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 "ScalarOptions.h"
24#include "llvm/ADT/APFloat.h"
25#include "llvm/ADT/APInt.h"
26#include "llvm/ADT/DenseMap.h"
27#include "llvm/ADT/PostOrderIterator.h"
28#include "llvm/ADT/SmallPtrSet.h"
29#include "llvm/ADT/SmallSet.h"
30#include "llvm/ADT/SmallVector.h"
31#include "llvm/ADT/Statistic.h"
32#include "llvm/Analysis/BasicAliasAnalysis.h"
33#include "llvm/Analysis/ConstantFolding.h"
34#include "llvm/Analysis/GlobalsModRef.h"
35#include "llvm/Analysis/ValueTracking.h"
36#include "llvm/IR/Argument.h"
37#include "llvm/IR/BasicBlock.h"
38#include "llvm/IR/CFG.h"
39#include "llvm/IR/Constant.h"
40#include "llvm/IR/Constants.h"
41#include "llvm/IR/Function.h"
42#include "llvm/IR/IRBuilder.h"
43#include "llvm/IR/InstrTypes.h"
44#include "llvm/IR/Instruction.h"
45#include "llvm/IR/Instructions.h"
46#include "llvm/IR/Operator.h"
47#include "llvm/IR/PassManager.h"
48#include "llvm/IR/PatternMatch.h"
49#include "llvm/IR/Type.h"
50#include "llvm/IR/User.h"
51#include "llvm/IR/Value.h"
52#include "llvm/IR/ValueHandle.h"
53#include "llvm/InitializePasses.h"
54#include "llvm/Pass.h"
55#include "llvm/Support/Casting.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
64using namespace llvm;
65using namespace reassociate;
66using namespace PatternMatch;
67
68#define DEBUG_TYPE "reassociate"
69
70STATISTIC(NumChanged, "Number of insts reassociated");
71STATISTIC(NumAnnihil, "Number of expr tree annihilated");
72STATISTIC(NumFactor , "Number of multiplies factored");
73
74#ifndef NDEBUG
75/// Print out the expression identified in the Ops list.
76static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
77 Module *M = I->getModule();
78 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
79 << *Ops[0].Op->getType() << '\t';
80 for (const ValueEntry &Op : Ops) {
81 dbgs() << "[ ";
82 Op.Op->printAsOperand(dbgs(), false, M);
83 dbgs() << ", #" << Op.Rank << "] ";
84 }
85}
86#endif
87
88/// Utility class representing a non-constant Xor-operand. We classify
89/// non-constant Xor-Operands into two categories:
90/// C1) The operand is in the form "X & C", where C is a constant and C != ~0
91/// C2)
92/// C2.1) The operand is in the form of "X | C", where C is a non-zero
93/// constant.
94/// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
95/// operand as "E | 0"
96class llvm::reassociate::XorOpnd {
97public:
98 XorOpnd(Value *V);
99
100 bool isInvalid() const { return SymbolicPart == nullptr; }
101 bool isOrExpr() const { return isOr; }
102 Value *getValue() const { return OrigVal; }
103 Value *getSymbolicPart() const { return SymbolicPart; }
104 unsigned getSymbolicRank() const { return SymbolicRank; }
105 const APInt &getConstPart() const { return ConstPart; }
106
107 void Invalidate() { SymbolicPart = OrigVal = nullptr; }
108 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
109
110private:
111 Value *OrigVal;
112 Value *SymbolicPart;
113 APInt ConstPart;
114 unsigned SymbolicRank;
115 bool isOr;
116};
117
118XorOpnd::XorOpnd(Value *V) {
119 assert(!isa<ConstantInt>(V) && "No ConstantInt");
120 OrigVal = V;
121 Instruction *I = dyn_cast<Instruction>(Val: V);
122 SymbolicRank = 0;
123
124 if (I && (I->getOpcode() == Instruction::Or ||
125 I->getOpcode() == Instruction::And)) {
126 Value *V0 = I->getOperand(i: 0);
127 Value *V1 = I->getOperand(i: 1);
128 const APInt *C;
129 if (match(V: V0, P: m_APInt(Res&: C)))
130 std::swap(a&: V0, b&: V1);
131
132 if (match(V: V1, P: m_APInt(Res&: C))) {
133 ConstPart = *C;
134 SymbolicPart = V0;
135 isOr = (I->getOpcode() == Instruction::Or);
136 return;
137 }
138 }
139
140 // view the operand as "V | 0"
141 SymbolicPart = V;
142 ConstPart = APInt::getZero(numBits: V->getType()->getScalarSizeInBits());
143 isOr = true;
144}
145
146/// Return true if I is an instruction with the FastMathFlags that are needed
147/// for general reassociation set. This is not the same as testing
148/// Instruction::isAssociative() because it includes operations like fsub.
149/// (This routine is only intended to be called for floating-point operations.)
150static bool hasFPAssociativeFlags(Instruction *I) {
151 assert(I && isa<FPMathOperator>(I) && "Should only check FP ops");
152 return I->hasAllowReassoc() && I->hasNoSignedZeros();
153}
154
155/// Return true if V is an instruction of the specified opcode and if it
156/// only has one use.
157static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
158 auto *BO = dyn_cast<BinaryOperator>(Val: V);
159 if (BO && BO->hasOneUse() && BO->getOpcode() == Opcode)
160 if (!isa<FPMathOperator>(Val: BO) || hasFPAssociativeFlags(I: BO))
161 return BO;
162 return nullptr;
163}
164
165static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
166 unsigned Opcode2) {
167 auto *BO = dyn_cast<BinaryOperator>(Val: V);
168 if (BO && BO->hasOneUse() &&
169 (BO->getOpcode() == Opcode1 || BO->getOpcode() == Opcode2))
170 if (!isa<FPMathOperator>(Val: BO) || hasFPAssociativeFlags(I: BO))
171 return BO;
172 return nullptr;
173}
174
175/// Return the fmul operand if V is a one-use fadd with a single one-use fmul
176/// operand, both allowing contraction. Such pairs can be fused into a single
177/// fma, so they are kept together as leaves of the enclosing expression tree
178/// instead of being linearized into it.
179///
180/// Do not keep the pair together if the other operand is itself a reassociable
181/// fadd. Treating the outer fadd as a leaf would hide the nested addition from
182/// reassociation and prevent the complete expression from being optimized.
183static BinaryOperator *isFMulAddCandidate(Value *V) {
184 BinaryOperator *FAdd = isReassociableOp(V, Opcode: Instruction::FAdd);
185 if (!FAdd || !FAdd->hasAllowContract())
186 return nullptr;
187 auto ContractableFMul = [](BinaryOperator *&FMul) {
188 return m_CombineAnd(Ps: m_AllowContract(SubPattern: m_OneUse(SubPattern: m_FMul(L: m_Value(), R: m_Value()))),
189 Ps: m_BinOp(I&: FMul));
190 };
191 BinaryOperator *Mul = nullptr;
192 Value *OtherOp = nullptr;
193 // Keep constants, nested additions, other contractible multiplies and shared
194 // addends visible to the enclosing expression so they can participate in
195 // folding, reassociation and factorization.
196 if (!match(V: FAdd, P: m_c_FAdd(L: ContractableFMul(Mul), R: m_Value(V&: OtherOp))) ||
197 isa<Constant>(Val: OtherOp) || isReassociableOp(V: OtherOp, Opcode: Instruction::FAdd) ||
198 match(V: OtherOp, P: m_AllowContract(SubPattern: m_FMul(L: m_Value(), R: m_Value()))) ||
199 !OtherOp->hasOneUse())
200 return nullptr;
201 return Mul;
202}
203
204void ReassociatePass::BuildRankMap(Function &F,
205 ReversePostOrderTraversal<Function*> &RPOT) {
206 unsigned Rank = 2;
207
208 // Assign distinct ranks to function arguments.
209 for (auto &Arg : F.args()) {
210 ValueRankMap[&Arg] = ++Rank;
211 LLVM_DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank
212 << "\n");
213 }
214
215 // Traverse basic blocks in ReversePostOrder.
216 for (BasicBlock *BB : RPOT) {
217 unsigned BBRank = RankMap[BB] = ++Rank << 16;
218
219 // Walk the basic block, adding precomputed ranks for any instructions that
220 // we cannot move. This ensures that the ranks for these instructions are
221 // all different in the block.
222 for (Instruction &I : *BB)
223 if (mayHaveNonDefUseDependency(I))
224 ValueRankMap[&I] = ++BBRank;
225 }
226}
227
228unsigned ReassociatePass::getRank(Value *V) {
229 // Return 1+MAX(rank(LHS), rank(RHS)) for expressions so we can reassociate
230 // expressions for code motion. Use an explicit worklist rather than native
231 // recursion so long acyclic use-def chains do not overflow the stack.
232 struct RankWorkItem {
233 Value *V;
234 unsigned OpNo;
235 unsigned Rank;
236 };
237
238 // Each item is one suspended recursive getRank() call.
239 // Completed ranks are folded back into the parent.
240 SmallVector<RankWorkItem, 16> Worklist;
241 Worklist.push_back(Elt: RankWorkItem{.V: V, .OpNo: 0, .Rank: 0});
242
243 while (true) {
244 RankWorkItem &Item = Worklist.back();
245 Instruction *I = dyn_cast<Instruction>(Val: Item.V);
246 unsigned Rank = 0;
247 if (!I) {
248 // Function argument, global or constant
249 Rank = isa<Argument>(Val: Item.V) ? ValueRankMap[Item.V] : 0;
250 } else if (ValueRankMap[I]) {
251 // Instruction that is not movable.
252 Rank = ValueRankMap[I];
253 } else if (Item.OpNo == I->getNumOperands() ||
254 Item.Rank == RankMap[I->getParent()]) {
255 // All operands were visited or the max block rank was reached.
256 Rank = Item.Rank;
257 // If this is a 'not' or 'neg' instruction, do not count it for rank.
258 // This assures us that X and ~X will have the same rank.
259 if (!match(V: I, P: m_Not(V: m_Value())) && !match(V: I, P: m_Neg(V: m_Value())) &&
260 !match(V: I, P: m_FNeg(X: m_Value())))
261 ++Rank;
262
263 LLVM_DEBUG(dbgs() << "Calculated Rank[" << I->getName() << "] = " << Rank
264 << "\n");
265
266 ValueRankMap[I] = Rank;
267 } else {
268 Worklist.push_back(Elt: RankWorkItem{.V: I->getOperand(i: Item.OpNo), .OpNo: 0, .Rank: 0});
269 continue;
270 }
271
272 // Once the current use-def node has a known rank, carry that rank back to
273 // the parent expression and advance past the operand that led here.
274 Worklist.pop_back();
275 if (Worklist.empty())
276 return Rank;
277
278 RankWorkItem &Parent = Worklist.back();
279 Parent.Rank = std::max(a: Parent.Rank, b: Rank);
280 ++Parent.OpNo;
281 }
282}
283
284// Canonicalize constants to RHS. Otherwise, sort the operands by rank.
285void ReassociatePass::canonicalizeOperands(Instruction *I) {
286 assert(isa<BinaryOperator>(I) && "Expected binary operator.");
287 assert(I->isCommutative() && "Expected commutative operator.");
288
289 Value *LHS = I->getOperand(i: 0);
290 Value *RHS = I->getOperand(i: 1);
291 if (LHS == RHS || isa<Constant>(Val: RHS))
292 return;
293 if (isa<Constant>(Val: LHS) || getRank(V: RHS) < getRank(V: LHS)) {
294 cast<BinaryOperator>(Val: I)->swapOperands();
295 MadeChange = true;
296 }
297}
298
299static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
300 BasicBlock::iterator InsertBefore,
301 Value *FlagsOp) {
302 if (S1->getType()->isIntOrIntVectorTy())
303 return BinaryOperator::CreateAdd(V1: S1, V2: S2, Name, InsertBefore);
304 else {
305 BinaryOperator *Res =
306 BinaryOperator::CreateFAdd(V1: S1, V2: S2, Name, InsertBefore);
307 Res->setFastMathFlags(cast<FPMathOperator>(Val: FlagsOp)->getFastMathFlags());
308 return Res;
309 }
310}
311
312static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
313 BasicBlock::iterator InsertBefore,
314 Value *FlagsOp) {
315 if (S1->getType()->isIntOrIntVectorTy())
316 return BinaryOperator::CreateMul(V1: S1, V2: S2, Name, InsertBefore);
317 else {
318 BinaryOperator *Res =
319 BinaryOperator::CreateFMul(V1: S1, V2: S2, Name, InsertBefore);
320 Res->setFastMathFlags(cast<FPMathOperator>(Val: FlagsOp)->getFastMathFlags());
321 return Res;
322 }
323}
324
325static Instruction *CreateNeg(Value *S1, const Twine &Name,
326 BasicBlock::iterator InsertBefore,
327 Value *FlagsOp) {
328 if (S1->getType()->isIntOrIntVectorTy())
329 return BinaryOperator::CreateNeg(Op: S1, Name, InsertBefore);
330
331 if (auto *FMFSource = dyn_cast<Instruction>(Val: FlagsOp))
332 return UnaryOperator::CreateFNegFMF(Op: S1, FMFSource, Name, InsertBefore);
333
334 return UnaryOperator::CreateFNeg(V: S1, Name, InsertBefore);
335}
336
337/// Replace 0-X with X*-1.
338static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
339 assert((isa<UnaryOperator>(Neg) || isa<BinaryOperator>(Neg)) &&
340 "Expected a Negate!");
341 // FIXME: It's not safe to lower a unary FNeg into a FMul by -1.0.
342 unsigned OpNo = isa<BinaryOperator>(Val: Neg) ? 1 : 0;
343 Type *Ty = Neg->getType();
344 Constant *NegOne = Ty->isIntOrIntVectorTy() ?
345 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, V: -1.0);
346
347 BinaryOperator *Res =
348 CreateMul(S1: Neg->getOperand(i: OpNo), S2: NegOne, Name: "", InsertBefore: Neg->getIterator(), FlagsOp: Neg);
349 Neg->setOperand(i: OpNo, Val: Constant::getNullValue(Ty)); // Drop use of op.
350 Res->takeName(V: Neg);
351 Neg->replaceAllUsesWith(V: Res);
352 Res->setDebugLoc(Neg->getDebugLoc());
353 return Res;
354}
355
356using RepeatedValue = std::pair<Value *, uint64_t>;
357
358/// Given an associative binary expression, return the leaf
359/// nodes in Ops along with their weights (how many times the leaf occurs). The
360/// original expression is the same as
361/// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
362/// op
363/// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
364/// op
365/// ...
366/// op
367/// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
368///
369/// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
370///
371/// This routine may modify the function, in which case it returns 'true'. The
372/// changes it makes may well be destructive, changing the value computed by 'I'
373/// to something completely different. Thus if the routine returns 'true' then
374/// you MUST either replace I with a new expression computed from the Ops array,
375/// or use RewriteExprTree to put the values back in.
376///
377/// A leaf node is either not a binary operation of the same kind as the root
378/// node 'I' (i.e. is not a binary operator at all, or is, but with a different
379/// opcode), or is the same kind of binary operator but has a use which either
380/// does not belong to the expression, or does belong to the expression but is
381/// a leaf node. Every leaf node has at least one use that is a non-leaf node
382/// of the expression, while for non-leaf nodes (except for the root 'I') every
383/// use is a non-leaf node of the expression.
