1//===- Reassociate.cpp - Reassociate binary expressions -------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This pass reassociates commutative expressions in an order that is designed
10// to promote better constant propagation, GCSE, LICM, PRE, etc.
11//
12// For example: 4 + (x + 5) -> x + (4 + 5)
13//
14// In the implementation of this algorithm, constants are assigned rank = 0,
15// function arguments are rank = 1, and other values are assigned ranks
16// corresponding to the reverse post order traversal of current function
17// (starting at 2), which effectively gives values in deep loops higher rank
18// than values not in loops.
19//
20//===----------------------------------------------------------------------===//
21
22#include "llvm/Transforms/Scalar/Reassociate.h"
23#include "llvm/ADT/APFloat.h"
24#include "llvm/ADT/APInt.h"
25#include "llvm/ADT/DenseMap.h"
26#include "llvm/ADT/PostOrderIterator.h"
27#include "llvm/ADT/SmallPtrSet.h"
28#include "llvm/ADT/SmallSet.h"
29#include "llvm/ADT/SmallVector.h"
30#include "llvm/ADT/Statistic.h"
31#include "llvm/Analysis/BasicAliasAnalysis.h"
32#include "llvm/Analysis/ConstantFolding.h"
33#include "llvm/Analysis/GlobalsModRef.h"
34#include "llvm/Analysis/ValueTracking.h"
35#include "llvm/IR/Argument.h"
36#include "llvm/IR/BasicBlock.h"
37#include "llvm/IR/CFG.h"
38#include "llvm/IR/Constant.h"
39#include "llvm/IR/Constants.h"
40#include "llvm/IR/Function.h"
41#include "llvm/IR/IRBuilder.h"
42#include "llvm/IR/InstrTypes.h"
43#include "llvm/IR/Instruction.h"
44#include "llvm/IR/Instructions.h"
45#include "llvm/IR/Operator.h"
46#include "llvm/IR/PassManager.h"
47#include "llvm/IR/PatternMatch.h"
48#include "llvm/IR/Type.h"
49#include "llvm/IR/User.h"
50#include "llvm/IR/Value.h"
51#include "llvm/IR/ValueHandle.h"
52#include "llvm/InitializePasses.h"
53#include "llvm/Pass.h"
54#include "llvm/Support/Casting.h"
55#include "llvm/Support/CommandLine.h"
56#include "llvm/Support/Debug.h"
57#include "llvm/Support/raw_ostream.h"
58#include "llvm/Transforms/Scalar.h"
59#include "llvm/Transforms/Utils/Local.h"
60#include <algorithm>
61#include <cassert>
62#include <utility>
63
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
74static cl::opt<bool>
75 UseCSELocalOpt(DEBUG_TYPE "-use-cse-local",
76 cl::desc("Only reorder expressions within a basic block "
77 "when exposing CSE opportunities"),
78 cl::init(Val: true), cl::Hidden);
79
80#ifndef NDEBUG
81/// Print out the expression identified in the Ops list.
82static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) {
83 Module *M = I->getModule();
84 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " "
85 << *Ops[0].Op->getType() << '\t';
86 for (const ValueEntry &Op : Ops) {
87 dbgs() << "[ ";
88 Op.Op->printAsOperand(dbgs(), false, M);
89 dbgs() << ", #" << Op.Rank << "] ";
90 }
91}
92#endif
93
94/// Utility class representing a non-constant Xor-operand. We classify
95/// non-constant Xor-Operands into two categories:
96/// C1) The operand is in the form "X & C", where C is a constant and C != ~0
97/// C2)
98/// C2.1) The operand is in the form of "X | C", where C is a non-zero
99/// constant.
100/// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this
101/// operand as "E | 0"
102class llvm::reassociate::XorOpnd {
103public:
104 XorOpnd(Value *V);
105
106 bool isInvalid() const { return SymbolicPart == nullptr; }
107 bool isOrExpr() const { return isOr; }
108 Value *getValue() const { return OrigVal; }
109 Value *getSymbolicPart() const { return SymbolicPart; }
110 unsigned getSymbolicRank() const { return SymbolicRank; }
111 const APInt &getConstPart() const { return ConstPart; }
112
113 void Invalidate() { SymbolicPart = OrigVal = nullptr; }
114 void setSymbolicRank(unsigned R) { SymbolicRank = R; }
115
116private:
117 Value *OrigVal;
118 Value *SymbolicPart;
119 APInt ConstPart;
120 unsigned SymbolicRank;
121 bool isOr;
122};
123
124XorOpnd::XorOpnd(Value *V) {
125 assert(!isa<ConstantInt>(V) && "No ConstantInt");
126 OrigVal = V;
127 Instruction *I = dyn_cast<Instruction>(Val: V);
128 SymbolicRank = 0;
129
130 if (I && (I->getOpcode() == Instruction::Or ||
131 I->getOpcode() == Instruction::And)) {
132 Value *V0 = I->getOperand(i: 0);
133 Value *V1 = I->getOperand(i: 1);
134 const APInt *C;
135 if (match(V: V0, P: m_APInt(Res&: C)))
136 std::swap(a&: V0, b&: V1);
137
138 if (match(V: V1, P: m_APInt(Res&: C))) {
139 ConstPart = *C;
140 SymbolicPart = V0;
141 isOr = (I->getOpcode() == Instruction::Or);
142 return;
143 }
144 }
145
146 // view the operand as "V | 0"
147 SymbolicPart = V;
148 ConstPart = APInt::getZero(numBits: V->getType()->getScalarSizeInBits());
149 isOr = true;
150}
151
152/// Return true if I is an instruction with the FastMathFlags that are needed
153/// for general reassociation set. This is not the same as testing
154/// Instruction::isAssociative() because it includes operations like fsub.
155/// (This routine is only intended to be called for floating-point operations.)
156static bool hasFPAssociativeFlags(Instruction *I) {
157 assert(I && isa<FPMathOperator>(I) && "Should only check FP ops");
158 return I->hasAllowReassoc() && I->hasNoSignedZeros();
159}
160
161/// Return true if V is an instruction of the specified opcode and if it
162/// only has one use.
163static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
164 auto *BO = dyn_cast<BinaryOperator>(Val: V);
165 if (BO && BO->hasOneUse() && BO->getOpcode() == Opcode)
166 if (!isa<FPMathOperator>(Val: BO) || hasFPAssociativeFlags(I: BO))
167 return BO;
168 return nullptr;
169}
170
171static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
172 unsigned Opcode2) {
173 auto *BO = dyn_cast<BinaryOperator>(Val: V);
174 if (BO && BO->hasOneUse() &&
175 (BO->getOpcode() == Opcode1 || BO->getOpcode() == Opcode2))
176 if (!isa<FPMathOperator>(Val: BO) || hasFPAssociativeFlags(I: BO))
177 return BO;
178 return nullptr;
179}
180
181/// Return the fmul operand if V is a one-use fadd with a single one-use fmul
182/// operand, both allowing contraction. Such pairs can be fused into a single
183/// fma, so they are kept together as leaves of the enclosing expression tree
184/// instead of being linearized into it.
185///
186/// Do not keep the pair together if the other operand is itself a reassociable
187/// fadd. Treating the outer fadd as a leaf would hide the nested addition from
188/// reassociation and prevent the complete expression from being optimized.
189static BinaryOperator *isFMulAddCandidate(Value *V) {
190 BinaryOperator *FAdd = isReassociableOp(V, Opcode: Instruction::FAdd);
191 if (!FAdd || !FAdd->hasAllowContract())
192 return nullptr;
193 auto ContractableFMul = [](BinaryOperator *&FMul) {
194 return m_CombineAnd(Ps: m_AllowContract(SubPattern: m_OneUse(SubPattern: m_FMul(L: m_Value(), R: m_Value()))),
195 Ps: m_BinOp(I&: FMul));
196 };
197 BinaryOperator *Mul = nullptr, *OtherMul = nullptr;
198 Value *OtherOp = nullptr;
199 // Keep constants and nested additions visible to the enclosing expression so
200 // they can participate in folding and reassociation.
201 if (!match(V: FAdd, P: m_c_FAdd(L: ContractableFMul(Mul), R: m_Value(V&: OtherOp))) ||
202 isa<Constant>(Val: OtherOp) || isReassociableOp(V: OtherOp, Opcode: Instruction::FAdd) ||
203 match(V: OtherOp, P: ContractableFMul(OtherMul)))
204 return nullptr;
205 return Mul;
206}
207
208void ReassociatePass::BuildRankMap(Function &F,
209 ReversePostOrderTraversal<Function*> &RPOT) {
210 unsigned Rank = 2;
211
212 // Assign distinct ranks to function arguments.
213 for (auto &Arg : F.args()) {
214 ValueRankMap[&Arg] = ++Rank;
215 LLVM_DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank
216 << "\n");
217 }
218
219 // Traverse basic blocks in ReversePostOrder.
220 for (BasicBlock *BB : RPOT) {
221 unsigned BBRank = RankMap[BB] = ++Rank << 16;
222
223 // Walk the basic block, adding precomputed ranks for any instructions that
224 // we cannot move. This ensures that the ranks for these instructions are
225 // all different in the block.
226 for (Instruction &I : *BB)
227 if (mayHaveNonDefUseDependency(I))
228 ValueRankMap[&I] = ++BBRank;
229 }
230}
231
232unsigned ReassociatePass::getRank(Value *V) {
233 // Return 1+MAX(rank(LHS), rank(RHS)) for expressions so we can reassociate
234 // expressions for code motion. Use an explicit worklist rather than native
235 // recursion so long acyclic use-def chains do not overflow the stack.
236 struct RankWorkItem {
237 Value *V;
238 unsigned OpNo;
239 unsigned Rank;
240 };
241
242 // Each item is one suspended recursive getRank() call.
243 // Completed ranks are folded back into the parent.
244 SmallVector<RankWorkItem, 16> Worklist;
245 Worklist.push_back(Elt: RankWorkItem{.V: V, .OpNo: 0, .Rank: 0});
246
247 while (true) {
248 RankWorkItem &Item = Worklist.back();
249 Instruction *I = dyn_cast<Instruction>(Val: Item.V);
250 unsigned Rank = 0;
251 if (!I) {
252 // Function argument, global or constant
253 Rank = isa<Argument>(Val: Item.V) ? ValueRankMap[Item.V] : 0;
254 } else if (ValueRankMap[I]) {
255 // Instruction that is not movable.
256 Rank = ValueRankMap[I];
257 } else if (Item.OpNo == I->getNumOperands() ||
258 Item.Rank == RankMap[I->getParent()]) {
259 // All operands were visited or the max block rank was reached.
260 Rank = Item.Rank;
261 // If this is a 'not' or 'neg' instruction, do not count it for rank.
262 // This assures us that X and ~X will have the same rank.
263 if (!match(V: I, P: m_Not(V: m_Value())) && !match(V: I, P: m_Neg(V: m_Value())) &&
264 !match(V: I, P: m_FNeg(X: m_Value())))
265 ++Rank;
266
267 LLVM_DEBUG(dbgs() << "Calculated Rank[" << I->getName() << "] = " << Rank
268 << "\n");
269
270 ValueRankMap[I] = Rank;
271 } else {
272 Worklist.push_back(Elt: RankWorkItem{.V: I->getOperand(i: Item.OpNo), .OpNo: 0, .Rank: 0});
273 continue;
274 }
275
276 // Once the current use-def node has a known rank, carry that rank back to
277 // the parent expression and advance past the operand that led here.
278 Worklist.pop_back();
279 if (Worklist.empty())
280 return Rank;
281
282 RankWorkItem &Parent = Worklist.back();
283 Parent.Rank = std::max(a: Parent.Rank, b: Rank);
284 ++Parent.OpNo;
285 }
286}
287
288// Canonicalize constants to RHS. Otherwise, sort the operands by rank.
289void ReassociatePass::canonicalizeOperands(Instruction *I) {
290 assert(isa<BinaryOperator>(I) && "Expected binary operator.");
291 assert(I->isCommutative() && "Expected commutative operator.");
292
293 Value *LHS = I->getOperand(i: 0);
294 Value *RHS = I->getOperand(i: 1);
295 if (LHS == RHS || isa<Constant>(Val: RHS))
296 return;
297 if (isa<Constant>(Val: LHS) || getRank(V: RHS) < getRank(V: LHS)) {
298 cast<BinaryOperator>(Val: I)->swapOperands();
299 MadeChange = true;
300 }
301}
302
303static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
304 BasicBlock::iterator InsertBefore,
305 Value *FlagsOp) {
306 if (S1->getType()->isIntOrIntVectorTy())
307 return BinaryOperator::CreateAdd(V1: S1, V2: S2, Name, InsertBefore);
308 else {
309 BinaryOperator *Res =
310 BinaryOperator::CreateFAdd(V1: S1, V2: S2, Name, InsertBefore);
311 Res->setFastMathFlags(cast<FPMathOperator>(Val: FlagsOp)->getFastMathFlags());
312 return Res;
313 }
314}
315
316static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
317 BasicBlock::iterator InsertBefore,
318 Value *FlagsOp) {
319 if (S1->getType()->isIntOrIntVectorTy())
320 return BinaryOperator::CreateMul(V1: S1, V2: S2, Name, InsertBefore);
321 else {
322 BinaryOperator *Res =
323 BinaryOperator::CreateFMul(V1: S1, V2: S2, Name, InsertBefore);
324 Res->setFastMathFlags(cast<FPMathOperator>(Val: FlagsOp)->getFastMathFlags());
325 return Res;
326 }
327}
328
329static Instruction *CreateNeg(Value *S1, const Twine &Name,
330 BasicBlock::iterator InsertBefore,
331 Value *FlagsOp) {
332 if (S1->getType()->isIntOrIntVectorTy())
333 return BinaryOperator::CreateNeg(Op: S1, Name, InsertBefore);
334
335 if (auto *FMFSource = dyn_cast<Instruction>(Val: FlagsOp))
336 return UnaryOperator::CreateFNegFMF(Op: S1, FMFSource, Name, InsertBefore);
337
338 return UnaryOperator::CreateFNeg(V: S1, Name, InsertBefore);
339}
340
341/// Replace 0-X with X*-1.
342static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
343 assert((isa<UnaryOperator>(Neg) || isa<BinaryOperator>(Neg)) &&
344 "Expected a Negate!");
345 // FIXME: It's not safe to lower a unary FNeg into a FMul by -1.0.
346 unsigned OpNo = isa<BinaryOperator>(Val: Neg) ? 1 : 0;
347 Type *Ty = Neg->getType();
348 Constant *NegOne = Ty->isIntOrIntVectorTy() ?
349 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, V: -1.0);
350
351 BinaryOperator *Res =
352 CreateMul(S1: Neg->getOperand(i: OpNo), S2: NegOne, Name: "", InsertBefore: Neg->getIterator(), FlagsOp: Neg);
353 Neg->setOperand(i: OpNo, Val: Constant::getNullValue(Ty)); // Drop use of op.
354 Res->takeName(V: Neg);
355 Neg->replaceAllUsesWith(V: Res);
356 Res->setDebugLoc(Neg->getDebugLoc());
357 return Res;
358}
359
360using RepeatedValue = std::pair<Value *, uint64_t>;
361
362/// Given an associative binary expression, return the leaf
363/// nodes in Ops along with their weights (how many times the leaf occurs). The
364/// original expression is the same as
365/// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times
366/// op
367/// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times
368/// op
369/// ...
