xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/Reassociate.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- Reassociate.cpp - Reassociate binary expressions -------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass reassociates commutative expressions in an order that is designed
10 // to promote better constant propagation, GCSE, LICM, PRE, etc.
11 //
12 // For example: 4 + (x + 5) -> x + (4 + 5)
13 //
14 // In the implementation of this algorithm, constants are assigned rank = 0,
15 // function arguments are rank = 1, and other values are assigned ranks
16 // corresponding to the reverse post order traversal of current function
17 // (starting at 2), which effectively gives values in deep loops higher rank
18 // than values not in loops.
19 //
20 //===----------------------------------------------------------------------===//
21 
22 #include "llvm/Transforms/Scalar/Reassociate.h"
23 #include "llvm/ADT/APFloat.h"
24 #include "llvm/ADT/APInt.h"
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/PostOrderIterator.h"
27 #include "llvm/ADT/SmallPtrSet.h"
28 #include "llvm/ADT/SmallSet.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/BasicAliasAnalysis.h"
32 #include "llvm/Analysis/ConstantFolding.h"
33 #include "llvm/Analysis/GlobalsModRef.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/Argument.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/CFG.h"
38 #include "llvm/IR/Constant.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/InstrTypes.h"
43 #include "llvm/IR/Instruction.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/Operator.h"
46 #include "llvm/IR/PassManager.h"
47 #include "llvm/IR/PatternMatch.h"
48 #include "llvm/IR/Type.h"
49 #include "llvm/IR/User.h"
50 #include "llvm/IR/Value.h"
51 #include "llvm/IR/ValueHandle.h"
52 #include "llvm/InitializePasses.h"
53 #include "llvm/Pass.h"
54 #include "llvm/Support/Casting.h"
55 #include "llvm/Support/CommandLine.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/Transforms/Scalar.h"
59 #include "llvm/Transforms/Utils/Local.h"
60 #include <algorithm>
61 #include <cassert>
62 #include <utility>
63 
64 using namespace llvm;
65 using namespace reassociate;
66 using namespace PatternMatch;
67 
68 #define DEBUG_TYPE "reassociate"
69 
70 STATISTIC(NumChanged, "Number of insts reassociated");
71 STATISTIC(NumAnnihil, "Number of expr tree annihilated");
72 STATISTIC(NumFactor , "Number of multiplies factored");
73 
74 static cl::opt<bool>
75     UseCSELocalOpt(DEBUG_TYPE "-use-cse-local",
76                    cl::desc("Only reorder expressions within a basic block "
77                             "when exposing CSE opportunities"),
78                    cl::init(true), cl::Hidden);
79 
80 #ifndef NDEBUG
81 /// Print out the expression identified in the Ops list.
PrintOps(Instruction * I,const SmallVectorImpl<ValueEntry> & Ops)82 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
83   Module *M = I->getModule();
84   dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
85        << *Ops[0].Op->getType() << '\t';
86   for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
87     dbgs() << "[ ";
88     Ops[i].Op->printAsOperand(dbgs(), false, M);
89     dbgs() << ", #" << Ops[i].Rank << "] ";
90   }
91 }
92 #endif
93 
94 /// Utility class representing a non-constant Xor-operand. We classify
95 /// non-constant Xor-Operands into two categories:
96 ///  C1) The operand is in the form "X & C", where C is a constant and C != ~0
97 ///  C2)
98 ///    C2.1) The operand is in the form of "X | C", where C is a non-zero
99 ///          constant.
100 ///    C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
101 ///          operand as "E | 0"
102 class llvm::reassociate::XorOpnd {
103 public:
104   XorOpnd(Value *V);
105 
isInvalid() const106   bool isInvalid() const { return SymbolicPart == nullptr; }
isOrExpr() const107   bool isOrExpr() const { return isOr; }
getValue() const108   Value *getValue() const { return OrigVal; }
getSymbolicPart() const109   Value *getSymbolicPart() const { return SymbolicPart; }
getSymbolicRank() const110   unsigned getSymbolicRank() const { return SymbolicRank; }
getConstPart() const111   const APInt &getConstPart() const { return ConstPart; }
112 
Invalidate()113   void Invalidate() { SymbolicPart = OrigVal = nullptr; }
setSymbolicRank(unsigned R)114   void setSymbolicRank(unsigned R) { SymbolicRank = R; }
115 
116 private:
117   Value *OrigVal;
118   Value *SymbolicPart;
119   APInt ConstPart;
120   unsigned SymbolicRank;
121   bool isOr;
122 };
123 
XorOpnd(Value * V)124 XorOpnd::XorOpnd(Value *V) {
125   assert(!isa<ConstantInt>(V) && "No ConstantInt");
126   OrigVal = V;
127   Instruction *I = dyn_cast<Instruction>(V);
128   SymbolicRank = 0;
129 
130   if (I && (I->getOpcode() == Instruction::Or ||
131             I->getOpcode() == Instruction::And)) {
132     Value *V0 = I->getOperand(0);
133     Value *V1 = I->getOperand(1);
134     const APInt *C;
135     if (match(V0, m_APInt(C)))
136       std::swap(V0, V1);
137 
138     if (match(V1, m_APInt(C))) {
139       ConstPart = *C;
140       SymbolicPart = V0;
141       isOr = (I->getOpcode() == Instruction::Or);
142       return;
143     }
144   }
145 
146   // view the operand as "V | 0"
147   SymbolicPart = V;
148   ConstPart = APInt::getZero(V->getType()->getScalarSizeInBits());
149   isOr = true;
150 }
151 
152 /// Return true if I is an instruction with the FastMathFlags that are needed
153 /// for general reassociation set.  This is not the same as testing
154 /// Instruction::isAssociative() because it includes operations like fsub.
155 /// (This routine is only intended to be called for floating-point operations.)
hasFPAssociativeFlags(Instruction * I)156 static bool hasFPAssociativeFlags(Instruction *I) {
157   assert(I && isa<FPMathOperator>(I) && "Should only check FP ops");
158   return I->hasAllowReassoc() && I->hasNoSignedZeros();
159 }
160 
161 /// Return true if V is an instruction of the specified opcode and if it
162 /// only has one use.
isReassociableOp(Value * V,unsigned Opcode)163 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
164   auto *BO = dyn_cast<BinaryOperator>(V);
165   if (BO && BO->hasOneUse() && BO->getOpcode() == Opcode)
166     if (!isa<FPMathOperator>(BO) || hasFPAssociativeFlags(BO))
167       return BO;
168   return nullptr;
169 }
170 
isReassociableOp(Value * V,unsigned Opcode1,unsigned Opcode2)171 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
172                                         unsigned Opcode2) {
173   auto *BO = dyn_cast<BinaryOperator>(V);
174   if (BO && BO->hasOneUse() &&
175       (BO->getOpcode() == Opcode1 || BO->getOpcode() == Opcode2))
176     if (!isa<FPMathOperator>(BO) || hasFPAssociativeFlags(BO))
177       return BO;
178   return nullptr;
179 }
180 
BuildRankMap(Function & F,ReversePostOrderTraversal<Function * > & RPOT)181 void ReassociatePass::BuildRankMap(Function &F,
182                                    ReversePostOrderTraversal<Function*> &RPOT) {
183   unsigned Rank = 2;
184 
185   // Assign distinct ranks to function arguments.
186   for (auto &Arg : F.args()) {
187     ValueRankMap[&Arg] = ++Rank;
188     LLVM_DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank
189                       << "\n");
190   }
191 
192   // Traverse basic blocks in ReversePostOrder.
193   for (BasicBlock *BB : RPOT) {
194     unsigned BBRank = RankMap[BB] = ++Rank << 16;
195 
196     // Walk the basic block, adding precomputed ranks for any instructions that
197     // we cannot move.  This ensures that the ranks for these instructions are
198     // all different in the block.
199     for (Instruction &I : *BB)
200       if (mayHaveNonDefUseDependency(I))
201         ValueRankMap[&I] = ++BBRank;
202   }
203 }
204 
getRank(Value * V)205 unsigned ReassociatePass::getRank(Value *V) {
206   Instruction *I = dyn_cast<Instruction>(V);
207   if (!I) {
208     if (isa<Argument>(V)) return ValueRankMap[V];   // Function argument.
209     return 0;  // Otherwise it's a global or constant, rank 0.
210   }
211 
212   if (unsigned Rank = ValueRankMap[I])
213     return Rank;    // Rank already known?
214 
215   // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that
216   // we can reassociate expressions for code motion!  Since we do not recurse
217   // for PHI nodes, we cannot have infinite recursion here, because there
218   // cannot be loops in the value graph that do not go through PHI nodes.
219   unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
220   for (unsigned i = 0, e = I->getNumOperands(); i != e && Rank != MaxRank; ++i)
221     Rank = std::max(Rank, getRank(I->getOperand(i)));
222 
223   // If this is a 'not' or 'neg' instruction, do not count it for rank. This
224   // assures us that X and ~X will have the same rank.
225   if (!match(I, m_Not(m_Value())) && !match(I, m_Neg(m_Value())) &&
226       !match(I, m_FNeg(m_Value())))
227     ++Rank;
228 
229   LLVM_DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank
230                     << "\n");
231 
232   return ValueRankMap[I] = Rank;
233 }
234 
235 // Canonicalize constants to RHS.  Otherwise, sort the operands by rank.
canonicalizeOperands(Instruction * I)236 void ReassociatePass::canonicalizeOperands(Instruction *I) {
237   assert(isa<BinaryOperator>(I) && "Expected binary operator.");
238   assert(I->isCommutative() && "Expected commutative operator.");
239 
240   Value *LHS = I->getOperand(0);
241   Value *RHS = I->getOperand(1);
242   if (LHS == RHS || isa<Constant>(RHS))
243     return;
244   if (isa<Constant>(LHS) || getRank(RHS) < getRank(LHS))
245     cast<BinaryOperator>(I)->swapOperands();
246 }
247 
CreateAdd(Value * S1,Value * S2,const Twine & Name,BasicBlock::iterator InsertBefore,Value * FlagsOp)248 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
249                                  BasicBlock::iterator InsertBefore,
250                                  Value *FlagsOp) {
251   if (S1->getType()->isIntOrIntVectorTy())
252     return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
253   else {
254     BinaryOperator *Res =
255         BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
256     Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
257     return Res;
258   }
259 }
260 
CreateMul(Value * S1,Value * S2,const Twine & Name,BasicBlock::iterator InsertBefore,Value * FlagsOp)261 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
262                                  BasicBlock::iterator InsertBefore,
263                                  Value *FlagsOp) {
264   if (S1->getType()->isIntOrIntVectorTy())
265     return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
266   else {
267     BinaryOperator *Res =
268       BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
269     Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
270     return Res;
271   }
272 }
273 
CreateNeg(Value * S1,const Twine & Name,BasicBlock::iterator InsertBefore,Value * FlagsOp)274 static Instruction *CreateNeg(Value *S1, const Twine &Name,
275                               BasicBlock::iterator InsertBefore,
276                               Value *FlagsOp) {
277   if (S1->getType()->isIntOrIntVectorTy())
278     return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
279 
280   if (auto *FMFSource = dyn_cast<Instruction>(FlagsOp))
281     return UnaryOperator::CreateFNegFMF(S1, FMFSource, Name, InsertBefore);
282 
283   return UnaryOperator::CreateFNeg(S1, Name, InsertBefore);
284 }
285 
286 /// Replace 0-X with X*-1.
LowerNegateToMultiply(Instruction * Neg)287 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
288   assert((isa<UnaryOperator>(Neg) || isa<BinaryOperator>(Neg)) &&
289          "Expected a Negate!");
290   // FIXME: It's not safe to lower a unary FNeg into a FMul by -1.0.
291   unsigned OpNo = isa<BinaryOperator>(Neg) ? 1 : 0;
292   Type *Ty = Neg->getType();
293   Constant *NegOne = Ty->isIntOrIntVectorTy() ?
294     ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
295 
296   BinaryOperator *Res =
297       CreateMul(Neg->getOperand(OpNo), NegOne, "", Neg->getIterator(), Neg);
298   Neg->setOperand(OpNo, Constant::getNullValue(Ty)); // Drop use of op.
299   Res->takeName(Neg);
300   Neg->replaceAllUsesWith(Res);
301   Res->setDebugLoc(Neg->getDebugLoc());
302   return Res;
303 }
304 
305 using RepeatedValue = std::pair<Value *, uint64_t>;
306 
307 /// Given an associative binary expression, return the leaf
308 /// nodes in Ops along with their weights (how many times the leaf occurs).  The
309 /// original expression is the same as
310 ///   (Ops[0].first op Ops[0].first op ... Ops[0].first)  <- Ops[0].second times
311 /// op
312 ///   (Ops[1].first op Ops[1].first op ... Ops[1].first)  <- Ops[1].second times
313 /// op
314 ///   ...
315 /// op
316 ///   (Ops[N].first op Ops[N].first op ... Ops[N].first)  <- Ops[N].second times
317 ///
318 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
319 ///
320 /// This routine may modify the function, in which case it returns 'true'.  The
321 /// changes it makes may well be destructive, changing the value computed by 'I'
322 /// to something completely different.  Thus if the routine returns 'true' then
323 /// you MUST either replace I with a new expression computed from the Ops array,
324 /// or use RewriteExprTree to put the values back in.
325 ///
326 /// A leaf node is either not a binary operation of the same kind as the root
327 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different
328 /// opcode), or is the same kind of binary operator but has a use which either
329 /// does not belong to the expression, or does belong to the expression but is
330 /// a leaf node.  Every leaf node has at least one use that is a non-leaf node
331 /// of the expression, while for non-leaf nodes (except for the root 'I') every
332 /// use is a non-leaf node of the expression.