384///
385/// For example:
386/// expression graph node names
387///
388/// + | I
389/// / \ |
390/// + + | A, B
391/// / \ / \ |
392/// * + * | C, D, E
393/// / \ / \ / \ |
394/// + * | F, G
395///
396/// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
397/// that order) (C, 1), (E, 1), (F, 2), (G, 2).
398///
399/// The expression is maximal: if some instruction is a binary operator of the
400/// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
401/// then the instruction also belongs to the expression, is not a leaf node of
402/// it, and its operands also belong to the expression (but may be leaf nodes).
403///
404/// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
405/// order to ensure that every non-root node in the expression has *exactly one*
406/// use by a non-leaf node of the expression. This destruction means that the
407/// caller MUST either replace 'I' with a new expression or use something like
408/// RewriteExprTree to put the values back in if the routine indicates that it
409/// made a change by returning 'true'.
410///
411/// In the above example either the right operand of A or the left operand of B
412/// will be replaced by undef. If it is B's operand then this gives:
413///
414/// + | I
415/// / \ |
416/// + + | A, B - operand of B replaced with undef
417/// / \ \ |
418/// * + * | C, D, E
419/// / \ / \ / \ |
420/// + * | F, G
421///
422/// Note that such undef operands can only be reached by passing through 'I'.
423/// For example, if you visit operands recursively starting from a leaf node
424/// then you will never see such an undef operand unless you get back to 'I',
425/// which requires passing through a phi node.
426///
427/// Note that this routine may also mutate binary operators of the wrong type
428/// that have all uses inside the expression (i.e. only used by non-leaf nodes
429/// of the expression) if it can turn them into binary operators of the right
430/// type and thus make the expression bigger.
431static bool LinearizeExprTree(Instruction *I,
432 SmallVectorImpl<RepeatedValue> &Ops,
433 ReassociatePass::OrderedSet &ToRedo,
434 OverflowTracking &Flags) {
435 assert((isa<UnaryOperator>(I) || isa<BinaryOperator>(I)) &&
436 "Expected a UnaryOperator or BinaryOperator!");
437 LLVM_DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
438 unsigned Opcode = I->getOpcode();
439 assert(I->isAssociative() && I->isCommutative() &&
440 "Expected an associative and commutative operation!");
441
442 // Visit all operands of the expression, keeping track of their weight (the
443 // number of paths from the expression root to the operand, or if you like
444 // the number of times that operand occurs in the linearized expression).
445 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
446 // while A has weight two.
447
448 // Worklist of non-leaf nodes (their operands are in the expression too) along
449 // with their weights, representing a certain number of paths to the operator.
450 // If an operator occurs in the worklist multiple times then we found multiple
451 // ways to get to it.
452 SmallVector<std::pair<Instruction *, uint64_t>, 8> Worklist; // (Op, Weight)
453 Worklist.push_back(Elt: std::make_pair(x&: I, y: 1));
454 bool Changed = false;
455
456 // Leaves of the expression are values that either aren't the right kind of
457 // operation (eg: a constant, or a multiply in an add tree), or are, but have
458 // some uses that are not inside the expression. For example, in I = X + X,
459 // X = A + B, the value X has two uses (by I) that are in the expression. If
460 // X has any other uses, for example in a return instruction, then we consider
461 // X to be a leaf, and won't analyze it further. When we first visit a value,
462 // if it has more than one use then at first we conservatively consider it to
463 // be a leaf. Later, as the expression is explored, we may discover some more
464 // uses of the value from inside the expression. If all uses turn out to be
465 // from within the expression (and the value is a binary operator of the right
466 // kind) then the value is no longer considered to be a leaf, and its operands
467 // are explored.
468
469 // Leaves - Keeps track of the set of putative leaves as well as the number of
470 // paths to each leaf seen so far.
471 using LeafMap = DenseMap<Value *, uint64_t>;
472 LeafMap Leaves; // Leaf -> Total weight so far.
473 SmallVector<Value *, 8> LeafOrder; // Ensure deterministic leaf output order.
474 const DataLayout &DL = I->getDataLayout();
475
476#ifndef NDEBUG
477 SmallPtrSet<Value *, 8> Visited; // For checking the iteration scheme.
478#endif
479 while (!Worklist.empty()) {
480 // We examine the operands of this binary operator.
481 auto [I, Weight] = Worklist.pop_back_val();
482
483 Flags.mergeFlags(I&: *I);
484
485 for (unsigned OpIdx = 0; OpIdx < I->getNumOperands(); ++OpIdx) { // Visit operands.
486 Value *Op = I->getOperand(i: OpIdx);
487 LLVM_DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
488 assert((!Op->hasUseList() || !Op->use_empty()) &&
489 "No uses, so how did we get to it?!");
490
491 // If this is a binary operation of the right kind with only one use then
492 // add its operands to the expression.
493 if (BinaryOperator *BO = isReassociableOp(V: Op, Opcode);
494 BO && (Opcode != Instruction::FAdd || !isFMulAddCandidate(V: BO))) {
495 assert(Visited.insert(Op).second && "Not first visit!");
496 LLVM_DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
497 Worklist.push_back(Elt: std::make_pair(x&: BO, y&: Weight));
498 continue;
499 }
500
501 // Appears to be a leaf. Is the operand already in the set of leaves?
502 LeafMap::iterator It = Leaves.find(Val: Op);
503 if (It == Leaves.end()) {
504 // Not in the leaf map. Must be the first time we saw this operand.
505 assert(Visited.insert(Op).second && "Not first visit!");
506 if (!Op->hasOneUse()) {
507 // This value has uses not accounted for by the expression, so it is
508 // not safe to modify. Mark it as being a leaf.
509 LLVM_DEBUG(dbgs()
510 << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
511 LeafOrder.push_back(Elt: Op);
512 Leaves[Op] = Weight;
513 continue;
514 }
515 // No uses outside the expression, try morphing it.
516 } else {
517 // Already in the leaf map.
518 assert(It != Leaves.end() && Visited.count(Op) &&
519 "In leaf map but not visited!");
520
521 // Update the number of paths to the leaf.
522 It->second += Weight;
523 assert(It->second >= Weight && "Weight overflows");
524
525 // If we still have uses that are not accounted for by the expression
526 // then it is not safe to modify the value.
527 if (!Op->hasOneUse())
528 continue;
529
530 // No uses outside the expression, try morphing it.
531 Weight = It->second;
532 Leaves.erase(I: It); // Since the value may be morphed below.
533 }
534
535 // At this point we have a value which, first of all, is not a binary
536 // expression of the right kind, and secondly, is only used inside the
537 // expression. This means that it can safely be modified. See if we
538 // can usefully morph it into an expression of the right kind.
539 assert((!isa<Instruction>(Op) ||
540 cast<Instruction>(Op)->getOpcode() != Opcode ||
541 (isa<FPMathOperator>(Op) &&
542 !hasFPAssociativeFlags(cast<Instruction>(Op))) ||
543 isFMulAddCandidate(Op)) &&
544 "Should have been handled above!");
545 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
546
547 // If this is a multiply expression, turn any internal negations into
548 // multiplies by -1 so they can be reassociated. Add any users of the
549 // newly created multiplication by -1 to the redo list, so any
550 // reassociation opportunities that are exposed will be reassociated
551 // further.
552 Instruction *Neg;
553 if (((Opcode == Instruction::Mul && match(V: Op, P: m_Neg(V: m_Value()))) ||
554 (Opcode == Instruction::FMul && match(V: Op, P: m_FNeg(X: m_Value())))) &&
555 match(V: Op, P: m_Instruction(I&: Neg))) {
556 LLVM_DEBUG(dbgs()
557 << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
558 Instruction *Mul = LowerNegateToMultiply(Neg);
559 LLVM_DEBUG(dbgs() << *Mul << '\n');
560 Worklist.push_back(Elt: std::make_pair(x&: Mul, y&: Weight));
561 for (User *U : Mul->users()) {
562 if (BinaryOperator *UserBO = dyn_cast<BinaryOperator>(Val: U))
563 ToRedo.insert(X: UserBO);
564 }
565 ToRedo.insert(X: Neg);
566 Changed = true;
567 continue;
568 }
569
570 // Failed to morph into an expression of the right type. This really is
571 // a leaf.
572 LLVM_DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
573 assert((!isReassociableOp(Op, Opcode) || isFMulAddCandidate(Op)) &&
574 "Value was morphed?");
575 LeafOrder.push_back(Elt: Op);
576 Leaves[Op] = Weight;
577 }
578 }
579
580 // The leaves, repeated according to their weights, represent the linearized
581 // form of the expression.
582 for (Value *V : LeafOrder) {
583 LeafMap::iterator It = Leaves.find(Val: V);
584 if (It == Leaves.end())
585 // Node initially thought to be a leaf wasn't.
586 continue;
587 assert((!isReassociableOp(V, Opcode) || isFMulAddCandidate(V)) &&
588 "Shouldn't be a leaf!");
589 uint64_t Weight = It->second;
590 // Ensure the leaf is only output once.
591 It->second = 0;
592 Ops.push_back(Elt: std::make_pair(x&: V, y&: Weight));
593 if (Opcode == Instruction::Add && Flags.AllKnownNonNegative && Flags.HasNSW)
594 Flags.AllKnownNonNegative &= isKnownNonNegative(V, SQ: SimplifyQuery(DL));
595 else if (Opcode == Instruction::Mul) {
596 // To preserve NUW we need all inputs non-zero.
597 // To preserve NSW we need all inputs strictly positive.
598 if (Flags.AllKnownNonZero &&
599 (Flags.HasNUW || (Flags.HasNSW && Flags.AllKnownNonNegative))) {
600 Flags.AllKnownNonZero &= isKnownNonZero(V, Q: SimplifyQuery(DL));
601 if (Flags.HasNSW && Flags.AllKnownNonNegative)
602 Flags.AllKnownNonNegative &= isKnownNonNegative(V, SQ: SimplifyQuery(DL));
603 }
604 }
605 }
606
607 // For nilpotent operations or addition there may be no operands, for example
608 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
609 // in both cases the weight reduces to 0 causing the value to be skipped.
610 if (Ops.empty()) {
611 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, Ty: I->getType());
612 assert(Identity && "Associative operation without identity!");
613 Ops.emplace_back(Args&: Identity, Args: 1);
614 }
615
616 return Changed;
617}
618
619/// Now that the operands for this expression tree are
620/// linearized and optimized, emit them in-order.
621void ReassociatePass::RewriteExprTree(BinaryOperator *I,
622 SmallVectorImpl<ValueEntry> &Ops,
623 OverflowTracking Flags) {
624 assert(Ops.size() > 1 && "Single values should be used directly!");
625
626 // Since our optimizations should never increase the number of operations, the
627 // new expression can usually be written reusing the existing binary operators
628 // from the original expression tree, without creating any new instructions,
629 // though the rewritten expression may have a completely different topology.
630 // We take care to not change anything if the new expression will be the same
631 // as the original. If more than trivial changes (like commuting operands)
632 // were made then we are obliged to clear out any optional subclass data like
633 // nsw flags.
634
635 /// NodesToRewrite - Nodes from the original expression available for writing
636 /// the new expression into.
637 SmallVector<BinaryOperator*, 8> NodesToRewrite;
638 unsigned Opcode = I->getOpcode();
639 BinaryOperator *Op = I;
640
641 /// NotRewritable - The operands being written will be the leaves of the new
642 /// expression and must not be used as inner nodes (via NodesToRewrite) by
643 /// mistake. Inner nodes are always reassociable, and usually leaves are not
644 /// (if they were they would have been incorporated into the expression and so
645 /// would not be leaves), so most of the time there is no danger of this. But
646 /// in rare cases a leaf may become reassociable if an optimization kills uses
647 /// of it, or it may momentarily become reassociable during rewriting (below)
648 /// due it being removed as an operand of one of its uses. Ensure that misuse
649 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
650 /// leaves and refusing to reuse any of them as inner nodes.
651 SmallPtrSet<Value*, 8> NotRewritable;
652 for (const ValueEntry &Op : Ops)
653 NotRewritable.insert(Ptr: Op.Op);
654
655 // ExpressionChangedStart - Non-null if the rewritten expression differs from
656 // the original in some non-trivial way, requiring the clearing of optional
657 // flags. Flags are cleared from the operator in ExpressionChangedStart up to
658 // ExpressionChangedEnd inclusive.
659 BinaryOperator *ExpressionChangedStart = nullptr,
660 *ExpressionChangedEnd = nullptr;
661 for (unsigned i = 0; ; ++i) {
662 // The last operation (which comes earliest in the IR) is special as both
663 // operands will come from Ops, rather than just one with the other being
664 // a subexpression.
665 if (i+2 == Ops.size()) {
666 Value *NewLHS = Ops[i].Op;
667 Value *NewRHS = Ops[i+1].Op;
668 Value *OldLHS = Op->getOperand(i_nocapture: 0);
669 Value *OldRHS = Op->getOperand(i_nocapture: 1);
670
671 if (NewLHS == OldLHS && NewRHS == OldRHS)
672 // Nothing changed, leave it alone.
673 break;
674
675 if (NewLHS == OldRHS && NewRHS == OldLHS) {
676 // The order of the operands was reversed. Swap them.
677 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
678 Op->swapOperands();
679 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
680 MadeChange = true;
681 ++NumChanged;
682 break;
683 }
684
685 // The new operation differs non-trivially from the original. Overwrite
686 // the old operands with the new ones.
687 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
688 if (NewLHS != OldLHS) {
689 BinaryOperator *BO = isReassociableOp(V: OldLHS, Opcode);
690 if (BO && !NotRewritable.count(Ptr: BO))
691 NodesToRewrite.push_back(Elt: BO);
692 salvageDebugInfo(I&: *Op);
693 Op->setOperand(i_nocapture: 0, Val_nocapture: NewLHS);
694 }
695 if (NewRHS != OldRHS) {
696 BinaryOperator *BO = isReassociableOp(V: OldRHS, Opcode);
697 if (BO && !NotRewritable.count(Ptr: BO))
698 NodesToRewrite.push_back(Elt: BO);
699 salvageDebugInfo(I&: *Op);
700 Op->setOperand(i_nocapture: 1, Val_nocapture: NewRHS);
701 }
702 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
703
704 ExpressionChangedStart = Op;
705 if (!ExpressionChangedEnd)
706 ExpressionChangedEnd = Op;
707 MadeChange = true;
708 ++NumChanged;
709
710 break;
711 }
712
713 // Not the last operation. The left-hand side will be a sub-expression
714 // while the right-hand side will be the current element of Ops.
715 Value *NewRHS = Ops[i].Op;
716 if (NewRHS != Op->getOperand(i_nocapture: 1)) {
717 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
718 if (NewRHS == Op->getOperand(i_nocapture: 0)) {
719 // The new right-hand side was already present as the left operand. If
720 // we are lucky then swapping the operands will sort out both of them.
721 Op->swapOperands();
722 } else {
723 // Overwrite with the new right-hand side.
