/Users/buildslave/jenkins/sharedspace/clang-stage2-coverage-R@2/llvm/lib/Transforms/Scalar/Reassociate.cpp
Line | Count | Source (jump to first uncovered line) |
1 | | //===- Reassociate.cpp - Reassociate binary expressions -------------------===// |
2 | | // |
3 | | // The LLVM Compiler Infrastructure |
4 | | // |
5 | | // This file is distributed under the University of Illinois Open Source |
6 | | // License. See LICENSE.TXT for details. |
7 | | // |
8 | | //===----------------------------------------------------------------------===// |
9 | | // |
10 | | // This pass reassociates commutative expressions in an order that is designed |
11 | | // to promote better constant propagation, GCSE, LICM, PRE, etc. |
12 | | // |
13 | | // For example: 4 + (x + 5) -> x + (4 + 5) |
14 | | // |
15 | | // In the implementation of this algorithm, constants are assigned rank = 0, |
16 | | // function arguments are rank = 1, and other values are assigned ranks |
17 | | // corresponding to the reverse post order traversal of current function |
18 | | // (starting at 2), which effectively gives values in deep loops higher rank |
19 | | // than values not in loops. |
20 | | // |
21 | | //===----------------------------------------------------------------------===// |
22 | | |
23 | | #include "llvm/Transforms/Scalar/Reassociate.h" |
24 | | #include "llvm/ADT/DenseMap.h" |
25 | | #include "llvm/ADT/PostOrderIterator.h" |
26 | | #include "llvm/ADT/STLExtras.h" |
27 | | #include "llvm/ADT/SetVector.h" |
28 | | #include "llvm/ADT/Statistic.h" |
29 | | #include "llvm/Analysis/GlobalsModRef.h" |
30 | | #include "llvm/Analysis/ValueTracking.h" |
31 | | #include "llvm/IR/CFG.h" |
32 | | #include "llvm/IR/Constants.h" |
33 | | #include "llvm/IR/DerivedTypes.h" |
34 | | #include "llvm/IR/Function.h" |
35 | | #include "llvm/IR/IRBuilder.h" |
36 | | #include "llvm/IR/Instructions.h" |
37 | | #include "llvm/IR/IntrinsicInst.h" |
38 | | #include "llvm/IR/PatternMatch.h" |
39 | | #include "llvm/IR/ValueHandle.h" |
40 | | #include "llvm/Pass.h" |
41 | | #include "llvm/Support/Debug.h" |
42 | | #include "llvm/Support/raw_ostream.h" |
43 | | #include "llvm/Transforms/Scalar.h" |
44 | | #include "llvm/Transforms/Utils/Local.h" |
45 | | #include <algorithm> |
46 | | using namespace llvm; |
47 | | using namespace reassociate; |
48 | | |
49 | | #define DEBUG_TYPE "reassociate" |
50 | | |
51 | | STATISTIC(NumChanged, "Number of insts reassociated"); |
52 | | STATISTIC(NumAnnihil, "Number of expr tree annihilated"); |
53 | | STATISTIC(NumFactor , "Number of multiplies factored"); |
54 | | |
55 | | #ifndef NDEBUG |
56 | | /// Print out the expression identified in the Ops list. |
57 | | /// |
58 | | static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) { |
59 | | Module *M = I->getModule(); |
60 | | dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " " |
61 | | << *Ops[0].Op->getType() << '\t'; |
62 | | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { |
63 | | dbgs() << "[ "; |
64 | | Ops[i].Op->printAsOperand(dbgs(), false, M); |
65 | | dbgs() << ", #" << Ops[i].Rank << "] "; |
66 | | } |
67 | | } |
68 | | #endif |
69 | | |
70 | | /// Utility class representing a non-constant Xor-operand. We classify |
71 | | /// non-constant Xor-Operands into two categories: |
72 | | /// C1) The operand is in the form "X & C", where C is a constant and C != ~0 |
73 | | /// C2) |
74 | | /// C2.1) The operand is in the form of "X | C", where C is a non-zero |
75 | | /// constant. |
76 | | /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this |
77 | | /// operand as "E | 0" |
78 | | class llvm::reassociate::XorOpnd { |
79 | | public: |
80 | | XorOpnd(Value *V); |
81 | | |
82 | 42 | bool isInvalid() const { return SymbolicPart == nullptr; } |
83 | 13.5k | bool isOrExpr() const { return isOr; } |
84 | 125 | Value *getValue() const { return OrigVal; } |
85 | 87.4k | Value *getSymbolicPart() const { return SymbolicPart; } |
86 | 125k | unsigned getSymbolicRank() const { return SymbolicRank; } |
87 | 13.0k | const APInt &getConstPart() const { return ConstPart; } |
88 | | |
89 | 22 | void Invalidate() { SymbolicPart = OrigVal = nullptr; } |
90 | 42.2k | void setSymbolicRank(unsigned R) { SymbolicRank = R; } |
91 | | |
92 | | private: |
93 | | Value *OrigVal; |
94 | | Value *SymbolicPart; |
95 | | APInt ConstPart; |
96 | | unsigned SymbolicRank; |
97 | | bool isOr; |
98 | | }; |
99 | | |
100 | 42.2k | XorOpnd::XorOpnd(Value *V) { |
101 | 42.2k | assert(!isa<ConstantInt>(V) && "No ConstantInt"); |
102 | 42.2k | OrigVal = V; |
103 | 42.2k | Instruction *I = dyn_cast<Instruction>(V); |
104 | 42.2k | SymbolicRank = 0; |
105 | 42.2k | |
106 | 42.2k | if (I && 42.2k (I->getOpcode() == Instruction::Or || |
107 | 42.2k | I->getOpcode() == Instruction::And39.4k )) { |
108 | 3.58k | Value *V0 = I->getOperand(0); |
109 | 3.58k | Value *V1 = I->getOperand(1); |
110 | 3.58k | const APInt *C; |
111 | 3.58k | if (match(V0, PatternMatch::m_APInt(C))) |
112 | 0 | std::swap(V0, V1); |
113 | 3.58k | |
114 | 3.58k | if (match(V1, PatternMatch::m_APInt(C))3.58k ) { |
115 | 1.59k | ConstPart = *C; |
116 | 1.59k | SymbolicPart = V0; |
117 | 1.59k | isOr = (I->getOpcode() == Instruction::Or); |
118 | 1.59k | return; |
119 | 1.59k | } |
120 | 40.6k | } |
121 | 40.6k | |
122 | 40.6k | // view the operand as "V | 0" |
123 | 40.6k | SymbolicPart = V; |
124 | 40.6k | ConstPart = APInt::getNullValue(V->getType()->getScalarSizeInBits()); |
125 | 40.6k | isOr = true; |
126 | 40.6k | } |
127 | | |
128 | | /// Return true if V is an instruction of the specified opcode and if it |
129 | | /// only has one use. |
130 | 3.17M | static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { |
131 | 3.17M | if (V->hasOneUse() && 3.17M isa<Instruction>(V)1.55M && |
132 | 1.43M | cast<Instruction>(V)->getOpcode() == Opcode && |
133 | 545k | (!isa<FPMathOperator>(V) || |
134 | 358 | cast<Instruction>(V)->hasUnsafeAlgebra())) |
135 | 545k | return cast<BinaryOperator>(V); |
136 | 2.63M | return nullptr; |
137 | 2.63M | } |
138 | | |
139 | | static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1, |
140 | 2.60M | unsigned Opcode2) { |
141 | 2.60M | if (V->hasOneUse() && 2.60M isa<Instruction>(V)1.09M && |
142 | 986k | (cast<Instruction>(V)->getOpcode() == Opcode1 || |
143 | 986k | cast<Instruction>(V)->getOpcode() == Opcode2) && |
144 | 274k | (!isa<FPMathOperator>(V) || |
145 | 279 | cast<Instruction>(V)->hasUnsafeAlgebra())) |
146 | 274k | return cast<BinaryOperator>(V); |
147 | 2.32M | return nullptr; |
148 | 2.32M | } |
149 | | |
150 | | void ReassociatePass::BuildRankMap(Function &F, |
151 | 516k | ReversePostOrderTraversal<Function*> &RPOT) { |
152 | 516k | unsigned Rank = 2; |
153 | 516k | |
154 | 516k | // Assign distinct ranks to function arguments. |
155 | 1.17M | for (auto &Arg : F.args()) { |
156 | 1.17M | ValueRankMap[&Arg] = ++Rank; |
157 | 1.17M | DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank |
158 | 1.17M | << "\n"); |
159 | 1.17M | } |
160 | 516k | |
161 | 516k | // Traverse basic blocks in ReversePostOrder |
162 | 3.62M | for (BasicBlock *BB : RPOT) { |
163 | 3.62M | unsigned BBRank = RankMap[BB] = ++Rank << 16; |
164 | 3.62M | |
165 | 3.62M | // Walk the basic block, adding precomputed ranks for any instructions that |
166 | 3.62M | // we cannot move. This ensures that the ranks for these instructions are |
167 | 3.62M | // all different in the block. |
168 | 3.62M | for (Instruction &I : *BB) |
169 | 18.7M | if (18.7M mayBeMemoryDependent(I)18.7M ) |
170 | 10.8M | ValueRankMap[&I] = ++BBRank; |
171 | 3.62M | } |
172 | 516k | } |
173 | | |
174 | 5.76M | unsigned ReassociatePass::getRank(Value *V) { |
175 | 5.76M | Instruction *I = dyn_cast<Instruction>(V); |
176 | 5.76M | if (!I5.76M ) { |
177 | 1.58M | if (isa<Argument>(V)1.58M ) return ValueRankMap[V]256k ; // Function argument. |
178 | 1.32M | return 0; // Otherwise it's a global or constant, rank 0. |
179 | 1.32M | } |
180 | 4.18M | |
181 | 4.18M | if (unsigned 4.18M Rank4.18M = ValueRankMap[I]) |
182 | 3.01M | return Rank; // Rank already known? |
183 | 1.16M | |
184 | 1.16M | // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that |
185 | 1.16M | // we can reassociate expressions for code motion! Since we do not recurse |
186 | 1.16M | // for PHI nodes, we cannot have infinite recursion here, because there |
187 | 1.16M | // cannot be loops in the value graph that do not go through PHI nodes. |
188 | 1.16M | unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; |
189 | 1.16M | for (unsigned i = 0, e = I->getNumOperands(); |
190 | 3.34M | i != e && 3.34M Rank != MaxRank2.17M ; ++i2.17M ) |
191 | 2.17M | Rank = std::max(Rank, getRank(I->getOperand(i))); |
192 | 1.16M | |
193 | 1.16M | // If this is a not or neg instruction, do not count it for rank. This |
194 | 1.16M | // assures us that X and ~X will have the same rank. |
195 | 1.16M | if (!BinaryOperator::isNot(I) && 1.16M !BinaryOperator::isNeg(I)1.15M && |
196 | 1.13M | !BinaryOperator::isFNeg(I)) |
197 | 1.13M | ++Rank; |
198 | 1.16M | |
199 | 1.16M | DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n"); |
200 | 5.76M | |
201 | 5.76M | return ValueRankMap[I] = Rank; |
202 | 5.76M | } |
203 | | |
204 | | // Canonicalize constants to RHS. Otherwise, sort the operands by rank. |
205 | 1.35M | void ReassociatePass::canonicalizeOperands(Instruction *I) { |
206 | 1.35M | assert(isa<BinaryOperator>(I) && "Expected binary operator."); |
207 | 1.35M | assert(I->isCommutative() && "Expected commutative operator."); |
208 | 1.35M | |
209 | 1.35M | Value *LHS = I->getOperand(0); |
210 | 1.35M | Value *RHS = I->getOperand(1); |
211 | 1.35M | if (LHS == RHS || 1.35M isa<Constant>(RHS)1.34M ) |
212 | 722k | return; |
213 | 631k | if (631k isa<Constant>(LHS) || 631k getRank(RHS) < getRank(LHS)631k ) |
214 | 290k | cast<BinaryOperator>(I)->swapOperands(); |
215 | 1.35M | } |
216 | | |
217 | | static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name, |
218 | 14.3k | Instruction *InsertBefore, Value *FlagsOp) { |
219 | 14.3k | if (S1->getType()->isIntOrIntVectorTy()) |
220 | 14.3k | return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore); |
221 | 48 | else { |
222 | 48 | BinaryOperator *Res = |
223 | 48 | BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore); |
224 | 48 | Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); |
225 | 48 | return Res; |
226 | 48 | } |
227 | 0 | } |
228 | | |
229 | | static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name, |
230 | 2.58k | Instruction *InsertBefore, Value *FlagsOp) { |
231 | 2.58k | if (S1->getType()->isIntOrIntVectorTy()) |
232 | 2.53k | return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore); |
233 | 52 | else { |
234 | 52 | BinaryOperator *Res = |
235 | 52 | BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore); |
236 | 52 | Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); |
237 | 52 | return Res; |
238 | 52 | } |
239 | 0 | } |
240 | | |
241 | | static BinaryOperator *CreateNeg(Value *S1, const Twine &Name, |
242 | 9.53k | Instruction *InsertBefore, Value *FlagsOp) { |
243 | 9.53k | if (S1->getType()->isIntOrIntVectorTy()) |
244 | 9.51k | return BinaryOperator::CreateNeg(S1, Name, InsertBefore); |
245 | 26 | else { |
246 | 26 | BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore); |
247 | 26 | Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); |
248 | 26 | return Res; |
249 | 26 | } |
250 | 0 | } |
251 | | |
252 | | /// Replace 0-X with X*-1. |
253 | 1.27k | static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) { |
254 | 1.27k | Type *Ty = Neg->getType(); |
255 | 1.27k | Constant *NegOne = Ty->isIntOrIntVectorTy() ? |
256 | 1.27k | ConstantInt::getAllOnesValue(Ty)1.25k : ConstantFP::get(Ty, -1.0)17 ; |
257 | 1.27k | |
258 | 1.27k | BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg); |
259 | 1.27k | Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op. |
260 | 1.27k | Res->takeName(Neg); |
261 | 1.27k | Neg->replaceAllUsesWith(Res); |
262 | 1.27k | Res->setDebugLoc(Neg->getDebugLoc()); |
263 | 1.27k | return Res; |
264 | 1.27k | } |
265 | | |
266 | | /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael |
267 | | /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for |
268 | | /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic. |
269 | | /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every |
270 | | /// even x in Bitwidth-bit arithmetic. |
271 | 4.43k | static unsigned CarmichaelShift(unsigned Bitwidth) { |
272 | 4.43k | if (Bitwidth < 3) |
273 | 0 | return Bitwidth - 1; |
274 | 4.43k | return Bitwidth - 2; |
275 | 4.43k | } |
276 | | |
277 | | /// Add the extra weight 'RHS' to the existing weight 'LHS', |
278 | | /// reducing the combined weight using any special properties of the operation. |
279 | | /// The existing weight LHS represents the computation X op X op ... op X where |
280 | | /// X occurs LHS times. The combined weight represents X op X op ... op X with |
281 | | /// X occurring LHS + RHS times. If op is "Xor" for example then the combined |
282 | | /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even; |
283 | | /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second. |
284 | 8.57k | static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) { |
285 | 8.57k | // If we were working with infinite precision arithmetic then the combined |
286 | 8.57k | // weight would be LHS + RHS. But we are using finite precision arithmetic, |
287 | 8.57k | // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct |
288 | 8.57k | // for nilpotent operations and addition, but not for idempotent operations |
289 | 8.57k | // and multiplication), so it is important to correctly reduce the combined |
290 | 8.57k | // weight back into range if wrapping would be wrong. |
291 | 8.57k | |
292 | 8.57k | // If RHS is zero then the weight didn't change. |
293 | 8.57k | if (RHS.isMinValue()) |
294 | 0 | return; |
295 | 8.57k | // If LHS is zero then the combined weight is RHS. |