333 ///
334 /// For example:
335 ///           expression graph        node names
336 ///
337 ///                     +        |        I
338 ///                    / \       |
339 ///                   +   +      |      A,  B
340 ///                  / \ / \     |
341 ///                 *   +   *    |    C,  D,  E
342 ///                / \ / \ / \   |
343 ///                   +   *      |      F,  G
344 ///
345 /// The leaf nodes are C, E, F and G.  The Ops array will contain (maybe not in
346 /// that order) (C, 1), (E, 1), (F, 2), (G, 2).
347 ///
348 /// The expression is maximal: if some instruction is a binary operator of the
349 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
350 /// then the instruction also belongs to the expression, is not a leaf node of
351 /// it, and its operands also belong to the expression (but may be leaf nodes).
352 ///
353 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
354 /// order to ensure that every non-root node in the expression has *exactly one*
355 /// use by a non-leaf node of the expression.  This destruction means that the
356 /// caller MUST either replace 'I' with a new expression or use something like
357 /// RewriteExprTree to put the values back in if the routine indicates that it
358 /// made a change by returning 'true'.
359 ///
360 /// In the above example either the right operand of A or the left operand of B
361 /// will be replaced by undef.  If it is B's operand then this gives:
362 ///
363 ///                     +        |        I
364 ///                    / \       |
365 ///                   +   +      |      A,  B - operand of B replaced with undef
366 ///                  / \   \     |
367 ///                 *   +   *    |    C,  D,  E
368 ///                / \ / \ / \   |
369 ///                   +   *      |      F,  G
370 ///
371 /// Note that such undef operands can only be reached by passing through 'I'.
372 /// For example, if you visit operands recursively starting from a leaf node
373 /// then you will never see such an undef operand unless you get back to 'I',
374 /// which requires passing through a phi node.
375 ///
376 /// Note that this routine may also mutate binary operators of the wrong type
377 /// that have all uses inside the expression (i.e. only used by non-leaf nodes
378 /// of the expression) if it can turn them into binary operators of the right
379 /// type and thus make the expression bigger.
LinearizeExprTree(Instruction * I,SmallVectorImpl<RepeatedValue> & Ops,ReassociatePass::OrderedSet & ToRedo,reassociate::OverflowTracking & Flags)380 static bool LinearizeExprTree(Instruction *I,
381                               SmallVectorImpl<RepeatedValue> &Ops,
382                               ReassociatePass::OrderedSet &ToRedo,
383                               reassociate::OverflowTracking &Flags) {
384   assert((isa<UnaryOperator>(I) || isa<BinaryOperator>(I)) &&
385          "Expected a UnaryOperator or BinaryOperator!");
386   LLVM_DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
387   unsigned Opcode = I->getOpcode();
388   assert(I->isAssociative() && I->isCommutative() &&
389          "Expected an associative and commutative operation!");
390 
391   // Visit all operands of the expression, keeping track of their weight (the
392   // number of paths from the expression root to the operand, or if you like
393   // the number of times that operand occurs in the linearized expression).
394   // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
395   // while A has weight two.
396 
397   // Worklist of non-leaf nodes (their operands are in the expression too) along
398   // with their weights, representing a certain number of paths to the operator.
399   // If an operator occurs in the worklist multiple times then we found multiple
400   // ways to get to it.
401   SmallVector<std::pair<Instruction *, uint64_t>, 8> Worklist; // (Op, Weight)
402   Worklist.push_back(std::make_pair(I, 1));
403   bool Changed = false;
404 
405   // Leaves of the expression are values that either aren't the right kind of
406   // operation (eg: a constant, or a multiply in an add tree), or are, but have
407   // some uses that are not inside the expression.  For example, in I = X + X,
408   // X = A + B, the value X has two uses (by I) that are in the expression.  If
409   // X has any other uses, for example in a return instruction, then we consider
410   // X to be a leaf, and won't analyze it further.  When we first visit a value,
411   // if it has more than one use then at first we conservatively consider it to
412   // be a leaf.  Later, as the expression is explored, we may discover some more
413   // uses of the value from inside the expression.  If all uses turn out to be
414   // from within the expression (and the value is a binary operator of the right
415   // kind) then the value is no longer considered to be a leaf, and its operands
416   // are explored.
417 
418   // Leaves - Keeps track of the set of putative leaves as well as the number of
419   // paths to each leaf seen so far.
420   using LeafMap = DenseMap<Value *, uint64_t>;
421   LeafMap Leaves; // Leaf -> Total weight so far.
422   SmallVector<Value *, 8> LeafOrder; // Ensure deterministic leaf output order.
423   const DataLayout DL = I->getDataLayout();
424 
425 #ifndef NDEBUG
426   SmallPtrSet<Value *, 8> Visited; // For checking the iteration scheme.
427 #endif
428   while (!Worklist.empty()) {
429     // We examine the operands of this binary operator.
430     auto [I, Weight] = Worklist.pop_back_val();
431 
432     if (isa<OverflowingBinaryOperator>(I)) {
433       Flags.HasNUW &= I->hasNoUnsignedWrap();
434       Flags.HasNSW &= I->hasNoSignedWrap();
435     }
436 
437     for (unsigned OpIdx = 0; OpIdx < I->getNumOperands(); ++OpIdx) { // Visit operands.
438       Value *Op = I->getOperand(OpIdx);
439       LLVM_DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
440       assert(!Op->use_empty() && "No uses, so how did we get to it?!");
441 
442       // If this is a binary operation of the right kind with only one use then
443       // add its operands to the expression.
444       if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) {
445         assert(Visited.insert(Op).second && "Not first visit!");
446         LLVM_DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
447         Worklist.push_back(std::make_pair(BO, Weight));
448         continue;
449       }
450 
451       // Appears to be a leaf.  Is the operand already in the set of leaves?
452       LeafMap::iterator It = Leaves.find(Op);
453       if (It == Leaves.end()) {
454         // Not in the leaf map.  Must be the first time we saw this operand.
455         assert(Visited.insert(Op).second && "Not first visit!");
456         if (!Op->hasOneUse()) {
457           // This value has uses not accounted for by the expression, so it is
458           // not safe to modify.  Mark it as being a leaf.
459           LLVM_DEBUG(dbgs()
460                      << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
461           LeafOrder.push_back(Op);
462           Leaves[Op] = Weight;
463           continue;
464         }
465         // No uses outside the expression, try morphing it.
466       } else {
467         // Already in the leaf map.
468         assert(It != Leaves.end() && Visited.count(Op) &&
469                "In leaf map but not visited!");
470 
471         // Update the number of paths to the leaf.
472         It->second += Weight;
473         assert(It->second >= Weight && "Weight overflows");
474 
475         // If we still have uses that are not accounted for by the expression
476         // then it is not safe to modify the value.
477         if (!Op->hasOneUse())
478           continue;
479 
480         // No uses outside the expression, try morphing it.
481         Weight = It->second;
482         Leaves.erase(It); // Since the value may be morphed below.
483       }
484 
485       // At this point we have a value which, first of all, is not a binary
486       // expression of the right kind, and secondly, is only used inside the
487       // expression.  This means that it can safely be modified.  See if we
488       // can usefully morph it into an expression of the right kind.
489       assert((!isa<Instruction>(Op) ||
490               cast<Instruction>(Op)->getOpcode() != Opcode
491               || (isa<FPMathOperator>(Op) &&
492                   !hasFPAssociativeFlags(cast<Instruction>(Op)))) &&
493              "Should have been handled above!");
494       assert(Op->hasOneUse() && "Has uses outside the expression tree!");
495 
496       // If this is a multiply expression, turn any internal negations into
497       // multiplies by -1 so they can be reassociated.  Add any users of the
498       // newly created multiplication by -1 to the redo list, so any
499       // reassociation opportunities that are exposed will be reassociated
500       // further.
501       Instruction *Neg;
502       if (((Opcode == Instruction::Mul && match(Op, m_Neg(m_Value()))) ||
503            (Opcode == Instruction::FMul && match(Op, m_FNeg(m_Value())))) &&
504            match(Op, m_Instruction(Neg))) {
505         LLVM_DEBUG(dbgs()
506                    << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
507         Instruction *Mul = LowerNegateToMultiply(Neg);
508         LLVM_DEBUG(dbgs() << *Mul << '\n');
509         Worklist.push_back(std::make_pair(Mul, Weight));
510         for (User *U : Mul->users()) {
511           if (BinaryOperator *UserBO = dyn_cast<BinaryOperator>(U))
512             ToRedo.insert(UserBO);
513         }
514         ToRedo.insert(Neg);
515         Changed = true;
516         continue;
517       }
518 
519       // Failed to morph into an expression of the right type.  This really is
520       // a leaf.
521       LLVM_DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
522       assert(!isReassociableOp(Op, Opcode) && "Value was morphed?");
523       LeafOrder.push_back(Op);
524       Leaves[Op] = Weight;
525     }
526   }
527 
528   // The leaves, repeated according to their weights, represent the linearized
529   // form of the expression.
530   for (Value *V : LeafOrder) {
531     LeafMap::iterator It = Leaves.find(V);
532     if (It == Leaves.end())
533       // Node initially thought to be a leaf wasn't.
534       continue;
535     assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!");
536     uint64_t Weight = It->second;
537     // Ensure the leaf is only output once.
538     It->second = 0;
539     Ops.push_back(std::make_pair(V, Weight));
540     if (Opcode == Instruction::Add && Flags.AllKnownNonNegative && Flags.HasNSW)
541       Flags.AllKnownNonNegative &= isKnownNonNegative(V, SimplifyQuery(DL));
542     else if (Opcode == Instruction::Mul) {
543       // To preserve NUW we need all inputs non-zero.
544       // To preserve NSW we need all inputs strictly positive.
545       if (Flags.AllKnownNonZero &&
546           (Flags.HasNUW || (Flags.HasNSW && Flags.AllKnownNonNegative))) {
547         Flags.AllKnownNonZero &= isKnownNonZero(V, SimplifyQuery(DL));
548         if (Flags.HasNSW && Flags.AllKnownNonNegative)
549           Flags.AllKnownNonNegative &= isKnownNonNegative(V, SimplifyQuery(DL));
550       }
551     }
552   }
553 
554   // For nilpotent operations or addition there may be no operands, for example
555   // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
556   // in both cases the weight reduces to 0 causing the value to be skipped.
557   if (Ops.empty()) {
558     Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType());
559     assert(Identity && "Associative operation without identity!");
560     Ops.emplace_back(Identity, 1);
561   }
562 
563   return Changed;
564 }
565 
566 /// Now that the operands for this expression tree are
567 /// linearized and optimized, emit them in-order.
RewriteExprTree(BinaryOperator * I,SmallVectorImpl<ValueEntry> & Ops,OverflowTracking Flags)568 void ReassociatePass::RewriteExprTree(BinaryOperator *I,
569                                       SmallVectorImpl<ValueEntry> &Ops,
570                                       OverflowTracking Flags) {
571   assert(Ops.size() > 1 && "Single values should be used directly!");
572 
573   // Since our optimizations should never increase the number of operations, the
574   // new expression can usually be written reusing the existing binary operators
575   // from the original expression tree, without creating any new instructions,
576   // though the rewritten expression may have a completely different topology.
577   // We take care to not change anything if the new expression will be the same
578   // as the original.  If more than trivial changes (like commuting operands)
579   // were made then we are obliged to clear out any optional subclass data like
580   // nsw flags.
581 
582   /// NodesToRewrite - Nodes from the original expression available for writing
583   /// the new expression into.
584   SmallVector<BinaryOperator*, 8> NodesToRewrite;
585   unsigned Opcode = I->getOpcode();
586   BinaryOperator *Op = I;
587 
588   /// NotRewritable - The operands being written will be the leaves of the new
589   /// expression and must not be used as inner nodes (via NodesToRewrite) by
590   /// mistake.  Inner nodes are always reassociable, and usually leaves are not
591   /// (if they were they would have been incorporated into the expression and so
592   /// would not be leaves), so most of the time there is no danger of this.  But
593   /// in rare cases a leaf may become reassociable if an optimization kills uses
594   /// of it, or it may momentarily become reassociable during rewriting (below)
595   /// due it being removed as an operand of one of its uses.  Ensure that misuse
596   /// of leaf nodes as inner nodes cannot occur by remembering all of the future
597   /// leaves and refusing to reuse any of them as inner nodes.
598   SmallPtrSet<Value*, 8> NotRewritable;
599   for (const ValueEntry &Op : Ops)
600     NotRewritable.insert(Op.Op);
601 
602   // ExpressionChangedStart - Non-null if the rewritten expression differs from
603   // the original in some non-trivial way, requiring the clearing of optional
604   // flags. Flags are cleared from the operator in ExpressionChangedStart up to
605   // ExpressionChangedEnd inclusive.
606   BinaryOperator *ExpressionChangedStart = nullptr,
607                  *ExpressionChangedEnd = nullptr;
608   for (unsigned i = 0; ; ++i) {
609     // The last operation (which comes earliest in the IR) is special as both
610     // operands will come from Ops, rather than just one with the other being
611     // a subexpression.
612     if (i+2 == Ops.size()) {
613       Value *NewLHS = Ops[i].Op;
614       Value *NewRHS = Ops[i+1].Op;
615       Value *OldLHS = Op->getOperand(0);
616       Value *OldRHS = Op->getOperand(1);
617 
618       if (NewLHS == OldLHS && NewRHS == OldRHS)
619         // Nothing changed, leave it alone.
620         break;
621 
622       if (NewLHS == OldRHS && NewRHS == OldLHS) {
623         // The order of the operands was reversed.  Swap them.
624         LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
625         Op->swapOperands();
626         LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
627         MadeChange = true;
628         ++NumChanged;
629         break;
630       }
631 
632       // The new operation differs non-trivially from the original. Overwrite
633       // the old operands with the new ones.
