1 | //===- Reassociate.cpp - Reassociate binary expressions -------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This pass reassociates commutative expressions in an order that is designed |
10 | // to promote better constant propagation, GCSE, LICM, PRE, etc. |
11 | // |
12 | // For example: 4 + (x + 5) -> x + (4 + 5) |
13 | // |
14 | // In the implementation of this algorithm, constants are assigned rank = 0, |
15 | // function arguments are rank = 1, and other values are assigned ranks |
16 | // corresponding to the reverse post order traversal of current function |
17 | // (starting at 2), which effectively gives values in deep loops higher rank |
18 | // than values not in loops. |
19 | // |
20 | //===----------------------------------------------------------------------===// |
21 | |
22 | #include "llvm/Transforms/Scalar/Reassociate.h" |
23 | #include "llvm/ADT/APFloat.h" |
24 | #include "llvm/ADT/APInt.h" |
25 | #include "llvm/ADT/DenseMap.h" |
26 | #include "llvm/ADT/PostOrderIterator.h" |
27 | #include "llvm/ADT/SmallPtrSet.h" |
28 | #include "llvm/ADT/SmallSet.h" |
29 | #include "llvm/ADT/SmallVector.h" |
30 | #include "llvm/ADT/Statistic.h" |
31 | #include "llvm/Analysis/BasicAliasAnalysis.h" |
32 | #include "llvm/Analysis/ConstantFolding.h" |
33 | #include "llvm/Analysis/GlobalsModRef.h" |
34 | #include "llvm/Analysis/ValueTracking.h" |
35 | #include "llvm/IR/Argument.h" |
36 | #include "llvm/IR/BasicBlock.h" |
37 | #include "llvm/IR/CFG.h" |
38 | #include "llvm/IR/Constant.h" |
39 | #include "llvm/IR/Constants.h" |
40 | #include "llvm/IR/Function.h" |
41 | #include "llvm/IR/IRBuilder.h" |
42 | #include "llvm/IR/InstrTypes.h" |
43 | #include "llvm/IR/Instruction.h" |
44 | #include "llvm/IR/Instructions.h" |
45 | #include "llvm/IR/Operator.h" |
46 | #include "llvm/IR/PassManager.h" |
47 | #include "llvm/IR/PatternMatch.h" |
48 | #include "llvm/IR/Type.h" |
49 | #include "llvm/IR/User.h" |
50 | #include "llvm/IR/Value.h" |
51 | #include "llvm/IR/ValueHandle.h" |
52 | #include "llvm/InitializePasses.h" |
53 | #include "llvm/Pass.h" |
54 | #include "llvm/Support/Casting.h" |
55 | #include "llvm/Support/CommandLine.h" |
56 | #include "llvm/Support/Debug.h" |
57 | #include "llvm/Support/raw_ostream.h" |
58 | #include "llvm/Transforms/Scalar.h" |
59 | #include "llvm/Transforms/Utils/Local.h" |
60 | #include <algorithm> |
61 | #include <cassert> |
62 | #include <utility> |
63 | |
64 | using namespace llvm; |
65 | using namespace reassociate; |
66 | using namespace PatternMatch; |
67 | |
68 | #define DEBUG_TYPE "reassociate" |
69 | |
70 | STATISTIC(NumChanged, "Number of insts reassociated" ); |
71 | STATISTIC(NumAnnihil, "Number of expr tree annihilated" ); |
72 | STATISTIC(NumFactor , "Number of multiplies factored" ); |
73 | |
74 | static cl::opt<bool> |
75 | UseCSELocalOpt(DEBUG_TYPE "-use-cse-local" , |
76 | cl::desc("Only reorder expressions within a basic block " |
77 | "when exposing CSE opportunities" ), |
78 | cl::init(Val: true), cl::Hidden); |
79 | |
80 | #ifndef NDEBUG |
81 | /// Print out the expression identified in the Ops list. |
82 | static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) { |
83 | Module *M = I->getModule(); |
84 | dbgs() << Instruction::getOpcodeName(Opcode: I->getOpcode()) << " " |
85 | << *Ops[0].Op->getType() << '\t'; |
86 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { |
87 | dbgs() << "[ " ; |
88 | Ops[i].Op->printAsOperand(O&: dbgs(), PrintType: false, M); |
89 | dbgs() << ", #" << Ops[i].Rank << "] " ; |
90 | } |
91 | } |
92 | #endif |
93 | |
94 | /// Utility class representing a non-constant Xor-operand. We classify |
95 | /// non-constant Xor-Operands into two categories: |
96 | /// C1) The operand is in the form "X & C", where C is a constant and C != ~0 |
97 | /// C2) |
98 | /// C2.1) The operand is in the form of "X | C", where C is a non-zero |
99 | /// constant. |
100 | /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this |
101 | /// operand as "E | 0" |
102 | class llvm::reassociate::XorOpnd { |
103 | public: |
104 | XorOpnd(Value *V); |
105 | |
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 | |
116 | private: |
117 | Value *OrigVal; |
118 | Value *SymbolicPart; |
119 | APInt ConstPart; |
120 | unsigned SymbolicRank; |
121 | bool isOr; |
122 | }; |
123 | |
124 | XorOpnd::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.) |
156 | static bool hasFPAssociativeFlags(Instruction *I) { |
157 | assert(I && isa<FPMathOperator>(I) && "Should only check FP ops" ); |
158 | return I->hasAllowReassoc() && I->hasNoSignedZeros(); |
159 | } |
160 | |
161 | /// Return true if V is an instruction of the specified opcode and if it |
162 | /// only has one use. |
163 | static 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 | |
171 | static 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 | void ReassociatePass::BuildRankMap(Function &F, |
182 | ReversePostOrderTraversal<Function*> &RPOT) { |
183 | unsigned Rank = 2; |
184 | |
185 | // Assign distinct ranks to function arguments. |
186 | for (auto &Arg : F.args()) { |
187 | ValueRankMap[&Arg] = ++Rank; |
188 | LLVM_DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank |
189 | << "\n" ); |
190 | } |
191 | |
192 | // Traverse basic blocks in ReversePostOrder. |
193 | for (BasicBlock *BB : RPOT) { |
194 | unsigned BBRank = RankMap[BB] = ++Rank << 16; |
195 | |
196 | // Walk the basic block, adding precomputed ranks for any instructions that |
197 | // we cannot move. This ensures that the ranks for these instructions are |
198 | // all different in the block. |
199 | for (Instruction &I : *BB) |
200 | if (mayHaveNonDefUseDependency(I)) |
201 | ValueRankMap[&I] = ++BBRank; |
202 | } |
203 | } |
204 | |
205 | unsigned ReassociatePass::getRank(Value *V) { |
206 | Instruction *I = dyn_cast<Instruction>(Val: V); |
207 | if (!I) { |
208 | if (isa<Argument>(Val: V)) return ValueRankMap[V]; // Function argument. |
209 | return 0; // Otherwise it's a global or constant, rank 0. |
210 | } |
211 | |
212 | if (unsigned Rank = ValueRankMap[I]) |
213 | return Rank; // Rank already known? |
214 | |
215 | // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that |
216 | // we can reassociate expressions for code motion! Since we do not recurse |
217 | // for PHI nodes, we cannot have infinite recursion here, because there |
218 | // cannot be loops in the value graph that do not go through PHI nodes. |
219 | unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; |
220 | for (unsigned i = 0, e = I->getNumOperands(); i != e && Rank != MaxRank; ++i) |
221 | Rank = std::max(a: Rank, b: getRank(V: I->getOperand(i))); |
222 | |
223 | // If this is a 'not' or 'neg' instruction, do not count it for rank. This |
224 | // assures us that X and ~X will have the same rank. |
225 | if (!match(V: I, P: m_Not(V: m_Value())) && !match(V: I, P: m_Neg(V: m_Value())) && |
226 | !match(V: I, P: m_FNeg(X: m_Value()))) |
227 | ++Rank; |
228 | |
229 | LLVM_DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank |
230 | << "\n" ); |
231 | |
232 | return ValueRankMap[I] = Rank; |
233 | } |
234 | |
235 | // Canonicalize constants to RHS. Otherwise, sort the operands by rank. |
236 | void ReassociatePass::canonicalizeOperands(Instruction *I) { |
237 | assert(isa<BinaryOperator>(I) && "Expected binary operator." ); |
238 | assert(I->isCommutative() && "Expected commutative operator." ); |
239 | |
240 | Value *LHS = I->getOperand(i: 0); |
241 | Value *RHS = I->getOperand(i: 1); |
242 | if (LHS == RHS || isa<Constant>(Val: RHS)) |
243 | return; |
244 | if (isa<Constant>(Val: LHS) || getRank(V: RHS) < getRank(V: LHS)) |
245 | cast<BinaryOperator>(Val: I)->swapOperands(); |
246 | } |
247 | |
248 | static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name, |
249 | BasicBlock::iterator InsertBefore, |
250 | Value *FlagsOp) { |
251 | if (S1->getType()->isIntOrIntVectorTy()) |
252 | return BinaryOperator::CreateAdd(V1: S1, V2: S2, Name, It: InsertBefore); |
253 | else { |
254 | BinaryOperator *Res = |
255 | BinaryOperator::CreateFAdd(V1: S1, V2: S2, Name, It: InsertBefore); |
256 | Res->setFastMathFlags(cast<FPMathOperator>(Val: FlagsOp)->getFastMathFlags()); |
257 | return Res; |
258 | } |
259 | } |
260 | |
261 | static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name, |
262 | BasicBlock::iterator InsertBefore, |
263 | Value *FlagsOp) { |
264 | if (S1->getType()->isIntOrIntVectorTy()) |
265 | return BinaryOperator::CreateMul(V1: S1, V2: S2, Name, It: InsertBefore); |
266 | else { |
267 | BinaryOperator *Res = |
268 | BinaryOperator::CreateFMul(V1: S1, V2: S2, Name, It: InsertBefore); |
269 | Res->setFastMathFlags(cast<FPMathOperator>(Val: FlagsOp)->getFastMathFlags()); |
270 | return Res; |
271 | } |
272 | } |
273 | |
274 | static Instruction *CreateNeg(Value *S1, const Twine &Name, |
275 | BasicBlock::iterator InsertBefore, |
276 | Value *FlagsOp) { |
277 | if (S1->getType()->isIntOrIntVectorTy()) |
278 | return BinaryOperator::CreateNeg(Op: S1, Name, InsertBefore); |
279 | |
280 | if (auto *FMFSource = dyn_cast<Instruction>(Val: FlagsOp)) |
281 | return UnaryOperator::CreateFNegFMF(Op: S1, FMFSource, Name, InsertBefore); |
282 | |
283 | return UnaryOperator::CreateFNeg(V: S1, Name, It: InsertBefore); |
284 | } |
285 | |
286 | /// Replace 0-X with X*-1. |
287 | static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) { |
288 | assert((isa<UnaryOperator>(Neg) || isa<BinaryOperator>(Neg)) && |
289 | "Expected a Negate!" ); |
290 | // FIXME: It's not safe to lower a unary FNeg into a FMul by -1.0. |
291 | unsigned OpNo = isa<BinaryOperator>(Val: Neg) ? 1 : 0; |
292 | Type *Ty = Neg->getType(); |
293 | Constant *NegOne = Ty->isIntOrIntVectorTy() ? |
294 | ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, V: -1.0); |
295 | |
296 | BinaryOperator *Res = |
297 | CreateMul(S1: Neg->getOperand(i: OpNo), S2: NegOne, Name: "" , InsertBefore: Neg->getIterator(), FlagsOp: Neg); |
298 | Neg->setOperand(i: OpNo, Val: Constant::getNullValue(Ty)); // Drop use of op. |
299 | Res->takeName(V: Neg); |
300 | Neg->replaceAllUsesWith(V: Res); |
301 | Res->setDebugLoc(Neg->getDebugLoc()); |
302 | return Res; |
303 | } |
304 | |
305 | /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael |
306 | /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for |
307 | /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic. |
308 | /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every |
309 | /// even x in Bitwidth-bit arithmetic. |
310 | static unsigned CarmichaelShift(unsigned Bitwidth) { |
311 | if (Bitwidth < 3) |
312 | return Bitwidth - 1; |
313 | return Bitwidth - 2; |
314 | } |
315 | |
316 | /// Add the extra weight 'RHS' to the existing weight 'LHS', |
317 | /// reducing the combined weight using any special properties of the operation. |
318 | /// The existing weight LHS represents the computation X op X op ... op X where |
319 | /// X occurs LHS times. The combined weight represents X op X op ... op X with |
320 | /// X occurring LHS + RHS times. If op is "Xor" for example then the combined |
321 | /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even; |
322 | /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second. |
323 | static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) { |
324 | // If we were working with infinite precision arithmetic then the combined |
325 | // weight would be LHS + RHS. But we are using finite precision arithmetic, |
326 | // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct |
327 | // for nilpotent operations and addition, but not for idempotent operations |
328 | // and multiplication), so it is important to correctly reduce the combined |
329 | // weight back into range if wrapping would be wrong. |
330 | |
331 | // If RHS is zero then the weight didn't change. |
332 | if (RHS.isMinValue()) |
333 | return; |
334 | // If LHS is zero then the combined weight is RHS. |
335 | if (LHS.isMinValue()) { |
336 | LHS = RHS; |
337 | return; |
338 | } |
339 | // From this point on we know that neither LHS nor RHS is zero. |
340 | |
341 | if (Instruction::isIdempotent(Opcode)) { |
342 | // Idempotent means X op X === X, so any non-zero weight is equivalent to a |
343 | // weight of 1. Keeping weights at zero or one also means that wrapping is |
344 | // not a problem. |
345 | assert(LHS == 1 && RHS == 1 && "Weights not reduced!" ); |
346 | return; // Return a weight of 1. |
347 | } |
348 | if (Instruction::isNilpotent(Opcode)) { |
349 | // Nilpotent means X op X === 0, so reduce weights modulo 2. |
350 | assert(LHS == 1 && RHS == 1 && "Weights not reduced!" ); |
351 | LHS = 0; // 1 + 1 === 0 modulo 2. |
352 | return; |
353 | } |
354 | if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) { |
355 | // TODO: Reduce the weight by exploiting nsw/nuw? |
356 | LHS += RHS; |
357 | return; |
358 | } |
359 | |
360 | assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) && |
361 | "Unknown associative operation!" ); |
362 | unsigned Bitwidth = LHS.getBitWidth(); |
363 | // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth |
364 | // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth |
365 | // bit number x, since either x is odd in which case x^CM = 1, or x is even in |
366 | // which case both x^W and x^(W - CM) are zero. By subtracting off multiples |
367 | // of CM like this weights can always be reduced to the range [0, CM+Bitwidth) |
368 | // which by a happy accident means that they can always be represented using |
369 | // Bitwidth bits. |
370 | // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than |
371 | // the Carmichael number). |
372 | if (Bitwidth > 3) { |
373 | /// CM - The value of Carmichael's lambda function. |
374 | APInt CM = APInt::getOneBitSet(numBits: Bitwidth, BitNo: CarmichaelShift(Bitwidth)); |
375 | // Any weight W >= Threshold can be replaced with W - CM. |
376 | APInt Threshold = CM + Bitwidth; |
377 | assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!" ); |
378 | // For Bitwidth 4 or more the following sum does not overflow. |
379 | LHS += RHS; |
380 | while (LHS.uge(RHS: Threshold)) |
381 | LHS -= CM; |
382 | } else { |
383 | // To avoid problems with overflow do everything the same as above but using |
384 | // a larger type. |
385 | unsigned CM = 1U << CarmichaelShift(Bitwidth); |
386 | unsigned Threshold = CM + Bitwidth; |
387 | assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && |
388 | "Weights not reduced!" ); |
389 | unsigned Total = LHS.getZExtValue() + RHS.getZExtValue(); |
390 | while (Total >= Threshold) |
391 | Total -= CM; |
392 | LHS = Total; |
393 | } |
394 | } |
395 | |
396 | using RepeatedValue = std::pair<Value*, APInt>; |
397 | |
398 | /// Given an associative binary expression, return the leaf |
399 | /// nodes in Ops along with their weights (how many times the leaf occurs). The |