724 BinaryOperator *BO = isReassociableOp(V: Op->getOperand(i_nocapture: 1), Opcode);
725 if (BO && !NotRewritable.count(Ptr: BO))
726 NodesToRewrite.push_back(Elt: BO);
727 salvageDebugInfo(I&: *Op);
728 Op->setOperand(i_nocapture: 1, Val_nocapture: NewRHS);
729 ExpressionChangedStart = Op;
730 if (!ExpressionChangedEnd)
731 ExpressionChangedEnd = Op;
732 }
733 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
734 MadeChange = true;
735 ++NumChanged;
736 }
737
738 // Now deal with the left-hand side. If this is already an operation node
739 // from the original expression then just rewrite the rest of the expression
740 // into it.
741 BinaryOperator *BO = isReassociableOp(V: Op->getOperand(i_nocapture: 0), Opcode);
742 if (BO && !NotRewritable.count(Ptr: BO)) {
743 Op = BO;
744 continue;
745 }
746
747 // Otherwise, grab a spare node from the original expression and use that as
748 // the left-hand side. If there are no nodes left then the optimizers made
749 // an expression with more nodes than the original! This usually means that
750 // they did something stupid but it might mean that the problem was just too
751 // hard (finding the mimimal number of multiplications needed to realize a
752 // multiplication expression is NP-complete). Whatever the reason, smart or
753 // stupid, create a new node if there are none left.
754 BinaryOperator *NewOp;
755 if (NodesToRewrite.empty()) {
756 Constant *Poison = PoisonValue::get(T: I->getType());
757 NewOp = BinaryOperator::Create(Op: Instruction::BinaryOps(Opcode), S1: Poison,
758 S2: Poison, Name: "", InsertBefore: I->getIterator());
759 if (isa<FPMathOperator>(Val: NewOp))
760 NewOp->setFastMathFlags(I->getFastMathFlags());
761 } else {
762 NewOp = NodesToRewrite.pop_back_val();
763 }
764
765 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
766 salvageDebugInfo(I&: *Op);
767 Op->setOperand(i_nocapture: 0, Val_nocapture: NewOp);
768 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
769 ExpressionChangedStart = Op;
770 if (!ExpressionChangedEnd)
771 ExpressionChangedEnd = Op;
772 MadeChange = true;
773 ++NumChanged;
774 Op = NewOp;
775 }
776
777 // If the expression changed non-trivially then clear out all subclass data
778 // starting from the operator specified in ExpressionChanged, and compactify
779 // the operators to just before the expression root to guarantee that the
780 // expression tree is dominated by all of Ops.
781 if (ExpressionChangedStart) {
782 bool ClearFlags = true;
783 do {
784 // Preserve flags.
785 if (ClearFlags) {
786 if (isa<FPMathOperator>(Val: I)) {
787 ExpressionChangedStart->copyFastMathFlags(FMF: I->getFastMathFlags());
788 } else {
789 Flags.applyFlags(I&: *ExpressionChangedStart);
790 }
791 }
792
793 if (ExpressionChangedStart == ExpressionChangedEnd)
794 ClearFlags = false;
795 if (ExpressionChangedStart == I)
796 break;
797
798 ExpressionChangedStart->moveBefore(InsertPos: I->getIterator());
799 ExpressionChangedStart =
800 cast<BinaryOperator>(Val: *ExpressionChangedStart->user_begin());
801 } while (true);
802 }
803
804 // Throw away any left over nodes from the original expression.
805 RedoInsts.insert_range(R&: NodesToRewrite);
806}
807
808/// Insert instructions before the instruction pointed to by BI,
809/// that computes the negative version of the value specified. The negative
810/// version of the value is returned, and BI is left pointing at the instruction
811/// that should be processed next by the reassociation pass.
812/// Also add intermediate instructions to the redo list that are modified while
813/// pushing the negates through adds. These will be revisited to see if
814/// additional opportunities have been exposed.
815static Value *NegateValue(Value *V, Instruction *BI,
816 ReassociatePass::OrderedSet &ToRedo) {
817 if (auto *C = dyn_cast<Constant>(Val: V)) {
818 const DataLayout &DL = BI->getDataLayout();
819 Constant *Res = C->getType()->isFPOrFPVectorTy()
820 ? ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)
821 : ConstantExpr::getNeg(C);
822 if (Res)
823 return Res;
824 }
825
826 // We are trying to expose opportunity for reassociation. One of the things
827 // that we want to do to achieve this is to push a negation as deep into an
828 // expression chain as possible, to expose the add instructions. In practice,
829 // this means that we turn this:
830 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
831 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
832 // the constants. We assume that instcombine will clean up the mess later if
833 // we introduce tons of unnecessary negation instructions.
834 //
835 if (BinaryOperator *I =
836 isReassociableOp(V, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd)) {
837 // Push the negates through the add.
838 I->setOperand(i_nocapture: 0, Val_nocapture: NegateValue(V: I->getOperand(i_nocapture: 0), BI, ToRedo));
839 I->setOperand(i_nocapture: 1, Val_nocapture: NegateValue(V: I->getOperand(i_nocapture: 1), BI, ToRedo));
840 if (I->getOpcode() == Instruction::Add) {
841 I->setHasNoUnsignedWrap(false);
842 I->setHasNoSignedWrap(false);
843 }
844
845 // We must move the add instruction here, because the neg instructions do
846 // not dominate the old add instruction in general. By moving it, we are
847 // assured that the neg instructions we just inserted dominate the
848 // instruction we are about to insert after them.
849 //
850 I->moveBefore(InsertPos: BI->getIterator());
851 I->setName(I->getName()+".neg");
852
853 // Add the intermediate negates to the redo list as processing them later
854 // could expose more reassociating opportunities.
855 ToRedo.insert(X: I);
856 return I;
857 }
858
859 // Okay, we need to materialize a negated version of V with an instruction.
860 // Scan the use lists of V to see if we have one already.
861 for (User *U : V->users()) {
862 if (!match(V: U, P: m_Neg(V: m_Value())) && !match(V: U, P: m_FNeg(X: m_Value())))
863 continue;
864
865 // We found one! Now we have to make sure that the definition dominates
866 // this use. We do this by moving it to the entry block (if it is a
867 // non-instruction value) or right after the definition. These negates will
868 // be zapped by reassociate later, so we don't need much finesse here.
869 Instruction *TheNeg = dyn_cast<Instruction>(Val: U);
870
871 // We can't safely propagate a vector zero constant with poison/undef lanes.
872 Constant *C;
873 if (match(V: TheNeg, P: m_BinOp(L: m_Constant(C), R: m_Value())) &&
874 C->containsUndefOrPoisonElement())
875 continue;
876
877 // Verify that the negate is in this function, V might be a constant expr.
878 if (!TheNeg ||
879 TheNeg->getParent()->getParent() != BI->getParent()->getParent())
880 continue;
881
882 BasicBlock::iterator InsertPt;
883 if (Instruction *InstInput = dyn_cast<Instruction>(Val: V)) {
884 auto InsertPtOpt = InstInput->getInsertionPointAfterDef();
885 if (!InsertPtOpt)
886 continue;
887 InsertPt = *InsertPtOpt;
888 } else {
889 InsertPt = TheNeg->getFunction()
890 ->getEntryBlock()
891 .getFirstNonPHIOrPseudoProbe()
892 ->getIterator();
893 }
894
895 // Check that if TheNeg is moved out of its parent block, we drop its
896 // debug location to avoid extra coverage.
897 // See test dropping_debugloc_the_neg.ll for a detailed example.
898 if (TheNeg->getParent() != InsertPt->getParent())
899 TheNeg->dropLocation();
900 TheNeg->moveBefore(BB&: *InsertPt->getParent(), I: InsertPt);
901
902 if (TheNeg->getOpcode() == Instruction::Sub) {
903 TheNeg->setHasNoUnsignedWrap(false);
904 TheNeg->setHasNoSignedWrap(false);
905 } else {
906 TheNeg->andIRFlags(V: BI);
907 }
908 ToRedo.insert(X: TheNeg);
909 return TheNeg;
910 }
911
912 // Insert a 'neg' instruction that subtracts the value from zero to get the
913 // negation.
914 Instruction *NewNeg =
915 CreateNeg(S1: V, Name: V->getName() + ".neg", InsertBefore: BI->getIterator(), FlagsOp: BI);
916 // NewNeg is generated to potentially replace BI, so use its DebugLoc.
917 NewNeg->setDebugLoc(BI->getDebugLoc());
918 ToRedo.insert(X: NewNeg);
919 return NewNeg;
920}
921
922// See if this `or` looks like an load widening reduction, i.e. that it
923// consists of an `or`/`shl`/`zext`/`load` nodes only. Note that we don't
924// ensure that the pattern is *really* a load widening reduction,
925// we do not ensure that it can really be replaced with a widened load,
926// only that it mostly looks like one.
927static bool isLoadCombineCandidate(Instruction *Or) {
928 SmallVector<Instruction *, 8> Worklist;
929 SmallPtrSet<Instruction *, 8> Visited;
930
931 auto Enqueue = [&](Value *V) {
932 auto *I = dyn_cast<Instruction>(Val: V);
933 // Each node of an `or` reduction must be an instruction,
934 if (!I)
935 return false; // Node is certainly not part of an `or` load reduction.
936 // Only process instructions we have never processed before.
937 if (Visited.insert(Ptr: I).second)
938 Worklist.emplace_back(Args&: I);
939 return true; // Will need to look at parent nodes.
940 };
941
942 if (!Enqueue(Or))
943 return false; // Not an `or` reduction pattern.
944
945 while (!Worklist.empty()) {
946 auto *I = Worklist.pop_back_val();
947
948 // Okay, which instruction is this node?
949 switch (I->getOpcode()) {
950 case Instruction::Or:
951 // Got an `or` node. That's fine, just recurse into it's operands.
952 for (Value *Op : I->operands())
953 if (!Enqueue(Op))
954 return false; // Not an `or` reduction pattern.
955 continue;
956
957 case Instruction::Shl:
958 case Instruction::ZExt:
959 // `shl`/`zext` nodes are fine, just recurse into their base operand.
960 if (!Enqueue(I->getOperand(i: 0)))
961 return false; // Not an `or` reduction pattern.
962 continue;
963
964 case Instruction::Load:
965 // Perfect, `load` node means we've reached an edge of the graph.
966 continue;
967
968 default: // Unknown node.
969 return false; // Not an `or` reduction pattern.
970 }
971 }
972
973 return true;
974}
975
976/// Return true if it may be profitable to convert this (X|Y) into (X+Y).
977static bool shouldConvertOrWithNoCommonBitsToAdd(Instruction *Or) {
978 // Don't bother to convert this up unless either the LHS is an associable add
979 // or subtract or mul or if this is only used by one of the above.
980 // This is only a compile-time improvement, it is not needed for correctness!
981 auto isInteresting = [](Value *V) {
982 for (auto Op : {Instruction::Add, Instruction::Sub, Instruction::Mul,
983 Instruction::Shl})
984 if (isReassociableOp(V, Opcode: Op))
985 return true;
986 return false;
987 };
988
989 if (any_of(Range: Or->operands(), P: isInteresting))
990 return true;
991
992 Value *VB = Or->user_back();
993 if (Or->hasOneUse() && isInteresting(VB))
994 return true;
995
996 return false;
997}
998
999/// If we have (X|Y), and iff X and Y have no common bits set,
1000/// transform this into (X+Y) to allow arithmetics reassociation.
1001static BinaryOperator *convertOrWithNoCommonBitsToAdd(Instruction *Or) {
1002 // Convert an or into an add.
1003 BinaryOperator *New = CreateAdd(S1: Or->getOperand(i: 0), S2: Or->getOperand(i: 1), Name: "",
1004 InsertBefore: Or->getIterator(), FlagsOp: Or);
1005 New->setHasNoSignedWrap();
1006 New->setHasNoUnsignedWrap();
1007 New->takeName(V: Or);
1008
1009 // Everyone now refers to the add instruction.
1010 Or->replaceAllUsesWith(V: New);
1011 New->setDebugLoc(Or->getDebugLoc());
1012
1013 LLVM_DEBUG(dbgs() << "Converted or into an add: " << *New << '\n');
1014 return New;
1015}
1016
1017/// Return true if Mul is of the form (X+Y)*C or (X-Y)*C where C is a
1018/// constant, and there exists a sibling instruction of the form X*C' or Y*C'
1019/// in the same expression — indicating that distribution followed by
1020/// factoring will reduce the instruction count.
1021static bool ShouldBreakUpDistribution(Instruction *Mul) {
1022 Value *A, *B;
1023 if (!match(V: Mul, P: m_OneUse(SubPattern: m_Mul(
1024 L: m_OneUse(SubPattern: m_CombineOr(Ps: m_Add(L: m_Value(V&: A), R: m_Value(V&: B)),
1025 Ps: m_Sub(L: m_Value(V&: A), R: m_Value(V&: B)))),
1026 R: m_ImmConstant()))))
1027 return false;
1028
1029 auto *MulUser = cast<Instruction>(Val: Mul->user_back());
1030 // The parent MUST be an Add or Sub to ensure the tree is flattened
1031 if (MulUser->getOpcode() != Instruction::Add &&
1032 MulUser->getOpcode() != Instruction::Sub)
1033 return false;
1034
1035 for (Value *Sibling : MulUser->operands()) {
1036 if (Sibling == Mul || !Sibling->hasOneUse())
1037 continue;
1038
1039 // Sibling must be NonConst * C'.
1040 Value *SibNC;
1041 if (match(V: Sibling, P: m_Mul(L: m_Value(V&: SibNC), R: m_ImmConstant())) &&
1042 (SibNC == A || SibNC == B) && !isa<Constant>(Val: SibNC))
1043 return true;
1044 }
1045 return false;
1046}
1047
1048/// Distribute Mul of the form (X+Y)*C into X*C + Y*C.
1049/// For the sub case (X-Y)*C, the second term uses -C to avoid
1050/// introducing a negation instruction.
1051static BinaryOperator *BreakUpDistribute(Instruction *Mul,
1052 ReassociatePass::OrderedSet &ToRedo) {
1053 Instruction *AddSub = cast<Instruction>(Val: Mul->getOperand(i: 0));
1054 Constant *C = cast<Constant>(Val: Mul->getOperand(i: 1));
1055 Constant *C2 =
1056 AddSub->getOpcode() == Instruction::Sub ? ConstantExpr::getNeg(C) : C;
1057
1058 BinaryOperator *M1 = BinaryOperator::CreateMul(V1: AddSub->getOperand(i: 0), V2: C,
1059 Name: "Mul1", InsertBefore: Mul->getIterator());
1060 BinaryOperator *M2 = BinaryOperator::CreateMul(V1: AddSub->getOperand(i: 1), V2: C2,
1061 Name: "Mul2", InsertBefore: Mul->getIterator());
1062 BinaryOperator *Result =
1063 BinaryOperator::CreateAdd(V1: M1, V2: M2, Name: "DistAdd", InsertBefore: Mul->getIterator());
1064
1065 Mul->replaceAllUsesWith(V: Result);
1066 Result->setDebugLoc(Mul->getDebugLoc());
1067
1068 ToRedo.insert(X: M1);
1069 ToRedo.insert(X: M2);
1070 ToRedo.insert(X: Result);
1071
1072 return Result;
1073}
1074
1075/// Return true if we should break up this subtract of X-Y into (X + -Y).
1076static bool ShouldBreakUpSubtract(Instruction *Sub) {
1077 // If this is a negation, we can't split it up!