370/// op
371/// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times
372///
373/// Note that the values Ops[0].first, ..., Ops[N].first are all distinct.
374///
375/// This routine may modify the function, in which case it returns 'true'. The
376/// changes it makes may well be destructive, changing the value computed by 'I'
377/// to something completely different. Thus if the routine returns 'true' then
378/// you MUST either replace I with a new expression computed from the Ops array,
379/// or use RewriteExprTree to put the values back in.
380///
381/// A leaf node is either not a binary operation of the same kind as the root
382/// node 'I' (i.e. is not a binary operator at all, or is, but with a different
383/// opcode), or is the same kind of binary operator but has a use which either
384/// does not belong to the expression, or does belong to the expression but is
385/// a leaf node. Every leaf node has at least one use that is a non-leaf node
386/// of the expression, while for non-leaf nodes (except for the root 'I') every
387/// use is a non-leaf node of the expression.
388///
389/// For example:
390/// expression graph node names
391///
392/// + | I
393/// / \ |
394/// + + | A, B
395/// / \ / \ |
396/// * + * | C, D, E
397/// / \ / \ / \ |
398/// + * | F, G
399///
400/// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in
401/// that order) (C, 1), (E, 1), (F, 2), (G, 2).
402///
403/// The expression is maximal: if some instruction is a binary operator of the
404/// same kind as 'I', and all of its uses are non-leaf nodes of the expression,
405/// then the instruction also belongs to the expression, is not a leaf node of
406/// it, and its operands also belong to the expression (but may be leaf nodes).
407///
408/// NOTE: This routine will set operands of non-leaf non-root nodes to undef in
409/// order to ensure that every non-root node in the expression has *exactly one*
410/// use by a non-leaf node of the expression. This destruction means that the
411/// caller MUST either replace 'I' with a new expression or use something like
412/// RewriteExprTree to put the values back in if the routine indicates that it
413/// made a change by returning 'true'.
414///
415/// In the above example either the right operand of A or the left operand of B
416/// will be replaced by undef. If it is B's operand then this gives:
417///
418/// + | I
419/// / \ |
420/// + + | A, B - operand of B replaced with undef
421/// / \ \ |
422/// * + * | C, D, E
423/// / \ / \ / \ |
424/// + * | F, G
425///
426/// Note that such undef operands can only be reached by passing through 'I'.
427/// For example, if you visit operands recursively starting from a leaf node
428/// then you will never see such an undef operand unless you get back to 'I',
429/// which requires passing through a phi node.
430///
431/// Note that this routine may also mutate binary operators of the wrong type
432/// that have all uses inside the expression (i.e. only used by non-leaf nodes
433/// of the expression) if it can turn them into binary operators of the right
434/// type and thus make the expression bigger.
435static bool LinearizeExprTree(Instruction *I,
436 SmallVectorImpl<RepeatedValue> &Ops,
437 ReassociatePass::OrderedSet &ToRedo,
438 OverflowTracking &Flags) {
439 assert((isa<UnaryOperator>(I) || isa<BinaryOperator>(I)) &&
440 "Expected a UnaryOperator or BinaryOperator!");
441 LLVM_DEBUG(dbgs() << "LINEARIZE: " << *I << '\n');
442 unsigned Opcode = I->getOpcode();
443 assert(I->isAssociative() && I->isCommutative() &&
444 "Expected an associative and commutative operation!");
445
446 // Visit all operands of the expression, keeping track of their weight (the
447 // number of paths from the expression root to the operand, or if you like
448 // the number of times that operand occurs in the linearized expression).
449 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1
450 // while A has weight two.
451
452 // Worklist of non-leaf nodes (their operands are in the expression too) along
453 // with their weights, representing a certain number of paths to the operator.
454 // If an operator occurs in the worklist multiple times then we found multiple
455 // ways to get to it.
456 SmallVector<std::pair<Instruction *, uint64_t>, 8> Worklist; // (Op, Weight)
457 Worklist.push_back(Elt: std::make_pair(x&: I, y: 1));
458 bool Changed = false;
459
460 // Leaves of the expression are values that either aren't the right kind of
461 // operation (eg: a constant, or a multiply in an add tree), or are, but have
462 // some uses that are not inside the expression. For example, in I = X + X,
463 // X = A + B, the value X has two uses (by I) that are in the expression. If
464 // X has any other uses, for example in a return instruction, then we consider
465 // X to be a leaf, and won't analyze it further. When we first visit a value,
466 // if it has more than one use then at first we conservatively consider it to
467 // be a leaf. Later, as the expression is explored, we may discover some more
468 // uses of the value from inside the expression. If all uses turn out to be
469 // from within the expression (and the value is a binary operator of the right
470 // kind) then the value is no longer considered to be a leaf, and its operands
471 // are explored.
472
473 // Leaves - Keeps track of the set of putative leaves as well as the number of
474 // paths to each leaf seen so far.
475 using LeafMap = DenseMap<Value *, uint64_t>;
476 LeafMap Leaves; // Leaf -> Total weight so far.
477 SmallVector<Value *, 8> LeafOrder; // Ensure deterministic leaf output order.
478 const DataLayout &DL = I->getDataLayout();
479
480#ifndef NDEBUG
481 SmallPtrSet<Value *, 8> Visited; // For checking the iteration scheme.
482#endif
483 while (!Worklist.empty()) {
484 // We examine the operands of this binary operator.
485 auto [I, Weight] = Worklist.pop_back_val();
486
487 Flags.mergeFlags(I&: *I);
488
489 for (unsigned OpIdx = 0; OpIdx < I->getNumOperands(); ++OpIdx) { // Visit operands.
490 Value *Op = I->getOperand(i: OpIdx);
491 LLVM_DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n");
492 assert((!Op->hasUseList() || !Op->use_empty()) &&
493 "No uses, so how did we get to it?!");
494
495 // If this is a binary operation of the right kind with only one use then
496 // add its operands to the expression.
497 if (BinaryOperator *BO = isReassociableOp(V: Op, Opcode);
498 BO && (Opcode != Instruction::FAdd || !isFMulAddCandidate(V: BO))) {
499 assert(Visited.insert(Op).second && "Not first visit!");
500 LLVM_DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n");
501 Worklist.push_back(Elt: std::make_pair(x&: BO, y&: Weight));
502 continue;
503 }
504
505 // Appears to be a leaf. Is the operand already in the set of leaves?
506 LeafMap::iterator It = Leaves.find(Val: Op);
507 if (It == Leaves.end()) {
508 // Not in the leaf map. Must be the first time we saw this operand.
509 assert(Visited.insert(Op).second && "Not first visit!");
510 if (!Op->hasOneUse()) {
511 // This value has uses not accounted for by the expression, so it is
512 // not safe to modify. Mark it as being a leaf.
513 LLVM_DEBUG(dbgs()
514 << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n");
515 LeafOrder.push_back(Elt: Op);
516 Leaves[Op] = Weight;
517 continue;
518 }
519 // No uses outside the expression, try morphing it.
520 } else {
521 // Already in the leaf map.
522 assert(It != Leaves.end() && Visited.count(Op) &&
523 "In leaf map but not visited!");
524
525 // Update the number of paths to the leaf.
526 It->second += Weight;
527 assert(It->second >= Weight && "Weight overflows");
528
529 // If we still have uses that are not accounted for by the expression
530 // then it is not safe to modify the value.
531 if (!Op->hasOneUse())
532 continue;
533
534 // No uses outside the expression, try morphing it.
535 Weight = It->second;
536 Leaves.erase(I: It); // Since the value may be morphed below.
537 }
538
539 // At this point we have a value which, first of all, is not a binary
540 // expression of the right kind, and secondly, is only used inside the
541 // expression. This means that it can safely be modified. See if we
542 // can usefully morph it into an expression of the right kind.
543 assert((!isa<Instruction>(Op) ||
544 cast<Instruction>(Op)->getOpcode() != Opcode ||
545 (isa<FPMathOperator>(Op) &&
546 !hasFPAssociativeFlags(cast<Instruction>(Op))) ||
547 isFMulAddCandidate(Op)) &&
548 "Should have been handled above!");
549 assert(Op->hasOneUse() && "Has uses outside the expression tree!");
550
551 // If this is a multiply expression, turn any internal negations into
552 // multiplies by -1 so they can be reassociated. Add any users of the
553 // newly created multiplication by -1 to the redo list, so any
554 // reassociation opportunities that are exposed will be reassociated
555 // further.
556 Instruction *Neg;
557 if (((Opcode == Instruction::Mul && match(V: Op, P: m_Neg(V: m_Value()))) ||
558 (Opcode == Instruction::FMul && match(V: Op, P: m_FNeg(X: m_Value())))) &&
559 match(V: Op, P: m_Instruction(I&: Neg))) {
560 LLVM_DEBUG(dbgs()
561 << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ");
562 Instruction *Mul = LowerNegateToMultiply(Neg);
563 LLVM_DEBUG(dbgs() << *Mul << '\n');
564 Worklist.push_back(Elt: std::make_pair(x&: Mul, y&: Weight));
565 for (User *U : Mul->users()) {
566 if (BinaryOperator *UserBO = dyn_cast<BinaryOperator>(Val: U))
567 ToRedo.insert(X: UserBO);
568 }
569 ToRedo.insert(X: Neg);
570 Changed = true;
571 continue;
572 }
573
574 // Failed to morph into an expression of the right type. This really is
575 // a leaf.
576 LLVM_DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n");
577 assert((!isReassociableOp(Op, Opcode) || isFMulAddCandidate(Op)) &&
578 "Value was morphed?");
579 LeafOrder.push_back(Elt: Op);
580 Leaves[Op] = Weight;
581 }
582 }
583
584 // The leaves, repeated according to their weights, represent the linearized
585 // form of the expression.
586 for (Value *V : LeafOrder) {
587 LeafMap::iterator It = Leaves.find(Val: V);
588 if (It == Leaves.end())
589 // Node initially thought to be a leaf wasn't.
590 continue;
591 assert((!isReassociableOp(V, Opcode) || isFMulAddCandidate(V)) &&
592 "Shouldn't be a leaf!");
593 uint64_t Weight = It->second;
594 // Ensure the leaf is only output once.
595 It->second = 0;
596 Ops.push_back(Elt: std::make_pair(x&: V, y&: Weight));
597 if (Opcode == Instruction::Add && Flags.AllKnownNonNegative && Flags.HasNSW)
598 Flags.AllKnownNonNegative &= isKnownNonNegative(V, SQ: SimplifyQuery(DL));
599 else if (Opcode == Instruction::Mul) {
600 // To preserve NUW we need all inputs non-zero.
601 // To preserve NSW we need all inputs strictly positive.
602 if (Flags.AllKnownNonZero &&
603 (Flags.HasNUW || (Flags.HasNSW && Flags.AllKnownNonNegative))) {
604 Flags.AllKnownNonZero &= isKnownNonZero(V, Q: SimplifyQuery(DL));
605 if (Flags.HasNSW && Flags.AllKnownNonNegative)
606 Flags.AllKnownNonNegative &= isKnownNonNegative(V, SQ: SimplifyQuery(DL));
607 }
608 }
609 }
610
611 // For nilpotent operations or addition there may be no operands, for example
612 // because the expression was "X xor X" or consisted of 2^Bitwidth additions:
613 // in both cases the weight reduces to 0 causing the value to be skipped.
614 if (Ops.empty()) {
615 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, Ty: I->getType());
616 assert(Identity && "Associative operation without identity!");
617 Ops.emplace_back(Args&: Identity, Args: 1);
618 }
619
620 return Changed;
621}
622
623/// Now that the operands for this expression tree are
624/// linearized and optimized, emit them in-order.
625void ReassociatePass::RewriteExprTree(BinaryOperator *I,
626 SmallVectorImpl<ValueEntry> &Ops,
627 OverflowTracking Flags) {
628 assert(Ops.size() > 1 && "Single values should be used directly!");
629
630 // Since our optimizations should never increase the number of operations, the
631 // new expression can usually be written reusing the existing binary operators
632 // from the original expression tree, without creating any new instructions,
633 // though the rewritten expression may have a completely different topology.
634 // We take care to not change anything if the new expression will be the same
635 // as the original. If more than trivial changes (like commuting operands)
636 // were made then we are obliged to clear out any optional subclass data like
637 // nsw flags.
638
639 /// NodesToRewrite - Nodes from the original expression available for writing
640 /// the new expression into.
641 SmallVector<BinaryOperator*, 8> NodesToRewrite;
642 unsigned Opcode = I->getOpcode();
643 BinaryOperator *Op = I;
644
645 /// NotRewritable - The operands being written will be the leaves of the new
646 /// expression and must not be used as inner nodes (via NodesToRewrite) by
647 /// mistake. Inner nodes are always reassociable, and usually leaves are not
648 /// (if they were they would have been incorporated into the expression and so
649 /// would not be leaves), so most of the time there is no danger of this. But
650 /// in rare cases a leaf may become reassociable if an optimization kills uses
651 /// of it, or it may momentarily become reassociable during rewriting (below)
652 /// due it being removed as an operand of one of its uses. Ensure that misuse
653 /// of leaf nodes as inner nodes cannot occur by remembering all of the future
654 /// leaves and refusing to reuse any of them as inner nodes.
655 SmallPtrSet<Value*, 8> NotRewritable;
656 for (const ValueEntry &Op : Ops)
657 NotRewritable.insert(Ptr: Op.Op);
658
659 // ExpressionChangedStart - Non-null if the rewritten expression differs from
660 // the original in some non-trivial way, requiring the clearing of optional
661 // flags. Flags are cleared from the operator in ExpressionChangedStart up to
662 // ExpressionChangedEnd inclusive.
663 BinaryOperator *ExpressionChangedStart = nullptr,
664 *ExpressionChangedEnd = nullptr;
665 for (unsigned i = 0; ; ++i) {
666 // The last operation (which comes earliest in the IR) is special as both
667 // operands will come from Ops, rather than just one with the other being
668 // a subexpression.
669 if (i+2 == Ops.size()) {
670 Value *NewLHS = Ops[i].Op;
671 Value *NewRHS = Ops[i+1].Op;
672 Value *OldLHS = Op->getOperand(i_nocapture: 0);
673 Value *OldRHS = Op->getOperand(i_nocapture: 1);
674
675 if (NewLHS == OldLHS && NewRHS == OldRHS)
676 // Nothing changed, leave it alone.
677 break;
678
679 if (NewLHS == OldRHS && NewRHS == OldLHS) {
680 // The order of the operands was reversed. Swap them.
681 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
682 Op->swapOperands();
683 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
684 MadeChange = true;
685 ++NumChanged;
686 break;
687 }
688
689 // The new operation differs non-trivially from the original. Overwrite
690 // the old operands with the new ones.