400 | /// original expression is the same as |
401 | /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times |
402 | /// op |
403 | /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times |
404 | /// op |
405 | /// ... |
406 | /// op |
407 | /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times |
408 | /// |
409 | /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct. |
410 | /// |
411 | /// This routine may modify the function, in which case it returns 'true'. The |
412 | /// changes it makes may well be destructive, changing the value computed by 'I' |
413 | /// to something completely different. Thus if the routine returns 'true' then |
414 | /// you MUST either replace I with a new expression computed from the Ops array, |
415 | /// or use RewriteExprTree to put the values back in. |
416 | /// |
417 | /// A leaf node is either not a binary operation of the same kind as the root |
418 | /// node 'I' (i.e. is not a binary operator at all, or is, but with a different |
419 | /// opcode), or is the same kind of binary operator but has a use which either |
420 | /// does not belong to the expression, or does belong to the expression but is |
421 | /// a leaf node. Every leaf node has at least one use that is a non-leaf node |
422 | /// of the expression, while for non-leaf nodes (except for the root 'I') every |
423 | /// use is a non-leaf node of the expression. |
424 | /// |
425 | /// For example: |
426 | /// expression graph node names |
427 | /// |
428 | /// + | I |
429 | /// / \ | |
430 | /// + + | A, B |
431 | /// / \ / \ | |
432 | /// * + * | C, D, E |
433 | /// / \ / \ / \ | |
434 | /// + * | F, G |
435 | /// |
436 | /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in |
437 | /// that order) (C, 1), (E, 1), (F, 2), (G, 2). |
438 | /// |
439 | /// The expression is maximal: if some instruction is a binary operator of the |
440 | /// same kind as 'I', and all of its uses are non-leaf nodes of the expression, |
441 | /// then the instruction also belongs to the expression, is not a leaf node of |
442 | /// it, and its operands also belong to the expression (but may be leaf nodes). |
443 | /// |
444 | /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in |
445 | /// order to ensure that every non-root node in the expression has *exactly one* |
446 | /// use by a non-leaf node of the expression. This destruction means that the |
447 | /// caller MUST either replace 'I' with a new expression or use something like |
448 | /// RewriteExprTree to put the values back in if the routine indicates that it |
449 | /// made a change by returning 'true'. |
450 | /// |
451 | /// In the above example either the right operand of A or the left operand of B |
452 | /// will be replaced by undef. If it is B's operand then this gives: |
453 | /// |
454 | /// + | I |
455 | /// / \ | |
456 | /// + + | A, B - operand of B replaced with undef |
457 | /// / \ \ | |
458 | /// * + * | C, D, E |
459 | /// / \ / \ / \ | |
460 | /// + * | F, G |
461 | /// |
462 | /// Note that such undef operands can only be reached by passing through 'I'. |
463 | /// For example, if you visit operands recursively starting from a leaf node |
464 | /// then you will never see such an undef operand unless you get back to 'I', |
465 | /// which requires passing through a phi node. |
466 | /// |
467 | /// Note that this routine may also mutate binary operators of the wrong type |
468 | /// that have all uses inside the expression (i.e. only used by non-leaf nodes |
469 | /// of the expression) if it can turn them into binary operators of the right |
470 | /// type and thus make the expression bigger. |
471 | static bool LinearizeExprTree(Instruction *I, |
472 | SmallVectorImpl<RepeatedValue> &Ops, |
473 | ReassociatePass::OrderedSet &ToRedo, |
474 | bool &HasNUW) { |
475 | assert((isa<UnaryOperator>(I) || isa<BinaryOperator>(I)) && |
476 | "Expected a UnaryOperator or BinaryOperator!" ); |
477 | LLVM_DEBUG(dbgs() << "LINEARIZE: " << *I << '\n'); |
478 | unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits(); |
479 | unsigned Opcode = I->getOpcode(); |
480 | assert(I->isAssociative() && I->isCommutative() && |
481 | "Expected an associative and commutative operation!" ); |
482 | |
483 | // Visit all operands of the expression, keeping track of their weight (the |
484 | // number of paths from the expression root to the operand, or if you like |
485 | // the number of times that operand occurs in the linearized expression). |
486 | // For example, if I = X + A, where X = A + B, then I, X and B have weight 1 |
487 | // while A has weight two. |
488 | |
489 | // Worklist of non-leaf nodes (their operands are in the expression too) along |
490 | // with their weights, representing a certain number of paths to the operator. |
491 | // If an operator occurs in the worklist multiple times then we found multiple |
492 | // ways to get to it. |
493 | SmallVector<std::pair<Instruction*, APInt>, 8> Worklist; // (Op, Weight) |
494 | Worklist.push_back(Elt: std::make_pair(x&: I, y: APInt(Bitwidth, 1))); |
495 | bool Changed = false; |
496 | |
497 | // Leaves of the expression are values that either aren't the right kind of |
498 | // operation (eg: a constant, or a multiply in an add tree), or are, but have |
499 | // some uses that are not inside the expression. For example, in I = X + X, |
500 | // X = A + B, the value X has two uses (by I) that are in the expression. If |
501 | // X has any other uses, for example in a return instruction, then we consider |
502 | // X to be a leaf, and won't analyze it further. When we first visit a value, |
503 | // if it has more than one use then at first we conservatively consider it to |
504 | // be a leaf. Later, as the expression is explored, we may discover some more |
505 | // uses of the value from inside the expression. If all uses turn out to be |
506 | // from within the expression (and the value is a binary operator of the right |
507 | // kind) then the value is no longer considered to be a leaf, and its operands |
508 | // are explored. |
509 | |
510 | // Leaves - Keeps track of the set of putative leaves as well as the number of |
511 | // paths to each leaf seen so far. |
512 | using LeafMap = DenseMap<Value *, APInt>; |
513 | LeafMap Leaves; // Leaf -> Total weight so far. |
514 | SmallVector<Value *, 8> LeafOrder; // Ensure deterministic leaf output order. |
515 | |
516 | #ifndef NDEBUG |
517 | SmallPtrSet<Value *, 8> Visited; // For checking the iteration scheme. |
518 | #endif |
519 | while (!Worklist.empty()) { |
520 | std::pair<Instruction*, APInt> P = Worklist.pop_back_val(); |
521 | I = P.first; // We examine the operands of this binary operator. |
522 | |
523 | if (isa<OverflowingBinaryOperator>(Val: I)) |
524 | HasNUW &= I->hasNoUnsignedWrap(); |
525 | |
526 | for (unsigned OpIdx = 0; OpIdx < I->getNumOperands(); ++OpIdx) { // Visit operands. |
527 | Value *Op = I->getOperand(i: OpIdx); |
528 | APInt Weight = P.second; // Number of paths to this operand. |
529 | LLVM_DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n" ); |
530 | assert(!Op->use_empty() && "No uses, so how did we get to it?!" ); |
531 | |
532 | // If this is a binary operation of the right kind with only one use then |
533 | // add its operands to the expression. |
534 | if (BinaryOperator *BO = isReassociableOp(V: Op, Opcode)) { |
535 | assert(Visited.insert(Op).second && "Not first visit!" ); |
536 | LLVM_DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n" ); |
537 | Worklist.push_back(Elt: std::make_pair(x&: BO, y&: Weight)); |
538 | continue; |
539 | } |
540 | |
541 | // Appears to be a leaf. Is the operand already in the set of leaves? |
542 | LeafMap::iterator It = Leaves.find(Val: Op); |
543 | if (It == Leaves.end()) { |
544 | // Not in the leaf map. Must be the first time we saw this operand. |
545 | assert(Visited.insert(Op).second && "Not first visit!" ); |
546 | if (!Op->hasOneUse()) { |
547 | // This value has uses not accounted for by the expression, so it is |
548 | // not safe to modify. Mark it as being a leaf. |
549 | LLVM_DEBUG(dbgs() |
550 | << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n" ); |
551 | LeafOrder.push_back(Elt: Op); |
552 | Leaves[Op] = Weight; |
553 | continue; |
554 | } |
555 | // No uses outside the expression, try morphing it. |
556 | } else { |
557 | // Already in the leaf map. |
558 | assert(It != Leaves.end() && Visited.count(Op) && |
559 | "In leaf map but not visited!" ); |
560 | |
561 | // Update the number of paths to the leaf. |
562 | IncorporateWeight(LHS&: It->second, RHS: Weight, Opcode); |
563 | |
564 | #if 0 // TODO: Re-enable once PR13021 is fixed. |
565 | // The leaf already has one use from inside the expression. As we want |
566 | // exactly one such use, drop this new use of the leaf. |
567 | assert(!Op->hasOneUse() && "Only one use, but we got here twice!" ); |
568 | I->setOperand(OpIdx, UndefValue::get(I->getType())); |
569 | Changed = true; |
570 | |
571 | // If the leaf is a binary operation of the right kind and we now see |
572 | // that its multiple original uses were in fact all by nodes belonging |
573 | // to the expression, then no longer consider it to be a leaf and add |
574 | // its operands to the expression. |
575 | if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { |
576 | LLVM_DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n" ); |
577 | Worklist.push_back(std::make_pair(BO, It->second)); |
578 | Leaves.erase(It); |
579 | continue; |
580 | } |
581 | #endif |
582 | |
583 | // If we still have uses that are not accounted for by the expression |
584 | // then it is not safe to modify the value. |
585 | if (!Op->hasOneUse()) |
586 | continue; |
587 | |
588 | // No uses outside the expression, try morphing it. |
589 | Weight = It->second; |
590 | Leaves.erase(I: It); // Since the value may be morphed below. |
591 | } |
592 | |
593 | // At this point we have a value which, first of all, is not a binary |
594 | // expression of the right kind, and secondly, is only used inside the |
595 | // expression. This means that it can safely be modified. See if we |
596 | // can usefully morph it into an expression of the right kind. |
597 | assert((!isa<Instruction>(Op) || |
598 | cast<Instruction>(Op)->getOpcode() != Opcode |
599 | || (isa<FPMathOperator>(Op) && |
600 | !hasFPAssociativeFlags(cast<Instruction>(Op)))) && |
601 | "Should have been handled above!" ); |
602 | assert(Op->hasOneUse() && "Has uses outside the expression tree!" ); |
603 | |
604 | // If this is a multiply expression, turn any internal negations into |
605 | // multiplies by -1 so they can be reassociated. Add any users of the |
606 | // newly created multiplication by -1 to the redo list, so any |
607 | // reassociation opportunities that are exposed will be reassociated |
608 | // further. |
609 | Instruction *Neg; |
610 | if (((Opcode == Instruction::Mul && match(V: Op, P: m_Neg(V: m_Value()))) || |
611 | (Opcode == Instruction::FMul && match(V: Op, P: m_FNeg(X: m_Value())))) && |
612 | match(V: Op, P: m_Instruction(I&: Neg))) { |
613 | LLVM_DEBUG(dbgs() |
614 | << "MORPH LEAF: " << *Op << " (" << Weight << ") TO " ); |
615 | Instruction *Mul = LowerNegateToMultiply(Neg); |
616 | LLVM_DEBUG(dbgs() << *Mul << '\n'); |
617 | Worklist.push_back(Elt: std::make_pair(x&: Mul, y&: Weight)); |
618 | for (User *U : Mul->users()) { |
619 | if (BinaryOperator *UserBO = dyn_cast<BinaryOperator>(Val: U)) |
620 | ToRedo.insert(X: UserBO); |
621 | } |
622 | ToRedo.insert(X: Neg); |
623 | Changed = true; |
624 | continue; |
625 | } |
626 | |
627 | // Failed to morph into an expression of the right type. This really is |
628 | // a leaf. |
629 | LLVM_DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n" ); |
630 | assert(!isReassociableOp(Op, Opcode) && "Value was morphed?" ); |
631 | LeafOrder.push_back(Elt: Op); |
632 | Leaves[Op] = Weight; |
633 | } |
634 | } |
635 | |
636 | // The leaves, repeated according to their weights, represent the linearized |
637 | // form of the expression. |
638 | for (Value *V : LeafOrder) { |
639 | LeafMap::iterator It = Leaves.find(Val: V); |
640 | if (It == Leaves.end()) |
641 | // Node initially thought to be a leaf wasn't. |
642 | continue; |
643 | assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!" ); |
644 | APInt Weight = It->second; |
645 | if (Weight.isMinValue()) |
646 | // Leaf already output or weight reduction eliminated it. |
647 | continue; |
648 | // Ensure the leaf is only output once. |
649 | It->second = 0; |
650 | Ops.push_back(Elt: std::make_pair(x&: V, y&: Weight)); |
651 | } |
652 | |
653 | // For nilpotent operations or addition there may be no operands, for example |
654 | // because the expression was "X xor X" or consisted of 2^Bitwidth additions: |
655 | // in both cases the weight reduces to 0 causing the value to be skipped. |
656 | if (Ops.empty()) { |
657 | Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, Ty: I->getType()); |
658 | assert(Identity && "Associative operation without identity!" ); |
659 | Ops.emplace_back(Args&: Identity, Args: APInt(Bitwidth, 1)); |
660 | } |
661 | |
662 | return Changed; |
663 | } |
664 | |
665 | /// Now that the operands for this expression tree are |
666 | /// linearized and optimized, emit them in-order. |
667 | void ReassociatePass::RewriteExprTree(BinaryOperator *I, |
668 | SmallVectorImpl<ValueEntry> &Ops, |
669 | bool HasNUW) { |
670 | assert(Ops.size() > 1 && "Single values should be used directly!" ); |
671 | |
672 | // Since our optimizations should never increase the number of operations, the |
673 | // new expression can usually be written reusing the existing binary operators |
674 | // from the original expression tree, without creating any new instructions, |
675 | // though the rewritten expression may have a completely different topology. |
676 | // We take care to not change anything if the new expression will be the same |
677 | // as the original. If more than trivial changes (like commuting operands) |
678 | // were made then we are obliged to clear out any optional subclass data like |
679 | // nsw flags. |
680 | |
681 | /// NodesToRewrite - Nodes from the original expression available for writing |
682 | /// the new expression into. |
683 | SmallVector<BinaryOperator*, 8> NodesToRewrite; |
684 | unsigned Opcode = I->getOpcode(); |
685 | BinaryOperator *Op = I; |
686 | |
687 | /// NotRewritable - The operands being written will be the leaves of the new |
688 | /// expression and must not be used as inner nodes (via NodesToRewrite) by |
689 | /// mistake. Inner nodes are always reassociable, and usually leaves are not |
690 | /// (if they were they would have been incorporated into the expression and so |
691 | /// would not be leaves), so most of the time there is no danger of this. But |
692 | /// in rare cases a leaf may become reassociable if an optimization kills uses |