634       LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
635       if (NewLHS != OldLHS) {
636         BinaryOperator *BO = isReassociableOp(OldLHS, Opcode);
637         if (BO && !NotRewritable.count(BO))
638           NodesToRewrite.push_back(BO);
639         Op->setOperand(0, NewLHS);
640       }
641       if (NewRHS != OldRHS) {
642         BinaryOperator *BO = isReassociableOp(OldRHS, Opcode);
643         if (BO && !NotRewritable.count(BO))
644           NodesToRewrite.push_back(BO);
645         Op->setOperand(1, NewRHS);
646       }
647       LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
648 
649       ExpressionChangedStart = Op;
650       if (!ExpressionChangedEnd)
651         ExpressionChangedEnd = Op;
652       MadeChange = true;
653       ++NumChanged;
654 
655       break;
656     }
657 
658     // Not the last operation.  The left-hand side will be a sub-expression
659     // while the right-hand side will be the current element of Ops.
660     Value *NewRHS = Ops[i].Op;
661     if (NewRHS != Op->getOperand(1)) {
662       LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
663       if (NewRHS == Op->getOperand(0)) {
664         // The new right-hand side was already present as the left operand.  If
665         // we are lucky then swapping the operands will sort out both of them.
666         Op->swapOperands();
667       } else {
668         // Overwrite with the new right-hand side.
669         BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode);
670         if (BO && !NotRewritable.count(BO))
671           NodesToRewrite.push_back(BO);
672         Op->setOperand(1, NewRHS);
673         ExpressionChangedStart = Op;
674         if (!ExpressionChangedEnd)
675           ExpressionChangedEnd = Op;
676       }
677       LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
678       MadeChange = true;
679       ++NumChanged;
680     }
681 
682     // Now deal with the left-hand side.  If this is already an operation node
683     // from the original expression then just rewrite the rest of the expression
684     // into it.
685     BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode);
686     if (BO && !NotRewritable.count(BO)) {
687       Op = BO;
688       continue;
689     }
690 
691     // Otherwise, grab a spare node from the original expression and use that as
692     // the left-hand side.  If there are no nodes left then the optimizers made
693     // an expression with more nodes than the original!  This usually means that
694     // they did something stupid but it might mean that the problem was just too
695     // hard (finding the mimimal number of multiplications needed to realize a
696     // multiplication expression is NP-complete).  Whatever the reason, smart or
697     // stupid, create a new node if there are none left.
698     BinaryOperator *NewOp;
699     if (NodesToRewrite.empty()) {
700       Constant *Poison = PoisonValue::get(I->getType());
701       NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode), Poison,
702                                      Poison, "", I->getIterator());
703       if (isa<FPMathOperator>(NewOp))
704         NewOp->setFastMathFlags(I->getFastMathFlags());
705     } else {
706       NewOp = NodesToRewrite.pop_back_val();
707     }
708 
709     LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
710     Op->setOperand(0, NewOp);
711     LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
712     ExpressionChangedStart = Op;
713     if (!ExpressionChangedEnd)
714       ExpressionChangedEnd = Op;
715     MadeChange = true;
716     ++NumChanged;
717     Op = NewOp;
718   }
719 
720   // If the expression changed non-trivially then clear out all subclass data
721   // starting from the operator specified in ExpressionChanged, and compactify
722   // the operators to just before the expression root to guarantee that the
723   // expression tree is dominated by all of Ops.
724   if (ExpressionChangedStart) {
725     bool ClearFlags = true;
726     do {
727       // Preserve flags.
728       if (ClearFlags) {
729         if (isa<FPMathOperator>(I)) {
730           FastMathFlags Flags = I->getFastMathFlags();
731           ExpressionChangedStart->clearSubclassOptionalData();
732           ExpressionChangedStart->setFastMathFlags(Flags);
733         } else {
734           ExpressionChangedStart->clearSubclassOptionalData();
735           if (ExpressionChangedStart->getOpcode() == Instruction::Add ||
736               (ExpressionChangedStart->getOpcode() == Instruction::Mul &&
737                Flags.AllKnownNonZero)) {
738             if (Flags.HasNUW)
739               ExpressionChangedStart->setHasNoUnsignedWrap();
740             if (Flags.HasNSW && (Flags.AllKnownNonNegative || Flags.HasNUW))
741               ExpressionChangedStart->setHasNoSignedWrap();
742           }
743         }
744       }
745 
746       if (ExpressionChangedStart == ExpressionChangedEnd)
747         ClearFlags = false;
748       if (ExpressionChangedStart == I)
749         break;
750 
751       // Discard any debug info related to the expressions that has changed (we
752       // can leave debug info related to the root and any operation that didn't
753       // change, since the result of the expression tree should be the same
754       // even after reassociation).
755       if (ClearFlags)
756         replaceDbgUsesWithUndef(ExpressionChangedStart);
757 
758       ExpressionChangedStart->moveBefore(I);
759       ExpressionChangedStart =
760           cast<BinaryOperator>(*ExpressionChangedStart->user_begin());
761     } while (true);
762   }
763 
764   // Throw away any left over nodes from the original expression.
765   for (BinaryOperator *BO : NodesToRewrite)
766     RedoInsts.insert(BO);
767 }
768 
769 /// Insert instructions before the instruction pointed to by BI,
770 /// that computes the negative version of the value specified.  The negative
771 /// version of the value is returned, and BI is left pointing at the instruction
772 /// that should be processed next by the reassociation pass.
773 /// Also add intermediate instructions to the redo list that are modified while
774 /// pushing the negates through adds.  These will be revisited to see if
775 /// additional opportunities have been exposed.
NegateValue(Value * V,Instruction * BI,ReassociatePass::OrderedSet & ToRedo)776 static Value *NegateValue(Value *V, Instruction *BI,
777                           ReassociatePass::OrderedSet &ToRedo) {
778   if (auto *C = dyn_cast<Constant>(V)) {
779     const DataLayout &DL = BI->getDataLayout();
780     Constant *Res = C->getType()->isFPOrFPVectorTy()
781                         ? ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)
782                         : ConstantExpr::getNeg(C);
783     if (Res)
784       return Res;
785   }
786 
787   // We are trying to expose opportunity for reassociation.  One of the things
788   // that we want to do to achieve this is to push a negation as deep into an
789   // expression chain as possible, to expose the add instructions.  In practice,
790   // this means that we turn this:
791   //   X = -(A+12+C+D)   into    X = -A + -12 + -C + -D = -12 + -A + -C + -D
792   // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
793   // the constants.  We assume that instcombine will clean up the mess later if
794   // we introduce tons of unnecessary negation instructions.
795   //
796   if (BinaryOperator *I =
797           isReassociableOp(V, Instruction::Add, Instruction::FAdd)) {
798     // Push the negates through the add.
799     I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo));
800     I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo));
801     if (I->getOpcode() == Instruction::Add) {
802       I->setHasNoUnsignedWrap(false);
803       I->setHasNoSignedWrap(false);
804     }
805 
806     // We must move the add instruction here, because the neg instructions do
807     // not dominate the old add instruction in general.  By moving it, we are
808     // assured that the neg instructions we just inserted dominate the
809     // instruction we are about to insert after them.
810     //
811     I->moveBefore(BI);
812     I->setName(I->getName()+".neg");
813 
814     // Add the intermediate negates to the redo list as processing them later
815     // could expose more reassociating opportunities.
816     ToRedo.insert(I);
817     return I;
818   }
819 
820   // Okay, we need to materialize a negated version of V with an instruction.
821   // Scan the use lists of V to see if we have one already.
822   for (User *U : V->users()) {
823     if (!match(U, m_Neg(m_Value())) && !match(U, m_FNeg(m_Value())))
824       continue;
825 
826     // We found one!  Now we have to make sure that the definition dominates
827     // this use.  We do this by moving it to the entry block (if it is a
828     // non-instruction value) or right after the definition.  These negates will
829     // be zapped by reassociate later, so we don't need much finesse here.
830     Instruction *TheNeg = dyn_cast<Instruction>(U);
831 
832     // We can't safely propagate a vector zero constant with poison/undef lanes.
833     Constant *C;
834     if (match(TheNeg, m_BinOp(m_Constant(C), m_Value())) &&
835         C->containsUndefOrPoisonElement())
836       continue;
837 
838     // Verify that the negate is in this function, V might be a constant expr.
839     if (!TheNeg ||
840         TheNeg->getParent()->getParent() != BI->getParent()->getParent())
841       continue;
842 
843     BasicBlock::iterator InsertPt;
844     if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
845       auto InsertPtOpt = InstInput->getInsertionPointAfterDef();
846       if (!InsertPtOpt)
847         continue;
848       InsertPt = *InsertPtOpt;
849     } else {
850       InsertPt = TheNeg->getFunction()
851                      ->getEntryBlock()
852                      .getFirstNonPHIOrDbg()
853                      ->getIterator();
854     }
855 
856     // Check that if TheNeg is moved out of its parent block, we drop its
857     // debug location to avoid extra coverage.
858     // See test dropping_debugloc_the_neg.ll for a detailed example.
859     if (TheNeg->getParent() != InsertPt->getParent())
860       TheNeg->dropLocation();
861     TheNeg->moveBefore(*InsertPt->getParent(), InsertPt);
862 
863     if (TheNeg->getOpcode() == Instruction::Sub) {
864       TheNeg->setHasNoUnsignedWrap(false);
865       TheNeg->setHasNoSignedWrap(false);
866     } else {
867       TheNeg->andIRFlags(BI);
868     }
869     ToRedo.insert(TheNeg);
870     return TheNeg;
871   }
872 
873   // Insert a 'neg' instruction that subtracts the value from zero to get the
874   // negation.
875   Instruction *NewNeg =
876       CreateNeg(V, V->getName() + ".neg", BI->getIterator(), BI);
877   ToRedo.insert(NewNeg);
878   return NewNeg;
879 }
880 
881 // See if this `or` looks like an load widening reduction, i.e. that it
882 // consists of an `or`/`shl`/`zext`/`load` nodes only. Note that we don't
883 // ensure that the pattern is *really* a load widening reduction,
884 // we do not ensure that it can really be replaced with a widened load,
885 // only that it mostly looks like one.
isLoadCombineCandidate(Instruction * Or)886 static bool isLoadCombineCandidate(Instruction *Or) {
887   SmallVector<Instruction *, 8> Worklist;
888   SmallSet<Instruction *, 8> Visited;
889 
890   auto Enqueue = [&](Value *V) {
891     auto *I = dyn_cast<Instruction>(V);
892     // Each node of an `or` reduction must be an instruction,
893     if (!I)
894       return false; // Node is certainly not part of an `or` load reduction.
895     // Only process instructions we have never processed before.
896     if (Visited.insert(I).second)
897       Worklist.emplace_back(I);
898     return true; // Will need to look at parent nodes.
899   };
900 
901   if (!Enqueue(Or))
902     return false; // Not an `or` reduction pattern.
903 
904   while (!Worklist.empty()) {
905     auto *I = Worklist.pop_back_val();
906 
907     // Okay, which instruction is this node?
908     switch (I->getOpcode()) {
909     case Instruction::Or:
910       // Got an `or` node. That's fine, just recurse into it's operands.
911       for (Value *Op : I->operands())
912         if (!Enqueue(Op))
913           return false; // Not an `or` reduction pattern.
914       continue;
915 
916     case Instruction::Shl:
917     case Instruction::ZExt:
918       // `shl`/`zext` nodes are fine, just recurse into their base operand.
919       if (!Enqueue(I->getOperand(0)))
920         return false; // Not an `or` reduction pattern.
921       continue;
922 
923     case Instruction::Load:
924       // Perfect, `load` node means we've reached an edge of the graph.
925       continue;
926 
927     default:        // Unknown node.
928       return false; // Not an `or` reduction pattern.
929     }
930   }
931 
932   return true;
933 }
934 
935 /// Return true if it may be profitable to convert this (X|Y) into (X+Y).
shouldConvertOrWithNoCommonBitsToAdd(Instruction * Or)936 static bool shouldConvertOrWithNoCommonBitsToAdd(Instruction *Or) {
937   // Don't bother to convert this up unless either the LHS is an associable add
938   // or subtract or mul or if this is only used by one of the above.
939   // This is only a compile-time improvement, it is not needed for correctness!
940   auto isInteresting = [](Value *V) {
941     for (auto Op : {Instruction::Add, Instruction::Sub, Instruction::Mul,
942                     Instruction::Shl})
943       if (isReassociableOp(V, Op))
944         return true;
945     return false;
946   };
947 
948   if (any_of(Or->operands(), isInteresting))
949     return true;
950 
951   Value *VB = Or->user_back();
952   if (Or->hasOneUse() && isInteresting(VB))
953     return true;
954 
955   return false;
956 }
957 
958 /// If we have (X|Y), and iff X and Y have no common bits set,
959 /// transform this into (X+Y) to allow arithmetics reassociation.
convertOrWithNoCommonBitsToAdd(Instruction * Or)960 static BinaryOperator *convertOrWithNoCommonBitsToAdd(Instruction *Or) {
961   // Convert an or into an add.
962   BinaryOperator *New = CreateAdd(Or->getOperand(0), Or->getOperand(1), "",
963                                   Or->getIterator(), Or);
964   New->setHasNoSignedWrap();
965   New->setHasNoUnsignedWrap();
966   New->takeName(Or);
967 
968   // Everyone now refers to the add instruction.
969   Or->replaceAllUsesWith(New);
970   New->setDebugLoc(Or->getDebugLoc());
971 
972   LLVM_DEBUG(dbgs() << "Converted or into an add: " << *New << '\n');
973   return New;
974 }
975 
976 /// Return true if we should break up this subtract of X-Y into (X + -Y).
ShouldBreakUpSubtract(Instruction * Sub)977 static bool ShouldBreakUpSubtract(Instruction *Sub) {
978   // If this is a negation, we can't split it up!
979   if (match(Sub, m_Neg(m_Value())) || match(Sub, m_FNeg(m_Value())))
980     return false;
981 
982   // Don't breakup X - undef.
983   if (isa<UndefValue>(Sub->getOperand(1)))
984     return false;
985 
986   // Don't bother to break this up unless either the LHS is an associable add or
987   // subtract or if this is only used by one.
