File: | lib/Transforms/Scalar/Reassociate.cpp |
Warning: | line 1468, column 40 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
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/APFloat.h" | |||
25 | #include "llvm/ADT/APInt.h" | |||
26 | #include "llvm/ADT/DenseMap.h" | |||
27 | #include "llvm/ADT/PostOrderIterator.h" | |||
28 | #include "llvm/ADT/SetVector.h" | |||
29 | #include "llvm/ADT/SmallPtrSet.h" | |||
30 | #include "llvm/ADT/SmallSet.h" | |||
31 | #include "llvm/ADT/SmallVector.h" | |||
32 | #include "llvm/ADT/Statistic.h" | |||
33 | #include "llvm/Analysis/GlobalsModRef.h" | |||
34 | #include "llvm/Analysis/Utils/Local.h" | |||
35 | #include "llvm/Analysis/ValueTracking.h" | |||
36 | #include "llvm/IR/Argument.h" | |||
37 | #include "llvm/IR/BasicBlock.h" | |||
38 | #include "llvm/IR/CFG.h" | |||
39 | #include "llvm/IR/Constant.h" | |||
40 | #include "llvm/IR/Constants.h" | |||
41 | #include "llvm/IR/Function.h" | |||
42 | #include "llvm/IR/IRBuilder.h" | |||
43 | #include "llvm/IR/InstrTypes.h" | |||
44 | #include "llvm/IR/Instruction.h" | |||
45 | #include "llvm/IR/Instructions.h" | |||
46 | #include "llvm/IR/Operator.h" | |||
47 | #include "llvm/IR/PassManager.h" | |||
48 | #include "llvm/IR/PatternMatch.h" | |||
49 | #include "llvm/IR/Type.h" | |||
50 | #include "llvm/IR/User.h" | |||
51 | #include "llvm/IR/Value.h" | |||
52 | #include "llvm/IR/ValueHandle.h" | |||
53 | #include "llvm/Pass.h" | |||
54 | #include "llvm/Support/Casting.h" | |||
55 | #include "llvm/Support/Debug.h" | |||
56 | #include "llvm/Support/ErrorHandling.h" | |||
57 | #include "llvm/Support/raw_ostream.h" | |||
58 | #include "llvm/Transforms/Scalar.h" | |||
59 | #include <algorithm> | |||
60 | #include <cassert> | |||
61 | #include <utility> | |||
62 | ||||
63 | using namespace llvm; | |||
64 | using namespace reassociate; | |||
65 | ||||
66 | #define DEBUG_TYPE"reassociate" "reassociate" | |||
67 | ||||
68 | STATISTIC(NumChanged, "Number of insts reassociated")static llvm::Statistic NumChanged = {"reassociate", "NumChanged" , "Number of insts reassociated", {0}, {false}}; | |||
69 | STATISTIC(NumAnnihil, "Number of expr tree annihilated")static llvm::Statistic NumAnnihil = {"reassociate", "NumAnnihil" , "Number of expr tree annihilated", {0}, {false}}; | |||
70 | STATISTIC(NumFactor , "Number of multiplies factored")static llvm::Statistic NumFactor = {"reassociate", "NumFactor" , "Number of multiplies factored", {0}, {false}}; | |||
71 | ||||
72 | #ifndef NDEBUG | |||
73 | /// Print out the expression identified in the Ops list. | |||
74 | static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) { | |||
75 | Module *M = I->getModule(); | |||
76 | dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " " | |||
77 | << *Ops[0].Op->getType() << '\t'; | |||
78 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { | |||
79 | dbgs() << "[ "; | |||
80 | Ops[i].Op->printAsOperand(dbgs(), false, M); | |||
81 | dbgs() << ", #" << Ops[i].Rank << "] "; | |||
82 | } | |||
83 | } | |||
84 | #endif | |||
85 | ||||
86 | /// Utility class representing a non-constant Xor-operand. We classify | |||
87 | /// non-constant Xor-Operands into two categories: | |||
88 | /// C1) The operand is in the form "X & C", where C is a constant and C != ~0 | |||
89 | /// C2) | |||
90 | /// C2.1) The operand is in the form of "X | C", where C is a non-zero | |||
91 | /// constant. | |||
92 | /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this | |||
93 | /// operand as "E | 0" | |||
94 | class llvm::reassociate::XorOpnd { | |||
95 | public: | |||
96 | XorOpnd(Value *V); | |||
97 | ||||
98 | bool isInvalid() const { return SymbolicPart == nullptr; } | |||
99 | bool isOrExpr() const { return isOr; } | |||
100 | Value *getValue() const { return OrigVal; } | |||
101 | Value *getSymbolicPart() const { return SymbolicPart; } | |||
102 | unsigned getSymbolicRank() const { return SymbolicRank; } | |||
103 | const APInt &getConstPart() const { return ConstPart; } | |||
104 | ||||
105 | void Invalidate() { SymbolicPart = OrigVal = nullptr; } | |||
106 | void setSymbolicRank(unsigned R) { SymbolicRank = R; } | |||
107 | ||||
108 | private: | |||
109 | Value *OrigVal; | |||
110 | Value *SymbolicPart; | |||
111 | APInt ConstPart; | |||
112 | unsigned SymbolicRank; | |||
113 | bool isOr; | |||
114 | }; | |||
115 | ||||
116 | XorOpnd::XorOpnd(Value *V) { | |||
117 | assert(!isa<ConstantInt>(V) && "No ConstantInt")(static_cast <bool> (!isa<ConstantInt>(V) && "No ConstantInt") ? void (0) : __assert_fail ("!isa<ConstantInt>(V) && \"No ConstantInt\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 117, __extension__ __PRETTY_FUNCTION__)); | |||
118 | OrigVal = V; | |||
119 | Instruction *I = dyn_cast<Instruction>(V); | |||
120 | SymbolicRank = 0; | |||
121 | ||||
122 | if (I && (I->getOpcode() == Instruction::Or || | |||
123 | I->getOpcode() == Instruction::And)) { | |||
124 | Value *V0 = I->getOperand(0); | |||
125 | Value *V1 = I->getOperand(1); | |||
126 | const APInt *C; | |||
127 | if (match(V0, PatternMatch::m_APInt(C))) | |||
128 | std::swap(V0, V1); | |||
129 | ||||
130 | if (match(V1, PatternMatch::m_APInt(C))) { | |||
131 | ConstPart = *C; | |||
132 | SymbolicPart = V0; | |||
133 | isOr = (I->getOpcode() == Instruction::Or); | |||
134 | return; | |||
135 | } | |||
136 | } | |||
137 | ||||
138 | // view the operand as "V | 0" | |||
139 | SymbolicPart = V; | |||
140 | ConstPart = APInt::getNullValue(V->getType()->getScalarSizeInBits()); | |||
141 | isOr = true; | |||
142 | } | |||
143 | ||||
144 | /// Return true if V is an instruction of the specified opcode and if it | |||
145 | /// only has one use. | |||
146 | static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { | |||
147 | auto *I = dyn_cast<Instruction>(V); | |||
148 | if (I && I->hasOneUse() && I->getOpcode() == Opcode) | |||
149 | if (!isa<FPMathOperator>(I) || I->isFast()) | |||
150 | return cast<BinaryOperator>(I); | |||
151 | return nullptr; | |||
152 | } | |||
153 | ||||
154 | static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1, | |||
155 | unsigned Opcode2) { | |||
156 | auto *I = dyn_cast<Instruction>(V); | |||
157 | if (I && I->hasOneUse() && | |||
158 | (I->getOpcode() == Opcode1 || I->getOpcode() == Opcode2)) | |||
159 | if (!isa<FPMathOperator>(I) || I->isFast()) | |||
160 | return cast<BinaryOperator>(I); | |||
161 | return nullptr; | |||
162 | } | |||
163 | ||||
164 | void ReassociatePass::BuildRankMap(Function &F, | |||
165 | ReversePostOrderTraversal<Function*> &RPOT) { | |||
166 | unsigned Rank = 2; | |||
167 | ||||
168 | // Assign distinct ranks to function arguments. | |||
169 | for (auto &Arg : F.args()) { | |||
170 | ValueRankMap[&Arg] = ++Rank; | |||
171 | DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rankdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank << "\n"; } } while (false) | |||
172 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank << "\n"; } } while (false); | |||
173 | } | |||
174 | ||||
175 | // Traverse basic blocks in ReversePostOrder | |||
176 | for (BasicBlock *BB : RPOT) { | |||
177 | unsigned BBRank = RankMap[BB] = ++Rank << 16; | |||
178 | ||||
179 | // Walk the basic block, adding precomputed ranks for any instructions that | |||
180 | // we cannot move. This ensures that the ranks for these instructions are | |||
181 | // all different in the block. | |||
182 | for (Instruction &I : *BB) | |||
183 | if (mayBeMemoryDependent(I)) | |||
184 | ValueRankMap[&I] = ++BBRank; | |||
185 | } | |||
186 | } | |||
187 | ||||
188 | unsigned ReassociatePass::getRank(Value *V) { | |||
189 | Instruction *I = dyn_cast<Instruction>(V); | |||
190 | if (!I) { | |||
191 | if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument. | |||
192 | return 0; // Otherwise it's a global or constant, rank 0. | |||
193 | } | |||
194 | ||||
195 | if (unsigned Rank = ValueRankMap[I]) | |||
196 | return Rank; // Rank already known? | |||
197 | ||||
198 | // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that | |||
199 | // we can reassociate expressions for code motion! Since we do not recurse | |||
200 | // for PHI nodes, we cannot have infinite recursion here, because there | |||
201 | // cannot be loops in the value graph that do not go through PHI nodes. | |||
202 | unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; | |||
203 | for (unsigned i = 0, e = I->getNumOperands(); | |||
204 | i != e && Rank != MaxRank; ++i) | |||
205 | Rank = std::max(Rank, getRank(I->getOperand(i))); | |||
206 | ||||
207 | // If this is a not or neg instruction, do not count it for rank. This | |||
208 | // assures us that X and ~X will have the same rank. | |||
209 | if (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) && | |||
210 | !BinaryOperator::isFNeg(I)) | |||
211 | ++Rank; | |||
212 | ||||
213 | DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n"; } } while (false); | |||
214 | ||||
215 | return ValueRankMap[I] = Rank; | |||
216 | } | |||
217 | ||||
218 | // Canonicalize constants to RHS. Otherwise, sort the operands by rank. | |||
219 | void ReassociatePass::canonicalizeOperands(Instruction *I) { | |||
220 | assert(isa<BinaryOperator>(I) && "Expected binary operator.")(static_cast <bool> (isa<BinaryOperator>(I) && "Expected binary operator.") ? void (0) : __assert_fail ("isa<BinaryOperator>(I) && \"Expected binary operator.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 220, __extension__ __PRETTY_FUNCTION__)); | |||
221 | assert(I->isCommutative() && "Expected commutative operator.")(static_cast <bool> (I->isCommutative() && "Expected commutative operator." ) ? void (0) : __assert_fail ("I->isCommutative() && \"Expected commutative operator.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 221, __extension__ __PRETTY_FUNCTION__)); | |||
222 | ||||
223 | Value *LHS = I->getOperand(0); | |||
224 | Value *RHS = I->getOperand(1); | |||
225 | if (LHS == RHS || isa<Constant>(RHS)) | |||
226 | return; | |||
227 | if (isa<Constant>(LHS) || getRank(RHS) < getRank(LHS)) | |||