296 | 8.57k | if (8.57k LHS.isMinValue()8.57k ) { |
297 | 2 | LHS = RHS; |
298 | 2 | return; |
299 | 2 | } |
300 | 8.57k | // From this point on we know that neither LHS nor RHS is zero. |
301 | 8.57k | |
302 | 8.57k | if (8.57k Instruction::isIdempotent(Opcode)8.57k ) { |
303 | 11 | // Idempotent means X op X === X, so any non-zero weight is equivalent to a |
304 | 11 | // weight of 1. Keeping weights at zero or one also means that wrapping is |
305 | 11 | // not a problem. |
306 | 11 | assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); |
307 | 11 | return; // Return a weight of 1. |
308 | 11 | } |
309 | 8.56k | if (8.56k Instruction::isNilpotent(Opcode)8.56k ) { |
310 | 6 | // Nilpotent means X op X === 0, so reduce weights modulo 2. |
311 | 6 | assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); |
312 | 6 | LHS = 0; // 1 + 1 === 0 modulo 2. |
313 | 6 | return; |
314 | 6 | } |
315 | 8.55k | if (8.55k Opcode == Instruction::Add || 8.55k Opcode == Instruction::FAdd4.46k ) { |
316 | 4.12k | // TODO: Reduce the weight by exploiting nsw/nuw? |
317 | 4.12k | LHS += RHS; |
318 | 4.12k | return; |
319 | 4.12k | } |
320 | 4.43k | |
321 | 8.55k | assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) && |
322 | 4.43k | "Unknown associative operation!"); |
323 | 4.43k | unsigned Bitwidth = LHS.getBitWidth(); |
324 | 4.43k | // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth |
325 | 4.43k | // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth |
326 | 4.43k | // bit number x, since either x is odd in which case x^CM = 1, or x is even in |
327 | 4.43k | // which case both x^W and x^(W - CM) are zero. By subtracting off multiples |
328 | 4.43k | // of CM like this weights can always be reduced to the range [0, CM+Bitwidth) |
329 | 4.43k | // which by a happy accident means that they can always be represented using |
330 | 4.43k | // Bitwidth bits. |
331 | 4.43k | // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than |
332 | 4.43k | // the Carmichael number). |
333 | 4.43k | if (Bitwidth > 34.43k ) { |
334 | 4.41k | /// CM - The value of Carmichael's lambda function. |
335 | 4.41k | APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth)); |
336 | 4.41k | // Any weight W >= Threshold can be replaced with W - CM. |
337 | 4.41k | APInt Threshold = CM + Bitwidth; |
338 | 4.41k | assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!"); |
339 | 4.41k | // For Bitwidth 4 or more the following sum does not overflow. |
340 | 4.41k | LHS += RHS; |
341 | 4.42k | while (LHS.uge(Threshold)) |
342 | 12 | LHS -= CM; |
343 | 4.43k | } else { |
344 | 23 | // To avoid problems with overflow do everything the same as above but using |
345 | 23 | // a larger type. |
346 | 23 | unsigned CM = 1U << CarmichaelShift(Bitwidth); |
347 | 23 | unsigned Threshold = CM + Bitwidth; |
348 | 23 | assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && |
349 | 23 | "Weights not reduced!"); |
350 | 23 | unsigned Total = LHS.getZExtValue() + RHS.getZExtValue(); |
351 | 27 | while (Total >= Threshold) |
352 | 4 | Total -= CM; |
353 | 23 | LHS = Total; |
354 | 23 | } |
355 | 8.57k | } |
356 | | |
357 | | typedef std::pair<Value*, APInt> RepeatedValue; |
358 | | |
359 | | /// Given an associative binary expression, return the leaf |
360 | | /// nodes in Ops along with their weights (how many times the leaf occurs). The |
361 | | /// original expression is the same as |
362 | | /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times |
363 | | /// op |
364 | | /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times |
365 | | /// op |
366 | | /// ... |
367 | | /// op |
368 | | /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times |
369 | | /// |
370 | | /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct. |
371 | | /// |
372 | | /// This routine may modify the function, in which case it returns 'true'. The |
373 | | /// changes it makes may well be destructive, changing the value computed by 'I' |
374 | | /// to something completely different. Thus if the routine returns 'true' then |
375 | | /// you MUST either replace I with a new expression computed from the Ops array, |
376 | | /// or use RewriteExprTree to put the values back in. |
377 | | /// |
378 | | /// A leaf node is either not a binary operation of the same kind as the root |
379 | | /// node 'I' (i.e. is not a binary operator at all, or is, but with a different |
380 | | /// opcode), or is the same kind of binary operator but has a use which either |
381 | | /// does not belong to the expression, or does belong to the expression but is |
382 | | /// a leaf node. Every leaf node has at least one use that is a non-leaf node |
383 | | /// of the expression, while for non-leaf nodes (except for the root 'I') every |
384 | | /// use is a non-leaf node of the expression. |
385 | | /// |
386 | | /// For example: |
387 | | /// expression graph node names |
388 | | /// |
389 | | /// + | I |
390 | | /// / \ | |
391 | | /// + + | A, B |
392 | | /// / \ / \ | |
393 | | /// * + * | C, D, E |
394 | | /// / \ / \ / \ | |
395 | | /// + * | F, G |
396 | | /// |
397 | | /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in |
398 | | /// that order) (C, 1), (E, 1), (F, 2), (G, 2). |
399 | | /// |
400 | | /// The expression is maximal: if some instruction is a binary operator of the |
401 | | /// same kind as 'I', and all of its uses are non-leaf nodes of the expression, |
402 | | /// then the instruction also belongs to the expression, is not a leaf node of |
403 | | /// it, and its operands also belong to the expression (but may be leaf nodes). |
404 | | /// |
405 | | /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in |
406 | | /// order to ensure that every non-root node in the expression has *exactly one* |
407 | | /// use by a non-leaf node of the expression. This destruction means that the |
408 | | /// caller MUST either replace 'I' with a new expression or use something like |
409 | | /// RewriteExprTree to put the values back in if the routine indicates that it |
410 | | /// made a change by returning 'true'. |
411 | | /// |
412 | | /// In the above example either the right operand of A or the left operand of B |
413 | | /// will be replaced by undef. If it is B's operand then this gives: |
414 | | /// |
415 | | /// + | I |
416 | | /// / \ | |
417 | | /// + + | A, B - operand of B replaced with undef |
418 | | /// / \ \ | |
419 | | /// * + * | C, D, E |
420 | | /// / \ / \ / \ | |
421 | | /// + * | F, G |
422 | | /// |
423 | | /// Note that such undef operands can only be reached by passing through 'I'. |
424 | | /// For example, if you visit operands recursively starting from a leaf node |
425 | | /// then you will never see such an undef operand unless you get back to 'I', |
426 | | /// which requires passing through a phi node. |
427 | | /// |
428 | | /// Note that this routine may also mutate binary operators of the wrong type |
429 | | /// that have all uses inside the expression (i.e. only used by non-leaf nodes |
430 | | /// of the expression) if it can turn them into binary operators of the right |
431 | | /// type and thus make the expression bigger. |
432 | | |
433 | | static bool LinearizeExprTree(BinaryOperator *I, |
434 | 1.01M | SmallVectorImpl<RepeatedValue> &Ops) { |
435 | 1.01M | DEBUG(dbgs() << "LINEARIZE: " << *I << '\n'); |
436 | 1.01M | unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits(); |
437 | 1.01M | unsigned Opcode = I->getOpcode(); |
438 | 1.01M | assert(I->isAssociative() && I->isCommutative() && |
439 | 1.01M | "Expected an associative and commutative operation!"); |
440 | 1.01M | |
441 | 1.01M | // Visit all operands of the expression, keeping track of their weight (the |
442 | 1.01M | // number of paths from the expression root to the operand, or if you like |
443 | 1.01M | // the number of times that operand occurs in the linearized expression). |
444 | 1.01M | // For example, if I = X + A, where X = A + B, then I, X and B have weight 1 |
445 | 1.01M | // while A has weight two. |
446 | 1.01M | |
447 | 1.01M | // Worklist of non-leaf nodes (their operands are in the expression too) along |
448 | 1.01M | // with their weights, representing a certain number of paths to the operator. |
449 | 1.01M | // If an operator occurs in the worklist multiple times then we found multiple |
450 | 1.01M | // ways to get to it. |
451 | 1.01M | SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight) |
452 | 1.01M | Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1))); |
453 | 1.01M | bool Changed = false; |
454 | 1.01M | |
455 | 1.01M | // Leaves of the expression are values that either aren't the right kind of |
456 | 1.01M | // operation (eg: a constant, or a multiply in an add tree), or are, but have |
457 | 1.01M | // some uses that are not inside the expression. For example, in I = X + X, |
458 | 1.01M | // X = A + B, the value X has two uses (by I) that are in the expression. If |
459 | 1.01M | // X has any other uses, for example in a return instruction, then we consider |
460 | 1.01M | // X to be a leaf, and won't analyze it further. When we first visit a value, |
461 | 1.01M | // if it has more than one use then at first we conservatively consider it to |
462 | 1.01M | // be a leaf. Later, as the expression is explored, we may discover some more |
463 | 1.01M | // uses of the value from inside the expression. If all uses turn out to be |
464 | 1.01M | // from within the expression (and the value is a binary operator of the right |
465 | 1.01M | // kind) then the value is no longer considered to be a leaf, and its operands |
466 | 1.01M | // are explored. |
467 | 1.01M | |
468 | 1.01M | // Leaves - Keeps track of the set of putative leaves as well as the number of |
469 | 1.01M | // paths to each leaf seen so far. |
470 | 1.01M | typedef DenseMap<Value*, APInt> LeafMap; |
471 | 1.01M | LeafMap Leaves; // Leaf -> Total weight so far. |
472 | 1.01M | SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order. |
473 | 1.01M | |
474 | | #ifndef NDEBUG |
475 | | SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme. |
476 | | #endif |
477 | 2.29M | while (!Worklist.empty()2.29M ) { |
478 | 1.27M | std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val(); |
479 | 1.27M | I = P.first; // We examine the operands of this binary operator. |
480 | 1.27M | |
481 | 3.83M | for (unsigned OpIdx = 0; OpIdx < 23.83M ; ++OpIdx2.55M ) { // Visit operands. |
482 | 2.55M | Value *Op = I->getOperand(OpIdx); |
483 | 2.55M | APInt Weight = P.second; // Number of paths to this operand. |
484 | 2.55M | DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n"); |
485 | 2.55M | assert(!Op->use_empty() && "No uses, so how did we get to it?!"); |
486 | 2.55M | |
487 | 2.55M | // If this is a binary operation of the right kind with only one use then |
488 | 2.55M | // add its operands to the expression. |
489 | 2.55M | if (BinaryOperator *BO2.55M = isReassociableOp(Op, Opcode)) { |
490 | 267k | assert(Visited.insert(Op).second && "Not first visit!"); |
491 | 267k | DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n"); |
492 | 267k | Worklist.push_back(std::make_pair(BO, Weight)); |
493 | 267k | continue; |
494 | 267k | } |
495 | 2.29M | |
496 | 2.29M | // Appears to be a leaf. Is the operand already in the set of leaves? |
497 | 2.29M | LeafMap::iterator It = Leaves.find(Op); |
498 | 2.29M | if (It == Leaves.end()2.29M ) { |
499 | 2.28M | // Not in the leaf map. Must be the first time we saw this operand. |
500 | 2.28M | assert(Visited.insert(Op).second && "Not first visit!"); |
501 | 2.28M | if (!Op->hasOneUse()2.28M ) { |
502 | 1.43M | // This value has uses not accounted for by the expression, so it is |
503 | 1.43M | // not safe to modify. Mark it as being a leaf. |
504 | 1.43M | DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n"); |
505 | 1.43M | LeafOrder.push_back(Op); |
506 | 1.43M | Leaves[Op] = Weight; |
507 | 1.43M | continue; |
508 | 1.43M | } |
509 | 2.29M | // No uses outside the expression, try morphing it. |
510 | 8.57k | } else { |
511 | 8.57k | // Already in the leaf map. |
512 | 8.57k | assert(It != Leaves.end() && Visited.count(Op) && |
513 | 8.57k | "In leaf map but not visited!"); |
514 | 8.57k | |
515 | 8.57k | // Update the number of paths to the leaf. |
516 | 8.57k | IncorporateWeight(It->second, Weight, Opcode); |
517 | 8.57k | |
518 | | #if 0 // TODO: Re-enable once PR13021 is fixed. |
519 | | // The leaf already has one use from inside the expression. As we want |
520 | | // exactly one such use, drop this new use of the leaf. |
521 | | assert(!Op->hasOneUse() && "Only one use, but we got here twice!"); |
522 | | I->setOperand(OpIdx, UndefValue::get(I->getType())); |
523 | | Changed = true; |
524 | | |
525 | | // If the leaf is a binary operation of the right kind and we now see |
526 | | // that its multiple original uses were in fact all by nodes belonging |
527 | | // to the expression, then no longer consider it to be a leaf and add |
528 | | // its operands to the expression. |
529 | | if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { |
530 | | DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n"); |
531 | | Worklist.push_back(std::make_pair(BO, It->second)); |
532 | | Leaves.erase(It); |
533 | | continue; |
534 | | } |
535 | | #endif |
536 | | |
537 | 8.57k | // If we still have uses that are not accounted for by the expression |
538 | 8.57k | // then it is not safe to modify the value. |
539 | 8.57k | if (!Op->hasOneUse()) |
540 | 8.57k | continue; |
541 | 1 | |
542 | 1 | // No uses outside the expression, try morphing it. |
543 | 1 | Weight = It->second; |
544 | 1 | Leaves.erase(It); // Since the value may be morphed below. |
545 | 1 | } |
546 | 2.29M | |
547 | 2.29M | // At this point we have a value which, first of all, is not a binary |
548 | 2.29M | // expression of the right kind, and secondly, is only used inside the |
549 | 2.29M | // expression. This means that it can safely be modified. See if we |
