruction>> &ToRedo) {
788 if (Constant *C = dyn_cast<Constant>(V)) {
789 if (C->getType()->isFPOrFPVectorTy()) {
790 return ConstantExpr::getFNeg(C);
792 return ConstantExpr::getNeg(C);
796 // We are trying to expose opportunity for reassociation. One of the things
797 // that we want to do to achieve this is to push a negation as deep into an
798 // expression chain as possible, to expose the add instructions. In practice,
799 // this means that we turn this:
800 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
801 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
802 // the constants. We assume that instcombine will clean up the mess later if
803 // we introduce tons of unnecessary negation instructions.
805 if (BinaryOperator *I =
806 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
807 // Push the negates through the add.
808 I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo));
809 I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo));
810 if (I->getOpcode() == Instruction::Add) {
811 I->setHasNoUnsignedWrap(false);
812 I->setHasNoSignedWrap(false);
815 // We must move the add instruction here, because the neg instructions do
816 // not dominate the old add instruction in general. By moving it, we are
817 // assured that the neg instructions we just inserted dominate the
818 // instruction we are about to insert after them.
821 I->setName(I->getName()+".neg");
823 // Add the intermediate negates to the redo list as processing them later
824 // could expose more reassociating opportunities.
829 // Okay, we need to materialize a negated version of V with an instruction.
830 // Scan the use lists of V to see if we have one already.
831 for (User *U : V->users()) {
832 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U))
835 // We found one! Now we have to make sure that the definition dominates
836 // this use. We do this by moving it to the entry block (if it is a
837 // non-instruction value) or right after the definition. These negates will
838 // be zapped by reassociate later, so we don't need much finesse here.
839 BinaryOperator *TheNeg = cast<BinaryOperator>(U);
841 // Verify that the negate is in this function, V might be a constant expr.
842 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
845 BasicBlock::iterator InsertPt;
846 if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
847 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
848 InsertPt = II->getNormalDest()->begin();
850 InsertPt = ++InstInput->getIterator();
852 while (isa<PHINode>(InsertPt)) ++InsertPt;
854 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
856 TheNeg->moveBefore(&*InsertPt);
857 if (TheNeg->getOpcode() == Instruction::Sub) {
858 TheNeg->setHasNoUnsignedWrap(false);
859 TheNeg->setHasNoSignedWrap(false);
861 TheNeg->andIRFlags(BI);
863 ToRedo.insert(TheNeg);
867 // Insert a 'neg' instruction that subtracts the value from zero to get the
869 BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI);
870 ToRedo.insert(NewNeg);
874 /// Return true if we should break up this subtract of X-Y into (X + -Y).
875 static bool ShouldBreakUpSubtract(Instruction *Sub) {
876 // If this is a negation, we can't split it up!
877 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub))
880 // Don't breakup X - undef.
881 if (isa<UndefValue>(Sub->getOperand(1)))
884 // Don't bother to break this up unless either the LHS is an associable add or
885 // subtract or if this is only used by one.
886 Value *V0 = Sub->getOperand(0);
887 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
888 isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
890 Value *V1 = Sub->getOperand(1);
891 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
892 isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
894 Value *VB = Sub->user_back();
895 if (Sub->hasOneUse() &&
896 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
897 isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
903 /// If we have (X-Y), and if either X is an add, or if this is only used by an
904 /// add, transform this into (X+(0-Y)) to promote better reassociation.
905 static BinaryOperator *
906 BreakUpSubtract(Instruction *Sub, SetVector<AssertingVH<Instruction>> &ToRedo) {
907 // Convert a subtract into an add and a neg instruction. This allows sub
908 // instructions to be commuted with other add instructions.
910 // Calculate the negative value of Operand 1 of the sub instruction,
911 // and set it as the RHS of the add instruction we just made.
913 Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo);
914 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub);
915 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
916 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
919 // Everyone now refers to the add instruction.
920 Sub->replaceAllUsesWith(New);
921 New->setDebugLoc(Sub->getDebugLoc());
923 DEBUG(dbgs() << "Negated: " << *New << '\n');
927 /// If this is a shift of a reassociable multiply or is used by one, change
928 /// this into a multiply by a constant to assist with further reassociation.
929 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
930 Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
931 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1)));
933 BinaryOperator *Mul =
934 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl);
935 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op.
938 // Everyone now refers to the mul instruction.
939 Shl->replaceAllUsesWith(Mul);
940 Mul->setDebugLoc(Shl->getDebugLoc());
942 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
943 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
945 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
946 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
948 Mul->setHasNoSignedWrap(true);
949 Mul->setHasNoUnsignedWrap(NUW);
953 /// Scan backwards and forwards among values with the same rank as element i
954 /// to see if X exists. If X does not exist, return i. This is useful when
955 /// scanning for 'x' when we see '-x' because they both get the same rank.
956 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i,
958 unsigned XRank = Ops[i].Rank;
959 unsigned e = Ops.size();
960 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
963 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
964 if (Instruction *I2 = dyn_cast<Instruction>(X))
965 if (I1->isIdenticalTo(I2))
969 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
972 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
973 if (Instruction *I2 = dyn_cast<Instruction>(X))
974 if (I1->isIdenticalTo(I2))
980 /// Emit a tree of add instructions, summing Ops together
981 /// and returning the result. Insert the tree before I.