691 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
692 if (NewLHS != OldLHS) {
693 BinaryOperator *BO = isReassociableOp(V: OldLHS, Opcode);
694 if (BO && !NotRewritable.count(Ptr: BO))
695 NodesToRewrite.push_back(Elt: BO);
696 salvageDebugInfo(I&: *Op);
697 Op->setOperand(i_nocapture: 0, Val_nocapture: NewLHS);
698 }
699 if (NewRHS != OldRHS) {
700 BinaryOperator *BO = isReassociableOp(V: OldRHS, Opcode);
701 if (BO && !NotRewritable.count(Ptr: BO))
702 NodesToRewrite.push_back(Elt: BO);
703 salvageDebugInfo(I&: *Op);
704 Op->setOperand(i_nocapture: 1, Val_nocapture: NewRHS);
705 }
706 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
707
708 ExpressionChangedStart = Op;
709 if (!ExpressionChangedEnd)
710 ExpressionChangedEnd = Op;
711 MadeChange = true;
712 ++NumChanged;
713
714 break;
715 }
716
717 // Not the last operation. The left-hand side will be a sub-expression
718 // while the right-hand side will be the current element of Ops.
719 Value *NewRHS = Ops[i].Op;
720 if (NewRHS != Op->getOperand(i_nocapture: 1)) {
721 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
722 if (NewRHS == Op->getOperand(i_nocapture: 0)) {
723 // The new right-hand side was already present as the left operand. If
724 // we are lucky then swapping the operands will sort out both of them.
725 Op->swapOperands();
726 } else {
727 // Overwrite with the new right-hand side.
728 BinaryOperator *BO = isReassociableOp(V: Op->getOperand(i_nocapture: 1), Opcode);
729 if (BO && !NotRewritable.count(Ptr: BO))
730 NodesToRewrite.push_back(Elt: BO);
731 salvageDebugInfo(I&: *Op);
732 Op->setOperand(i_nocapture: 1, Val_nocapture: NewRHS);
733 ExpressionChangedStart = Op;
734 if (!ExpressionChangedEnd)
735 ExpressionChangedEnd = Op;
736 }
737 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
738 MadeChange = true;
739 ++NumChanged;
740 }
741
742 // Now deal with the left-hand side. If this is already an operation node
743 // from the original expression then just rewrite the rest of the expression
744 // into it.
745 BinaryOperator *BO = isReassociableOp(V: Op->getOperand(i_nocapture: 0), Opcode);
746 if (BO && !NotRewritable.count(Ptr: BO)) {
747 Op = BO;
748 continue;
749 }
750
751 // Otherwise, grab a spare node from the original expression and use that as
752 // the left-hand side. If there are no nodes left then the optimizers made
753 // an expression with more nodes than the original! This usually means that
754 // they did something stupid but it might mean that the problem was just too
755 // hard (finding the mimimal number of multiplications needed to realize a
756 // multiplication expression is NP-complete). Whatever the reason, smart or
757 // stupid, create a new node if there are none left.
758 BinaryOperator *NewOp;
759 if (NodesToRewrite.empty()) {
760 Constant *Poison = PoisonValue::get(T: I->getType());
761 NewOp = BinaryOperator::Create(Op: Instruction::BinaryOps(Opcode), S1: Poison,
762 S2: Poison, Name: "", InsertBefore: I->getIterator());
763 if (isa<FPMathOperator>(Val: NewOp))
764 NewOp->setFastMathFlags(I->getFastMathFlags());
765 } else {
766 NewOp = NodesToRewrite.pop_back_val();
767 }
768
769 LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n');
770 salvageDebugInfo(I&: *Op);
771 Op->setOperand(i_nocapture: 0, Val_nocapture: NewOp);
772 LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n');
773 ExpressionChangedStart = Op;
774 if (!ExpressionChangedEnd)
775 ExpressionChangedEnd = Op;
776 MadeChange = true;
777 ++NumChanged;
778 Op = NewOp;
779 }
780
781 // If the expression changed non-trivially then clear out all subclass data
782 // starting from the operator specified in ExpressionChanged, and compactify
783 // the operators to just before the expression root to guarantee that the
784 // expression tree is dominated by all of Ops.
785 if (ExpressionChangedStart) {
786 bool ClearFlags = true;
787 do {
788 // Preserve flags.
789 if (ClearFlags) {
790 if (isa<FPMathOperator>(Val: I)) {
791 ExpressionChangedStart->copyFastMathFlags(FMF: I->getFastMathFlags());
792 } else {
793 Flags.applyFlags(I&: *ExpressionChangedStart);
794 }
795 }
796
797 if (ExpressionChangedStart == ExpressionChangedEnd)
798 ClearFlags = false;
799 if (ExpressionChangedStart == I)
800 break;
801
802 ExpressionChangedStart->moveBefore(InsertPos: I->getIterator());
803 ExpressionChangedStart =
804 cast<BinaryOperator>(Val: *ExpressionChangedStart->user_begin());
805 } while (true);
806 }
807
808 // Throw away any left over nodes from the original expression.
809 RedoInsts.insert_range(R&: NodesToRewrite);
810}
811
812/// Insert instructions before the instruction pointed to by BI,
813/// that computes the negative version of the value specified. The negative
814/// version of the value is returned, and BI is left pointing at the instruction
815/// that should be processed next by the reassociation pass.
816/// Also add intermediate instructions to the redo list that are modified while
817/// pushing the negates through adds. These will be revisited to see if
818/// additional opportunities have been exposed.
819static Value *NegateValue(Value *V, Instruction *BI,
820 ReassociatePass::OrderedSet &ToRedo) {
821 if (auto *C = dyn_cast<Constant>(Val: V)) {
822 const DataLayout &DL = BI->getDataLayout();
823 Constant *Res = C->getType()->isFPOrFPVectorTy()
824 ? ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)
825 : ConstantExpr::getNeg(C);
826 if (Res)
827 return Res;
828 }
829
830 // We are trying to expose opportunity for reassociation. One of the things
831 // that we want to do to achieve this is to push a negation as deep into an
832 // expression chain as possible, to expose the add instructions. In practice,
833 // this means that we turn this:
834 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
835 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
836 // the constants. We assume that instcombine will clean up the mess later if
837 // we introduce tons of unnecessary negation instructions.
838 //
839 if (BinaryOperator *I =
840 isReassociableOp(V, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd)) {
841 // Push the negates through the add.
842 I->setOperand(i_nocapture: 0, Val_nocapture: NegateValue(V: I->getOperand(i_nocapture: 0), BI, ToRedo));
843 I->setOperand(i_nocapture: 1, Val_nocapture: NegateValue(V: I->getOperand(i_nocapture: 1), BI, ToRedo));
844 if (I->getOpcode() == Instruction::Add) {
845 I->setHasNoUnsignedWrap(false);
846 I->setHasNoSignedWrap(false);
847 }
848
849 // We must move the add instruction here, because the neg instructions do
850 // not dominate the old add instruction in general. By moving it, we are
851 // assured that the neg instructions we just inserted dominate the
852 // instruction we are about to insert after them.
853 //
854 I->moveBefore(InsertPos: BI->getIterator());
855 I->setName(I->getName()+".neg");
856
857 // Add the intermediate negates to the redo list as processing them later
858 // could expose more reassociating opportunities.
859 ToRedo.insert(X: I);
860 return I;
861 }
862
863 // Okay, we need to materialize a negated version of V with an instruction.
864 // Scan the use lists of V to see if we have one already.
865 for (User *U : V->users()) {
866 if (!match(V: U, P: m_Neg(V: m_Value())) && !match(V: U, P: m_FNeg(X: m_Value())))
867 continue;
868
869 // We found one! Now we have to make sure that the definition dominates
870 // this use. We do this by moving it to the entry block (if it is a
871 // non-instruction value) or right after the definition. These negates will
872 // be zapped by reassociate later, so we don't need much finesse here.
873 Instruction *TheNeg = dyn_cast<Instruction>(Val: U);
874
875 // We can't safely propagate a vector zero constant with poison/undef lanes.
876 Constant *C;
877 if (match(V: TheNeg, P: m_BinOp(L: m_Constant(C), R: m_Value())) &&
878 C->containsUndefOrPoisonElement())
879 continue;
880
881 // Verify that the negate is in this function, V might be a constant expr.
882 if (!TheNeg ||
883 TheNeg->getParent()->getParent() != BI->getParent()->getParent())
884 continue;
885
886 BasicBlock::iterator InsertPt;
887 if (Instruction *InstInput = dyn_cast<Instruction>(Val: V)) {
888 auto InsertPtOpt = InstInput->getInsertionPointAfterDef();
889 if (!InsertPtOpt)
890 continue;
891 InsertPt = *InsertPtOpt;
892 } else {
893 InsertPt = TheNeg->getFunction()
894 ->getEntryBlock()
895 .getFirstNonPHIOrDbg()
896 ->getIterator();
897 }
898
899 // Check that if TheNeg is moved out of its parent block, we drop its
900 // debug location to avoid extra coverage.
901 // See test dropping_debugloc_the_neg.ll for a detailed example.
902 if (TheNeg->getParent() != InsertPt->getParent())
903 TheNeg->dropLocation();
904 TheNeg->moveBefore(BB&: *InsertPt->getParent(), I: InsertPt);
905
906 if (TheNeg->getOpcode() == Instruction::Sub) {
907 TheNeg->setHasNoUnsignedWrap(false);
908 TheNeg->setHasNoSignedWrap(false);
909 } else {
910 TheNeg->andIRFlags(V: BI);
911 }
912 ToRedo.insert(X: TheNeg);
913 return TheNeg;
914 }
915
916 // Insert a 'neg' instruction that subtracts the value from zero to get the
917 // negation.
918 Instruction *NewNeg =
919 CreateNeg(S1: V, Name: V->getName() + ".neg", InsertBefore: BI->getIterator(), FlagsOp: BI);
920 // NewNeg is generated to potentially replace BI, so use its DebugLoc.
921 NewNeg->setDebugLoc(BI->getDebugLoc());
922 ToRedo.insert(X: NewNeg);
923 return NewNeg;
924}
925
926// See if this `or` looks like an load widening reduction, i.e. that it
927// consists of an `or`/`shl`/`zext`/`load` nodes only. Note that we don't
928// ensure that the pattern is *really* a load widening reduction,
929// we do not ensure that it can really be replaced with a widened load,
930// only that it mostly looks like one.
931static bool isLoadCombineCandidate(Instruction *Or) {
932 SmallVector<Instruction *, 8> Worklist;
933 SmallPtrSet<Instruction *, 8> Visited;
934
935 auto Enqueue = [&](Value *V) {
936 auto *I = dyn_cast<Instruction>(Val: V);
937 // Each node of an `or` reduction must be an instruction,
938 if (!I)
939 return false; // Node is certainly not part of an `or` load reduction.
940 // Only process instructions we have never processed before.
941 if (Visited.insert(Ptr: I).second)
942 Worklist.emplace_back(Args&: I);
943 return true; // Will need to look at parent nodes.
944 };
945
946 if (!Enqueue(Or))
947 return false; // Not an `or` reduction pattern.
948
949 while (!Worklist.empty()) {
950 auto *I = Worklist.pop_back_val();
951
952 // Okay, which instruction is this node?
953 switch (I->getOpcode()) {
954 case Instruction::Or:
955 // Got an `or` node. That's fine, just recurse into it's operands.
956 for (Value *Op : I->operands())
957 if (!Enqueue(Op))
958 return false; // Not an `or` reduction pattern.
959 continue;
960
961 case Instruction::Shl:
962 case Instruction::ZExt:
963 // `shl`/`zext` nodes are fine, just recurse into their base operand.
964 if (!Enqueue(I->getOperand(i: 0)))
965 return false; // Not an `or` reduction pattern.
966 continue;
967
968 case Instruction::Load:
969 // Perfect, `load` node means we've reached an edge of the graph.
970 continue;
971
972 default: // Unknown node.
973 return false; // Not an `or` reduction pattern.
974 }
975 }
976
977 return true;
978}
979
980/// Return true if it may be profitable to convert this (X|Y) into (X+Y).
981static bool shouldConvertOrWithNoCommonBitsToAdd(Instruction *Or) {
982 // Don't bother to convert this up unless either the LHS is an associable add
983 // or subtract or mul or if this is only used by one of the above.
984 // This is only a compile-time improvement, it is not needed for correctness!
985 auto isInteresting = [](Value *V) {
986 for (auto Op : {Instruction::Add, Instruction::Sub, Instruction::Mul,
987 Instruction::Shl})
988 if (isReassociableOp(V, Opcode: Op))
989 return true;
990 return false;
991 };
992
993 if (any_of(Range: Or->operands(), P: isInteresting))
994 return true;
995
996 Value *VB = Or->user_back();
997 if (Or->hasOneUse() && isInteresting(VB))
998 return true;
999
1000 return false;
1001}
1002
1003/// If we have (X|Y), and iff X and Y have no common bits set,
1004/// transform this into (X+Y) to allow arithmetics reassociation.
1005static BinaryOperator *convertOrWithNoCommonBitsToAdd(Instruction *Or) {
1006 // Convert an or into an add.
1007 BinaryOperator *New = CreateAdd(S1: Or->getOperand(i: 0), S2: Or->getOperand(i: 1), Name: "",
1008 InsertBefore: Or->getIterator(), FlagsOp: Or);
1009 New->setHasNoSignedWrap();
1010 New->setHasNoUnsignedWrap();
1011 New->takeName(V: Or);
1012
1013 // Everyone now refers to the add instruction.
1014 Or->replaceAllUsesWith(V: New);
1015 New->setDebugLoc(Or->getDebugLoc());
1016
1017 LLVM_DEBUG(dbgs() << "Converted or into an add: " << *New << '\n');
1018 return New;
1019}
1020
1021/// Return true if Mul is of the form (X+Y)*C or (X-Y)*C where C is a
1022/// constant, and there exists a sibling instruction of the form X*C' or Y*C'
1023/// in the same expression — indicating that distribution followed by
1024/// factoring will reduce the instruction count.
1025static bool ShouldBreakUpDistribution(Instruction *Mul) {
1026 Value *A, *B;
1027 if (!match(V: Mul, P: m_OneUse(SubPattern: m_Mul(
1028 L: m_OneUse(SubPattern: m_CombineOr(Ps: m_Add(L: m_Value(V&: A), R: m_Value(V&: B)),
1029 Ps: m_Sub(L: m_Value(V&: A), R: m_Value(V&: B)))),
1030 R: m_ImmConstant()))))
1031 return false;
1032
1033 auto *MulUser = cast<Instruction>(Val: Mul->user_back());
1034 // The parent MUST be an Add or Sub to ensure the tree is flattened
1035 if (MulUser->getOpcode() != Instruction::Add &&
1036 MulUser->getOpcode() != Instruction::Sub)
1037 return false;
1038
1039 for (Value *Sibling : MulUser->operands()) {
1040 if (Sibling == Mul || !Sibling->hasOneUse())
1041 continue;
1042
1043 // Sibling must be NonConst * C'.