550 | 2.29M | // can usefully morph it into an expression of the right kind. |
551 | 844k | assert((!isa<Instruction>(Op) || |
552 | 844k | cast<Instruction>(Op)->getOpcode() != Opcode |
553 | 844k | || (isa<FPMathOperator>(Op) && |
554 | 844k | !cast<Instruction>(Op)->hasUnsafeAlgebra())) && |
555 | 844k | "Should have been handled above!"); |
556 | 844k | assert(Op->hasOneUse() && "Has uses outside the expression tree!"); |
557 | 844k | |
558 | 844k | // If this is a multiply expression, turn any internal negations into |
559 | 844k | // multiplies by -1 so they can be reassociated. |
560 | 844k | if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) |
561 | 273k | if (273k (Opcode == Instruction::Mul && 273k BinaryOperator::isNeg(BO)16.4k ) || |
562 | 273k | (Opcode == Instruction::FMul && 273k BinaryOperator::isFNeg(BO)38 )) { |
563 | 120 | DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO "); |
564 | 120 | BO = LowerNegateToMultiply(BO); |
565 | 120 | DEBUG(dbgs() << *BO << '\n'); |
566 | 273k | Worklist.push_back(std::make_pair(BO, Weight)); |
567 | 273k | Changed = true; |
568 | 273k | continue; |
569 | 273k | } |
570 | 844k | |
571 | 844k | // Failed to morph into an expression of the right type. This really is |
572 | 844k | // a leaf. |
573 | 844k | DEBUG844k (dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n"); |
574 | 844k | assert(!isReassociableOp(Op, Opcode) && "Value was morphed?"); |
575 | 844k | LeafOrder.push_back(Op); |
576 | 844k | Leaves[Op] = Weight; |
577 | 844k | } |
578 | 1.27M | } |
579 | 1.01M | |
580 | 1.01M | // The leaves, repeated according to their weights, represent the linearized |
581 | 1.01M | // form of the expression. |
582 | 3.29M | for (unsigned i = 0, e = LeafOrder.size(); i != e3.29M ; ++i2.28M ) { |
583 | 2.28M | Value *V = LeafOrder[i]; |
584 | 2.28M | LeafMap::iterator It = Leaves.find(V); |
585 | 2.28M | if (It == Leaves.end()) |
586 | 2.28M | // Node initially thought to be a leaf wasn't. |
587 | 0 | continue; |
588 | 2.28M | assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!"); |
589 | 2.28M | APInt Weight = It->second; |
590 | 2.28M | if (Weight.isMinValue()) |
591 | 2.28M | // Leaf already output or weight reduction eliminated it. |
592 | 5 | continue; |
593 | 2.28M | // Ensure the leaf is only output once. |
594 | 2.28M | It->second = 0; |
595 | 2.28M | Ops.push_back(std::make_pair(V, Weight)); |
596 | 2.28M | } |
597 | 1.01M | |
598 | 1.01M | // For nilpotent operations or addition there may be no operands, for example |
599 | 1.01M | // because the expression was "X xor X" or consisted of 2^Bitwidth additions: |
600 | 1.01M | // in both cases the weight reduces to 0 causing the value to be skipped. |
601 | 1.01M | if (Ops.empty()1.01M ) { |
602 | 3 | Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType()); |
603 | 3 | assert(Identity && "Associative operation without identity!"); |
604 | 3 | Ops.emplace_back(Identity, APInt(Bitwidth, 1)); |
605 | 3 | } |
606 | 1.01M | |
607 | 1.01M | return Changed; |
608 | 1.01M | } |
609 | | |
610 | | /// Now that the operands for this expression tree are |
611 | | /// linearized and optimized, emit them in-order. |
612 | | void ReassociatePass::RewriteExprTree(BinaryOperator *I, |
613 | 1.00M | SmallVectorImpl<ValueEntry> &Ops) { |
614 | 1.00M | assert(Ops.size() > 1 && "Single values should be used directly!"); |
615 | 1.00M | |
616 | 1.00M | // Since our optimizations should never increase the number of operations, the |
617 | 1.00M | // new expression can usually be written reusing the existing binary operators |
618 | 1.00M | // from the original expression tree, without creating any new instructions, |
619 | 1.00M | // though the rewritten expression may have a completely different topology. |
620 | 1.00M | // We take care to not change anything if the new expression will be the same |
621 | 1.00M | // as the original. If more than trivial changes (like commuting operands) |
622 | 1.00M | // were made then we are obliged to clear out any optional subclass data like |
623 | 1.00M | // nsw flags. |
624 | 1.00M | |
625 | 1.00M | /// NodesToRewrite - Nodes from the original expression available for writing |
626 | 1.00M | /// the new expression into. |
627 | 1.00M | SmallVector<BinaryOperator*, 8> NodesToRewrite; |
628 | 1.00M | unsigned Opcode = I->getOpcode(); |
629 | 1.00M | BinaryOperator *Op = I; |
630 | 1.00M | |
631 | 1.00M | /// NotRewritable - The operands being written will be the leaves of the new |
632 | 1.00M | /// expression and must not be used as inner nodes (via NodesToRewrite) by |
633 | 1.00M | /// mistake. Inner nodes are always reassociable, and usually leaves are not |
634 | 1.00M | /// (if they were they would have been incorporated into the expression and so |
635 | 1.00M | /// would not be leaves), so most of the time there is no danger of this. But |
636 | 1.00M | /// in rare cases a leaf may become reassociable if an optimization kills uses |
637 | 1.00M | /// of it, or it may momentarily become reassociable during rewriting (below) |
638 | 1.00M | /// due it being removed as an operand of one of its uses. Ensure that misuse |
639 | 1.00M | /// of leaf nodes as inner nodes cannot occur by remembering all of the future |
640 | 1.00M | /// leaves and refusing to reuse any of them as inner nodes. |
641 | 1.00M | SmallPtrSet<Value*, 8> NotRewritable; |
642 | 3.27M | for (unsigned i = 0, e = Ops.size(); i != e3.27M ; ++i2.27M ) |
643 | 2.27M | NotRewritable.insert(Ops[i].Op); |
644 | 1.00M | |
645 | 1.00M | // ExpressionChanged - Non-null if the rewritten expression differs from the |
646 | 1.00M | // original in some non-trivial way, requiring the clearing of optional flags. |
647 | 1.00M | // Flags are cleared from the operator in ExpressionChanged up to I inclusive. |
648 | 1.00M | BinaryOperator *ExpressionChanged = nullptr; |
649 | 1.00M | for (unsigned i = 0; ; ++i256k ) { |
650 | 1.26M | // The last operation (which comes earliest in the IR) is special as both |
651 | 1.26M | // operands will come from Ops, rather than just one with the other being |
652 | 1.26M | // a subexpression. |
653 | 1.26M | if (i+2 == Ops.size()1.26M ) { |
654 | 1.00M | Value *NewLHS = Ops[i].Op; |
655 | 1.00M | Value *NewRHS = Ops[i+1].Op; |
656 | 1.00M | Value *OldLHS = Op->getOperand(0); |
657 | 1.00M | Value *OldRHS = Op->getOperand(1); |
658 | 1.00M | |
659 | 1.00M | if (NewLHS == OldLHS && 1.00M NewRHS == OldRHS764k ) |
660 | 1.00M | // Nothing changed, leave it alone. |
661 | 745k | break; |
662 | 261k | |
663 | 261k | if (261k NewLHS == OldRHS && 261k NewRHS == OldLHS238k ) { |
664 | 236k | // The order of the operands was reversed. Swap them. |
665 | 236k | DEBUG(dbgs() << "RA: " << *Op << '\n'); |
666 | 236k | Op->swapOperands(); |
667 | 236k | DEBUG(dbgs() << "TO: " << *Op << '\n'); |
668 | 236k | MadeChange = true; |
669 | 236k | ++NumChanged; |
670 | 236k | break; |
671 | 236k | } |
672 | 25.7k | |
673 | 25.7k | // The new operation differs non-trivially from the original. Overwrite |
674 | 25.7k | // the old operands with the new ones. |
675 | 25.7k | DEBUG25.7k (dbgs() << "RA: " << *Op << '\n'); |
676 | 25.7k | if (NewLHS != OldLHS25.7k ) { |
677 | 7.00k | BinaryOperator *BO = isReassociableOp(OldLHS, Opcode); |
678 | 7.00k | if (BO && 7.00k !NotRewritable.count(BO)2.31k ) |
679 | 2.31k | NodesToRewrite.push_back(BO); |
680 | 7.00k | Op->setOperand(0, NewLHS); |
681 | 7.00k | } |
682 | 25.7k | if (NewRHS != OldRHS25.7k ) { |
683 | 24.3k | BinaryOperator *BO = isReassociableOp(OldRHS, Opcode); |
684 | 24.3k | if (BO && 24.3k !NotRewritable.count(BO)175 ) |
685 | 175 | NodesToRewrite.push_back(BO); |
686 | 24.3k | Op->setOperand(1, NewRHS); |
687 | 24.3k | } |
688 | 25.7k | DEBUG(dbgs() << "TO: " << *Op << '\n'); |
689 | 1.00M | |
690 | 1.00M | ExpressionChanged = Op; |
691 | 1.00M | MadeChange = true; |
692 | 1.00M | ++NumChanged; |
693 | 1.00M | |
694 | 1.00M | break; |
695 | 1.00M | } |
696 | 256k | |
697 | 256k | // Not the last operation. The left-hand side will be a sub-expression |
698 | 256k | // while the right-hand side will be the current element of Ops. |
699 | 256k | Value *NewRHS = Ops[i].Op; |
700 | 256k | if (NewRHS != Op->getOperand(1)256k ) { |
701 | 91.3k | DEBUG(dbgs() << "RA: " << *Op << '\n'); |
702 | 91.3k | if (NewRHS == Op->getOperand(0)91.3k ) { |
703 | 57.7k | // The new right-hand side was already present as the left operand. If |
704 | 57.7k | // we are lucky then swapping the operands will sort out both of them. |
705 | 57.7k | Op->swapOperands(); |
706 | 91.3k | } else { |
707 | 33.5k | // Overwrite with the new right-hand side. |
708 | 33.5k | BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode); |
709 | 33.5k | if (BO && 33.5k !NotRewritable.count(BO)6.58k ) |
710 | 6.58k | NodesToRewrite.push_back(BO); |
711 | 33.5k | Op->setOperand(1, NewRHS); |
712 | 33.5k | ExpressionChanged = Op; |
713 | 33.5k | } |
714 | 91.3k | DEBUG(dbgs() << "TO: " << *Op << '\n'); |
715 | 91.3k | MadeChange = true; |
716 | 91.3k | ++NumChanged; |
717 | 91.3k | } |
718 | 256k | |
719 | 256k | // Now deal with the left-hand side. If this is already an operation node |
720 | 256k | // from the original expression then just rewrite the rest of the expression |
721 | 256k | // into it. |
722 | 256k | BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode); |
723 | 256k | if (BO && 256k !NotRewritable.count(BO)250k ) { |
724 | 250k | Op = BO; |
725 | 250k | continue; |
726 | 250k | } |
727 | 6.53k | |
728 | 6.53k | // Otherwise, grab a spare node from the original expression and use that as |
729 | 6.53k | // the left-hand side. If there are no nodes left then the optimizers made |
730 | 6.53k | // an expression with more nodes than the original! This usually means that |
731 | 6.53k | // they did something stupid but it might mean that the problem was just too |
732 | 6.53k | // hard (finding the mimimal number of multiplications needed to realize a |
733 | 6.53k | // multiplication expression is NP-complete). Whatever the reason, smart or |
734 | 6.53k | // stupid, create a new node if there are none left. |
735 | 6.53k | BinaryOperator *NewOp; |
736 | 6.53k | if (NodesToRewrite.empty()6.53k ) { |
737 | 0 | Constant *Undef = UndefValue::get(I->getType()); |
738 | 0 | NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode), |
739 | 0 | Undef, Undef, "", I); |
740 | 0 | if (NewOp->getType()->isFPOrFPVectorTy()) |
741 | 0 | NewOp->setFastMathFlags(I->getFastMathFlags()); |
742 | 6.53k | } else { |
743 | 6.53k | NewOp = NodesToRewrite.pop_back_val(); |
744 | 6.53k | } |
745 | 6.53k | |
746 | 6.53k | DEBUG(dbgs() << "RA: " << *Op << '\n'); |
747 | 6.53k | Op->setOperand(0, NewOp); |
748 | 6.53k | DEBUG(dbgs() << "TO: " << *Op << '\n'); |
749 | 1.26M | ExpressionChanged = Op; |
750 | 1.26M | MadeChange = true; |
751 | 1.26M | ++NumChanged; |
752 | 1.26M | Op = NewOp; |
753 | 1.26M | } |
754 | 1.00M | |
755 | 1.00M | // If the expression changed non-trivially then clear out all subclass data |
756 | 1.00M | // starting from the operator specified in ExpressionChanged, and compactify |
757 | 1.00M | // the operators to just before the expression root to guarantee that the |
758 | 1.00M | // expression tree is dominated by all of Ops. |
759 | 1.00M | if (ExpressionChanged) |
760 | 26.5k | do 26.5k { |
761 | 61.1k | // Preserve FastMathFlags. |
762 | 61.1k | if (isa<FPMathOperator>(I)61.1k ) { |
763 | 130 | FastMathFlags Flags = I->getFastMathFlags(); |
764 | 130 | ExpressionChanged->clearSubclassOptionalData(); |
765 | 130 | ExpressionChanged->setFastMathFlags(Flags); |
766 | 130 | } else |
767 | 61.0k | ExpressionChanged->clearSubclassOptionalData(); |
768 | 61.1k | |
769 | 61.1k | if (ExpressionChanged == I) |
770 | 26.5k | break; |
771 | 34.6k | ExpressionChanged->moveBefore(I); |
772 | 34.6k | ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin()); |
773 | 34.6k | } while (1); |
774 | 1.00M | |
775 | 1.00M | // Throw away any left over nodes from the original expression. |
776 | 1.00M | for (unsigned i = 0, e = NodesToRewrite.size(); i != e1.00M ; ++i2.54k ) |
777 | 2.54k | RedoInsts.insert(NodesToRewrite[i]); |
778 | 1.00M | } |
779 | | |
780 | | /// Insert instructions before the instruction pointed to by BI, |
781 | | /// that computes the negative version of the value specified. The negative |
782 | | /// version of the value is returned, and BI is left pointing at the instruction |
783 | | /// that should be processed next by the reassociation pass. |
784 | | /// Also add intermediate instructions to the redo list that are modified while |
785 | | /// pushing the negates through adds. These will be revisited to see if |
786 | | /// additional opportunities have been exposed. |
787 | | static Value *NegateValue(Value *V, Instruction *BI, |
788 | 11.4k | SetVector<AssertingVH<Instruction>> &ToRedo) { |
789 | 11.4k | if (Constant *C11.4k = dyn_cast<Constant>(V)) { |
790 | 163 | if (C->getType()->isFPOrFPVectorTy()163 ) { |
791 | 7 | return ConstantExpr::getFNeg(C); |
792 | 7 | } |
793 | 156 | return ConstantExpr::getNeg(C); |
794 | 156 | } |
795 | 11.2k | |
796 | 11.2k | |
797 | 11.2k | // We are trying to expose opportunity for reassociation. One of the things |
798 | 11.2k | // that we want to do to achieve this is to push a negation as deep into an |
799 | 11.2k | // expression chain as possible, to expose the add instructions. In practice, |
800 | 11.2k | // this means that we turn this: |
801 | 11.2k | // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D |
802 | 11.2k | // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate |
803 | 11.2k | // the constants. We assume that instcombine will clean up the mess later if |
804 | 11.2k | // we introduce tons of unnecessary negation instructions. |
805 | 11.2k | // |
806 | 11.2k | if (BinaryOperator *11.2k I11.2k = |
807 | 1.04k | isReassociableOp(V, Instruction::Add, Instruction::FAdd)) { |
808 | 1.04k | // Push the negates through the add. |