1078 if (match(V: Sub, P: m_Neg(V: m_Value())) || match(V: Sub, P: m_FNeg(X: m_Value())))
1079 return false;
1080
1081 // Don't breakup X - undef.
1082 if (isa<UndefValue>(Val: Sub->getOperand(i: 1)))
1083 return false;
1084
1085 // Don't bother to break this up unless either the LHS is an associable add or
1086 // subtract or if this is only used by one.
1087 Value *V0 = Sub->getOperand(i: 0);
1088 if (isReassociableOp(V: V0, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd) ||
1089 isReassociableOp(V: V0, Opcode1: Instruction::Sub, Opcode2: Instruction::FSub))
1090 return true;
1091 Value *V1 = Sub->getOperand(i: 1);
1092 if (isReassociableOp(V: V1, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd) ||
1093 isReassociableOp(V: V1, Opcode1: Instruction::Sub, Opcode2: Instruction::FSub))
1094 return true;
1095 Value *VB = Sub->user_back();
1096 if (Sub->hasOneUse() &&
1097 (isReassociableOp(V: VB, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd) ||
1098 isReassociableOp(V: VB, Opcode1: Instruction::Sub, Opcode2: Instruction::FSub)))
1099 return true;
1100
1101 return false;
1102}
1103
1104/// If we have (X-Y), and if either X is an add, or if this is only used by an
1105/// add, transform this into (X+(0-Y)) to promote better reassociation.
1106static BinaryOperator *BreakUpSubtract(Instruction *Sub,
1107 ReassociatePass::OrderedSet &ToRedo) {
1108 // Convert a subtract into an add and a neg instruction. This allows sub
1109 // instructions to be commuted with other add instructions.
1110 //
1111 // Calculate the negative value of Operand 1 of the sub instruction,
1112 // and set it as the RHS of the add instruction we just made.
1113 Value *NegVal = NegateValue(V: Sub->getOperand(i: 1), BI: Sub, ToRedo);
1114 BinaryOperator *New =
1115 CreateAdd(S1: Sub->getOperand(i: 0), S2: NegVal, Name: "", InsertBefore: Sub->getIterator(), FlagsOp: Sub);
1116 Sub->setOperand(i: 0, Val: Constant::getNullValue(Ty: Sub->getType())); // Drop use of op.
1117 Sub->setOperand(i: 1, Val: Constant::getNullValue(Ty: Sub->getType())); // Drop use of op.
1118 New->takeName(V: Sub);
1119
1120 // Everyone now refers to the add instruction.
1121 Sub->replaceAllUsesWith(V: New);
1122 New->setDebugLoc(Sub->getDebugLoc());
1123
1124 LLVM_DEBUG(dbgs() << "Negated: " << *New << '\n');
1125 return New;
1126}
1127
1128/// If this is a shift of a reassociable multiply or is used by one, change
1129/// this into a multiply by a constant to assist with further reassociation.
1130static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1131 Constant *MulCst = ConstantInt::get(Ty: Shl->getType(), V: 1);
1132 auto *SA = cast<ConstantInt>(Val: Shl->getOperand(i: 1));
1133 MulCst = ConstantFoldBinaryInstruction(Opcode: Instruction::Shl, V1: MulCst, V2: SA);
1134 assert(MulCst && "Constant folding of immediate constants failed");
1135
1136 BinaryOperator *Mul = BinaryOperator::CreateMul(V1: Shl->getOperand(i: 0), V2: MulCst,
1137 Name: "", InsertBefore: Shl->getIterator());
1138 Shl->setOperand(i: 0, Val: PoisonValue::get(T: Shl->getType())); // Drop use of op.
1139 Mul->takeName(V: Shl);
1140
1141 // Everyone now refers to the mul instruction.
1142 Shl->replaceAllUsesWith(V: Mul);
1143 Mul->setDebugLoc(Shl->getDebugLoc());
1144
1145 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
1146 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
1147 // handling. It can be preserved as long as we're not left shifting by
1148 // bitwidth - 1.
1149 bool NSW = cast<BinaryOperator>(Val: Shl)->hasNoSignedWrap();
1150 bool NUW = cast<BinaryOperator>(Val: Shl)->hasNoUnsignedWrap();
1151 unsigned BitWidth = Shl->getType()->getScalarSizeInBits();
1152 if (NSW && (NUW || SA->getValue().ult(RHS: BitWidth - 1)))
1153 Mul->setHasNoSignedWrap(true);
1154 Mul->setHasNoUnsignedWrap(NUW);
1155 return Mul;
1156}
1157
1158/// Scan backwards and forwards among values with the same rank as element i
1159/// to see if X exists. If X does not exist, return i. This is useful when
1160/// scanning for 'x' when we see '-x' because they both get the same rank.
1161static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops,
1162 unsigned i, Value *X) {
1163 unsigned XRank = Ops[i].Rank;
1164 unsigned e = Ops.size();
1165 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
1166 if (Ops[j].Op == X)
1167 return j;
1168 if (Instruction *I1 = dyn_cast<Instruction>(Val: Ops[j].Op))
1169 if (Instruction *I2 = dyn_cast<Instruction>(Val: X))
1170 if (I1->isIdenticalTo(I: I2))
1171 return j;
1172 }
1173 // Scan backwards.
1174 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
1175 if (Ops[j].Op == X)
1176 return j;
1177 if (Instruction *I1 = dyn_cast<Instruction>(Val: Ops[j].Op))
1178 if (Instruction *I2 = dyn_cast<Instruction>(Val: X))
1179 if (I1->isIdenticalTo(I: I2))
1180 return j;
1181 }
1182 return i;
1183}
1184
1185/// Emit a tree of add instructions, summing Ops together
1186/// and returning the result. Insert the tree before I.
1187static Value *EmitAddTreeOfValues(Instruction *I,
1188 SmallVectorImpl<WeakTrackingVH> &Ops) {
1189 if (Ops.size() == 1) return Ops.back();
1190
1191 Value *V1 = Ops.pop_back_val();
1192 Value *V2 = EmitAddTreeOfValues(I, Ops);
1193 auto *NewAdd = CreateAdd(S1: V2, S2: V1, Name: "reass.add", InsertBefore: I->getIterator(), FlagsOp: I);
1194 NewAdd->setDebugLoc(I->getDebugLoc());
1195 return NewAdd;
1196}
1197
1198/// If V is an expression tree that is a multiplication sequence,
1199/// and if this sequence contains a multiply by Factor,
1200/// remove Factor from the tree and return the new tree.
1201/// If new instructions are inserted to generate this tree, DL should be used
1202/// as the DebugLoc for these instructions.
1203Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor,
1204 DebugLoc DL) {
1205 BinaryOperator *BO = isReassociableOp(V, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul);
1206 if (!BO)
1207 return nullptr;
1208
1209 SmallVector<RepeatedValue, 8> Tree;
1210 OverflowTracking Flags;
1211 MadeChange |= LinearizeExprTree(I: BO, Ops&: Tree, ToRedo&: RedoInsts, Flags);
1212 SmallVector<ValueEntry, 8> Factors;
1213 Factors.reserve(N: Tree.size());
1214 for (const RepeatedValue &E : Tree)
1215 Factors.append(NumInputs: E.second, Elt: ValueEntry(getRank(V: E.first), E.first));
1216
1217 bool FoundFactor = false;
1218 bool NeedsNegate = false;
1219 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1220 if (Factors[i].Op == Factor) {
1221 FoundFactor = true;
1222 Factors.erase(CI: Factors.begin()+i);
1223 break;
1224 }
1225
1226 // If this is a negative version of this factor, remove it.
1227 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Val: Factor)) {
1228 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Val: Factors[i].Op))
1229 if (FC1->getValue() == -FC2->getValue()) {
1230 FoundFactor = NeedsNegate = true;
1231 Factors.erase(CI: Factors.begin()+i);
1232 break;
1233 }
1234 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Val: Factor)) {
1235 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Val: Factors[i].Op)) {
1236 const APFloat &F1 = FC1->getValueAPF();
1237 APFloat F2(FC2->getValueAPF());
1238 F2.changeSign();
1239 if (F1 == F2) {
1240 FoundFactor = NeedsNegate = true;
1241 Factors.erase(CI: Factors.begin() + i);
1242 break;
1243 }
1244 }
1245 }
1246 }
1247
1248 if (!FoundFactor) {
1249 // Make sure to restore the operands to the expression tree.
1250 RewriteExprTree(I: BO, Ops&: Factors, Flags);
1251 return nullptr;
1252 }
1253
1254 BasicBlock::iterator InsertPt = ++BO->getIterator();
1255
1256 // If this was just a single multiply, remove the multiply and return the only
1257 // remaining operand.
1258 if (Factors.size() == 1) {
1259 RedoInsts.insert(X: BO);
1260 V = Factors[0].Op;
1261 } else {
1262 RewriteExprTree(I: BO, Ops&: Factors, Flags);
1263 V = BO;
1264 }
1265
1266 if (NeedsNegate) {
1267 V = CreateNeg(S1: V, Name: "neg", InsertBefore: InsertPt, FlagsOp: BO);
1268 cast<Instruction>(Val: V)->setDebugLoc(DL);
1269 }
1270
1271 return V;
1272}
1273
1274/// If V is a single-use multiply, recursively add its operands as factors,
1275/// otherwise add V to the list of factors.
1276///
1277/// Ops is the top-level list of add operands we're trying to factor.
1278static void FindSingleUseMultiplyFactors(Value *V,
1279 SmallVectorImpl<Value*> &Factors) {
1280 BinaryOperator *BO = isReassociableOp(V, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul);
1281 if (!BO) {
1282 Factors.push_back(Elt: V);
1283 return;
1284 }
1285
1286 // Otherwise, add the LHS and RHS to the list of factors.
1287 FindSingleUseMultiplyFactors(V: BO->getOperand(i_nocapture: 1), Factors);
1288 FindSingleUseMultiplyFactors(V: BO->getOperand(i_nocapture: 0), Factors);
1289}
1290
1291/// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1292/// This optimizes based on identities. If it can be reduced to a single Value,
1293/// it is returned, otherwise the Ops list is mutated as necessary.
1294static Value *OptimizeAndOrXor(unsigned Opcode,
1295 SmallVectorImpl<ValueEntry> &Ops) {
1296 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1297 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1298 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1299 // First, check for X and ~X in the operand list.
1300 assert(i < Ops.size());
1301 Value *X;
1302 if (match(V: Ops[i].Op, P: m_Not(V: m_Value(V&: X)))) { // Cannot occur for ^.
1303 unsigned FoundX = FindInOperandList(Ops, i, X);
1304 if (FoundX != i) {
1305 if (Opcode == Instruction::And) // ...&X&~X = 0
1306 return Constant::getNullValue(Ty: X->getType());
1307
1308 if (Opcode == Instruction::Or) // ...|X|~X = -1
1309 return Constant::getAllOnesValue(Ty: X->getType());
1310 }
1311 }
1312
1313 // Next, check for duplicate pairs of values, which we assume are next to
1314 // each other, due to our sorting criteria.
1315 assert(i < Ops.size());
1316 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1317 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1318 // Drop duplicate values for And and Or.
1319 Ops.erase(CI: Ops.begin()+i);
1320 --i; --e;
1321 ++NumAnnihil;
1322 continue;
1323 }
1324
1325 // Drop pairs of values for Xor.
1326 assert(Opcode == Instruction::Xor);
1327 if (e == 2)
1328 return Constant::getNullValue(Ty: Ops[0].Op->getType());
1329
1330 // Y ^ X^X -> Y
1331 Ops.erase(CS: Ops.begin()+i, CE: Ops.begin()+i+2);
1332 i -= 1; e -= 2;
1333 ++NumAnnihil;
1334 }
1335 }
1336 return nullptr;
1337}
1338
1339/// Helper function of CombineXorOpnd(). It creates a bitwise-and
1340/// instruction with the given two operands, and return the resulting
1341/// instruction. There are two special cases: 1) if the constant operand is 0,
1342/// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1343/// be returned.
1344static Value *createAndInstr(BasicBlock::iterator InsertBefore, Value *Opnd,
1345 const APInt &ConstOpnd) {
1346 if (ConstOpnd.isZero())
1347 return nullptr;
1348
1349 if (ConstOpnd.isAllOnes())
1350 return Opnd;
1351
1352 Instruction *I = BinaryOperator::CreateAnd(
1353 V1: Opnd, V2: ConstantInt::get(Ty: Opnd->getType(), V: ConstOpnd), Name: "and.ra",
1354 InsertBefore);
1355 I->setDebugLoc(InsertBefore->getDebugLoc());
1356 return I;
1357}
1358
1359// Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1360// into "R ^ C", where C would be 0, and R is a symbolic value.
1361//
1362// If it was successful, true is returned, and the "R" and "C" is returned
1363// via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1364// and both "Res" and "ConstOpnd" remain unchanged.
1365bool ReassociatePass::CombineXorOpnd(BasicBlock::iterator It, XorOpnd *Opnd1,
1366 APInt &ConstOpnd, Value *&Res) {
1367 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1368 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1369 // = (x & ~c1) ^ (c1 ^ c2)
1370 // It is useful only when c1 == c2.
1371 if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isZero())
1372 return false;
1373
1374 if (!Opnd1->getValue()->hasOneUse())
1375 return false;
1376
1377 const APInt &C1 = Opnd1->getConstPart();
1378 if (C1 != ConstOpnd)
1379 return false;
1380
1381 Value *X = Opnd1->getSymbolicPart();
1382 Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: ~C1);
1383 // ConstOpnd was C2, now C1 ^ C2.
1384 ConstOpnd ^= C1;
1385
1386 if (Instruction *T = dyn_cast<Instruction>(Val: Opnd1->getValue()))
1387 RedoInsts.insert(X: T);
1388 return true;
1389}
1390
1391// Helper function of OptimizeXor(). It tries to simplify
1392// "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1393// symbolic value.
1394//
1395// If it was successful, true is returned, and the "R" and "C" is returned
1396// via "Res" and "ConstOpnd", respectively (If the entire expression is
1397// evaluated to a constant, the Res is set to NULL); otherwise, false is
1398// returned, and both "Res" and "ConstOpnd" remain unchanged.
1399bool ReassociatePass::CombineXorOpnd(BasicBlock::iterator It, XorOpnd *Opnd1,
1400 XorOpnd *Opnd2, APInt &ConstOpnd,
1401 Value *&Res) {
1402 Value *X = Opnd1->getSymbolicPart();
1403 if (X != Opnd2->getSymbolicPart())
1404 return false;
1405
1406 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1407 int DeadInstNum = 1;
1408 if (Opnd1->getValue()->hasOneUse())
1409 DeadInstNum++;
1410 if (Opnd2->getValue()->hasOneUse())
1411 DeadInstNum++;
1412
1413 // Xor-Rule 2:
1414 // (x | c1) ^ (x & c2)
1415 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1416 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1417 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1418 //
1419 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1420 if (Opnd2->isOrExpr())
1421 std::swap(a&: Opnd1, b&: Opnd2);
1422
1423 const APInt &C1 = Opnd1->getConstPart();
1424 const APInt &C2 = Opnd2->getConstPart();
1425 APInt C3((~C1) ^ C2);
1426
1427 // Do not increase code size!