1044 Value *SibNC;
1045 if (match(V: Sibling, P: m_Mul(L: m_Value(V&: SibNC), R: m_ImmConstant())) &&
1046 (SibNC == A || SibNC == B) && !isa<Constant>(Val: SibNC))
1047 return true;
1048 }
1049 return false;
1050}
1051
1052/// Distribute Mul of the form (X+Y)*C into X*C + Y*C.
1053/// For the sub case (X-Y)*C, the second term uses -C to avoid
1054/// introducing a negation instruction.
1055static BinaryOperator *BreakUpDistribute(Instruction *Mul,
1056 ReassociatePass::OrderedSet &ToRedo) {
1057 Instruction *AddSub = cast<Instruction>(Val: Mul->getOperand(i: 0));
1058 Constant *C = cast<Constant>(Val: Mul->getOperand(i: 1));
1059 Constant *C2 =
1060 AddSub->getOpcode() == Instruction::Sub ? ConstantExpr::getNeg(C) : C;
1061
1062 BinaryOperator *M1 = BinaryOperator::CreateMul(V1: AddSub->getOperand(i: 0), V2: C,
1063 Name: "Mul1", InsertBefore: Mul->getIterator());
1064 BinaryOperator *M2 = BinaryOperator::CreateMul(V1: AddSub->getOperand(i: 1), V2: C2,
1065 Name: "Mul2", InsertBefore: Mul->getIterator());
1066 BinaryOperator *Result =
1067 BinaryOperator::CreateAdd(V1: M1, V2: M2, Name: "DistAdd", InsertBefore: Mul->getIterator());
1068
1069 Mul->replaceAllUsesWith(V: Result);
1070 Result->setDebugLoc(Mul->getDebugLoc());
1071
1072 ToRedo.insert(X: M1);
1073 ToRedo.insert(X: M2);
1074 ToRedo.insert(X: Result);
1075
1076 return Result;
1077}
1078
1079/// Return true if we should break up this subtract of X-Y into (X + -Y).
1080static bool ShouldBreakUpSubtract(Instruction *Sub) {
1081 // If this is a negation, we can't split it up!
1082 if (match(V: Sub, P: m_Neg(V: m_Value())) || match(V: Sub, P: m_FNeg(X: m_Value())))
1083 return false;
1084
1085 // Don't breakup X - undef.
1086 if (isa<UndefValue>(Val: Sub->getOperand(i: 1)))
1087 return false;
1088
1089 // Don't bother to break this up unless either the LHS is an associable add or
1090 // subtract or if this is only used by one.
1091 Value *V0 = Sub->getOperand(i: 0);
1092 if (isReassociableOp(V: V0, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd) ||
1093 isReassociableOp(V: V0, Opcode1: Instruction::Sub, Opcode2: Instruction::FSub))
1094 return true;
1095 Value *V1 = Sub->getOperand(i: 1);
1096 if (isReassociableOp(V: V1, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd) ||
1097 isReassociableOp(V: V1, Opcode1: Instruction::Sub, Opcode2: Instruction::FSub))
1098 return true;
1099 Value *VB = Sub->user_back();
1100 if (Sub->hasOneUse() &&
1101 (isReassociableOp(V: VB, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd) ||
1102 isReassociableOp(V: VB, Opcode1: Instruction::Sub, Opcode2: Instruction::FSub)))
1103 return true;
1104
1105 return false;
1106}
1107
1108/// If we have (X-Y), and if either X is an add, or if this is only used by an
1109/// add, transform this into (X+(0-Y)) to promote better reassociation.
1110static BinaryOperator *BreakUpSubtract(Instruction *Sub,
1111 ReassociatePass::OrderedSet &ToRedo) {
1112 // Convert a subtract into an add and a neg instruction. This allows sub
1113 // instructions to be commuted with other add instructions.
1114 //
1115 // Calculate the negative value of Operand 1 of the sub instruction,
1116 // and set it as the RHS of the add instruction we just made.
1117 Value *NegVal = NegateValue(V: Sub->getOperand(i: 1), BI: Sub, ToRedo);
1118 BinaryOperator *New =
1119 CreateAdd(S1: Sub->getOperand(i: 0), S2: NegVal, Name: "", InsertBefore: Sub->getIterator(), FlagsOp: Sub);
1120 Sub->setOperand(i: 0, Val: Constant::getNullValue(Ty: Sub->getType())); // Drop use of op.
1121 Sub->setOperand(i: 1, Val: Constant::getNullValue(Ty: Sub->getType())); // Drop use of op.
1122 New->takeName(V: Sub);
1123
1124 // Everyone now refers to the add instruction.
1125 Sub->replaceAllUsesWith(V: New);
1126 New->setDebugLoc(Sub->getDebugLoc());
1127
1128 LLVM_DEBUG(dbgs() << "Negated: " << *New << '\n');
1129 return New;
1130}
1131
1132/// If this is a shift of a reassociable multiply or is used by one, change
1133/// this into a multiply by a constant to assist with further reassociation.
1134static BinaryOperator *ConvertShiftToMul(Instruction *Shl) {
1135 Constant *MulCst = ConstantInt::get(Ty: Shl->getType(), V: 1);
1136 auto *SA = cast<ConstantInt>(Val: Shl->getOperand(i: 1));
1137 MulCst = ConstantFoldBinaryInstruction(Opcode: Instruction::Shl, V1: MulCst, V2: SA);
1138 assert(MulCst && "Constant folding of immediate constants failed");
1139
1140 BinaryOperator *Mul = BinaryOperator::CreateMul(V1: Shl->getOperand(i: 0), V2: MulCst,
1141 Name: "", InsertBefore: Shl->getIterator());
1142 Shl->setOperand(i: 0, Val: PoisonValue::get(T: Shl->getType())); // Drop use of op.
1143 Mul->takeName(V: Shl);
1144
1145 // Everyone now refers to the mul instruction.
1146 Shl->replaceAllUsesWith(V: Mul);
1147 Mul->setDebugLoc(Shl->getDebugLoc());
1148
1149 // We can safely preserve the nuw flag in all cases. It's also safe to turn a
1150 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special
1151 // handling. It can be preserved as long as we're not left shifting by
1152 // bitwidth - 1.
1153 bool NSW = cast<BinaryOperator>(Val: Shl)->hasNoSignedWrap();
1154 bool NUW = cast<BinaryOperator>(Val: Shl)->hasNoUnsignedWrap();
1155 unsigned BitWidth = Shl->getType()->getScalarSizeInBits();
1156 if (NSW && (NUW || SA->getValue().ult(RHS: BitWidth - 1)))
1157 Mul->setHasNoSignedWrap(true);
1158 Mul->setHasNoUnsignedWrap(NUW);
1159 return Mul;
1160}
1161
1162/// Scan backwards and forwards among values with the same rank as element i
1163/// to see if X exists. If X does not exist, return i. This is useful when
1164/// scanning for 'x' when we see '-x' because they both get the same rank.
1165static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops,
1166 unsigned i, Value *X) {
1167 unsigned XRank = Ops[i].Rank;
1168 unsigned e = Ops.size();
1169 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) {
1170 if (Ops[j].Op == X)
1171 return j;
1172 if (Instruction *I1 = dyn_cast<Instruction>(Val: Ops[j].Op))
1173 if (Instruction *I2 = dyn_cast<Instruction>(Val: X))
1174 if (I1->isIdenticalTo(I: I2))
1175 return j;
1176 }
1177 // Scan backwards.
1178 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) {
1179 if (Ops[j].Op == X)
1180 return j;
1181 if (Instruction *I1 = dyn_cast<Instruction>(Val: Ops[j].Op))
1182 if (Instruction *I2 = dyn_cast<Instruction>(Val: X))
1183 if (I1->isIdenticalTo(I: I2))
1184 return j;
1185 }
1186 return i;
1187}
1188
1189/// Emit a tree of add instructions, summing Ops together
1190/// and returning the result. Insert the tree before I.
1191static Value *EmitAddTreeOfValues(Instruction *I,
1192 SmallVectorImpl<WeakTrackingVH> &Ops) {
1193 if (Ops.size() == 1) return Ops.back();
1194
1195 Value *V1 = Ops.pop_back_val();
1196 Value *V2 = EmitAddTreeOfValues(I, Ops);
1197 auto *NewAdd = CreateAdd(S1: V2, S2: V1, Name: "reass.add", InsertBefore: I->getIterator(), FlagsOp: I);
1198 NewAdd->setDebugLoc(I->getDebugLoc());
1199 return NewAdd;
1200}
1201
1202/// If V is an expression tree that is a multiplication sequence,
1203/// and if this sequence contains a multiply by Factor,
1204/// remove Factor from the tree and return the new tree.
1205/// If new instructions are inserted to generate this tree, DL should be used
1206/// as the DebugLoc for these instructions.
1207Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor,
1208 DebugLoc DL) {
1209 BinaryOperator *BO = isReassociableOp(V, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul);
1210 if (!BO)
1211 return nullptr;
1212
1213 SmallVector<RepeatedValue, 8> Tree;
1214 OverflowTracking Flags;
1215 MadeChange |= LinearizeExprTree(I: BO, Ops&: Tree, ToRedo&: RedoInsts, Flags);
1216 SmallVector<ValueEntry, 8> Factors;
1217 Factors.reserve(N: Tree.size());
1218 for (const RepeatedValue &E : Tree)
1219 Factors.append(NumInputs: E.second, Elt: ValueEntry(getRank(V: E.first), E.first));
1220
1221 bool FoundFactor = false;
1222 bool NeedsNegate = false;
1223 for (unsigned i = 0, e = Factors.size(); i != e; ++i) {
1224 if (Factors[i].Op == Factor) {
1225 FoundFactor = true;
1226 Factors.erase(CI: Factors.begin()+i);
1227 break;
1228 }
1229
1230 // If this is a negative version of this factor, remove it.
1231 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Val: Factor)) {
1232 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Val: Factors[i].Op))
1233 if (FC1->getValue() == -FC2->getValue()) {
1234 FoundFactor = NeedsNegate = true;
1235 Factors.erase(CI: Factors.begin()+i);
1236 break;
1237 }
1238 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Val: Factor)) {
1239 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Val: Factors[i].Op)) {
1240 const APFloat &F1 = FC1->getValueAPF();
1241 APFloat F2(FC2->getValueAPF());
1242 F2.changeSign();
1243 if (F1 == F2) {
1244 FoundFactor = NeedsNegate = true;
1245 Factors.erase(CI: Factors.begin() + i);
1246 break;
1247 }
1248 }
1249 }
1250 }
1251
1252 if (!FoundFactor) {
1253 // Make sure to restore the operands to the expression tree.
1254 RewriteExprTree(I: BO, Ops&: Factors, Flags);
1255 return nullptr;
1256 }
1257
1258 BasicBlock::iterator InsertPt = ++BO->getIterator();
1259
1260 // If this was just a single multiply, remove the multiply and return the only
1261 // remaining operand.
1262 if (Factors.size() == 1) {
1263 RedoInsts.insert(X: BO);
1264 V = Factors[0].Op;
1265 } else {
1266 RewriteExprTree(I: BO, Ops&: Factors, Flags);
1267 V = BO;
1268 }
1269
1270 if (NeedsNegate) {
1271 V = CreateNeg(S1: V, Name: "neg", InsertBefore: InsertPt, FlagsOp: BO);
1272 cast<Instruction>(Val: V)->setDebugLoc(DL);
1273 }
1274
1275 return V;
1276}
1277
1278/// If V is a single-use multiply, recursively add its operands as factors,
1279/// otherwise add V to the list of factors.
1280///
1281/// Ops is the top-level list of add operands we're trying to factor.
1282static void FindSingleUseMultiplyFactors(Value *V,
1283 SmallVectorImpl<Value*> &Factors) {
1284 BinaryOperator *BO = isReassociableOp(V, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul);
1285 if (!BO) {
1286 Factors.push_back(Elt: V);
1287 return;
1288 }
1289
1290 // Otherwise, add the LHS and RHS to the list of factors.
1291 FindSingleUseMultiplyFactors(V: BO->getOperand(i_nocapture: 1), Factors);
1292 FindSingleUseMultiplyFactors(V: BO->getOperand(i_nocapture: 0), Factors);
1293}
1294
1295/// Optimize a series of operands to an 'and', 'or', or 'xor' instruction.
1296/// This optimizes based on identities. If it can be reduced to a single Value,
1297/// it is returned, otherwise the Ops list is mutated as necessary.
1298static Value *OptimizeAndOrXor(unsigned Opcode,
1299 SmallVectorImpl<ValueEntry> &Ops) {
1300 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs.
1301 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1.
1302 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1303 // First, check for X and ~X in the operand list.
1304 assert(i < Ops.size());
1305 Value *X;
1306 if (match(V: Ops[i].Op, P: m_Not(V: m_Value(V&: X)))) { // Cannot occur for ^.
1307 unsigned FoundX = FindInOperandList(Ops, i, X);
1308 if (FoundX != i) {
1309 if (Opcode == Instruction::And) // ...&X&~X = 0
1310 return Constant::getNullValue(Ty: X->getType());
1311
1312 if (Opcode == Instruction::Or) // ...|X|~X = -1
1313 return Constant::getAllOnesValue(Ty: X->getType());
1314 }
1315 }
1316
1317 // Next, check for duplicate pairs of values, which we assume are next to
1318 // each other, due to our sorting criteria.
1319 assert(i < Ops.size());
1320 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) {
1321 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
1322 // Drop duplicate values for And and Or.
1323 Ops.erase(CI: Ops.begin()+i);
1324 --i; --e;
1325 ++NumAnnihil;
1326 continue;
1327 }
1328
1329 // Drop pairs of values for Xor.
1330 assert(Opcode == Instruction::Xor);
1331 if (e == 2)
1332 return Constant::getNullValue(Ty: Ops[0].Op->getType());
1333
1334 // Y ^ X^X -> Y
1335 Ops.erase(CS: Ops.begin()+i, CE: Ops.begin()+i+2);
1336 i -= 1; e -= 2;
1337 ++NumAnnihil;
1338 }
1339 }
1340 return nullptr;
1341}
1342
1343/// Helper function of CombineXorOpnd(). It creates a bitwise-and
1344/// instruction with the given two operands, and return the resulting
1345/// instruction. There are two special cases: 1) if the constant operand is 0,
1346/// it will return NULL. 2) if the constant is ~0, the symbolic operand will
1347/// be returned.
1348static Value *createAndInstr(BasicBlock::iterator InsertBefore, Value *Opnd,
1349 const APInt &ConstOpnd) {
1350 if (ConstOpnd.isZero())
1351 return nullptr;
1352
1353 if (ConstOpnd.isAllOnes())
1354 return Opnd;
1355
1356 Instruction *I = BinaryOperator::CreateAnd(
1357 V1: Opnd, V2: ConstantInt::get(Ty: Opnd->getType(), V: ConstOpnd), Name: "and.ra",
1358 InsertBefore);
1359 I->setDebugLoc(InsertBefore->getDebugLoc());
1360 return I;
1361}
1362
1363// Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd"
1364// into "R ^ C", where C would be 0, and R is a symbolic value.
1365//
1366// If it was successful, true is returned, and the "R" and "C" is returned
1367// via "Res" and "ConstOpnd", respectively; otherwise, false is returned,
1368// and both "Res" and "ConstOpnd" remain unchanged.