809 | 1.04k | I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo)); |
810 | 1.04k | I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo)); |
811 | 1.04k | if (I->getOpcode() == Instruction::Add1.04k ) { |
812 | 1.03k | I->setHasNoUnsignedWrap(false); |
813 | 1.03k | I->setHasNoSignedWrap(false); |
814 | 1.03k | } |
815 | 1.04k | |
816 | 1.04k | // We must move the add instruction here, because the neg instructions do |
817 | 1.04k | // not dominate the old add instruction in general. By moving it, we are |
818 | 1.04k | // assured that the neg instructions we just inserted dominate the |
819 | 1.04k | // instruction we are about to insert after them. |
820 | 1.04k | // |
821 | 1.04k | I->moveBefore(BI); |
822 | 1.04k | I->setName(I->getName()+".neg"); |
823 | 1.04k | |
824 | 1.04k | // Add the intermediate negates to the redo list as processing them later |
825 | 1.04k | // could expose more reassociating opportunities. |
826 | 1.04k | ToRedo.insert(I); |
827 | 1.04k | return I; |
828 | 1.04k | } |
829 | 10.2k | |
830 | 10.2k | // Okay, we need to materialize a negated version of V with an instruction. |
831 | 10.2k | // Scan the use lists of V to see if we have one already. |
832 | 10.2k | for (User *U : V->users()) 10.2k { |
833 | 18.5k | if (!BinaryOperator::isNeg(U) && 18.5k !BinaryOperator::isFNeg(U)17.8k ) |
834 | 17.8k | continue; |
835 | 716 | |
836 | 716 | // We found one! Now we have to make sure that the definition dominates |
837 | 716 | // this use. We do this by moving it to the entry block (if it is a |
838 | 716 | // non-instruction value) or right after the definition. These negates will |
839 | 716 | // be zapped by reassociate later, so we don't need much finesse here. |
840 | 716 | BinaryOperator *TheNeg = cast<BinaryOperator>(U); |
841 | 716 | |
842 | 716 | // Verify that the negate is in this function, V might be a constant expr. |
843 | 716 | if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) |
844 | 0 | continue; |
845 | 716 | |
846 | 716 | BasicBlock::iterator InsertPt; |
847 | 716 | if (Instruction *InstInput716 = dyn_cast<Instruction>(V)) { |
848 | 632 | if (InvokeInst *II632 = dyn_cast<InvokeInst>(InstInput)) { |
849 | 0 | InsertPt = II->getNormalDest()->begin(); |
850 | 632 | } else { |
851 | 632 | InsertPt = ++InstInput->getIterator(); |
852 | 632 | } |
853 | 796 | while (isa<PHINode>(InsertPt)796 ) ++InsertPt164 ; |
854 | 716 | } else { |
855 | 84 | InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); |
856 | 84 | } |
857 | 716 | TheNeg->moveBefore(&*InsertPt); |
858 | 716 | if (TheNeg->getOpcode() == Instruction::Sub716 ) { |
859 | 711 | TheNeg->setHasNoUnsignedWrap(false); |
860 | 711 | TheNeg->setHasNoSignedWrap(false); |
861 | 716 | } else { |
862 | 5 | TheNeg->andIRFlags(BI); |
863 | 5 | } |
864 | 18.5k | ToRedo.insert(TheNeg); |
865 | 18.5k | return TheNeg; |
866 | 18.5k | } |
867 | 9.49k | |
868 | 9.49k | // Insert a 'neg' instruction that subtracts the value from zero to get the |
869 | 9.49k | // negation. |
870 | 9.49k | BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI); |
871 | 9.49k | ToRedo.insert(NewNeg); |
872 | 9.49k | return NewNeg; |
873 | 9.49k | } |
874 | | |
875 | | /// Return true if we should break up this subtract of X-Y into (X + -Y). |
876 | 155k | static bool ShouldBreakUpSubtract(Instruction *Sub) { |
877 | 155k | // If this is a negation, we can't split it up! |
878 | 155k | if (BinaryOperator::isNeg(Sub) || 155k BinaryOperator::isFNeg(Sub)118k ) |
879 | 36.7k | return false; |
880 | 118k | |
881 | 118k | // Don't breakup X - undef. |
882 | 118k | if (118k isa<UndefValue>(Sub->getOperand(1))118k ) |
883 | 3 | return false; |
884 | 118k | |
885 | 118k | // Don't bother to break this up unless either the LHS is an associable add or |
886 | 118k | // subtract or if this is only used by one. |
887 | 118k | Value *V0 = Sub->getOperand(0); |
888 | 118k | if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) || |
889 | 115k | isReassociableOp(V0, Instruction::Sub, Instruction::FSub)) |
890 | 2.89k | return true; |
891 | 115k | Value *V1 = Sub->getOperand(1); |
892 | 115k | if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) || |
893 | 115k | isReassociableOp(V1, Instruction::Sub, Instruction::FSub)) |
894 | 246 | return true; |
895 | 115k | Value *VB = Sub->user_back(); |
896 | 115k | if (Sub->hasOneUse() && |
897 | 102k | (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) || |
898 | 96.5k | isReassociableOp(VB, Instruction::Sub, Instruction::FSub))) |
899 | 6.19k | return true; |
900 | 109k | |
901 | 109k | return false; |
902 | 109k | } |
903 | | |
904 | | /// If we have (X-Y), and if either X is an add, or if this is only used by an |
905 | | /// add, transform this into (X+(0-Y)) to promote better reassociation. |
906 | | static BinaryOperator * |
907 | 9.33k | BreakUpSubtract(Instruction *Sub, SetVector<AssertingVH<Instruction>> &ToRedo) { |
908 | 9.33k | // Convert a subtract into an add and a neg instruction. This allows sub |
909 | 9.33k | // instructions to be commuted with other add instructions. |
910 | 9.33k | // |
911 | 9.33k | // Calculate the negative value of Operand 1 of the sub instruction, |
912 | 9.33k | // and set it as the RHS of the add instruction we just made. |
913 | 9.33k | // |
914 | 9.33k | Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo); |
915 | 9.33k | BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub); |
916 | 9.33k | Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op. |
917 | 9.33k | Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op. |
918 | 9.33k | New->takeName(Sub); |
919 | 9.33k | |
920 | 9.33k | // Everyone now refers to the add instruction. |
921 | 9.33k | Sub->replaceAllUsesWith(New); |
922 | 9.33k | New->setDebugLoc(Sub->getDebugLoc()); |
923 | 9.33k | |
924 | 9.33k | DEBUG(dbgs() << "Negated: " << *New << '\n'); |
925 | 9.33k | return New; |
926 | 9.33k | } |
927 | | |
928 | | /// If this is a shift of a reassociable multiply or is used by one, change |
929 | | /// this into a multiply by a constant to assist with further reassociation. |
930 | 18.4k | static BinaryOperator *ConvertShiftToMul(Instruction *Shl) { |
931 | 18.4k | Constant *MulCst = ConstantInt::get(Shl->getType(), 1); |
932 | 18.4k | MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); |
933 | 18.4k | |
934 | 18.4k | BinaryOperator *Mul = |
935 | 18.4k | BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); |
936 | 18.4k | Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op. |
937 | 18.4k | Mul->takeName(Shl); |
938 | 18.4k | |
939 | 18.4k | // Everyone now refers to the mul instruction. |
940 | 18.4k | Shl->replaceAllUsesWith(Mul); |
941 | 18.4k | Mul->setDebugLoc(Shl->getDebugLoc()); |
942 | 18.4k | |
943 | 18.4k | // We can safely preserve the nuw flag in all cases. It's also safe to turn a |
944 | 18.4k | // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special |
945 | 18.4k | // handling. |
946 | 18.4k | bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap(); |
947 | 18.4k | bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap(); |
948 | 18.4k | if (NSW && 18.4k NUW7.86k ) |
949 | 6.69k | Mul->setHasNoSignedWrap(true); |
950 | 18.4k | Mul->setHasNoUnsignedWrap(NUW); |
951 | 18.4k | return Mul; |
952 | 18.4k | } |
953 | | |
954 | | /// Scan backwards and forwards among values with the same rank as element i |
955 | | /// to see if X exists. If X does not exist, return i. This is useful when |
956 | | /// scanning for 'x' when we see '-x' because they both get the same rank. |
957 | | static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops, |
958 | 26.5k | unsigned i, Value *X) { |
959 | 26.5k | unsigned XRank = Ops[i].Rank; |
960 | 26.5k | unsigned e = Ops.size(); |
961 | 26.7k | for (unsigned j = i+1; j != e && 26.7k Ops[j].Rank == XRank22.0k ; ++j289 ) { |
962 | 308 | if (Ops[j].Op == X) |
963 | 18 | return j; |
964 | 290 | if (Instruction *290 I1290 = dyn_cast<Instruction>(Ops[j].Op)) |
965 | 290 | if (Instruction *290 I2290 = dyn_cast<Instruction>(X)) |
966 | 290 | if (290 I1->isIdenticalTo(I2)290 ) |
967 | 1 | return j; |
968 | 308 | } |
969 | 26.5k | // Scan backwards. |
970 | 32.5k | for (unsigned j = i-1; 26.4k j != ~0U && 32.5k Ops[j].Rank == XRank20.2k ; --j6.03k ) { |
971 | 6.03k | if (Ops[j].Op == X) |
972 | 3 | return j; |
973 | 6.03k | if (Instruction *6.03k I16.03k = dyn_cast<Instruction>(Ops[j].Op)) |
974 | 6.02k | if (Instruction *6.02k I26.02k = dyn_cast<Instruction>(X)) |
975 | 6.01k | if (6.01k I1->isIdenticalTo(I2)6.01k ) |
976 | 0 | return j; |
977 | 6.03k | } |
978 | 26.4k | return i; |
979 | 26.5k | } |
980 | | |
981 | | /// Emit a tree of add instructions, summing Ops together |
982 | | /// and returning the result. Insert the tree before I. |
983 | | static Value *EmitAddTreeOfValues(Instruction *I, |
984 | 6.25k | SmallVectorImpl<WeakTrackingVH> &Ops) { |
985 | 6.25k | if (Ops.size() == 16.25k ) return Ops.back()1.21k ; |
986 | 5.03k | |
987 | 5.03k | Value *V1 = Ops.back(); |
988 | 5.03k | Ops.pop_back(); |
989 | 5.03k | Value *V2 = EmitAddTreeOfValues(I, Ops); |
990 | 5.03k | return CreateAdd(V2, V1, "tmp", I, I); |
991 | 5.03k | } |
992 | | |
993 | | /// If V is an expression tree that is a multiplication sequence, |
994 | | /// and if this sequence contains a multiply by Factor, |
995 | | /// remove Factor from the tree and return the new tree. |
996 | 6.55k | Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) { |
997 | 6.55k | BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); |
998 | 6.55k | if (!BO) |
999 | 0 | return nullptr; |
1000 | 6.55k | |
1001 | 6.55k | SmallVector<RepeatedValue, 8> Tree; |
1002 | 6.55k | MadeChange |= LinearizeExprTree(BO, Tree); |
1003 | 6.55k | SmallVector<ValueEntry, 8> Factors; |
1004 | 6.55k | Factors.reserve(Tree.size()); |
1005 | 20.6k | for (unsigned i = 0, e = Tree.size(); i != e20.6k ; ++i14.0k ) { |
1006 | 14.0k | RepeatedValue E = Tree[i]; |
1007 | 14.0k | Factors.append(E.second.getZExtValue(), |
1008 | 14.0k | ValueEntry(getRank(E.first), E.first)); |
1009 | 14.0k | } |
1010 | 6.55k | |
1011 | 6.55k | bool FoundFactor = false; |
1012 | 6.55k | bool NeedsNegate = false; |
1013 | 13.9k | for (unsigned i = 0, e = Factors.size(); i != e13.9k ; ++i7.39k ) { |
1014 | 13.6k | if (Factors[i].Op == Factor13.6k ) { |
1015 | 6.21k | FoundFactor = true; |
1016 | 6.21k | Factors.erase(Factors.begin()+i); |
1017 | 6.21k | break; |
1018 | 6.21k | } |
1019 | 7.43k | |
1020 | 7.43k | // If this is a negative version of this factor, remove it. |
1021 | 7.43k | if (ConstantInt *7.43k FC17.43k = dyn_cast<ConstantInt>(Factor)) { |
1022 | 4.88k | if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) |
1023 | 199 | if (199 FC1->getValue() == -FC2->getValue()199 ) { |
1024 | 36 | FoundFactor = NeedsNegate = true; |
1025 | 36 | Factors.erase(Factors.begin()+i); |
1026 | 36 | break; |
1027 | 36 | } |
1028 | 2.55k | } else if (ConstantFP *2.55k FC12.55k = dyn_cast<ConstantFP>(Factor)) { |
1029 | 9 | if (ConstantFP *FC29 = dyn_cast<ConstantFP>(Factors[i].Op)) { |
1030 | 3 | const APFloat &F1 = FC1->getValueAPF(); |
1031 | 3 | APFloat F2(FC2->getValueAPF()); |
1032 | 3 | F2.changeSign(); |
1033 | 3 | if (F1.compare(F2) == APFloat::cmpEqual3 ) { |
1034 | 3 | FoundFactor = NeedsNegate = true; |
1035 | 3 | Factors.erase(Factors.begin() + i); |
1036 | 3 | break; |
1037 | 3 | } |
1038 | 3 | } |
1039 | 2.55k | } |
1040 | 13.6k | } |
1041 | 6.55k | |
1042 | 6.55k | if (!FoundFactor6.55k ) { |
1043 | 304 | // Make sure to restore the operands to the expression tree. |
1044 | 304 | RewriteExprTree(BO, Factors); |
1045 | 304 | return nullptr; |
1046 | 304 | } |
1047 | 6.25k | |
1048 | 6.25k | BasicBlock::iterator InsertPt = ++BO->getIterator(); |
1049 | 6.25k | |
1050 | 6.25k | // If this was just a single multiply, remove the multiply and return the only |
1051 | 6.25k | // remaining operand. |
1052 | 6.25k | if (Factors.size() == 16.25k ) { |
1053 | 5.33k | RedoInsts.insert(BO); |
1054 | 5.33k | V = Factors[0].Op; |
1055 | 6.25k | } else { |
1056 | 921 | RewriteExprTree(BO, Factors); |
1057 | 921 | V = BO; |
1058 | 921 | } |
1059 | 6.25k | |
1060 | 6.25k | if (NeedsNegate) |
1061 | 39 | V = CreateNeg(V, "neg", &*InsertPt, BO); |
1062 | 6.55k | |
1063 | 6.55k | return V; |
1064 | 6.55k | } |
1065 | | |
1066 | | /// If V is a single-use multiply, recursively add its operands as factors, |
1067 | | /// otherwise add V to the list of factors. |
1068 | | /// |
1069 | | /// Ops is the top-level list of add operands we're trying to factor. |
1070 | | static void FindSingleUseMultiplyFactors(Value *V, |
1071 | 379k | SmallVectorImpl<Value*> &Factors) { |
1072 | 379k | BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); |
1073 | 379k | if (!BO379k ) { |
1074 | 251k | Factors.push_back(V); |
1075 | 251k | return; |
1076 | 251k | } |
1077 | 128k | |
1078 | 128k | // Otherwise, add the LHS and RHS to the list of factors. |
1079 | 128k | FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); |
1080 | 128k | FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); |
1081 | 128k | } |
1082 | | |
1083 | | /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction. |
1084 | | /// This optimizes based on identities. If it can be reduced to a single Value, |
1085 | | /// it is returned, otherwise the Ops list is mutated as necessary. |
1086 | | static Value *OptimizeAndOrXor(unsigned Opcode, |
1087 | 181k | SmallVectorImpl<ValueEntry> &Ops) { |
1088 | 181k | // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. |
1089 | 181k | // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. |
1090 | 582k | for (unsigned i = 0, e = Ops.size(); i != e582k ; ++i400k ) { |
1091 | 400k | // First, check for X and ~X in the operand list. |
1092 | 400k | assert(i < Ops.size()); |
1093 | 400k | if (BinaryOperator::isNot(Ops[i].Op)400k ) { // Cannot occur for ^. |
1094 | 4.07k | Value *X = BinaryOperator::getNotArgument(Ops[i].Op); |