1428 if (!C3.isZero() && !C3.isAllOnes()) {
1429 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1430 if (NewInstNum > DeadInstNum)
1431 return false;
1432 }
1433
1434 Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: C3);
1435 ConstOpnd ^= C1;
1436 } else if (Opnd1->isOrExpr()) {
1437 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1438 //
1439 const APInt &C1 = Opnd1->getConstPart();
1440 const APInt &C2 = Opnd2->getConstPart();
1441 APInt C3 = C1 ^ C2;
1442
1443 // Do not increase code size
1444 if (!C3.isZero() && !C3.isAllOnes()) {
1445 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1446 if (NewInstNum > DeadInstNum)
1447 return false;
1448 }
1449
1450 Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: C3);
1451 ConstOpnd ^= C3;
1452 } else {
1453 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1454 //
1455 const APInt &C1 = Opnd1->getConstPart();
1456 const APInt &C2 = Opnd2->getConstPart();
1457 APInt C3 = C1 ^ C2;
1458 Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: C3);
1459 }
1460
1461 // Put the original operands in the Redo list; hope they will be deleted
1462 // as dead code.
1463 if (Instruction *T = dyn_cast<Instruction>(Val: Opnd1->getValue()))
1464 RedoInsts.insert(X: T);
1465 if (Instruction *T = dyn_cast<Instruction>(Val: Opnd2->getValue()))
1466 RedoInsts.insert(X: T);
1467
1468 return true;
1469}
1470
1471/// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1472/// to a single Value, it is returned, otherwise the Ops list is mutated as
1473/// necessary.
1474Value *ReassociatePass::OptimizeXor(Instruction *I,
1475 SmallVectorImpl<ValueEntry> &Ops) {
1476 if (Value *V = OptimizeAndOrXor(Opcode: Instruction::Xor, Ops))
1477 return V;
1478
1479 if (Ops.size() == 1)
1480 return nullptr;
1481
1482 SmallVector<XorOpnd, 8> Opnds;
1483 SmallVector<XorOpnd*, 8> OpndPtrs;
1484 Type *Ty = Ops[0].Op->getType();
1485 APInt ConstOpnd(Ty->getScalarSizeInBits(), 0);
1486
1487 // Step 1: Convert ValueEntry to XorOpnd
1488 for (const ValueEntry &Op : Ops) {
1489 Value *V = Op.Op;
1490 const APInt *C;
1491 // TODO: Support non-splat vectors.
1492 if (match(V, P: m_APInt(Res&: C))) {
1493 ConstOpnd ^= *C;
1494 } else {
1495 XorOpnd O(V);
1496 O.setSymbolicRank(getRank(V: O.getSymbolicPart()));
1497 Opnds.push_back(Elt: O);
1498 }
1499 }
1500
1501 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1502 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1503 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1504 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1505 // when new elements are added to the vector.
1506 for (XorOpnd &Op : Opnds)
1507 OpndPtrs.push_back(Elt: &Op);
1508
1509 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1510 // the same symbolic value cluster together. For instance, the input operand
1511 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1512 // ("x | 123", "x & 789", "y & 456").
1513 //
1514 // The purpose is twofold:
1515 // 1) Cluster together the operands sharing the same symbolic-value.
1516 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
1517 // could potentially shorten crital path, and expose more loop-invariants.
1518 // Note that values' rank are basically defined in RPO order (FIXME).
1519 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1520 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1521 // "z" in the order of X-Y-Z is better than any other orders.
1522 llvm::stable_sort(Range&: OpndPtrs, C: [](XorOpnd *LHS, XorOpnd *RHS) {
1523 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1524 });
1525
1526 // Step 3: Combine adjacent operands
1527 XorOpnd *PrevOpnd = nullptr;
1528 bool Changed = false;
1529 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1530 XorOpnd *CurrOpnd = OpndPtrs[i];
1531 // The combined value
1532 Value *CV;
1533
1534 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1535 if (!ConstOpnd.isZero() &&
1536 CombineXorOpnd(It: I->getIterator(), Opnd1: CurrOpnd, ConstOpnd, Res&: CV)) {
1537 Changed = true;
1538 if (CV)
1539 *CurrOpnd = XorOpnd(CV);
1540 else {
1541 CurrOpnd->Invalidate();
1542 continue;
1543 }
1544 }
1545
1546 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1547 PrevOpnd = CurrOpnd;
1548 continue;
1549 }
1550
1551 // step 3.2: When previous and current operands share the same symbolic
1552 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1553 if (CombineXorOpnd(It: I->getIterator(), Opnd1: CurrOpnd, Opnd2: PrevOpnd, ConstOpnd, Res&: CV)) {
1554 // Remove previous operand
1555 PrevOpnd->Invalidate();
1556 if (CV) {
1557 *CurrOpnd = XorOpnd(CV);
1558 PrevOpnd = CurrOpnd;
1559 } else {
1560 CurrOpnd->Invalidate();
1561 PrevOpnd = nullptr;
1562 }
1563 Changed = true;
1564 }
1565 }
1566
1567 // Step 4: Reassemble the Ops
1568 if (Changed) {
1569 Ops.clear();
1570 for (const XorOpnd &O : Opnds) {
1571 if (O.isInvalid())
1572 continue;
1573 ValueEntry VE(getRank(V: O.getValue()), O.getValue());
1574 Ops.push_back(Elt: VE);
1575 }
1576 if (!ConstOpnd.isZero()) {
1577 Value *C = ConstantInt::get(Ty, V: ConstOpnd);
1578 ValueEntry VE(getRank(V: C), C);
1579 Ops.push_back(Elt: VE);
1580 }
1581 unsigned Sz = Ops.size();
1582 if (Sz == 1)
1583 return Ops.back().Op;
1584 if (Sz == 0) {
1585 assert(ConstOpnd.isZero());
1586 return ConstantInt::get(Ty, V: ConstOpnd);
1587 }
1588 }
1589
1590 return nullptr;
1591}
1592
1593/// Optimize a series of operands to an 'add' instruction. This
1594/// optimizes based on identities. If it can be reduced to a single Value, it
1595/// is returned, otherwise the Ops list is mutated as necessary.
1596Value *ReassociatePass::OptimizeAdd(Instruction *I,
1597 SmallVectorImpl<ValueEntry> &Ops) {
1598 // Scan the operand lists looking for X and -X pairs. If we find any, we
1599 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1600 // scan for any
1601 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1602
1603 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1604 Value *TheOp = Ops[i].Op;
1605 // Check to see if we've seen this operand before. If so, we factor all
1606 // instances of the operand together. Due to our sorting criteria, we know
1607 // that these need to be next to each other in the vector.
1608 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1609 // Rescan the list, remove all instances of this operand from the expr.
1610 unsigned NumFound = 0;
1611 do {
1612 Ops.erase(CI: Ops.begin()+i);
1613 ++NumFound;
1614 } while (i != Ops.size() && Ops[i].Op == TheOp);
1615
1616 LLVM_DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp
1617 << '\n');
1618 ++NumFactor;
1619
1620 // Insert a new multiply.
1621 Type *Ty = TheOp->getType();
1622 // Truncate if NumFound overflows the type.
1623 Constant *C = Ty->isIntOrIntVectorTy()
1624 ? ConstantInt::get(Ty, V: NumFound, /*IsSigned=*/false,
1625 /*ImplicitTrunc=*/true)
1626 : ConstantFP::get(Ty, V: NumFound);
1627 Instruction *Mul = CreateMul(S1: TheOp, S2: C, Name: "factor", InsertBefore: I->getIterator(), FlagsOp: I);
1628 Mul->setDebugLoc(I->getDebugLoc());
1629
1630 // Now that we have inserted a multiply, optimize it. This allows us to
1631 // handle cases that require multiple factoring steps, such as this:
1632 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1633 RedoInsts.insert(X: Mul);
1634
1635 // If every add operand was a duplicate, return the multiply.
1636 if (Ops.empty())
1637 return Mul;
1638
1639 // Otherwise, we had some input that didn't have the dupe, such as
1640 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1641 // things being added by this operation.
1642 Ops.insert(I: Ops.begin(), Elt: ValueEntry(getRank(V: Mul), Mul));
1643
1644 --i;
1645 e = Ops.size();
1646 continue;
1647 }
1648
1649 // Check for X and -X or X and ~X in the operand list.
1650 Value *X;
1651 if (!match(V: TheOp, P: m_Neg(V: m_Value(V&: X))) && !match(V: TheOp, P: m_Not(V: m_Value(V&: X))) &&
1652 !match(V: TheOp, P: m_FNeg(X: m_Value(V&: X))))
1653 continue;
1654
1655 unsigned FoundX = FindInOperandList(Ops, i, X);
1656 if (FoundX == i)
1657 continue;
1658
1659 // Remove X and -X from the operand list.
1660 if (Ops.size() == 2 &&
1661 (match(V: TheOp, P: m_Neg(V: m_Value())) || match(V: TheOp, P: m_FNeg(X: m_Value()))))
1662 return Constant::getNullValue(Ty: X->getType());
1663
1664 // Remove X and ~X from the operand list.
1665 if (Ops.size() == 2 && match(V: TheOp, P: m_Not(V: m_Value())))
1666 return Constant::getAllOnesValue(Ty: X->getType());
1667
1668 Ops.erase(CI: Ops.begin()+i);
1669 if (i < FoundX)
1670 --FoundX;
1671 else
1672 --i; // Need to back up an extra one.
1673 Ops.erase(CI: Ops.begin()+FoundX);
1674 ++NumAnnihil;
1675 --i; // Revisit element.
1676 e -= 2; // Removed two elements.
1677
1678 // if X and ~X we append -1 to the operand list.
1679 if (match(V: TheOp, P: m_Not(V: m_Value()))) {
1680 Value *V = Constant::getAllOnesValue(Ty: X->getType());
1681 Ops.insert(I: Ops.end(), Elt: ValueEntry(getRank(V), V));
1682 e += 1;
1683 }
1684 }
1685
1686 // Scan the operand list, checking to see if there are any common factors
1687 // between operands. Consider something like A*A+A*B*C+D. We would like to
1688 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1689 // To efficiently find this, we count the number of times a factor occurs
1690 // for any ADD operands that are MULs.
1691 DenseMap<Value*, unsigned> FactorOccurrences;
1692
1693 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1694 // where they are actually the same multiply.
1695 unsigned MaxOcc = 0;
1696 Value *MaxOccVal = nullptr;
1697
1698 // Prefer a non-constant factor over a constant when occurrence counts
1699 // tie. Factoring out a variable (e.g., X from X*C1 + X*C2) exposes
1700 // downstream constant folding; factoring out a constant does not.
1701 auto IsBetterFactor = [](Value *Factor, Value *MaxOccVal, unsigned Occ,
1702 unsigned MaxOcc) {
1703 return Occ > MaxOcc ||
1704 (Occ == MaxOcc &&
1705 (isa<Instruction>(Val: Factor) || isa<Argument>(Val: Factor)) &&
1706 isa<Constant>(Val: MaxOccVal) && !isa<UndefValue>(Val: MaxOccVal));
1707 };
1708 auto CountFactors = [&](BinaryOperator *BOp) {
1709 // Compute all of the factors of this added value.
1710 SmallVector<Value*, 8> Factors;
1711 FindSingleUseMultiplyFactors(V: BOp, Factors);
1712 assert(Factors.size() > 1 && "Bad linearize!");
1713
1714 // Add one to FactorOccurrences for each unique factor in this op.
1715 SmallPtrSet<Value*, 8> Duplicates;
1716 for (Value *Factor : Factors) {
1717 if (!Duplicates.insert(Ptr: Factor).second)
1718 continue;
1719
1720 unsigned Occ = ++FactorOccurrences[Factor];
1721 if (IsBetterFactor(Factor, MaxOccVal, Occ, MaxOcc)) {
1722 MaxOcc = Occ;
1723 MaxOccVal = Factor;
1724 }
1725
1726 // If Factor is a negative constant, add the negated value as a factor
1727 // because we can percolate the negate out. Watch for minint, which
1728 // cannot be positivified.
1729 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Factor)) {
1730 if (CI->isNegative() && !CI->isMinValue(IsSigned: true)) {
1731 Factor = ConstantInt::get(Context&: CI->getContext(), V: -CI->getValue());
1732 if (!Duplicates.insert(Ptr: Factor).second)
1733 continue;
1734 unsigned Occ = ++FactorOccurrences[Factor];
1735 if (IsBetterFactor(Factor, MaxOccVal, Occ, MaxOcc)) {
1736 MaxOcc = Occ;
1737 MaxOccVal = Factor;
1738 }
1739 }
1740 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Val: Factor)) {
1741 if (CF->isNegative()) {
1742 APFloat F(CF->getValueAPF());
1743 F.changeSign();
1744 Factor = ConstantFP::get(Ty: CF->getType(), V: F);
1745 if (!Duplicates.insert(Ptr: Factor).second)
1746 continue;
1747 unsigned Occ = ++FactorOccurrences[Factor];
1748 if (IsBetterFactor(Factor, MaxOccVal, Occ, MaxOcc)) {
1749 MaxOcc = Occ;
1750 MaxOccVal = Factor;
1751 }
1752 }
1753 }
1754 }
1755 };
1756
1757 // fmul/fadd pairs kept together for fma hide their muls; count the factors
1758 // of the reassociable ones as well and break those pairs up if a repeated
1759 // factor exists, so that factorization still applies.
1760 SmallVector<Value *> FMulAddCands;
1761 for (const ValueEntry &Entry : Ops) {
1762 if (BinaryOperator *BOp =
1763 isReassociableOp(V: Entry.Op, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul)) {
1764 CountFactors(BOp);
1765 continue;
1766 }
1767 if (BinaryOperator *BOp = isFMulAddCandidate(V: Entry.Op);
1768 BOp && hasFPAssociativeFlags(I: BOp)) {
1769 FMulAddCands.push_back(Elt: Entry.Op);
1770 CountFactors(BOp);
1771 }
1772 }
1773
1774 if (MaxOcc > 1) {
1775 for (Value *V : FMulAddCands) {
1776 erase_if(C&: Ops, P: [V](const ValueEntry &E) { return E.Op == V; });
1777 for (Value *Op : cast<BinaryOperator>(Val: V)->operands())
1778 Ops.emplace_back(Args: getRank(V: Op), Args&: Op);
1779 }
1780 }
1781
1782 // If any factor occurred more than one time, we can pull it out.
1783 if (MaxOcc > 1) {
1784 LLVM_DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal
1785 << '\n');
1786 ++NumFactor;
1787
1788 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1789 // this, we could otherwise run into situations where removing a factor
1790 // from an expression will drop a use of maxocc, and this can cause
1791 // RemoveFactorFromExpression on successive values to behave differently.
1792 Instruction *DummyInst =
1793 I->getType()->isIntOrIntVectorTy()
1794 ? BinaryOperator::CreateAdd(V1: MaxOccVal, V2: MaxOccVal)
1795 : BinaryOperator::CreateFAdd(V1: MaxOccVal, V2: MaxOccVal);
1796
1797 SmallVector<WeakTrackingVH, 4> NewMulOps;
1798 for (unsigned i = 0; i != Ops.size(); ++i) {
1799 // Only try to remove factors from expressions we're allowed to.
1800 BinaryOperator *BOp =
1801 isReassociableOp(V: Ops[i].Op, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul);
1802 if (!BOp)
1803 continue;
1804
1805 if (Value *V = RemoveFactorFromExpression(V: Ops[i].Op, Factor: MaxOccVal,
1806 DL: I->getDebugLoc())) {
1807 // The factorized operand may occur several times. Convert them all in
1808 // one fell swoop.
