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