228 | cast<BinaryOperator>(I)->swapOperands(); | |||
229 | } | |||
230 | ||||
231 | static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name, | |||
232 | Instruction *InsertBefore, Value *FlagsOp) { | |||
233 | if (S1->getType()->isIntOrIntVectorTy()) | |||
234 | return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore); | |||
235 | else { | |||
236 | BinaryOperator *Res = | |||
237 | BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore); | |||
238 | Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); | |||
239 | return Res; | |||
240 | } | |||
241 | } | |||
242 | ||||
243 | static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name, | |||
244 | Instruction *InsertBefore, Value *FlagsOp) { | |||
245 | if (S1->getType()->isIntOrIntVectorTy()) | |||
246 | return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore); | |||
247 | else { | |||
248 | BinaryOperator *Res = | |||
249 | BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore); | |||
250 | Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); | |||
251 | return Res; | |||
252 | } | |||
253 | } | |||
254 | ||||
255 | static BinaryOperator *CreateNeg(Value *S1, const Twine &Name, | |||
256 | Instruction *InsertBefore, Value *FlagsOp) { | |||
257 | if (S1->getType()->isIntOrIntVectorTy()) | |||
258 | return BinaryOperator::CreateNeg(S1, Name, InsertBefore); | |||
259 | else { | |||
260 | BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore); | |||
261 | Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); | |||
262 | return Res; | |||
263 | } | |||
264 | } | |||
265 | ||||
266 | /// Replace 0-X with X*-1. | |||
267 | static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) { | |||
268 | Type *Ty = Neg->getType(); | |||
269 | Constant *NegOne = Ty->isIntOrIntVectorTy() ? | |||
270 | ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0); | |||
271 | ||||
272 | BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg); | |||
273 | Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op. | |||
274 | Res->takeName(Neg); | |||
275 | Neg->replaceAllUsesWith(Res); | |||
276 | Res->setDebugLoc(Neg->getDebugLoc()); | |||
277 | return Res; | |||
278 | } | |||
279 | ||||
280 | /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael | |||
281 | /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for | |||
282 | /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic. | |||
283 | /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every | |||
284 | /// even x in Bitwidth-bit arithmetic. | |||
285 | static unsigned CarmichaelShift(unsigned Bitwidth) { | |||
286 | if (Bitwidth < 3) | |||
287 | return Bitwidth - 1; | |||
288 | return Bitwidth - 2; | |||
289 | } | |||
290 | ||||
291 | /// Add the extra weight 'RHS' to the existing weight 'LHS', | |||
292 | /// reducing the combined weight using any special properties of the operation. | |||
293 | /// The existing weight LHS represents the computation X op X op ... op X where | |||
294 | /// X occurs LHS times. The combined weight represents X op X op ... op X with | |||
295 | /// X occurring LHS + RHS times. If op is "Xor" for example then the combined | |||
296 | /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even; | |||
297 | /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second. | |||
298 | static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) { | |||
299 | // If we were working with infinite precision arithmetic then the combined | |||
300 | // weight would be LHS + RHS. But we are using finite precision arithmetic, | |||
301 | // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct | |||
302 | // for nilpotent operations and addition, but not for idempotent operations | |||
303 | // and multiplication), so it is important to correctly reduce the combined | |||
304 | // weight back into range if wrapping would be wrong. | |||
305 | ||||
306 | // If RHS is zero then the weight didn't change. | |||
307 | if (RHS.isMinValue()) | |||
308 | return; | |||
309 | // If LHS is zero then the combined weight is RHS. | |||
310 | if (LHS.isMinValue()) { | |||
311 | LHS = RHS; | |||
312 | return; | |||
313 | } | |||
314 | // From this point on we know that neither LHS nor RHS is zero. | |||
315 | ||||
316 | if (Instruction::isIdempotent(Opcode)) { | |||
317 | // Idempotent means X op X === X, so any non-zero weight is equivalent to a | |||
318 | // weight of 1. Keeping weights at zero or one also means that wrapping is | |||
319 | // not a problem. | |||
320 | assert(LHS == 1 && RHS == 1 && "Weights not reduced!")(static_cast <bool> (LHS == 1 && RHS == 1 && "Weights not reduced!") ? void (0) : __assert_fail ("LHS == 1 && RHS == 1 && \"Weights not reduced!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 320, __extension__ __PRETTY_FUNCTION__)); | |||
321 | return; // Return a weight of 1. | |||
322 | } | |||
323 | if (Instruction::isNilpotent(Opcode)) { | |||
324 | // Nilpotent means X op X === 0, so reduce weights modulo 2. | |||
325 | assert(LHS == 1 && RHS == 1 && "Weights not reduced!")(static_cast <bool> (LHS == 1 && RHS == 1 && "Weights not reduced!") ? void (0) : __assert_fail ("LHS == 1 && RHS == 1 && \"Weights not reduced!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 325, __extension__ __PRETTY_FUNCTION__)); | |||
326 | LHS = 0; // 1 + 1 === 0 modulo 2. | |||
327 | return; | |||
328 | } | |||
329 | if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) { | |||
330 | // TODO: Reduce the weight by exploiting nsw/nuw? | |||
331 | LHS += RHS; | |||
332 | return; | |||
333 | } | |||
334 | ||||
335 | assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) &&(static_cast <bool> ((Opcode == Instruction::Mul || Opcode == Instruction::FMul) && "Unknown associative operation!" ) ? void (0) : __assert_fail ("(Opcode == Instruction::Mul || Opcode == Instruction::FMul) && \"Unknown associative operation!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 336, __extension__ __PRETTY_FUNCTION__)) | |||
336 | "Unknown associative operation!")(static_cast <bool> ((Opcode == Instruction::Mul || Opcode == Instruction::FMul) && "Unknown associative operation!" ) ? void (0) : __assert_fail ("(Opcode == Instruction::Mul || Opcode == Instruction::FMul) && \"Unknown associative operation!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 336, __extension__ __PRETTY_FUNCTION__)); | |||
337 | unsigned Bitwidth = LHS.getBitWidth(); | |||
338 | // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth | |||
339 | // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth | |||
340 | // bit number x, since either x is odd in which case x^CM = 1, or x is even in | |||
341 | // which case both x^W and x^(W - CM) are zero. By subtracting off multiples | |||
342 | // of CM like this weights can always be reduced to the range [0, CM+Bitwidth) | |||
343 | // which by a happy accident means that they can always be represented using | |||
344 | // Bitwidth bits. | |||
345 | // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than | |||
346 | // the Carmichael number). | |||
347 | if (Bitwidth > 3) { | |||
348 | /// CM - The value of Carmichael's lambda function. | |||
349 | APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth)); | |||
350 | // Any weight W >= Threshold can be replaced with W - CM. | |||
351 | APInt Threshold = CM + Bitwidth; | |||
352 | assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!")(static_cast <bool> (LHS.ult(Threshold) && RHS. ult(Threshold) && "Weights not reduced!") ? void (0) : __assert_fail ("LHS.ult(Threshold) && RHS.ult(Threshold) && \"Weights not reduced!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 352, __extension__ __PRETTY_FUNCTION__)); | |||
353 | // For Bitwidth 4 or more the following sum does not overflow. | |||
354 | LHS += RHS; | |||
355 | while (LHS.uge(Threshold)) | |||
356 | LHS -= CM; | |||
357 | } else { | |||
358 | // To avoid problems with overflow do everything the same as above but using | |||
359 | // a larger type. | |||
360 | unsigned CM = 1U << CarmichaelShift(Bitwidth); | |||
361 | unsigned Threshold = CM + Bitwidth; | |||
362 | assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold &&(static_cast <bool> (LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && "Weights not reduced!" ) ? void (0) : __assert_fail ("LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && \"Weights not reduced!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 363, __extension__ __PRETTY_FUNCTION__)) | |||
363 | "Weights not reduced!")(static_cast <bool> (LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && "Weights not reduced!" ) ? void (0) : __assert_fail ("LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && \"Weights not reduced!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 363, __extension__ __PRETTY_FUNCTION__)); | |||
364 | unsigned Total = LHS.getZExtValue() + RHS.getZExtValue(); | |||
365 | while (Total >= Threshold) | |||
366 | Total -= CM; | |||
367 | LHS = Total; | |||
368 | } | |||
369 | } | |||
370 | ||||
371 | using RepeatedValue = std::pair<Value*, APInt>; | |||
372 | ||||
373 | /// Given an associative binary expression, return the leaf | |||
374 | /// nodes in Ops along with their weights (how many times the leaf occurs). The | |||
375 | /// original expression is the same as | |||
376 | /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times | |||
377 | /// op | |||
378 | /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times | |||
379 | /// op | |||
380 | /// ... | |||
381 | /// op | |||
382 | /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times | |||
383 | /// | |||
384 | /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct. | |||
385 | /// | |||
386 | /// This routine may modify the function, in which case it returns 'true'. The | |||
387 | /// changes it makes may well be destructive, changing the value computed by 'I' | |||
388 | /// to something completely different. Thus if the routine returns 'true' then | |||