988   Value *V0 = Sub->getOperand(0);
989   if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) ||
990       isReassociableOp(V0, Instruction::Sub, Instruction::FSub))
991     return true;
992   Value *V1 = Sub->getOperand(1);
993   if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) ||
994       isReassociableOp(V1, Instruction::Sub, Instruction::FSub))
995     return true;
996   Value *VB = Sub->user_back();
997   if (Sub->hasOneUse() &&
998       (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) ||
999        isReassociableOp(VB, Instruction::Sub, Instruction::FSub)))
1000     return true;
1001 
1002   return false;
1003 }
1004 
1005 /// If we have (X-Y), and if either X is an add, or if this is only used by an
1006 /// add, transform this into (X+(0-Y)) to promote better reassociation.
BreakUpSubtract(Instruction * Sub,ReassociatePass::OrderedSet & ToRedo)1007 static BinaryOperator *BreakUpSubtract(Instruction *Sub,
1008                                        ReassociatePass::OrderedSet &ToRedo) {
1009   // Convert a subtract into an add and a neg instruction. This allows sub
1010   // instructions to be commuted with other add instructions.
1011   //
1012   // Calculate the negative value of Operand 1 of the sub instruction,
1013   // and set it as the RHS of the add instruction we just made.
1014   Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo);
1015   BinaryOperator *New =
1016       CreateAdd(Sub->getOperand(0), NegVal, "", Sub->getIterator(), Sub);
1017   Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op.
1018   Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op.
1019   New->takeName(Sub);
1020 
1021   // Everyone now refers to the add instruction.
1022   Sub->replaceAllUsesWith(New);
1023   New->setDebugLoc(Sub->getDebugLoc());
1024 
1025   LLVM_DEBUG(dbgs() << "Negated: " << *New << '\n');
1026   return New;
1027 }
1028 
1029 /// If this is a shift of a reassociable multiply or is used by one, change
1030 /// this into a multiply by a constant to assist with further reassociation.
ConvertShiftToMul(Instruction * Shl)1031 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1032   Constant *MulCst = ConstantInt::get(Shl->getType(), 1);
1033   auto *SA = cast<ConstantInt>(Shl->getOperand(1));
1034   MulCst = ConstantFoldBinaryInstruction(Instruction::Shl, MulCst, SA);
1035   assert(MulCst && "Constant folding of immediate constants failed");
1036 
1037   BinaryOperator *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst,
1038                                                   "", Shl->getIterator());
1039   Shl->setOperand(0, PoisonValue::get(Shl->getType())); // Drop use of op.
1040   Mul->takeName(Shl);
1041 
1042   // Everyone now refers to the mul instruction.
1043   Shl->replaceAllUsesWith(Mul);
1044   Mul->setDebugLoc(Shl->getDebugLoc());
1045 
1046   // We can safely preserve the nuw flag in all cases.  It's also safe to turn a
1047   // nuw nsw shl into a nuw nsw mul.  However, nsw in isolation requires special
1048   // handling.  It can be preserved as long as we're not left shifting by
1049   // bitwidth - 1.
1050   bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap();
1051   bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap();
1052   unsigned BitWidth = Shl->getType()->getIntegerBitWidth();
1053   if (NSW && (NUW || SA->getValue().ult(BitWidth - 1)))
1054     Mul->setHasNoSignedWrap(true);
1055   Mul->setHasNoUnsignedWrap(NUW);
1056   return Mul;
1057 }
1058 
1059 /// Scan backwards and forwards among values with the same rank as element i
1060 /// to see if X exists.  If X does not exist, return i.  This is useful when
1061 /// scanning for 'x' when we see '-x' because they both get the same rank.
FindInOperandList(const SmallVectorImpl<ValueEntry> & Ops,unsigned i,Value * X)1062 static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops,
1063                                   unsigned i, Value *X) {
1064   unsigned XRank = Ops[i].Rank;
1065   unsigned e = Ops.size();
1066   for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
1067     if (Ops[j].Op == X)
1068       return j;
1069     if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1070       if (Instruction *I2 = dyn_cast<Instruction>(X))
1071         if (I1->isIdenticalTo(I2))
1072           return j;
1073   }
1074   // Scan backwards.
1075   for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
1076     if (Ops[j].Op == X)
1077       return j;
1078     if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op))
1079       if (Instruction *I2 = dyn_cast<Instruction>(X))
1080         if (I1->isIdenticalTo(I2))
1081           return j;
1082   }
1083   return i;
1084 }
1085 
1086 /// Emit a tree of add instructions, summing Ops together
1087 /// and returning the result.  Insert the tree before I.
EmitAddTreeOfValues(BasicBlock::iterator It,SmallVectorImpl<WeakTrackingVH> & Ops)1088 static Value *EmitAddTreeOfValues(BasicBlock::iterator It,
1089                                   SmallVectorImpl<WeakTrackingVH> &Ops) {
1090   if (Ops.size() == 1) return Ops.back();
1091 
1092   Value *V1 = Ops.pop_back_val();
1093   Value *V2 = EmitAddTreeOfValues(It, Ops);
1094   return CreateAdd(V2, V1, "reass.add", It, &*It);
1095 }
1096 
1097 /// If V is an expression tree that is a multiplication sequence,
1098 /// and if this sequence contains a multiply by Factor,
1099 /// remove Factor from the tree and return the new tree.
RemoveFactorFromExpression(Value * V,Value * Factor)1100 Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) {
1101   BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1102   if (!BO)
1103     return nullptr;
1104 
1105   SmallVector<RepeatedValue, 8> Tree;
1106   OverflowTracking Flags;
1107   MadeChange |= LinearizeExprTree(BO, Tree, RedoInsts, Flags);
1108   SmallVector<ValueEntry, 8> Factors;
1109   Factors.reserve(Tree.size());
1110   for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
1111     RepeatedValue E = Tree[i];
1112     Factors.append(E.second, ValueEntry(getRank(E.first), E.first));
1113   }
1114 
1115   bool FoundFactor = false;
1116   bool NeedsNegate = false;
1117   for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1118     if (Factors[i].Op == Factor) {
1119       FoundFactor = true;
1120       Factors.erase(Factors.begin()+i);
1121       break;
1122     }
1123 
1124     // If this is a negative version of this factor, remove it.
1125     if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) {
1126       if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op))
1127         if (FC1->getValue() == -FC2->getValue()) {
1128           FoundFactor = NeedsNegate = true;
1129           Factors.erase(Factors.begin()+i);
1130           break;
1131         }
1132     } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) {
1133       if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) {
1134         const APFloat &F1 = FC1->getValueAPF();
1135         APFloat F2(FC2->getValueAPF());
1136         F2.changeSign();
1137         if (F1 == F2) {
1138           FoundFactor = NeedsNegate = true;
1139           Factors.erase(Factors.begin() + i);
1140           break;
1141         }
1142       }
1143     }
1144   }
1145 
1146   if (!FoundFactor) {
1147     // Make sure to restore the operands to the expression tree.
1148     RewriteExprTree(BO, Factors, Flags);
1149     return nullptr;
1150   }
1151 
1152   BasicBlock::iterator InsertPt = ++BO->getIterator();
1153 
1154   // If this was just a single multiply, remove the multiply and return the only
1155   // remaining operand.
1156   if (Factors.size() == 1) {
1157     RedoInsts.insert(BO);
1158     V = Factors[0].Op;
1159   } else {
1160     RewriteExprTree(BO, Factors, Flags);
1161     V = BO;
1162   }
1163 
1164   if (NeedsNegate)
1165     V = CreateNeg(V, "neg", InsertPt, BO);
1166 
1167   return V;
1168 }
1169 
1170 /// If V is a single-use multiply, recursively add its operands as factors,
1171 /// otherwise add V to the list of factors.
1172 ///
1173 /// Ops is the top-level list of add operands we're trying to factor.
FindSingleUseMultiplyFactors(Value * V,SmallVectorImpl<Value * > & Factors)1174 static void FindSingleUseMultiplyFactors(Value *V,
1175                                          SmallVectorImpl<Value*> &Factors) {
1176   BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul);
1177   if (!BO) {
1178     Factors.push_back(V);
1179     return;
1180   }
1181 
1182   // Otherwise, add the LHS and RHS to the list of factors.
1183   FindSingleUseMultiplyFactors(BO->getOperand(1), Factors);
1184   FindSingleUseMultiplyFactors(BO->getOperand(0), Factors);
1185 }
1186 
1187 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1188 /// This optimizes based on identities.  If it can be reduced to a single Value,
1189 /// it is returned, otherwise the Ops list is mutated as necessary.
OptimizeAndOrXor(unsigned Opcode,SmallVectorImpl<ValueEntry> & Ops)1190 static Value *OptimizeAndOrXor(unsigned Opcode,
1191                                SmallVectorImpl<ValueEntry> &Ops) {
1192   // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1193   // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1194   for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1195     // First, check for X and ~X in the operand list.
1196     assert(i < Ops.size());
1197     Value *X;
1198     if (match(Ops[i].Op, m_Not(m_Value(X)))) {    // Cannot occur for ^.
1199       unsigned FoundX = FindInOperandList(Ops, i, X);
1200       if (FoundX != i) {
1201         if (Opcode == Instruction::And)   // ...&X&~X = 0
1202           return Constant::getNullValue(X->getType());
1203 
1204         if (Opcode == Instruction::Or)    // ...|X|~X = -1
1205           return Constant::getAllOnesValue(X->getType());
1206       }
1207     }
1208 
1209     // Next, check for duplicate pairs of values, which we assume are next to
1210     // each other, due to our sorting criteria.
1211     assert(i < Ops.size());
1212     if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1213       if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1214         // Drop duplicate values for And and Or.
1215         Ops.erase(Ops.begin()+i);
1216         --i; --e;
1217         ++NumAnnihil;
1218         continue;
1219       }
1220 
1221       // Drop pairs of values for Xor.
1222       assert(Opcode == Instruction::Xor);
1223       if (e == 2)
1224         return Constant::getNullValue(Ops[0].Op->getType());
1225 
1226       // Y ^ X^X -> Y
1227       Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1228       i -= 1; e -= 2;
1229       ++NumAnnihil;
1230     }
1231   }
1232   return nullptr;
1233 }
1234 
1235 /// Helper function of CombineXorOpnd(). It creates a bitwise-and
1236 /// instruction with the given two operands, and return the resulting
1237 /// instruction. There are two special cases: 1) if the constant operand is 0,
1238 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1239 /// be returned.
createAndInstr(BasicBlock::iterator InsertBefore,Value * Opnd,const APInt & ConstOpnd)1240 static Value *createAndInstr(BasicBlock::iterator InsertBefore, Value *Opnd,
1241                              const APInt &ConstOpnd) {
1242   if (ConstOpnd.isZero())
1243     return nullptr;
1244 
1245   if (ConstOpnd.isAllOnes())
1246     return Opnd;
1247 
1248   Instruction *I = BinaryOperator::CreateAnd(
1249       Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra",
1250       InsertBefore);
1251   I->setDebugLoc(InsertBefore->getDebugLoc());
1252   return I;
1253 }
1254 
1255 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1256 // into "R ^ C", where C would be 0, and R is a symbolic value.
1257 //
1258 // If it was successful, true is returned, and the "R" and "C" is returned
1259 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1260 // and both "Res" and "ConstOpnd" remain unchanged.
CombineXorOpnd(BasicBlock::iterator It,XorOpnd * Opnd1,APInt & ConstOpnd,Value * & Res)1261 bool ReassociatePass::CombineXorOpnd(BasicBlock::iterator It, XorOpnd *Opnd1,
1262                                      APInt &ConstOpnd, Value *&Res) {
1263   // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1264   //                       = ((x | c1) ^ c1) ^ (c1 ^ c2)
1265   //                       = (x & ~c1) ^ (c1 ^ c2)
1266   // It is useful only when c1 == c2.
1267   if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isZero())
1268     return false;
1269 
1270   if (!Opnd1->getValue()->hasOneUse())
1271     return false;
1272 
1273   const APInt &C1 = Opnd1->getConstPart();
1274   if (C1 != ConstOpnd)
1275     return false;
1276 
1277   Value *X = Opnd1->getSymbolicPart();
1278   Res = createAndInstr(It, X, ~C1);
1279   // ConstOpnd was C2, now C1 ^ C2.
1280   ConstOpnd ^= C1;
1281 
1282   if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1283     RedoInsts.insert(T);
1284   return true;
1285 }
1286 
1287 // Helper function of OptimizeXor(). It tries to simplify
1288 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1289 // symbolic value.
1290 //
1291 // If it was successful, true is returned, and the "R" and "C" is returned
1292 // via "Res" and "ConstOpnd", respectively (If the entire expression is
1293 // evaluated to a constant, the Res is set to NULL); otherwise, false is
1294 // returned, and both "Res" and "ConstOpnd" remain unchanged.
CombineXorOpnd(BasicBlock::iterator It,XorOpnd * Opnd1,XorOpnd * Opnd2,APInt & ConstOpnd,Value * & Res)1295 bool ReassociatePass::CombineXorOpnd(BasicBlock::iterator It, XorOpnd *Opnd1,
1296                                      XorOpnd *Opnd2, APInt &ConstOpnd,
1297                                      Value *&Res) {
1298   Value *X = Opnd1->getSymbolicPart();
1299   if (X != Opnd2->getSymbolicPart())
1300     return false;
1301 
1302   // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1303   int DeadInstNum = 1;
1304   if (Opnd1->getValue()->hasOneUse())
1305     DeadInstNum++;
1306   if (Opnd2->getValue()->hasOneUse())
1307     DeadInstNum++;
1308 
1309   // Xor-Rule 2:
1310   //  (x | c1) ^ (x & c2)
1311   //   = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1312   //   = (x & ~c1) ^ (x & c2) ^ c1               // Xor-Rule 1
1313   //   = (x & c3) ^ c1, where c3 = ~c1 ^ c2      // Xor-rule 3
1314   //
1315   if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1316     if (Opnd2->isOrExpr())
1317       std::swap(Opnd1, Opnd2);
1318 
1319     const APInt &C1 = Opnd1->getConstPart();
1320     const APInt &C2 = Opnd2->getConstPart();
1321     APInt C3((~C1) ^ C2);
1322 
1323     // Do not increase code size!