982 static Value *EmitAddTreeOfValues(Instruction *I,
983 SmallVectorImpl<WeakVH> &Ops){
984 if (Ops.size() == 1) return Ops.back();
986 Value *V1 = Ops.back();
988 Value *V2 = EmitAddTreeOfValues(I, Ops);
989 return CreateAdd(V2, V1, "tmp", I, I);
992 /// If V is an expression tree that is a multiplication sequence,
993 /// and if this sequence contains a multiply by Factor,
994 /// remove Factor from the tree and return the new tree.
995 Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) {
996 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1000 SmallVector<RepeatedValue, 8> Tree;
1001 MadeChange |= LinearizeExprTree(BO, Tree);
1002 SmallVector<ValueEntry, 8> Factors;
1003 Factors.reserve(Tree.size());
1004 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1005 RepeatedValue E = Tree[i];
1006 Factors.append(E.second.getZExtValue(),
1007 ValueEntry(getRank(E.first), E.first));
1010 bool FoundFactor = false;
1011 bool NeedsNegate = false;
1012 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1013 if (Factors[i].Op == Factor) {
1015 Factors.erase(Factors.begin()+i);
1019 // If this is a negative version of this factor, remove it.
1020 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1021 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1022 if (FC1->getValue() == -FC2->getValue()) {
1023 FoundFactor = NeedsNegate = true;
1024 Factors.erase(Factors.begin()+i);
1027 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1028 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1029 const APFloat &F1 = FC1->getValueAPF();
1030 APFloat F2(FC2->getValueAPF());
1032 if (F1.compare(F2) == APFloat::cmpEqual) {
1033 FoundFactor = NeedsNegate = true;
1034 Factors.erase(Factors.begin() + i);
1042 // Make sure to restore the operands to the expression tree.
1043 RewriteExprTree(BO, Factors);
1047 BasicBlock::iterator InsertPt = ++BO->getIterator();
1049 // If this was just a single multiply, remove the multiply and return the only
1050 // remaining operand.
1051 if (Factors.size() == 1) {
1052 RedoInsts.insert(BO);
1055 RewriteExprTree(BO, Factors);
1060 V = CreateNeg(V, "neg", &*InsertPt, BO);
1065 /// If V is a single-use multiply, recursively add its operands as factors,
1066 /// otherwise add V to the list of factors.
1068 /// Ops is the top-level list of add operands we're trying to factor.
1069 static void FindSingleUseMultiplyFactors(Value *V,
1070 SmallVectorImpl<Value*> &Factors,
1071 const SmallVectorImpl<ValueEntry> &Ops) {
1072 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1074 Factors.push_back(V);
1078 // Otherwise, add the LHS and RHS to the list of factors.
1079 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops);
1080 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops);
1083 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1084 /// This optimizes based on identities. If it can be reduced to a single Value,
1085 /// it is returned, otherwise the Ops list is mutated as necessary.
1086 static Value *OptimizeAndOrXor(unsigned Opcode,
1087 SmallVectorImpl<ValueEntry> &Ops) {
1088 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1089 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1090 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1091 // First, check for X and ~X in the operand list.
1092 assert(i < Ops.size());
1093 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^.
1094 Value *X = BinaryOperator::getNotArgument(Ops[i].Op);
1095 unsigned FoundX = FindInOperandList(Ops, i, X);
1097 if (Opcode == Instruction::And) // ...&X&~X = 0
1098 return Constant::getNullValue(X->getType());
1100 if (Opcode == Instruction::Or) // ...|X|~X = -1
1101 return Constant::getAllOnesValue(X->getType());
1105 // Next, check for duplicate pairs of values, which we assume are next to
1106 // each other, due to our sorting criteria.
1107 assert(i < Ops.size());
1108 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1109 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1110 // Drop duplicate values for And and Or.
1111 Ops.erase(Ops.begin()+i);
1117 // Drop pairs of values for Xor.
1118 assert(Opcode == Instruction::Xor);
1120 return Constant::getNullValue(Ops[0].Op->getType());
1123 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1131 /// Helper function of CombineXorOpnd(). It creates a bitwise-and
1132 /// instruction with the given two operands, and return the resulting
1133 /// instruction. There are two special cases: 1) if the constant operand is 0,
1134 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1136 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd,
1137 const APInt &ConstOpnd) {
1138 if (ConstOpnd != 0) {
1139 if (!ConstOpnd.isAllOnesValue()) {
1140 LLVMContext &Ctx = Opnd->getType()->getContext();
1142 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd),
1143 "and.ra", InsertBefore);
1144 I->setDebugLoc(InsertBefore->getDebugLoc());
1152 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1153 // into "R ^ C", where C would be 0, and R is a symbolic value.
1155 // If it was successful, true is returned, and the "R" and "C" is returned
1156 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1157 // and both "Res" and "ConstOpnd" remain unchanged.
1159 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1160 APInt &ConstOpnd, Value *&Res) {
1161 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1162 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1163 // = (x & ~c1) ^ (c1 ^ c2)
1164 // It is useful only when c1 == c2.