389 | /// you MUST either replace I with a new expression computed from the Ops array, | |||
390 | /// or use RewriteExprTree to put the values back in. | |||
391 | /// | |||
392 | /// A leaf node is either not a binary operation of the same kind as the root | |||
393 | /// node 'I' (i.e. is not a binary operator at all, or is, but with a different | |||
394 | /// opcode), or is the same kind of binary operator but has a use which either | |||
395 | /// does not belong to the expression, or does belong to the expression but is | |||
396 | /// a leaf node. Every leaf node has at least one use that is a non-leaf node | |||
397 | /// of the expression, while for non-leaf nodes (except for the root 'I') every | |||
398 | /// use is a non-leaf node of the expression. | |||
399 | /// | |||
400 | /// For example: | |||
401 | /// expression graph node names | |||
402 | /// | |||
403 | /// + | I | |||
404 | /// / \ | | |||
405 | /// + + | A, B | |||
406 | /// / \ / \ | | |||
407 | /// * + * | C, D, E | |||
408 | /// / \ / \ / \ | | |||
409 | /// + * | F, G | |||
410 | /// | |||
411 | /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in | |||
412 | /// that order) (C, 1), (E, 1), (F, 2), (G, 2). | |||
413 | /// | |||
414 | /// The expression is maximal: if some instruction is a binary operator of the | |||
415 | /// same kind as 'I', and all of its uses are non-leaf nodes of the expression, | |||
416 | /// then the instruction also belongs to the expression, is not a leaf node of | |||
417 | /// it, and its operands also belong to the expression (but may be leaf nodes). | |||
418 | /// | |||
419 | /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in | |||
420 | /// order to ensure that every non-root node in the expression has *exactly one* | |||
421 | /// use by a non-leaf node of the expression. This destruction means that the | |||
422 | /// caller MUST either replace 'I' with a new expression or use something like | |||
423 | /// RewriteExprTree to put the values back in if the routine indicates that it | |||
424 | /// made a change by returning 'true'. | |||
425 | /// | |||
426 | /// In the above example either the right operand of A or the left operand of B | |||
427 | /// will be replaced by undef. If it is B's operand then this gives: | |||
428 | /// | |||
429 | /// + | I | |||
430 | /// / \ | | |||
431 | /// + + | A, B - operand of B replaced with undef | |||
432 | /// / \ \ | | |||
433 | /// * + * | C, D, E | |||
434 | /// / \ / \ / \ | | |||
435 | /// + * | F, G | |||
436 | /// | |||
437 | /// Note that such undef operands can only be reached by passing through 'I'. | |||
438 | /// For example, if you visit operands recursively starting from a leaf node | |||
439 | /// then you will never see such an undef operand unless you get back to 'I', | |||
440 | /// which requires passing through a phi node. | |||
441 | /// | |||
442 | /// Note that this routine may also mutate binary operators of the wrong type | |||
443 | /// that have all uses inside the expression (i.e. only used by non-leaf nodes | |||
444 | /// of the expression) if it can turn them into binary operators of the right | |||
445 | /// type and thus make the expression bigger. | |||
446 | static bool LinearizeExprTree(BinaryOperator *I, | |||
447 | SmallVectorImpl<RepeatedValue> &Ops) { | |||
448 | DEBUG(dbgs() << "LINEARIZE: " << *I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "LINEARIZE: " << *I << '\n'; } } while (false); | |||
449 | unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits(); | |||
450 | unsigned Opcode = I->getOpcode(); | |||
451 | assert(I->isAssociative() && I->isCommutative() &&(static_cast <bool> (I->isAssociative() && I ->isCommutative() && "Expected an associative and commutative operation!" ) ? void (0) : __assert_fail ("I->isAssociative() && I->isCommutative() && \"Expected an associative and commutative operation!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 452, __extension__ __PRETTY_FUNCTION__)) | |||
452 | "Expected an associative and commutative operation!")(static_cast <bool> (I->isAssociative() && I ->isCommutative() && "Expected an associative and commutative operation!" ) ? void (0) : __assert_fail ("I->isAssociative() && I->isCommutative() && \"Expected an associative and commutative operation!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 452, __extension__ __PRETTY_FUNCTION__)); | |||
453 | ||||
454 | // Visit all operands of the expression, keeping track of their weight (the | |||
455 | // number of paths from the expression root to the operand, or if you like | |||
456 | // the number of times that operand occurs in the linearized expression). | |||
457 | // For example, if I = X + A, where X = A + B, then I, X and B have weight 1 | |||
458 | // while A has weight two. | |||
459 | ||||
460 | // Worklist of non-leaf nodes (their operands are in the expression too) along | |||
461 | // with their weights, representing a certain number of paths to the operator. | |||
462 | // If an operator occurs in the worklist multiple times then we found multiple | |||
463 | // ways to get to it. | |||
464 | SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight) | |||
465 | Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1))); | |||
466 | bool Changed = false; | |||
467 | ||||
468 | // Leaves of the expression are values that either aren't the right kind of | |||
469 | // operation (eg: a constant, or a multiply in an add tree), or are, but have | |||
470 | // some uses that are not inside the expression. For example, in I = X + X, | |||
471 | // X = A + B, the value X has two uses (by I) that are in the expression. If | |||
472 | // X has any other uses, for example in a return instruction, then we consider | |||
473 | // X to be a leaf, and won't analyze it further. When we first visit a value, | |||
474 | // if it has more than one use then at first we conservatively consider it to | |||
475 | // be a leaf. Later, as the expression is explored, we may discover some more | |||
476 | // uses of the value from inside the expression. If all uses turn out to be | |||
477 | // from within the expression (and the value is a binary operator of the right | |||
478 | // kind) then the value is no longer considered to be a leaf, and its operands | |||
479 | // are explored. | |||
480 | ||||
481 | // Leaves - Keeps track of the set of putative leaves as well as the number of | |||
482 | // paths to each leaf seen so far. | |||
483 | using LeafMap = DenseMap<Value *, APInt>; | |||
484 | LeafMap Leaves; // Leaf -> Total weight so far. | |||
485 | SmallVector<Value *, 8> LeafOrder; // Ensure deterministic leaf output order. | |||
486 | ||||
487 | #ifndef NDEBUG | |||
488 | SmallPtrSet<Value *, 8> Visited; // For sanity checking the iteration scheme. | |||
489 | #endif | |||
490 | while (!Worklist.empty()) { | |||
491 | std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val(); | |||
492 | I = P.first; // We examine the operands of this binary operator. | |||
493 | ||||
494 | for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands. | |||
495 | Value *Op = I->getOperand(OpIdx); | |||
496 | APInt Weight = P.second; // Number of paths to this operand. | |||
497 | DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n"; } } while (false); | |||
498 | assert(!Op->use_empty() && "No uses, so how did we get to it?!")(static_cast <bool> (!Op->use_empty() && "No uses, so how did we get to it?!" ) ? void (0) : __assert_fail ("!Op->use_empty() && \"No uses, so how did we get to it?!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 498, __extension__ __PRETTY_FUNCTION__)); | |||
499 | ||||
500 | // If this is a binary operation of the right kind with only one use then | |||
501 | // add its operands to the expression. | |||
502 | if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { | |||
503 | assert(Visited.insert(Op).second && "Not first visit!")(static_cast <bool> (Visited.insert(Op).second && "Not first visit!") ? void (0) : __assert_fail ("Visited.insert(Op).second && \"Not first visit!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 503, __extension__ __PRETTY_FUNCTION__)); | |||
504 | DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n"; } } while (false ); | |||
505 | Worklist.push_back(std::make_pair(BO, Weight)); | |||
506 | continue; | |||
507 | } | |||
508 | ||||
509 | // Appears to be a leaf. Is the operand already in the set of leaves? | |||
510 | LeafMap::iterator It = Leaves.find(Op); | |||
511 | if (It == Leaves.end()) { | |||
512 | // Not in the leaf map. Must be the first time we saw this operand. | |||
513 | assert(Visited.insert(Op).second && "Not first visit!")(static_cast <bool> (Visited.insert(Op).second && "Not first visit!") ? void (0) : __assert_fail ("Visited.insert(Op).second && \"Not first visit!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 513, __extension__ __PRETTY_FUNCTION__)); | |||
514 | if (!Op->hasOneUse()) { | |||
515 | // This value has uses not accounted for by the expression, so it is | |||
516 | // not safe to modify. Mark it as being a leaf. | |||
517 | DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n"; } } while ( false); | |||
518 | LeafOrder.push_back(Op); | |||
519 | Leaves[Op] = Weight; | |||
520 | continue; | |||
521 | } | |||
522 | // No uses outside the expression, try morphing it. | |||
523 | } else { | |||
524 | // Already in the leaf map. | |||
525 | assert(It != Leaves.end() && Visited.count(Op) &&(static_cast <bool> (It != Leaves.end() && Visited .count(Op) && "In leaf map but not visited!") ? void ( 0) : __assert_fail ("It != Leaves.end() && Visited.count(Op) && \"In leaf map but not visited!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 526, __extension__ __PRETTY_FUNCTION__)) | |||
526 | "In leaf map but not visited!")(static_cast <bool> (It != Leaves.end() && Visited .count(Op) && "In leaf map but not visited!") ? void ( 0) : __assert_fail ("It != Leaves.end() && Visited.count(Op) && \"In leaf map but not visited!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 526, __extension__ __PRETTY_FUNCTION__)); | |||