1324     if (!C3.isZero() && !C3.isAllOnes()) {
1325       int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1326       if (NewInstNum > DeadInstNum)
1327         return false;
1328     }
1329 
1330     Res = createAndInstr(It, X, C3);
1331     ConstOpnd ^= C1;
1332   } else if (Opnd1->isOrExpr()) {
1333     // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1334     //
1335     const APInt &C1 = Opnd1->getConstPart();
1336     const APInt &C2 = Opnd2->getConstPart();
1337     APInt C3 = C1 ^ C2;
1338 
1339     // Do not increase code size
1340     if (!C3.isZero() && !C3.isAllOnes()) {
1341       int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1342       if (NewInstNum > DeadInstNum)
1343         return false;
1344     }
1345 
1346     Res = createAndInstr(It, X, C3);
1347     ConstOpnd ^= C3;
1348   } else {
1349     // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1350     //
1351     const APInt &C1 = Opnd1->getConstPart();
1352     const APInt &C2 = Opnd2->getConstPart();
1353     APInt C3 = C1 ^ C2;
1354     Res = createAndInstr(It, X, C3);
1355   }
1356 
1357   // Put the original operands in the Redo list; hope they will be deleted
1358   // as dead code.
1359   if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue()))
1360     RedoInsts.insert(T);
1361   if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue()))
1362     RedoInsts.insert(T);
1363 
1364   return true;
1365 }
1366 
1367 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1368 /// to a single Value, it is returned, otherwise the Ops list is mutated as
1369 /// necessary.
OptimizeXor(Instruction * I,SmallVectorImpl<ValueEntry> & Ops)1370 Value *ReassociatePass::OptimizeXor(Instruction *I,
1371                                     SmallVectorImpl<ValueEntry> &Ops) {
1372   if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops))
1373     return V;
1374 
1375   if (Ops.size() == 1)
1376     return nullptr;
1377 
1378   SmallVector<XorOpnd, 8> Opnds;
1379   SmallVector<XorOpnd*, 8> OpndPtrs;
1380   Type *Ty = Ops[0].Op->getType();
1381   APInt ConstOpnd(Ty->getScalarSizeInBits(), 0);
1382 
1383   // Step 1: Convert ValueEntry to XorOpnd
1384   for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1385     Value *V = Ops[i].Op;
1386     const APInt *C;
1387     // TODO: Support non-splat vectors.
1388     if (match(V, m_APInt(C))) {
1389       ConstOpnd ^= *C;
1390     } else {
1391       XorOpnd O(V);
1392       O.setSymbolicRank(getRank(O.getSymbolicPart()));
1393       Opnds.push_back(O);
1394     }
1395   }
1396 
1397   // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1398   //  It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1399   //  the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1400   //  with the previous loop --- the iterator of the "Opnds" may be invalidated
1401   //  when new elements are added to the vector.
1402   for (XorOpnd &Op : Opnds)
1403     OpndPtrs.push_back(&Op);
1404 
1405   // Step 2: Sort the Xor-Operands in a way such that the operands containing
1406   //  the same symbolic value cluster together. For instance, the input operand
1407   //  sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1408   //  ("x | 123", "x & 789", "y & 456").
1409   //
1410   //  The purpose is twofold:
1411   //  1) Cluster together the operands sharing the same symbolic-value.
1412   //  2) Operand having smaller symbolic-value-rank is permuted earlier, which
1413   //     could potentially shorten crital path, and expose more loop-invariants.
1414   //     Note that values' rank are basically defined in RPO order (FIXME).
1415   //     So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1416   //     than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1417   //     "z" in the order of X-Y-Z is better than any other orders.
1418   llvm::stable_sort(OpndPtrs, [](XorOpnd *LHS, XorOpnd *RHS) {
1419     return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1420   });
1421 
1422   // Step 3: Combine adjacent operands
1423   XorOpnd *PrevOpnd = nullptr;
1424   bool Changed = false;
1425   for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1426     XorOpnd *CurrOpnd = OpndPtrs[i];
1427     // The combined value
1428     Value *CV;
1429 
1430     // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1431     if (!ConstOpnd.isZero() &&
1432         CombineXorOpnd(I->getIterator(), CurrOpnd, ConstOpnd, CV)) {
1433       Changed = true;
1434       if (CV)
1435         *CurrOpnd = XorOpnd(CV);
1436       else {
1437         CurrOpnd->Invalidate();
1438         continue;
1439       }
1440     }
1441 
1442     if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1443       PrevOpnd = CurrOpnd;
1444       continue;
1445     }
1446 
1447     // step 3.2: When previous and current operands share the same symbolic
1448     //  value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1449     if (CombineXorOpnd(I->getIterator(), CurrOpnd, PrevOpnd, ConstOpnd, CV)) {
1450       // Remove previous operand
1451       PrevOpnd->Invalidate();
1452       if (CV) {
1453         *CurrOpnd = XorOpnd(CV);
1454         PrevOpnd = CurrOpnd;
1455       } else {
1456         CurrOpnd->Invalidate();
1457         PrevOpnd = nullptr;
1458       }
1459       Changed = true;
1460     }
1461   }
1462 
1463   // Step 4: Reassemble the Ops
1464   if (Changed) {
1465     Ops.clear();
1466     for (const XorOpnd &O : Opnds) {
1467       if (O.isInvalid())
1468         continue;
1469       ValueEntry VE(getRank(O.getValue()), O.getValue());
1470       Ops.push_back(VE);
1471     }
1472     if (!ConstOpnd.isZero()) {
1473       Value *C = ConstantInt::get(Ty, ConstOpnd);
1474       ValueEntry VE(getRank(C), C);
1475       Ops.push_back(VE);
1476     }
1477     unsigned Sz = Ops.size();
1478     if (Sz == 1)
1479       return Ops.back().Op;
1480     if (Sz == 0) {
1481       assert(ConstOpnd.isZero());
1482       return ConstantInt::get(Ty, ConstOpnd);
1483     }
1484   }
1485 
1486   return nullptr;
1487 }
1488 
1489 /// Optimize a series of operands to an 'add' instruction.  This
1490 /// optimizes based on identities.  If it can be reduced to a single Value, it
1491 /// is returned, otherwise the Ops list is mutated as necessary.
OptimizeAdd(Instruction * I,SmallVectorImpl<ValueEntry> & Ops)1492 Value *ReassociatePass::OptimizeAdd(Instruction *I,
1493                                     SmallVectorImpl<ValueEntry> &Ops) {
1494   // Scan the operand lists looking for X and -X pairs.  If we find any, we
1495   // can simplify expressions like X+-X == 0 and X+~X ==-1.  While we're at it,
1496   // scan for any
1497   // duplicates.  We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1498 
1499   for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1500     Value *TheOp = Ops[i].Op;
1501     // Check to see if we've seen this operand before.  If so, we factor all
1502     // instances of the operand together.  Due to our sorting criteria, we know
1503     // that these need to be next to each other in the vector.
1504     if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1505       // Rescan the list, remove all instances of this operand from the expr.
1506       unsigned NumFound = 0;
1507       do {
1508         Ops.erase(Ops.begin()+i);
1509         ++NumFound;
1510       } while (i != Ops.size() && Ops[i].Op == TheOp);
1511 
1512       LLVM_DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp
1513                         << '\n');
1514       ++NumFactor;
1515 
1516       // Insert a new multiply.
1517       Type *Ty = TheOp->getType();
1518       Constant *C = Ty->isIntOrIntVectorTy() ?
1519         ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
1520       Instruction *Mul = CreateMul(TheOp, C, "factor", I->getIterator(), I);
1521 
1522       // Now that we have inserted a multiply, optimize it. This allows us to
1523       // handle cases that require multiple factoring steps, such as this:
1524       // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1525       RedoInsts.insert(Mul);
1526 
1527       // If every add operand was a duplicate, return the multiply.
1528       if (Ops.empty())
1529         return Mul;
1530 
1531       // Otherwise, we had some input that didn't have the dupe, such as
1532       // "A + A + B" -> "A*2 + B".  Add the new multiply to the list of
1533       // things being added by this operation.
1534       Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul));
1535 
1536       --i;
1537       e = Ops.size();
1538       continue;
1539     }
1540 
1541     // Check for X and -X or X and ~X in the operand list.
1542     Value *X;
1543     if (!match(TheOp, m_Neg(m_Value(X))) && !match(TheOp, m_Not(m_Value(X))) &&
1544         !match(TheOp, m_FNeg(m_Value(X))))
1545       continue;
1546 
1547     unsigned FoundX = FindInOperandList(Ops, i, X);
1548     if (FoundX == i)
1549       continue;
1550 
1551     // Remove X and -X from the operand list.
1552     if (Ops.size() == 2 &&
1553         (match(TheOp, m_Neg(m_Value())) || match(TheOp, m_FNeg(m_Value()))))
1554       return Constant::getNullValue(X->getType());
1555 
1556     // Remove X and ~X from the operand list.
1557     if (Ops.size() == 2 && match(TheOp, m_Not(m_Value())))
1558       return Constant::getAllOnesValue(X->getType());
1559 
1560     Ops.erase(Ops.begin()+i);
1561     if (i < FoundX)
1562       --FoundX;
1563     else
1564       --i;   // Need to back up an extra one.
1565     Ops.erase(Ops.begin()+FoundX);
1566     ++NumAnnihil;
1567     --i;     // Revisit element.
1568     e -= 2;  // Removed two elements.
1569 
1570     // if X and ~X we append -1 to the operand list.
1571     if (match(TheOp, m_Not(m_Value()))) {
1572       Value *V = Constant::getAllOnesValue(X->getType());
1573       Ops.insert(Ops.end(), ValueEntry(getRank(V), V));
1574       e += 1;
1575     }
1576   }
1577 
1578   // Scan the operand list, checking to see if there are any common factors
1579   // between operands.  Consider something like A*A+A*B*C+D.  We would like to
1580   // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1581   // To efficiently find this, we count the number of times a factor occurs
1582   // for any ADD operands that are MULs.
1583   DenseMap<Value*, unsigned> FactorOccurrences;
1584 
1585   // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1586   // where they are actually the same multiply.
1587   unsigned MaxOcc = 0;
1588   Value *MaxOccVal = nullptr;
1589   for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1590     BinaryOperator *BOp =
1591         isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1592     if (!BOp)
1593       continue;
1594 
1595     // Compute all of the factors of this added value.
1596     SmallVector<Value*, 8> Factors;
1597     FindSingleUseMultiplyFactors(BOp, Factors);
1598     assert(Factors.size() > 1 && "Bad linearize!");
1599 
1600     // Add one to FactorOccurrences for each unique factor in this op.
1601     SmallPtrSet<Value*, 8> Duplicates;
1602     for (Value *Factor : Factors) {
1603       if (!Duplicates.insert(Factor).second)
1604         continue;
1605 
1606       unsigned Occ = ++FactorOccurrences[Factor];
1607       if (Occ > MaxOcc) {
1608         MaxOcc = Occ;
1609         MaxOccVal = Factor;
1610       }
1611 
1612       // If Factor is a negative constant, add the negated value as a factor
1613       // because we can percolate the negate out.  Watch for minint, which
1614       // cannot be positivified.
1615       if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) {
1616         if (CI->isNegative() && !CI->isMinValue(true)) {
1617           Factor = ConstantInt::get(CI->getContext(), -CI->getValue());
1618           if (!Duplicates.insert(Factor).second)
1619             continue;
1620           unsigned Occ = ++FactorOccurrences[Factor];
1621           if (Occ > MaxOcc) {
1622             MaxOcc = Occ;
1623             MaxOccVal = Factor;
1624           }
1625         }
1626       } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) {
1627         if (CF->isNegative()) {
1628           APFloat F(CF->getValueAPF());
1629           F.changeSign();
1630           Factor = ConstantFP::get(CF->getContext(), F);
1631           if (!Duplicates.insert(Factor).second)
1632             continue;
1633           unsigned Occ = ++FactorOccurrences[Factor];
1634           if (Occ > MaxOcc) {
1635             MaxOcc = Occ;
1636             MaxOccVal = Factor;
1637           }
1638         }
1639       }
1640     }
1641   }
1642 
1643   // If any factor occurred more than one time, we can pull it out.
1644   if (MaxOcc > 1) {
1645     LLVM_DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal
1646                       << '\n');
1647     ++NumFactor;
1648 
1649     // Create a new instruction that uses the MaxOccVal twice.  If we don't do
1650     // this, we could otherwise run into situations where removing a factor
1651     // from an expression will drop a use of maxocc, and this can cause
1652     // RemoveFactorFromExpression on successive values to behave differently.
1653     Instruction *DummyInst =
1654         I->getType()->isIntOrIntVectorTy()
1655             ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
1656             : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
1657 
1658     SmallVector<WeakTrackingVH, 4> NewMulOps;
1659     for (unsigned i = 0; i != Ops.size(); ++i) {
1660       // Only try to remove factors from expressions we're allowed to.
1661       BinaryOperator *BOp =
1662           isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);
1663       if (!BOp)
1664         continue;
1665 
1666       if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) {
1667         // The factorized operand may occur several times.  Convert them all in
1668         // one fell swoop.
1669         for (unsigned j = Ops.size(); j != i;) {
1670           --j;
1671           if (Ops[j].Op == Ops[i].Op) {
1672             NewMulOps.push_back(V);
1673             Ops.erase(Ops.begin()+j);
1674           }
1675         }
1676         --i;
1677       }
1678     }
1679 
1680     // No need for extra uses anymore.