1369bool ReassociatePass::CombineXorOpnd(BasicBlock::iterator It, XorOpnd *Opnd1,
1370 APInt &ConstOpnd, Value *&Res) {
1371 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2
1372 // = ((x | c1) ^ c1) ^ (c1 ^ c2)
1373 // = (x & ~c1) ^ (c1 ^ c2)
1374 // It is useful only when c1 == c2.
1375 if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isZero())
1376 return false;
1377
1378 if (!Opnd1->getValue()->hasOneUse())
1379 return false;
1380
1381 const APInt &C1 = Opnd1->getConstPart();
1382 if (C1 != ConstOpnd)
1383 return false;
1384
1385 Value *X = Opnd1->getSymbolicPart();
1386 Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: ~C1);
1387 // ConstOpnd was C2, now C1 ^ C2.
1388 ConstOpnd ^= C1;
1389
1390 if (Instruction *T = dyn_cast<Instruction>(Val: Opnd1->getValue()))
1391 RedoInsts.insert(X: T);
1392 return true;
1393}
1394
1395// Helper function of OptimizeXor(). It tries to simplify
1396// "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a
1397// symbolic value.
1398//
1399// If it was successful, true is returned, and the "R" and "C" is returned
1400// via "Res" and "ConstOpnd", respectively (If the entire expression is
1401// evaluated to a constant, the Res is set to NULL); otherwise, false is
1402// returned, and both "Res" and "ConstOpnd" remain unchanged.
1403bool ReassociatePass::CombineXorOpnd(BasicBlock::iterator It, XorOpnd *Opnd1,
1404 XorOpnd *Opnd2, APInt &ConstOpnd,
1405 Value *&Res) {
1406 Value *X = Opnd1->getSymbolicPart();
1407 if (X != Opnd2->getSymbolicPart())
1408 return false;
1409
1410 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.)
1411 int DeadInstNum = 1;
1412 if (Opnd1->getValue()->hasOneUse())
1413 DeadInstNum++;
1414 if (Opnd2->getValue()->hasOneUse())
1415 DeadInstNum++;
1416
1417 // Xor-Rule 2:
1418 // (x | c1) ^ (x & c2)
1419 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1
1420 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1
1421 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3
1422 //
1423 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) {
1424 if (Opnd2->isOrExpr())
1425 std::swap(a&: Opnd1, b&: Opnd2);
1426
1427 const APInt &C1 = Opnd1->getConstPart();
1428 const APInt &C2 = Opnd2->getConstPart();
1429 APInt C3((~C1) ^ C2);
1430
1431 // Do not increase code size!
1432 if (!C3.isZero() && !C3.isAllOnes()) {
1433 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1434 if (NewInstNum > DeadInstNum)
1435 return false;
1436 }
1437
1438 Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: C3);
1439 ConstOpnd ^= C1;
1440 } else if (Opnd1->isOrExpr()) {
1441 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
1442 //
1443 const APInt &C1 = Opnd1->getConstPart();
1444 const APInt &C2 = Opnd2->getConstPart();
1445 APInt C3 = C1 ^ C2;
1446
1447 // Do not increase code size
1448 if (!C3.isZero() && !C3.isAllOnes()) {
1449 int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2;
1450 if (NewInstNum > DeadInstNum)
1451 return false;
1452 }
1453
1454 Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: C3);
1455 ConstOpnd ^= C3;
1456 } else {
1457 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2))
1458 //
1459 const APInt &C1 = Opnd1->getConstPart();
1460 const APInt &C2 = Opnd2->getConstPart();
1461 APInt C3 = C1 ^ C2;
1462 Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: C3);
1463 }
1464
1465 // Put the original operands in the Redo list; hope they will be deleted
1466 // as dead code.
1467 if (Instruction *T = dyn_cast<Instruction>(Val: Opnd1->getValue()))
1468 RedoInsts.insert(X: T);
1469 if (Instruction *T = dyn_cast<Instruction>(Val: Opnd2->getValue()))
1470 RedoInsts.insert(X: T);
1471
1472 return true;
1473}
1474
1475/// Optimize a series of operands to an 'xor' instruction. If it can be reduced
1476/// to a single Value, it is returned, otherwise the Ops list is mutated as
1477/// necessary.
1478Value *ReassociatePass::OptimizeXor(Instruction *I,
1479 SmallVectorImpl<ValueEntry> &Ops) {
1480 if (Value *V = OptimizeAndOrXor(Opcode: Instruction::Xor, Ops))
1481 return V;
1482
1483 if (Ops.size() == 1)
1484 return nullptr;
1485
1486 SmallVector<XorOpnd, 8> Opnds;
1487 SmallVector<XorOpnd*, 8> OpndPtrs;
1488 Type *Ty = Ops[0].Op->getType();
1489 APInt ConstOpnd(Ty->getScalarSizeInBits(), 0);
1490
1491 // Step 1: Convert ValueEntry to XorOpnd
1492 for (const ValueEntry &Op : Ops) {
1493 Value *V = Op.Op;
1494 const APInt *C;
1495 // TODO: Support non-splat vectors.
1496 if (match(V, P: m_APInt(Res&: C))) {
1497 ConstOpnd ^= *C;
1498 } else {
1499 XorOpnd O(V);
1500 O.setSymbolicRank(getRank(V: O.getSymbolicPart()));
1501 Opnds.push_back(Elt: O);
1502 }
1503 }
1504
1505 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds".
1506 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate
1507 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop
1508 // with the previous loop --- the iterator of the "Opnds" may be invalidated
1509 // when new elements are added to the vector.
1510 for (XorOpnd &Op : Opnds)
1511 OpndPtrs.push_back(Elt: &Op);
1512
1513 // Step 2: Sort the Xor-Operands in a way such that the operands containing
1514 // the same symbolic value cluster together. For instance, the input operand
1515 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into:
1516 // ("x | 123", "x & 789", "y & 456").
1517 //
1518 // The purpose is twofold:
1519 // 1) Cluster together the operands sharing the same symbolic-value.
1520 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which
1521 // could potentially shorten crital path, and expose more loop-invariants.
1522 // Note that values' rank are basically defined in RPO order (FIXME).
1523 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier
1524 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2",
1525 // "z" in the order of X-Y-Z is better than any other orders.
1526 llvm::stable_sort(Range&: OpndPtrs, C: [](XorOpnd *LHS, XorOpnd *RHS) {
1527 return LHS->getSymbolicRank() < RHS->getSymbolicRank();
1528 });
1529
1530 // Step 3: Combine adjacent operands
1531 XorOpnd *PrevOpnd = nullptr;
1532 bool Changed = false;
1533 for (unsigned i = 0, e = Opnds.size(); i < e; i++) {
1534 XorOpnd *CurrOpnd = OpndPtrs[i];
1535 // The combined value
1536 Value *CV;
1537
1538 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd"
1539 if (!ConstOpnd.isZero() &&
1540 CombineXorOpnd(It: I->getIterator(), Opnd1: CurrOpnd, ConstOpnd, Res&: CV)) {
1541 Changed = true;
1542 if (CV)
1543 *CurrOpnd = XorOpnd(CV);
1544 else {
1545 CurrOpnd->Invalidate();
1546 continue;
1547 }
1548 }
1549
1550 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) {
1551 PrevOpnd = CurrOpnd;
1552 continue;
1553 }
1554
1555 // step 3.2: When previous and current operands share the same symbolic
1556 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd"
1557 if (CombineXorOpnd(It: I->getIterator(), Opnd1: CurrOpnd, Opnd2: PrevOpnd, ConstOpnd, Res&: CV)) {
1558 // Remove previous operand
1559 PrevOpnd->Invalidate();
1560 if (CV) {
1561 *CurrOpnd = XorOpnd(CV);
1562 PrevOpnd = CurrOpnd;
1563 } else {
1564 CurrOpnd->Invalidate();
1565 PrevOpnd = nullptr;
1566 }
1567 Changed = true;
1568 }
1569 }
1570
1571 // Step 4: Reassemble the Ops
1572 if (Changed) {
1573 Ops.clear();
1574 for (const XorOpnd &O : Opnds) {
1575 if (O.isInvalid())
1576 continue;
1577 ValueEntry VE(getRank(V: O.getValue()), O.getValue());
1578 Ops.push_back(Elt: VE);
1579 }
1580 if (!ConstOpnd.isZero()) {
1581 Value *C = ConstantInt::get(Ty, V: ConstOpnd);
1582 ValueEntry VE(getRank(V: C), C);
1583 Ops.push_back(Elt: VE);
1584 }
1585 unsigned Sz = Ops.size();
1586 if (Sz == 1)
1587 return Ops.back().Op;
1588 if (Sz == 0) {
1589 assert(ConstOpnd.isZero());
1590 return ConstantInt::get(Ty, V: ConstOpnd);
1591 }
1592 }
1593
1594 return nullptr;
1595}
1596
1597/// Optimize a series of operands to an 'add' instruction. This
1598/// optimizes based on identities. If it can be reduced to a single Value, it
1599/// is returned, otherwise the Ops list is mutated as necessary.
1600Value *ReassociatePass::OptimizeAdd(Instruction *I,
1601 SmallVectorImpl<ValueEntry> &Ops) {
1602 // Scan the operand lists looking for X and -X pairs. If we find any, we
1603 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it,
1604 // scan for any
1605 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z.
1606
1607 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1608 Value *TheOp = Ops[i].Op;
1609 // Check to see if we've seen this operand before. If so, we factor all
1610 // instances of the operand together. Due to our sorting criteria, we know
1611 // that these need to be next to each other in the vector.
1612 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
1613 // Rescan the list, remove all instances of this operand from the expr.
1614 unsigned NumFound = 0;
1615 do {
1616 Ops.erase(CI: Ops.begin()+i);
1617 ++NumFound;
1618 } while (i != Ops.size() && Ops[i].Op == TheOp);
1619
1620 LLVM_DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp
1621 << '\n');
1622 ++NumFactor;
1623
1624 // Insert a new multiply.
1625 Type *Ty = TheOp->getType();
1626 // Truncate if NumFound overflows the type.
1627 Constant *C = Ty->isIntOrIntVectorTy()
1628 ? ConstantInt::get(Ty, V: NumFound, /*IsSigned=*/false,
1629 /*ImplicitTrunc=*/true)
1630 : ConstantFP::get(Ty, V: NumFound);
1631 Instruction *Mul = CreateMul(S1: TheOp, S2: C, Name: "factor", InsertBefore: I->getIterator(), FlagsOp: I);
1632 Mul->setDebugLoc(I->getDebugLoc());
1633
1634 // Now that we have inserted a multiply, optimize it. This allows us to
1635 // handle cases that require multiple factoring steps, such as this:
1636 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
1637 RedoInsts.insert(X: Mul);
1638
1639 // If every add operand was a duplicate, return the multiply.
1640 if (Ops.empty())
1641 return Mul;
1642
1643 // Otherwise, we had some input that didn't have the dupe, such as
1644 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of
1645 // things being added by this operation.
1646 Ops.insert(I: Ops.begin(), Elt: ValueEntry(getRank(V: Mul), Mul));
1647
1648 --i;
1649 e = Ops.size();
1650 continue;
1651 }
1652
1653 // Check for X and -X or X and ~X in the operand list.
1654 Value *X;
1655 if (!match(V: TheOp, P: m_Neg(V: m_Value(V&: X))) && !match(V: TheOp, P: m_Not(V: m_Value(V&: X))) &&
1656 !match(V: TheOp, P: m_FNeg(X: m_Value(V&: X))))
1657 continue;
1658
1659 unsigned FoundX = FindInOperandList(Ops, i, X);
1660 if (FoundX == i)
1661 continue;
1662
1663 // Remove X and -X from the operand list.
1664 if (Ops.size() == 2 &&
1665 (match(V: TheOp, P: m_Neg(V: m_Value())) || match(V: TheOp, P: m_FNeg(X: m_Value()))))
1666 return Constant::getNullValue(Ty: X->getType());
1667
1668 // Remove X and ~X from the operand list.
1669 if (Ops.size() == 2 && match(V: TheOp, P: m_Not(V: m_Value())))
1670 return Constant::getAllOnesValue(Ty: X->getType());
1671
1672 Ops.erase(CI: Ops.begin()+i);
1673 if (i < FoundX)
1674 --FoundX;
1675 else
1676 --i; // Need to back up an extra one.
1677 Ops.erase(CI: Ops.begin()+FoundX);
1678 ++NumAnnihil;
1679 --i; // Revisit element.
1680 e -= 2; // Removed two elements.
1681
1682 // if X and ~X we append -1 to the operand list.
1683 if (match(V: TheOp, P: m_Not(V: m_Value()))) {
1684 Value *V = Constant::getAllOnesValue(Ty: X->getType());
1685 Ops.insert(I: Ops.end(), Elt: ValueEntry(getRank(V), V));
1686 e += 1;
1687 }
1688 }
1689
1690 // Scan the operand list, checking to see if there are any common factors
1691 // between operands. Consider something like A*A+A*B*C+D. We would like to
1692 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies.
1693 // To efficiently find this, we count the number of times a factor occurs
1694 // for any ADD operands that are MULs.
1695 DenseMap<Value*, unsigned> FactorOccurrences;
1696
1697 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4)
1698 // where they are actually the same multiply.
1699 unsigned MaxOcc = 0;
1700 Value *MaxOccVal = nullptr;
1701
1702 // Prefer a non-constant factor over a constant when occurrence counts
1703 // tie. Factoring out a variable (e.g., X from X*C1 + X*C2) exposes
1704 // downstream constant folding; factoring out a constant does not.
1705 auto IsBetterFactor = [](Value *Factor, Value *MaxOccVal, unsigned Occ,
1706 unsigned MaxOcc) {
1707 return Occ > MaxOcc ||
1708 (Occ == MaxOcc &&
1709 (isa<Instruction>(Val: Factor) || isa<Argument>(Val: Factor)) &&
1710 isa<Constant>(Val: MaxOccVal) && !isa<UndefValue>(Val: MaxOccVal));
1711 };
1712 auto CountFactors = [&](BinaryOperator *BOp) {
1713 // Compute all of the factors of this added value.
1714 SmallVector<Value*, 8> Factors;
1715 FindSingleUseMultiplyFactors(V: BOp, Factors);
1716 assert(Factors.size() > 1 && "Bad linearize!");
1717
1718 // Add one to FactorOccurrences for each unique factor in this op.
1719 SmallPtrSet<Value*, 8> Duplicates;
1720 for (Value *Factor : Factors) {
1721 if (!Duplicates.insert(Ptr: Factor).second)
1722 continue;
1723
1724 unsigned Occ = ++FactorOccurrences[Factor];
1725 if (IsBetterFactor(Factor, MaxOccVal, Occ, MaxOcc)) {
1726 MaxOcc = Occ;
1727 MaxOccVal = Factor;
1728 }
1729
1730 // If Factor is a negative constant, add the negated value as a factor
1731 // because we can percolate the negate out. Watch for minint, which
1732 // cannot be positivified.