1165 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) {
1166 if (!Opnd1->getValue()->hasOneUse())
1169 const APInt &C1 = Opnd1->getConstPart();
1170 if (C1 != ConstOpnd)
1173 Value *X = Opnd1->getSymbolicPart();
1174 Res = createAndInstr(I, X, ~C1);
1175 // ConstOpnd was C2, now C1 ^ C2.
1178 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1179 RedoInsts.insert(T);
1186 // Helper function of OptimizeXor(). It tries to simplify
1187 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1190 // If it was successful, true is returned, and the "R" and "C" is returned
1191 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1192 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1193 // returned, and both "Res" and "ConstOpnd" remain unchanged.
1194 bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1,
1195 XorOpnd *Opnd2, APInt &ConstOpnd,
1197 Value *X = Opnd1->getSymbolicPart();
1198 if (X != Opnd2->getSymbolicPart())
1201 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1202 int DeadInstNum = 1;
1203 if (Opnd1->getValue()->hasOneUse())
1205 if (Opnd2->getValue()->hasOneUse())
1209 // (x | c1) ^ (x & c2)
1210 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1211 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1212 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1214 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1215 if (Opnd2->isOrExpr())
1216 std::swap(Opnd1, Opnd2);
1218 const APInt &C1 = Opnd1->getConstPart();
1219 const APInt &C2 = Opnd2->getConstPart();
1220 APInt C3((~C1) ^ C2);
1222 // Do not increase code size!
1223 if (C3 != 0 && !C3.isAllOnesValue()) {
1224 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1225 if (NewInstNum > DeadInstNum)
1229 Res = createAndInstr(I, X, C3);
1232 } else if (Opnd1->isOrExpr()) {
1233 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1235 const APInt &C1 = Opnd1->getConstPart();
1236 const APInt &C2 = Opnd2->getConstPart();
1239 // Do not increase code size
1240 if (C3 != 0 && !C3.isAllOnesValue()) {
1241 int NewInstNum = ConstOpnd != 0 ? 1 : 2;
1242 if (NewInstNum > DeadInstNum)
1246 Res = createAndInstr(I, X, C3);
1249 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1251 const APInt &C1 = Opnd1->getConstPart();
1252 const APInt &C2 = Opnd2->getConstPart();
1254 Res = createAndInstr(I, X, C3);
1257 // Put the original operands in the Redo list; hope they will be deleted
1259 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1260 RedoInsts.insert(T);
1261 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1262 RedoInsts.insert(T);
1267 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1268 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1270 Value *ReassociatePass::OptimizeXor(Instruction *I,
1271 SmallVectorImpl<ValueEntry> &Ops) {
1272 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1275 if (Ops.size() == 1)
1278 SmallVector<XorOpnd, 8> Opnds;
1279 SmallVector<XorOpnd*, 8> OpndPtrs;
1280 Type *Ty = Ops[0].Op->getType();
1281 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0);
1283 // Step 1: Convert ValueEntry to XorOpnd
1284 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1285 Value *V = Ops[i].Op;
1286 if (!isa<ConstantInt>(V)) {
1288 O.setSymbolicRank(getRank(O.getSymbolicPart()));
1291 ConstOpnd ^= cast<ConstantInt>(V)->getValue();
1294 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1295 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1296 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1297 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1298 // when new elements are added to the vector.
1299 for (unsigned i = 0, e = Opnds.size(); i != e; ++i)
1300 OpndPtrs.push_back(&Opnds[i]);
1302 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1303 // the same symbolic value cluster together. For instance, the input operand
1304 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1305 // ("x | 123", "x & 789", "y & 456").
1307 // The purpose is twofold:
1308 // 1) Cluster together the operands sharing the same symbolic-value.
1309 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
1310 // could potentially shorten crital path, and expose more loop-invariants.
1311 // Note that values' rank are basically defined in RPO order (FIXME).
1312 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1313 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1314 // "z" in the order of X-Y-Z is better than any other orders.
1315 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(),
1316 [](XorOpnd *LHS, XorOpnd *RHS) {
1317 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1320 // Step 3: Combine adjacent operands
1321 XorOpnd *PrevOpnd = nullptr;
1322 bool Changed = false;
1323 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1324 XorOpnd *CurrOpnd = OpndPtrs[i];
1325 // The combined value
1328 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1329 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) {
1332 *CurrOpnd = XorOpnd(CV);
1334 CurrOpnd->Invalidate();
1339 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1340 PrevOpnd = CurrOpnd;
1344 // step 3.2: When previous and current operands share the same symbolic
1345 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1347 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1348 // Remove previous operand
1349 PrevOpnd->Invalidate();
1351 *CurrOpnd = XorOpnd(CV);
1352 PrevOpnd = CurrOpnd;
1354 CurrOpnd->Invalidate();
1361 // Step 4: Reassemble the Ops
1364 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) {
1365 XorOpnd &O = Opnds[i];
1368 ValueEntry VE(getRank(O.getValue()), O.getValue());
1371 if (ConstOpnd != 0) {
1372 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd);
1373 ValueEntry VE(getRank(C), C);
1376 int Sz = Ops.size();
1378 return Ops.back().Op;
1380 assert(ConstOpnd == 0);
1381 return ConstantInt::get(Ty->getContext(), ConstOpnd);
1388 /// Optimize a series of operands to an 'add' instruction. This
1389 /// optimizes based on identities. If it can be reduced to a single Value, it
1390 /// is returned, otherwise the Ops list is mutated as necessary.