693 | /// of it, or it may momentarily become reassociable during rewriting (below) |
694 | /// due it being removed as an operand of one of its uses. Ensure that misuse |
695 | /// of leaf nodes as inner nodes cannot occur by remembering all of the future |
696 | /// leaves and refusing to reuse any of them as inner nodes. |
697 | SmallPtrSet<Value*, 8> NotRewritable; |
698 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) |
699 | NotRewritable.insert(Ptr: Ops[i].Op); |
700 | |
701 | // ExpressionChangedStart - Non-null if the rewritten expression differs from |
702 | // the original in some non-trivial way, requiring the clearing of optional |
703 | // flags. Flags are cleared from the operator in ExpressionChangedStart up to |
704 | // ExpressionChangedEnd inclusive. |
705 | BinaryOperator *ExpressionChangedStart = nullptr, |
706 | *ExpressionChangedEnd = nullptr; |
707 | for (unsigned i = 0; ; ++i) { |
708 | // The last operation (which comes earliest in the IR) is special as both |
709 | // operands will come from Ops, rather than just one with the other being |
710 | // a subexpression. |
711 | if (i+2 == Ops.size()) { |
712 | Value *NewLHS = Ops[i].Op; |
713 | Value *NewRHS = Ops[i+1].Op; |
714 | Value *OldLHS = Op->getOperand(i_nocapture: 0); |
715 | Value *OldRHS = Op->getOperand(i_nocapture: 1); |
716 | |
717 | if (NewLHS == OldLHS && NewRHS == OldRHS) |
718 | // Nothing changed, leave it alone. |
719 | break; |
720 | |
721 | if (NewLHS == OldRHS && NewRHS == OldLHS) { |
722 | // The order of the operands was reversed. Swap them. |
723 | LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n'); |
724 | Op->swapOperands(); |
725 | LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n'); |
726 | MadeChange = true; |
727 | ++NumChanged; |
728 | break; |
729 | } |
730 | |
731 | // The new operation differs non-trivially from the original. Overwrite |
732 | // the old operands with the new ones. |
733 | LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n'); |
734 | if (NewLHS != OldLHS) { |
735 | BinaryOperator *BO = isReassociableOp(V: OldLHS, Opcode); |
736 | if (BO && !NotRewritable.count(Ptr: BO)) |
737 | NodesToRewrite.push_back(Elt: BO); |
738 | Op->setOperand(i_nocapture: 0, Val_nocapture: NewLHS); |
739 | } |
740 | if (NewRHS != OldRHS) { |
741 | BinaryOperator *BO = isReassociableOp(V: OldRHS, Opcode); |
742 | if (BO && !NotRewritable.count(Ptr: BO)) |
743 | NodesToRewrite.push_back(Elt: BO); |
744 | Op->setOperand(i_nocapture: 1, Val_nocapture: NewRHS); |
745 | } |
746 | LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n'); |
747 | |
748 | ExpressionChangedStart = Op; |
749 | if (!ExpressionChangedEnd) |
750 | ExpressionChangedEnd = Op; |
751 | MadeChange = true; |
752 | ++NumChanged; |
753 | |
754 | break; |
755 | } |
756 | |
757 | // Not the last operation. The left-hand side will be a sub-expression |
758 | // while the right-hand side will be the current element of Ops. |
759 | Value *NewRHS = Ops[i].Op; |
760 | if (NewRHS != Op->getOperand(i_nocapture: 1)) { |
761 | LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n'); |
762 | if (NewRHS == Op->getOperand(i_nocapture: 0)) { |
763 | // The new right-hand side was already present as the left operand. If |
764 | // we are lucky then swapping the operands will sort out both of them. |
765 | Op->swapOperands(); |
766 | } else { |
767 | // Overwrite with the new right-hand side. |
768 | BinaryOperator *BO = isReassociableOp(V: Op->getOperand(i_nocapture: 1), Opcode); |
769 | if (BO && !NotRewritable.count(Ptr: BO)) |
770 | NodesToRewrite.push_back(Elt: BO); |
771 | Op->setOperand(i_nocapture: 1, Val_nocapture: NewRHS); |
772 | ExpressionChangedStart = Op; |
773 | if (!ExpressionChangedEnd) |
774 | ExpressionChangedEnd = Op; |
775 | } |
776 | LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n'); |
777 | MadeChange = true; |
778 | ++NumChanged; |
779 | } |
780 | |
781 | // Now deal with the left-hand side. If this is already an operation node |
782 | // from the original expression then just rewrite the rest of the expression |
783 | // into it. |
784 | BinaryOperator *BO = isReassociableOp(V: Op->getOperand(i_nocapture: 0), Opcode); |
785 | if (BO && !NotRewritable.count(Ptr: BO)) { |
786 | Op = BO; |
787 | continue; |
788 | } |
789 | |
790 | // Otherwise, grab a spare node from the original expression and use that as |
791 | // the left-hand side. If there are no nodes left then the optimizers made |
792 | // an expression with more nodes than the original! This usually means that |
793 | // they did something stupid but it might mean that the problem was just too |
794 | // hard (finding the mimimal number of multiplications needed to realize a |
795 | // multiplication expression is NP-complete). Whatever the reason, smart or |
796 | // stupid, create a new node if there are none left. |
797 | BinaryOperator *NewOp; |
798 | if (NodesToRewrite.empty()) { |
799 | Constant *Undef = UndefValue::get(T: I->getType()); |
800 | NewOp = BinaryOperator::Create(Op: Instruction::BinaryOps(Opcode), S1: Undef, |
801 | S2: Undef, Name: "" , InsertBefore: I->getIterator()); |
802 | if (isa<FPMathOperator>(Val: NewOp)) |
803 | NewOp->setFastMathFlags(I->getFastMathFlags()); |
804 | } else { |
805 | NewOp = NodesToRewrite.pop_back_val(); |
806 | } |
807 | |
808 | LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n'); |
809 | Op->setOperand(i_nocapture: 0, Val_nocapture: NewOp); |
810 | LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n'); |
811 | ExpressionChangedStart = Op; |
812 | if (!ExpressionChangedEnd) |
813 | ExpressionChangedEnd = Op; |
814 | MadeChange = true; |
815 | ++NumChanged; |
816 | Op = NewOp; |
817 | } |
818 | |
819 | // If the expression changed non-trivially then clear out all subclass data |
820 | // starting from the operator specified in ExpressionChanged, and compactify |
821 | // the operators to just before the expression root to guarantee that the |
822 | // expression tree is dominated by all of Ops. |
823 | if (ExpressionChangedStart) { |
824 | bool ClearFlags = true; |
825 | do { |
826 | // Preserve flags. |
827 | if (ClearFlags) { |
828 | if (isa<FPMathOperator>(Val: I)) { |
829 | FastMathFlags Flags = I->getFastMathFlags(); |
830 | ExpressionChangedStart->clearSubclassOptionalData(); |
831 | ExpressionChangedStart->setFastMathFlags(Flags); |
832 | } else { |
833 | ExpressionChangedStart->clearSubclassOptionalData(); |
834 | // Note that it doesn't hold for mul if one of the operands is zero. |
835 | // TODO: We can preserve NUW flag if we prove that all mul operands |
836 | // are non-zero. |
837 | if (HasNUW && ExpressionChangedStart->getOpcode() == Instruction::Add) |
838 | ExpressionChangedStart->setHasNoUnsignedWrap(); |
839 | } |
840 | } |
841 | |
842 | if (ExpressionChangedStart == ExpressionChangedEnd) |
843 | ClearFlags = false; |
844 | if (ExpressionChangedStart == I) |
845 | break; |
846 | |
847 | // Discard any debug info related to the expressions that has changed (we |
848 | // can leave debug info related to the root and any operation that didn't |
849 | // change, since the result of the expression tree should be the same |
850 | // even after reassociation). |
851 | if (ClearFlags) |
852 | replaceDbgUsesWithUndef(I: ExpressionChangedStart); |
853 | |
854 | ExpressionChangedStart->moveBefore(MovePos: I); |
855 | ExpressionChangedStart = |
856 | cast<BinaryOperator>(Val: *ExpressionChangedStart->user_begin()); |
857 | } while (true); |
858 | } |
859 | |
860 | // Throw away any left over nodes from the original expression. |
861 | for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i) |
862 | RedoInsts.insert(X: NodesToRewrite[i]); |
863 | } |
864 | |
865 | /// Insert instructions before the instruction pointed to by BI, |
866 | /// that computes the negative version of the value specified. The negative |
867 | /// version of the value is returned, and BI is left pointing at the instruction |
868 | /// that should be processed next by the reassociation pass. |
869 | /// Also add intermediate instructions to the redo list that are modified while |
870 | /// pushing the negates through adds. These will be revisited to see if |
871 | /// additional opportunities have been exposed. |
872 | static Value *NegateValue(Value *V, Instruction *BI, |
873 | ReassociatePass::OrderedSet &ToRedo) { |
874 | if (auto *C = dyn_cast<Constant>(Val: V)) { |
875 | const DataLayout &DL = BI->getModule()->getDataLayout(); |
876 | Constant *Res = C->getType()->isFPOrFPVectorTy() |
877 | ? ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL) |
878 | : ConstantExpr::getNeg(C); |
879 | if (Res) |
880 | return Res; |
881 | } |
882 | |
883 | // We are trying to expose opportunity for reassociation. One of the things |
884 | // that we want to do to achieve this is to push a negation as deep into an |
885 | // expression chain as possible, to expose the add instructions. In practice, |
886 | // this means that we turn this: |
887 | // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D |
888 | // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate |
889 | // the constants. We assume that instcombine will clean up the mess later if |
890 | // we introduce tons of unnecessary negation instructions. |
891 | // |
892 | if (BinaryOperator *I = |
893 | isReassociableOp(V, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd)) { |
894 | // Push the negates through the add. |
895 | I->setOperand(i_nocapture: 0, Val_nocapture: NegateValue(V: I->getOperand(i_nocapture: 0), BI, ToRedo)); |
896 | I->setOperand(i_nocapture: 1, Val_nocapture: NegateValue(V: I->getOperand(i_nocapture: 1), BI, ToRedo)); |
897 | if (I->getOpcode() == Instruction::Add) { |
898 | I->setHasNoUnsignedWrap(false); |
899 | I->setHasNoSignedWrap(false); |
900 | } |
901 | |
902 | // We must move the add instruction here, because the neg instructions do |
903 | // not dominate the old add instruction in general. By moving it, we are |
904 | // assured that the neg instructions we just inserted dominate the |
905 | // instruction we are about to insert after them. |
906 | // |
907 | I->moveBefore(MovePos: BI); |
908 | I->setName(I->getName()+".neg" ); |
909 | |
910 | // Add the intermediate negates to the redo list as processing them later |
911 | // could expose more reassociating opportunities. |
912 | ToRedo.insert(X: I); |
913 | return I; |
914 | } |
915 | |
916 | // Okay, we need to materialize a negated version of V with an instruction. |
917 | // Scan the use lists of V to see if we have one already. |
918 | for (User *U : V->users()) { |
919 | if (!match(V: U, P: m_Neg(V: m_Value())) && !match(V: U, P: m_FNeg(X: m_Value()))) |
920 | continue; |
921 | |
922 | // We found one! Now we have to make sure that the definition dominates |
923 | // this use. We do this by moving it to the entry block (if it is a |
924 | // non-instruction value) or right after the definition. These negates will |
925 | // be zapped by reassociate later, so we don't need much finesse here. |
926 | Instruction *TheNeg = dyn_cast<Instruction>(Val: U); |
927 | |
928 | // We can't safely propagate a vector zero constant with poison/undef lanes. |
929 | Constant *C; |
930 | if (match(V: TheNeg, P: m_BinOp(L: m_Constant(C), R: m_Value())) && |
931 | C->containsUndefOrPoisonElement()) |
932 | continue; |
933 | |
934 | // Verify that the negate is in this function, V might be a constant expr. |
935 | if (!TheNeg || |
936 | TheNeg->getParent()->getParent() != BI->getParent()->getParent()) |
937 | continue; |
938 | |
939 | BasicBlock::iterator InsertPt; |
940 | if (Instruction *InstInput = dyn_cast<Instruction>(Val: V)) { |
941 | auto InsertPtOpt = InstInput->getInsertionPointAfterDef(); |
942 | if (!InsertPtOpt) |
943 | continue; |
944 | InsertPt = *InsertPtOpt; |
945 | } else { |
946 | InsertPt = TheNeg->getFunction() |
947 | ->getEntryBlock() |
948 | .getFirstNonPHIOrDbg() |
949 | ->getIterator(); |
950 | } |
951 | |
952 | TheNeg->moveBefore(BB&: *InsertPt->getParent(), I: InsertPt); |
953 | if (TheNeg->getOpcode() == Instruction::Sub) { |
954 | TheNeg->setHasNoUnsignedWrap(false); |
955 | TheNeg->setHasNoSignedWrap(false); |
956 | } else { |
957 | TheNeg->andIRFlags(V: BI); |
958 | } |
959 | ToRedo.insert(X: TheNeg); |
960 | return TheNeg; |
961 | } |
962 | |
963 | // Insert a 'neg' instruction that subtracts the value from zero to get the |
964 | // negation. |
965 | Instruction *NewNeg = |
966 | CreateNeg(S1: V, Name: V->getName() + ".neg" , InsertBefore: BI->getIterator(), FlagsOp: BI); |
967 | ToRedo.insert(X: NewNeg); |
968 | return NewNeg; |
969 | } |
970 | |
971 | // See if this `or` looks like an load widening reduction, i.e. that it |
972 | // consists of an `or`/`shl`/`zext`/`load` nodes only. Note that we don't |
973 | // ensure that the pattern is *really* a load widening reduction, |
974 | // we do not ensure that it can really be replaced with a widened load, |
975 | // only that it mostly looks like one. |
976 | static bool isLoadCombineCandidate(Instruction *Or) { |
977 | SmallVector<Instruction *, 8> Worklist; |
978 | SmallSet<Instruction *, 8> Visited; |
979 | |
980 | auto Enqueue = [&](Value *V) { |
981 | auto *I = dyn_cast<Instruction>(Val: V); |
982 | // Each node of an `or` reduction must be an instruction, |
983 | if (!I) |
984 | return false; // Node is certainly not part of an `or` load reduction. |
985 | // Only process instructions we have never processed before. |
986 | if (Visited.insert(Ptr: I).second) |
987 | Worklist.emplace_back(Args&: I); |
988 | return true; // Will need to look at parent nodes. |
989 | }; |
990 | |
991 | if (!Enqueue(Or)) |
992 | return false; // Not an `or` reduction pattern. |
993 | |
994 | while (!Worklist.empty()) { |
995 | auto *I = Worklist.pop_back_val(); |
996 | |
997 | // Okay, which instruction is this node? |
998 | switch (I->getOpcode()) { |
999 | case Instruction::Or: |
1000 | // Got an `or` node. That's fine, just recurse into it's operands. |
1001 | for (Value *Op : I->operands()) |
1002 | if (!Enqueue(Op)) |
1003 | return false; // Not an `or` reduction pattern. |
1004 | continue; |
1005 | |
1006 | case Instruction::Shl: |
1007 | case Instruction::ZExt: |
1008 | // `shl`/`zext` nodes are fine, just recurse into their base operand. |
1009 | if (!Enqueue(I->getOperand(i: 0))) |
1010 | return false; // Not an `or` reduction pattern. |
1011 | continue; |
1012 | |
1013 | case Instruction::Load: |
1014 | // Perfect, `load` node means we've reached an edge of the graph. |
1015 | continue; |
1016 | |
1017 | default: // Unknown node. |
1018 | return false; // Not an `or` reduction pattern. |
1019 | } |
1020 | } |
1021 | |
1022 | return true; |
1023 | } |
1024 | |
1025 | /// Return true if it may be profitable to convert this (X|Y) into (X+Y). |
1026 | static bool shouldConvertOrWithNoCommonBitsToAdd(Instruction *Or) { |