1733 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Factor)) {
1734 if (CI->isNegative() && !CI->isMinValue(IsSigned: true)) {
1735 Factor = ConstantInt::get(Context&: CI->getContext(), V: -CI->getValue());
1736 if (!Duplicates.insert(Ptr: Factor).second)
1737 continue;
1738 unsigned Occ = ++FactorOccurrences[Factor];
1739 if (IsBetterFactor(Factor, MaxOccVal, Occ, MaxOcc)) {
1740 MaxOcc = Occ;
1741 MaxOccVal = Factor;
1742 }
1743 }
1744 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Val: Factor)) {
1745 if (CF->isNegative()) {
1746 APFloat F(CF->getValueAPF());
1747 F.changeSign();
1748 Factor = ConstantFP::get(Ty: CF->getType(), V: F);
1749 if (!Duplicates.insert(Ptr: Factor).second)
1750 continue;
1751 unsigned Occ = ++FactorOccurrences[Factor];
1752 if (IsBetterFactor(Factor, MaxOccVal, Occ, MaxOcc)) {
1753 MaxOcc = Occ;
1754 MaxOccVal = Factor;
1755 }
1756 }
1757 }
1758 }
1759 };
1760
1761 // fmul/fadd pairs kept together for fma hide their muls; count the factors
1762 // of the reassociable ones as well and break those pairs up if a repeated
1763 // factor exists, so that factorization still applies.
1764 SmallVector<Value *> FMulAddCands;
1765 for (const ValueEntry &Entry : Ops) {
1766 if (BinaryOperator *BOp =
1767 isReassociableOp(V: Entry.Op, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul)) {
1768 CountFactors(BOp);
1769 continue;
1770 }
1771 if (BinaryOperator *BOp = isFMulAddCandidate(V: Entry.Op);
1772 BOp && hasFPAssociativeFlags(I: BOp)) {
1773 FMulAddCands.push_back(Elt: Entry.Op);
1774 CountFactors(BOp);
1775 }
1776 }
1777
1778 if (MaxOcc > 1) {
1779 for (Value *V : FMulAddCands) {
1780 erase_if(C&: Ops, P: [V](const ValueEntry &E) { return E.Op == V; });
1781 for (Value *Op : cast<BinaryOperator>(Val: V)->operands())
1782 Ops.emplace_back(Args: getRank(V: Op), Args&: Op);
1783 }
1784 }
1785
1786 // If any factor occurred more than one time, we can pull it out.
1787 if (MaxOcc > 1) {
1788 LLVM_DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal
1789 << '\n');
1790 ++NumFactor;
1791
1792 // Create a new instruction that uses the MaxOccVal twice. If we don't do
1793 // this, we could otherwise run into situations where removing a factor
1794 // from an expression will drop a use of maxocc, and this can cause
1795 // RemoveFactorFromExpression on successive values to behave differently.
1796 Instruction *DummyInst =
1797 I->getType()->isIntOrIntVectorTy()
1798 ? BinaryOperator::CreateAdd(V1: MaxOccVal, V2: MaxOccVal)
1799 : BinaryOperator::CreateFAdd(V1: MaxOccVal, V2: MaxOccVal);
1800
1801 SmallVector<WeakTrackingVH, 4> NewMulOps;
1802 for (unsigned i = 0; i != Ops.size(); ++i) {
1803 // Only try to remove factors from expressions we're allowed to.
1804 BinaryOperator *BOp =
1805 isReassociableOp(V: Ops[i].Op, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul);
1806 if (!BOp)
1807 continue;
1808
1809 if (Value *V = RemoveFactorFromExpression(V: Ops[i].Op, Factor: MaxOccVal,
1810 DL: I->getDebugLoc())) {
1811 // The factorized operand may occur several times. Convert them all in
1812 // one fell swoop.
1813 for (unsigned j = Ops.size(); j != i;) {
1814 --j;
1815 if (Ops[j].Op == Ops[i].Op) {
1816 NewMulOps.push_back(Elt: V);
1817 Ops.erase(CI: Ops.begin()+j);
1818 }
1819 }
1820 --i;
1821 }
1822 }
1823
1824 // No need for extra uses anymore.
1825 DummyInst->deleteValue();
1826
1827 unsigned NumAddedValues = NewMulOps.size();
1828 Value *V = EmitAddTreeOfValues(I, Ops&: NewMulOps);
1829
1830 // Now that we have inserted the add tree, optimize it. This allows us to
1831 // handle cases that require multiple factoring steps, such as this:
1832 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C))
1833 assert(NumAddedValues > 1 && "Each occurrence should contribute a value");
1834 (void)NumAddedValues;
1835 if (Instruction *VI = dyn_cast<Instruction>(Val: V))
1836 RedoInsts.insert(X: VI);
1837
1838 // Create the multiply.
1839 Instruction *V2 = CreateMul(S1: V, S2: MaxOccVal, Name: "reass.mul", InsertBefore: I->getIterator(), FlagsOp: I);
1840 V2->setDebugLoc(I->getDebugLoc());
1841
1842 // Rerun associate on the multiply in case the inner expression turned into
1843 // a multiply. We want to make sure that we keep things in canonical form.
1844 RedoInsts.insert(X: V2);
1845
1846 // If every add operand included the factor (e.g. "A*B + A*C"), then the
1847 // entire result expression is just the multiply "A*(B+C)".
1848 if (Ops.empty())
1849 return V2;
1850
1851 // Otherwise, we had some input that didn't have the factor, such as
1852 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of
1853 // things being added by this operation.
1854 Ops.insert(I: Ops.begin(), Elt: ValueEntry(getRank(V: V2), V2));
1855 }
1856
1857 return nullptr;
1858}
1859
1860/// Build up a vector of value/power pairs factoring a product.
1861///
1862/// Given a series of multiplication operands, build a vector of factors and
1863/// the powers each is raised to when forming the final product. Sort them in
1864/// the order of descending power.
1865///
1866/// (x*x) -> [(x, 2)]
1867/// ((x*x)*x) -> [(x, 3)]
1868/// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)]
1869///
1870/// \returns Whether any factors have a power greater than one.
1871static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops,
1872 SmallVectorImpl<Factor> &Factors) {
1873 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this.
1874 // Compute the sum of powers of simplifiable factors.
1875 unsigned FactorPowerSum = 0;
1876 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) {
1877 Value *Op = Ops[Idx-1].Op;
1878
1879 // Count the number of occurrences of this value.
1880 unsigned Count = 1;
1881 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx)
1882 ++Count;
1883 // Track for simplification all factors which occur 2 or more times.
1884 if (Count > 1)
1885 FactorPowerSum += Count;
1886 }
1887
1888 // We can only simplify factors if the sum of the powers of our simplifiable
1889 // factors is 4 or higher. When that is the case, we will *always* have
1890 // a simplification. This is an important invariant to prevent cyclicly
1891 // trying to simplify already minimal formations.
1892 if (FactorPowerSum < 4)
1893 return false;
1894
1895 // Now gather the simplifiable factors, removing them from Ops.
1896 FactorPowerSum = 0;
1897 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) {
1898 Value *Op = Ops[Idx-1].Op;
1899
1900 // Count the number of occurrences of this value.
1901 unsigned Count = 1;
1902 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx)
1903 ++Count;
1904 if (Count == 1)
1905 continue;
1906 // Move an even number of occurrences to Factors.
1907 Count &= ~1U;
1908 Idx -= Count;
1909 FactorPowerSum += Count;
1910 Factors.push_back(Elt: Factor(Op, Count));
1911 Ops.erase(CS: Ops.begin()+Idx, CE: Ops.begin()+Idx+Count);
1912 }
1913
1914 // None of the adjustments above should have reduced the sum of factor powers
1915 // below our mininum of '4'.
1916 assert(FactorPowerSum >= 4);
1917
1918 llvm::stable_sort(Range&: Factors, C: [](const Factor &LHS, const Factor &RHS) {
1919 return LHS.Power > RHS.Power;
1920 });
1921 return true;
1922}
1923
1924/// Build a tree of multiplies, computing the product of Ops.
1925static Value *buildMultiplyTree(IRBuilderBase &Builder,
1926 SmallVectorImpl<Value*> &Ops) {
1927 if (Ops.size() == 1)
1928 return Ops.back();
1929
1930 Value *LHS = Ops.pop_back_val();
1931 do {
1932 if (LHS->getType()->isIntOrIntVectorTy())
1933 LHS = Builder.CreateMul(LHS, RHS: Ops.pop_back_val());
1934 else
1935 LHS = Builder.CreateFMul(L: LHS, R: Ops.pop_back_val());
1936 } while (!Ops.empty());
1937
1938 return LHS;
1939}
1940
1941/// Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*...
1942///
1943/// Given a vector of values raised to various powers, where no two values are
1944/// equal and the powers are sorted in decreasing order, compute the minimal
1945/// DAG of multiplies to compute the final product, and return that product
1946/// value.
1947Value *
1948ReassociatePass::buildMinimalMultiplyDAG(IRBuilderBase &Builder,
1949 SmallVectorImpl<Factor> &Factors) {
1950 assert(Factors[0].Power);
1951 SmallVector<Value *, 4> OuterProduct;
1952 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size();
1953 Idx < Size && Factors[Idx].Power > 0; ++Idx) {
1954 if (Factors[Idx].Power != Factors[LastIdx].Power) {
1955 LastIdx = Idx;
1956 continue;
1957 }
1958
1959 // We want to multiply across all the factors with the same power so that
1960 // we can raise them to that power as a single entity. Build a mini tree
1961 // for that.
1962 SmallVector<Value *, 4> InnerProduct;
1963 InnerProduct.push_back(Elt: Factors[LastIdx].Base);
1964 do {
1965 InnerProduct.push_back(Elt: Factors[Idx].Base);
1966 ++Idx;
1967 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power);
1968
1969 // Reset the base value of the first factor to the new expression tree.
1970 // We'll remove all the factors with the same power in a second pass.
1971 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, Ops&: InnerProduct);
1972 if (Instruction *MI = dyn_cast<Instruction>(Val: M))
1973 RedoInsts.insert(X: MI);
1974
1975 LastIdx = Idx;
1976 }
1977 // Unique factors with equal powers -- we've folded them into the first one's
1978 // base.
1979 Factors.erase(CS: llvm::unique(R&: Factors,
1980 P: [](const Factor &LHS, const Factor &RHS) {
1981 return LHS.Power == RHS.Power;
1982 }),
1983 CE: Factors.end());
1984
1985 // Iteratively collect the base of each factor with an add power into the
1986 // outer product, and halve each power in preparation for squaring the
1987 // expression.
1988 for (Factor &F : Factors) {
1989 if (F.Power & 1)
1990 OuterProduct.push_back(Elt: F.Base);
1991 F.Power >>= 1;
1992 }
1993 if (Factors[0].Power) {
1994 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors);
1995 OuterProduct.push_back(Elt: SquareRoot);
1996 OuterProduct.push_back(Elt: SquareRoot);
1997 }
1998 if (OuterProduct.size() == 1)
1999 return OuterProduct.front();
2000
2001 Value *V = buildMultiplyTree(Builder, Ops&: OuterProduct);
2002 return V;
2003}
2004
2005Value *ReassociatePass::OptimizeMul(BinaryOperator *I,
2006 SmallVectorImpl<ValueEntry> &Ops) {
2007 // We can only optimize the multiplies when there is a chain of more than
2008 // three, such that a balanced tree might require fewer total multiplies.
2009 if (Ops.size() < 4)
2010 return nullptr;
2011
2012 // Try to turn linear trees of multiplies without other uses of the
2013 // intermediate stages into minimal multiply DAGs with perfect sub-expression
2014 // re-use.
2015 SmallVector<Factor, 4> Factors;
2016 if (!collectMultiplyFactors(Ops, Factors))
2017 return nullptr; // All distinct factors, so nothing left for us to do.
2018
2019 IRBuilder<> Builder(I);
2020 // The reassociate transformation for FP operations is performed only
2021 // if unsafe algebra is permitted by FastMathFlags. Propagate those flags
2022 // to the newly generated operations.
2023 if (auto FPI = dyn_cast<FPMathOperator>(Val: I))
2024 Builder.setFastMathFlags(FPI->getFastMathFlags());
2025
2026 Value *V = buildMinimalMultiplyDAG(Builder, Factors);
2027 if (Ops.empty())
2028 return V;
2029
2030 ValueEntry NewEntry = ValueEntry(getRank(V), V);
2031 Ops.insert(I: llvm::lower_bound(Range&: Ops, Value&: NewEntry), Elt: NewEntry);
2032 return nullptr;
2033}
2034
2035Value *ReassociatePass::OptimizeExpression(BinaryOperator *I,
2036 SmallVectorImpl<ValueEntry> &Ops) {
2037 // Now that we have the linearized expression tree, try to optimize it.
2038 // Start by folding any constants that we found.
2039 const DataLayout &DL = I->getDataLayout();
2040 Constant *Cst = nullptr;
2041 unsigned Opcode = I->getOpcode();
2042 while (!Ops.empty()) {
2043 if (auto *C = dyn_cast<Constant>(Val: Ops.back().Op)) {
2044 if (!Cst) {
2045 Ops.pop_back();
2046 Cst = C;
2047 continue;
2048 }
2049 if (Constant *Res = ConstantFoldBinaryOpOperands(Opcode, LHS: C, RHS: Cst, DL)) {
2050 Ops.pop_back();
2051 Cst = Res;
2052 continue;
2053 }
2054 }
2055 break;
2056 }
2057 // If there was nothing but constants then we are done.
2058 if (Ops.empty())
2059 return Cst;
2060
2061 // Put the combined constant back at the end of the operand list, except if
2062 // there is no point. For example, an add of 0 gets dropped here, while a
2063 // multiplication by zero turns the whole expression into zero.
2064 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, Ty: I->getType())) {
2065 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, Ty: I->getType()))
2066 return Cst;
2067 Ops.push_back(Elt: ValueEntry(0, Cst));
2068 }
2069
2070 if (Ops.size() == 1) return Ops[0].Op;
2071
2072 // Handle destructive annihilation due to identities between elements in the
2073 // argument list here.
2074 unsigned NumOps = Ops.size();
2075 switch (Opcode) {
2076 default: break;
2077 case Instruction::And:
2078 case Instruction::Or:
2079 if (Value *Result = OptimizeAndOrXor(Opcode, Ops))
2080 return Result;
2081 break;
2082
2083 case Instruction::Xor:
2084 if (Value *Result = OptimizeXor(I, Ops))
2085 return Result;
2086 break;
2087
2088 case Instruction::Add:
2089 case Instruction::FAdd:
2090 if (Value *Result = OptimizeAdd(I, Ops))
2091 return Result;
2092 break;
2093
2094 case Instruction::Mul:
2095 case Instruction::FMul:
2096 if (Value *Result = OptimizeMul(I, Ops))
2097 return Result;
2098 break;
2099 }
2100
2101 if (Ops.size() != NumOps)
2102 return OptimizeExpression(I, Ops);
2103 return nullptr;
2104}
2105
2106// Remove dead instructions and if any operands are trivially dead add them to
2107// Insts so they will be removed as well.