1095 | 4.07k | unsigned FoundX = FindInOperandList(Ops, i, X); |
1096 | 4.07k | if (FoundX != i4.07k ) { |
1097 | 2 | if (Opcode == Instruction::And) // ...&X&~X = 0 |
1098 | 2 | return Constant::getNullValue(X->getType()); |
1099 | 0 |
|
1100 | 0 | if (0 Opcode == Instruction::Or0 ) // ...|X|~X = -1 |
1101 | 0 | return Constant::getAllOnesValue(X->getType()); |
1102 | 400k | } |
1103 | 4.07k | } |
1104 | 400k | |
1105 | 400k | // Next, check for duplicate pairs of values, which we assume are next to |
1106 | 400k | // each other, due to our sorting criteria. |
1107 | 400k | assert(i < Ops.size()); |
1108 | 400k | if (i+1 != Ops.size() && 400k Ops[i+1].Op == Ops[i].Op218k ) { |
1109 | 0 | if (Opcode == Instruction::And || 0 Opcode == Instruction::Or0 ) { |
1110 | 0 | // Drop duplicate values for And and Or. |
1111 | 0 | Ops.erase(Ops.begin()+i); |
1112 | 0 | --i; --e; |
1113 | 0 | ++NumAnnihil; |
1114 | 0 | continue; |
1115 | 0 | } |
1116 | 0 |
|
1117 | 0 | // Drop pairs of values for Xor. |
1118 | 0 | assert(Opcode == Instruction::Xor); |
1119 | 0 | if (e == 2) |
1120 | 0 | return Constant::getNullValue(Ops[0].Op->getType()); |
1121 | 0 |
|
1122 | 0 | // Y ^ X^X -> Y |
1123 | 0 | Ops.erase(Ops.begin()+i, Ops.begin()+i+2); |
1124 | 0 | i -= 1; e -= 2; |
1125 | 0 | ++NumAnnihil; |
1126 | 0 | } |
1127 | 400k | } |
1128 | 181k | return nullptr; |
1129 | 181k | } |
1130 | | |
1131 | | /// Helper function of CombineXorOpnd(). It creates a bitwise-and |
1132 | | /// instruction with the given two operands, and return the resulting |
1133 | | /// instruction. There are two special cases: 1) if the constant operand is 0, |
1134 | | /// it will return NULL. 2) if the constant is ~0, the symbolic operand will |
1135 | | /// be returned. |
1136 | | static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd, |
1137 | 16 | const APInt &ConstOpnd) { |
1138 | 16 | if (ConstOpnd.isNullValue()) |
1139 | 6 | return nullptr; |
1140 | 10 | |
1141 | 10 | if (10 ConstOpnd.isAllOnesValue()10 ) |
1142 | 2 | return Opnd; |
1143 | 8 | |
1144 | 8 | Instruction *I = BinaryOperator::CreateAnd( |
1145 | 8 | Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra", |
1146 | 8 | InsertBefore); |
1147 | 8 | I->setDebugLoc(InsertBefore->getDebugLoc()); |
1148 | 8 | return I; |
1149 | 8 | } |
1150 | | |
1151 | | // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd" |
1152 | | // into "R ^ C", where C would be 0, and R is a symbolic value. |
1153 | | // |
1154 | | // If it was successful, true is returned, and the "R" and "C" is returned |
1155 | | // via "Res" and "ConstOpnd", respectively; otherwise, false is returned, |
1156 | | // and both "Res" and "ConstOpnd" remain unchanged. |
1157 | | // |
1158 | | bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, |
1159 | 13.5k | APInt &ConstOpnd, Value *&Res) { |
1160 | 13.5k | // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2 |
1161 | 13.5k | // = ((x | c1) ^ c1) ^ (c1 ^ c2) |
1162 | 13.5k | // = (x & ~c1) ^ (c1 ^ c2) |
1163 | 13.5k | // It is useful only when c1 == c2. |
1164 | 13.5k | if (!Opnd1->isOrExpr() || 13.5k Opnd1->getConstPart().isNullValue()12.9k ) |
1165 | 13.4k | return false; |
1166 | 11 | |
1167 | 11 | if (11 !Opnd1->getValue()->hasOneUse()11 ) |
1168 | 6 | return false; |
1169 | 5 | |
1170 | 5 | const APInt &C1 = Opnd1->getConstPart(); |
1171 | 5 | if (C1 != ConstOpnd) |
1172 | 5 | return false; |
1173 | 0 |
|
1174 | 0 | Value *X = Opnd1->getSymbolicPart(); |
1175 | 0 | Res = createAndInstr(I, X, ~C1); |
1176 | 0 | // ConstOpnd was C2, now C1 ^ C2. |
1177 | 0 | ConstOpnd ^= C1; |
1178 | 0 |
|
1179 | 0 | if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue())) |
1180 | 0 | RedoInsts.insert(T); |
1181 | 13.5k | return true; |
1182 | 13.5k | } |
1183 | | |
1184 | | |
1185 | | // Helper function of OptimizeXor(). It tries to simplify |
1186 | | // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a |
1187 | | // symbolic value. |
1188 | | // |
1189 | | // If it was successful, true is returned, and the "R" and "C" is returned |
1190 | | // via "Res" and "ConstOpnd", respectively (If the entire expression is |
1191 | | // evaluated to a constant, the Res is set to NULL); otherwise, false is |
1192 | | // returned, and both "Res" and "ConstOpnd" remain unchanged. |
1193 | | bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, |
1194 | | XorOpnd *Opnd2, APInt &ConstOpnd, |
1195 | 21 | Value *&Res) { |
1196 | 21 | Value *X = Opnd1->getSymbolicPart(); |
1197 | 21 | if (X != Opnd2->getSymbolicPart()) |
1198 | 0 | return false; |
1199 | 21 | |
1200 | 21 | // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.) |
1201 | 21 | int DeadInstNum = 1; |
1202 | 21 | if (Opnd1->getValue()->hasOneUse()) |
1203 | 16 | DeadInstNum++; |
1204 | 21 | if (Opnd2->getValue()->hasOneUse()) |
1205 | 15 | DeadInstNum++; |
1206 | 21 | |
1207 | 21 | // Xor-Rule 2: |
1208 | 21 | // (x | c1) ^ (x & c2) |
1209 | 21 | // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1 |
1210 | 21 | // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1 |
1211 | 21 | // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3 |
1212 | 21 | // |
1213 | 21 | if (Opnd1->isOrExpr() != Opnd2->isOrExpr()21 ) { |
1214 | 11 | if (Opnd2->isOrExpr()) |
1215 | 1 | std::swap(Opnd1, Opnd2); |
1216 | 11 | |
1217 | 11 | const APInt &C1 = Opnd1->getConstPart(); |
1218 | 11 | const APInt &C2 = Opnd2->getConstPart(); |
1219 | 11 | APInt C3((~C1) ^ C2); |
1220 | 11 | |
1221 | 11 | // Do not increase code size! |
1222 | 11 | if (!C3.isNullValue() && 11 !C3.isAllOnesValue()9 ) { |
1223 | 7 | int NewInstNum = ConstOpnd.getBoolValue() ? 10 : 27 ; |
1224 | 7 | if (NewInstNum > DeadInstNum) |
1225 | 4 | return false; |
1226 | 7 | } |
1227 | 7 | |
1228 | 7 | Res = createAndInstr(I, X, C3); |
1229 | 7 | ConstOpnd ^= C1; |
1230 | 7 | |
1231 | 21 | } else if (10 Opnd1->isOrExpr()10 ) { |
1232 | 6 | // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2 |
1233 | 6 | // |
1234 | 6 | const APInt &C1 = Opnd1->getConstPart(); |
1235 | 6 | const APInt &C2 = Opnd2->getConstPart(); |
1236 | 6 | APInt C3 = C1 ^ C2; |
1237 | 6 | |
1238 | 6 | // Do not increase code size |
1239 | 6 | if (!C3.isNullValue() && 6 !C3.isAllOnesValue()4 ) { |
1240 | 4 | int NewInstNum = ConstOpnd.getBoolValue() ? 10 : 24 ; |
1241 | 4 | if (NewInstNum > DeadInstNum) |
1242 | 1 | return false; |
1243 | 5 | } |
1244 | 5 | |
1245 | 5 | Res = createAndInstr(I, X, C3); |
1246 | 5 | ConstOpnd ^= C3; |
1247 | 10 | } else { |
1248 | 4 | // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2)) |
1249 | 4 | // |
1250 | 4 | const APInt &C1 = Opnd1->getConstPart(); |
1251 | 4 | const APInt &C2 = Opnd2->getConstPart(); |
1252 | 4 | APInt C3 = C1 ^ C2; |
1253 | 4 | Res = createAndInstr(I, X, C3); |
1254 | 4 | } |
1255 | 21 | |
1256 | 21 | // Put the original operands in the Redo list; hope they will be deleted |
1257 | 21 | // as dead code. |
1258 | 16 | if (Instruction *16 T16 = dyn_cast<Instruction>(Opnd1->getValue())) |
1259 | 16 | RedoInsts.insert(T); |
1260 | 16 | if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue())) |
1261 | 16 | RedoInsts.insert(T); |
1262 | 16 | |
1263 | 16 | return true; |
1264 | 21 | } |
1265 | | |
1266 | | /// Optimize a series of operands to an 'xor' instruction. If it can be reduced |
1267 | | /// to a single Value, it is returned, otherwise the Ops list is mutated as |
1268 | | /// necessary. |
1269 | | Value *ReassociatePass::OptimizeXor(Instruction *I, |
1270 | 19.6k | SmallVectorImpl<ValueEntry> &Ops) { |
1271 | 19.6k | if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops)) |
1272 | 0 | return V; |
1273 | 19.6k | |
1274 | 19.6k | if (19.6k Ops.size() == 119.6k ) |
1275 | 0 | return nullptr; |
1276 | 19.6k | |
1277 | 19.6k | SmallVector<XorOpnd, 8> Opnds; |
1278 | 19.6k | SmallVector<XorOpnd*, 8> OpndPtrs; |
1279 | 19.6k | Type *Ty = Ops[0].Op->getType(); |
1280 | 19.6k | APInt ConstOpnd(Ty->getScalarSizeInBits(), 0); |
1281 | 19.6k | |
1282 | 19.6k | // Step 1: Convert ValueEntry to XorOpnd |
1283 | 73.2k | for (unsigned i = 0, e = Ops.size(); i != e73.2k ; ++i53.5k ) { |
1284 | 53.5k | Value *V = Ops[i].Op; |
1285 | 53.5k | const APInt *C; |
1286 | 53.5k | // TODO: Support non-splat vectors. |
1287 | 53.5k | if (match(V, PatternMatch::m_APInt(C))53.5k ) { |
1288 | 11.3k | ConstOpnd ^= *C; |
1289 | 53.5k | } else { |
1290 | 42.2k | XorOpnd O(V); |
1291 | 42.2k | O.setSymbolicRank(getRank(O.getSymbolicPart())); |
1292 | 42.2k | Opnds.push_back(O); |
1293 | 42.2k | } |
1294 | 53.5k | } |
1295 | 19.6k | |
1296 | 19.6k | // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds". |
1297 | 19.6k | // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate |
1298 | 19.6k | // the "OpndPtrs" as well. For the similar reason, do not fuse this loop |
1299 | 19.6k | // with the previous loop --- the iterator of the "Opnds" may be invalidated |
1300 | 19.6k | // when new elements are added to the vector. |
1301 | 61.9k | for (unsigned i = 0, e = Opnds.size(); i != e61.9k ; ++i42.2k ) |
1302 | 42.2k | OpndPtrs.push_back(&Opnds[i]); |
1303 | 19.6k | |
1304 | 19.6k | // Step 2: Sort the Xor-Operands in a way such that the operands containing |
1305 | 19.6k | // the same symbolic value cluster together. For instance, the input operand |
1306 | 19.6k | // sequence ("x | 123", "y & 456", "x & 789") will be sorted into: |
1307 | 19.6k | // ("x | 123", "x & 789", "y & 456"). |
1308 | 19.6k | // |
1309 | 19.6k | // The purpose is twofold: |
1310 | 19.6k | // 1) Cluster together the operands sharing the same symbolic-value. |
1311 | 19.6k | // 2) Operand having smaller symbolic-value-rank is permuted earlier, which |
1312 | 19.6k | // could potentially shorten crital path, and expose more loop-invariants. |
1313 | 19.6k | // Note that values' rank are basically defined in RPO order (FIXME). |
1314 | 19.6k | // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier |
1315 | 19.6k | // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2", |
1316 | 19.6k | // "z" in the order of X-Y-Z is better than any other orders. |
1317 | 19.6k | std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), |
1318 | 62.6k | [](XorOpnd *LHS, XorOpnd *RHS) { |
1319 | 62.6k | return LHS->getSymbolicRank() < RHS->getSymbolicRank(); |
1320 | 62.6k | }); |
1321 | 19.6k | |
1322 | 19.6k | // Step 3: Combine adjacent operands |
1323 | 19.6k | XorOpnd *PrevOpnd = nullptr; |
1324 | 19.6k | bool Changed = false; |
1325 | 61.9k | for (unsigned i = 0, e = Opnds.size(); i < e61.9k ; i++42.2k ) { |
1326 | 42.2k | XorOpnd *CurrOpnd = OpndPtrs[i]; |
1327 | 42.2k | // The combined value |
1328 | 42.2k | Value *CV; |
1329 | 42.2k | |
1330 | 42.2k | // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd" |
1331 | 42.2k | if (!ConstOpnd.isNullValue() && |
1332 | 42.2k | CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)13.5k ) { |
1333 | 0 | Changed = true; |
1334 | 0 | if (CV) |
1335 | 0 | *CurrOpnd = XorOpnd(CV); |
1336 | 0 | else { |
1337 | 0 | CurrOpnd->Invalidate(); |
1338 | 0 | continue; |
1339 | 0 | } |
1340 | 42.2k | } |
1341 | 42.2k | |
1342 | 42.2k | if (42.2k !PrevOpnd || 42.2k CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()22.5k ) { |
1343 | 42.2k | PrevOpnd = CurrOpnd; |
1344 | 42.2k | continue; |
1345 | 42.2k | } |
1346 | 21 | |
1347 | 21 | // step 3.2: When previous and current operands share the same symbolic |
1348 | 21 | // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd" |
1349 | 21 | // |
1350 | 21 | if (21 CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)21 ) { |
1351 | 16 | // Remove previous operand |
1352 | 16 | PrevOpnd->Invalidate(); |
1353 | 16 | if (CV16 ) { |
1354 | 10 | *CurrOpnd = XorOpnd(CV); |
1355 | 10 | PrevOpnd = CurrOpnd; |
1356 | 16 | } else { |
1357 | 6 | CurrOpnd->Invalidate(); |
1358 | 6 | PrevOpnd = nullptr; |
1359 | 6 | } |
1360 | 16 | Changed = true; |
1361 | 16 | } |
1362 | 42.2k | } |
1363 | 19.6k | |
1364 | 19.6k | // Step 4: Reassemble the Ops |
1365 | 19.6k | if (Changed19.6k ) { |
1366 | 16 | Ops.clear(); |
1367 | 58 | for (unsigned int i = 0, e = Opnds.size(); i < e58 ; i++42 ) { |
1368 | 42 | XorOpnd &O = Opnds[i]; |
1369 | 42 | if (O.isInvalid()) |
1370 | 22 | continue; |
1371 | 20 | ValueEntry VE(getRank(O.getValue()), O.getValue()); |
1372 | 20 | Ops.push_back(VE); |
1373 | 20 | } |
1374 | 16 | if (!ConstOpnd.isNullValue()16 ) { |
1375 | 10 | Value *C = ConstantInt::get(Ty, ConstOpnd); |
1376 | 10 | ValueEntry VE(getRank(C), C); |
1377 | 10 | Ops.push_back(VE); |
1378 | 10 | } |
1379 | 16 | unsigned Sz = Ops.size(); |
1380 | 16 | if (Sz == 1) |
1381 | 0 | return Ops.back().Op; |
1382 | 16 | if (16 Sz == 016 ) { |
1383 | 4 | assert(ConstOpnd.isNullValue()); |
1384 | 4 | return ConstantInt::get(Ty, ConstOpnd); |
1385 | 4 | } |
1386 | 19.6k | } |
1387 | 19.6k | |
1388 | 19.6k | return nullptr; |
1389 | 19.6k | } |
1390 | | |
1391 | | /// Optimize a series of operands to an 'add' instruction. This |
1392 | | /// optimizes based on identities. If it can be reduced to a single Value, it |
1393 | | /// is returned, otherwise the Ops list is mutated as necessary. |
1394 | | Value *ReassociatePass::OptimizeAdd(Instruction *I, |
1395 | 663k | SmallVectorImpl<ValueEntry> &Ops) { |
1396 | 663k | // Scan the operand lists looking for X and -X pairs. If we find any, we |
1397 | 663k | // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it, |
1398 | 663k | // scan for any |
1399 | 663k | // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. |
1400 | 663k | |
1401 | 2.19M | for (unsigned i = 0, e = Ops.size(); i != e2.19M ; ++i1.52M ) { |
1402 | 1.52M | Value *TheOp = Ops[i].Op; |
1403 | 1.52M | // Check to see if we've seen this operand before. If so, we factor all |
1404 | 1.52M | // instances of the operand together. Due to our sorting criteria, we know |
1405 | 1.52M | // that these need to be next to each other in the vector. |
1406 | 1.52M | if (i+1 != Ops.size() && 1.52M Ops[i+1].Op == TheOp865k ) { |
1407 | 98 | // Rescan the list, remove all instances of this operand from the expr. |