1027 | // Don't bother to convert this up unless either the LHS is an associable add |
1028 | // or subtract or mul or if this is only used by one of the above. |
1029 | // This is only a compile-time improvement, it is not needed for correctness! |
1030 | auto isInteresting = [](Value *V) { |
1031 | for (auto Op : {Instruction::Add, Instruction::Sub, Instruction::Mul, |
1032 | Instruction::Shl}) |
1033 | if (isReassociableOp(V, Opcode: Op)) |
1034 | return true; |
1035 | return false; |
1036 | }; |
1037 | |
1038 | if (any_of(Range: Or->operands(), P: isInteresting)) |
1039 | return true; |
1040 | |
1041 | Value *VB = Or->user_back(); |
1042 | if (Or->hasOneUse() && isInteresting(VB)) |
1043 | return true; |
1044 | |
1045 | return false; |
1046 | } |
1047 | |
1048 | /// If we have (X|Y), and iff X and Y have no common bits set, |
1049 | /// transform this into (X+Y) to allow arithmetics reassociation. |
1050 | static BinaryOperator *convertOrWithNoCommonBitsToAdd(Instruction *Or) { |
1051 | // Convert an or into an add. |
1052 | BinaryOperator *New = CreateAdd(S1: Or->getOperand(i: 0), S2: Or->getOperand(i: 1), Name: "" , |
1053 | InsertBefore: Or->getIterator(), FlagsOp: Or); |
1054 | New->setHasNoSignedWrap(); |
1055 | New->setHasNoUnsignedWrap(); |
1056 | New->takeName(V: Or); |
1057 | |
1058 | // Everyone now refers to the add instruction. |
1059 | Or->replaceAllUsesWith(V: New); |
1060 | New->setDebugLoc(Or->getDebugLoc()); |
1061 | |
1062 | LLVM_DEBUG(dbgs() << "Converted or into an add: " << *New << '\n'); |
1063 | return New; |
1064 | } |
1065 | |
1066 | /// Return true if we should break up this subtract of X-Y into (X + -Y). |
1067 | static bool ShouldBreakUpSubtract(Instruction *Sub) { |
1068 | // If this is a negation, we can't split it up! |
1069 | if (match(V: Sub, P: m_Neg(V: m_Value())) || match(V: Sub, P: m_FNeg(X: m_Value()))) |
1070 | return false; |
1071 | |
1072 | // Don't breakup X - undef. |
1073 | if (isa<UndefValue>(Val: Sub->getOperand(i: 1))) |
1074 | return false; |
1075 | |
1076 | // Don't bother to break this up unless either the LHS is an associable add or |
1077 | // subtract or if this is only used by one. |
1078 | Value *V0 = Sub->getOperand(i: 0); |
1079 | if (isReassociableOp(V: V0, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd) || |
1080 | isReassociableOp(V: V0, Opcode1: Instruction::Sub, Opcode2: Instruction::FSub)) |
1081 | return true; |
1082 | Value *V1 = Sub->getOperand(i: 1); |
1083 | if (isReassociableOp(V: V1, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd) || |
1084 | isReassociableOp(V: V1, Opcode1: Instruction::Sub, Opcode2: Instruction::FSub)) |
1085 | return true; |
1086 | Value *VB = Sub->user_back(); |
1087 | if (Sub->hasOneUse() && |
1088 | (isReassociableOp(V: VB, Opcode1: Instruction::Add, Opcode2: Instruction::FAdd) || |
1089 | isReassociableOp(V: VB, Opcode1: Instruction::Sub, Opcode2: Instruction::FSub))) |
1090 | return true; |
1091 | |
1092 | return false; |
1093 | } |
1094 | |
1095 | /// If we have (X-Y), and if either X is an add, or if this is only used by an |
1096 | /// add, transform this into (X+(0-Y)) to promote better reassociation. |
1097 | static BinaryOperator *BreakUpSubtract(Instruction *Sub, |
1098 | ReassociatePass::OrderedSet &ToRedo) { |
1099 | // Convert a subtract into an add and a neg instruction. This allows sub |
1100 | // instructions to be commuted with other add instructions. |
1101 | // |
1102 | // Calculate the negative value of Operand 1 of the sub instruction, |
1103 | // and set it as the RHS of the add instruction we just made. |
1104 | Value *NegVal = NegateValue(V: Sub->getOperand(i: 1), BI: Sub, ToRedo); |
1105 | BinaryOperator *New = |
1106 | CreateAdd(S1: Sub->getOperand(i: 0), S2: NegVal, Name: "" , InsertBefore: Sub->getIterator(), FlagsOp: Sub); |
1107 | Sub->setOperand(i: 0, Val: Constant::getNullValue(Ty: Sub->getType())); // Drop use of op. |
1108 | Sub->setOperand(i: 1, Val: Constant::getNullValue(Ty: Sub->getType())); // Drop use of op. |
1109 | New->takeName(V: Sub); |
1110 | |
1111 | // Everyone now refers to the add instruction. |
1112 | Sub->replaceAllUsesWith(V: New); |
1113 | New->setDebugLoc(Sub->getDebugLoc()); |
1114 | |
1115 | LLVM_DEBUG(dbgs() << "Negated: " << *New << '\n'); |
1116 | return New; |
1117 | } |
1118 | |
1119 | /// If this is a shift of a reassociable multiply or is used by one, change |
1120 | /// this into a multiply by a constant to assist with further reassociation. |
1121 | static BinaryOperator *ConvertShiftToMul(Instruction *Shl) { |
1122 | Constant *MulCst = ConstantInt::get(Ty: Shl->getType(), V: 1); |
1123 | auto *SA = cast<ConstantInt>(Val: Shl->getOperand(i: 1)); |
1124 | MulCst = ConstantExpr::getShl(C1: MulCst, C2: SA); |
1125 | |
1126 | BinaryOperator *Mul = BinaryOperator::CreateMul(V1: Shl->getOperand(i: 0), V2: MulCst, |
1127 | Name: "" , It: Shl->getIterator()); |
1128 | Shl->setOperand(i: 0, Val: PoisonValue::get(T: Shl->getType())); // Drop use of op. |
1129 | Mul->takeName(V: Shl); |
1130 | |
1131 | // Everyone now refers to the mul instruction. |
1132 | Shl->replaceAllUsesWith(V: Mul); |
1133 | Mul->setDebugLoc(Shl->getDebugLoc()); |
1134 | |
1135 | // We can safely preserve the nuw flag in all cases. It's also safe to turn a |
1136 | // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special |
1137 | // handling. It can be preserved as long as we're not left shifting by |
1138 | // bitwidth - 1. |
1139 | bool NSW = cast<BinaryOperator>(Val: Shl)->hasNoSignedWrap(); |
1140 | bool NUW = cast<BinaryOperator>(Val: Shl)->hasNoUnsignedWrap(); |
1141 | unsigned BitWidth = Shl->getType()->getIntegerBitWidth(); |
1142 | if (NSW && (NUW || SA->getValue().ult(RHS: BitWidth - 1))) |
1143 | Mul->setHasNoSignedWrap(true); |
1144 | Mul->setHasNoUnsignedWrap(NUW); |
1145 | return Mul; |
1146 | } |
1147 | |
1148 | /// Scan backwards and forwards among values with the same rank as element i |
1149 | /// to see if X exists. If X does not exist, return i. This is useful when |
1150 | /// scanning for 'x' when we see '-x' because they both get the same rank. |
1151 | static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops, |
1152 | unsigned i, Value *X) { |
1153 | unsigned XRank = Ops[i].Rank; |
1154 | unsigned e = Ops.size(); |
1155 | for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) { |
1156 | if (Ops[j].Op == X) |
1157 | return j; |
1158 | if (Instruction *I1 = dyn_cast<Instruction>(Val: Ops[j].Op)) |
1159 | if (Instruction *I2 = dyn_cast<Instruction>(Val: X)) |
1160 | if (I1->isIdenticalTo(I: I2)) |
1161 | return j; |
1162 | } |
1163 | // Scan backwards. |
1164 | for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) { |
1165 | if (Ops[j].Op == X) |
1166 | return j; |
1167 | if (Instruction *I1 = dyn_cast<Instruction>(Val: Ops[j].Op)) |
1168 | if (Instruction *I2 = dyn_cast<Instruction>(Val: X)) |
1169 | if (I1->isIdenticalTo(I: I2)) |
1170 | return j; |
1171 | } |
1172 | return i; |
1173 | } |
1174 | |
1175 | /// Emit a tree of add instructions, summing Ops together |
1176 | /// and returning the result. Insert the tree before I. |
1177 | static Value *EmitAddTreeOfValues(BasicBlock::iterator It, |
1178 | SmallVectorImpl<WeakTrackingVH> &Ops) { |
1179 | if (Ops.size() == 1) return Ops.back(); |
1180 | |
1181 | Value *V1 = Ops.pop_back_val(); |
1182 | Value *V2 = EmitAddTreeOfValues(It, Ops); |
1183 | return CreateAdd(S1: V2, S2: V1, Name: "reass.add" , InsertBefore: It, FlagsOp: &*It); |
1184 | } |
1185 | |
1186 | /// If V is an expression tree that is a multiplication sequence, |
1187 | /// and if this sequence contains a multiply by Factor, |
1188 | /// remove Factor from the tree and return the new tree. |
1189 | Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) { |
1190 | BinaryOperator *BO = isReassociableOp(V, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul); |
1191 | if (!BO) |
1192 | return nullptr; |
1193 | |
1194 | SmallVector<RepeatedValue, 8> Tree; |
1195 | bool HasNUW = true; |
1196 | MadeChange |= LinearizeExprTree(I: BO, Ops&: Tree, ToRedo&: RedoInsts, HasNUW); |
1197 | SmallVector<ValueEntry, 8> Factors; |
1198 | Factors.reserve(N: Tree.size()); |
1199 | for (unsigned i = 0, e = Tree.size(); i != e; ++i) { |
1200 | RepeatedValue E = Tree[i]; |
1201 | Factors.append(NumInputs: E.second.getZExtValue(), |
1202 | Elt: ValueEntry(getRank(V: E.first), E.first)); |
1203 | } |
1204 | |
1205 | bool FoundFactor = false; |
1206 | bool NeedsNegate = false; |
1207 | for (unsigned i = 0, e = Factors.size(); i != e; ++i) { |
1208 | if (Factors[i].Op == Factor) { |
1209 | FoundFactor = true; |
1210 | Factors.erase(CI: Factors.begin()+i); |
1211 | break; |
1212 | } |
1213 | |
1214 | // If this is a negative version of this factor, remove it. |
1215 | if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Val: Factor)) { |
1216 | if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Val: Factors[i].Op)) |
1217 | if (FC1->getValue() == -FC2->getValue()) { |
1218 | FoundFactor = NeedsNegate = true; |
1219 | Factors.erase(CI: Factors.begin()+i); |
1220 | break; |
1221 | } |
1222 | } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Val: Factor)) { |
1223 | if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Val: Factors[i].Op)) { |
1224 | const APFloat &F1 = FC1->getValueAPF(); |
1225 | APFloat F2(FC2->getValueAPF()); |
1226 | F2.changeSign(); |
1227 | if (F1 == F2) { |
1228 | FoundFactor = NeedsNegate = true; |
1229 | Factors.erase(CI: Factors.begin() + i); |
1230 | break; |
1231 | } |
1232 | } |
1233 | } |
1234 | } |
1235 | |
1236 | if (!FoundFactor) { |
1237 | // Make sure to restore the operands to the expression tree. |
1238 | RewriteExprTree(I: BO, Ops&: Factors, HasNUW); |
1239 | return nullptr; |
1240 | } |
1241 | |
1242 | BasicBlock::iterator InsertPt = ++BO->getIterator(); |
1243 | |
1244 | // If this was just a single multiply, remove the multiply and return the only |
1245 | // remaining operand. |
1246 | if (Factors.size() == 1) { |
1247 | RedoInsts.insert(X: BO); |
1248 | V = Factors[0].Op; |
1249 | } else { |
1250 | RewriteExprTree(I: BO, Ops&: Factors, HasNUW); |
1251 | V = BO; |
1252 | } |
1253 | |
1254 | if (NeedsNegate) |
1255 | V = CreateNeg(S1: V, Name: "neg" , InsertBefore: InsertPt, FlagsOp: BO); |
1256 | |
1257 | return V; |
1258 | } |
1259 | |
1260 | /// If V is a single-use multiply, recursively add its operands as factors, |
1261 | /// otherwise add V to the list of factors. |
1262 | /// |
1263 | /// Ops is the top-level list of add operands we're trying to factor. |
1264 | static void FindSingleUseMultiplyFactors(Value *V, |
1265 | SmallVectorImpl<Value*> &Factors) { |
1266 | BinaryOperator *BO = isReassociableOp(V, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul); |
1267 | if (!BO) { |
1268 | Factors.push_back(Elt: V); |
1269 | return; |
1270 | } |
1271 | |
1272 | // Otherwise, add the LHS and RHS to the list of factors. |
1273 | FindSingleUseMultiplyFactors(V: BO->getOperand(i_nocapture: 1), Factors); |
1274 | FindSingleUseMultiplyFactors(V: BO->getOperand(i_nocapture: 0), Factors); |
1275 | } |
1276 | |
1277 | /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction. |
1278 | /// This optimizes based on identities. If it can be reduced to a single Value, |
1279 | /// it is returned, otherwise the Ops list is mutated as necessary. |
1280 | static Value *OptimizeAndOrXor(unsigned Opcode, |
1281 | SmallVectorImpl<ValueEntry> &Ops) { |
1282 | // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. |
1283 | // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. |
1284 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { |
1285 | // First, check for X and ~X in the operand list. |
1286 | assert(i < Ops.size()); |
1287 | Value *X; |
1288 | if (match(V: Ops[i].Op, P: m_Not(V: m_Value(V&: X)))) { // Cannot occur for ^. |
1289 | unsigned FoundX = FindInOperandList(Ops, i, X); |
1290 | if (FoundX != i) { |
1291 | if (Opcode == Instruction::And) // ...&X&~X = 0 |
1292 | return Constant::getNullValue(Ty: X->getType()); |
1293 | |
1294 | if (Opcode == Instruction::Or) // ...|X|~X = -1 |
1295 | return Constant::getAllOnesValue(Ty: X->getType()); |
1296 | } |
1297 | } |
1298 | |
1299 | // Next, check for duplicate pairs of values, which we assume are next to |
1300 | // each other, due to our sorting criteria. |
1301 | assert(i < Ops.size()); |
1302 | if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { |
1303 | if (Opcode == Instruction::And || Opcode == Instruction::Or) { |
1304 | // Drop duplicate values for And and Or. |
1305 | Ops.erase(CI: Ops.begin()+i); |
1306 | --i; --e; |
1307 | ++NumAnnihil; |
1308 | continue; |
1309 | } |
1310 | |
1311 | // Drop pairs of values for Xor. |
1312 | assert(Opcode == Instruction::Xor); |
1313 | if (e == 2) |
1314 | return Constant::getNullValue(Ty: Ops[0].Op->getType()); |
1315 | |
1316 | // Y ^ X^X -> Y |
1317 | Ops.erase(CS: Ops.begin()+i, CE: Ops.begin()+i+2); |
1318 | i -= 1; e -= 2; |
1319 | ++NumAnnihil; |
1320 | } |
1321 | } |
1322 | return nullptr; |
1323 | } |
1324 | |
1325 | /// Helper function of CombineXorOpnd(). It creates a bitwise-and |
1326 | /// instruction with the given two operands, and return the resulting |
1327 | /// instruction. There are two special cases: 1) if the constant operand is 0, |
1328 | /// it will return NULL. 2) if the constant is ~0, the symbolic operand will |
1329 | /// be returned. |
1330 | static Value *createAndInstr(BasicBlock::iterator InsertBefore, Value *Opnd, |
1331 | const APInt &ConstOpnd) { |
1332 | if (ConstOpnd.isZero()) |
1333 | return nullptr; |
1334 | |
1335 | if (ConstOpnd.isAllOnes()) |
1336 | return Opnd; |
1337 | |
1338 | Instruction *I = BinaryOperator::CreateAnd( |
1339 | V1: Opnd, V2: ConstantInt::get(Ty: Opnd->getType(), V: ConstOpnd), Name: "and.ra" , |
1340 | It: InsertBefore); |
1341 | I->setDebugLoc(InsertBefore->getDebugLoc()); |
1342 | return I; |
1343 | } |
1344 | |
1345 | // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd" |
1346 | // into "R ^ C", where C would be 0, and R is a symbolic value. |
1347 | // |
1348 | // If it was successful, true is returned, and the "R" and "C" is returned |
1349 | // via "Res" and "ConstOpnd", respectively; otherwise, false is returned, |
1350 | // and both "Res" and "ConstOpnd" remain unchanged. |
1351 | bool ReassociatePass::CombineXorOpnd(BasicBlock::iterator It, XorOpnd *Opnd1, |
1352 | APInt &ConstOpnd, Value *&Res) { |
1353 | // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2 |
1354 | // = ((x | c1) ^ c1) ^ (c1 ^ c2) |
1355 | // = (x & ~c1) ^ (c1 ^ c2) |
1356 | // It is useful only when c1 == c2. |
1357 | if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isZero()) |
1358 | return false; |
1359 | |
1360 | if (!Opnd1->getValue()->hasOneUse()) |
1361 | return false; |
1362 | |