1391 Value *ReassociatePass::OptimizeAdd(Instruction *I,
1392 SmallVectorImpl<ValueEntry> &Ops) {
1393 // Scan the operand lists looking for X and -X pairs. If we find any, we
1394 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1396 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1398 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1399 Value *TheOp = Ops[i].Op;
1400 // Check to see if we've seen this operand before. If so, we factor all
1401 // instances of the operand together. Due to our sorting criteria, we know
1402 // that these need to be next to each other in the vector.
1403 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1404 // Rescan the list, remove all instances of this operand from the expr.
1405 unsigned NumFound = 0;
1407 Ops.erase(Ops.begin()+i);
1409 } while (i != Ops.size() && Ops[i].Op == TheOp);
1411 DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
1414 // Insert a new multiply.
1415 Type *Ty = TheOp->getType();
1416 Constant *C = Ty->isIntOrIntVectorTy() ?
1417 ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1418 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
1420 // Now that we have inserted a multiply, optimize it. This allows us to
1421 // handle cases that require multiple factoring steps, such as this:
1422 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1423 RedoInsts.insert(Mul);
1425 // If every add operand was a duplicate, return the multiply.
1429 // Otherwise, we had some input that didn't have the dupe, such as
1430 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1431 // things being added by this operation.
1432 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1439 // Check for X and -X or X and ~X in the operand list.
1440 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) &&
1441 !BinaryOperator::isNot(TheOp))
1445 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))
1446 X = BinaryOperator::getNegArgument(TheOp);
1447 else if (BinaryOperator::isNot(TheOp))
1448 X = BinaryOperator::getNotArgument(TheOp);
1450 unsigned FoundX = FindInOperandList(Ops, i, X);
1454 // Remove X and -X from the operand list.
1455 if (Ops.size() == 2 &&
1456 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)))
1457 return Constant::getNullValue(X->getType());
1459 // Remove X and ~X from the operand list.
1460 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp))
1461 return Constant::getAllOnesValue(X->getType());
1463 Ops.erase(Ops.begin()+i);
1467 --i; // Need to back up an extra one.
1468 Ops.erase(Ops.begin()+FoundX);
1470 --i; // Revisit element.
1471 e -= 2; // Removed two elements.
1473 // if X and ~X we append -1 to the operand list.
1474 if (BinaryOperator::isNot(TheOp)) {
1475 Value *V = Constant::getAllOnesValue(X->getType());
1476 Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1481 // Scan the operand list, checking to see if there are any common factors
1482 // between operands. Consider something like A*A+A*B*C+D. We would like to
1483 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1484 // To efficiently find this, we count the number of times a factor occurs
1485 // for any ADD operands that are MULs.
1486 DenseMap<Value*, unsigned> FactorOccurrences;
1488 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1489 // where they are actually the same multiply.
1490 unsigned MaxOcc = 0;
1491 Value *MaxOccVal = nullptr;
1492 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1493 BinaryOperator *BOp =
1494 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1498 // Compute all of the factors of this added value.
1499 SmallVector<Value*, 8> Factors;
1500 FindSingleUseMultiplyFactors(BOp, Factors, Ops);
1501 assert(Factors.size() > 1 && "Bad linearize!");
1503 // Add one to FactorOccurrences for each unique factor in this op.
1504 SmallPtrSet<Value*, 8> Duplicates;
1505 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1506 Value *Factor = Factors[i];
1507 if (!Duplicates.insert(Factor).second)
1510 unsigned Occ = ++FactorOccurrences[Factor];
1516 // If Factor is a negative constant, add the negated value as a factor
1517 // because we can percolate the negate out. Watch for minint, which
1518 // cannot be positivified.
1519 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1520 if (CI->isNegative() && !CI->isMinValue(true)) {
1521 Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1522 assert(!Duplicates.count(Factor) &&
1523 "Shouldn't have two constant factors, missed a canonicalize");
1524 unsigned Occ = ++FactorOccurrences[Factor];
1530 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1531 if (CF->isNegative()) {
1532 APFloat F(CF->getValueAPF());
1534 Factor = ConstantFP::get(CF->getContext(), F);
1535 assert(!Duplicates.count(Factor) &&
1536 "Shouldn't have two constant factors, missed a canonicalize");
1537 unsigned Occ = ++FactorOccurrences[Factor];
1547 // If any factor occurred more than one time, we can pull it out.
1549 DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
1552 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1553 // this, we could otherwise run into situations where removing a factor
1554 // from an expression will drop a use of maxocc, and this can cause
1555 // RemoveFactorFromExpression on successive values to behave differently.
1556 Instruction *DummyInst =
1557 I->getType()->isIntOrIntVectorTy()
1558 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1559 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1561 SmallVector<WeakVH, 4> NewMulOps;
1562 for (unsigned i = 0; i != Ops.size(); ++i) {
1563 // Only try to remove factors from expressions we're allowed to.
1564 BinaryOperator *BOp =
1565 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1569 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1570 // The factorized operand may occur several times. Convert them all in
1572 for (unsigned j = Ops.size(); j != i;) {
1574 if (Ops[j].Op == Ops[i].Op) {
1575 NewMulOps.push_back(V);
1576 Ops.erase(Ops.begin()+j);
1583 // No need for extra uses anymore.