1809 for (unsigned j = Ops.size(); j != i;) {
1810 --j;
1811 if (Ops[j].Op == Ops[i].Op) {
1812 NewMulOps.push_back(Elt: V);
1813 Ops.erase(CI: Ops.begin()+j);
1814 }
1815 }
1816 --i;
1817 }
1818 }
1819
1820 // No need for extra uses anymore.
1821 DummyInst->deleteValue();
1822
1823 unsigned NumAddedValues = NewMulOps.size();
1824 Value *V = EmitAddTreeOfValues(I, Ops&: NewMulOps);
1825
1826 // Now that we have inserted the add tree, optimize it. This allows us to
1827 // handle cases that require multiple factoring steps, such as this:
1828 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1829 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1830 (void)NumAddedValues;
1831 if (Instruction *VI = dyn_cast<Instruction>(Val: V))
1832 RedoInsts.insert(X: VI);
1833
1834 // Create the multiply.
1835 Instruction *V2 = CreateMul(S1: V, S2: MaxOccVal, Name: "reass.mul", InsertBefore: I->getIterator(), FlagsOp: I);
1836 V2->setDebugLoc(I->getDebugLoc());
1837
1838 // Rerun associate on the multiply in case the inner expression turned into
1839 // a multiply. We want to make sure that we keep things in canonical form.
1840 RedoInsts.insert(X: V2);
1841
1842 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1843 // entire result expression is just the multiply "A*(B+C)".
1844 if (Ops.empty())
1845 return V2;
1846
1847 // Otherwise, we had some input that didn't have the factor, such as
1848 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1849 // things being added by this operation.
1850 Ops.insert(I: Ops.begin(), Elt: ValueEntry(getRank(V: V2), V2));
1851 }
1852
1853 return nullptr;
1854}
1855
1856/// Build up a vector of value/power pairs factoring a product.
1857///
1858/// Given a series of multiplication operands, build a vector of factors and
1859/// the powers each is raised to when forming the final product. Sort them in
1860/// the order of descending power.
1861///
1862/// (x*x) -> [(x, 2)]
1863/// ((x*x)*x) -> [(x, 3)]
1864/// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1865///
1866/// \returns Whether any factors have a power greater than one.
1867static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1868 SmallVectorImpl<Factor> &Factors) {
1869 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1870 // Compute the sum of powers of simplifiable factors.
1871 unsigned FactorPowerSum = 0;
1872 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1873 Value *Op = Ops[Idx-1].Op;
1874
1875 // Count the number of occurrences of this value.
1876 unsigned Count = 1;
1877 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1878 ++Count;
1879 // Track for simplification all factors which occur 2 or more times.
1880 if (Count > 1)
1881 FactorPowerSum += Count;
1882 }
1883
1884 // We can only simplify factors if the sum of the powers of our simplifiable
1885 // factors is 4 or higher. When that is the case, we will *always* have
1886 // a simplification. This is an important invariant to prevent cyclicly
1887 // trying to simplify already minimal formations.
1888 if (FactorPowerSum < 4)
1889 return false;
1890
1891 // Now gather the simplifiable factors, removing them from Ops.
1892 FactorPowerSum = 0;
1893 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1894 Value *Op = Ops[Idx-1].Op;
1895
1896 // Count the number of occurrences of this value.
1897 unsigned Count = 1;
1898 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1899 ++Count;
1900 if (Count == 1)
1901 continue;
1902 // Move an even number of occurrences to Factors.
1903 Count &= ~1U;
1904 Idx -= Count;
1905 FactorPowerSum += Count;
1906 Factors.push_back(Elt: Factor(Op, Count));
1907 Ops.erase(CS: Ops.begin()+Idx, CE: Ops.begin()+Idx+Count);
1908 }
1909
1910 // None of the adjustments above should have reduced the sum of factor powers
1911 // below our mininum of '4'.
1912 assert(FactorPowerSum >= 4);
1913
1914 llvm::stable_sort(Range&: Factors, C: [](const Factor &LHS, const Factor &RHS) {
1915 return LHS.Power > RHS.Power;
1916 });
1917 return true;
1918}
1919
1920/// Build a tree of multiplies, computing the product of Ops.
1921static Value *buildMultiplyTree(IRBuilderBase &Builder,
1922 SmallVectorImpl<Value*> &Ops) {
1923 if (Ops.size() == 1)
1924 return Ops.back();
1925
1926 Value *LHS = Ops.pop_back_val();
1927 do {
1928 if (LHS->getType()->isIntOrIntVectorTy())
1929 LHS = Builder.CreateMul(LHS, RHS: Ops.pop_back_val());
1930 else
1931 LHS = Builder.CreateFMul(L: LHS, R: Ops.pop_back_val());
1932 } while (!Ops.empty());
1933
1934 return LHS;
1935}
1936
1937/// Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1938///
1939/// Given a vector of values raised to various powers, where no two values are
1940/// equal and the powers are sorted in decreasing order, compute the minimal
1941/// DAG of multiplies to compute the final product, and return that product
1942/// value.
1943Value *
1944ReassociatePass::buildMinimalMultiplyDAG(IRBuilderBase &Builder,
1945 SmallVectorImpl<Factor> &Factors) {
1946 assert(Factors[0].Power);
1947 SmallVector<Value *, 4> OuterProduct;
1948 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1949 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1950 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1951 LastIdx = Idx;
1952 continue;
1953 }
1954
1955 // We want to multiply across all the factors with the same power so that
1956 // we can raise them to that power as a single entity. Build a mini tree
1957 // for that.
1958 SmallVector<Value *, 4> InnerProduct;
1959 InnerProduct.push_back(Elt: Factors[LastIdx].Base);
1960 do {
1961 InnerProduct.push_back(Elt: Factors[Idx].Base);
1962 ++Idx;
1963 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1964
1965 // Reset the base value of the first factor to the new expression tree.
1966 // We'll remove all the factors with the same power in a second pass.
1967 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, Ops&: InnerProduct);
1968 if (Instruction *MI = dyn_cast<Instruction>(Val: M))
1969 RedoInsts.insert(X: MI);
1970
1971 LastIdx = Idx;
1972 }
1973 // Unique factors with equal powers -- we've folded them into the first one's
1974 // base.
1975 Factors.erase(CS: llvm::unique(R&: Factors,
1976 P: [](const Factor &LHS, const Factor &RHS) {
1977 return LHS.Power == RHS.Power;
1978 }),
1979 CE: Factors.end());
1980
1981 // Iteratively collect the base of each factor with an add power into the
1982 // outer product, and halve each power in preparation for squaring the
1983 // expression.
1984 for (Factor &F : Factors) {
1985 if (F.Power & 1)
1986 OuterProduct.push_back(Elt: F.Base);
1987 F.Power >>= 1;
1988 }
1989 if (Factors[0].Power) {
1990 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1991 OuterProduct.push_back(Elt: SquareRoot);
1992 OuterProduct.push_back(Elt: SquareRoot);
1993 }
1994 if (OuterProduct.size() == 1)
1995 return OuterProduct.front();
1996
1997 Value *V = buildMultiplyTree(Builder, Ops&: OuterProduct);
1998 return V;
1999}
2000
2001Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
2002 SmallVectorImpl<ValueEntry> &Ops) {
2003 // We can only optimize the multiplies when there is a chain of more than
2004 // three, such that a balanced tree might require fewer total multiplies.
2005 if (Ops.size() < 4)
2006 return nullptr;
2007
2008 // Try to turn linear trees of multiplies without other uses of the
2009 // intermediate stages into minimal multiply DAGs with perfect sub-expression
2010 // re-use.
2011 SmallVector<Factor, 4> Factors;
2012 if (!collectMultiplyFactors(Ops, Factors))
2013 return nullptr; // All distinct factors, so nothing left for us to do.
2014
2015 IRBuilder<> Builder(I);
2016 // The reassociate transformation for FP operations is performed only
2017 // if unsafe algebra is permitted by FastMathFlags. Propagate those flags
2018 // to the newly generated operations.
2019 if (auto FPI = dyn_cast<FPMathOperator>(Val: I))
2020 Builder.setFastMathFlags(FPI->getFastMathFlags());
2021
2022 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
2023 if (Ops.empty())
2024 return V;
2025
2026 ValueEntry NewEntry = ValueEntry(getRank(V), V);
2027 Ops.insert(I: llvm::lower_bound(Range&: Ops, Value&: NewEntry), Elt: NewEntry);
2028 return nullptr;
2029}
2030
2031Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
2032 SmallVectorImpl<ValueEntry> &Ops) {
2033 // Now that we have the linearized expression tree, try to optimize it.
2034 // Start by folding any constants that we found.
2035 const DataLayout &DL = I->getDataLayout();
2036 Constant *Cst = nullptr;
2037 unsigned Opcode = I->getOpcode();
2038 while (!Ops.empty()) {
2039 if (auto *C = dyn_cast<Constant>(Val: Ops.back().Op)) {
2040 if (!Cst) {
2041 Ops.pop_back();
2042 Cst = C;
2043 continue;
2044 }
2045 if (Constant *Res = ConstantFoldBinaryOpOperands(Opcode, LHS: C, RHS: Cst, DL)) {
2046 Ops.pop_back();
2047 Cst = Res;
2048 continue;
2049 }
2050 }
2051 break;
2052 }
2053 // If there was nothing but constants then we are done.
2054 if (Ops.empty())
2055 return Cst;
2056
2057 // Put the combined constant back at the end of the operand list, except if
2058 // there is no point. For example, an add of 0 gets dropped here, while a
2059 // multiplication by zero turns the whole expression into zero.
2060 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, Ty: I->getType())) {
2061 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, Ty: I->getType()))
2062 return Cst;
2063 Ops.push_back(Elt: ValueEntry(0, Cst));
2064 }
2065
2066 if (Ops.size() == 1) return Ops[0].Op;
2067
2068 // Handle destructive annihilation due to identities between elements in the
2069 // argument list here.
2070 unsigned NumOps = Ops.size();
2071 switch (Opcode) {
2072 default: break;
2073 case Instruction::And:
2074 case Instruction::Or:
2075 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
2076 return Result;
2077 break;
2078
2079 case Instruction::Xor:
2080 if (Value *Result = OptimizeXor(I, Ops))
2081 return Result;
2082 break;
2083
2084 case Instruction::Add:
2085 case Instruction::FAdd:
2086 if (Value *Result = OptimizeAdd(I, Ops))
2087 return Result;
2088 break;
2089
2090 case Instruction::Mul:
2091 case Instruction::FMul:
2092 if (Value *Result = OptimizeMul(I, Ops))
2093 return Result;
2094 break;
2095 }
2096
2097 if (Ops.size() != NumOps)
2098 return OptimizeExpression(I, Ops);
2099 return nullptr;
2100}
2101
2102// Remove dead instructions and if any operands are trivially dead add them to
2103// Insts so they will be removed as well.
2104void ReassociatePass::RecursivelyEraseDeadInsts(Instruction *I,
2105 OrderedSet &Insts) {
2106 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
2107 SmallVector<Value *, 4> Ops(I->operands());
2108 ValueRankMap.erase(Val: I);
2109 Insts.remove(X: I);
2110 RedoInsts.remove(X: I);
2111 if (UA)
2112 UA->forgetValue(V: I);
2113 llvm::salvageDebugInfo(I&: *I);
2114 I->eraseFromParent();
2115 for (auto *Op : Ops)
2116 if (Instruction *OpInst = dyn_cast<Instruction>(Val: Op))
2117 if (OpInst->use_empty())
2118 Insts.insert(X: OpInst);
2119}
2120
2121/// Zap the given instruction, adding interesting operands to the work list.
2122void ReassociatePass::EraseInst(Instruction *I) {
2123 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
2124 LLVM_DEBUG(dbgs() << "Erasing dead inst: "; I->dump());
2125
2126 SmallVector<Value *, 8> Ops(I->operands());
2127 // Erase the dead instruction.
2128 ValueRankMap.erase(Val: I);
2129 RedoInsts.remove(X: I);
2130 if (UA)
2131 UA->forgetValue(V: I);
2132 llvm::salvageDebugInfo(I&: *I);
2133 I->eraseFromParent();
2134 // Optimize its operands.
2135 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
2136 for (Value *V : Ops)
2137 if (Instruction *Op = dyn_cast<Instruction>(Val: V)) {
2138 // If this is a node in an expression tree, climb to the expression root
2139 // and add that since that's where optimization actually happens.
2140 unsigned Opcode = Op->getOpcode();
2141 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
2142 Visited.insert(Ptr: Op).second)
2143 Op = Op->user_back();
2144
2145 // The instruction we're going to push may be coming from a
2146 // dead block, and Reassociate skips the processing of unreachable
2147 // blocks because it's a waste of time and also because it can
2148 // lead to infinite loop due to LLVM's non-standard definition
2149 // of dominance.
2150 if (ValueRankMap.contains(Val: Op))
2151 RedoInsts.insert(X: Op);
2152 }
2153
2154 MadeChange = true;
2155}
2156
2157/// Recursively analyze an expression to build a list of instructions that have
2158/// negative floating-point constant operands. The caller can then transform
2159/// the list to create positive constants for better reassociation and CSE.
2160static void getNegatibleInsts(Value *V,
2161 SmallVectorImpl<Instruction *> &Candidates) {
2162 // Handle only one-use instructions. Combining negations does not justify
2163 // replicating instructions.
2164 Instruction *I;
2165 if (!match(V, P: m_OneUse(SubPattern: m_Instruction(I))))
2166 return;
2167
2168 // Handle expressions of multiplications and divisions.
2169 // TODO: This could look through floating-point casts.
2170 const APFloat *C;
2171 switch (I->getOpcode()) {
2172 case Instruction::FMul:
2173 // Not expecting non-canonical code here. Bail out and wait.
2174 if (match(V: I->getOperand(i: 0), P: m_Constant()))
2175 break;
2176
2177 if (match(V: I->getOperand(i: 1), P: m_APFloat(Res&: C)) && C->isNegative()) {
2178 Candidates.push_back(Elt: I);
2179 LLVM_DEBUG(dbgs() << "FMul with negative constant: " << *I << '\n');
2180 }
2181 getNegatibleInsts(V: I->getOperand(i: 0), Candidates);
2182 getNegatibleInsts(V: I->getOperand(i: 1), Candidates);
2183 break;
2184 case Instruction::FDiv:
2185 // Not expecting non-canonical code here. Bail out and wait.
2186 if (match(V: I->getOperand(i: 0), P: m_Constant()) &&
2187 match(V: I->getOperand(i: 1), P: m_Constant()))
2188 break;
2189
2190 if ((match(V: I->getOperand(i: 0), P: m_APFloat(Res&: C)) && C->isNegative()) ||
2191 (match(V: I->getOperand(i: 1), P: m_APFloat(Res&: C)) && C->isNegative())) {
2192 Candidates.push_back(Elt: I);
2193 LLVM_DEBUG(dbgs() << "FDiv with negative constant: " << *I << '\n');
2194 }
2195 getNegatibleInsts(V: I->getOperand(i: 0), Candidates);
2196 getNegatibleInsts(V: I->getOperand(i: 1), Candidates);
2197 break;
2198 default:
2199 break;
2200 }
2201}
2202
2203/// Given an fadd/fsub with an operand that is a one-use instruction
2204/// (the fadd/fsub), try to change negative floating-point constants into
2205/// positive constants to increase potential for reassociation and CSE.