1363 | const APInt &C1 = Opnd1->getConstPart(); |
1364 | if (C1 != ConstOpnd) |
1365 | return false; |
1366 | |
1367 | Value *X = Opnd1->getSymbolicPart(); |
1368 | Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: ~C1); |
1369 | // ConstOpnd was C2, now C1 ^ C2. |
1370 | ConstOpnd ^= C1; |
1371 | |
1372 | if (Instruction *T = dyn_cast<Instruction>(Val: Opnd1->getValue())) |
1373 | RedoInsts.insert(X: T); |
1374 | return true; |
1375 | } |
1376 | |
1377 | // Helper function of OptimizeXor(). It tries to simplify |
1378 | // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a |
1379 | // symbolic value. |
1380 | // |
1381 | // If it was successful, true is returned, and the "R" and "C" is returned |
1382 | // via "Res" and "ConstOpnd", respectively (If the entire expression is |
1383 | // evaluated to a constant, the Res is set to NULL); otherwise, false is |
1384 | // returned, and both "Res" and "ConstOpnd" remain unchanged. |
1385 | bool ReassociatePass::CombineXorOpnd(BasicBlock::iterator It, XorOpnd *Opnd1, |
1386 | XorOpnd *Opnd2, APInt &ConstOpnd, |
1387 | Value *&Res) { |
1388 | Value *X = Opnd1->getSymbolicPart(); |
1389 | if (X != Opnd2->getSymbolicPart()) |
1390 | return false; |
1391 | |
1392 | // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.) |
1393 | int DeadInstNum = 1; |
1394 | if (Opnd1->getValue()->hasOneUse()) |
1395 | DeadInstNum++; |
1396 | if (Opnd2->getValue()->hasOneUse()) |
1397 | DeadInstNum++; |
1398 | |
1399 | // Xor-Rule 2: |
1400 | // (x | c1) ^ (x & c2) |
1401 | // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1 |
1402 | // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1 |
1403 | // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3 |
1404 | // |
1405 | if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) { |
1406 | if (Opnd2->isOrExpr()) |
1407 | std::swap(a&: Opnd1, b&: Opnd2); |
1408 | |
1409 | const APInt &C1 = Opnd1->getConstPart(); |
1410 | const APInt &C2 = Opnd2->getConstPart(); |
1411 | APInt C3((~C1) ^ C2); |
1412 | |
1413 | // Do not increase code size! |
1414 | if (!C3.isZero() && !C3.isAllOnes()) { |
1415 | int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2; |
1416 | if (NewInstNum > DeadInstNum) |
1417 | return false; |
1418 | } |
1419 | |
1420 | Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: C3); |
1421 | ConstOpnd ^= C1; |
1422 | } else if (Opnd1->isOrExpr()) { |
1423 | // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2 |
1424 | // |
1425 | const APInt &C1 = Opnd1->getConstPart(); |
1426 | const APInt &C2 = Opnd2->getConstPart(); |
1427 | APInt C3 = C1 ^ C2; |
1428 | |
1429 | // Do not increase code size |
1430 | if (!C3.isZero() && !C3.isAllOnes()) { |
1431 | int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2; |
1432 | if (NewInstNum > DeadInstNum) |
1433 | return false; |
1434 | } |
1435 | |
1436 | Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: C3); |
1437 | ConstOpnd ^= C3; |
1438 | } else { |
1439 | // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2)) |
1440 | // |
1441 | const APInt &C1 = Opnd1->getConstPart(); |
1442 | const APInt &C2 = Opnd2->getConstPart(); |
1443 | APInt C3 = C1 ^ C2; |
1444 | Res = createAndInstr(InsertBefore: It, Opnd: X, ConstOpnd: C3); |
1445 | } |
1446 | |
1447 | // Put the original operands in the Redo list; hope they will be deleted |
1448 | // as dead code. |
1449 | if (Instruction *T = dyn_cast<Instruction>(Val: Opnd1->getValue())) |
1450 | RedoInsts.insert(X: T); |
1451 | if (Instruction *T = dyn_cast<Instruction>(Val: Opnd2->getValue())) |
1452 | RedoInsts.insert(X: T); |
1453 | |
1454 | return true; |
1455 | } |
1456 | |
1457 | /// Optimize a series of operands to an 'xor' instruction. If it can be reduced |
1458 | /// to a single Value, it is returned, otherwise the Ops list is mutated as |
1459 | /// necessary. |
1460 | Value *ReassociatePass::OptimizeXor(Instruction *I, |
1461 | SmallVectorImpl<ValueEntry> &Ops) { |
1462 | if (Value *V = OptimizeAndOrXor(Opcode: Instruction::Xor, Ops)) |
1463 | return V; |
1464 | |
1465 | if (Ops.size() == 1) |
1466 | return nullptr; |
1467 | |
1468 | SmallVector<XorOpnd, 8> Opnds; |
1469 | SmallVector<XorOpnd*, 8> OpndPtrs; |
1470 | Type *Ty = Ops[0].Op->getType(); |
1471 | APInt ConstOpnd(Ty->getScalarSizeInBits(), 0); |
1472 | |
1473 | // Step 1: Convert ValueEntry to XorOpnd |
1474 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { |
1475 | Value *V = Ops[i].Op; |
1476 | const APInt *C; |
1477 | // TODO: Support non-splat vectors. |
1478 | if (match(V, P: m_APInt(Res&: C))) { |
1479 | ConstOpnd ^= *C; |
1480 | } else { |
1481 | XorOpnd O(V); |
1482 | O.setSymbolicRank(getRank(V: O.getSymbolicPart())); |
1483 | Opnds.push_back(Elt: O); |
1484 | } |
1485 | } |
1486 | |
1487 | // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds". |
1488 | // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate |
1489 | // the "OpndPtrs" as well. For the similar reason, do not fuse this loop |
1490 | // with the previous loop --- the iterator of the "Opnds" may be invalidated |
1491 | // when new elements are added to the vector. |
1492 | for (unsigned i = 0, e = Opnds.size(); i != e; ++i) |
1493 | OpndPtrs.push_back(Elt: &Opnds[i]); |
1494 | |
1495 | // Step 2: Sort the Xor-Operands in a way such that the operands containing |
1496 | // the same symbolic value cluster together. For instance, the input operand |
1497 | // sequence ("x | 123", "y & 456", "x & 789") will be sorted into: |
1498 | // ("x | 123", "x & 789", "y & 456"). |
1499 | // |
1500 | // The purpose is twofold: |
1501 | // 1) Cluster together the operands sharing the same symbolic-value. |
1502 | // 2) Operand having smaller symbolic-value-rank is permuted earlier, which |
1503 | // could potentially shorten crital path, and expose more loop-invariants. |
1504 | // Note that values' rank are basically defined in RPO order (FIXME). |
1505 | // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier |
1506 | // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2", |
1507 | // "z" in the order of X-Y-Z is better than any other orders. |
1508 | llvm::stable_sort(Range&: OpndPtrs, C: [](XorOpnd *LHS, XorOpnd *RHS) { |
1509 | return LHS->getSymbolicRank() < RHS->getSymbolicRank(); |
1510 | }); |
1511 | |
1512 | // Step 3: Combine adjacent operands |
1513 | XorOpnd *PrevOpnd = nullptr; |
1514 | bool Changed = false; |
1515 | for (unsigned i = 0, e = Opnds.size(); i < e; i++) { |
1516 | XorOpnd *CurrOpnd = OpndPtrs[i]; |
1517 | // The combined value |
1518 | Value *CV; |
1519 | |
1520 | // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd" |
1521 | if (!ConstOpnd.isZero() && |
1522 | CombineXorOpnd(It: I->getIterator(), Opnd1: CurrOpnd, ConstOpnd, Res&: CV)) { |
1523 | Changed = true; |
1524 | if (CV) |
1525 | *CurrOpnd = XorOpnd(CV); |
1526 | else { |
1527 | CurrOpnd->Invalidate(); |
1528 | continue; |
1529 | } |
1530 | } |
1531 | |
1532 | if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) { |
1533 | PrevOpnd = CurrOpnd; |
1534 | continue; |
1535 | } |
1536 | |
1537 | // step 3.2: When previous and current operands share the same symbolic |
1538 | // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd" |
1539 | if (CombineXorOpnd(It: I->getIterator(), Opnd1: CurrOpnd, Opnd2: PrevOpnd, ConstOpnd, Res&: CV)) { |
1540 | // Remove previous operand |
1541 | PrevOpnd->Invalidate(); |
1542 | if (CV) { |
1543 | *CurrOpnd = XorOpnd(CV); |
1544 | PrevOpnd = CurrOpnd; |
1545 | } else { |
1546 | CurrOpnd->Invalidate(); |
1547 | PrevOpnd = nullptr; |
1548 | } |
1549 | Changed = true; |
1550 | } |
1551 | } |
1552 | |
1553 | // Step 4: Reassemble the Ops |
1554 | if (Changed) { |
1555 | Ops.clear(); |
1556 | for (const XorOpnd &O : Opnds) { |
1557 | if (O.isInvalid()) |
1558 | continue; |
1559 | ValueEntry VE(getRank(V: O.getValue()), O.getValue()); |
1560 | Ops.push_back(Elt: VE); |
1561 | } |
1562 | if (!ConstOpnd.isZero()) { |
1563 | Value *C = ConstantInt::get(Ty, V: ConstOpnd); |
1564 | ValueEntry VE(getRank(V: C), C); |
1565 | Ops.push_back(Elt: VE); |
1566 | } |
1567 | unsigned Sz = Ops.size(); |
1568 | if (Sz == 1) |
1569 | return Ops.back().Op; |
1570 | if (Sz == 0) { |
1571 | assert(ConstOpnd.isZero()); |
1572 | return ConstantInt::get(Ty, V: ConstOpnd); |
1573 | } |
1574 | } |
1575 | |
1576 | return nullptr; |
1577 | } |
1578 | |
1579 | /// Optimize a series of operands to an 'add' instruction. This |
1580 | /// optimizes based on identities. If it can be reduced to a single Value, it |
1581 | /// is returned, otherwise the Ops list is mutated as necessary. |
1582 | Value *ReassociatePass::OptimizeAdd(Instruction *I, |
1583 | SmallVectorImpl<ValueEntry> &Ops) { |
1584 | // Scan the operand lists looking for X and -X pairs. If we find any, we |
1585 | // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it, |
1586 | // scan for any |
1587 | // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. |
1588 | |
1589 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { |
1590 | Value *TheOp = Ops[i].Op; |
1591 | // Check to see if we've seen this operand before. If so, we factor all |
1592 | // instances of the operand together. Due to our sorting criteria, we know |
1593 | // that these need to be next to each other in the vector. |
1594 | if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { |
1595 | // Rescan the list, remove all instances of this operand from the expr. |
1596 | unsigned NumFound = 0; |
1597 | do { |
1598 | Ops.erase(CI: Ops.begin()+i); |
1599 | ++NumFound; |
1600 | } while (i != Ops.size() && Ops[i].Op == TheOp); |
1601 | |
1602 | LLVM_DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp |
1603 | << '\n'); |
1604 | ++NumFactor; |
1605 | |
1606 | // Insert a new multiply. |
1607 | Type *Ty = TheOp->getType(); |
1608 | Constant *C = Ty->isIntOrIntVectorTy() ? |
1609 | ConstantInt::get(Ty, V: NumFound) : ConstantFP::get(Ty, V: NumFound); |
1610 | Instruction *Mul = CreateMul(S1: TheOp, S2: C, Name: "factor" , InsertBefore: I->getIterator(), FlagsOp: I); |
1611 | |
1612 | // Now that we have inserted a multiply, optimize it. This allows us to |
1613 | // handle cases that require multiple factoring steps, such as this: |
1614 | // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 |
1615 | RedoInsts.insert(X: Mul); |
1616 | |
1617 | // If every add operand was a duplicate, return the multiply. |
1618 | if (Ops.empty()) |
1619 | return Mul; |
1620 | |
1621 | // Otherwise, we had some input that didn't have the dupe, such as |
1622 | // "A + A + B" -> "A*2 + B". Add the new multiply to the list of |
1623 | // things being added by this operation. |
1624 | Ops.insert(I: Ops.begin(), Elt: ValueEntry(getRank(V: Mul), Mul)); |
1625 | |
1626 | --i; |
1627 | e = Ops.size(); |
1628 | continue; |
1629 | } |
1630 | |
1631 | // Check for X and -X or X and ~X in the operand list. |
1632 | Value *X; |
1633 | if (!match(V: TheOp, P: m_Neg(V: m_Value(V&: X))) && !match(V: TheOp, P: m_Not(V: m_Value(V&: X))) && |
1634 | !match(V: TheOp, P: m_FNeg(X: m_Value(V&: X)))) |
1635 | continue; |
1636 | |
1637 | unsigned FoundX = FindInOperandList(Ops, i, X); |
1638 | if (FoundX == i) |
1639 | continue; |
1640 | |
1641 | // Remove X and -X from the operand list. |
1642 | if (Ops.size() == 2 && |
1643 | (match(V: TheOp, P: m_Neg(V: m_Value())) || match(V: TheOp, P: m_FNeg(X: m_Value())))) |
1644 | return Constant::getNullValue(Ty: X->getType()); |
1645 | |
1646 | // Remove X and ~X from the operand list. |
1647 | if (Ops.size() == 2 && match(V: TheOp, P: m_Not(V: m_Value()))) |
1648 | return Constant::getAllOnesValue(Ty: X->getType()); |
1649 | |
1650 | Ops.erase(CI: Ops.begin()+i); |
1651 | if (i < FoundX) |
1652 | --FoundX; |
1653 | else |
1654 | --i; // Need to back up an extra one. |
1655 | Ops.erase(CI: Ops.begin()+FoundX); |
1656 | ++NumAnnihil; |
1657 | --i; // Revisit element. |
1658 | e -= 2; // Removed two elements. |
1659 | |
1660 | // if X and ~X we append -1 to the operand list. |
1661 | if (match(V: TheOp, P: m_Not(V: m_Value()))) { |
1662 | Value *V = Constant::getAllOnesValue(Ty: X->getType()); |
1663 | Ops.insert(I: Ops.end(), Elt: ValueEntry(getRank(V), V)); |
1664 | e += 1; |
1665 | } |
1666 | } |
1667 | |
1668 | // Scan the operand list, checking to see if there are any common factors |
1669 | // between operands. Consider something like A*A+A*B*C+D. We would like to |
1670 | // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. |
1671 | // To efficiently find this, we count the number of times a factor occurs |
1672 | // for any ADD operands that are MULs. |
1673 | DenseMap<Value*, unsigned> FactorOccurrences; |
1674 | |
1675 | // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) |
1676 | // where they are actually the same multiply. |
1677 | unsigned MaxOcc = 0; |
1678 | Value *MaxOccVal = nullptr; |
1679 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { |
1680 | BinaryOperator *BOp = |
1681 | isReassociableOp(V: Ops[i].Op, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul); |
1682 | if (!BOp) |
1683 | continue; |
1684 | |
1685 | // Compute all of the factors of this added value. |
1686 | SmallVector<Value*, 8> Factors; |
1687 | FindSingleUseMultiplyFactors(V: BOp, Factors); |
1688 | assert(Factors.size() > 1 && "Bad linearize!" ); |
1689 | |
1690 | // Add one to FactorOccurrences for each unique factor in this op. |
1691 | SmallPtrSet<Value*, 8> Duplicates; |
1692 | for (Value *Factor : Factors) { |
1693 | if (!Duplicates.insert(Ptr: Factor).second) |
1694 | continue; |
1695 | |
1696 | unsigned Occ = ++FactorOccurrences[Factor]; |
1697 | if (Occ > MaxOcc) { |
1698 | MaxOcc = Occ; |
1699 | MaxOccVal = Factor; |
1700 | } |
1701 | |
1702 | // If Factor is a negative constant, add the negated value as a factor |
1703 | // because we can percolate the negate out. Watch for minint, which |
1704 | // cannot be positivified. |
1705 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Factor)) { |
1706 | if (CI->isNegative() && !CI->isMinValue(IsSigned: true)) { |
1707 | Factor = ConstantInt::get(Context&: CI->getContext(), V: -CI->getValue()); |
1708 | if (!Duplicates.insert(Ptr: Factor).second) |
1709 | continue; |
1710 | unsigned Occ = ++FactorOccurrences[Factor]; |
1711 | if (Occ > MaxOcc) { |
1712 | MaxOcc = Occ; |
1713 | MaxOccVal = Factor; |
1714 | } |
1715 | } |
1716 | } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Val: Factor)) { |
1717 | if (CF->isNegative()) { |
1718 | APFloat F(CF->getValueAPF()); |
1719 | F.changeSign(); |
1720 | Factor = ConstantFP::get(Context&: CF->getContext(), V: F); |
1721 | if (!Duplicates.insert(Ptr: Factor).second) |
1722 | continue; |
1723 | unsigned Occ = ++FactorOccurrences[Factor]; |
1724 | if (Occ > MaxOcc) { |
1725 | MaxOcc = Occ; |
1726 | MaxOccVal = Factor; |
1727 | } |
1728 | } |
1729 | } |
1730 | } |
1731 | } |
1732 | |
1733 | // If any factor occurred more than one time, we can pull it out. |