1586 unsigned NumAddedValues = NewMulOps.size();
1587 Value *V = EmitAddTreeOfValues(I, NewMulOps);
1589 // Now that we have inserted the add tree, optimize it. This allows us to
1590 // handle cases that require multiple factoring steps, such as this:
1591 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1592 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1593 (void)NumAddedValues;
1594 if (Instruction *VI = dyn_cast<Instruction>(V))
1595 RedoInsts.insert(VI);
1597 // Create the multiply.
1598 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I);
1600 // Rerun associate on the multiply in case the inner expression turned into
1601 // a multiply. We want to make sure that we keep things in canonical form.
1602 RedoInsts.insert(V2);
1604 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1605 // entire result expression is just the multiply "A*(B+C)".
1609 // Otherwise, we had some input that didn't have the factor, such as
1610 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1611 // things being added by this operation.
1612 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1618 /// \brief Build up a vector of value/power pairs factoring a product.
1620 /// Given a series of multiplication operands, build a vector of factors and
1621 /// the powers each is raised to when forming the final product. Sort them in
1622 /// the order of descending power.
1624 /// (x*x) -> [(x, 2)]
1625 /// ((x*x)*x) -> [(x, 3)]
1626 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1628 /// \returns Whether any factors have a power greater than one.
1629 bool ReassociatePass::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1630 SmallVectorImpl<Factor> &Factors) {
1631 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1632 // Compute the sum of powers of simplifiable factors.
1633 unsigned FactorPowerSum = 0;
1634 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1635 Value *Op = Ops[Idx-1].Op;
1637 // Count the number of occurrences of this value.
1639 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1641 // Track for simplification all factors which occur 2 or more times.
1643 FactorPowerSum += Count;
1646 // We can only simplify factors if the sum of the powers of our simplifiable
1647 // factors is 4 or higher. When that is the case, we will *always* have
1648 // a simplification. This is an important invariant to prevent cyclicly
1649 // trying to simplify already minimal formations.
1650 if (FactorPowerSum < 4)
1653 // Now gather the simplifiable factors, removing them from Ops.
1655 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1656 Value *Op = Ops[Idx-1].Op;
1658 // Count the number of occurrences of this value.
1660 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1664 // Move an even number of occurrences to Factors.
1667 FactorPowerSum += Count;
1668 Factors.push_back(Factor(Op, Count));
1669 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1672 // None of the adjustments above should have reduced the sum of factor powers
1673 // below our mininum of '4'.
1674 assert(FactorPowerSum >= 4);
1676 std::stable_sort(Factors.begin(), Factors.end(),
1677 [](const Factor &LHS, const Factor &RHS) {
1678 return LHS.Power > RHS.Power;
1683 /// \brief Build a tree of multiplies, computing the product of Ops.
1684 static Value *buildMultiplyTree(IRBuilder<> &Builder,
1685 SmallVectorImpl<Value*> &Ops) {
1686 if (Ops.size() == 1)
1689 Value *LHS = Ops.pop_back_val();
1691 if (LHS->getType()->isIntOrIntVectorTy())
1692 LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1694 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1695 } while (!Ops.empty());
1700 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1702 /// Given a vector of values raised to various powers, where no two values are
1703 /// equal and the powers are sorted in decreasing order, compute the minimal
1704 /// DAG of multiplies to compute the final product, and return that product
1707 ReassociatePass::buildMinimalMultiplyDAG(IRBuilder<> &Builder,
1708 SmallVectorImpl<Factor> &Factors) {
1709 assert(Factors[0].Power);
1710 SmallVector<Value *, 4> OuterProduct;
1711 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1712 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1713 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1718 // We want to multiply across all the factors with the same power so that
1719 // we can raise them to that power as a single entity. Build a mini tree
1721 SmallVector<Value *, 4> InnerProduct;
1722 InnerProduct.push_back(Factors[LastIdx].Base);
1724 InnerProduct.push_back(Factors[Idx].Base);
1726 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1728 // Reset the base value of the first factor to the new expression tree.
1729 // We'll remove all the factors with the same power in a second pass.
1730 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1731 if (Instruction *MI = dyn_cast<Instruction>(M))
1732 RedoInsts.insert(MI);
1736 // Unique factors with equal powers -- we've folded them into the first one's
1738 Factors.erase(std::unique(Factors.begin(), Factors.end(),
1739 [](const Factor &LHS, const Factor &RHS) {
1740 return LHS.Power == RHS.Power;
1744 // Iteratively collect the base of each factor with an add power into the
1745 // outer product, and halve each power in preparation for squaring the
1747 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) {
1748 if (Factors[Idx].Power & 1)
1749 OuterProduct.push_back(Factors[Idx].Base);
1750 Factors[Idx].Power >>= 1;
1752 if (Factors[0].Power) {
1753 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1754 OuterProduct.push_back(SquareRoot);
1755 OuterProduct.push_back(SquareRoot);
1757 if (OuterProduct.size() == 1)
1758 return OuterProduct.front();
1760 Value *V = buildMultiplyTree(Builder, OuterProduct);
1764 Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
1765 SmallVectorImpl<ValueEntry> &Ops) {
1766 // We can only optimize the multiplies when there is a chain of more than
1767 // three, such that a balanced tree might require fewer total multiplies.