527 | ||||
528 | // Update the number of paths to the leaf. | |||
529 | IncorporateWeight(It->second, Weight, Opcode); | |||
530 | ||||
531 | #if 0 // TODO: Re-enable once PR13021 is fixed. | |||
532 | // The leaf already has one use from inside the expression. As we want | |||
533 | // exactly one such use, drop this new use of the leaf. | |||
534 | assert(!Op->hasOneUse() && "Only one use, but we got here twice!")(static_cast <bool> (!Op->hasOneUse() && "Only one use, but we got here twice!" ) ? void (0) : __assert_fail ("!Op->hasOneUse() && \"Only one use, but we got here twice!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 534, __extension__ __PRETTY_FUNCTION__)); | |||
535 | I->setOperand(OpIdx, UndefValue::get(I->getType())); | |||
536 | Changed = true; | |||
537 | ||||
538 | // If the leaf is a binary operation of the right kind and we now see | |||
539 | // that its multiple original uses were in fact all by nodes belonging | |||
540 | // to the expression, then no longer consider it to be a leaf and add | |||
541 | // its operands to the expression. | |||
542 | if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { | |||
543 | DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n"; } } while (false ); | |||
544 | Worklist.push_back(std::make_pair(BO, It->second)); | |||
545 | Leaves.erase(It); | |||
546 | continue; | |||
547 | } | |||
548 | #endif | |||
549 | ||||
550 | // If we still have uses that are not accounted for by the expression | |||
551 | // then it is not safe to modify the value. | |||
552 | if (!Op->hasOneUse()) | |||
553 | continue; | |||
554 | ||||
555 | // No uses outside the expression, try morphing it. | |||
556 | Weight = It->second; | |||
557 | Leaves.erase(It); // Since the value may be morphed below. | |||
558 | } | |||
559 | ||||
560 | // At this point we have a value which, first of all, is not a binary | |||
561 | // expression of the right kind, and secondly, is only used inside the | |||
562 | // expression. This means that it can safely be modified. See if we | |||
563 | // can usefully morph it into an expression of the right kind. | |||
564 | assert((!isa<Instruction>(Op) ||(static_cast <bool> ((!isa<Instruction>(Op) || cast <Instruction>(Op)->getOpcode() != Opcode || (isa< FPMathOperator>(Op) && !cast<Instruction>(Op )->isFast())) && "Should have been handled above!" ) ? void (0) : __assert_fail ("(!isa<Instruction>(Op) || cast<Instruction>(Op)->getOpcode() != Opcode || (isa<FPMathOperator>(Op) && !cast<Instruction>(Op)->isFast())) && \"Should have been handled above!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 568, __extension__ __PRETTY_FUNCTION__)) | |||
565 | cast<Instruction>(Op)->getOpcode() != Opcode(static_cast <bool> ((!isa<Instruction>(Op) || cast <Instruction>(Op)->getOpcode() != Opcode || (isa< FPMathOperator>(Op) && !cast<Instruction>(Op )->isFast())) && "Should have been handled above!" ) ? void (0) : __assert_fail ("(!isa<Instruction>(Op) || cast<Instruction>(Op)->getOpcode() != Opcode || (isa<FPMathOperator>(Op) && !cast<Instruction>(Op)->isFast())) && \"Should have been handled above!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 568, __extension__ __PRETTY_FUNCTION__)) | |||
566 | || (isa<FPMathOperator>(Op) &&(static_cast <bool> ((!isa<Instruction>(Op) || cast <Instruction>(Op)->getOpcode() != Opcode || (isa< FPMathOperator>(Op) && !cast<Instruction>(Op )->isFast())) && "Should have been handled above!" ) ? void (0) : __assert_fail ("(!isa<Instruction>(Op) || cast<Instruction>(Op)->getOpcode() != Opcode || (isa<FPMathOperator>(Op) && !cast<Instruction>(Op)->isFast())) && \"Should have been handled above!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 568, __extension__ __PRETTY_FUNCTION__)) | |||
567 | !cast<Instruction>(Op)->isFast())) &&(static_cast <bool> ((!isa<Instruction>(Op) || cast <Instruction>(Op)->getOpcode() != Opcode || (isa< FPMathOperator>(Op) && !cast<Instruction>(Op )->isFast())) && "Should have been handled above!" ) ? void (0) : __assert_fail ("(!isa<Instruction>(Op) || cast<Instruction>(Op)->getOpcode() != Opcode || (isa<FPMathOperator>(Op) && !cast<Instruction>(Op)->isFast())) && \"Should have been handled above!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 568, __extension__ __PRETTY_FUNCTION__)) | |||
568 | "Should have been handled above!")(static_cast <bool> ((!isa<Instruction>(Op) || cast <Instruction>(Op)->getOpcode() != Opcode || (isa< FPMathOperator>(Op) && !cast<Instruction>(Op )->isFast())) && "Should have been handled above!" ) ? void (0) : __assert_fail ("(!isa<Instruction>(Op) || cast<Instruction>(Op)->getOpcode() != Opcode || (isa<FPMathOperator>(Op) && !cast<Instruction>(Op)->isFast())) && \"Should have been handled above!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 568, __extension__ __PRETTY_FUNCTION__)); | |||
569 | assert(Op->hasOneUse() && "Has uses outside the expression tree!")(static_cast <bool> (Op->hasOneUse() && "Has uses outside the expression tree!" ) ? void (0) : __assert_fail ("Op->hasOneUse() && \"Has uses outside the expression tree!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 569, __extension__ __PRETTY_FUNCTION__)); | |||
570 | ||||
571 | // If this is a multiply expression, turn any internal negations into | |||
572 | // multiplies by -1 so they can be reassociated. | |||
573 | if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) | |||
574 | if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) || | |||
575 | (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) { | |||
576 | DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO "; } } while (false ); | |||
577 | BO = LowerNegateToMultiply(BO); | |||
578 | DEBUG(dbgs() << *BO << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << *BO << '\n'; } } while (false); | |||
579 | Worklist.push_back(std::make_pair(BO, Weight)); | |||
580 | Changed = true; | |||
581 | continue; | |||
582 | } | |||
583 | ||||
584 | // Failed to morph into an expression of the right type. This really is | |||
585 | // a leaf. | |||
586 | DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n"; } } while (false); | |||
587 | assert(!isReassociableOp(Op, Opcode) && "Value was morphed?")(static_cast <bool> (!isReassociableOp(Op, Opcode) && "Value was morphed?") ? void (0) : __assert_fail ("!isReassociableOp(Op, Opcode) && \"Value was morphed?\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 587, __extension__ __PRETTY_FUNCTION__)); | |||
588 | LeafOrder.push_back(Op); | |||
589 | Leaves[Op] = Weight; | |||
590 | } | |||
591 | } | |||
592 | ||||
593 | // The leaves, repeated according to their weights, represent the linearized | |||
594 | // form of the expression. | |||
595 | for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) { | |||
596 | Value *V = LeafOrder[i]; | |||
597 | LeafMap::iterator It = Leaves.find(V); | |||
598 | if (It == Leaves.end()) | |||
599 | // Node initially thought to be a leaf wasn't. | |||
600 | continue; | |||
601 | assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!")(static_cast <bool> (!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!") ? void (0) : __assert_fail ("!isReassociableOp(V, Opcode) && \"Shouldn't be a leaf!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 601, __extension__ __PRETTY_FUNCTION__)); | |||
602 | APInt Weight = It->second; | |||
603 | if (Weight.isMinValue()) | |||
604 | // Leaf already output or weight reduction eliminated it. | |||
605 | continue; | |||
606 | // Ensure the leaf is only output once. | |||
607 | It->second = 0; | |||
608 | Ops.push_back(std::make_pair(V, Weight)); | |||
609 | } | |||
610 | ||||
611 | // For nilpotent operations or addition there may be no operands, for example | |||
612 | // because the expression was "X xor X" or consisted of 2^Bitwidth additions: | |||
613 | // in both cases the weight reduces to 0 causing the value to be skipped. | |||
614 | if (Ops.empty()) { | |||
615 | Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType()); | |||
616 | assert(Identity && "Associative operation without identity!")(static_cast <bool> (Identity && "Associative operation without identity!" ) ? void (0) : __assert_fail ("Identity && \"Associative operation without identity!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 616, __extension__ __PRETTY_FUNCTION__)); | |||
617 | Ops.emplace_back(Identity, APInt(Bitwidth, 1)); | |||
618 | } | |||
619 | ||||
620 | return Changed; | |||
621 | } | |||
622 | ||||
623 | /// Now that the operands for this expression tree are | |||
624 | /// linearized and optimized, emit them in-order. | |||
625 | void ReassociatePass::RewriteExprTree(BinaryOperator *I, | |||
626 | SmallVectorImpl<ValueEntry> &Ops) { | |||
627 | assert(Ops.size() > 1 && "Single values should be used directly!")(static_cast <bool> (Ops.size() > 1 && "Single values should be used directly!" ) ? void (0) : __assert_fail ("Ops.size() > 1 && \"Single values should be used directly!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 627, __extension__ __PRETTY_FUNCTION__)); | |||
628 | ||||
629 | // Since our optimizations should never increase the number of operations, the | |||
630 | // new expression can usually be written reusing the existing binary operators | |||
631 | // from the original expression tree, without creating any new instructions, | |||
632 | // though the rewritten expression may have a completely different topology. | |||
633 | // We take care to not change anything if the new expression will be the same | |||
634 | // as the original. If more than trivial changes (like commuting operands) | |||
635 | // were made then we are obliged to clear out any optional subclass data like | |||
636 | // nsw flags. | |||
637 | ||||
638 | /// NodesToRewrite - Nodes from the original expression available for writing | |||
639 | /// the new expression into. | |||
640 | SmallVector<BinaryOperator*, 8> NodesToRewrite; | |||
641 | unsigned Opcode = I->getOpcode(); | |||