1734 | if (MaxOcc > 1) { |
1735 | LLVM_DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal |
1736 | << '\n'); |
1737 | ++NumFactor; |
1738 | |
1739 | // Create a new instruction that uses the MaxOccVal twice. If we don't do |
1740 | // this, we could otherwise run into situations where removing a factor |
1741 | // from an expression will drop a use of maxocc, and this can cause |
1742 | // RemoveFactorFromExpression on successive values to behave differently. |
1743 | Instruction *DummyInst = |
1744 | I->getType()->isIntOrIntVectorTy() |
1745 | ? BinaryOperator::CreateAdd(V1: MaxOccVal, V2: MaxOccVal) |
1746 | : BinaryOperator::CreateFAdd(V1: MaxOccVal, V2: MaxOccVal); |
1747 | |
1748 | SmallVector<WeakTrackingVH, 4> NewMulOps; |
1749 | for (unsigned i = 0; i != Ops.size(); ++i) { |
1750 | // Only try to remove factors from expressions we're allowed to. |
1751 | BinaryOperator *BOp = |
1752 | isReassociableOp(V: Ops[i].Op, Opcode1: Instruction::Mul, Opcode2: Instruction::FMul); |
1753 | if (!BOp) |
1754 | continue; |
1755 | |
1756 | if (Value *V = RemoveFactorFromExpression(V: Ops[i].Op, Factor: MaxOccVal)) { |
1757 | // The factorized operand may occur several times. Convert them all in |
1758 | // one fell swoop. |
1759 | for (unsigned j = Ops.size(); j != i;) { |
1760 | --j; |
1761 | if (Ops[j].Op == Ops[i].Op) { |
1762 | NewMulOps.push_back(Elt: V); |
1763 | Ops.erase(CI: Ops.begin()+j); |
1764 | } |
1765 | } |
1766 | --i; |
1767 | } |
1768 | } |
1769 | |
1770 | // No need for extra uses anymore. |
1771 | DummyInst->deleteValue(); |
1772 | |
1773 | unsigned NumAddedValues = NewMulOps.size(); |
1774 | Value *V = EmitAddTreeOfValues(It: I->getIterator(), Ops&: NewMulOps); |
1775 | |
1776 | // Now that we have inserted the add tree, optimize it. This allows us to |
1777 | // handle cases that require multiple factoring steps, such as this: |
1778 | // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) |
1779 | assert(NumAddedValues > 1 && "Each occurrence should contribute a value" ); |
1780 | (void)NumAddedValues; |
1781 | if (Instruction *VI = dyn_cast<Instruction>(Val: V)) |
1782 | RedoInsts.insert(X: VI); |
1783 | |
1784 | // Create the multiply. |
1785 | Instruction *V2 = CreateMul(S1: V, S2: MaxOccVal, Name: "reass.mul" , InsertBefore: I->getIterator(), FlagsOp: I); |
1786 | |
1787 | // Rerun associate on the multiply in case the inner expression turned into |
1788 | // a multiply. We want to make sure that we keep things in canonical form. |
1789 | RedoInsts.insert(X: V2); |
1790 | |
1791 | // If every add operand included the factor (e.g. "A*B + A*C"), then the |
1792 | // entire result expression is just the multiply "A*(B+C)". |
1793 | if (Ops.empty()) |
1794 | return V2; |
1795 | |
1796 | // Otherwise, we had some input that didn't have the factor, such as |
1797 | // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of |
1798 | // things being added by this operation. |
1799 | Ops.insert(I: Ops.begin(), Elt: ValueEntry(getRank(V: V2), V2)); |
1800 | } |
1801 | |
1802 | return nullptr; |
1803 | } |
1804 | |
1805 | /// Build up a vector of value/power pairs factoring a product. |
1806 | /// |
1807 | /// Given a series of multiplication operands, build a vector of factors and |
1808 | /// the powers each is raised to when forming the final product. Sort them in |
1809 | /// the order of descending power. |
1810 | /// |
1811 | /// (x*x) -> [(x, 2)] |
1812 | /// ((x*x)*x) -> [(x, 3)] |
1813 | /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)] |
1814 | /// |
1815 | /// \returns Whether any factors have a power greater than one. |
1816 | static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, |
1817 | SmallVectorImpl<Factor> &Factors) { |
1818 | // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this. |
1819 | // Compute the sum of powers of simplifiable factors. |
1820 | unsigned FactorPowerSum = 0; |
1821 | for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) { |
1822 | Value *Op = Ops[Idx-1].Op; |
1823 | |
1824 | // Count the number of occurrences of this value. |
1825 | unsigned Count = 1; |
1826 | for (; Idx < Size && Ops[Idx].Op == Op; ++Idx) |
1827 | ++Count; |
1828 | // Track for simplification all factors which occur 2 or more times. |
1829 | if (Count > 1) |
1830 | FactorPowerSum += Count; |
1831 | } |
1832 | |
1833 | // We can only simplify factors if the sum of the powers of our simplifiable |
1834 | // factors is 4 or higher. When that is the case, we will *always* have |
1835 | // a simplification. This is an important invariant to prevent cyclicly |
1836 | // trying to simplify already minimal formations. |
1837 | if (FactorPowerSum < 4) |
1838 | return false; |
1839 | |
1840 | // Now gather the simplifiable factors, removing them from Ops. |
1841 | FactorPowerSum = 0; |
1842 | for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) { |
1843 | Value *Op = Ops[Idx-1].Op; |
1844 | |
1845 | // Count the number of occurrences of this value. |
1846 | unsigned Count = 1; |
1847 | for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx) |
1848 | ++Count; |
1849 | if (Count == 1) |
1850 | continue; |
1851 | // Move an even number of occurrences to Factors. |
1852 | Count &= ~1U; |
1853 | Idx -= Count; |
1854 | FactorPowerSum += Count; |
1855 | Factors.push_back(Elt: Factor(Op, Count)); |
1856 | Ops.erase(CS: Ops.begin()+Idx, CE: Ops.begin()+Idx+Count); |
1857 | } |
1858 | |
1859 | // None of the adjustments above should have reduced the sum of factor powers |
1860 | // below our mininum of '4'. |
1861 | assert(FactorPowerSum >= 4); |
1862 | |
1863 | llvm::stable_sort(Range&: Factors, C: [](const Factor &LHS, const Factor &RHS) { |
1864 | return LHS.Power > RHS.Power; |
1865 | }); |
1866 | return true; |
1867 | } |
1868 | |
1869 | /// Build a tree of multiplies, computing the product of Ops. |
1870 | static Value *buildMultiplyTree(IRBuilderBase &Builder, |
1871 | SmallVectorImpl<Value*> &Ops) { |
1872 | if (Ops.size() == 1) |
1873 | return Ops.back(); |
1874 | |
1875 | Value *LHS = Ops.pop_back_val(); |
1876 | do { |
1877 | if (LHS->getType()->isIntOrIntVectorTy()) |
1878 | LHS = Builder.CreateMul(LHS, RHS: Ops.pop_back_val()); |
1879 | else |
1880 | LHS = Builder.CreateFMul(L: LHS, R: Ops.pop_back_val()); |
1881 | } while (!Ops.empty()); |
1882 | |
1883 | return LHS; |
1884 | } |
1885 | |
1886 | /// Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*... |
1887 | /// |
1888 | /// Given a vector of values raised to various powers, where no two values are |
1889 | /// equal and the powers are sorted in decreasing order, compute the minimal |
1890 | /// DAG of multiplies to compute the final product, and return that product |
1891 | /// value. |
1892 | Value * |
1893 | ReassociatePass::buildMinimalMultiplyDAG(IRBuilderBase &Builder, |
1894 | SmallVectorImpl<Factor> &Factors) { |
1895 | assert(Factors[0].Power); |
1896 | SmallVector<Value *, 4> OuterProduct; |
1897 | for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size(); |
1898 | Idx < Size && Factors[Idx].Power > 0; ++Idx) { |
1899 | if (Factors[Idx].Power != Factors[LastIdx].Power) { |
1900 | LastIdx = Idx; |
1901 | continue; |
1902 | } |
1903 | |
1904 | // We want to multiply across all the factors with the same power so that |
1905 | // we can raise them to that power as a single entity. Build a mini tree |
1906 | // for that. |
1907 | SmallVector<Value *, 4> InnerProduct; |
1908 | InnerProduct.push_back(Elt: Factors[LastIdx].Base); |
1909 | do { |
1910 | InnerProduct.push_back(Elt: Factors[Idx].Base); |
1911 | ++Idx; |
1912 | } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power); |
1913 | |
1914 | // Reset the base value of the first factor to the new expression tree. |
1915 | // We'll remove all the factors with the same power in a second pass. |
1916 | Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, Ops&: InnerProduct); |
1917 | if (Instruction *MI = dyn_cast<Instruction>(Val: M)) |
1918 | RedoInsts.insert(X: MI); |
1919 | |
1920 | LastIdx = Idx; |
1921 | } |
1922 | // Unique factors with equal powers -- we've folded them into the first one's |
1923 | // base. |
1924 | Factors.erase(CS: std::unique(first: Factors.begin(), last: Factors.end(), |
1925 | binary_pred: [](const Factor &LHS, const Factor &RHS) { |
1926 | return LHS.Power == RHS.Power; |
1927 | }), |
1928 | CE: Factors.end()); |
1929 | |
1930 | // Iteratively collect the base of each factor with an add power into the |
1931 | // outer product, and halve each power in preparation for squaring the |
1932 | // expression. |
1933 | for (Factor &F : Factors) { |
1934 | if (F.Power & 1) |
1935 | OuterProduct.push_back(Elt: F.Base); |
1936 | F.Power >>= 1; |
1937 | } |
1938 | if (Factors[0].Power) { |
1939 | Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors); |
1940 | OuterProduct.push_back(Elt: SquareRoot); |
1941 | OuterProduct.push_back(Elt: SquareRoot); |
1942 | } |
1943 | if (OuterProduct.size() == 1) |
1944 | return OuterProduct.front(); |
1945 | |
1946 | Value *V = buildMultiplyTree(Builder, Ops&: OuterProduct); |
1947 | return V; |
1948 | } |
1949 | |
1950 | Value *ReassociatePass::OptimizeMul(BinaryOperator *I, |
1951 | SmallVectorImpl<ValueEntry> &Ops) { |
1952 | // We can only optimize the multiplies when there is a chain of more than |
1953 | // three, such that a balanced tree might require fewer total multiplies. |
1954 | if (Ops.size() < 4) |
1955 | return nullptr; |
1956 | |
1957 | // Try to turn linear trees of multiplies without other uses of the |
1958 | // intermediate stages into minimal multiply DAGs with perfect sub-expression |
1959 | // re-use. |
1960 | SmallVector<Factor, 4> Factors; |
1961 | if (!collectMultiplyFactors(Ops, Factors)) |
1962 | return nullptr; // All distinct factors, so nothing left for us to do. |
1963 | |
1964 | IRBuilder<> Builder(I); |
1965 | // The reassociate transformation for FP operations is performed only |
1966 | // if unsafe algebra is permitted by FastMathFlags. Propagate those flags |
1967 | // to the newly generated operations. |
1968 | if (auto FPI = dyn_cast<FPMathOperator>(Val: I)) |
1969 | Builder.setFastMathFlags(FPI->getFastMathFlags()); |
1970 | |
1971 | Value *V = buildMinimalMultiplyDAG(Builder, Factors); |
1972 | if (Ops.empty()) |
1973 | return V; |
1974 | |
1975 | ValueEntry NewEntry = ValueEntry(getRank(V), V); |
1976 | Ops.insert(I: llvm::lower_bound(Range&: Ops, Value&: NewEntry), Elt: NewEntry); |
1977 | return nullptr; |
1978 | } |
1979 | |
1980 | Value *ReassociatePass::OptimizeExpression(BinaryOperator *I, |
1981 | SmallVectorImpl<ValueEntry> &Ops) { |
1982 | // Now that we have the linearized expression tree, try to optimize it. |
1983 | // Start by folding any constants that we found. |
1984 | const DataLayout &DL = I->getModule()->getDataLayout(); |
1985 | Constant *Cst = nullptr; |
1986 | unsigned Opcode = I->getOpcode(); |
1987 | while (!Ops.empty()) { |
1988 | if (auto *C = dyn_cast<Constant>(Val: Ops.back().Op)) { |
1989 | if (!Cst) { |
1990 | Ops.pop_back(); |
1991 | Cst = C; |
1992 | continue; |
1993 | } |
1994 | if (Constant *Res = ConstantFoldBinaryOpOperands(Opcode, LHS: C, RHS: Cst, DL)) { |
1995 | Ops.pop_back(); |
1996 | Cst = Res; |
1997 | continue; |
1998 | } |
1999 | } |
2000 | break; |
2001 | } |
2002 | // If there was nothing but constants then we are done. |
2003 | if (Ops.empty()) |
2004 | return Cst; |
2005 | |
2006 | // Put the combined constant back at the end of the operand list, except if |
2007 | // there is no point. For example, an add of 0 gets dropped here, while a |
2008 | // multiplication by zero turns the whole expression into zero. |
2009 | if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, Ty: I->getType())) { |
2010 | if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, Ty: I->getType())) |
2011 | return Cst; |
2012 | Ops.push_back(Elt: ValueEntry(0, Cst)); |
2013 | } |
2014 | |
2015 | if (Ops.size() == 1) return Ops[0].Op; |
2016 | |
2017 | // Handle destructive annihilation due to identities between elements in the |
2018 | // argument list here. |
2019 | unsigned NumOps = Ops.size(); |
2020 | switch (Opcode) { |
2021 | default: break; |
2022 | case Instruction::And: |
2023 | case Instruction::Or: |
2024 | if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) |
2025 | return Result; |
2026 | break; |
2027 | |
2028 | case Instruction::Xor: |
2029 | if (Value *Result = OptimizeXor(I, Ops)) |
2030 | return Result; |
2031 | break; |
2032 | |
2033 | case Instruction::Add: |
2034 | case Instruction::FAdd: |
2035 | if (Value *Result = OptimizeAdd(I, Ops)) |
2036 | return Result; |
2037 | break; |
2038 | |
2039 | case Instruction::Mul: |
2040 | case Instruction::FMul: |
2041 | if (Value *Result = OptimizeMul(I, Ops)) |
2042 | return Result; |
2043 | break; |
2044 | } |
2045 | |
2046 | if (Ops.size() != NumOps) |
2047 | return OptimizeExpression(I, Ops); |
2048 | return nullptr; |
2049 | } |
2050 | |
2051 | // Remove dead instructions and if any operands are trivially dead add them to |
2052 | // Insts so they will be removed as well. |
2053 | void ReassociatePass::RecursivelyEraseDeadInsts(Instruction *I, |
2054 | OrderedSet &Insts) { |
2055 | assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!" ); |
2056 | SmallVector<Value *, 4> Ops(I->operands()); |
2057 | ValueRankMap.erase(Val: I); |
2058 | Insts.remove(X: I); |
2059 | RedoInsts.remove(X: I); |
2060 | llvm::salvageDebugInfo(I&: *I); |
2061 | I->eraseFromParent(); |
2062 | for (auto *Op : Ops) |
2063 | if (Instruction *OpInst = dyn_cast<Instruction>(Val: Op)) |
2064 | if (OpInst->use_empty()) |
2065 | Insts.insert(X: OpInst); |
2066 | } |
2067 | |
2068 | /// Zap the given instruction, adding interesting operands to the work list. |
2069 | void ReassociatePass::EraseInst(Instruction *I) { |
2070 | assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!" ); |
2071 | LLVM_DEBUG(dbgs() << "Erasing dead inst: " ; I->dump()); |
2072 | |
2073 | SmallVector<Value *, 8> Ops(I->operands()); |
2074 | // Erase the dead instruction. |
2075 | ValueRankMap.erase(Val: I); |
2076 | RedoInsts.remove(X: I); |
2077 | llvm::salvageDebugInfo(I&: *I); |
2078 | I->eraseFromParent(); |
2079 | // Optimize its operands. |
2080 | SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes. |
2081 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) |
2082 | if (Instruction *Op = dyn_cast<Instruction>(Val: Ops[i])) { |
2083 | // If this is a node in an expression tree, climb to the expression root |
2084 | // and add that since that's where optimization actually happens. |
2085 | unsigned Opcode = Op->getOpcode(); |
2086 | while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode && |
2087 | Visited.insert(Ptr: Op).second) |
2088 | Op = Op->user_back(); |
2089 | |
2090 | // The instruction we're going to push may be coming from a |
2091 | // dead block, and Reassociate skips the processing of unreachable |