1771 // Try to turn linear trees of multiplies without other uses of the
1772 // intermediate stages into minimal multiply DAGs with perfect sub-expression
1774 SmallVector<Factor, 4> Factors;
1775 if (!collectMultiplyFactors(Ops, Factors))
1776 return nullptr; // All distinct factors, so nothing left for us to do.
1778 IRBuilder<> Builder(I);
1779 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1783 ValueEntry NewEntry = ValueEntry(getRank(V), V);
1784 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry);
1788 Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
1789 SmallVectorImpl<ValueEntry> &Ops) {
1790 // Now that we have the linearized expression tree, try to optimize it.
1791 // Start by folding any constants that we found.
1792 Constant *Cst = nullptr;
1793 unsigned Opcode = I->getOpcode();
1794 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) {
1795 Constant *C = cast<Constant>(Ops.pop_back_val().Op);
1796 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C;
1798 // If there was nothing but constants then we are done.
1802 // Put the combined constant back at the end of the operand list, except if
1803 // there is no point. For example, an add of 0 gets dropped here, while a
1804 // multiplication by zero turns the whole expression into zero.
1805 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1806 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1808 Ops.push_back(ValueEntry(0, Cst));
1811 if (Ops.size() == 1) return Ops[0].Op;
1813 // Handle destructive annihilation due to identities between elements in the
1814 // argument list here.
1815 unsigned NumOps = Ops.size();
1818 case Instruction::And:
1819 case Instruction::Or:
1820 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1824 case Instruction::Xor:
1825 if (Value *Result = OptimizeXor(I, Ops))
1829 case Instruction::Add:
1830 case Instruction::FAdd:
1831 if (Value *Result = OptimizeAdd(I, Ops))
1835 case Instruction::Mul:
1836 case Instruction::FMul:
1837 if (Value *Result = OptimizeMul(I, Ops))
1842 if (Ops.size() != NumOps)
1843 return OptimizeExpression(I, Ops);
1847 // Remove dead instructions and if any operands are trivially dead add them to
1848 // Insts so they will be removed as well.
1849 void ReassociatePass::RecursivelyEraseDeadInsts(
1850 Instruction *I, SetVector<AssertingVH<Instruction>> &Insts) {
1851 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1852 SmallVector<Value *, 4> Ops(I->op_begin(), I->op_end());
1853 ValueRankMap.erase(I);
1855 RedoInsts.remove(I);
1856 I->eraseFromParent();
1858 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1859 if (OpInst->use_empty())
1860 Insts.insert(OpInst);
1863 /// Zap the given instruction, adding interesting operands to the work list.
1864 void ReassociatePass::EraseInst(Instruction *I) {
1865 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1866 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
1867 // Erase the dead instruction.
1868 ValueRankMap.erase(I);
1869 RedoInsts.remove(I);
1870 I->eraseFromParent();
1871 // Optimize its operands.
1872 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1873 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1874 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) {
1875 // If this is a node in an expression tree, climb to the expression root
1876 // and add that since that's where optimization actually happens.
1877 unsigned Opcode = Op->getOpcode();
1878 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1879 Visited.insert(Op).second)
1880 Op = Op->user_back();
1881 RedoInsts.insert(Op);
1885 // Canonicalize expressions of the following form:
1886 // x + (-Constant * y) -> x - (Constant * y)
1887 // x - (-Constant * y) -> x + (Constant * y)
1888 Instruction *ReassociatePass::canonicalizeNegConstExpr(Instruction *I) {
1889 if (!I->hasOneUse() || I->getType()->isVectorTy())
1892 // Must be a fmul or fdiv instruction.
1893 unsigned Opcode = I->getOpcode();
1894 if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv)
1897 auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0));
1898 auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1));
1900 // Both operands are constant, let it get constant folded away.
1904 ConstantFP *CF = C0 ? C0 : C1;
1906 // Must have one constant operand.
1910 // Must be a negative ConstantFP.
1911 if (!CF->isNegative())
1914 // User must be a binary operator with one or more uses.
1915 Instruction *User = I->user_back();
1916 if (!isa<BinaryOperator>(User) || !User->hasNUsesOrMore(1))
1919 unsigned UserOpcode = User->getOpcode();
1920 if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub)
1923 // Subtraction is not commutative. Explicitly, the following transform is
1924 // not valid: (-Constant * y) - x -> x + (Constant * y)
1925 if (!User->isCommutative() && User->getOperand(1) != I)
1928 // Change the sign of the constant.
1929 APFloat Val = CF->getValueAPF();
1931 I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val));
1933 // Canonicalize I to RHS to simplify the next bit of logic. E.g.,
1934 // ((-Const*y) + x) -> (x + (-Const*y)).
1935 if (User->getOperand(0) == I && User->isCommutative())
1936 cast<BinaryOperator>(User)->swapOperands();
1938 Value *Op0 = User->getOperand(0);
1939 Value *Op1 = User->getOperand(1);
1941 switch (UserOpcode) {
1943 llvm_unreachable("Unexpected Opcode!");
1944 case Instruction::FAdd:
1945 NI = BinaryOperator::CreateFSub(Op0, Op1);
1946 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1948 case Instruction::FSub:
1949 NI = BinaryOperator::CreateFAdd(Op0, Op1);
1950 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags());
1954 NI->insertBefore(User);
1955 NI->setName(User->getName());
1956 User->replaceAllUsesWith(NI);
1957 NI->setDebugLoc(I->getDebugLoc());
1958 RedoInsts.insert(I);
1963 /// Inspect and optimize the given instruction. Note that erasing
1964 /// instructions is not allowed.