1681     DummyInst->deleteValue();
1682 
1683     unsigned NumAddedValues = NewMulOps.size();
1684     Value *V = EmitAddTreeOfValues(I->getIterator(), NewMulOps);
1685 
1686     // Now that we have inserted the add tree, optimize it. This allows us to
1687     // handle cases that require multiple factoring steps, such as this:
1688     // A*A*B + A*A*C   -->   A*(A*B+A*C)   -->   A*(A*(B+C))
1689     assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1690     (void)NumAddedValues;
1691     if (Instruction *VI = dyn_cast<Instruction>(V))
1692       RedoInsts.insert(VI);
1693 
1694     // Create the multiply.
1695     Instruction *V2 = CreateMul(V, MaxOccVal, "reass.mul", I->getIterator(), I);
1696 
1697     // Rerun associate on the multiply in case the inner expression turned into
1698     // a multiply.  We want to make sure that we keep things in canonical form.
1699     RedoInsts.insert(V2);
1700 
1701     // If every add operand included the factor (e.g. "A*B + A*C"), then the
1702     // entire result expression is just the multiply "A*(B+C)".
1703     if (Ops.empty())
1704       return V2;
1705 
1706     // Otherwise, we had some input that didn't have the factor, such as
1707     // "A*B + A*C + D" -> "A*(B+C) + D".  Add the new multiply to the list of
1708     // things being added by this operation.
1709     Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2));
1710   }
1711 
1712   return nullptr;
1713 }
1714 
1715 /// Build up a vector of value/power pairs factoring a product.
1716 ///
1717 /// Given a series of multiplication operands, build a vector of factors and
1718 /// the powers each is raised to when forming the final product. Sort them in
1719 /// the order of descending power.
1720 ///
1721 ///      (x*x)          -> [(x, 2)]
1722 ///     ((x*x)*x)       -> [(x, 3)]
1723 ///   ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1724 ///
1725 /// \returns Whether any factors have a power greater than one.
collectMultiplyFactors(SmallVectorImpl<ValueEntry> & Ops,SmallVectorImpl<Factor> & Factors)1726 static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1727                                    SmallVectorImpl<Factor> &Factors) {
1728   // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1729   // Compute the sum of powers of simplifiable factors.
1730   unsigned FactorPowerSum = 0;
1731   for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1732     Value *Op = Ops[Idx-1].Op;
1733 
1734     // Count the number of occurrences of this value.
1735     unsigned Count = 1;
1736     for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1737       ++Count;
1738     // Track for simplification all factors which occur 2 or more times.
1739     if (Count > 1)
1740       FactorPowerSum += Count;
1741   }
1742 
1743   // We can only simplify factors if the sum of the powers of our simplifiable
1744   // factors is 4 or higher. When that is the case, we will *always* have
1745   // a simplification. This is an important invariant to prevent cyclicly
1746   // trying to simplify already minimal formations.
1747   if (FactorPowerSum < 4)
1748     return false;
1749 
1750   // Now gather the simplifiable factors, removing them from Ops.
1751   FactorPowerSum = 0;
1752   for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1753     Value *Op = Ops[Idx-1].Op;
1754 
1755     // Count the number of occurrences of this value.
1756     unsigned Count = 1;
1757     for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1758       ++Count;
1759     if (Count == 1)
1760       continue;
1761     // Move an even number of occurrences to Factors.
1762     Count &= ~1U;
1763     Idx -= Count;
1764     FactorPowerSum += Count;
1765     Factors.push_back(Factor(Op, Count));
1766     Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count);
1767   }
1768 
1769   // None of the adjustments above should have reduced the sum of factor powers
1770   // below our mininum of '4'.
1771   assert(FactorPowerSum >= 4);
1772 
1773   llvm::stable_sort(Factors, [](const Factor &LHS, const Factor &RHS) {
1774     return LHS.Power > RHS.Power;
1775   });
1776   return true;
1777 }
1778 
1779 /// Build a tree of multiplies, computing the product of Ops.
buildMultiplyTree(IRBuilderBase & Builder,SmallVectorImpl<Value * > & Ops)1780 static Value *buildMultiplyTree(IRBuilderBase &Builder,
1781                                 SmallVectorImpl<Value*> &Ops) {
1782   if (Ops.size() == 1)
1783     return Ops.back();
1784 
1785   Value *LHS = Ops.pop_back_val();
1786   do {
1787     if (LHS->getType()->isIntOrIntVectorTy())
1788       LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
1789     else
1790       LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
1791   } while (!Ops.empty());
1792 
1793   return LHS;
1794 }
1795 
1796 /// Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1797 ///
1798 /// Given a vector of values raised to various powers, where no two values are
1799 /// equal and the powers are sorted in decreasing order, compute the minimal
1800 /// DAG of multiplies to compute the final product, and return that product
1801 /// value.
1802 Value *
buildMinimalMultiplyDAG(IRBuilderBase & Builder,SmallVectorImpl<Factor> & Factors)1803 ReassociatePass::buildMinimalMultiplyDAG(IRBuilderBase &Builder,
1804                                          SmallVectorImpl<Factor> &Factors) {
1805   assert(Factors[0].Power);
1806   SmallVector<Value *, 4> OuterProduct;
1807   for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1808        Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1809     if (Factors[Idx].Power != Factors[LastIdx].Power) {
1810       LastIdx = Idx;
1811       continue;
1812     }
1813 
1814     // We want to multiply across all the factors with the same power so that
1815     // we can raise them to that power as a single entity. Build a mini tree
1816     // for that.
1817     SmallVector<Value *, 4> InnerProduct;
1818     InnerProduct.push_back(Factors[LastIdx].Base);
1819     do {
1820       InnerProduct.push_back(Factors[Idx].Base);
1821       ++Idx;
1822     } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1823 
1824     // Reset the base value of the first factor to the new expression tree.
1825     // We'll remove all the factors with the same power in a second pass.
1826     Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct);
1827     if (Instruction *MI = dyn_cast<Instruction>(M))
1828       RedoInsts.insert(MI);
1829 
1830     LastIdx = Idx;
1831   }
1832   // Unique factors with equal powers -- we've folded them into the first one's
1833   // base.
1834   Factors.erase(llvm::unique(Factors,
1835                              [](const Factor &LHS, const Factor &RHS) {
1836                                return LHS.Power == RHS.Power;
1837                              }),
1838                 Factors.end());
1839 
1840   // Iteratively collect the base of each factor with an add power into the
1841   // outer product, and halve each power in preparation for squaring the
1842   // expression.
1843   for (Factor &F : Factors) {
1844     if (F.Power & 1)
1845       OuterProduct.push_back(F.Base);
1846     F.Power >>= 1;
1847   }
1848   if (Factors[0].Power) {
1849     Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1850     OuterProduct.push_back(SquareRoot);
1851     OuterProduct.push_back(SquareRoot);
1852   }
1853   if (OuterProduct.size() == 1)
1854     return OuterProduct.front();
1855 
1856   Value *V = buildMultiplyTree(Builder, OuterProduct);
1857   return V;
1858 }
1859 
OptimizeMul(BinaryOperator * I,SmallVectorImpl<ValueEntry> & Ops)1860 Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
1861                                     SmallVectorImpl<ValueEntry> &Ops) {
1862   // We can only optimize the multiplies when there is a chain of more than
1863   // three, such that a balanced tree might require fewer total multiplies.
1864   if (Ops.size() < 4)
1865     return nullptr;
1866 
1867   // Try to turn linear trees of multiplies without other uses of the
1868   // intermediate stages into minimal multiply DAGs with perfect sub-expression
1869   // re-use.
1870   SmallVector<Factor, 4> Factors;
1871   if (!collectMultiplyFactors(Ops, Factors))
1872     return nullptr; // All distinct factors, so nothing left for us to do.
1873 
1874   IRBuilder<> Builder(I);
1875   // The reassociate transformation for FP operations is performed only
1876   // if unsafe algebra is permitted by FastMathFlags. Propagate those flags
1877   // to the newly generated operations.
1878   if (auto FPI = dyn_cast<FPMathOperator>(I))
1879     Builder.setFastMathFlags(FPI->getFastMathFlags());
1880 
1881   Value *V = buildMinimalMultiplyDAG(Builder, Factors);
1882   if (Ops.empty())
1883     return V;
1884 
1885   ValueEntry NewEntry = ValueEntry(getRank(V), V);
1886   Ops.insert(llvm::lower_bound(Ops, NewEntry), NewEntry);
1887   return nullptr;
1888 }
1889 
OptimizeExpression(BinaryOperator * I,SmallVectorImpl<ValueEntry> & Ops)1890 Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
1891                                            SmallVectorImpl<ValueEntry> &Ops) {
1892   // Now that we have the linearized expression tree, try to optimize it.
1893   // Start by folding any constants that we found.
1894   const DataLayout &DL = I->getDataLayout();
1895   Constant *Cst = nullptr;
1896   unsigned Opcode = I->getOpcode();
1897   while (!Ops.empty()) {
1898     if (auto *C = dyn_cast<Constant>(Ops.back().Op)) {
1899       if (!Cst) {
1900         Ops.pop_back();
1901         Cst = C;
1902         continue;
1903       }
1904       if (Constant *Res = ConstantFoldBinaryOpOperands(Opcode, C, Cst, DL)) {
1905         Ops.pop_back();
1906         Cst = Res;
1907         continue;
1908       }
1909     }
1910     break;
1911   }
1912   // If there was nothing but constants then we are done.
1913   if (Ops.empty())
1914     return Cst;
1915 
1916   // Put the combined constant back at the end of the operand list, except if
1917   // there is no point.  For example, an add of 0 gets dropped here, while a
1918   // multiplication by zero turns the whole expression into zero.
1919   if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) {
1920     if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType()))
1921       return Cst;
1922     Ops.push_back(ValueEntry(0, Cst));
1923   }
1924 
1925   if (Ops.size() == 1) return Ops[0].Op;
1926 
1927   // Handle destructive annihilation due to identities between elements in the
1928   // argument list here.
1929   unsigned NumOps = Ops.size();
1930   switch (Opcode) {
1931   default: break;
1932   case Instruction::And:
1933   case Instruction::Or:
1934     if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
1935       return Result;
1936     break;
1937 
1938   case Instruction::Xor:
1939     if (Value *Result = OptimizeXor(I, Ops))
1940       return Result;
1941     break;
1942 
1943   case Instruction::Add:
1944   case Instruction::FAdd:
1945     if (Value *Result = OptimizeAdd(I, Ops))
1946       return Result;
1947     break;
1948 
1949   case Instruction::Mul:
1950   case Instruction::FMul:
1951     if (Value *Result = OptimizeMul(I, Ops))
1952       return Result;
1953     break;
1954   }
1955 
1956   if (Ops.size() != NumOps)
1957     return OptimizeExpression(I, Ops);
1958   return nullptr;
1959 }
1960 
1961 // Remove dead instructions and if any operands are trivially dead add them to
1962 // Insts so they will be removed as well.
RecursivelyEraseDeadInsts(Instruction * I,OrderedSet & Insts)1963 void ReassociatePass::RecursivelyEraseDeadInsts(Instruction *I,
1964                                                 OrderedSet &Insts) {
1965   assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1966   SmallVector<Value *, 4> Ops(I->operands());
1967   ValueRankMap.erase(I);
1968   Insts.remove(I);
1969   RedoInsts.remove(I);
1970   llvm::salvageDebugInfo(*I);
1971   I->eraseFromParent();
1972   for (auto *Op : Ops)
1973     if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1974       if (OpInst->use_empty())
1975         Insts.insert(OpInst);
1976 }
1977 
1978 /// Zap the given instruction, adding interesting operands to the work list.
EraseInst(Instruction * I)1979 void ReassociatePass::EraseInst(Instruction *I) {
1980   assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
1981   LLVM_DEBUG(dbgs() << "Erasing dead inst: "; I->dump());
1982 
1983   SmallVector<Value *, 8> Ops(I->operands());
1984   // Erase the dead instruction.
1985   ValueRankMap.erase(I);
1986   RedoInsts.remove(I);
1987   llvm::salvageDebugInfo(*I);
1988   I->eraseFromParent();
1989   // Optimize its operands.
1990   SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
1991   for (Value *V : Ops)
1992     if (Instruction *Op = dyn_cast<Instruction>(V)) {
1993       // If this is a node in an expression tree, climb to the expression root
1994       // and add that since that's where optimization actually happens.
1995       unsigned Opcode = Op->getOpcode();
1996       while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
1997              Visited.insert(Op).second)
1998         Op = Op->user_back();
1999 
2000       // The instruction we're going to push may be coming from a
2001       // dead block, and Reassociate skips the processing of unreachable
2002       // blocks because it's a waste of time and also because it can
2003       // lead to infinite loop due to LLVM's non-standard definition
2004       // of dominance.
2005       if (ValueRankMap.contains(Op))
2006         RedoInsts.insert(Op);
2007     }
2008 
2009   MadeChange = true;
2010 }
2011 
2012 /// Recursively analyze an expression to build a list of instructions that have
2013 /// negative floating-point constant operands. The caller can then transform
2014 /// the list to create positive constants for better reassociation and CSE.
getNegatibleInsts(Value * V,SmallVectorImpl<Instruction * > & Candidates)2015 static void getNegatibleInsts(Value *V,
2016                               SmallVectorImpl<Instruction *> &Candidates) {
2017   // Handle only one-use instructions. Combining negations does not justify
2018   // replicating instructions.
2019   Instruction *I;
2020   if (!match(V, m_OneUse(m_Instruction(I))))
2021     return;
2022 
2023   // Handle expressions of multiplications and divisions.
2024   // TODO: This could look through floating-point casts.
2025   const APFloat *C;
2026   switch (I->getOpcode()) {
2027     case Instruction::FMul:
2028       // Not expecting non-canonical code here. Bail out and wait.
2029       if (match(I->getOperand(0), m_Constant()))
2030         break;
2031 
2032       if (match(I->getOperand(1), m_APFloat(C)) && C->isNegative()) {
2033         Candidates.push_back(I);
2034         LLVM_DEBUG(dbgs() << "FMul with negative constant: " << *I << '\n');
2035       }
2036       getNegatibleInsts(I->getOperand(0), Candidates);
2037       getNegatibleInsts(I->getOperand(1), Candidates);
2038       break;
2039     case Instruction::FDiv:
2040       // Not expecting non-canonical code here. Bail out and wait.