642 | BinaryOperator *Op = I; | |||
643 | ||||
644 | /// NotRewritable - The operands being written will be the leaves of the new | |||
645 | /// expression and must not be used as inner nodes (via NodesToRewrite) by | |||
646 | /// mistake. Inner nodes are always reassociable, and usually leaves are not | |||
647 | /// (if they were they would have been incorporated into the expression and so | |||
648 | /// would not be leaves), so most of the time there is no danger of this. But | |||
649 | /// in rare cases a leaf may become reassociable if an optimization kills uses | |||
650 | /// of it, or it may momentarily become reassociable during rewriting (below) | |||
651 | /// due it being removed as an operand of one of its uses. Ensure that misuse | |||
652 | /// of leaf nodes as inner nodes cannot occur by remembering all of the future | |||
653 | /// leaves and refusing to reuse any of them as inner nodes. | |||
654 | SmallPtrSet<Value*, 8> NotRewritable; | |||
655 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) | |||
656 | NotRewritable.insert(Ops[i].Op); | |||
657 | ||||
658 | // ExpressionChanged - Non-null if the rewritten expression differs from the | |||
659 | // original in some non-trivial way, requiring the clearing of optional flags. | |||
660 | // Flags are cleared from the operator in ExpressionChanged up to I inclusive. | |||
661 | BinaryOperator *ExpressionChanged = nullptr; | |||
662 | for (unsigned i = 0; ; ++i) { | |||
663 | // The last operation (which comes earliest in the IR) is special as both | |||
664 | // operands will come from Ops, rather than just one with the other being | |||
665 | // a subexpression. | |||
666 | if (i+2 == Ops.size()) { | |||
667 | Value *NewLHS = Ops[i].Op; | |||
668 | Value *NewRHS = Ops[i+1].Op; | |||
669 | Value *OldLHS = Op->getOperand(0); | |||
670 | Value *OldRHS = Op->getOperand(1); | |||
671 | ||||
672 | if (NewLHS == OldLHS && NewRHS == OldRHS) | |||
673 | // Nothing changed, leave it alone. | |||
674 | break; | |||
675 | ||||
676 | if (NewLHS == OldRHS && NewRHS == OldLHS) { | |||
677 | // The order of the operands was reversed. Swap them. | |||
678 | DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "RA: " << *Op << '\n'; } } while (false); | |||
679 | Op->swapOperands(); | |||
680 | DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "TO: " << *Op << '\n'; } } while (false); | |||
681 | MadeChange = true; | |||
682 | ++NumChanged; | |||
683 | break; | |||
684 | } | |||
685 | ||||
686 | // The new operation differs non-trivially from the original. Overwrite | |||
687 | // the old operands with the new ones. | |||
688 | DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "RA: " << *Op << '\n'; } } while (false); | |||
689 | if (NewLHS != OldLHS) { | |||
690 | BinaryOperator *BO = isReassociableOp(OldLHS, Opcode); | |||
691 | if (BO && !NotRewritable.count(BO)) | |||
692 | NodesToRewrite.push_back(BO); | |||
693 | Op->setOperand(0, NewLHS); | |||
694 | } | |||
695 | if (NewRHS != OldRHS) { | |||
696 | BinaryOperator *BO = isReassociableOp(OldRHS, Opcode); | |||
697 | if (BO && !NotRewritable.count(BO)) | |||
698 | NodesToRewrite.push_back(BO); | |||
699 | Op->setOperand(1, NewRHS); | |||
700 | } | |||
701 | DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "TO: " << *Op << '\n'; } } while (false); | |||
702 | ||||
703 | ExpressionChanged = Op; | |||
704 | MadeChange = true; | |||
705 | ++NumChanged; | |||
706 | ||||
707 | break; | |||
708 | } | |||
709 | ||||
710 | // Not the last operation. The left-hand side will be a sub-expression | |||
711 | // while the right-hand side will be the current element of Ops. | |||
712 | Value *NewRHS = Ops[i].Op; | |||
713 | if (NewRHS != Op->getOperand(1)) { | |||
714 | DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "RA: " << *Op << '\n'; } } while (false); | |||
715 | if (NewRHS == Op->getOperand(0)) { | |||
716 | // The new right-hand side was already present as the left operand. If | |||
717 | // we are lucky then swapping the operands will sort out both of them. | |||
718 | Op->swapOperands(); | |||
719 | } else { | |||
720 | // Overwrite with the new right-hand side. | |||
721 | BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode); | |||
722 | if (BO && !NotRewritable.count(BO)) | |||
723 | NodesToRewrite.push_back(BO); | |||
724 | Op->setOperand(1, NewRHS); | |||
725 | ExpressionChanged = Op; | |||
726 | } | |||
727 | DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "TO: " << *Op << '\n'; } } while (false); | |||
728 | MadeChange = true; | |||
729 | ++NumChanged; | |||
730 | } | |||
731 | ||||
732 | // Now deal with the left-hand side. If this is already an operation node | |||
733 | // from the original expression then just rewrite the rest of the expression | |||
734 | // into it. | |||
735 | BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode); | |||
736 | if (BO && !NotRewritable.count(BO)) { | |||
737 | Op = BO; | |||
738 | continue; | |||
739 | } | |||
740 | ||||
741 | // Otherwise, grab a spare node from the original expression and use that as | |||
742 | // the left-hand side. If there are no nodes left then the optimizers made | |||
743 | // an expression with more nodes than the original! This usually means that | |||
744 | // they did something stupid but it might mean that the problem was just too | |||
745 | // hard (finding the mimimal number of multiplications needed to realize a | |||
746 | // multiplication expression is NP-complete). Whatever the reason, smart or | |||
747 | // stupid, create a new node if there are none left. | |||
748 | BinaryOperator *NewOp; | |||
749 | if (NodesToRewrite.empty()) { | |||
750 | Constant *Undef = UndefValue::get(I->getType()); | |||
751 | NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode), | |||
752 | Undef, Undef, "", I); | |||
753 | if (NewOp->getType()->isFPOrFPVectorTy()) | |||
754 | NewOp->setFastMathFlags(I->getFastMathFlags()); | |||
755 | } else { | |||
756 | NewOp = NodesToRewrite.pop_back_val(); | |||
757 | } | |||
758 | ||||
759 | DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "RA: " << *Op << '\n'; } } while (false); | |||
760 | Op->setOperand(0, NewOp); | |||
761 | DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "TO: " << *Op << '\n'; } } while (false); | |||
762 | ExpressionChanged = Op; | |||
763 | MadeChange = true; | |||
764 | ++NumChanged; | |||
765 | Op = NewOp; | |||
766 | } | |||
767 | ||||
768 | // If the expression changed non-trivially then clear out all subclass data | |||
769 | // starting from the operator specified in ExpressionChanged, and compactify | |||
770 | // the operators to just before the expression root to guarantee that the | |||
771 | // expression tree is dominated by all of Ops. | |||
772 | if (ExpressionChanged) | |||
773 | do { | |||
774 | // Preserve FastMathFlags. | |||
775 | if (isa<FPMathOperator>(I)) { | |||
776 | FastMathFlags Flags = I->getFastMathFlags(); | |||
777 | ExpressionChanged->clearSubclassOptionalData(); | |||
778 | ExpressionChanged->setFastMathFlags(Flags); | |||
779 | } else | |||
780 | ExpressionChanged->clearSubclassOptionalData(); | |||
781 | ||||
782 | if (ExpressionChanged == I) | |||
783 | break; | |||
784 | ExpressionChanged->moveBefore(I); | |||
785 | ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin()); | |||
786 | } while (true); | |||
787 | ||||
788 | // Throw away any left over nodes from the original expression. | |||
789 | for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i) | |||
790 | RedoInsts.insert(NodesToRewrite[i]); | |||
791 | } | |||
792 | ||||
793 | /// Insert instructions before the instruction pointed to by BI, | |||
794 | /// that computes the negative version of the value specified. The negative | |||
795 | /// version of the value is returned, and BI is left pointing at the instruction | |||
796 | /// that should be processed next by the reassociation pass. | |||
797 | /// Also add intermediate instructions to the redo list that are modified while | |||
798 | /// pushing the negates through adds. These will be revisited to see if | |||
799 | /// additional opportunities have been exposed. | |||
800 | static Value *NegateValue(Value *V, Instruction *BI, | |||
801 | SetVector<AssertingVH<Instruction>> &ToRedo) { | |||
802 | if (auto *C = dyn_cast<Constant>(V)) | |||
803 | return C->getType()->isFPOrFPVectorTy() ? ConstantExpr::getFNeg(C) : | |||
804 | ConstantExpr::getNeg(C); | |||
805 | ||||
806 | // We are trying to expose opportunity for reassociation. One of the things | |||
807 | // that we want to do to achieve this is to push a negation as deep into an | |||
808 | // expression chain as possible, to expose the add instructions. In practice, | |||
809 | // this means that we turn this: | |||
810 | // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D | |||
811 | // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate | |||
812 | // the constants. We assume that instcombine will clean up the mess later if | |||
813 | // we introduce tons of unnecessary negation instructions. | |||
814 | // | |||
815 | if (BinaryOperator *I = | |||
816 | isReassociableOp(V, Instruction::Add, Instruction::FAdd)) { | |||
817 | // Push the negates through the add. | |||
818 | I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo)); | |||
819 | I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo)); | |||
820 | if (I->getOpcode() == Instruction::Add) { | |||
821 | I->setHasNoUnsignedWrap(false); | |||
822 | I->setHasNoSignedWrap(false); | |||
823 | } | |||
824 | ||||
825 | // We must move the add instruction here, because the neg instructions do | |||
826 | // not dominate the old add instruction in general. By moving it, we are | |||
827 | // assured that the neg instructions we just inserted dominate the | |||
828 | // instruction we are about to insert after them. | |||
829 | // | |||
830 | I->moveBefore(BI); | |||
831 | I->setName(I->getName()+".neg"); | |||
832 | ||||