1408 | 98 | unsigned NumFound = 0; |
1409 | 222 | do { |
1410 | 222 | Ops.erase(Ops.begin()+i); |
1411 | 222 | ++NumFound; |
1412 | 222 | } while (i != Ops.size() && 222 Ops[i].Op == TheOp171 ); |
1413 | 98 | |
1414 | 98 | DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); |
1415 | 98 | ++NumFactor; |
1416 | 98 | |
1417 | 98 | // Insert a new multiply. |
1418 | 98 | Type *Ty = TheOp->getType(); |
1419 | 98 | Constant *C = Ty->isIntOrIntVectorTy() ? |
1420 | 98 | ConstantInt::get(Ty, NumFound)82 : ConstantFP::get(Ty, NumFound)16 ; |
1421 | 98 | Instruction *Mul = CreateMul(TheOp, C, "factor", I, I); |
1422 | 98 | |
1423 | 98 | // Now that we have inserted a multiply, optimize it. This allows us to |
1424 | 98 | // handle cases that require multiple factoring steps, such as this: |
1425 | 98 | // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 |
1426 | 98 | RedoInsts.insert(Mul); |
1427 | 98 | |
1428 | 98 | // If every add operand was a duplicate, return the multiply. |
1429 | 98 | if (Ops.empty()) |
1430 | 14 | return Mul; |
1431 | 84 | |
1432 | 84 | // Otherwise, we had some input that didn't have the dupe, such as |
1433 | 84 | // "A + A + B" -> "A*2 + B". Add the new multiply to the list of |
1434 | 84 | // things being added by this operation. |
1435 | 84 | Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); |
1436 | 84 | |
1437 | 84 | --i; |
1438 | 84 | e = Ops.size(); |
1439 | 84 | continue; |
1440 | 84 | } |
1441 | 1.52M | |
1442 | 1.52M | // Check for X and -X or X and ~X in the operand list. |
1443 | 1.52M | if (1.52M !BinaryOperator::isNeg(TheOp) && 1.52M !BinaryOperator::isFNeg(TheOp)1.50M && |
1444 | 1.50M | !BinaryOperator::isNot(TheOp)) |
1445 | 1.50M | continue; |
1446 | 22.4k | |
1447 | 22.4k | Value *X = nullptr; |
1448 | 22.4k | if (BinaryOperator::isNeg(TheOp) || 22.4k BinaryOperator::isFNeg(TheOp)223 ) |
1449 | 22.2k | X = BinaryOperator::getNegArgument(TheOp); |
1450 | 181 | else if (181 BinaryOperator::isNot(TheOp)181 ) |
1451 | 181 | X = BinaryOperator::getNotArgument(TheOp); |
1452 | 22.4k | |
1453 | 22.4k | unsigned FoundX = FindInOperandList(Ops, i, X); |
1454 | 22.4k | if (FoundX == i) |
1455 | 22.4k | continue; |
1456 | 20 | |
1457 | 20 | // Remove X and -X from the operand list. |
1458 | 20 | if (20 Ops.size() == 2 && |
1459 | 4 | (BinaryOperator::isNeg(TheOp) || 4 BinaryOperator::isFNeg(TheOp)1 )) |
1460 | 4 | return Constant::getNullValue(X->getType()); |
1461 | 16 | |
1462 | 16 | // Remove X and ~X from the operand list. |
1463 | 16 | if (16 Ops.size() == 2 && 16 BinaryOperator::isNot(TheOp)0 ) |
1464 | 0 | return Constant::getAllOnesValue(X->getType()); |
1465 | 16 | |
1466 | 16 | Ops.erase(Ops.begin()+i); |
1467 | 16 | if (i < FoundX) |
1468 | 14 | --FoundX; |
1469 | 16 | else |
1470 | 2 | --i; // Need to back up an extra one. |
1471 | 16 | Ops.erase(Ops.begin()+FoundX); |
1472 | 16 | ++NumAnnihil; |
1473 | 16 | --i; // Revisit element. |
1474 | 16 | e -= 2; // Removed two elements. |
1475 | 16 | |
1476 | 16 | // if X and ~X we append -1 to the operand list. |
1477 | 16 | if (BinaryOperator::isNot(TheOp)16 ) { |
1478 | 1 | Value *V = Constant::getAllOnesValue(X->getType()); |
1479 | 1 | Ops.insert(Ops.end(), ValueEntry(getRank(V), V)); |
1480 | 1 | e += 1; |
1481 | 1 | } |
1482 | 1.52M | } |
1483 | 663k | |
1484 | 663k | // Scan the operand list, checking to see if there are any common factors |
1485 | 663k | // between operands. Consider something like A*A+A*B*C+D. We would like to |
1486 | 663k | // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. |
1487 | 663k | // To efficiently find this, we count the number of times a factor occurs |
1488 | 663k | // for any ADD operands that are MULs. |
1489 | 663k | DenseMap<Value*, unsigned> FactorOccurrences; |
1490 | 663k | |
1491 | 663k | // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) |
1492 | 663k | // where they are actually the same multiply. |
1493 | 663k | unsigned MaxOcc = 0; |
1494 | 663k | Value *MaxOccVal = nullptr; |
1495 | 2.19M | for (unsigned i = 0, e = Ops.size(); i != e2.19M ; ++i1.52M ) { |
1496 | 1.52M | BinaryOperator *BOp = |
1497 | 1.52M | isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); |
1498 | 1.52M | if (!BOp) |
1499 | 1.40M | continue; |
1500 | 123k | |
1501 | 123k | // Compute all of the factors of this added value. |
1502 | 123k | SmallVector<Value*, 8> Factors; |
1503 | 123k | FindSingleUseMultiplyFactors(BOp, Factors); |
1504 | 123k | assert(Factors.size() > 1 && "Bad linearize!"); |
1505 | 123k | |
1506 | 123k | // Add one to FactorOccurrences for each unique factor in this op. |
1507 | 123k | SmallPtrSet<Value*, 8> Duplicates; |
1508 | 374k | for (unsigned i = 0, e = Factors.size(); i != e374k ; ++i251k ) { |
1509 | 251k | Value *Factor = Factors[i]; |
1510 | 251k | if (!Duplicates.insert(Factor).second) |
1511 | 10.3k | continue; |
1512 | 240k | |
1513 | 240k | unsigned Occ = ++FactorOccurrences[Factor]; |
1514 | 240k | if (Occ > MaxOcc240k ) { |
1515 | 72.2k | MaxOcc = Occ; |
1516 | 72.2k | MaxOccVal = Factor; |
1517 | 72.2k | } |
1518 | 240k | |
1519 | 240k | // If Factor is a negative constant, add the negated value as a factor |
1520 | 240k | // because we can percolate the negate out. Watch for minint, which |
1521 | 240k | // cannot be positivified. |
1522 | 240k | if (ConstantInt *CI240k = dyn_cast<ConstantInt>(Factor)) { |
1523 | 54.1k | if (CI->isNegative() && 54.1k !CI->isMinValue(true)11.4k ) { |
1524 | 11.4k | Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); |
1525 | 11.4k | if (!Duplicates.insert(Factor).second) |
1526 | 0 | continue; |
1527 | 11.4k | unsigned Occ = ++FactorOccurrences[Factor]; |
1528 | 11.4k | if (Occ > MaxOcc11.4k ) { |
1529 | 28 | MaxOcc = Occ; |
1530 | 28 | MaxOccVal = Factor; |
1531 | 28 | } |
1532 | 11.4k | } |
1533 | 240k | } else if (ConstantFP *186k CF186k = dyn_cast<ConstantFP>(Factor)) { |
1534 | 25 | if (CF->isNegative()25 ) { |
1535 | 12 | APFloat F(CF->getValueAPF()); |
1536 | 12 | F.changeSign(); |
1537 | 12 | Factor = ConstantFP::get(CF->getContext(), F); |
1538 | 12 | if (!Duplicates.insert(Factor).second) |
1539 | 0 | continue; |
1540 | 12 | unsigned Occ = ++FactorOccurrences[Factor]; |
1541 | 12 | if (Occ > MaxOcc12 ) { |
1542 | 3 | MaxOcc = Occ; |
1543 | 3 | MaxOccVal = Factor; |
1544 | 3 | } |
1545 | 12 | } |
1546 | 186k | } |
1547 | 251k | } |
1548 | 1.52M | } |
1549 | 663k | |
1550 | 663k | // If any factor occurred more than one time, we can pull it out. |
1551 | 663k | if (MaxOcc > 1663k ) { |
1552 | 1.21k | DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); |
1553 | 1.21k | ++NumFactor; |
1554 | 1.21k | |
1555 | 1.21k | // Create a new instruction that uses the MaxOccVal twice. If we don't do |
1556 | 1.21k | // this, we could otherwise run into situations where removing a factor |
1557 | 1.21k | // from an expression will drop a use of maxocc, and this can cause |
1558 | 1.21k | // RemoveFactorFromExpression on successive values to behave differently. |
1559 | 1.21k | Instruction *DummyInst = |
1560 | 1.21k | I->getType()->isIntOrIntVectorTy() |
1561 | 1.20k | ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal) |
1562 | 19 | : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal); |
1563 | 1.21k | |
1564 | 1.21k | SmallVector<WeakTrackingVH, 4> NewMulOps; |
1565 | 10.7k | for (unsigned i = 0; i != Ops.size()10.7k ; ++i9.50k ) { |
1566 | 9.50k | // Only try to remove factors from expressions we're allowed to. |
1567 | 9.50k | BinaryOperator *BOp = |
1568 | 9.50k | isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); |
1569 | 9.50k | if (!BOp) |
1570 | 2.94k | continue; |
1571 | 6.55k | |
1572 | 6.55k | if (Value *6.55k V6.55k = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { |
1573 | 6.25k | // The factorized operand may occur several times. Convert them all in |
1574 | 6.25k | // one fell swoop. |
1575 | 70.0k | for (unsigned j = Ops.size(); j != i70.0k ;) { |
1576 | 63.7k | --j; |
1577 | 63.7k | if (Ops[j].Op == Ops[i].Op63.7k ) { |
1578 | 6.25k | NewMulOps.push_back(V); |
1579 | 6.25k | Ops.erase(Ops.begin()+j); |
1580 | 6.25k | } |
1581 | 63.7k | } |
1582 | 6.25k | --i; |
1583 | 6.25k | } |
1584 | 9.50k | } |
1585 | 1.21k | |
1586 | 1.21k | // No need for extra uses anymore. |
1587 | 1.21k | DummyInst->deleteValue(); |
1588 | 1.21k | |
1589 | 1.21k | unsigned NumAddedValues = NewMulOps.size(); |
1590 | 1.21k | Value *V = EmitAddTreeOfValues(I, NewMulOps); |
1591 | 1.21k | |
1592 | 1.21k | // Now that we have inserted the add tree, optimize it. This allows us to |
1593 | 1.21k | // handle cases that require multiple factoring steps, such as this: |
1594 | 1.21k | // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) |
1595 | 1.21k | assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); |
1596 | 1.21k | (void)NumAddedValues; |
1597 | 1.21k | if (Instruction *VI = dyn_cast<Instruction>(V)) |
1598 | 1.21k | RedoInsts.insert(VI); |
1599 | 1.21k | |
1600 | 1.21k | // Create the multiply. |
1601 | 1.21k | Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I); |
1602 | 1.21k | |
1603 | 1.21k | // Rerun associate on the multiply in case the inner expression turned into |
1604 | 1.21k | // a multiply. We want to make sure that we keep things in canonical form. |
1605 | 1.21k | RedoInsts.insert(V2); |
1606 | 1.21k | |
1607 | 1.21k | // If every add operand included the factor (e.g. "A*B + A*C"), then the |
1608 | 1.21k | // entire result expression is just the multiply "A*(B+C)". |
1609 | 1.21k | if (Ops.empty()) |
1610 | 32 | return V2; |
1611 | 1.18k | |
1612 | 1.18k | // Otherwise, we had some input that didn't have the factor, such as |
1613 | 1.18k | // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of |
1614 | 1.18k | // things being added by this operation. |
1615 | 1.18k | Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); |
1616 | 1.18k | } |
1617 | 663k | |
1618 | 663k | return nullptr; |
1619 | 663k | } |
1620 | | |
1621 | | /// \brief Build up a vector of value/power pairs factoring a product. |
1622 | | /// |
1623 | | /// Given a series of multiplication operands, build a vector of factors and |
1624 | | /// the powers each is raised to when forming the final product. Sort them in |
1625 | | /// the order of descending power. |
1626 | | /// |
1627 | | /// (x*x) -> [(x, 2)] |
1628 | | /// ((x*x)*x) -> [(x, 3)] |
1629 | | /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)] |
1630 | | /// |
1631 | | /// \returns Whether any factors have a power greater than one. |
1632 | | static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, |
1633 | 8.63k | SmallVectorImpl<Factor> &Factors) { |
1634 | 8.63k | // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this. |
1635 | 8.63k | // Compute the sum of powers of simplifiable factors. |
1636 | 8.63k | unsigned FactorPowerSum = 0; |
1637 | 34.6k | for (unsigned Idx = 1, Size = Ops.size(); Idx < Size34.6k ; ++Idx25.9k ) { |
1638 | 25.9k | Value *Op = Ops[Idx-1].Op; |
1639 | 25.9k | |
1640 | 25.9k | // Count the number of occurrences of this value. |
1641 | 25.9k | unsigned Count = 1; |
1642 | 26.0k | for (; Idx < Size && 26.0k Ops[Idx].Op == Op26.0k ; ++Idx99 ) |
1643 | 99 | ++Count; |
1644 | 25.9k | // Track for simplification all factors which occur 2 or more times. |
1645 | 25.9k | if (Count > 1) |
1646 | 49 | FactorPowerSum += Count; |
1647 | 25.9k | } |
1648 | 8.63k | |
1649 | 8.63k | // We can only simplify factors if the sum of the powers of our simplifiable |
1650 | 8.63k | // factors is 4 or higher. When that is the case, we will *always* have |
1651 | 8.63k | // a simplification. This is an important invariant to prevent cyclicly |
1652 | 8.63k | // trying to simplify already minimal formations. |
1653 | 8.63k | if (FactorPowerSum < 4) |
1654 | 8.61k | return false; |
1655 | 17 | |
1656 | 17 | // Now gather the simplifiable factors, removing them from Ops. |
1657 | 17 | FactorPowerSum = 0; |
1658 | 42 | for (unsigned Idx = 1; Idx < Ops.size()42 ; ++Idx25 ) { |
1659 | 25 | Value *Op = Ops[Idx-1].Op; |
1660 | 25 | |
1661 | 25 | // Count the number of occurrences of this value. |
1662 | 25 | unsigned Count = 1; |
1663 | 99 | for (; Idx < Ops.size() && 99 Ops[Idx].Op == Op82 ; ++Idx74 ) |
1664 | 74 | ++Count; |
1665 | 25 | if (Count == 1) |
1666 | 1 | continue; |
1667 | 24 | // Move an even number of occurrences to Factors. |
1668 | 24 | Count &= ~1U; |
1669 | 24 | Idx -= Count; |
1670 | 24 | FactorPowerSum += Count; |
1671 | 24 | Factors.push_back(Factor(Op, Count)); |
1672 | 24 | Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count); |
1673 | 24 | } |
1674 | 17 | |
1675 | 17 | // None of the adjustments above should have reduced the sum of factor powers |
1676 | 17 | // below our mininum of '4'. |
1677 | 17 | assert(FactorPowerSum >= 4); |
1678 | 17 | |
1679 | 17 | std::stable_sort(Factors.begin(), Factors.end(), |
1680 | 8 | [](const Factor &LHS, const Factor &RHS) { |
1681 | 8 | return LHS.Power > RHS.Power; |
1682 | 8 | }); |
1683 | 8.63k | return true; |
1684 | 8.63k | } |
1685 | | |
1686 | | /// \brief Build a tree of multiplies, computing the product of Ops. |
1687 | | static Value *buildMultiplyTree(IRBuilder<> &Builder, |
1688 | 36 | SmallVectorImpl<Value*> &Ops) { |
1689 | 36 | if (Ops.size() == 1) |
1690 | 0 | return Ops.back(); |
1691 | 36 | |
1692 | 36 | Value *LHS = Ops.pop_back_val(); |
1693 | 44 | do { |
1694 | 44 | if (LHS->getType()->isIntOrIntVectorTy()) |
1695 | 42 | LHS = Builder.CreateMul(LHS, Ops.pop_back_val()); |
1696 | 44 | else |
1697 | 2 | LHS = Builder.CreateFMul(LHS, Ops.pop_back_val()); |
1698 | 44 | } while (!Ops.empty()); |
1699 | 36 | |
1700 | 36 | return LHS; |
1701 | 36 | } |
1702 | | |
1703 | | /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*... |
1704 | | /// |
1705 | | /// Given a vector of values raised to various powers, where no two values are |
1706 | | /// equal and the powers are sorted in decreasing order, compute the minimal |
1707 | | /// DAG of multiplies to compute the final product, and return that product |