2041       if (match(I->getOperand(0), m_Constant()) &&
2042           match(I->getOperand(1), m_Constant()))
2043         break;
2044 
2045       if ((match(I->getOperand(0), m_APFloat(C)) && C->isNegative()) ||
2046           (match(I->getOperand(1), m_APFloat(C)) && C->isNegative())) {
2047         Candidates.push_back(I);
2048         LLVM_DEBUG(dbgs() << "FDiv with negative constant: " << *I << '\n');
2049       }
2050       getNegatibleInsts(I->getOperand(0), Candidates);
2051       getNegatibleInsts(I->getOperand(1), Candidates);
2052       break;
2053     default:
2054       break;
2055   }
2056 }
2057 
2058 /// Given an fadd/fsub with an operand that is a one-use instruction
2059 /// (the fadd/fsub), try to change negative floating-point constants into
2060 /// positive constants to increase potential for reassociation and CSE.
canonicalizeNegFPConstantsForOp(Instruction * I,Instruction * Op,Value * OtherOp)2061 Instruction *ReassociatePass::canonicalizeNegFPConstantsForOp(Instruction *I,
2062                                                               Instruction *Op,
2063                                                               Value *OtherOp) {
2064   assert((I->getOpcode() == Instruction::FAdd ||
2065           I->getOpcode() == Instruction::FSub) && "Expected fadd/fsub");
2066 
2067   // Collect instructions with negative FP constants from the subtree that ends
2068   // in Op.
2069   SmallVector<Instruction *, 4> Candidates;
2070   getNegatibleInsts(Op, Candidates);
2071   if (Candidates.empty())
2072     return nullptr;
2073 
2074   // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the
2075   // resulting subtract will be broken up later.  This can get us into an
2076   // infinite loop during reassociation.
2077   bool IsFSub = I->getOpcode() == Instruction::FSub;
2078   bool NeedsSubtract = !IsFSub && Candidates.size() % 2 == 1;
2079   if (NeedsSubtract && ShouldBreakUpSubtract(I))
2080     return nullptr;
2081 
2082   for (Instruction *Negatible : Candidates) {
2083     const APFloat *C;
2084     if (match(Negatible->getOperand(0), m_APFloat(C))) {
2085       assert(!match(Negatible->getOperand(1), m_Constant()) &&
2086              "Expecting only 1 constant operand");
2087       assert(C->isNegative() && "Expected negative FP constant");
2088       Negatible->setOperand(0, ConstantFP::get(Negatible->getType(), abs(*C)));
2089       MadeChange = true;
2090     }
2091     if (match(Negatible->getOperand(1), m_APFloat(C))) {
2092       assert(!match(Negatible->getOperand(0), m_Constant()) &&
2093              "Expecting only 1 constant operand");
2094       assert(C->isNegative() && "Expected negative FP constant");
2095       Negatible->setOperand(1, ConstantFP::get(Negatible->getType(), abs(*C)));
2096       MadeChange = true;
2097     }
2098   }
2099   assert(MadeChange == true && "Negative constant candidate was not changed");
2100 
2101   // Negations cancelled out.
2102   if (Candidates.size() % 2 == 0)
2103     return I;
2104 
2105   // Negate the final operand in the expression by flipping the opcode of this
2106   // fadd/fsub.
2107   assert(Candidates.size() % 2 == 1 && "Expected odd number");
2108   IRBuilder<> Builder(I);
2109   Value *NewInst = IsFSub ? Builder.CreateFAddFMF(OtherOp, Op, I)
2110                           : Builder.CreateFSubFMF(OtherOp, Op, I);
2111   I->replaceAllUsesWith(NewInst);
2112   RedoInsts.insert(I);
2113   return dyn_cast<Instruction>(NewInst);
2114 }
2115 
2116 /// Canonicalize expressions that contain a negative floating-point constant
2117 /// of the following form:
2118 ///   OtherOp + (subtree) -> OtherOp {+/-} (canonical subtree)
2119 ///   (subtree) + OtherOp -> OtherOp {+/-} (canonical subtree)
2120 ///   OtherOp - (subtree) -> OtherOp {+/-} (canonical subtree)
2121 ///
2122 /// The fadd/fsub opcode may be switched to allow folding a negation into the
2123 /// input instruction.
canonicalizeNegFPConstants(Instruction * I)2124 Instruction *ReassociatePass::canonicalizeNegFPConstants(Instruction *I) {
2125   LLVM_DEBUG(dbgs() << "Combine negations for: " << *I << '\n');
2126   Value *X;
2127   Instruction *Op;
2128   if (match(I, m_FAdd(m_Value(X), m_OneUse(m_Instruction(Op)))))
2129     if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X))
2130       I = R;
2131   if (match(I, m_FAdd(m_OneUse(m_Instruction(Op)), m_Value(X))))
2132     if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X))
2133       I = R;
2134   if (match(I, m_FSub(m_Value(X), m_OneUse(m_Instruction(Op)))))
2135     if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X))
2136       I = R;
2137   return I;
2138 }
2139 
2140 /// Inspect and optimize the given instruction. Note that erasing
2141 /// instructions is not allowed.
OptimizeInst(Instruction * I)2142 void ReassociatePass::OptimizeInst(Instruction *I) {
2143   // Only consider operations that we understand.
2144   if (!isa<UnaryOperator>(I) && !isa<BinaryOperator>(I))
2145     return;
2146 
2147   if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1)))
2148     // If an operand of this shift is a reassociable multiply, or if the shift
2149     // is used by a reassociable multiply or add, turn into a multiply.
2150     if (isReassociableOp(I->getOperand(0), Instruction::Mul) ||
2151         (I->hasOneUse() &&
2152          (isReassociableOp(I->user_back(), Instruction::Mul) ||
2153           isReassociableOp(I->user_back(), Instruction::Add)))) {
2154       Instruction *NI = ConvertShiftToMul(I);
2155       RedoInsts.insert(I);
2156       MadeChange = true;
2157       I = NI;
2158     }
2159 
2160   // Commute binary operators, to canonicalize the order of their operands.
2161   // This can potentially expose more CSE opportunities, and makes writing other
2162   // transformations simpler.
2163   if (I->isCommutative())
2164     canonicalizeOperands(I);
2165 
2166   // Canonicalize negative constants out of expressions.
2167   if (Instruction *Res = canonicalizeNegFPConstants(I))
2168     I = Res;
2169 
2170   // Don't optimize floating-point instructions unless they have the
2171   // appropriate FastMathFlags for reassociation enabled.
2172   if (isa<FPMathOperator>(I) && !hasFPAssociativeFlags(I))
2173     return;
2174 
2175   // Do not reassociate boolean (i1) expressions.  We want to preserve the
2176   // original order of evaluation for short-circuited comparisons that
2177   // SimplifyCFG has folded to AND/OR expressions.  If the expression
2178   // is not further optimized, it is likely to be transformed back to a
2179   // short-circuited form for code gen, and the source order may have been
2180   // optimized for the most likely conditions.
2181   if (I->getType()->isIntegerTy(1))
2182     return;
2183 
2184   // If this is a bitwise or instruction of operands
2185   // with no common bits set, convert it to X+Y.
2186   if (I->getOpcode() == Instruction::Or &&
2187       shouldConvertOrWithNoCommonBitsToAdd(I) && !isLoadCombineCandidate(I) &&
2188       (cast<PossiblyDisjointInst>(I)->isDisjoint() ||
2189        haveNoCommonBitsSet(I->getOperand(0), I->getOperand(1),
2190                            SimplifyQuery(I->getDataLayout(),
2191                                          /*DT=*/nullptr, /*AC=*/nullptr, I)))) {
2192     Instruction *NI = convertOrWithNoCommonBitsToAdd(I);
2193     RedoInsts.insert(I);
2194     MadeChange = true;
2195     I = NI;
2196   }
2197 
2198   // If this is a subtract instruction which is not already in negate form,
2199   // see if we can convert it to X+-Y.
2200   if (I->getOpcode() == Instruction::Sub) {
2201     if (ShouldBreakUpSubtract(I)) {
2202       Instruction *NI = BreakUpSubtract(I, RedoInsts);
2203       RedoInsts.insert(I);
2204       MadeChange = true;
2205       I = NI;
2206     } else if (match(I, m_Neg(m_Value()))) {
2207       // Otherwise, this is a negation.  See if the operand is a multiply tree
2208       // and if this is not an inner node of a multiply tree.
2209       if (isReassociableOp(I->getOperand(1), Instruction::Mul) &&
2210           (!I->hasOneUse() ||
2211            !isReassociableOp(I->user_back(), Instruction::Mul))) {
2212         Instruction *NI = LowerNegateToMultiply(I);
2213         // If the negate was simplified, revisit the users to see if we can
2214         // reassociate further.
2215         for (User *U : NI->users()) {
2216           if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2217             RedoInsts.insert(Tmp);
2218         }
2219         RedoInsts.insert(I);
2220         MadeChange = true;
2221         I = NI;
2222       }
2223     }
2224   } else if (I->getOpcode() == Instruction::FNeg ||
2225              I->getOpcode() == Instruction::FSub) {
2226     if (ShouldBreakUpSubtract(I)) {
2227       Instruction *NI = BreakUpSubtract(I, RedoInsts);
2228       RedoInsts.insert(I);
2229       MadeChange = true;
2230       I = NI;
2231     } else if (match(I, m_FNeg(m_Value()))) {
2232       // Otherwise, this is a negation.  See if the operand is a multiply tree
2233       // and if this is not an inner node of a multiply tree.
2234       Value *Op = isa<BinaryOperator>(I) ? I->getOperand(1) :
2235                                            I->getOperand(0);
2236       if (isReassociableOp(Op, Instruction::FMul) &&
2237           (!I->hasOneUse() ||
2238            !isReassociableOp(I->user_back(), Instruction::FMul))) {
2239         // If the negate was simplified, revisit the users to see if we can
2240         // reassociate further.
2241         Instruction *NI = LowerNegateToMultiply(I);
2242         for (User *U : NI->users()) {
2243           if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U))
2244             RedoInsts.insert(Tmp);
2245         }
2246         RedoInsts.insert(I);
2247         MadeChange = true;
2248         I = NI;
2249       }
2250     }
2251   }
2252 
2253   // If this instruction is an associative binary operator, process it.
2254   if (!I->isAssociative()) return;
2255   BinaryOperator *BO = cast<BinaryOperator>(I);
2256 
2257   // If this is an interior node of a reassociable tree, ignore it until we
2258   // get to the root of the tree, to avoid N^2 analysis.
2259   unsigned Opcode = BO->getOpcode();
2260   if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2261     // During the initial run we will get to the root of the tree.
2262     // But if we get here while we are redoing instructions, there is no
2263     // guarantee that the root will be visited. So Redo later
2264     if (BO->user_back() != BO &&
2265         BO->getParent() == BO->user_back()->getParent())
2266       RedoInsts.insert(BO->user_back());
2267     return;
2268   }
2269 
2270   // If this is an add tree that is used by a sub instruction, ignore it
2271   // until we process the subtract.
2272   if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2273       cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub)
2274     return;
2275   if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2276       cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
2277     return;
2278 
2279   ReassociateExpression(BO);
2280 }
2281 
ReassociateExpression(BinaryOperator * I)2282 void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2283   // First, walk the expression tree, linearizing the tree, collecting the
2284   // operand information.
2285   SmallVector<RepeatedValue, 8> Tree;
2286   OverflowTracking Flags;
2287   MadeChange |= LinearizeExprTree(I, Tree, RedoInsts, Flags);
2288   SmallVector<ValueEntry, 8> Ops;
2289   Ops.reserve(Tree.size());
2290   for (const RepeatedValue &E : Tree)
2291     Ops.append(E.second, ValueEntry(getRank(E.first), E.first));
2292 
2293   LLVM_DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2294 
2295   // Now that we have linearized the tree to a list and have gathered all of
2296   // the operands and their ranks, sort the operands by their rank.  Use a
2297   // stable_sort so that values with equal ranks will have their relative
2298   // positions maintained (and so the compiler is deterministic).  Note that
2299   // this sorts so that the highest ranking values end up at the beginning of
2300   // the vector.
2301   llvm::stable_sort(Ops);
2302 
2303   // Now that we have the expression tree in a convenient
2304   // sorted form, optimize it globally if possible.
2305   if (Value *V = OptimizeExpression(I, Ops)) {
2306     if (V == I)
2307       // Self-referential expression in unreachable code.
2308       return;
2309     // This expression tree simplified to something that isn't a tree,
2310     // eliminate it.
2311     LLVM_DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2312     I->replaceAllUsesWith(V);
2313     if (Instruction *VI = dyn_cast<Instruction>(V))
2314       if (I->getDebugLoc())
2315         VI->setDebugLoc(I->getDebugLoc());
2316     RedoInsts.insert(I);
2317     ++NumAnnihil;
2318     return;
2319   }
2320 
2321   // We want to sink immediates as deeply as possible except in the case where
2322   // this is a multiply tree used only by an add, and the immediate is a -1.
2323   // In this case we reassociate to put the negation on the outside so that we
2324   // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2325   if (I->hasOneUse()) {
2326     if (I->getOpcode() == Instruction::Mul &&
2327         cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add &&
2328         isa<ConstantInt>(Ops.back().Op) &&
2329         cast<ConstantInt>(Ops.back().Op)->isMinusOne()) {
2330       ValueEntry Tmp = Ops.pop_back_val();
2331       Ops.insert(Ops.begin(), Tmp);
2332     } else if (I->getOpcode() == Instruction::FMul &&
2333                cast<Instruction>(I->user_back())->getOpcode() ==
2334                    Instruction::FAdd &&
2335                isa<ConstantFP>(Ops.back().Op) &&
2336                cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) {
2337       ValueEntry Tmp = Ops.pop_back_val();
2338       Ops.insert(Ops.begin(), Tmp);
2339     }
2340   }
2341 
2342   LLVM_DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2343 
2344   if (Ops.size() == 1) {
2345     if (Ops[0].Op == I)
2346       // Self-referential expression in unreachable code.