2108void ReassociatePass::RecursivelyEraseDeadInsts(Instruction *I,
2109 OrderedSet &Insts) {
2110 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
2111 SmallVector<Value *, 4> Ops(I->operands());
2112 ValueRankMap.erase(Val: I);
2113 Insts.remove(X: I);
2114 RedoInsts.remove(X: I);
2115 if (UA)
2116 UA->forgetValue(V: I);
2117 llvm::salvageDebugInfo(I&: *I);
2118 I->eraseFromParent();
2119 for (auto *Op : Ops)
2120 if (Instruction *OpInst = dyn_cast<Instruction>(Val: Op))
2121 if (OpInst->use_empty())
2122 Insts.insert(X: OpInst);
2123}
2124
2125/// Zap the given instruction, adding interesting operands to the work list.
2126void ReassociatePass::EraseInst(Instruction *I) {
2127 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!");
2128 LLVM_DEBUG(dbgs() << "Erasing dead inst: "; I->dump());
2129
2130 SmallVector<Value *, 8> Ops(I->operands());
2131 // Erase the dead instruction.
2132 ValueRankMap.erase(Val: I);
2133 RedoInsts.remove(X: I);
2134 if (UA)
2135 UA->forgetValue(V: I);
2136 llvm::salvageDebugInfo(I&: *I);
2137 I->eraseFromParent();
2138 // Optimize its operands.
2139 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes.
2140 for (Value *V : Ops)
2141 if (Instruction *Op = dyn_cast<Instruction>(Val: V)) {
2142 // If this is a node in an expression tree, climb to the expression root
2143 // and add that since that's where optimization actually happens.
2144 unsigned Opcode = Op->getOpcode();
2145 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode &&
2146 Visited.insert(Ptr: Op).second)
2147 Op = Op->user_back();
2148
2149 // The instruction we're going to push may be coming from a
2150 // dead block, and Reassociate skips the processing of unreachable
2151 // blocks because it's a waste of time and also because it can
2152 // lead to infinite loop due to LLVM's non-standard definition
2153 // of dominance.
2154 if (ValueRankMap.contains(Val: Op))
2155 RedoInsts.insert(X: Op);
2156 }
2157
2158 MadeChange = true;
2159}
2160
2161/// Recursively analyze an expression to build a list of instructions that have
2162/// negative floating-point constant operands. The caller can then transform
2163/// the list to create positive constants for better reassociation and CSE.
2164static void getNegatibleInsts(Value *V,
2165 SmallVectorImpl<Instruction *> &Candidates) {
2166 // Handle only one-use instructions. Combining negations does not justify
2167 // replicating instructions.
2168 Instruction *I;
2169 if (!match(V, P: m_OneUse(SubPattern: m_Instruction(I))))
2170 return;
2171
2172 // Handle expressions of multiplications and divisions.
2173 // TODO: This could look through floating-point casts.
2174 const APFloat *C;
2175 switch (I->getOpcode()) {
2176 case Instruction::FMul:
2177 // Not expecting non-canonical code here. Bail out and wait.
2178 if (match(V: I->getOperand(i: 0), P: m_Constant()))
2179 break;
2180
2181 if (match(V: I->getOperand(i: 1), P: m_APFloat(Res&: C)) && C->isNegative()) {
2182 Candidates.push_back(Elt: I);
2183 LLVM_DEBUG(dbgs() << "FMul with negative constant: " << *I << '\n');
2184 }
2185 getNegatibleInsts(V: I->getOperand(i: 0), Candidates);
2186 getNegatibleInsts(V: I->getOperand(i: 1), Candidates);
2187 break;
2188 case Instruction::FDiv:
2189 // Not expecting non-canonical code here. Bail out and wait.
2190 if (match(V: I->getOperand(i: 0), P: m_Constant()) &&
2191 match(V: I->getOperand(i: 1), P: m_Constant()))
2192 break;
2193
2194 if ((match(V: I->getOperand(i: 0), P: m_APFloat(Res&: C)) && C->isNegative()) ||
2195 (match(V: I->getOperand(i: 1), P: m_APFloat(Res&: C)) && C->isNegative())) {
2196 Candidates.push_back(Elt: I);
2197 LLVM_DEBUG(dbgs() << "FDiv with negative constant: " << *I << '\n');
2198 }
2199 getNegatibleInsts(V: I->getOperand(i: 0), Candidates);
2200 getNegatibleInsts(V: I->getOperand(i: 1), Candidates);
2201 break;
2202 default:
2203 break;
2204 }
2205}
2206
2207/// Given an fadd/fsub with an operand that is a one-use instruction
2208/// (the fadd/fsub), try to change negative floating-point constants into
2209/// positive constants to increase potential for reassociation and CSE.
2210Instruction *ReassociatePass::canonicalizeNegFPConstantsForOp(Instruction *I,
2211 Instruction *Op,
2212 Value *OtherOp) {
2213 assert((I->getOpcode() == Instruction::FAdd ||
2214 I->getOpcode() == Instruction::FSub) && "Expected fadd/fsub");
2215
2216 // Collect instructions with negative FP constants from the subtree that ends
2217 // in Op.
2218 SmallVector<Instruction *, 4> Candidates;
2219 getNegatibleInsts(V: Op, Candidates);
2220 if (Candidates.empty())
2221 return nullptr;
2222
2223 // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the
2224 // resulting subtract will be broken up later. This can get us into an
2225 // infinite loop during reassociation.
2226 bool IsFSub = I->getOpcode() == Instruction::FSub;
2227 bool NeedsSubtract = !IsFSub && Candidates.size() % 2 == 1;
2228 if (NeedsSubtract && ShouldBreakUpSubtract(Sub: I))
2229 return nullptr;
2230
2231 for (Instruction *Negatible : Candidates) {
2232 const APFloat *C;
2233 if (match(V: Negatible->getOperand(i: 0), P: m_APFloat(Res&: C))) {
2234 assert(!match(Negatible->getOperand(1), m_Constant()) &&
2235 "Expecting only 1 constant operand");
2236 assert(C->isNegative() && "Expected negative FP constant");
2237 Negatible->setOperand(i: 0, Val: ConstantFP::get(Ty: Negatible->getType(), V: abs(X: *C)));
2238 MadeChange = true;
2239 }
2240 if (match(V: Negatible->getOperand(i: 1), P: m_APFloat(Res&: C))) {
2241 assert(!match(Negatible->getOperand(0), m_Constant()) &&
2242 "Expecting only 1 constant operand");
2243 assert(C->isNegative() && "Expected negative FP constant");
2244 Negatible->setOperand(i: 1, Val: ConstantFP::get(Ty: Negatible->getType(), V: abs(X: *C)));
2245 MadeChange = true;
2246 }
2247 }
2248 assert(MadeChange == true && "Negative constant candidate was not changed");
2249
2250 // Negations cancelled out.
2251 if (Candidates.size() % 2 == 0)
2252 return I;
2253
2254 // Negate the final operand in the expression by flipping the opcode of this
2255 // fadd/fsub.
2256 assert(Candidates.size() % 2 == 1 && "Expected odd number");
2257 IRBuilder<> Builder(I);
2258 Value *NewInst = IsFSub ? Builder.CreateFAddFMF(L: OtherOp, R: Op, FMFSource: I)
2259 : Builder.CreateFSubFMF(L: OtherOp, R: Op, FMFSource: I);
2260 I->replaceAllUsesWith(V: NewInst);
2261 RedoInsts.insert(X: I);
2262 return dyn_cast<Instruction>(Val: NewInst);
2263}
2264
2265/// Canonicalize expressions that contain a negative floating-point constant
2266/// of the following form:
2267/// OtherOp + (subtree) -> OtherOp {+/-} (canonical subtree)
2268/// (subtree) + OtherOp -> OtherOp {+/-} (canonical subtree)
2269/// OtherOp - (subtree) -> OtherOp {+/-} (canonical subtree)
2270///
2271/// The fadd/fsub opcode may be switched to allow folding a negation into the
2272/// input instruction.
2273Instruction *ReassociatePass::canonicalizeNegFPConstants(Instruction *I) {
2274 LLVM_DEBUG(dbgs() << "Combine negations for: " << *I << '\n');
2275 Value *X;
2276 Instruction *Op;
2277 if (match(V: I, P: m_FAdd(L: m_Value(V&: X), R: m_OneUse(SubPattern: m_Instruction(I&: Op)))))
2278 if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, OtherOp: X))
2279 I = R;
2280 if (match(V: I, P: m_FAdd(L: m_OneUse(SubPattern: m_Instruction(I&: Op)), R: m_Value(V&: X))))
2281 if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, OtherOp: X))
2282 I = R;
2283 if (match(V: I, P: m_FSub(L: m_Value(V&: X), R: m_OneUse(SubPattern: m_Instruction(I&: Op)))))
2284 if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, OtherOp: X))
2285 I = R;
2286 return I;
2287}
2288
2289/// Inspect and optimize the given instruction. Note that erasing
2290/// instructions is not allowed.
2291void ReassociatePass::OptimizeInst(Instruction *I) {
2292 // Only consider operations that we understand.
2293 if (!isa<UnaryOperator>(Val: I) && !isa<BinaryOperator>(Val: I))
2294 return;
2295
2296 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(Val: I->getOperand(i: 1)))
2297 // If an operand of this shift is a reassociable multiply, or if the shift
2298 // is used by a reassociable multiply or add, turn into a multiply.
2299 if (isReassociableOp(V: I->getOperand(i: 0), Opcode: Instruction::Mul) ||
2300 (I->hasOneUse() &&
2301 (isReassociableOp(V: I->user_back(), Opcode: Instruction::Mul) ||
2302 isReassociableOp(V: I->user_back(), Opcode: Instruction::Add)))) {
2303 Instruction *NI = ConvertShiftToMul(Shl: I);
2304 RedoInsts.insert(X: I);
2305 MadeChange = true;
2306 I = NI;
2307 }
2308
2309 // Commute binary operators, to canonicalize the order of their operands.
2310 // This can potentially expose more CSE opportunities, and makes writing other
2311 // transformations simpler.
2312 if (I->isCommutative())
2313 canonicalizeOperands(I);
2314
2315 // Canonicalize negative constants out of expressions.
2316 if (Instruction *Res = canonicalizeNegFPConstants(I))
2317 I = Res;
2318
2319 // Don't optimize floating-point instructions unless they have the
2320 // appropriate FastMathFlags for reassociation enabled.
2321 if (isa<FPMathOperator>(Val: I) && !hasFPAssociativeFlags(I))
2322 return;
2323
2324 // Do not reassociate boolean (i1/vXi1) expressions. We want to preserve the
2325 // original order of evaluation for short-circuited comparisons that
2326 // SimplifyCFG has folded to AND/OR expressions. If the expression
2327 // is not further optimized, it is likely to be transformed back to a
2328 // short-circuited form for code gen, and the source order may have been
2329 // optimized for the most likely conditions. For vector boolean expressions,
2330 // we should be optimizing for ILP and not serializing the logical operations.
2331 if (I->getType()->isIntOrIntVectorTy(BitWidth: 1))
2332 return;
2333
2334 // If this is a bitwise or instruction of operands
2335 // with no common bits set, convert it to X+Y.
2336 if (I->getOpcode() == Instruction::Or &&
2337 shouldConvertOrWithNoCommonBitsToAdd(Or: I) && !isLoadCombineCandidate(Or: I) &&
2338 (cast<PossiblyDisjointInst>(Val: I)->isDisjoint() ||
2339 haveNoCommonBitsSet(LHSCache: I->getOperand(i: 0), RHSCache: I->getOperand(i: 1),
2340 SQ: SimplifyQuery(I->getDataLayout(),
2341 /*DT=*/nullptr, /*AC=*/nullptr, I)))) {
2342 Instruction *NI = convertOrWithNoCommonBitsToAdd(Or: I);
2343 RedoInsts.insert(X: I);
2344 MadeChange = true;
2345 I = NI;
2346 }
2347
2348 if (I->getOpcode() == Instruction::Mul && ShouldBreakUpDistribution(Mul: I)) {
2349 Instruction *MulUser = cast<Instruction>(Val: I->user_back());
2350 Instruction *NI = BreakUpDistribute(Mul: I, ToRedo&: RedoInsts);
2351 RedoInsts.insert(X: I);
2352 RedoInsts.insert(X: MulUser);
2353 MadeChange = true;
2354 I = NI;
2355 }
2356
2357 // If this is a subtract instruction which is not already in negate form,
2358 // see if we can convert it to X+-Y.
2359 if (I->getOpcode() == Instruction::Sub) {
2360 if (ShouldBreakUpSubtract(Sub: I)) {
2361 Instruction *NI = BreakUpSubtract(Sub: I, ToRedo&: RedoInsts);
2362 RedoInsts.insert(X: I);
2363 MadeChange = true;
2364 I = NI;
2365 } else if (match(V: I, P: m_Neg(V: m_Value()))) {
2366 // Otherwise, this is a negation. See if the operand is a multiply tree
2367 // and if this is not an inner node of a multiply tree.
2368 if (isReassociableOp(V: I->getOperand(i: 1), Opcode: Instruction::Mul) &&
2369 (!I->hasOneUse() ||
2370 !isReassociableOp(V: I->user_back(), Opcode: Instruction::Mul))) {
2371 Instruction *NI = LowerNegateToMultiply(Neg: I);
2372 // If the negate was simplified, revisit the users to see if we can
2373 // reassociate further.
2374 for (User *U : NI->users()) {
2375 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(Val: U))
2376 RedoInsts.insert(X: Tmp);
2377 }
2378 RedoInsts.insert(X: I);
2379 MadeChange = true;
2380 I = NI;
2381 }
2382 }
2383 } else if (I->getOpcode() == Instruction::FNeg ||
2384 I->getOpcode() == Instruction::FSub) {
2385 if (ShouldBreakUpSubtract(Sub: I)) {
2386 Instruction *NI = BreakUpSubtract(Sub: I, ToRedo&: RedoInsts);
2387 RedoInsts.insert(X: I);
2388 MadeChange = true;
2389 I = NI;
2390 } else if (match(V: I, P: m_FNeg(X: m_Value()))) {
2391 // Otherwise, this is a negation. See if the operand is a multiply tree
2392 // and if this is not an inner node of a multiply tree.
2393 Value *Op = isa<BinaryOperator>(Val: I) ? I->getOperand(i: 1) :
2394 I->getOperand(i: 0);
2395 if (isReassociableOp(V: Op, Opcode: Instruction::FMul) &&
2396 (!I->hasOneUse() ||
2397 !isReassociableOp(V: I->user_back(), Opcode: Instruction::FMul))) {
2398 // If the negate was simplified, revisit the users to see if we can
2399 // reassociate further.
2400 Instruction *NI = LowerNegateToMultiply(Neg: I);
2401 for (User *U : NI->users()) {
2402 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(Val: U))
2403 RedoInsts.insert(X: Tmp);
2404 }
2405 RedoInsts.insert(X: I);
2406 MadeChange = true;
2407 I = NI;
2408 }
2409 }
2410 }
2411
2412 // If this instruction is an associative binary operator, process it.