1708 | | /// value. |
1709 | | Value * |
1710 | | ReassociatePass::buildMinimalMultiplyDAG(IRBuilder<> &Builder, |
1711 | 49 | SmallVectorImpl<Factor> &Factors) { |
1712 | 49 | assert(Factors[0].Power); |
1713 | 49 | SmallVector<Value *, 4> OuterProduct; |
1714 | 49 | for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size(); |
1715 | 59 | Idx < Size && 59 Factors[Idx].Power > 013 ; ++Idx10 ) { |
1716 | 10 | if (Factors[Idx].Power != Factors[LastIdx].Power10 ) { |
1717 | 6 | LastIdx = Idx; |
1718 | 6 | continue; |
1719 | 6 | } |
1720 | 4 | |
1721 | 4 | // We want to multiply across all the factors with the same power so that |
1722 | 4 | // we can raise them to that power as a single entity. Build a mini tree |
1723 | 4 | // for that. |
1724 | 4 | SmallVector<Value *, 4> InnerProduct; |
1725 | 4 | InnerProduct.push_back(Factors[LastIdx].Base); |
1726 | 4 | do { |
1727 | 4 | InnerProduct.push_back(Factors[Idx].Base); |
1728 | 4 | ++Idx; |
1729 | 4 | } while (Idx < Size && 4 Factors[Idx].Power == Factors[LastIdx].Power0 ); |
1730 | 4 | |
1731 | 4 | // Reset the base value of the first factor to the new expression tree. |
1732 | 4 | // We'll remove all the factors with the same power in a second pass. |
1733 | 4 | Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct); |
1734 | 4 | if (Instruction *MI = dyn_cast<Instruction>(M)) |
1735 | 4 | RedoInsts.insert(MI); |
1736 | 10 | |
1737 | 10 | LastIdx = Idx; |
1738 | 10 | } |
1739 | 49 | // Unique factors with equal powers -- we've folded them into the first one's |
1740 | 49 | // base. |
1741 | 49 | Factors.erase(std::unique(Factors.begin(), Factors.end(), |
1742 | 13 | [](const Factor &LHS, const Factor &RHS) { |
1743 | 13 | return LHS.Power == RHS.Power; |
1744 | 13 | }), |
1745 | 49 | Factors.end()); |
1746 | 49 | |
1747 | 49 | // Iteratively collect the base of each factor with an add power into the |
1748 | 49 | // outer product, and halve each power in preparation for squaring the |
1749 | 49 | // expression. |
1750 | 107 | for (unsigned Idx = 0, Size = Factors.size(); Idx != Size107 ; ++Idx58 ) { |
1751 | 58 | if (Factors[Idx].Power & 1) |
1752 | 25 | OuterProduct.push_back(Factors[Idx].Base); |
1753 | 58 | Factors[Idx].Power >>= 1; |
1754 | 58 | } |
1755 | 49 | if (Factors[0].Power49 ) { |
1756 | 32 | Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors); |
1757 | 32 | OuterProduct.push_back(SquareRoot); |
1758 | 32 | OuterProduct.push_back(SquareRoot); |
1759 | 32 | } |
1760 | 49 | if (OuterProduct.size() == 1) |
1761 | 17 | return OuterProduct.front(); |
1762 | 32 | |
1763 | 32 | Value *V = buildMultiplyTree(Builder, OuterProduct); |
1764 | 32 | return V; |
1765 | 32 | } |
1766 | | |
1767 | | Value *ReassociatePass::OptimizeMul(BinaryOperator *I, |
1768 | 162k | SmallVectorImpl<ValueEntry> &Ops) { |
1769 | 162k | // We can only optimize the multiplies when there is a chain of more than |
1770 | 162k | // three, such that a balanced tree might require fewer total multiplies. |
1771 | 162k | if (Ops.size() < 4) |
1772 | 153k | return nullptr; |
1773 | 8.63k | |
1774 | 8.63k | // Try to turn linear trees of multiplies without other uses of the |
1775 | 8.63k | // intermediate stages into minimal multiply DAGs with perfect sub-expression |
1776 | 8.63k | // re-use. |
1777 | 8.63k | SmallVector<Factor, 4> Factors; |
1778 | 8.63k | if (!collectMultiplyFactors(Ops, Factors)) |
1779 | 8.61k | return nullptr; // All distinct factors, so nothing left for us to do. |
1780 | 16 | |
1781 | 16 | IRBuilder<> Builder(I); |
1782 | 16 | // The reassociate transformation for FP operations is performed only |
1783 | 16 | // if unsafe algebra is permitted by FastMathFlags. Propagate those flags |
1784 | 16 | // to the newly generated operations. |
1785 | 16 | if (auto FPI = dyn_cast<FPMathOperator>(I)) |
1786 | 1 | Builder.setFastMathFlags(FPI->getFastMathFlags()); |
1787 | 16 | |
1788 | 16 | Value *V = buildMinimalMultiplyDAG(Builder, Factors); |
1789 | 16 | if (Ops.empty()) |
1790 | 10 | return V; |
1791 | 6 | |
1792 | 6 | ValueEntry NewEntry = ValueEntry(getRank(V), V); |
1793 | 6 | Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry); |
1794 | 6 | return nullptr; |
1795 | 6 | } |
1796 | | |
1797 | | Value *ReassociatePass::OptimizeExpression(BinaryOperator *I, |
1798 | 1.00M | SmallVectorImpl<ValueEntry> &Ops) { |
1799 | 1.00M | // Now that we have the linearized expression tree, try to optimize it. |
1800 | 1.00M | // Start by folding any constants that we found. |
1801 | 1.00M | Constant *Cst = nullptr; |
1802 | 1.00M | unsigned Opcode = I->getOpcode(); |
1803 | 1.71M | while (!Ops.empty() && 1.71M isa<Constant>(Ops.back().Op)1.71M ) { |
1804 | 705k | Constant *C = cast<Constant>(Ops.pop_back_val().Op); |
1805 | 705k | Cst = Cst ? ConstantExpr::get(Opcode, C, Cst)4.40k : C700k ; |
1806 | 705k | } |
1807 | 1.00M | // If there was nothing but constants then we are done. |
1808 | 1.00M | if (Ops.empty()) |
1809 | 19 | return Cst; |
1810 | 1.00M | |
1811 | 1.00M | // Put the combined constant back at the end of the operand list, except if |
1812 | 1.00M | // there is no point. For example, an add of 0 gets dropped here, while a |
1813 | 1.00M | // multiplication by zero turns the whole expression into zero. |
1814 | 1.00M | if (1.00M Cst && 1.00M Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())700k ) { |
1815 | 700k | if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType())) |
1816 | 1 | return Cst; |
1817 | 700k | Ops.push_back(ValueEntry(0, Cst)); |
1818 | 700k | } |
1819 | 1.00M | |
1820 | 1.00M | if (1.00M Ops.size() == 11.00M ) return Ops[0].Op7 ; |
1821 | 1.00M | |
1822 | 1.00M | // Handle destructive annihilation due to identities between elements in the |
1823 | 1.00M | // argument list here. |
1824 | 1.00M | unsigned NumOps = Ops.size(); |
1825 | 1.00M | switch (Opcode) { |
1826 | 0 | default: break; |
1827 | 162k | case Instruction::And: |
1828 | 162k | case Instruction::Or: |
1829 | 162k | if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) |
1830 | 2 | return Result; |
1831 | 162k | break; |
1832 | 162k | |
1833 | 19.6k | case Instruction::Xor: |
1834 | 19.6k | if (Value *Result = OptimizeXor(I, Ops)) |
1835 | 4 | return Result; |
1836 | 19.6k | break; |
1837 | 19.6k | |
1838 | 663k | case Instruction::Add: |
1839 | 663k | case Instruction::FAdd: |
1840 | 663k | if (Value *Result = OptimizeAdd(I, Ops)) |
1841 | 50 | return Result; |
1842 | 663k | break; |
1843 | 663k | |
1844 | 162k | case Instruction::Mul: |
1845 | 162k | case Instruction::FMul: |
1846 | 162k | if (Value *Result = OptimizeMul(I, Ops)) |
1847 | 10 | return Result; |
1848 | 162k | break; |
1849 | 1.00M | } |
1850 | 1.00M | |
1851 | 1.00M | if (1.00M Ops.size() != NumOps1.00M ) |
1852 | 1.26k | return OptimizeExpression(I, Ops); |
1853 | 1.00M | return nullptr; |
1854 | 1.00M | } |
1855 | | |
1856 | | // Remove dead instructions and if any operands are trivially dead add them to |
1857 | | // Insts so they will be removed as well. |
1858 | | void ReassociatePass::RecursivelyEraseDeadInsts( |
1859 | 41.5k | Instruction *I, SetVector<AssertingVH<Instruction>> &Insts) { |
1860 | 41.5k | assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!"); |
1861 | 41.5k | SmallVector<Value *, 4> Ops(I->op_begin(), I->op_end()); |
1862 | 41.5k | ValueRankMap.erase(I); |
1863 | 41.5k | Insts.remove(I); |
1864 | 41.5k | RedoInsts.remove(I); |
1865 | 41.5k | I->eraseFromParent(); |
1866 | 41.5k | for (auto Op : Ops) |
1867 | 83.1k | if (Instruction *83.1k OpInst83.1k = dyn_cast<Instruction>(Op)) |
1868 | 18.0k | if (18.0k OpInst->use_empty()18.0k ) |
1869 | 12.2k | Insts.insert(OpInst); |
1870 | 41.5k | } |
1871 | | |
1872 | | /// Zap the given instruction, adding interesting operands to the work list. |
1873 | 3.89k | void ReassociatePass::EraseInst(Instruction *I) { |
1874 | 3.89k | assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!"); |
1875 | 3.89k | DEBUG(dbgs() << "Erasing dead inst: "; I->dump()); |
1876 | 3.89k | |
1877 | 3.89k | SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); |
1878 | 3.89k | // Erase the dead instruction. |
1879 | 3.89k | ValueRankMap.erase(I); |
1880 | 3.89k | RedoInsts.remove(I); |
1881 | 3.89k | I->eraseFromParent(); |
1882 | 3.89k | // Optimize its operands. |
1883 | 3.89k | SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes. |
1884 | 11.7k | for (unsigned i = 0, e = Ops.size(); i != e11.7k ; ++i7.81k ) |
1885 | 7.81k | if (Instruction *7.81k Op7.81k = dyn_cast<Instruction>(Ops[i])) { |
1886 | 5.04k | // If this is a node in an expression tree, climb to the expression root |
1887 | 5.04k | // and add that since that's where optimization actually happens. |
1888 | 5.04k | unsigned Opcode = Op->getOpcode(); |
1889 | 5.12k | while (Op->hasOneUse() && 5.12k Op->user_back()->getOpcode() == Opcode3.17k && |
1890 | 111 | Visited.insert(Op).second) |
1891 | 78 | Op = Op->user_back(); |
1892 | 7.81k | RedoInsts.insert(Op); |
1893 | 7.81k | } |
1894 | 3.89k | |
1895 | 3.89k | MadeChange = true; |
1896 | 3.89k | } |
1897 | | |
1898 | | // Canonicalize expressions of the following form: |
1899 | | // x + (-Constant * y) -> x - (Constant * y) |
1900 | | // x - (-Constant * y) -> x + (Constant * y) |
1901 | 1.79M | Instruction *ReassociatePass::canonicalizeNegConstExpr(Instruction *I) { |
1902 | 1.79M | if (!I->hasOneUse() || 1.79M I->getType()->isVectorTy()1.36M ) |
1903 | 427k | return nullptr; |
1904 | 1.36M | |
1905 | 1.36M | // Must be a fmul or fdiv instruction. |
1906 | 1.36M | unsigned Opcode = I->getOpcode(); |
1907 | 1.36M | if (Opcode != Instruction::FMul && 1.36M Opcode != Instruction::FDiv1.32M ) |
1908 | 1.29M | return nullptr; |
1909 | 65.3k | |
1910 | 65.3k | auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0)); |
1911 | 65.3k | auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1)); |
1912 | 65.3k | |
1913 | 65.3k | // Both operands are constant, let it get constant folded away. |
1914 | 65.3k | if (C0 && 65.3k C114.6k ) |
1915 | 0 | return nullptr; |
1916 | 65.3k | |
1917 | 65.3k | ConstantFP *CF = C0 ? 65.3k C014.6k : C150.7k ; |
1918 | 65.3k | |
1919 | 65.3k | // Must have one constant operand. |
1920 | 65.3k | if (!CF) |
1921 | 32.9k | return nullptr; |
1922 | 32.4k | |
1923 | 32.4k | // Must be a negative ConstantFP. |
1924 | 32.4k | if (32.4k !CF->isNegative()32.4k ) |
1925 | 32.3k | return nullptr; |
1926 | 131 | |
1927 | 131 | // User must be a binary operator with one or more uses. |
1928 | 131 | Instruction *User = I->user_back(); |
1929 | 131 | if (!isa<BinaryOperator>(User) || 131 User->use_empty()93 ) |
1930 | 39 | return nullptr; |
1931 | 92 | |
1932 | 92 | unsigned UserOpcode = User->getOpcode(); |
1933 | 92 | if (UserOpcode != Instruction::FAdd && 92 UserOpcode != Instruction::FSub46 ) |
1934 | 31 | return nullptr; |
1935 | 61 | |
1936 | 61 | // Subtraction is not commutative. Explicitly, the following transform is |
1937 | 61 | // not valid: (-Constant * y) - x -> x + (Constant * y) |
1938 | 61 | if (61 !User->isCommutative() && 61 User->getOperand(1) != I15 ) |
1939 | 4 | return nullptr; |
1940 | 57 | |
1941 | 57 | // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the |
1942 | 57 | // resulting subtract will be broken up later. This can get us into an |
1943 | 57 | // infinite loop during reassociation. |
1944 | 57 | if (57 UserOpcode == Instruction::FAdd && 57 ShouldBreakUpSubtract(User)46 ) |
1945 | 1 | return nullptr; |
1946 | 56 | |
1947 | 56 | // Change the sign of the constant. |
1948 | 56 | APFloat Val = CF->getValueAPF(); |
1949 | 56 | Val.changeSign(); |
1950 | 56 | I->setOperand(C0 ? 05 : 151 , ConstantFP::get(CF->getContext(), Val)); |
1951 | 56 | |
1952 | 56 | // Canonicalize I to RHS to simplify the next bit of logic. E.g., |
1953 | 56 | // ((-Const*y) + x) -> (x + (-Const*y)). |
1954 | 56 | if (User->getOperand(0) == I && 56 User->isCommutative()29 ) |
1955 | 29 | cast<BinaryOperator>(User)->swapOperands(); |
1956 | 56 | |
1957 | 56 | Value *Op0 = User->getOperand(0); |
1958 | 56 | Value *Op1 = User->getOperand(1); |
1959 | 56 | BinaryOperator *NI; |
1960 | 56 | switch (UserOpcode) { |
1961 | 0 | default: |
1962 | 0 | llvm_unreachable("Unexpected Opcode!"); |
1963 | 45 | case Instruction::FAdd: |
1964 | 45 | NI = BinaryOperator::CreateFSub(Op0, Op1); |
1965 | 45 | NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags()); |
1966 | 45 | break; |
1967 | 11 | case Instruction::FSub: |
1968 | 11 | NI = BinaryOperator::CreateFAdd(Op0, Op1); |
1969 | 11 | NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags()); |
1970 | 11 | break; |
1971 | 56 | } |
1972 | 56 | |
1973 | 56 | NI->insertBefore(User); |
1974 | 56 | NI->setName(User->getName()); |
1975 | 56 | User->replaceAllUsesWith(NI); |
1976 | 56 | NI->setDebugLoc(I->getDebugLoc()); |
1977 | 56 | RedoInsts.insert(I); |
1978 | 56 | MadeChange = true; |
1979 | 56 | return NI; |
1980 | 56 | } |
1981 | | |
1982 | | /// Inspect and optimize the given instruction. Note that erasing |
1983 | | /// instructions is not allowed. |
1984 | 18.8M | void ReassociatePass::OptimizeInst(Instruction *I) { |
1985 | 18.8M | // Only consider operations that we understand. |
1986 | 18.8M | if (!isa<BinaryOperator>(I)) |
1987 | 17.0M | return; |
1988 | 1.79M | |
1989 | 1.79M | if (1.79M I->getOpcode() == Instruction::Shl && 1.79M isa<ConstantInt>(I->getOperand(1))112k ) |
1990 | 1.79M | // If an operand of this shift is a reassociable multiply, or if the shift |
1991 | 1.79M | // is used by a reassociable multiply or add, turn into a multiply. |
1992 | 92.8k | if (92.8k isReassociableOp(I->getOperand(0), Instruction::Mul) || |
1993 | 92.6k | (I->hasOneUse() && |
1994 | 82.0k | (isReassociableOp(I->user_back(), Instruction::Mul) || |
1995 | 92.8k | isReassociableOp(I->user_back(), Instruction::Add)81.9k ))) { |
1996 | 18.4k | Instruction *NI = ConvertShiftToMul(I); |
1997 | 18.4k | RedoInsts.insert(I); |
1998 | 18.4k | MadeChange = true; |
1999 | 18.4k | I = NI; |