1965 void ReassociatePass::OptimizeInst(Instruction *I) {
1966 // Only consider operations that we understand.
1967 if (!isa<BinaryOperator>(I))
1970 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
1971 // If an operand of this shift is a reassociable multiply, or if the shift
1972 // is used by a reassociable multiply or add, turn into a multiply.
1973 if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
1975 (isReassociableOp(I->user_back(), Instruction::Mul) ||
1976 isReassociableOp(I->user_back(), Instruction::Add)))) {
1977 Instruction *NI = ConvertShiftToMul(I);
1978 RedoInsts.insert(I);
1983 // Canonicalize negative constants out of expressions.
1984 if (Instruction *Res = canonicalizeNegConstExpr(I))
1987 // Commute binary operators, to canonicalize the order of their operands.
1988 // This can potentially expose more CSE opportunities, and makes writing other
1989 // transformations simpler.
1990 if (I->isCommutative())
1991 canonicalizeOperands(I);
1993 // TODO: We should optimize vector Xor instructions, but they are
1994 // currently unsupported.
1995 if (I->getType()->isVectorTy() && I->getOpcode() == Instruction::Xor)
1998 // Don't optimize floating point instructions that don't have unsafe algebra.
1999 if (I->getType()->isFPOrFPVectorTy() && !I->hasUnsafeAlgebra())
2002 // Do not reassociate boolean (i1) expressions. We want to preserve the
2003 // original order of evaluation for short-circuited comparisons that
2004 // SimplifyCFG has folded to AND/OR expressions. If the expression
2005 // is not further optimized, it is likely to be transformed back to a
2006 // short-circuited form for code gen, and the source order may have been
2007 // optimized for the most likely conditions.
2008 if (I->getType()->isIntegerTy(1))
2011 // If this is a subtract instruction which is not already in negate form,
2012 // see if we can convert it to X+-Y.
2013 if (I->getOpcode() == Instruction::Sub) {
2014 if (ShouldBreakUpSubtract(I)) {
2015 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2016 RedoInsts.insert(I);
2019 } else if (BinaryOperator::isNeg(I)) {
2020 // Otherwise, this is a negation. See if the operand is a multiply tree
2021 // and if this is not an inner node of a multiply tree.
2022 if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2024 !isReassociableOp(I->user_back(), Instruction::Mul))) {
2025 Instruction *NI = LowerNegateToMultiply(I);
2026 // If the negate was simplified, revisit the users to see if we can
2027 // reassociate further.
2028 for (User *U : NI->users()) {
2029 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2030 RedoInsts.insert(Tmp);
2032 RedoInsts.insert(I);
2037 } else if (I->getOpcode() == Instruction::FSub) {
2038 if (ShouldBreakUpSubtract(I)) {
2039 Instruction *NI = BreakUpSubtract(I, RedoInsts);
2040 RedoInsts.insert(I);
2043 } else if (BinaryOperator::isFNeg(I)) {
2044 // Otherwise, this is a negation. See if the operand is a multiply tree
2045 // and if this is not an inner node of a multiply tree.
2046 if (isReassociableOp(I->getOperand(1), Instruction::FMul) &&
2048 !isReassociableOp(I->user_back(), Instruction::FMul))) {
2049 // If the negate was simplified, revisit the users to see if we can
2050 // reassociate further.
2051 Instruction *NI = LowerNegateToMultiply(I);
2052 for (User *U : NI->users()) {
2053 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2054 RedoInsts.insert(Tmp);
2056 RedoInsts.insert(I);
2063 // If this instruction is an associative binary operator, process it.
2064 if (!I->isAssociative()) return;
2065 BinaryOperator *BO = cast<BinaryOperator>(I);
2067 // If this is an interior node of a reassociable tree, ignore it until we
2068 // get to the root of the tree, to avoid N^2 analysis.
2069 unsigned Opcode = BO->getOpcode();
2070 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2071 // During the initial run we will get to the root of the tree.
2072 // But if we get here while we are redoing instructions, there is no
2073 // guarantee that the root will be visited. So Redo later
2074 if (BO->user_back() != BO &&
2075 BO->getParent() == BO->user_back()->getParent())
2076 RedoInsts.insert(BO->user_back());
2080 // If this is an add tree that is used by a sub instruction, ignore it
2081 // until we process the subtract.
2082 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2083 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2085 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2086 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2089 ReassociateExpression(BO);
2092 void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2093 // First, walk the expression tree, linearizing the tree, collecting the
2094 // operand information.