2092 | // blocks because it's a waste of time and also because it can |
2093 | // lead to infinite loop due to LLVM's non-standard definition |
2094 | // of dominance. |
2095 | if (ValueRankMap.contains(Val: Op)) |
2096 | RedoInsts.insert(X: Op); |
2097 | } |
2098 | |
2099 | MadeChange = true; |
2100 | } |
2101 | |
2102 | /// Recursively analyze an expression to build a list of instructions that have |
2103 | /// negative floating-point constant operands. The caller can then transform |
2104 | /// the list to create positive constants for better reassociation and CSE. |
2105 | static void getNegatibleInsts(Value *V, |
2106 | SmallVectorImpl<Instruction *> &Candidates) { |
2107 | // Handle only one-use instructions. Combining negations does not justify |
2108 | // replicating instructions. |
2109 | Instruction *I; |
2110 | if (!match(V, P: m_OneUse(SubPattern: m_Instruction(I)))) |
2111 | return; |
2112 | |
2113 | // Handle expressions of multiplications and divisions. |
2114 | // TODO: This could look through floating-point casts. |
2115 | const APFloat *C; |
2116 | switch (I->getOpcode()) { |
2117 | case Instruction::FMul: |
2118 | // Not expecting non-canonical code here. Bail out and wait. |
2119 | if (match(V: I->getOperand(i: 0), P: m_Constant())) |
2120 | break; |
2121 | |
2122 | if (match(V: I->getOperand(i: 1), P: m_APFloat(Res&: C)) && C->isNegative()) { |
2123 | Candidates.push_back(Elt: I); |
2124 | LLVM_DEBUG(dbgs() << "FMul with negative constant: " << *I << '\n'); |
2125 | } |
2126 | getNegatibleInsts(V: I->getOperand(i: 0), Candidates); |
2127 | getNegatibleInsts(V: I->getOperand(i: 1), Candidates); |
2128 | break; |
2129 | case Instruction::FDiv: |
2130 | // Not expecting non-canonical code here. Bail out and wait. |
2131 | if (match(V: I->getOperand(i: 0), P: m_Constant()) && |
2132 | match(V: I->getOperand(i: 1), P: m_Constant())) |
2133 | break; |
2134 | |
2135 | if ((match(V: I->getOperand(i: 0), P: m_APFloat(Res&: C)) && C->isNegative()) || |
2136 | (match(V: I->getOperand(i: 1), P: m_APFloat(Res&: C)) && C->isNegative())) { |
2137 | Candidates.push_back(Elt: I); |
2138 | LLVM_DEBUG(dbgs() << "FDiv with negative constant: " << *I << '\n'); |
2139 | } |
2140 | getNegatibleInsts(V: I->getOperand(i: 0), Candidates); |
2141 | getNegatibleInsts(V: I->getOperand(i: 1), Candidates); |
2142 | break; |
2143 | default: |
2144 | break; |
2145 | } |
2146 | } |
2147 | |
2148 | /// Given an fadd/fsub with an operand that is a one-use instruction |
2149 | /// (the fadd/fsub), try to change negative floating-point constants into |
2150 | /// positive constants to increase potential for reassociation and CSE. |
2151 | Instruction *ReassociatePass::canonicalizeNegFPConstantsForOp(Instruction *I, |
2152 | Instruction *Op, |
2153 | Value *OtherOp) { |
2154 | assert((I->getOpcode() == Instruction::FAdd || |
2155 | I->getOpcode() == Instruction::FSub) && "Expected fadd/fsub" ); |
2156 | |
2157 | // Collect instructions with negative FP constants from the subtree that ends |
2158 | // in Op. |
2159 | SmallVector<Instruction *, 4> Candidates; |
2160 | getNegatibleInsts(V: Op, Candidates); |
2161 | if (Candidates.empty()) |
2162 | return nullptr; |
2163 | |
2164 | // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the |
2165 | // resulting subtract will be broken up later. This can get us into an |
2166 | // infinite loop during reassociation. |
2167 | bool IsFSub = I->getOpcode() == Instruction::FSub; |
2168 | bool NeedsSubtract = !IsFSub && Candidates.size() % 2 == 1; |
2169 | if (NeedsSubtract && ShouldBreakUpSubtract(Sub: I)) |
2170 | return nullptr; |
2171 | |
2172 | for (Instruction *Negatible : Candidates) { |
2173 | const APFloat *C; |
2174 | if (match(V: Negatible->getOperand(i: 0), P: m_APFloat(Res&: C))) { |
2175 | assert(!match(Negatible->getOperand(1), m_Constant()) && |
2176 | "Expecting only 1 constant operand" ); |
2177 | assert(C->isNegative() && "Expected negative FP constant" ); |
2178 | Negatible->setOperand(i: 0, Val: ConstantFP::get(Ty: Negatible->getType(), V: abs(X: *C))); |
2179 | MadeChange = true; |
2180 | } |
2181 | if (match(V: Negatible->getOperand(i: 1), P: m_APFloat(Res&: C))) { |
2182 | assert(!match(Negatible->getOperand(0), m_Constant()) && |
2183 | "Expecting only 1 constant operand" ); |
2184 | assert(C->isNegative() && "Expected negative FP constant" ); |
2185 | Negatible->setOperand(i: 1, Val: ConstantFP::get(Ty: Negatible->getType(), V: abs(X: *C))); |
2186 | MadeChange = true; |
2187 | } |
2188 | } |
2189 | assert(MadeChange == true && "Negative constant candidate was not changed" ); |
2190 | |
2191 | // Negations cancelled out. |
2192 | if (Candidates.size() % 2 == 0) |
2193 | return I; |
2194 | |
2195 | // Negate the final operand in the expression by flipping the opcode of this |
2196 | // fadd/fsub. |
2197 | assert(Candidates.size() % 2 == 1 && "Expected odd number" ); |
2198 | IRBuilder<> Builder(I); |
2199 | Value *NewInst = IsFSub ? Builder.CreateFAddFMF(L: OtherOp, R: Op, FMFSource: I) |
2200 | : Builder.CreateFSubFMF(L: OtherOp, R: Op, FMFSource: I); |
2201 | I->replaceAllUsesWith(V: NewInst); |
2202 | RedoInsts.insert(X: I); |
2203 | return dyn_cast<Instruction>(Val: NewInst); |
2204 | } |
2205 | |
2206 | /// Canonicalize expressions that contain a negative floating-point constant |
2207 | /// of the following form: |
2208 | /// OtherOp + (subtree) -> OtherOp {+/-} (canonical subtree) |
2209 | /// (subtree) + OtherOp -> OtherOp {+/-} (canonical subtree) |
2210 | /// OtherOp - (subtree) -> OtherOp {+/-} (canonical subtree) |
2211 | /// |
2212 | /// The fadd/fsub opcode may be switched to allow folding a negation into the |
2213 | /// input instruction. |
2214 | Instruction *ReassociatePass::canonicalizeNegFPConstants(Instruction *I) { |
2215 | LLVM_DEBUG(dbgs() << "Combine negations for: " << *I << '\n'); |
2216 | Value *X; |
2217 | Instruction *Op; |
2218 | if (match(V: I, P: m_FAdd(L: m_Value(V&: X), R: m_OneUse(SubPattern: m_Instruction(I&: Op))))) |
2219 | if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, OtherOp: X)) |
2220 | I = R; |
2221 | if (match(V: I, P: m_FAdd(L: m_OneUse(SubPattern: m_Instruction(I&: Op)), R: m_Value(V&: X)))) |
2222 | if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, OtherOp: X)) |
2223 | I = R; |
2224 | if (match(V: I, P: m_FSub(L: m_Value(V&: X), R: m_OneUse(SubPattern: m_Instruction(I&: Op))))) |
2225 | if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, OtherOp: X)) |
2226 | I = R; |
2227 | return I; |
2228 | } |
2229 | |
2230 | /// Inspect and optimize the given instruction. Note that erasing |
2231 | /// instructions is not allowed. |
2232 | void ReassociatePass::OptimizeInst(Instruction *I) { |
2233 | // Only consider operations that we understand. |
2234 | if (!isa<UnaryOperator>(Val: I) && !isa<BinaryOperator>(Val: I)) |
2235 | return; |
2236 | |
2237 | if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(Val: I->getOperand(i: 1))) |
2238 | // If an operand of this shift is a reassociable multiply, or if the shift |
2239 | // is used by a reassociable multiply or add, turn into a multiply. |
2240 | if (isReassociableOp(V: I->getOperand(i: 0), Opcode: Instruction::Mul) || |
2241 | (I->hasOneUse() && |
2242 | (isReassociableOp(V: I->user_back(), Opcode: Instruction::Mul) || |
2243 | isReassociableOp(V: I->user_back(), Opcode: Instruction::Add)))) { |
2244 | Instruction *NI = ConvertShiftToMul(Shl: I); |
2245 | RedoInsts.insert(X: I); |
2246 | MadeChange = true; |
2247 | I = NI; |
2248 | } |
2249 | |
2250 | // Commute binary operators, to canonicalize the order of their operands. |
2251 | // This can potentially expose more CSE opportunities, and makes writing other |
2252 | // transformations simpler. |
2253 | if (I->isCommutative()) |
2254 | canonicalizeOperands(I); |
2255 | |
2256 | // Canonicalize negative constants out of expressions. |
2257 | if (Instruction *Res = canonicalizeNegFPConstants(I)) |
2258 | I = Res; |
2259 | |
2260 | // Don't optimize floating-point instructions unless they have the |
2261 | // appropriate FastMathFlags for reassociation enabled. |
2262 | if (isa<FPMathOperator>(Val: I) && !hasFPAssociativeFlags(I)) |
2263 | return; |
2264 | |
2265 | // Do not reassociate boolean (i1) expressions. We want to preserve the |
2266 | // original order of evaluation for short-circuited comparisons that |
2267 | // SimplifyCFG has folded to AND/OR expressions. If the expression |
2268 | // is not further optimized, it is likely to be transformed back to a |
2269 | // short-circuited form for code gen, and the source order may have been |
2270 | // optimized for the most likely conditions. |
2271 | if (I->getType()->isIntegerTy(Bitwidth: 1)) |
2272 | return; |
2273 | |
2274 | // If this is a bitwise or instruction of operands |
2275 | // with no common bits set, convert it to X+Y. |
2276 | if (I->getOpcode() == Instruction::Or && |
2277 | shouldConvertOrWithNoCommonBitsToAdd(Or: I) && !isLoadCombineCandidate(Or: I) && |
2278 | (cast<PossiblyDisjointInst>(Val: I)->isDisjoint() || |
2279 | haveNoCommonBitsSet(LHSCache: I->getOperand(i: 0), RHSCache: I->getOperand(i: 1), |
2280 | SQ: SimplifyQuery(I->getModule()->getDataLayout(), |
2281 | /*DT=*/nullptr, /*AC=*/nullptr, I)))) { |
2282 | Instruction *NI = convertOrWithNoCommonBitsToAdd(Or: I); |
2283 | RedoInsts.insert(X: I); |
2284 | MadeChange = true; |
2285 | I = NI; |
2286 | } |
2287 | |
2288 | // If this is a subtract instruction which is not already in negate form, |
2289 | // see if we can convert it to X+-Y. |
2290 | if (I->getOpcode() == Instruction::Sub) { |
2291 | if (ShouldBreakUpSubtract(Sub: I)) { |
2292 | Instruction *NI = BreakUpSubtract(Sub: I, ToRedo&: RedoInsts); |
2293 | RedoInsts.insert(X: I); |
2294 | MadeChange = true; |
2295 | I = NI; |
2296 | } else if (match(V: I, P: m_Neg(V: m_Value()))) { |
2297 | // Otherwise, this is a negation. See if the operand is a multiply tree |
2298 | // and if this is not an inner node of a multiply tree. |
2299 | if (isReassociableOp(V: I->getOperand(i: 1), Opcode: Instruction::Mul) && |
2300 | (!I->hasOneUse() || |
2301 | !isReassociableOp(V: I->user_back(), Opcode: Instruction::Mul))) { |
2302 | Instruction *NI = LowerNegateToMultiply(Neg: I); |
2303 | // If the negate was simplified, revisit the users to see if we can |
2304 | // reassociate further. |
2305 | for (User *U : NI->users()) { |
2306 | if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(Val: U)) |
2307 | RedoInsts.insert(X: Tmp); |
2308 | } |
2309 | RedoInsts.insert(X: I); |
2310 | MadeChange = true; |
2311 | I = NI; |
2312 | } |
2313 | } |
2314 | } else if (I->getOpcode() == Instruction::FNeg || |
2315 | I->getOpcode() == Instruction::FSub) { |
2316 | if (ShouldBreakUpSubtract(Sub: I)) { |
2317 | Instruction *NI = BreakUpSubtract(Sub: I, ToRedo&: RedoInsts); |
2318 | RedoInsts.insert(X: I); |
2319 | MadeChange = true; |
2320 | I = NI; |
2321 | } else if (match(V: I, P: m_FNeg(X: m_Value()))) { |
2322 | // Otherwise, this is a negation. See if the operand is a multiply tree |
2323 | // and if this is not an inner node of a multiply tree. |
2324 | Value *Op = isa<BinaryOperator>(Val: I) ? I->getOperand(i: 1) : |
2325 | I->getOperand(i: 0); |
2326 | if (isReassociableOp(V: Op, Opcode: Instruction::FMul) && |
2327 | (!I->hasOneUse() || |
2328 | !isReassociableOp(V: I->user_back(), Opcode: Instruction::FMul))) { |
2329 | // If the negate was simplified, revisit the users to see if we can |
2330 | // reassociate further. |
2331 | Instruction *NI = LowerNegateToMultiply(Neg: I); |
2332 | for (User *U : NI->users()) { |
2333 | if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(Val: U)) |
2334 | RedoInsts.insert(X: Tmp); |
2335 | } |
2336 | RedoInsts.insert(X: I); |
2337 | MadeChange = true; |
2338 | I = NI; |
2339 | } |
2340 | } |
2341 | } |
2342 | |
2343 | // If this instruction is an associative binary operator, process it. |
2344 | if (!I->isAssociative()) return; |
2345 | BinaryOperator *BO = cast<BinaryOperator>(Val: I); |
2346 | |
2347 | // If this is an interior node of a reassociable tree, ignore it until we |
2348 | // get to the root of the tree, to avoid N^2 analysis. |
2349 | unsigned Opcode = BO->getOpcode(); |
2350 | if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) { |
2351 | // During the initial run we will get to the root of the tree. |
2352 | // But if we get here while we are redoing instructions, there is no |
2353 | // guarantee that the root will be visited. So Redo later |
2354 | if (BO->user_back() != BO && |
2355 | BO->getParent() == BO->user_back()->getParent()) |
2356 | RedoInsts.insert(X: BO->user_back()); |
2357 | return; |
2358 | } |
2359 | |
2360 | // If this is an add tree that is used by a sub instruction, ignore it |
2361 | // until we process the subtract. |
2362 | if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add && |
2363 | cast<Instruction>(Val: BO->user_back())->getOpcode() == Instruction::Sub) |
2364 | return; |
2365 | if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd && |
2366 | cast<Instruction>(Val: BO->user_back())->getOpcode() == Instruction::FSub) |
2367 | return; |
2368 | |
2369 | ReassociateExpression(I: BO); |
2370 | } |
2371 | |
2372 | void ReassociatePass::ReassociateExpression(BinaryOperator *I) { |
2373 | // First, walk the expression tree, linearizing the tree, collecting the |
2374 | // operand information. |
2375 | SmallVector<RepeatedValue, 8> Tree; |
2376 | bool HasNUW = true; |
2377 | MadeChange |= LinearizeExprTree(I, Ops&: Tree, ToRedo&: RedoInsts, HasNUW); |
2378 | SmallVector<ValueEntry, 8> Ops; |
2379 | Ops.reserve(N: Tree.size()); |
2380 | for (const RepeatedValue &E : Tree) |
2381 | Ops.append(NumInputs: E.second.getZExtValue(), Elt: ValueEntry(getRank(V: E.first), E.first)); |
2382 | |
2383 | LLVM_DEBUG(dbgs() << "RAIn:\t" ; PrintOps(I, Ops); dbgs() << '\n'); |
2384 | |
2385 | // Now that we have linearized the tree to a list and have gathered all of |
2386 | // the operands and their ranks, sort the operands by their rank. Use a |
2387 | // stable_sort so that values with equal ranks will have their relative |
2388 | // positions maintained (and so the compiler is deterministic). Note that |
2389 | // this sorts so that the highest ranking values end up at the beginning of |
2390 | // the vector. |
2391 | llvm::stable_sort(Range&: Ops); |
2392 | |
2393 | // Now that we have the expression tree in a convenient |
2394 | // sorted form, optimize it globally if possible. |
2395 | if (Value *V = OptimizeExpression(I, Ops)) { |
2396 | if (V == I) |
2397 | // Self-referential expression in unreachable code. |
2398 | return; |
2399 | // This expression tree simplified to something that isn't a tree, |
2400 | // eliminate it. |
2401 | LLVM_DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n'); |