833 | // Add the intermediate negates to the redo list as processing them later | |||
834 | // could expose more reassociating opportunities. | |||
835 | ToRedo.insert(I); | |||
836 | return I; | |||
837 | } | |||
838 | ||||
839 | // Okay, we need to materialize a negated version of V with an instruction. | |||
840 | // Scan the use lists of V to see if we have one already. | |||
841 | for (User *U : V->users()) { | |||
842 | if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U)) | |||
843 | continue; | |||
844 | ||||
845 | // We found one! Now we have to make sure that the definition dominates | |||
846 | // this use. We do this by moving it to the entry block (if it is a | |||
847 | // non-instruction value) or right after the definition. These negates will | |||
848 | // be zapped by reassociate later, so we don't need much finesse here. | |||
849 | BinaryOperator *TheNeg = cast<BinaryOperator>(U); | |||
850 | ||||
851 | // Verify that the negate is in this function, V might be a constant expr. | |||
852 | if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) | |||
853 | continue; | |||
854 | ||||
855 | BasicBlock::iterator InsertPt; | |||
856 | if (Instruction *InstInput = dyn_cast<Instruction>(V)) { | |||
857 | if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { | |||
858 | InsertPt = II->getNormalDest()->begin(); | |||
859 | } else { | |||
860 | InsertPt = ++InstInput->getIterator(); | |||
861 | } | |||
862 | while (isa<PHINode>(InsertPt)) ++InsertPt; | |||
863 | } else { | |||
864 | InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); | |||
865 | } | |||
866 | TheNeg->moveBefore(&*InsertPt); | |||
867 | if (TheNeg->getOpcode() == Instruction::Sub) { | |||
868 | TheNeg->setHasNoUnsignedWrap(false); | |||
869 | TheNeg->setHasNoSignedWrap(false); | |||
870 | } else { | |||
871 | TheNeg->andIRFlags(BI); | |||
872 | } | |||
873 | ToRedo.insert(TheNeg); | |||
874 | return TheNeg; | |||
875 | } | |||
876 | ||||
877 | // Insert a 'neg' instruction that subtracts the value from zero to get the | |||
878 | // negation. | |||
879 | BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI); | |||
880 | ToRedo.insert(NewNeg); | |||
881 | return NewNeg; | |||
882 | } | |||
883 | ||||
884 | /// Return true if we should break up this subtract of X-Y into (X + -Y). | |||
885 | static bool ShouldBreakUpSubtract(Instruction *Sub) { | |||
886 | // If this is a negation, we can't split it up! | |||
887 | if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub)) | |||
888 | return false; | |||
889 | ||||
890 | // Don't breakup X - undef. | |||
891 | if (isa<UndefValue>(Sub->getOperand(1))) | |||
892 | return false; | |||
893 | ||||
894 | // Don't bother to break this up unless either the LHS is an associable add or | |||
895 | // subtract or if this is only used by one. | |||
896 | Value *V0 = Sub->getOperand(0); | |||
897 | if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) || | |||
898 | isReassociableOp(V0, Instruction::Sub, Instruction::FSub)) | |||
899 | return true; | |||
900 | Value *V1 = Sub->getOperand(1); | |||
901 | if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) || | |||
902 | isReassociableOp(V1, Instruction::Sub, Instruction::FSub)) | |||
903 | return true; | |||
904 | Value *VB = Sub->user_back(); | |||
905 | if (Sub->hasOneUse() && | |||
906 | (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) || | |||
907 | isReassociableOp(VB, Instruction::Sub, Instruction::FSub))) | |||
908 | return true; | |||
909 | ||||
910 | return false; | |||
911 | } | |||
912 | ||||
913 | /// If we have (X-Y), and if either X is an add, or if this is only used by an | |||
914 | /// add, transform this into (X+(0-Y)) to promote better reassociation. | |||
915 | static BinaryOperator * | |||
916 | BreakUpSubtract(Instruction *Sub, SetVector<AssertingVH<Instruction>> &ToRedo) { | |||
917 | // Convert a subtract into an add and a neg instruction. This allows sub | |||
918 | // instructions to be commuted with other add instructions. | |||
919 | // | |||
920 | // Calculate the negative value of Operand 1 of the sub instruction, | |||
921 | // and set it as the RHS of the add instruction we just made. | |||
922 | Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo); | |||
923 | BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub); | |||
924 | Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op. | |||
925 | Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op. | |||
926 | New->takeName(Sub); | |||
927 | ||||
928 | // Everyone now refers to the add instruction. | |||
929 | Sub->replaceAllUsesWith(New); | |||
930 | New->setDebugLoc(Sub->getDebugLoc()); | |||
931 | ||||
932 | DEBUG(dbgs() << "Negated: " << *New << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Negated: " << *New << '\n'; } } while (false); | |||
933 | return New; | |||
934 | } | |||
935 | ||||
936 | /// If this is a shift of a reassociable multiply or is used by one, change | |||
937 | /// this into a multiply by a constant to assist with further reassociation. | |||
938 | static BinaryOperator *ConvertShiftToMul(Instruction *Shl) { | |||
939 | Constant *MulCst = ConstantInt::get(Shl->getType(), 1); | |||
940 | MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); | |||
941 | ||||
942 | BinaryOperator *Mul = | |||
943 | BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); | |||
944 | Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op. | |||
945 | Mul->takeName(Shl); | |||
946 | ||||
947 | // Everyone now refers to the mul instruction. | |||
948 | Shl->replaceAllUsesWith(Mul); | |||
949 | Mul->setDebugLoc(Shl->getDebugLoc()); | |||
950 | ||||
951 | // We can safely preserve the nuw flag in all cases. It's also safe to turn a | |||
952 | // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special | |||
953 | // handling. | |||
954 | bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap(); | |||
955 | bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap(); | |||
956 | if (NSW && NUW) | |||
957 | Mul->setHasNoSignedWrap(true); | |||
958 | Mul->setHasNoUnsignedWrap(NUW); | |||
959 | return Mul; | |||
960 | } | |||
961 | ||||
962 | /// Scan backwards and forwards among values with the same rank as element i | |||
963 | /// to see if X exists. If X does not exist, return i. This is useful when | |||
964 | /// scanning for 'x' when we see '-x' because they both get the same rank. | |||
965 | static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops, | |||
966 | unsigned i, Value *X) { | |||
967 | unsigned XRank = Ops[i].Rank; | |||
968 | unsigned e = Ops.size(); | |||
969 | for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) { | |||
970 | if (Ops[j].Op == X) | |||
971 | return j; | |||
972 | if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op)) | |||
973 | if (Instruction *I2 = dyn_cast<Instruction>(X)) | |||
974 | if (I1->isIdenticalTo(I2)) | |||
975 | return j; | |||
976 | } | |||
977 | // Scan backwards. | |||
978 | for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) { | |||
979 | if (Ops[j].Op == X) | |||
980 | return j; | |||
981 | if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op)) | |||
982 | if (Instruction *I2 = dyn_cast<Instruction>(X)) | |||
983 | if (I1->isIdenticalTo(I2)) | |||
984 | return j; | |||
985 | } | |||
986 | return i; | |||
987 | } | |||
988 | ||||
989 | /// Emit a tree of add instructions, summing Ops together | |||
990 | /// and returning the result. Insert the tree before I. | |||
991 | static Value *EmitAddTreeOfValues(Instruction *I, | |||
992 | SmallVectorImpl<WeakTrackingVH> &Ops) { | |||
993 | if (Ops.size() == 1) return Ops.back(); | |||
994 | ||||
995 | Value *V1 = Ops.back(); | |||
996 | Ops.pop_back(); | |||
997 | Value *V2 = EmitAddTreeOfValues(I, Ops); | |||
998 | return CreateAdd(V2, V1, "reass.add", I, I); | |||
999 | } | |||
1000 | ||||
1001 | /// If V is an expression tree that is a multiplication sequence, | |||
1002 | /// and if this sequence contains a multiply by Factor, | |||
1003 | /// remove Factor from the tree and return the new tree. | |||
1004 | Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) { | |||
1005 | BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); | |||
1006 | if (!BO) | |||
1007 | return nullptr; | |||
1008 | ||||
1009 | SmallVector<RepeatedValue, 8> Tree; | |||
1010 | MadeChange |= LinearizeExprTree(BO, Tree); | |||
1011 | SmallVector<ValueEntry, 8> Factors; | |||
1012 | Factors.reserve(Tree.size()); | |||
1013 | for (unsigned i = 0, e = Tree.size(); i != e; ++i) { | |||
1014 | RepeatedValue E = Tree[i]; | |||
1015 | Factors.append(E.second.getZExtValue(), | |||
1016 | ValueEntry(getRank(E.first), E.first)); | |||
1017 | } | |||
1018 | ||||
1019 | bool FoundFactor = false; | |||
1020 | bool NeedsNegate = false; | |||
1021 | for (unsigned i = 0, e = Factors.size(); i != e; ++i) { | |||
1022 | if (Factors[i].Op == Factor) { | |||
1023 | FoundFactor = true; | |||
1024 | Factors.erase(Factors.begin()+i); | |||
1025 | break; | |||
1026 | } | |||
1027 | ||||
1028 | // If this is a negative version of this factor, remove it. | |||
1029 | if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) { | |||
1030 | if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) | |||
1031 | if (FC1->getValue() == -FC2->getValue()) { | |||
1032 | FoundFactor = NeedsNegate = true; | |||
1033 | Factors.erase(Factors.begin()+i); | |||
1034 | break; | |||
1035 | } | |||
1036 | } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) { | |||
1037 | if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) { | |||
1038 | const APFloat &F1 = FC1->getValueAPF(); | |||
1039 | APFloat F2(FC2->getValueAPF()); | |||
1040 | F2.changeSign(); | |||
1041 | if (F1.compare(F2) == APFloat::cmpEqual) { | |||
1042 | FoundFactor = NeedsNegate = true; | |||
1043 | Factors.erase(Factors.begin() + i); | |||
1044 | break; | |||
1045 | } | |||
1046 | } | |||
1047 | } | |||
1048 | } | |||
1049 | ||||
1050 | if (!FoundFactor) { | |||
1051 | // Make sure to restore the operands to the expression tree. | |||