2000 | 18.4k | } |
2001 | 1.79M | |
2002 | 1.79M | // Canonicalize negative constants out of expressions. |
2003 | 1.79M | if (Instruction *Res = canonicalizeNegConstExpr(I)) |
2004 | 56 | I = Res; |
2005 | 1.79M | |
2006 | 1.79M | // Commute binary operators, to canonicalize the order of their operands. |
2007 | 1.79M | // This can potentially expose more CSE opportunities, and makes writing other |
2008 | 1.79M | // transformations simpler. |
2009 | 1.79M | if (I->isCommutative()) |
2010 | 1.35M | canonicalizeOperands(I); |
2011 | 1.79M | |
2012 | 1.79M | // Don't optimize floating point instructions that don't have unsafe algebra. |
2013 | 1.79M | if (I->getType()->isFPOrFPVectorTy() && 1.79M !I->hasUnsafeAlgebra()147k ) |
2014 | 146k | return; |
2015 | 1.64M | |
2016 | 1.64M | // Do not reassociate boolean (i1) expressions. We want to preserve the |
2017 | 1.64M | // original order of evaluation for short-circuited comparisons that |
2018 | 1.64M | // SimplifyCFG has folded to AND/OR expressions. If the expression |
2019 | 1.64M | // is not further optimized, it is likely to be transformed back to a |
2020 | 1.64M | // short-circuited form for code gen, and the source order may have been |
2021 | 1.64M | // optimized for the most likely conditions. |
2022 | 1.64M | if (1.64M I->getType()->isIntegerTy(1)1.64M ) |
2023 | 123k | return; |
2024 | 1.52M | |
2025 | 1.52M | // If this is a subtract instruction which is not already in negate form, |
2026 | 1.52M | // see if we can convert it to X+-Y. |
2027 | 1.52M | if (1.52M I->getOpcode() == Instruction::Sub1.52M ) { |
2028 | 155k | if (ShouldBreakUpSubtract(I)155k ) { |
2029 | 9.31k | Instruction *NI = BreakUpSubtract(I, RedoInsts); |
2030 | 9.31k | RedoInsts.insert(I); |
2031 | 9.31k | MadeChange = true; |
2032 | 9.31k | I = NI; |
2033 | 155k | } else if (145k BinaryOperator::isNeg(I)145k ) { |
2034 | 36.6k | // Otherwise, this is a negation. See if the operand is a multiply tree |
2035 | 36.6k | // and if this is not an inner node of a multiply tree. |
2036 | 36.6k | if (isReassociableOp(I->getOperand(1), Instruction::Mul) && |
2037 | 1.14k | (!I->hasOneUse() || |
2038 | 36.6k | !isReassociableOp(I->user_back(), Instruction::Mul)1.10k )) { |
2039 | 1.14k | Instruction *NI = LowerNegateToMultiply(I); |
2040 | 1.14k | // If the negate was simplified, revisit the users to see if we can |
2041 | 1.14k | // reassociate further. |
2042 | 1.19k | for (User *U : NI->users()) { |
2043 | 1.19k | if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U)) |
2044 | 1.19k | RedoInsts.insert(Tmp); |
2045 | 1.19k | } |
2046 | 1.14k | RedoInsts.insert(I); |
2047 | 1.14k | MadeChange = true; |
2048 | 1.14k | I = NI; |
2049 | 1.14k | } |
2050 | 145k | } |
2051 | 1.52M | } else if (1.36M I->getOpcode() == Instruction::FSub1.36M ) { |
2052 | 57 | if (ShouldBreakUpSubtract(I)57 ) { |
2053 | 23 | Instruction *NI = BreakUpSubtract(I, RedoInsts); |
2054 | 23 | RedoInsts.insert(I); |
2055 | 23 | MadeChange = true; |
2056 | 23 | I = NI; |
2057 | 57 | } else if (34 BinaryOperator::isFNeg(I)34 ) { |
2058 | 26 | // Otherwise, this is a negation. See if the operand is a multiply tree |
2059 | 26 | // and if this is not an inner node of a multiply tree. |
2060 | 26 | if (isReassociableOp(I->getOperand(1), Instruction::FMul) && |
2061 | 10 | (!I->hasOneUse() || |
2062 | 26 | !isReassociableOp(I->user_back(), Instruction::FMul)10 )) { |
2063 | 8 | // If the negate was simplified, revisit the users to see if we can |
2064 | 8 | // reassociate further. |
2065 | 8 | Instruction *NI = LowerNegateToMultiply(I); |
2066 | 8 | for (User *U : NI->users()) { |
2067 | 8 | if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U)) |
2068 | 5 | RedoInsts.insert(Tmp); |
2069 | 8 | } |
2070 | 8 | RedoInsts.insert(I); |
2071 | 8 | MadeChange = true; |
2072 | 8 | I = NI; |
2073 | 8 | } |
2074 | 34 | } |
2075 | 1.36M | } |
2076 | 1.52M | |
2077 | 1.52M | // If this instruction is an associative binary operator, process it. |
2078 | 1.52M | if (!I->isAssociative()1.52M ) return373k ; |
2079 | 1.14M | BinaryOperator *BO = cast<BinaryOperator>(I); |
2080 | 1.14M | |
2081 | 1.14M | // If this is an interior node of a reassociable tree, ignore it until we |
2082 | 1.14M | // get to the root of the tree, to avoid N^2 analysis. |
2083 | 1.14M | unsigned Opcode = BO->getOpcode(); |
2084 | 1.14M | if (BO->hasOneUse() && 1.14M BO->user_back()->getOpcode() == Opcode815k ) { |
2085 | 136k | // During the initial run we will get to the root of the tree. |
2086 | 136k | // But if we get here while we are redoing instructions, there is no |
2087 | 136k | // guarantee that the root will be visited. So Redo later |
2088 | 136k | if (BO->user_back() != BO && |
2089 | 136k | BO->getParent() == BO->user_back()->getParent()) |
2090 | 135k | RedoInsts.insert(BO->user_back()); |
2091 | 136k | return; |
2092 | 136k | } |
2093 | 1.01M | |
2094 | 1.01M | // If this is an add tree that is used by a sub instruction, ignore it |
2095 | 1.01M | // until we process the subtract. |
2096 | 1.01M | if (1.01M BO->hasOneUse() && 1.01M BO->getOpcode() == Instruction::Add679k && |
2097 | 418k | cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub) |
2098 | 3.86k | return; |
2099 | 1.00M | if (1.00M BO->hasOneUse() && 1.00M BO->getOpcode() == Instruction::FAdd675k && |
2100 | 117 | cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub) |
2101 | 18 | return; |
2102 | 1.00M | |
2103 | 1.00M | ReassociateExpression(BO); |
2104 | 1.00M | } |
2105 | | |
2106 | 1.00M | void ReassociatePass::ReassociateExpression(BinaryOperator *I) { |
2107 | 1.00M | // First, walk the expression tree, linearizing the tree, collecting the |
2108 | 1.00M | // operand information. |
2109 | 1.00M | SmallVector<RepeatedValue, 8> Tree; |
2110 | 1.00M | MadeChange |= LinearizeExprTree(I, Tree); |
2111 | 1.00M | SmallVector<ValueEntry, 8> Ops; |
2112 | 1.00M | Ops.reserve(Tree.size()); |
2113 | 3.27M | for (unsigned i = 0, e = Tree.size(); i != e3.27M ; ++i2.27M ) { |
2114 | 2.27M | RepeatedValue E = Tree[i]; |
2115 | 2.27M | Ops.append(E.second.getZExtValue(), |
2116 | 2.27M | ValueEntry(getRank(E.first), E.first)); |
2117 | 2.27M | } |
2118 | 1.00M | |
2119 | 1.00M | DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n'); |
2120 | 1.00M | |
2121 | 1.00M | // Now that we have linearized the tree to a list and have gathered all of |
2122 | 1.00M | // the operands and their ranks, sort the operands by their rank. Use a |
2123 | 1.00M | // stable_sort so that values with equal ranks will have their relative |
2124 | 1.00M | // positions maintained (and so the compiler is deterministic). Note that |
2125 | 1.00M | // this sorts so that the highest ranking values end up at the beginning of |
2126 | 1.00M | // the vector. |
2127 | 1.00M | std::stable_sort(Ops.begin(), Ops.end()); |
2128 | 1.00M | |
2129 | 1.00M | // Now that we have the expression tree in a convenient |
2130 | 1.00M | // sorted form, optimize it globally if possible. |
2131 | 1.00M | if (Value *V1.00M = OptimizeExpression(I, Ops)) { |
2132 | 93 | if (V == I) |
2133 | 93 | // Self-referential expression in unreachable code. |
2134 | 0 | return; |
2135 | 93 | // This expression tree simplified to something that isn't a tree, |
2136 | 93 | // eliminate it. |
2137 | 93 | DEBUG93 (dbgs() << "Reassoc to scalar: " << *V << '\n'); |
2138 | 93 | I->replaceAllUsesWith(V); |
2139 | 93 | if (Instruction *VI = dyn_cast<Instruction>(V)) |
2140 | 60 | if (60 I->getDebugLoc()60 ) |
2141 | 0 | VI->setDebugLoc(I->getDebugLoc()); |
2142 | 93 | RedoInsts.insert(I); |
2143 | 93 | ++NumAnnihil; |
2144 | 93 | return; |
2145 | 93 | } |
2146 | 1.00M | |
2147 | 1.00M | // We want to sink immediates as deeply as possible except in the case where |
2148 | 1.00M | // this is a multiply tree used only by an add, and the immediate is a -1. |
2149 | 1.00M | // In this case we reassociate to put the negation on the outside so that we |
2150 | 1.00M | // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y |
2151 | 1.00M | if (1.00M I->hasOneUse()1.00M ) { |
2152 | 675k | if (I->getOpcode() == Instruction::Mul && |
2153 | 130k | cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add && |
2154 | 68.6k | isa<ConstantInt>(Ops.back().Op) && |
2155 | 675k | cast<ConstantInt>(Ops.back().Op)->isMinusOne()43.3k ) { |
2156 | 1.87k | ValueEntry Tmp = Ops.pop_back_val(); |
2157 | 1.87k | Ops.insert(Ops.begin(), Tmp); |
2158 | 675k | } else if (673k I->getOpcode() == Instruction::FMul && |
2159 | 141 | cast<Instruction>(I->user_back())->getOpcode() == |
2160 | 141 | Instruction::FAdd && |
2161 | 52 | isa<ConstantFP>(Ops.back().Op) && |
2162 | 673k | cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)21 ) { |
2163 | 2 | ValueEntry Tmp = Ops.pop_back_val(); |
2164 | 2 | Ops.insert(Ops.begin(), Tmp); |
2165 | 2 | } |
2166 | 675k | } |
2167 | 1.00M | |
2168 | 1.00M | DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n'); |
2169 | 1.00M | |
2170 | 1.00M | if (Ops.size() == 11.00M ) { |
2171 | 0 | if (Ops[0].Op == I) |
2172 | 0 | // Self-referential expression in unreachable code. |
2173 | 0 | return; |
2174 | 0 |
|
2175 | 0 | // This expression tree simplified to something that isn't a tree, |
2176 | 0 | // eliminate it. |
2177 | 0 | I->replaceAllUsesWith(Ops[0].Op); |
2178 | 0 | if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op)) |
2179 | 0 | OI->setDebugLoc(I->getDebugLoc()); |
2180 | 0 | RedoInsts.insert(I); |
2181 | 0 | return; |
2182 | 0 | } |
2183 | 1.00M | |
2184 | 1.00M | // Now that we ordered and optimized the expressions, splat them back into |
2185 | 1.00M | // the expression tree, removing any unneeded nodes. |
2186 | 1.00M | RewriteExprTree(I, Ops); |
2187 | 1.00M | } |
2188 | | |
2189 | 516k | PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) { |
2190 | 516k | // Get the functions basic blocks in Reverse Post Order. This order is used by |
2191 | 516k | // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic |
2192 | 516k | // blocks (it has been seen that the analysis in this pass could hang when |
2193 | 516k | // analysing dead basic blocks). |
2194 | 516k | ReversePostOrderTraversal<Function *> RPOT(&F); |
2195 | 516k | |
2196 | 516k | // Calculate the rank map for F. |
2197 | 516k | BuildRankMap(F, RPOT); |
2198 | 516k | |
2199 | 516k | MadeChange = false; |
2200 | 516k | // Traverse the same blocks that was analysed by BuildRankMap. |
2201 | 3.62M | for (BasicBlock *BI : RPOT) { |
2202 | 3.62M | assert(RankMap.count(&*BI) && "BB should be ranked."); |
2203 | 3.62M | // Optimize every instruction in the basic block. |
2204 | 22.3M | for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;) |
2205 | 18.7M | if (18.7M isInstructionTriviallyDead(&*II)18.7M ) { |
2206 | 199 | EraseInst(&*II++); |
2207 | 18.7M | } else { |
2208 | 18.7M | OptimizeInst(&*II); |
2209 | 18.7M | assert(II->getParent() == &*BI && "Moved to a different block!"); |
2210 | 18.7M | ++II; |
2211 | 18.7M | } |
2212 | 3.62M | |
2213 | 3.62M | // Make a copy of all the instructions to be redone so we can remove dead |
2214 | 3.62M | // instructions. |
2215 | 3.62M | SetVector<AssertingVH<Instruction>> ToRedo(RedoInsts); |
2216 | 3.62M | // Iterate over all instructions to be reevaluated and remove trivially dead |
2217 | 3.62M | // instructions. If any operand of the trivially dead instruction becomes |
2218 | 3.62M | // dead mark it for deletion as well. Continue this process until all |
2219 | 3.62M | // trivially dead instructions have been removed. |
2220 | 3.75M | while (!ToRedo.empty()3.75M ) { |
2221 | 133k | Instruction *I = ToRedo.pop_back_val(); |
2222 | 133k | if (isInstructionTriviallyDead(I)133k ) { |
2223 | 41.5k | RecursivelyEraseDeadInsts(I, ToRedo); |
2224 | 41.5k | MadeChange = true; |
2225 | 41.5k | } |
2226 | 133k | } |
2227 | 3.62M | |
2228 | 3.62M | // Now that we have removed dead instructions, we can reoptimize the |
2229 | 3.62M | // remaining instructions. |
2230 | 3.77M | while (!RedoInsts.empty()3.77M ) { |
2231 | 148k | Instruction *I = RedoInsts.pop_back_val(); |
2232 | 148k | if (isInstructionTriviallyDead(I)) |
2233 | 3.69k | EraseInst(I); |
2234 | 148k | else |
2235 | 145k | OptimizeInst(I); |
2236 | 148k | } |
2237 | 3.62M | } |
2238 | 516k | |
2239 | 516k | // We are done with the rank map. |
2240 | 516k | RankMap.clear(); |
2241 | 516k | ValueRankMap.clear(); |
2242 | 516k | |
2243 | 516k | if (MadeChange516k ) { |
2244 | 68.2k | PreservedAnalyses PA; |
2245 | 68.2k | PA.preserveSet<CFGAnalyses>(); |
2246 | 68.2k | PA.preserve<GlobalsAA>(); |
2247 | 68.2k | return PA; |
2248 | 68.2k | } |
2249 | 448k | |
2250 | 448k | return PreservedAnalyses::all(); |
2251 | 448k | } |
2252 | | |
2253 | | namespace { |
2254 | | class ReassociateLegacyPass : public FunctionPass { |
2255 | | ReassociatePass Impl; |
2256 | | public: |
2257 | | static char ID; // Pass identification, replacement for typeid |
2258 | 17.3k | ReassociateLegacyPass() : FunctionPass(ID) { |
2259 | 17.3k | initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry()); |
2260 | 17.3k | } |
2261 | | |
2262 | 516k | bool runOnFunction(Function &F) override { |
2263 | 516k | if (skipFunction(F)) |
2264 | 58 | return false; |
2265 | 516k | |
2266 | 516k | FunctionAnalysisManager DummyFAM; |
2267 | 516k | auto PA = Impl.run(F, DummyFAM); |
2268 | 516k | return !PA.areAllPreserved(); |
2269 | 516k | } |
2270 | | |
2271 | 17.3k | void getAnalysisUsage(AnalysisUsage &AU) const override { |
2272 | 17.3k | AU.setPreservesCFG(); |
2273 | 17.3k | AU.addPreserved<GlobalsAAWrapperPass>(); |
2274 | 17.3k | } |
2275 | | }; |
2276 | | } |
2277 | | |
2278 | | char ReassociateLegacyPass::ID = 0; |
2279 | | INITIALIZE_PASS(ReassociateLegacyPass, "reassociate", |
2280 | | "Reassociate expressions", false, false) |
2281 | | |
2282 | | // Public interface to the Reassociate pass |
2283 | 17.2k | FunctionPass *llvm::createReassociatePass() { |
2284 | 17.2k | return new ReassociateLegacyPass(); |
2285 | 17.2k | } |