2413 if (!I->isAssociative()) return;
2414 BinaryOperator *BO = cast<BinaryOperator>(Val: I);
2415
2416 // If this is an interior node of a reassociable tree, ignore it until we
2417 // get to the root of the tree, to avoid N^2 analysis.
2418 unsigned Opcode = BO->getOpcode();
2419 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) {
2420 // During the initial run we will get to the root of the tree.
2421 // But if we get here while we are redoing instructions, there is no
2422 // guarantee that the root will be visited. So Redo later
2423 if (BO->user_back() != BO &&
2424 BO->getParent() == BO->user_back()->getParent())
2425 RedoInsts.insert(X: BO->user_back());
2426 return;
2427 }
2428
2429 // If this is an add tree that is used by a sub instruction, ignore it
2430 // until we process the subtract.
2431 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add &&
2432 cast<Instruction>(Val: BO->user_back())->getOpcode() == Instruction::Sub)
2433 return;
2434 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
2435 cast<Instruction>(Val: BO->user_back())->getOpcode() == Instruction::FSub)
2436 return;
2437
2438 ReassociateExpression(I: BO);
2439}
2440
2441void ReassociatePass::ReassociateExpression(BinaryOperator *I) {
2442 // First, walk the expression tree, linearizing the tree, collecting the
2443 // operand information.
2444 SmallVector<RepeatedValue, 8> Tree;
2445 OverflowTracking Flags;
2446 MadeChange |= LinearizeExprTree(I, Ops&: Tree, ToRedo&: RedoInsts, Flags);
2447 SmallVector<ValueEntry, 8> Ops;
2448 Ops.reserve(N: Tree.size());
2449 for (const RepeatedValue &E : Tree)
2450 Ops.append(NumInputs: E.second, Elt: ValueEntry(getRank(V: E.first), E.first));
2451
2452 LLVM_DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n');
2453
2454 // Boost the rank of divergent operands so they sort towards the root of the
2455 // expression tree, clustering uniform operands together at the leaves. On
2456 // targets without divergence UniformityInfo is empty and this is a no-op.
2457 //
2458 // Example: (uniform1 + divergent) + uniform2
2459 // -> (uniform1 + uniform2) + divergent
2460 if (UA && Ops.size() > 2) {
2461 constexpr unsigned DivergentRankOffset = 1U << 28;
2462 BasicBlock *ParentBB = I->getParent();
2463 for (ValueEntry &Entry : Ops) {
2464 if (isa<Constant>(Val: Entry.Op))
2465 continue;
2466 bool Divergent = false;
2467 for (const Use &U : Entry.Op->uses()) {
2468 Instruction *Usr = dyn_cast<Instruction>(Val: U.getUser());
2469 if (Usr && Usr->getParent() == ParentBB) {
2470 Divergent = UA->isDivergentAtUse(U);
2471 break;
2472 }
2473 }
2474 if (Divergent)
2475 Entry.Rank += DivergentRankOffset;
2476 }
2477 }
2478
2479 // Now that we have linearized the tree to a list and have gathered all of
2480 // the operands and their ranks, sort the operands by their rank. Use a
2481 // stable_sort so that values with equal ranks will have their relative
2482 // positions maintained (and so the compiler is deterministic). Note that
2483 // this sorts so that the highest ranking values end up at the beginning of
2484 // the vector.
2485 llvm::stable_sort(Range&: Ops);
2486
2487 // Now that we have the expression tree in a convenient
2488 // sorted form, optimize it globally if possible.
2489 if (Value *V = OptimizeExpression(I, Ops)) {
2490 if (V == I)
2491 // Self-referential expression in unreachable code.
2492 return;
2493 // This expression tree simplified to something that isn't a tree,
2494 // eliminate it.
2495 LLVM_DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n');
2496 I->replaceAllUsesWith(V);
2497 if (Instruction *VI = dyn_cast<Instruction>(Val: V))
2498 if (I->getDebugLoc())
2499 VI->setDebugLoc(I->getDebugLoc());
2500 RedoInsts.insert(X: I);
2501 ++NumAnnihil;
2502 return;
2503 }
2504
2505 // We want to sink immediates as deeply as possible except in the case where
2506 // this is a multiply tree used only by an add, and the immediate is a -1.
2507 // In this case we reassociate to put the negation on the outside so that we
2508 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y
2509 if (I->hasOneUse()) {
2510 if (I->getOpcode() == Instruction::Mul &&
2511 cast<Instruction>(Val: I->user_back())->getOpcode() == Instruction::Add &&
2512 isa<ConstantInt>(Val: Ops.back().Op) &&
2513 cast<ConstantInt>(Val: Ops.back().Op)->isMinusOne()) {
2514 ValueEntry Tmp = Ops.pop_back_val();
2515 Ops.insert(I: Ops.begin(), Elt: Tmp);
2516 } else if (I->getOpcode() == Instruction::FMul &&
2517 cast<Instruction>(Val: I->user_back())->getOpcode() ==
2518 Instruction::FAdd &&
2519 isa<ConstantFP>(Val: Ops.back().Op) &&
2520 cast<ConstantFP>(Val: Ops.back().Op)->isMinusOne()) {
2521 ValueEntry Tmp = Ops.pop_back_val();
2522 Ops.insert(I: Ops.begin(), Elt: Tmp);
2523 }
2524 }
2525
2526 LLVM_DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n');
2527
2528 if (Ops.size() == 1) {
2529 if (Ops[0].Op == I)
2530 // Self-referential expression in unreachable code.
2531 return;
2532
2533 // This expression tree simplified to something that isn't a tree,
2534 // eliminate it.
2535 I->replaceAllUsesWith(V: Ops[0].Op);
2536 if (Instruction *OI = dyn_cast<Instruction>(Val: Ops[0].Op))
2537 OI->setDebugLoc(I->getDebugLoc());
2538 RedoInsts.insert(X: I);
2539 return;
2540 }
2541
2542 if (Ops.size() > 2 && Ops.size() <= GlobalReassociateLimit) {
2543 // Find the pair with the highest count in the pairmap and move it to the
2544 // back of the list so that it can later be CSE'd.
2545 // example:
2546 // a*b*c*d*e
2547 // if c*e is the most "popular" pair, we can express this as
2548 // (((c*e)*d)*b)*a
2549 unsigned Max = 1;
2550 unsigned BestRank = 0;
2551 std::pair<unsigned, unsigned> BestPair;
2552 unsigned Idx = I->getOpcode() - Instruction::BinaryOpsBegin;
2553 unsigned LimitIdx = 0;
2554 // With the CSE-driven heuristic, we are about to slap two values at the
2555 // beginning of the expression whereas they could live very late in the CFG.
2556 // When using the CSE-local heuristic we avoid creating dependences from
2557 // completely unrelated part of the CFG by limiting the expression
2558 // reordering on the values that live in the first seen basic block.
2559 // The main idea is that we want to avoid forming expressions that would
2560 // become loop dependent.
2561 if (UseCSELocalOpt) {
2562 const BasicBlock *FirstSeenBB = nullptr;
2563 int StartIdx = Ops.size() - 1;
2564 // Skip the first value of the expression since we need at least two
2565 // values to materialize an expression. I.e., even if this value is
2566 // anchored in a different basic block, the actual first sub expression
2567 // will be anchored on the second value.
2568 for (int i = StartIdx - 1; i != -1; --i) {
2569 const Value *Val = Ops[i].Op;
2570 const auto *CurrLeafInstr = dyn_cast<Instruction>(Val);
2571 const BasicBlock *SeenBB = nullptr;
2572 if (!CurrLeafInstr) {
2573 // The value is free of any CFG dependencies.
2574 // Do as if it lives in the entry block.
2575 //
2576 // We do this to make sure all the values falling on this path are
2577 // seen through the same anchor point. The rationale is these values
2578 // can be combined together to from a sub expression free of any CFG
2579 // dependencies so we want them to stay together.
2580 // We could be cleverer and postpone the anchor down to the first
2581 // anchored value, but that's likely complicated to get right.
2582 // E.g., we wouldn't want to do that if that means being stuck in a
2583 // loop.
2584 //
2585 // For instance, we wouldn't want to change:
2586 // res = arg1 op arg2 op arg3 op ... op loop_val1 op loop_val2 ...
2587 // into
2588 // res = loop_val1 op arg1 op arg2 op arg3 op ... op loop_val2 ...
2589 // Because all the sub expressions with arg2..N would be stuck between
2590 // two loop dependent values.
2591 SeenBB = &I->getParent()->getParent()->getEntryBlock();
2592 } else {
2593 SeenBB = CurrLeafInstr->getParent();
2594 }
2595
2596 if (!FirstSeenBB) {
2597 FirstSeenBB = SeenBB;
2598 continue;
2599 }
2600 if (FirstSeenBB != SeenBB) {
2601 // ith value is in a different basic block.
2602 // Rewind the index once to point to the last value on the same basic
2603 // block.
2604 LimitIdx = i + 1;
2605 LLVM_DEBUG(dbgs() << "CSE reordering: Consider values between ["
2606 << LimitIdx << ", " << StartIdx << "]\n");
2607 break;
2608 }
2609 }
2610 }
2611 for (unsigned i = Ops.size() - 1; i > LimitIdx; --i) {
2612 // We must use int type to go below zero when LimitIdx is 0.
2613 for (int j = i - 1; j >= (int)LimitIdx; --j) {
2614 unsigned Score = 0;
2615 Value *Op0 = Ops[i].Op;
2616 Value *Op1 = Ops[j].Op;
2617 if (std::less<Value *>()(Op1, Op0))
2618 std::swap(a&: Op0, b&: Op1);
2619 auto it = PairMap[Idx].find(Val: {Op0, Op1});
2620 if (it != PairMap[Idx].end()) {
2621 // Functions like BreakUpSubtract() can erase the Values we're using
2622 // as keys and create new Values after we built the PairMap. There's a
2623 // small chance that the new nodes can have the same address as
2624 // something already in the table. We shouldn't accumulate the stored
2625 // score in that case as it refers to the wrong Value.
2626 if (it->second.isValid())
2627 Score += it->second.Score;
2628 }
2629
2630 unsigned MaxRank = std::max(a: Ops[i].Rank, b: Ops[j].Rank);
2631
2632 // By construction, the operands are sorted in reverse order of their
2633 // topological order.
2634 // So we tend to form (sub) expressions with values that are close to
2635 // each other.
2636 //
2637 // Now to expose more CSE opportunities we want to expose the pair of
2638 // operands that occur the most (as statically computed in
2639 // BuildPairMap.) as the first sub-expression.
2640 //
2641 // If two pairs occur as many times, we pick the one with the
2642 // lowest rank, meaning the one with both operands appearing first in
2643 // the topological order.
2644 if (Score > Max || (Score == Max && MaxRank < BestRank)) {
2645 BestPair = {j, i};
2646 Max = Score;
2647 BestRank = MaxRank;
2648 }
2649 }
2650 }
2651 if (Max > 1) {
2652 auto Op0 = Ops[BestPair.first];
2653 auto Op1 = Ops[BestPair.second];
2654 Ops.erase(CI: &Ops[BestPair.second]);
2655 Ops.erase(CI: &Ops[BestPair.first]);
2656 Ops.push_back(Elt: Op0);
2657 Ops.push_back(Elt: Op1);
2658 }
2659 }
2660 LLVM_DEBUG(dbgs() << "RAOut after CSE reorder:\t"; PrintOps(I, Ops);
2661 dbgs() << '\n');
2662 // Now that we ordered and optimized the expressions, splat them back into
2663 // the expression tree, removing any unneeded nodes.
2664 RewriteExprTree(I, Ops, Flags);
2665}
2666
2667void
2668ReassociatePass::BuildPairMap(ReversePostOrderTraversal<Function *> &RPOT) {
2669 // Make a "pairmap" of how often each operand pair occurs.
2670 for (BasicBlock *BI : RPOT) {
2671 for (Instruction &I : *BI) {
2672 if (!I.isAssociative() || !I.isBinaryOp())
2673 continue;
2674
2675 // Ignore nodes that aren't at the root of trees.
2676 if (I.hasOneUse() && I.user_back()->getOpcode() == I.getOpcode())
2677 continue;
2678
2679 // Collect all operands in a single reassociable expression.
2680 // Since Reassociate has already been run once, we can assume things
2681 // are already canonical according to Reassociation's regime.
2682 SmallVector<Value *, 8> Worklist = { I.getOperand(i: 0), I.getOperand(i: 1) };
2683 SmallVector<Value *, 8> Ops;
2684 while (!Worklist.empty() && Ops.size() <= GlobalReassociateLimit) {
2685 Value *Op = Worklist.pop_back_val();
2686 Instruction *OpI = dyn_cast<Instruction>(Val: Op);
2687 if (!OpI || OpI->getOpcode() != I.getOpcode() || !OpI->hasOneUse()) {
2688 Ops.push_back(Elt: Op);
2689 continue;
2690 }
2691 // Be paranoid about self-referencing expressions in unreachable code.
2692 if (OpI->getOperand(i: 0) != OpI)
2693 Worklist.push_back(Elt: OpI->getOperand(i: 0));
2694 if (OpI->getOperand(i: 1) != OpI)
2695 Worklist.push_back(Elt: OpI->getOperand(i: 1));
2696 }
2697 // Skip extremely long expressions.
2698 if (Ops.size() > GlobalReassociateLimit)
2699 continue;
2700
2701 // Add all pairwise combinations of operands to the pair map.
2702 unsigned BinaryIdx = I.getOpcode() - Instruction::BinaryOpsBegin;
2703 SmallSet<std::pair<Value *, Value*>, 32> Visited;
2704 for (unsigned i = 0; i < Ops.size() - 1; ++i) {
2705 for (unsigned j = i + 1; j < Ops.size(); ++j) {
2706 // Canonicalize operand orderings.
2707 Value *Op0 = Ops[i];
2708 Value *Op1 = Ops[j];
2709 if (std::less<Value *>()(Op1, Op0))
2710 std::swap(a&: Op0, b&: Op1);
2711 if (!Visited.insert(V: {Op0, Op1}).second)
2712 continue;
2713 auto res = PairMap[BinaryIdx].insert(KV: {{Op0, Op1}, {.Value1: Op0, .Value2: Op1, .Score: 1}});
2714 if (!res.second) {
2715 // If either key value has been erased then we've got the same
2716 // address by coincidence. That can't happen here because nothing is
2717 // erasing values but it can happen by the time we're querying the
2718 // map.
2719 assert(res.first->second.isValid() && "WeakVH invalidated");
2720 ++res.first->second.Score;
2721 }
2722 }
2723 }
2724 }
2725 }
2726}
2727
2728PreservedAnalyses ReassociatePass::run(Function &F,
2729 FunctionAnalysisManager &AM) {
2730 // UniformityInfo is empty (and cheap) on targets without branch divergence,
2731 // so request it unconditionally.
2732 UniformityInfo &UI = AM.getResult<UniformityInfoAnalysis>(IR&: F);
2733 return runImpl(F, UI);
2734}
2735
2736PreservedAnalyses ReassociatePass::runImpl(Function &F, UniformityInfo &UI) {
2737 UA = &UI;
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