2095 SmallVector<RepeatedValue, 8> Tree;
2096 MadeChange |= LinearizeExprTree(I, Tree);
2097 SmallVector<ValueEntry, 8> Ops;
2098 Ops.reserve(Tree.size());
2099 for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
2100 RepeatedValue E = Tree[i];
2101 Ops.append(E.second.getZExtValue(),
2102 ValueEntry(getRank(E.first), E.first));
2105 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2107 // Now that we have linearized the tree to a list and have gathered all of
2108 // the operands and their ranks, sort the operands by their rank. Use a
2109 // stable_sort so that values with equal ranks will have their relative
2110 // positions maintained (and so the compiler is deterministic). Note that
2111 // this sorts so that the highest ranking values end up at the beginning of
2113 std::stable_sort(Ops.begin(), Ops.end());
2115 // Now that we have the expression tree in a convenient
2116 // sorted form, optimize it globally if possible.
2117 if (Value *V = OptimizeExpression(I, Ops)) {
2119 // Self-referential expression in unreachable code.
2121 // This expression tree simplified to something that isn't a tree,
2123 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2124 I->replaceAllUsesWith(V);
2125 if (Instruction *VI = dyn_cast<Instruction>(V))
2126 VI->setDebugLoc(I->getDebugLoc());
2127 RedoInsts.insert(I);
2132 // We want to sink immediates as deeply as possible except in the case where
2133 // this is a multiply tree used only by an add, and the immediate is a -1.
2134 // In this case we reassociate to put the negation on the outside so that we
2135 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2136 if (I->hasOneUse()) {
2137 if (I->getOpcode() == Instruction::Mul &&
2138 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2139 isa<ConstantInt>(Ops.back().Op) &&
2140 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) {
2141 ValueEntry Tmp = Ops.pop_back_val();
2142 Ops.insert(Ops.begin(), Tmp);
2143 } else if (I->getOpcode() == Instruction::FMul &&
2144 cast<Instruction>(I->user_back())->getOpcode() ==
2145 Instruction::FAdd &&
2146 isa<ConstantFP>(Ops.back().Op) &&
2147 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2148 ValueEntry Tmp = Ops.pop_back_val();
2149 Ops.insert(Ops.begin(), Tmp);
2153 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2155 if (Ops.size() == 1) {
2157 // Self-referential expression in unreachable code.
2160 // This expression tree simplified to something that isn't a tree,
2162 I->replaceAllUsesWith(Ops[0].Op);
2163 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2164 OI->setDebugLoc(I->getDebugLoc());
2165 RedoInsts.insert(I);
2169 // Now that we ordered and optimized the expressions, splat them back into
2170 // the expression tree, removing any unneeded nodes.
2171 RewriteExprTree(I, Ops);
2174 PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) {
2175 // Reassociate needs for each instruction to have its operands already
2176 // processed, so we first perform a RPOT of the basic blocks so that
2177 // when we process a basic block, all its dominators have been processed
2179 ReversePostOrderTraversal<Function *> RPOT(&F);
2180 BuildRankMap(F, RPOT);
2183 for (BasicBlock *BI : RPOT) {
2184 // Use a worklist to keep track of which instructions have been processed
2185 // (and which insts won't be optimized again) so when redoing insts,
2186 // optimize insts rightaway which won't be processed later.
2187 SmallSet<Instruction *, 8> Worklist;
2189 // Insert all instructions in the BB
2190 for (Instruction &I : *BI)
2191 Worklist.insert(&I);
2193 // Optimize every instruction in the basic block.
2194 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;) {
2195 // This instruction has been processed.
2196 Worklist.erase(&*II);
2197 if (isInstructionTriviallyDead(&*II)) {
2201 assert(II->getParent() == &*BI && "Moved to a different block!");
2205 // If the above optimizations produced new instructions to optimize or
2206 // made modifications which need to be redone, do them now if they won't
2207 // be handled later.
2208 while (!RedoInsts.empty()) {
2209 Instruction *I = RedoInsts.pop_back_val();
2210 // Process instructions that won't be processed later, either
2211 // inside the block itself or in another basic block (based on rank),
2212 // since these will be processed later.
2213 if ((I->getParent() != BI || !Worklist.count(I)) &&
2214 RankMap[I->getParent()] <= RankMap[BI]) {
2215 if (isInstructionTriviallyDead(I))
2224 // We are done with the rank map.
2226 ValueRankMap.clear();
2229 // FIXME: Reassociate should also 'preserve the CFG'.
2230 // The new pass manager has currently no way to do it.
2231 auto PA = PreservedAnalyses();
2232 PA.preserve<GlobalsAA>();
2236 return PreservedAnalyses::all();
2240 class ReassociateLegacyPass : public FunctionPass {
2241 ReassociatePass Impl;
2243 static char ID; // Pass identification, replacement for typeid
2244 ReassociateLegacyPass() : FunctionPass(ID) {
2245 initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry());
2248 bool runOnFunction(Function &F) override {
2249 if (skipFunction(F))
2252 FunctionAnalysisManager DummyFAM;
2253 auto PA = Impl.run(F, DummyFAM);
2254 return !PA.areAllPreserved();
2257 void getAnalysisUsage(AnalysisUsage &AU) const override {
2258 AU.setPreservesCFG();
2259 AU.addPreserved<GlobalsAAWrapperPass>();
2264 char ReassociateLegacyPass::ID = 0;
2265 INITIALIZE_PASS(ReassociateLegacyPass, "reassociate",
2266 "Reassociate expressions", false, false)
2268 // Public interface to the Reassociate pass
2269 FunctionPass *llvm::createReassociatePass() {
2270 return new ReassociateLegacyPass();