2206Instruction *ReassociatePass::canonicalizeNegFPConstantsForOp(Instruction *I,
2207 Instruction *Op,
2208 Value *OtherOp) {
2209 assert((I->getOpcode() == Instruction::FAdd ||
2210 I->getOpcode() == Instruction::FSub) && "Expected fadd/fsub");
2211
2212 // Collect instructions with negative FP constants from the subtree that ends
2213 // in Op.
2214 SmallVector<Instruction *, 4> Candidates;
2215 getNegatibleInsts(V: Op, Candidates);
2216 if (Candidates.empty())
2217 return nullptr;
2218
2219 // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the
2220 // resulting subtract will be broken up later. This can get us into an
2221 // infinite loop during reassociation.
2222 bool IsFSub = I->getOpcode() == Instruction::FSub;
2223 bool NeedsSubtract = !IsFSub && Candidates.size() % 2 == 1;
2224 if (NeedsSubtract && ShouldBreakUpSubtract(Sub: I))
2225 return nullptr;
2226
2227 for (Instruction *Negatible : Candidates) {
2228 const APFloat *C;
2229 if (match(V: Negatible->getOperand(i: 0), P: m_APFloat(Res&: C))) {
2230 assert(!match(Negatible->getOperand(1), m_Constant()) &&
2231 "Expecting only 1 constant operand");
2232 assert(C->isNegative() && "Expected negative FP constant");
2233 Negatible->setOperand(i: 0, Val: ConstantFP::get(Ty: Negatible->getType(), V: abs(X: *C)));
2234 MadeChange = true;
2235 }
2236 if (match(V: Negatible->getOperand(i: 1), P: m_APFloat(Res&: C))) {
2237 assert(!match(Negatible->getOperand(0), m_Constant()) &&
2238 "Expecting only 1 constant operand");
2239 assert(C->isNegative() && "Expected negative FP constant");
2240 Negatible->setOperand(i: 1, Val: ConstantFP::get(Ty: Negatible->getType(), V: abs(X: *C)));
2241 MadeChange = true;
2242 }
2243 }
2244 assert(MadeChange == true && "Negative constant candidate was not changed");
2245
2246 // Negations cancelled out.
2247 if (Candidates.size() % 2 == 0)
2248 return I;
2249
2250 // Negate the final operand in the expression by flipping the opcode of this
2251 // fadd/fsub.
2252 assert(Candidates.size() % 2 == 1 && "Expected odd number");
2253 IRBuilder<> Builder(I);
2254 Value *NewInst = IsFSub ? Builder.CreateFAddFMF(L: OtherOp, R: Op, FMFSource: I)
2255 : Builder.CreateFSubFMF(L: OtherOp, R: Op, FMFSource: I);
2256 I->replaceAllUsesWith(V: NewInst);
2257 RedoInsts.insert(X: I);
2258 return dyn_cast<Instruction>(Val: NewInst);
2259}
2260
2261/// Canonicalize expressions that contain a negative floating-point constant
2262/// of the following form:
2263/// OtherOp + (subtree) -> OtherOp {+/-} (canonical subtree)
2264/// (subtree) + OtherOp -> OtherOp {+/-} (canonical subtree)
2265/// OtherOp - (subtree) -> OtherOp {+/-} (canonical subtree)
2266///
2267/// The fadd/fsub opcode may be switched to allow folding a negation into the
2268/// input instruction.
2269Instruction *ReassociatePass::canonicalizeNegFPConstants(Instruction *I) {
2270 LLVM_DEBUG(dbgs() << "Combine negations for: " << *I << '\n');
2271 Value *X;
2272 Instruction *Op;
2273 if (match(V: I, P: m_FAdd(L: m_Value(V&: X), R: m_OneUse(SubPattern: m_Instruction(I&: Op)))))
2274 if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, OtherOp: X))
2275 I = R;
2276 if (match(V: I, P: m_FAdd(L: m_OneUse(SubPattern: m_Instruction(I&: Op)), R: m_Value(V&: X))))
2277 if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, OtherOp: X))
2278 I = R;
2279 if (match(V: I, P: m_FSub(L: m_Value(V&: X), R: m_OneUse(SubPattern: m_Instruction(I&: Op)))))
2280 if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, OtherOp: X))
2281 I = R;
2282 return I;
2283}
2284
2285/// Inspect and optimize the given instruction. Note that erasing
2286/// instructions is not allowed.
2287void ReassociatePass::OptimizeInst(Instruction *I) {
2288 // Only consider operations that we understand.
2289 if (!isa<UnaryOperator>(Val: I) && !isa<BinaryOperator>(Val: I))
2290 return;
2291
2292 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(Val: I->getOperand(i: 1)))
2293 // If an operand of this shift is a reassociable multiply, or if the shift
2294 // is used by a reassociable multiply or add, turn into a multiply.
2295 if (isReassociableOp(V: I->getOperand(i: 0), Opcode: Instruction::Mul) ||
2296 (I->hasOneUse() &&
2297 (isReassociableOp(V: I->user_back(), Opcode: Instruction::Mul) ||
2298 isReassociableOp(V: I->user_back(), Opcode: Instruction::Add)))) {
2299 Instruction *NI = ConvertShiftToMul(Shl: I);
2300 RedoInsts.insert(X: I);
2301 MadeChange = true;
2302 I = NI;
2303 }
2304
2305 // Commute binary operators, to canonicalize the order of their operands.
2306 // This can potentially expose more CSE opportunities, and makes writing other
2307 // transformations simpler.
2308 if (I->isCommutative())
2309 canonicalizeOperands(I);
2310
2311 // Canonicalize negative constants out of expressions.
2312 if (Instruction *Res = canonicalizeNegFPConstants(I))
2313 I = Res;
2314
2315 // Don't optimize floating-point instructions unless they have the
2316 // appropriate FastMathFlags for reassociation enabled.
2317 if (isa<FPMathOperator>(Val: I) && !hasFPAssociativeFlags(I))
2318 return;
2319
2320 // Do not reassociate boolean (i1/vXi1) expressions. We want to preserve the
2321 // original order of evaluation for short-circuited comparisons that
2322 // SimplifyCFG has folded to AND/OR expressions. If the expression
2323 // is not further optimized, it is likely to be transformed back to a
2324 // short-circuited form for code gen, and the source order may have been
2325 // optimized for the most likely conditions. For vector boolean expressions,
2326 // we should be optimizing for ILP and not serializing the logical operations.
2327 if (I->getType()->isIntOrIntVectorTy(BitWidth: 1))
2328 return;
2329
2330 // If this is a bitwise or instruction of operands
2331 // with no common bits set, convert it to X+Y.
2332 if (I->getOpcode() == Instruction::Or &&
2333 shouldConvertOrWithNoCommonBitsToAdd(Or: I) && !isLoadCombineCandidate(Or: I) &&
2334 (cast<PossiblyDisjointInst>(Val: I)->isDisjoint() ||
2335 haveNoCommonBitsSet(LHSCache: I->getOperand(i: 0), RHSCache: I->getOperand(i: 1),
2336 SQ: SimplifyQuery(I->getDataLayout(),
2337 /*DT=*/nullptr, /*AC=*/nullptr, I)))) {
2338 Instruction *NI = convertOrWithNoCommonBitsToAdd(Or: I);
2339 RedoInsts.insert(X: I);
2340 MadeChange = true;
2341 I = NI;
2342 }
2343
2344 if (I->getOpcode() == Instruction::Mul && ShouldBreakUpDistribution(Mul: I)) {
2345 Instruction *MulUser = cast<Instruction>(Val: I->user_back());
2346 Instruction *NI = BreakUpDistribute(Mul: I, ToRedo&: RedoInsts);
2347 RedoInsts.insert(X: I);
2348 RedoInsts.insert(X: MulUser);
2349 MadeChange = true;
2350 I = NI;
2351 }
2352
2353 // If this is a subtract instruction which is not already in negate form,
2354 // see if we can convert it to X+-Y.
2355 if (I->getOpcode() == Instruction::Sub) {
2356 if (ShouldBreakUpSubtract(Sub: I)) {
2357 Instruction *NI = BreakUpSubtract(Sub: I, ToRedo&: RedoInsts);
2358 RedoInsts.insert(X: I);
2359 MadeChange = true;
2360 I = NI;
2361 } else if (match(V: I, P: m_Neg(V: m_Value()))) {
2362 // Otherwise, this is a negation. See if the operand is a multiply tree
2363 // and if this is not an inner node of a multiply tree.
2364 if (isReassociableOp(V: I->getOperand(i: 1), Opcode: Instruction::Mul) &&
2365 (!I->hasOneUse() ||
2366 !isReassociableOp(V: I->user_back(), Opcode: Instruction::Mul))) {
2367 Instruction *NI = LowerNegateToMultiply(Neg: I);
2368 // If the negate was simplified, revisit the users to see if we can
2369 // reassociate further.
2370 for (User *U : NI->users()) {
2371 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(Val: U))
2372 RedoInsts.insert(X: Tmp);
2373 }
2374 RedoInsts.insert(X: I);
2375 MadeChange = true;
2376 I = NI;
2377 }
2378 }
2379 } else if (I->getOpcode() == Instruction::FNeg ||
2380 I->getOpcode() == Instruction::FSub) {
2381 if (ShouldBreakUpSubtract(Sub: I)) {
2382 Instruction *NI = BreakUpSubtract(Sub: I, ToRedo&: RedoInsts);
2383 RedoInsts.insert(X: I);
2384 MadeChange = true;
2385 I = NI;
2386 } else if (match(V: I, P: m_FNeg(X: m_Value()))) {
2387 // Otherwise, this is a negation. See if the operand is a multiply tree
2388 // and if this is not an inner node of a multiply tree.
2389 Value *Op = isa<BinaryOperator>(Val: I) ? I->getOperand(i: 1) :
2390 I->getOperand(i: 0);
2391 if (isReassociableOp(V: Op, Opcode: Instruction::FMul) &&
2392 (!I->hasOneUse() ||
2393 !isReassociableOp(V: I->user_back(), Opcode: Instruction::FMul))) {
2394 // If the negate was simplified, revisit the users to see if we can
2395 // reassociate further.
2396 Instruction *NI = LowerNegateToMultiply(Neg: I);
2397 for (User *U : NI->users()) {
2398 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(Val: U))
2399 RedoInsts.insert(X: Tmp);
2400 }
2401 RedoInsts.insert(X: I);
2402 MadeChange = true;
2403 I = NI;
2404 }
2405 }
2406 }
2407
2408 // If this instruction is an associative binary operator, process it.
2409 if (!I->isAssociative()) return;
2410 BinaryOperator *BO = cast<BinaryOperator>(Val: I);
2411
2412 // If this is an interior node of a reassociable tree, ignore it until we
2413 // get to the root of the tree, to avoid N^2 analysis.
2414 unsigned Opcode = BO->getOpcode();
2415 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2416 // During the initial run we will get to the root of the tree.
2417 // But if we get here while we are redoing instructions, there is no
2418 // guarantee that the root will be visited. So Redo later
2419 if (BO->user_back() != BO &&
2420 BO->getParent() == BO->user_back()->getParent())
2421 RedoInsts.insert(X: BO->user_back());
2422 return;
2423 }
2424
2425 // If this is an add tree that is used by a sub instruction, ignore it
2426 // until we process the subtract.
2427 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2428 cast<Instruction>(Val: BO->user_back())->getOpcode() == Instruction::Sub)
2429 return;
2430 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2431 cast<Instruction>(Val: BO->user_back())->getOpcode() == Instruction::FSub)
2432 return;
2433
2434 ReassociateExpression(I: BO);
2435}
2436
2437void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2438 // First, walk the expression tree, linearizing the tree, collecting the
2439 // operand information.
2440 SmallVector<RepeatedValue, 8> Tree;
2441 OverflowTracking Flags;
2442 MadeChange |= LinearizeExprTree(I, Ops&: Tree, ToRedo&: RedoInsts, Flags);
2443 SmallVector<ValueEntry, 8> Ops;
2444 Ops.reserve(N: Tree.size());
2445 for (const RepeatedValue &E : Tree)
2446 Ops.append(NumInputs: E.second, Elt: ValueEntry(getRank(V: E.first), E.first));
2447
2448 LLVM_DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2449
2450 // Boost the rank of divergent operands so they sort towards the root of the
2451 // expression tree, clustering uniform operands together at the leaves. On
2452 // targets without divergence UniformityInfo is empty and this is a no-op.
2453 //
2454 // Example: (uniform1 + divergent) + uniform2
2455 // -> (uniform1 + uniform2) + divergent
2456 if (UA && Ops.size() > 2) {
2457 constexpr unsigned DivergentRankOffset = 1U << 28;
2458 BasicBlock *ParentBB = I->getParent();
2459 for (ValueEntry &Entry : Ops) {
2460 if (isa<Constant>(Val: Entry.Op))
2461 continue;
2462 bool Divergent = false;
2463 for (const Use &U : Entry.Op->uses()) {
2464 Instruction *Usr = dyn_cast<Instruction>(Val: U.getUser());
2465 if (Usr && Usr->getParent() == ParentBB) {
2466 Divergent = UA->isDivergentAtUse(U);
2467 break;
2468 }
2469 }
2470 if (Divergent)
2471 Entry.Rank += DivergentRankOffset;
2472 }
2473 }
2474
2475 // Now that we have linearized the tree to a list and have gathered all of
2476 // the operands and their ranks, sort the operands by their rank. Use a
2477 // stable_sort so that values with equal ranks will have their relative
2478 // positions maintained (and so the compiler is deterministic). Note that
2479 // this sorts so that the highest ranking values end up at the beginning of
2480 // the vector.
2481 llvm::stable_sort(Range&: Ops);
2482
2483 // Now that we have the expression tree in a convenient
2484 // sorted form, optimize it globally if possible.
2485 if (Value *V = OptimizeExpression(I, Ops)) {
2486 if (V == I)
2487 // Self-referential expression in unreachable code.
2488 return;
2489 // This expression tree simplified to something that isn't a tree,
2490 // eliminate it.
2491 LLVM_DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2492 I->replaceAllUsesWith(V);
2493 if (Instruction *VI = dyn_cast<Instruction>(Val: V))
2494 if (I->getDebugLoc())
2495 VI->setDebugLoc(I->getDebugLoc());
2496 RedoInsts.insert(X: I);
2497 ++NumAnnihil;
2498 return;
2499 }
2500
2501 // We want to sink immediates as deeply as possible except in the case where
2502 // this is a multiply tree used only by an add, and the immediate is a -1.