2402 | I->replaceAllUsesWith(V); |
2403 | if (Instruction *VI = dyn_cast<Instruction>(Val: V)) |
2404 | if (I->getDebugLoc()) |
2405 | VI->setDebugLoc(I->getDebugLoc()); |
2406 | RedoInsts.insert(X: I); |
2407 | ++NumAnnihil; |
2408 | return; |
2409 | } |
2410 | |
2411 | // We want to sink immediates as deeply as possible except in the case where |
2412 | // this is a multiply tree used only by an add, and the immediate is a -1. |
2413 | // In this case we reassociate to put the negation on the outside so that we |
2414 | // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y |
2415 | if (I->hasOneUse()) { |
2416 | if (I->getOpcode() == Instruction::Mul && |
2417 | cast<Instruction>(Val: I->user_back())->getOpcode() == Instruction::Add && |
2418 | isa<ConstantInt>(Val: Ops.back().Op) && |
2419 | cast<ConstantInt>(Val: Ops.back().Op)->isMinusOne()) { |
2420 | ValueEntry Tmp = Ops.pop_back_val(); |
2421 | Ops.insert(I: Ops.begin(), Elt: Tmp); |
2422 | } else if (I->getOpcode() == Instruction::FMul && |
2423 | cast<Instruction>(Val: I->user_back())->getOpcode() == |
2424 | Instruction::FAdd && |
2425 | isa<ConstantFP>(Val: Ops.back().Op) && |
2426 | cast<ConstantFP>(Val: Ops.back().Op)->isExactlyValue(V: -1.0)) { |
2427 | ValueEntry Tmp = Ops.pop_back_val(); |
2428 | Ops.insert(I: Ops.begin(), Elt: Tmp); |
2429 | } |
2430 | } |
2431 | |
2432 | LLVM_DEBUG(dbgs() << "RAOut:\t" ; PrintOps(I, Ops); dbgs() << '\n'); |
2433 | |
2434 | if (Ops.size() == 1) { |
2435 | if (Ops[0].Op == I) |
2436 | // Self-referential expression in unreachable code. |
2437 | return; |
2438 | |
2439 | // This expression tree simplified to something that isn't a tree, |
2440 | // eliminate it. |
2441 | I->replaceAllUsesWith(V: Ops[0].Op); |
2442 | if (Instruction *OI = dyn_cast<Instruction>(Val: Ops[0].Op)) |
2443 | OI->setDebugLoc(I->getDebugLoc()); |
2444 | RedoInsts.insert(X: I); |
2445 | return; |
2446 | } |
2447 | |
2448 | if (Ops.size() > 2 && Ops.size() <= GlobalReassociateLimit) { |
2449 | // Find the pair with the highest count in the pairmap and move it to the |
2450 | // back of the list so that it can later be CSE'd. |
2451 | // example: |
2452 | // a*b*c*d*e |
2453 | // if c*e is the most "popular" pair, we can express this as |
2454 | // (((c*e)*d)*b)*a |
2455 | unsigned Max = 1; |
2456 | unsigned BestRank = 0; |
2457 | std::pair<unsigned, unsigned> BestPair; |
2458 | unsigned Idx = I->getOpcode() - Instruction::BinaryOpsBegin; |
2459 | unsigned LimitIdx = 0; |
2460 | // With the CSE-driven heuristic, we are about to slap two values at the |
2461 | // beginning of the expression whereas they could live very late in the CFG. |
2462 | // When using the CSE-local heuristic we avoid creating dependences from |
2463 | // completely unrelated part of the CFG by limiting the expression |
2464 | // reordering on the values that live in the first seen basic block. |
2465 | // The main idea is that we want to avoid forming expressions that would |
2466 | // become loop dependent. |
2467 | if (UseCSELocalOpt) { |
2468 | const BasicBlock *FirstSeenBB = nullptr; |
2469 | int StartIdx = Ops.size() - 1; |
2470 | // Skip the first value of the expression since we need at least two |
2471 | // values to materialize an expression. I.e., even if this value is |
2472 | // anchored in a different basic block, the actual first sub expression |
2473 | // will be anchored on the second value. |
2474 | for (int i = StartIdx - 1; i != -1; --i) { |
2475 | const Value *Val = Ops[i].Op; |
2476 | const auto *CurrLeafInstr = dyn_cast<Instruction>(Val); |
2477 | const BasicBlock *SeenBB = nullptr; |
2478 | if (!CurrLeafInstr) { |
2479 | // The value is free of any CFG dependencies. |
2480 | // Do as if it lives in the entry block. |
2481 | // |
2482 | // We do this to make sure all the values falling on this path are |
2483 | // seen through the same anchor point. The rationale is these values |
2484 | // can be combined together to from a sub expression free of any CFG |
2485 | // dependencies so we want them to stay together. |
2486 | // We could be cleverer and postpone the anchor down to the first |
2487 | // anchored value, but that's likely complicated to get right. |
2488 | // E.g., we wouldn't want to do that if that means being stuck in a |
2489 | // loop. |
2490 | // |
2491 | // For instance, we wouldn't want to change: |
2492 | // res = arg1 op arg2 op arg3 op ... op loop_val1 op loop_val2 ... |
2493 | // into |
2494 | // res = loop_val1 op arg1 op arg2 op arg3 op ... op loop_val2 ... |
2495 | // Because all the sub expressions with arg2..N would be stuck between |
2496 | // two loop dependent values. |
2497 | SeenBB = &I->getParent()->getParent()->getEntryBlock(); |
2498 | } else { |
2499 | SeenBB = CurrLeafInstr->getParent(); |
2500 | } |
2501 | |
2502 | if (!FirstSeenBB) { |
2503 | FirstSeenBB = SeenBB; |
2504 | continue; |
2505 | } |
2506 | if (FirstSeenBB != SeenBB) { |
2507 | // ith value is in a different basic block. |
2508 | // Rewind the index once to point to the last value on the same basic |
2509 | // block. |
2510 | LimitIdx = i + 1; |
2511 | LLVM_DEBUG(dbgs() << "CSE reordering: Consider values between [" |
2512 | << LimitIdx << ", " << StartIdx << "]\n" ); |
2513 | break; |
2514 | } |
2515 | } |
2516 | } |
2517 | for (unsigned i = Ops.size() - 1; i > LimitIdx; --i) { |
2518 | // We must use int type to go below zero when LimitIdx is 0. |
2519 | for (int j = i - 1; j >= (int)LimitIdx; --j) { |
2520 | unsigned Score = 0; |
2521 | Value *Op0 = Ops[i].Op; |
2522 | Value *Op1 = Ops[j].Op; |
2523 | if (std::less<Value *>()(Op1, Op0)) |
2524 | std::swap(a&: Op0, b&: Op1); |
2525 | auto it = PairMap[Idx].find(Val: {Op0, Op1}); |
2526 | if (it != PairMap[Idx].end()) { |
2527 | // Functions like BreakUpSubtract() can erase the Values we're using |
2528 | // as keys and create new Values after we built the PairMap. There's a |
2529 | // small chance that the new nodes can have the same address as |
2530 | // something already in the table. We shouldn't accumulate the stored |
2531 | // score in that case as it refers to the wrong Value. |
2532 | if (it->second.isValid()) |
2533 | Score += it->second.Score; |
2534 | } |
2535 | |
2536 | unsigned MaxRank = std::max(a: Ops[i].Rank, b: Ops[j].Rank); |
2537 | |
2538 | // By construction, the operands are sorted in reverse order of their |
2539 | // topological order. |
2540 | // So we tend to form (sub) expressions with values that are close to |
2541 | // each other. |
2542 | // |
2543 | // Now to expose more CSE opportunities we want to expose the pair of |
2544 | // operands that occur the most (as statically computed in |
2545 | // BuildPairMap.) as the first sub-expression. |
2546 | // |
2547 | // If two pairs occur as many times, we pick the one with the |
2548 | // lowest rank, meaning the one with both operands appearing first in |
2549 | // the topological order. |
2550 | if (Score > Max || (Score == Max && MaxRank < BestRank)) { |
2551 | BestPair = {j, i}; |
2552 | Max = Score; |
2553 | BestRank = MaxRank; |
2554 | } |
2555 | } |
2556 | } |
2557 | if (Max > 1) { |
2558 | auto Op0 = Ops[BestPair.first]; |
2559 | auto Op1 = Ops[BestPair.second]; |
2560 | Ops.erase(CI: &Ops[BestPair.second]); |
2561 | Ops.erase(CI: &Ops[BestPair.first]); |
2562 | Ops.push_back(Elt: Op0); |
2563 | Ops.push_back(Elt: Op1); |
2564 | } |
2565 | } |
2566 | LLVM_DEBUG(dbgs() << "RAOut after CSE reorder:\t" ; PrintOps(I, Ops); |
2567 | dbgs() << '\n'); |
2568 | // Now that we ordered and optimized the expressions, splat them back into |
2569 | // the expression tree, removing any unneeded nodes. |
2570 | RewriteExprTree(I, Ops, HasNUW); |
2571 | } |
2572 | |
2573 | void |
2574 | ReassociatePass::BuildPairMap(ReversePostOrderTraversal<Function *> &RPOT) { |
2575 | // Make a "pairmap" of how often each operand pair occurs. |
2576 | for (BasicBlock *BI : RPOT) { |
2577 | for (Instruction &I : *BI) { |
2578 | if (!I.isAssociative() || !I.isBinaryOp()) |
2579 | continue; |
2580 | |
2581 | // Ignore nodes that aren't at the root of trees. |
2582 | if (I.hasOneUse() && I.user_back()->getOpcode() == I.getOpcode()) |
2583 | continue; |
2584 | |
2585 | // Collect all operands in a single reassociable expression. |
2586 | // Since Reassociate has already been run once, we can assume things |
2587 | // are already canonical according to Reassociation's regime. |
2588 | SmallVector<Value *, 8> Worklist = { I.getOperand(i: 0), I.getOperand(i: 1) }; |
2589 | SmallVector<Value *, 8> Ops; |
2590 | while (!Worklist.empty() && Ops.size() <= GlobalReassociateLimit) { |
2591 | Value *Op = Worklist.pop_back_val(); |
2592 | Instruction *OpI = dyn_cast<Instruction>(Val: Op); |
2593 | if (!OpI || OpI->getOpcode() != I.getOpcode() || !OpI->hasOneUse()) { |
2594 | Ops.push_back(Elt: Op); |
2595 | continue; |
2596 | } |
2597 | // Be paranoid about self-referencing expressions in unreachable code. |
2598 | if (OpI->getOperand(i: 0) != OpI) |
2599 | Worklist.push_back(Elt: OpI->getOperand(i: 0)); |
2600 | if (OpI->getOperand(i: 1) != OpI) |
2601 | Worklist.push_back(Elt: OpI->getOperand(i: 1)); |
2602 | } |
2603 | // Skip extremely long expressions. |
2604 | if (Ops.size() > GlobalReassociateLimit) |
2605 | continue; |
2606 | |
2607 | // Add all pairwise combinations of operands to the pair map. |
2608 | unsigned BinaryIdx = I.getOpcode() - Instruction::BinaryOpsBegin; |
2609 | SmallSet<std::pair<Value *, Value*>, 32> Visited; |
2610 | for (unsigned i = 0; i < Ops.size() - 1; ++i) { |
2611 | for (unsigned j = i + 1; j < Ops.size(); ++j) { |
2612 | // Canonicalize operand orderings. |
2613 | Value *Op0 = Ops[i]; |
2614 | Value *Op1 = Ops[j]; |
2615 | if (std::less<Value *>()(Op1, Op0)) |
2616 | std::swap(a&: Op0, b&: Op1); |
2617 | if (!Visited.insert(V: {Op0, Op1}).second) |
2618 | continue; |
2619 | auto res = PairMap[BinaryIdx].insert(KV: {{Op0, Op1}, {.Value1: Op0, .Value2: Op1, .Score: 1}}); |
2620 | if (!res.second) { |
2621 | // If either key value has been erased then we've got the same |
2622 | // address by coincidence. That can't happen here because nothing is |
2623 | // erasing values but it can happen by the time we're querying the |
2624 | // map. |
2625 | assert(res.first->second.isValid() && "WeakVH invalidated" ); |
2626 | ++res.first->second.Score; |
2627 | } |
2628 | } |
2629 | } |
2630 | } |
2631 | } |
2632 | } |
2633 | |
2634 | PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) { |
2635 | // Get the functions basic blocks in Reverse Post Order. This order is used by |
2636 | // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic |
2637 | // blocks (it has been seen that the analysis in this pass could hang when |
2638 | // analysing dead basic blocks). |
2639 | ReversePostOrderTraversal<Function *> RPOT(&F); |
2640 | |
2641 | // Calculate the rank map for F. |
2642 | BuildRankMap(F, RPOT); |
2643 | |
2644 | // Build the pair map before running reassociate. |
2645 | // Technically this would be more accurate if we did it after one round |
2646 | // of reassociation, but in practice it doesn't seem to help much on |
2647 | // real-world code, so don't waste the compile time running reassociate |
2648 | // twice. |
2649 | // If a user wants, they could expicitly run reassociate twice in their |
2650 | // pass pipeline for further potential gains. |
2651 | // It might also be possible to update the pair map during runtime, but the |
2652 | // overhead of that may be large if there's many reassociable chains. |
2653 | BuildPairMap(RPOT); |
2654 | |
2655 | MadeChange = false; |
2656 | |
2657 | // Traverse the same blocks that were analysed by BuildRankMap. |
2658 | for (BasicBlock *BI : RPOT) { |
2659 | assert(RankMap.count(&*BI) && "BB should be ranked." ); |
2660 | // Optimize every instruction in the basic block. |
2661 | for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;) |
2662 | if (isInstructionTriviallyDead(I: &*II)) { |
2663 | EraseInst(I: &*II++); |
2664 | } else { |
2665 | OptimizeInst(I: &*II); |
2666 | assert(II->getParent() == &*BI && "Moved to a different block!" ); |
2667 | ++II; |
2668 | } |
2669 | |
2670 | // Make a copy of all the instructions to be redone so we can remove dead |
2671 | // instructions. |
2672 | OrderedSet ToRedo(RedoInsts); |
2673 | // Iterate over all instructions to be reevaluated and remove trivially dead |
2674 | // instructions. If any operand of the trivially dead instruction becomes |
2675 | // dead mark it for deletion as well. Continue this process until all |
2676 | // trivially dead instructions have been removed. |
2677 | while (!ToRedo.empty()) { |
2678 | Instruction *I = ToRedo.pop_back_val(); |
2679 | if (isInstructionTriviallyDead(I)) { |
2680 | RecursivelyEraseDeadInsts(I, Insts&: ToRedo); |
2681 | MadeChange = true; |
2682 | } |
2683 | } |
2684 | |
2685 | // Now that we have removed dead instructions, we can reoptimize the |
2686 | // remaining instructions. |
2687 | while (!RedoInsts.empty()) { |
2688 | Instruction *I = RedoInsts.front(); |
2689 | RedoInsts.erase(I: RedoInsts.begin()); |
2690 | if (isInstructionTriviallyDead(I)) |
2691 | EraseInst(I); |
2692 | else |
2693 | OptimizeInst(I); |
2694 | } |
2695 | } |
2696 | |
2697 | // We are done with the rank map and pair map. |
2698 | RankMap.clear(); |
2699 | ValueRankMap.clear(); |
2700 | for (auto &Entry : PairMap) |
2701 | Entry.clear(); |
2702 | |
2703 | if (MadeChange) { |
2704 | PreservedAnalyses PA; |
2705 | PA.preserveSet<CFGAnalyses>(); |
2706 | return PA; |
2707 | } |
2708 | |
2709 | return PreservedAnalyses::all(); |
2710 | } |
2711 | |
2712 | namespace { |
2713 | |
2714 | class ReassociateLegacyPass : public FunctionPass { |
2715 | ReassociatePass Impl; |
2716 | |
2717 | public: |
2718 | static char ID; // Pass identification, replacement for typeid |
2719 | |
2720 | ReassociateLegacyPass() : FunctionPass(ID) { |
2721 | initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry()); |
2722 | } |
2723 | |
2724 | bool runOnFunction(Function &F) override { |
2725 | if (skipFunction(F)) |
2726 | return false; |
2727 | |
2728 | FunctionAnalysisManager DummyFAM; |
2729 | auto PA = Impl.run(F, DummyFAM); |
2730 | return !PA.areAllPreserved(); |
2731 | } |
2732 | |
2733 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
2734 | AU.setPreservesCFG(); |
2735 | AU.addPreserved<AAResultsWrapperPass>(); |
2736 | AU.addPreserved<BasicAAWrapperPass>(); |
2737 | AU.addPreserved<GlobalsAAWrapperPass>(); |
2738 | } |
2739 | }; |
2740 | |
2741 | } // end anonymous namespace |
2742 | |
2743 | char ReassociateLegacyPass::ID = 0; |
2744 | |
2745 | INITIALIZE_PASS(ReassociateLegacyPass, "reassociate" , |
2746 | "Reassociate expressions" , false, false) |
2747 | |
2748 | // Public interface to the Reassociate pass |
2749 | FunctionPass *llvm::createReassociatePass() { |
2750 | return new ReassociateLegacyPass(); |
2751 | } |
2752 | |