2347       return;
2348 
2349     // This expression tree simplified to something that isn't a tree,
2350     // eliminate it.
2351     I->replaceAllUsesWith(Ops[0].Op);
2352     if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op))
2353       OI->setDebugLoc(I->getDebugLoc());
2354     RedoInsts.insert(I);
2355     return;
2356   }
2357 
2358   if (Ops.size() > 2 && Ops.size() <= GlobalReassociateLimit) {
2359     // Find the pair with the highest count in the pairmap and move it to the
2360     // back of the list so that it can later be CSE'd.
2361     // example:
2362     //   a*b*c*d*e
2363     // if c*e is the most "popular" pair, we can express this as
2364     //   (((c*e)*d)*b)*a
2365     unsigned Max = 1;
2366     unsigned BestRank = 0;
2367     std::pair<unsigned, unsigned> BestPair;
2368     unsigned Idx = I->getOpcode() - Instruction::BinaryOpsBegin;
2369     unsigned LimitIdx = 0;
2370     // With the CSE-driven heuristic, we are about to slap two values at the
2371     // beginning of the expression whereas they could live very late in the CFG.
2372     // When using the CSE-local heuristic we avoid creating dependences from
2373     // completely unrelated part of the CFG by limiting the expression
2374     // reordering on the values that live in the first seen basic block.
2375     // The main idea is that we want to avoid forming expressions that would
2376     // become loop dependent.
2377     if (UseCSELocalOpt) {
2378       const BasicBlock *FirstSeenBB = nullptr;
2379       int StartIdx = Ops.size() - 1;
2380       // Skip the first value of the expression since we need at least two
2381       // values to materialize an expression. I.e., even if this value is
2382       // anchored in a different basic block, the actual first sub expression
2383       // will be anchored on the second value.
2384       for (int i = StartIdx - 1; i != -1; --i) {
2385         const Value *Val = Ops[i].Op;
2386         const auto *CurrLeafInstr = dyn_cast<Instruction>(Val);
2387         const BasicBlock *SeenBB = nullptr;
2388         if (!CurrLeafInstr) {
2389           // The value is free of any CFG dependencies.
2390           // Do as if it lives in the entry block.
2391           //
2392           // We do this to make sure all the values falling on this path are
2393           // seen through the same anchor point. The rationale is these values
2394           // can be combined together to from a sub expression free of any CFG
2395           // dependencies so we want them to stay together.
2396           // We could be cleverer and postpone the anchor down to the first
2397           // anchored value, but that's likely complicated to get right.
2398           // E.g., we wouldn't want to do that if that means being stuck in a
2399           // loop.
2400           //
2401           // For instance, we wouldn't want to change:
2402           // res = arg1 op arg2 op arg3 op ... op loop_val1 op loop_val2 ...
2403           // into
2404           // res = loop_val1 op arg1 op arg2 op arg3 op ... op loop_val2 ...
2405           // Because all the sub expressions with arg2..N would be stuck between
2406           // two loop dependent values.
2407           SeenBB = &I->getParent()->getParent()->getEntryBlock();
2408         } else {
2409           SeenBB = CurrLeafInstr->getParent();
2410         }
2411 
2412         if (!FirstSeenBB) {
2413           FirstSeenBB = SeenBB;
2414           continue;
2415         }
2416         if (FirstSeenBB != SeenBB) {
2417           // ith value is in a different basic block.
2418           // Rewind the index once to point to the last value on the same basic
2419           // block.
2420           LimitIdx = i + 1;
2421           LLVM_DEBUG(dbgs() << "CSE reordering: Consider values between ["
2422                             << LimitIdx << ", " << StartIdx << "]\n");
2423           break;
2424         }
2425       }
2426     }
2427     for (unsigned i = Ops.size() - 1; i > LimitIdx; --i) {
2428       // We must use int type to go below zero when LimitIdx is 0.
2429       for (int j = i - 1; j >= (int)LimitIdx; --j) {
2430         unsigned Score = 0;
2431         Value *Op0 = Ops[i].Op;
2432         Value *Op1 = Ops[j].Op;
2433         if (std::less<Value *>()(Op1, Op0))
2434           std::swap(Op0, Op1);
2435         auto it = PairMap[Idx].find({Op0, Op1});
2436         if (it != PairMap[Idx].end()) {
2437           // Functions like BreakUpSubtract() can erase the Values we're using
2438           // as keys and create new Values after we built the PairMap. There's a
2439           // small chance that the new nodes can have the same address as
2440           // something already in the table. We shouldn't accumulate the stored
2441           // score in that case as it refers to the wrong Value.
2442           if (it->second.isValid())
2443             Score += it->second.Score;
2444         }
2445 
2446         unsigned MaxRank = std::max(Ops[i].Rank, Ops[j].Rank);
2447 
2448         // By construction, the operands are sorted in reverse order of their
2449         // topological order.
2450         // So we tend to form (sub) expressions with values that are close to
2451         // each other.
2452         //
2453         // Now to expose more CSE opportunities we want to expose the pair of
2454         // operands that occur the most (as statically computed in
2455         // BuildPairMap.) as the first sub-expression.
2456         //
2457         // If two pairs occur as many times, we pick the one with the
2458         // lowest rank, meaning the one with both operands appearing first in
2459         // the topological order.
2460         if (Score > Max || (Score == Max && MaxRank < BestRank)) {
2461           BestPair = {j, i};
2462           Max = Score;
2463           BestRank = MaxRank;
2464         }
2465       }
2466     }
2467     if (Max > 1) {
2468       auto Op0 = Ops[BestPair.first];
2469       auto Op1 = Ops[BestPair.second];
2470       Ops.erase(&Ops[BestPair.second]);
2471       Ops.erase(&Ops[BestPair.first]);
2472       Ops.push_back(Op0);
2473       Ops.push_back(Op1);
2474     }
2475   }
2476   LLVM_DEBUG(dbgs() << "RAOut after CSE reorder:\t"; PrintOps(I, Ops);
2477              dbgs() << '\n');
2478   // Now that we ordered and optimized the expressions, splat them back into
2479   // the expression tree, removing any unneeded nodes.
2480   RewriteExprTree(I, Ops, Flags);
2481 }
2482 
2483 void
BuildPairMap(ReversePostOrderTraversal<Function * > & RPOT)2484 ReassociatePass::BuildPairMap(ReversePostOrderTraversal<Function *> &RPOT) {
2485   // Make a "pairmap" of how often each operand pair occurs.
2486   for (BasicBlock *BI : RPOT) {
2487     for (Instruction &I : *BI) {
2488       if (!I.isAssociative() || !I.isBinaryOp())
2489         continue;
2490 
2491       // Ignore nodes that aren't at the root of trees.
2492       if (I.hasOneUse() && I.user_back()->getOpcode() == I.getOpcode())
2493         continue;
2494 
2495       // Collect all operands in a single reassociable expression.
2496       // Since Reassociate has already been run once, we can assume things
2497       // are already canonical according to Reassociation's regime.
2498       SmallVector<Value *, 8> Worklist = { I.getOperand(0), I.getOperand(1) };
2499       SmallVector<Value *, 8> Ops;
2500       while (!Worklist.empty() && Ops.size() <= GlobalReassociateLimit) {
2501         Value *Op = Worklist.pop_back_val();
2502         Instruction *OpI = dyn_cast<Instruction>(Op);
2503         if (!OpI || OpI->getOpcode() != I.getOpcode() || !OpI->hasOneUse()) {
2504           Ops.push_back(Op);
2505           continue;
2506         }
2507         // Be paranoid about self-referencing expressions in unreachable code.
2508         if (OpI->getOperand(0) != OpI)
2509           Worklist.push_back(OpI->getOperand(0));
2510         if (OpI->getOperand(1) != OpI)
2511           Worklist.push_back(OpI->getOperand(1));
2512       }
2513       // Skip extremely long expressions.
2514       if (Ops.size() > GlobalReassociateLimit)
2515         continue;
2516 
2517       // Add all pairwise combinations of operands to the pair map.
2518       unsigned BinaryIdx = I.getOpcode() - Instruction::BinaryOpsBegin;
2519       SmallSet<std::pair<Value *, Value*>, 32> Visited;
2520       for (unsigned i = 0; i < Ops.size() - 1; ++i) {
2521         for (unsigned j = i + 1; j < Ops.size(); ++j) {
2522           // Canonicalize operand orderings.
2523           Value *Op0 = Ops[i];
2524           Value *Op1 = Ops[j];
2525           if (std::less<Value *>()(Op1, Op0))
2526             std::swap(Op0, Op1);
2527           if (!Visited.insert({Op0, Op1}).second)
2528             continue;
2529           auto res = PairMap[BinaryIdx].insert({{Op0, Op1}, {Op0, Op1, 1}});
2530           if (!res.second) {
2531             // If either key value has been erased then we've got the same
2532             // address by coincidence. That can't happen here because nothing is
2533             // erasing values but it can happen by the time we're querying the
2534             // map.
2535             assert(res.first->second.isValid() && "WeakVH invalidated");
2536             ++res.first->second.Score;
2537           }
2538         }
2539       }
2540     }
2541   }
2542 }
2543 
run(Function & F,FunctionAnalysisManager &)2544 PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) {
2545   // Get the functions basic blocks in Reverse Post Order. This order is used by
2546   // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic
2547   // blocks (it has been seen that the analysis in this pass could hang when
2548   // analysing dead basic blocks).
2549   ReversePostOrderTraversal<Function *> RPOT(&F);
2550 
2551   // Calculate the rank map for F.
2552   BuildRankMap(F, RPOT);
2553 
2554   // Build the pair map before running reassociate.
2555   // Technically this would be more accurate if we did it after one round
2556   // of reassociation, but in practice it doesn't seem to help much on
2557   // real-world code, so don't waste the compile time running reassociate
2558   // twice.
2559   // If a user wants, they could expicitly run reassociate twice in their
2560   // pass pipeline for further potential gains.
2561   // It might also be possible to update the pair map during runtime, but the
2562   // overhead of that may be large if there's many reassociable chains.
2563   BuildPairMap(RPOT);
2564 
2565   MadeChange = false;
2566 
2567   // Traverse the same blocks that were analysed by BuildRankMap.
2568   for (BasicBlock *BI : RPOT) {
2569     assert(RankMap.count(&*BI) && "BB should be ranked.");
2570     // Optimize every instruction in the basic block.
2571     for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;)
2572       if (isInstructionTriviallyDead(&*II)) {
2573         EraseInst(&*II++);
2574       } else {
2575         OptimizeInst(&*II);
2576         assert(II->getParent() == &*BI && "Moved to a different block!");
2577         ++II;
2578       }
2579 
2580     // Make a copy of all the instructions to be redone so we can remove dead
2581     // instructions.
2582     OrderedSet ToRedo(RedoInsts);
2583     // Iterate over all instructions to be reevaluated and remove trivially dead
2584     // instructions. If any operand of the trivially dead instruction becomes
2585     // dead mark it for deletion as well. Continue this process until all
2586     // trivially dead instructions have been removed.
2587     while (!ToRedo.empty()) {
2588       Instruction *I = ToRedo.pop_back_val();
2589       if (isInstructionTriviallyDead(I)) {
2590         RecursivelyEraseDeadInsts(I, ToRedo);
2591         MadeChange = true;
2592       }
2593     }
2594 
2595     // Now that we have removed dead instructions, we can reoptimize the
2596     // remaining instructions.
2597     while (!RedoInsts.empty()) {
2598       Instruction *I = RedoInsts.front();
2599       RedoInsts.erase(RedoInsts.begin());
2600       if (isInstructionTriviallyDead(I))
2601         EraseInst(I);
2602       else
2603         OptimizeInst(I);
2604     }
2605   }
2606 
2607   // We are done with the rank map and pair map.
2608   RankMap.clear();
2609   ValueRankMap.clear();
2610   for (auto &Entry : PairMap)
2611     Entry.clear();
2612 
2613   if (MadeChange) {
2614     PreservedAnalyses PA;
2615     PA.preserveSet<CFGAnalyses>();
2616     return PA;
2617   }
2618 
2619   return PreservedAnalyses::all();
2620 }
2621 
2622 namespace {
2623 
2624   class ReassociateLegacyPass : public FunctionPass {
2625     ReassociatePass Impl;
2626 
2627   public:
2628     static char ID; // Pass identification, replacement for typeid
2629 
ReassociateLegacyPass()2630     ReassociateLegacyPass() : FunctionPass(ID) {
2631       initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry());
2632     }
2633 
runOnFunction(Function & F)2634     bool runOnFunction(Function &F) override {
2635       if (skipFunction(F))
2636         return false;
2637 
2638       FunctionAnalysisManager DummyFAM;
2639       auto PA = Impl.run(F, DummyFAM);
2640       return !PA.areAllPreserved();
2641     }
2642 
getAnalysisUsage(AnalysisUsage & AU) const2643     void getAnalysisUsage(AnalysisUsage &AU) const override {
2644       AU.setPreservesCFG();
2645       AU.addPreserved<AAResultsWrapperPass>();
2646       AU.addPreserved<BasicAAWrapperPass>();
2647       AU.addPreserved<GlobalsAAWrapperPass>();
2648     }
2649   };
2650 
2651 } // end anonymous namespace
2652 
2653 char ReassociateLegacyPass::ID = 0;
2654 
2655 INITIALIZE_PASS(ReassociateLegacyPass, "reassociate",
2656                 "Reassociate expressions", false, false)
2657 
2658 // Public interface to the Reassociate pass
createReassociatePass()2659 FunctionPass *llvm::createReassociatePass() {
2660   return new ReassociateLegacyPass();
2661 }
2662