1052 | RewriteExprTree(BO, Factors); | |||
1053 | return nullptr; | |||
1054 | } | |||
1055 | ||||
1056 | BasicBlock::iterator InsertPt = ++BO->getIterator(); | |||
1057 | ||||
1058 | // If this was just a single multiply, remove the multiply and return the only | |||
1059 | // remaining operand. | |||
1060 | if (Factors.size() == 1) { | |||
1061 | RedoInsts.insert(BO); | |||
1062 | V = Factors[0].Op; | |||
1063 | } else { | |||
1064 | RewriteExprTree(BO, Factors); | |||
1065 | V = BO; | |||
1066 | } | |||
1067 | ||||
1068 | if (NeedsNegate) | |||
1069 | V = CreateNeg(V, "neg", &*InsertPt, BO); | |||
1070 | ||||
1071 | return V; | |||
1072 | } | |||
1073 | ||||
1074 | /// If V is a single-use multiply, recursively add its operands as factors, | |||
1075 | /// otherwise add V to the list of factors. | |||
1076 | /// | |||
1077 | /// Ops is the top-level list of add operands we're trying to factor. | |||
1078 | static void FindSingleUseMultiplyFactors(Value *V, | |||
1079 | SmallVectorImpl<Value*> &Factors) { | |||
1080 | BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); | |||
1081 | if (!BO) { | |||
1082 | Factors.push_back(V); | |||
1083 | return; | |||
1084 | } | |||
1085 | ||||
1086 | // Otherwise, add the LHS and RHS to the list of factors. | |||
1087 | FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); | |||
1088 | FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); | |||
1089 | } | |||
1090 | ||||
1091 | /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction. | |||
1092 | /// This optimizes based on identities. If it can be reduced to a single Value, | |||
1093 | /// it is returned, otherwise the Ops list is mutated as necessary. | |||
1094 | static Value *OptimizeAndOrXor(unsigned Opcode, | |||
1095 | SmallVectorImpl<ValueEntry> &Ops) { | |||
1096 | // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. | |||
1097 | // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. | |||
1098 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { | |||
1099 | // First, check for X and ~X in the operand list. | |||
1100 | assert(i < Ops.size())(static_cast <bool> (i < Ops.size()) ? void (0) : __assert_fail ("i < Ops.size()", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 1100, __extension__ __PRETTY_FUNCTION__)); | |||
1101 | if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. | |||
1102 | Value *X = BinaryOperator::getNotArgument(Ops[i].Op); | |||
1103 | unsigned FoundX = FindInOperandList(Ops, i, X); | |||
1104 | if (FoundX != i) { | |||
1105 | if (Opcode == Instruction::And) // ...&X&~X = 0 | |||
1106 | return Constant::getNullValue(X->getType()); | |||
1107 | ||||
1108 | if (Opcode == Instruction::Or) // ...|X|~X = -1 | |||
1109 | return Constant::getAllOnesValue(X->getType()); | |||
1110 | } | |||
1111 | } | |||
1112 | ||||
1113 | // Next, check for duplicate pairs of values, which we assume are next to | |||
1114 | // each other, due to our sorting criteria. | |||
1115 | assert(i < Ops.size())(static_cast <bool> (i < Ops.size()) ? void (0) : __assert_fail ("i < Ops.size()", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 1115, __extension__ __PRETTY_FUNCTION__)); | |||
1116 | if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { | |||
1117 | if (Opcode == Instruction::And || Opcode == Instruction::Or) { | |||
1118 | // Drop duplicate values for And and Or. | |||
1119 | Ops.erase(Ops.begin()+i); | |||
1120 | --i; --e; | |||
1121 | ++NumAnnihil; | |||
1122 | continue; | |||
1123 | } | |||
1124 | ||||
1125 | // Drop pairs of values for Xor. | |||
1126 | assert(Opcode == Instruction::Xor)(static_cast <bool> (Opcode == Instruction::Xor) ? void (0) : __assert_fail ("Opcode == Instruction::Xor", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 1126, __extension__ __PRETTY_FUNCTION__)); | |||
1127 | if (e == 2) | |||
1128 | return Constant::getNullValue(Ops[0].Op->getType()); | |||
1129 | ||||
1130 | // Y ^ X^X -> Y | |||
1131 | Ops.erase(Ops.begin()+i, Ops.begin()+i+2); | |||
1132 | i -= 1; e -= 2; | |||
1133 | ++NumAnnihil; | |||
1134 | } | |||
1135 | } | |||
1136 | return nullptr; | |||
1137 | } | |||
1138 | ||||
1139 | /// Helper function of CombineXorOpnd(). It creates a bitwise-and | |||
1140 | /// instruction with the given two operands, and return the resulting | |||
1141 | /// instruction. There are two special cases: 1) if the constant operand is 0, | |||
1142 | /// it will return NULL. 2) if the constant is ~0, the symbolic operand will | |||
1143 | /// be returned. | |||
1144 | static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd, | |||
1145 | const APInt &ConstOpnd) { | |||
1146 | if (ConstOpnd.isNullValue()) | |||
1147 | return nullptr; | |||
1148 | ||||
1149 | if (ConstOpnd.isAllOnesValue()) | |||
1150 | return Opnd; | |||
1151 | ||||
1152 | Instruction *I = BinaryOperator::CreateAnd( | |||
1153 | Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra", | |||
1154 | InsertBefore); | |||
1155 | I->setDebugLoc(InsertBefore->getDebugLoc()); | |||
1156 | return I; | |||
1157 | } | |||
1158 | ||||
1159 | // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd" | |||
1160 | // into "R ^ C", where C would be 0, and R is a symbolic value. | |||
1161 | // | |||
1162 | // If it was successful, true is returned, and the "R" and "C" is returned | |||
1163 | // via "Res" and "ConstOpnd", respectively; otherwise, false is returned, | |||
1164 | // and both "Res" and "ConstOpnd" remain unchanged. | |||
1165 | bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, | |||
1166 | APInt &ConstOpnd, Value *&Res) { | |||
1167 | // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2 | |||
1168 | // = ((x | c1) ^ c1) ^ (c1 ^ c2) | |||
1169 | // = (x & ~c1) ^ (c1 ^ c2) | |||
1170 | // It is useful only when c1 == c2. | |||
1171 | if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isNullValue()) | |||
1172 | return false; | |||
1173 | ||||
1174 | if (!Opnd1->getValue()->hasOneUse()) | |||
1175 | return false; | |||
1176 | ||||
1177 | const APInt &C1 = Opnd1->getConstPart(); | |||
1178 | if (C1 != ConstOpnd) | |||
1179 | return false; | |||
1180 | ||||
1181 | Value *X = Opnd1->getSymbolicPart(); | |||
1182 | Res = createAndInstr(I, X, ~C1); | |||
1183 | // ConstOpnd was C2, now C1 ^ C2. | |||
1184 | ConstOpnd ^= C1; | |||
1185 | ||||
1186 | if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue())) | |||
1187 | RedoInsts.insert(T); | |||
1188 | return true; | |||
1189 | } | |||
1190 | ||||
1191 | // Helper function of OptimizeXor(). It tries to simplify | |||
1192 | // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a | |||
1193 | // symbolic value. | |||
1194 | // | |||
1195 | // If it was successful, true is returned, and the "R" and "C" is returned | |||
1196 | // via "Res" and "ConstOpnd", respectively (If the entire expression is | |||
1197 | // evaluated to a constant, the Res is set to NULL); otherwise, false is | |||
1198 | // returned, and both "Res" and "ConstOpnd" remain unchanged. | |||
1199 | bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, | |||
1200 | XorOpnd *Opnd2, APInt &ConstOpnd, | |||
1201 | Value *&Res) { | |||
1202 | Value *X = Opnd1->getSymbolicPart(); | |||
1203 | if (X != Opnd2->getSymbolicPart()) | |||
1204 | return false; | |||
1205 | ||||
1206 | // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.) | |||
1207 | int DeadInstNum = 1; | |||
1208 | if (Opnd1->getValue()->hasOneUse()) | |||
1209 | DeadInstNum++; | |||
1210 | if (Opnd2->getValue()->hasOneUse()) | |||
1211 | DeadInstNum++; | |||
1212 | ||||
1213 | // Xor-Rule 2: | |||
1214 | // (x | c1) ^ (x & c2) | |||
1215 | // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1 | |||
1216 | // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1 | |||
1217 | // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3 | |||
1218 | // | |||
1219 | if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) { | |||
1220 | if (Opnd2->isOrExpr()) | |||
1221 | std::swap(Opnd1, Opnd2); | |||
1222 | ||||
1223 | const APInt &C1 = Opnd1->getConstPart(); | |||
1224 | const APInt &C2 = Opnd2->getConstPart(); | |||
1225 | APInt C3((~C1) ^ C2); | |||
1226 | ||||
1227 | // Do not increase code size! | |||
1228 | if (!C3.isNullValue() && !C3.isAllOnesValue()) { | |||
1229 | int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2; | |||
1230 | if (NewInstNum > DeadInstNum) | |||
1231 | return false; | |||
1232 | } | |||
1233 | ||||
1234 | Res = createAndInstr(I, X, C3); | |||
1235 | ConstOpnd ^= C1; | |||
1236 | } else if (Opnd1->isOrExpr()) { | |||
1237 | // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2 | |||
1238 | // | |||
1239 | const APInt &C1 = Opnd1->getConstPart(); | |||
1240 | const APInt &C2 = Opnd2->getConstPart(); | |||
1241 | APInt C3 = C1 ^ C2; | |||
1242 | ||||
1243 | // Do not increase code size | |||
1244 | if (!C3.isNullValue() && !C3.isAllOnesValue()) { | |||
1245 | int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2; | |||
1246 | if (NewInstNum > DeadInstNum) | |||
1247 | return false; | |||
1248 | } | |||
1249 | ||||
1250 | Res = createAndInstr(I, X, C3); | |||
1251 | ConstOpnd ^= C3; | |||
1252 | } else { | |||
1253 | // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2)) | |||
1254 | // | |||
1255 | const APInt &C1 = Opnd1->getConstPart(); | |||
1256 | const APInt &C2 = Opnd2->getConstPart(); | |||
1257 | APInt C3 = C1 ^ C2; | |||
1258 | Res = createAndInstr(I, X, C3); | |||
1259 | } | |||
1260 | ||||
1261 | // Put the original operands in the Redo list; hope they will be deleted | |||
1262 | // as dead code. | |||
1263 | if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue())) | |||
1264 | RedoInsts.insert(T); | |||
1265 | if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue())) | |||
1266 | RedoInsts.insert(T); | |||
1267 | ||||
1268 | return true; | |||
1269 | } | |||
1270 | ||||
1271 | /// Optimize a series of operands to an 'xor' instruction. If it can be reduced | |||
1272 | /// to a single Value, it is returned, otherwise the Ops list is mutated as | |||