2503 // In this case we reassociate to put the negation on the outside so that we
2504 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2505 if (I->hasOneUse()) {
2506 if (I->getOpcode() == Instruction::Mul &&
2507 cast<Instruction>(Val: I->user_back())->getOpcode() == Instruction::Add &&
2508 isa<ConstantInt>(Val: Ops.back().Op) &&
2509 cast<ConstantInt>(Val: Ops.back().Op)->isMinusOne()) {
2510 ValueEntry Tmp = Ops.pop_back_val();
2511 Ops.insert(I: Ops.begin(), Elt: Tmp);
2512 } else if (I->getOpcode() == Instruction::FMul &&
2513 cast<Instruction>(Val: I->user_back())->getOpcode() ==
2514 Instruction::FAdd &&
2515 isa<ConstantFP>(Val: Ops.back().Op) &&
2516 cast<ConstantFP>(Val: Ops.back().Op)->isMinusOne()) {
2517 ValueEntry Tmp = Ops.pop_back_val();
2518 Ops.insert(I: Ops.begin(), Elt: Tmp);
2519 }
2520 }
2521
2522 LLVM_DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2523
2524 if (Ops.size() == 1) {
2525 if (Ops[0].Op == I)
2526 // Self-referential expression in unreachable code.
2527 return;
2528
2529 // This expression tree simplified to something that isn't a tree,
2530 // eliminate it.
2531 I->replaceAllUsesWith(V: Ops[0].Op);
2532 if (Instruction *OI = dyn_cast<Instruction>(Val: Ops[0].Op))
2533 OI->setDebugLoc(I->getDebugLoc());
2534 RedoInsts.insert(X: I);
2535 return;
2536 }
2537
2538 if (Ops.size() > 2 && Ops.size() <= GlobalReassociateLimit) {
2539 // Find the pair with the highest count in the pairmap and move it to the
2540 // back of the list so that it can later be CSE'd.
2541 // example:
2542 // a*b*c*d*e
2543 // if c*e is the most "popular" pair, we can express this as
2544 // (((c*e)*d)*b)*a
2545 unsigned Max = 1;
2546 unsigned BestRank = 0;
2547 std::pair<unsigned, unsigned> BestPair;
2548 unsigned Idx = I->getOpcode() - Instruction::BinaryOpsBegin;
2549 unsigned LimitIdx = 0;
2550 // With the CSE-driven heuristic, we are about to slap two values at the
2551 // beginning of the expression whereas they could live very late in the CFG.
2552 // When using the CSE-local heuristic we avoid creating dependences from
2553 // completely unrelated part of the CFG by limiting the expression
2554 // reordering on the values that live in the first seen basic block.
2555 // The main idea is that we want to avoid forming expressions that would
2556 // become loop dependent.
2557 if (UseCSELocal) {
2558 const BasicBlock *FirstSeenBB = nullptr;
2559 int StartIdx = Ops.size() - 1;
2560 // Skip the first value of the expression since we need at least two
2561 // values to materialize an expression. I.e., even if this value is
2562 // anchored in a different basic block, the actual first sub expression
2563 // will be anchored on the second value.
2564 for (int i = StartIdx - 1; i != -1; --i) {
2565 const Value *Val = Ops[i].Op;
2566 const auto *CurrLeafInstr = dyn_cast<Instruction>(Val);
2567 const BasicBlock *SeenBB = nullptr;
2568 if (!CurrLeafInstr) {
2569 // The value is free of any CFG dependencies.
2570 // Do as if it lives in the entry block.
2571 //
2572 // We do this to make sure all the values falling on this path are
2573 // seen through the same anchor point. The rationale is these values
2574 // can be combined together to from a sub expression free of any CFG
2575 // dependencies so we want them to stay together.
2576 // We could be cleverer and postpone the anchor down to the first
2577 // anchored value, but that's likely complicated to get right.
2578 // E.g., we wouldn't want to do that if that means being stuck in a
2579 // loop.
2580 //
2581 // For instance, we wouldn't want to change:
2582 // res = arg1 op arg2 op arg3 op ... op loop_val1 op loop_val2 ...
2583 // into
2584 // res = loop_val1 op arg1 op arg2 op arg3 op ... op loop_val2 ...
2585 // Because all the sub expressions with arg2..N would be stuck between
2586 // two loop dependent values.
2587 SeenBB = &I->getParent()->getParent()->getEntryBlock();
2588 } else {
2589 SeenBB = CurrLeafInstr->getParent();
2590 }
2591
2592 if (!FirstSeenBB) {
2593 FirstSeenBB = SeenBB;
2594 continue;
2595 }
2596 if (FirstSeenBB != SeenBB) {
2597 // ith value is in a different basic block.
2598 // Rewind the index once to point to the last value on the same basic
2599 // block.
2600 LimitIdx = i + 1;
2601 LLVM_DEBUG(dbgs() << "CSE reordering: Consider values between ["
2602 << LimitIdx << ", " << StartIdx << "]\n");
2603 break;
2604 }
2605 }
2606 }
2607 for (unsigned i = Ops.size() - 1; i > LimitIdx; --i) {
2608 // We must use int type to go below zero when LimitIdx is 0.
2609 for (int j = i - 1; j >= (int)LimitIdx; --j) {
2610 unsigned Score = 0;
2611 Value *Op0 = Ops[i].Op;
2612 Value *Op1 = Ops[j].Op;
2613 if (std::less<Value *>()(Op1, Op0))
2614 std::swap(a&: Op0, b&: Op1);
2615 auto it = PairMap[Idx].find(Val: {Op0, Op1});
2616 if (it != PairMap[Idx].end()) {
2617 // Functions like BreakUpSubtract() can erase the Values we're using
2618 // as keys and create new Values after we built the PairMap. There's a
2619 // small chance that the new nodes can have the same address as
2620 // something already in the table. We shouldn't accumulate the stored
2621 // score in that case as it refers to the wrong Value.
2622 if (it->second.isValid())
2623 Score += it->second.Score;
2624 }
2625
2626 unsigned MaxRank = std::max(a: Ops[i].Rank, b: Ops[j].Rank);
2627
2628 // By construction, the operands are sorted in reverse order of their
2629 // topological order.
2630 // So we tend to form (sub) expressions with values that are close to
2631 // each other.
2632 //
2633 // Now to expose more CSE opportunities we want to expose the pair of
2634 // operands that occur the most (as statically computed in
2635 // BuildPairMap.) as the first sub-expression.
2636 //
2637 // If two pairs occur as many times, we pick the one with the
2638 // lowest rank, meaning the one with both operands appearing first in
2639 // the topological order.
2640 if (Score > Max || (Score == Max && MaxRank < BestRank)) {
2641 BestPair = {j, i};
2642 Max = Score;
2643 BestRank = MaxRank;
2644 }
2645 }
2646 }
2647 if (Max > 1) {
2648 auto Op0 = Ops[BestPair.first];
2649 auto Op1 = Ops[BestPair.second];
2650 Ops.erase(CI: &Ops[BestPair.second]);
2651 Ops.erase(CI: &Ops[BestPair.first]);
2652 Ops.push_back(Elt: Op0);
2653 Ops.push_back(Elt: Op1);
2654 }
2655 }
2656 LLVM_DEBUG(dbgs() << "RAOut after CSE reorder:\t"; PrintOps(I, Ops);
2657 dbgs() << '\n');
2658 // Now that we ordered and optimized the expressions, splat them back into
2659 // the expression tree, removing any unneeded nodes.
2660 RewriteExprTree(I, Ops, Flags);
2661}
2662
2663void
2664ReassociatePass::BuildPairMap(ReversePostOrderTraversal<Function *> &RPOT) {
2665 // Make a "pairmap" of how often each operand pair occurs.
2666 for (BasicBlock *BI : RPOT) {
2667 for (Instruction &I : *BI) {
2668 if (!I.isAssociative() || !I.isBinaryOp())
2669 continue;
2670
2671 // Ignore nodes that aren't at the root of trees.
2672 if (I.hasOneUse() && I.user_back()->getOpcode() == I.getOpcode())
2673 continue;
2674
2675 // Collect all operands in a single reassociable expression.
2676 // Since Reassociate has already been run once, we can assume things
2677 // are already canonical according to Reassociation's regime.
2678 SmallVector<Value *, 8> Worklist = { I.getOperand(i: 0), I.getOperand(i: 1) };
2679 SmallVector<Value *, 8> Ops;
2680 while (!Worklist.empty() && Ops.size() <= GlobalReassociateLimit) {
2681 Value *Op = Worklist.pop_back_val();
2682 Instruction *OpI = dyn_cast<Instruction>(Val: Op);
2683 if (!OpI || OpI->getOpcode() != I.getOpcode() || !OpI->hasOneUse()) {
2684 Ops.push_back(Elt: Op);
2685 continue;
2686 }
2687 // Be paranoid about self-referencing expressions in unreachable code.
2688 if (OpI->getOperand(i: 0) != OpI)
2689 Worklist.push_back(Elt: OpI->getOperand(i: 0));
2690 if (OpI->getOperand(i: 1) != OpI)
2691 Worklist.push_back(Elt: OpI->getOperand(i: 1));
2692 }
2693 // Skip extremely long expressions.
2694 if (Ops.size() > GlobalReassociateLimit)
2695 continue;
2696
2697 // Add all pairwise combinations of operands to the pair map.
2698 unsigned BinaryIdx = I.getOpcode() - Instruction::BinaryOpsBegin;
2699 SmallSet<std::pair<Value *, Value*>, 32> Visited;
2700 for (unsigned i = 0; i < Ops.size() - 1; ++i) {
2701 for (unsigned j = i + 1; j < Ops.size(); ++j) {
2702 // Canonicalize operand orderings.
2703 Value *Op0 = Ops[i];
2704 Value *Op1 = Ops[j];
2705 if (std::less<Value *>()(Op1, Op0))
2706 std::swap(a&: Op0, b&: Op1);
2707 if (!Visited.insert(V: {Op0, Op1}).second)
2708 continue;
2709 auto res = PairMap[BinaryIdx].insert(KV: {{Op0, Op1}, {.Value1: Op0, .Value2: Op1, .Score: 1}});
2710 if (!res.second) {
2711 // If either key value has been erased then we've got the same
2712 // address by coincidence. That can't happen here because nothing is
2713 // erasing values but it can happen by the time we're querying the
2714 // map.
2715 assert(res.first->second.isValid() && "WeakVH invalidated");
2716 ++res.first->second.Score;
2717 }
2718 }
2719 }
2720 }
2721 }
2722}
2723
2724PreservedAnalyses ReassociatePass::run(Function &F,
2725 FunctionAnalysisManager &AM) {
2726 // UniformityInfo is empty (and cheap) on targets without branch divergence,
2727 // so request it unconditionally.
2728 UniformityInfo &UI = AM.getResult<UniformityInfoAnalysis>(IR&: F);
2729 return runImpl(F, UI);
2730}
2731
2732PreservedAnalyses ReassociatePass::runImpl(Function &F, UniformityInfo &UI) {
2733 UA = &UI;
2734 UseCSELocal = F.getContext()
2735 .getLibraryOptions()
2736 .get<ScalarOptions>()
2737 .reassociate_use_cse_local;
2738
2739 // Get the functions basic blocks in Reverse Post Order. This order is used by
2740 // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic
2741 // blocks (it has been seen that the analysis in this pass could hang when
2742 // analysing dead basic blocks).
2743 ReversePostOrderTraversal<Function *> RPOT(&F);
2744
2745 // Calculate the rank map for F.
2746 BuildRankMap(F, RPOT);
2747
2748 // Build the pair map before running reassociate.
2749 // Technically this would be more accurate if we did it after one round
2750 // of reassociation, but in practice it doesn't seem to help much on
2751 // real-world code, so don't waste the compile time running reassociate
2752 // twice.
2753 // If a user wants, they could expicitly run reassociate twice in their
2754 // pass pipeline for further potential gains.
2755 // It might also be possible to update the pair map during runtime, but the
2756 // overhead of that may be large if there's many reassociable chains.
2757 BuildPairMap(RPOT);
2758
2759 MadeChange = false;
2760
2761 // Traverse the same blocks that were analysed by BuildRankMap.
2762 for (BasicBlock *BI : RPOT) {
2763 assert(RankMap.count(&*BI) && "BB should be ranked.");
2764 // Optimize every instruction in the basic block.
2765 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;)
2766 if (isInstructionTriviallyDead(I: &*II)) {
2767 EraseInst(I: &*II++);
2768 } else {
2769 OptimizeInst(I: &*II);
2770 assert(II->getParent() == &*BI && "Moved to a different block!");
2771 ++II;
2772 }
2773
2774 // Make a copy of all the instructions to be redone so we can remove dead
2775 // instructions.
2776 OrderedSet ToRedo(RedoInsts);
2777 // Iterate over all instructions to be reevaluated and remove trivially dead
2778 // instructions. If any operand of the trivially dead instruction becomes
2779 // dead mark it for deletion as well. Continue this process until all
2780 // trivially dead instructions have been removed.
2781 while (!ToRedo.empty()) {
2782 Instruction *I = ToRedo.pop_back_val();
2783 if (isInstructionTriviallyDead(I)) {
2784 RecursivelyEraseDeadInsts(I, Insts&: ToRedo);
2785 MadeChange = true;
2786 }
2787 }
2788
2789 // Now that we have removed dead instructions, we can reoptimize the
2790 // remaining instructions.
2791 while (!RedoInsts.empty()) {
2792 Instruction *I = RedoInsts.front();
2793 RedoInsts.erase(I: RedoInsts.begin());
2794 if (isInstructionTriviallyDead(I))
2795 EraseInst(I);
2796 else
2797 OptimizeInst(I);
2798 }
2799 }
2800
2801 // We are done with the rank map, pair map, and uniformity info.
2802 RankMap.clear();
2803 ValueRankMap.clear();
2804 for (auto &Entry : PairMap)
2805 Entry.clear();
2806 UA = nullptr;
2807
2808 if (MadeChange) {
2809 PreservedAnalyses PA;
2810 PA.preserveSet<CFGAnalyses>();
2811 return PA;
2812 }
2813
2814 return PreservedAnalyses::all();
2815}
2816
2817namespace {
2818
2819class ReassociateLegacyPass : public FunctionPass {
2820 ReassociatePass Impl;
2821
2822public:
2823 static char ID; // Pass identification, replacement for typeid
2824
2825 ReassociateLegacyPass() : FunctionPass(ID) {
2826 initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry());
2827 }
2828
2829 bool runOnFunction(Function &F) override {
2830 if (skipFunction(F))
2831 return false;
2832
2833 UniformityInfo &UI =
2834 getAnalysis<UniformityInfoWrapperPass>().getUniformityInfo();
2835
2836 PreservedAnalyses PA = Impl.runImpl(F, UI);
2837 return !PA.areAllPreserved();
2838 }
2839
2840 void getAnalysisUsage(AnalysisUsage &AU) const override {
2841 AU.setPreservesCFG();
2842 AU.addRequired<UniformityInfoWrapperPass>();
2843 AU.addPreserved<AAResultsWrapperPass>();
2844 AU.addPreserved<GlobalsAAWrapperPass>();
2845 }
2846};
2847
2848} // end anonymous namespace
2849
2850char ReassociateLegacyPass::ID = 0;
2851
2852INITIALIZE_PASS_BEGIN(ReassociateLegacyPass, "reassociate",
2853 "Reassociate expressions", false, false)
2854INITIALIZE_PASS_DEPENDENCY(UniformityInfoWrapperPass)
2855INITIALIZE_PASS_END(ReassociateLegacyPass, "reassociate",
2856 "Reassociate expressions", false, false)
2857
2858// Public interface to the Reassociate pass
2859FunctionPass *llvm::createReassociatePass() {
2860 return new ReassociateLegacyPass();
2861}
2862