1273 | /// necessary. | |||
1274 | Value *ReassociatePass::OptimizeXor(Instruction *I, | |||
1275 | SmallVectorImpl<ValueEntry> &Ops) { | |||
1276 | if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops)) | |||
1277 | return V; | |||
1278 | ||||
1279 | if (Ops.size() == 1) | |||
1280 | return nullptr; | |||
1281 | ||||
1282 | SmallVector<XorOpnd, 8> Opnds; | |||
1283 | SmallVector<XorOpnd*, 8> OpndPtrs; | |||
1284 | Type *Ty = Ops[0].Op->getType(); | |||
1285 | APInt ConstOpnd(Ty->getScalarSizeInBits(), 0); | |||
1286 | ||||
1287 | // Step 1: Convert ValueEntry to XorOpnd | |||
1288 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { | |||
1289 | Value *V = Ops[i].Op; | |||
1290 | const APInt *C; | |||
1291 | // TODO: Support non-splat vectors. | |||
1292 | if (match(V, PatternMatch::m_APInt(C))) { | |||
1293 | ConstOpnd ^= *C; | |||
1294 | } else { | |||
1295 | XorOpnd O(V); | |||
1296 | O.setSymbolicRank(getRank(O.getSymbolicPart())); | |||
1297 | Opnds.push_back(O); | |||
1298 | } | |||
1299 | } | |||
1300 | ||||
1301 | // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds". | |||
1302 | // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate | |||
1303 | // the "OpndPtrs" as well. For the similar reason, do not fuse this loop | |||
1304 | // with the previous loop --- the iterator of the "Opnds" may be invalidated | |||
1305 | // when new elements are added to the vector. | |||
1306 | for (unsigned i = 0, e = Opnds.size(); i != e; ++i) | |||
1307 | OpndPtrs.push_back(&Opnds[i]); | |||
1308 | ||||
1309 | // Step 2: Sort the Xor-Operands in a way such that the operands containing | |||
1310 | // the same symbolic value cluster together. For instance, the input operand | |||
1311 | // sequence ("x | 123", "y & 456", "x & 789") will be sorted into: | |||
1312 | // ("x | 123", "x & 789", "y & 456"). | |||
1313 | // | |||
1314 | // The purpose is twofold: | |||
1315 | // 1) Cluster together the operands sharing the same symbolic-value. | |||
1316 | // 2) Operand having smaller symbolic-value-rank is permuted earlier, which | |||
1317 | // could potentially shorten crital path, and expose more loop-invariants. | |||
1318 | // Note that values' rank are basically defined in RPO order (FIXME). | |||
1319 | // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier | |||
1320 | // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2", | |||
1321 | // "z" in the order of X-Y-Z is better than any other orders. | |||
1322 | std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), | |||
1323 | [](XorOpnd *LHS, XorOpnd *RHS) { | |||
1324 | return LHS->getSymbolicRank() < RHS->getSymbolicRank(); | |||
1325 | }); | |||
1326 | ||||
1327 | // Step 3: Combine adjacent operands | |||
1328 | XorOpnd *PrevOpnd = nullptr; | |||
1329 | bool Changed = false; | |||
1330 | for (unsigned i = 0, e = Opnds.size(); i < e; i++) { | |||
1331 | XorOpnd *CurrOpnd = OpndPtrs[i]; | |||
1332 | // The combined value | |||
1333 | Value *CV; | |||
1334 | ||||
1335 | // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd" | |||
1336 | if (!ConstOpnd.isNullValue() && | |||
1337 | CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) { | |||
1338 | Changed = true; | |||
1339 | if (CV) | |||
1340 | *CurrOpnd = XorOpnd(CV); | |||
1341 | else { | |||
1342 | CurrOpnd->Invalidate(); | |||
1343 | continue; | |||
1344 | } | |||
1345 | } | |||
1346 | ||||
1347 | if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) { | |||
1348 | PrevOpnd = CurrOpnd; | |||
1349 | continue; | |||
1350 | } | |||
1351 | ||||
1352 | // step 3.2: When previous and current operands share the same symbolic | |||
1353 | // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd" | |||
1354 | if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) { | |||
1355 | // Remove previous operand | |||
1356 | PrevOpnd->Invalidate(); | |||
1357 | if (CV) { | |||
1358 | *CurrOpnd = XorOpnd(CV); | |||
1359 | PrevOpnd = CurrOpnd; | |||
1360 | } else { | |||
1361 | CurrOpnd->Invalidate(); | |||
1362 | PrevOpnd = nullptr; | |||
1363 | } | |||
1364 | Changed = true; | |||
1365 | } | |||
1366 | } | |||
1367 | ||||
1368 | // Step 4: Reassemble the Ops | |||
1369 | if (Changed) { | |||
1370 | Ops.clear(); | |||
1371 | for (unsigned int i = 0, e = Opnds.size(); i < e; i++) { | |||
1372 | XorOpnd &O = Opnds[i]; | |||
1373 | if (O.isInvalid()) | |||
1374 | continue; | |||
1375 | ValueEntry VE(getRank(O.getValue()), O.getValue()); | |||
1376 | Ops.push_back(VE); | |||
1377 | } | |||
1378 | if (!ConstOpnd.isNullValue()) { | |||
1379 | Value *C = ConstantInt::get(Ty, ConstOpnd); | |||
1380 | ValueEntry VE(getRank(C), C); | |||
1381 | Ops.push_back(VE); | |||
1382 | } | |||
1383 | unsigned Sz = Ops.size(); | |||
1384 | if (Sz == 1) | |||
1385 | return Ops.back().Op; | |||
1386 | if (Sz == 0) { | |||
1387 | assert(ConstOpnd.isNullValue())(static_cast <bool> (ConstOpnd.isNullValue()) ? void (0 ) : __assert_fail ("ConstOpnd.isNullValue()", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 1387, __extension__ __PRETTY_FUNCTION__)); | |||
1388 | return ConstantInt::get(Ty, ConstOpnd); | |||
1389 | } | |||
1390 | } | |||
1391 | ||||
1392 | return nullptr; | |||
1393 | } | |||
1394 | ||||
1395 | /// Optimize a series of operands to an 'add' instruction. This | |||
1396 | /// optimizes based on identities. If it can be reduced to a single Value, it | |||
1397 | /// is returned, otherwise the Ops list is mutated as necessary. | |||
1398 | Value *ReassociatePass::OptimizeAdd(Instruction *I, | |||
1399 | SmallVectorImpl<ValueEntry> &Ops) { | |||
1400 | // Scan the operand lists looking for X and -X pairs. If we find any, we | |||
1401 | // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it, | |||
1402 | // scan for any | |||
1403 | // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. | |||
1404 | ||||
1405 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { | |||
| ||||
1406 | Value *TheOp = Ops[i].Op; | |||
1407 | // Check to see if we've seen this operand before. If so, we factor all | |||
1408 | // instances of the operand together. Due to our sorting criteria, we know | |||
1409 | // that these need to be next to each other in the vector. | |||
1410 | if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { | |||
1411 | // Rescan the list, remove all instances of this operand from the expr. | |||
1412 | unsigned NumFound = 0; | |||
1413 | do { | |||
1414 | Ops.erase(Ops.begin()+i); | |||
1415 | ++NumFound; | |||
1416 | } while (i != Ops.size() && Ops[i].Op == TheOp); | |||
1417 | ||||
1418 | DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'; } } while (false ); | |||
1419 | ++NumFactor; | |||
1420 | ||||
1421 | // Insert a new multiply. | |||
1422 | Type *Ty = TheOp->getType(); | |||
1423 | Constant *C = Ty->isIntOrIntVectorTy() ? | |||
1424 | ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound); | |||
1425 | Instruction *Mul = CreateMul(TheOp, C, "factor", I, I); | |||
1426 | ||||
1427 | // Now that we have inserted a multiply, optimize it. This allows us to | |||
1428 | // handle cases that require multiple factoring steps, such as this: | |||
1429 | // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 | |||
1430 | RedoInsts.insert(Mul); | |||
1431 | ||||
1432 | // If every add operand was a duplicate, return the multiply. | |||
1433 | if (Ops.empty()) | |||
1434 | return Mul; | |||
1435 | ||||
1436 | // Otherwise, we had some input that didn't have the dupe, such as | |||
1437 | // "A + A + B" -> "A*2 + B". Add the new multiply to the list of | |||
1438 | // things being added by this operation. | |||
1439 | Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); | |||
1440 | ||||
1441 | --i; | |||
1442 | e = Ops.size(); | |||
1443 | continue; | |||
1444 | } | |||
1445 | ||||
1446 | // Check for X and -X or X and ~X in the operand list. | |||
1447 | if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) && | |||
1448 | !BinaryOperator::isNot(TheOp)) | |||
1449 | continue; | |||
1450 | ||||
1451 | Value *X = nullptr; | |||
1452 | if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)) | |||
1453 | X = BinaryOperator::getNegArgument(TheOp); | |||
1454 | else if (BinaryOperator::isNot(TheOp)) | |||
1455 | X = BinaryOperator::getNotArgument(TheOp); | |||
1456 | ||||
1457 | unsigned FoundX = FindInOperandList(Ops, i, X); | |||
1458 | if (FoundX == i) | |||
1459 | continue; | |||
1460 | ||||
1461 | // Remove X and -X from the operand list. | |||
1462 | if (Ops.size() == 2 && | |||
1463 | (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))) | |||
1464 | return Constant::getNullValue(X->getType()); | |||
1465 | ||||
1466 | // Remove X and ~X from the operand list. | |||
1467 | if (Ops.size() == 2 && BinaryOperator::isNot(TheOp)) | |||
1468 | return Constant::getAllOnesValue(X->getType()); | |||
| ||||
1469 | ||||
1470 | Ops.erase(Ops.begin()+i); | |||
1471 | if (i < FoundX) | |||
1472 | --FoundX; | |||
1473 | else | |||
1474 | --i; // Need to back up an extra one. | |||
1475 | Ops.erase(Ops.begin()+FoundX); | |||
1476 | ++NumAnnihil; | |||
1477 | --i; // Revisit element. | |||
1478 | e -= 2; // Removed two elements. | |||
1479 | ||||
1480 | // if X and ~X we append -1 to the operand list. | |||
1481 | if (BinaryOperator::isNot(TheOp)) { | |||
1482 | Value *V = Constant::getAllOnesValue(X->getType()); | |||
1483 | Ops.insert(Ops.end(), ValueEntry(getRank(V), V)); | |||
1484 | e += 1; | |||
1485 | } | |||
1486 | } | |||
1487 | ||||
1488 | // Scan the operand list, checking to see if there are any common factors | |||
1489 | // between operands. Consider something like A*A+A*B*C+D. We would like to | |||
1490 | // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. | |||
1491 | // To efficiently find this, we count the number of times a factor occurs | |||
1492 | // for any ADD operands that are MULs. | |||
1493 | DenseMap<Value*, unsigned> FactorOccurrences; | |||
1494 | ||||
1495 | // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) | |||
1496 | // where they are actually the same multiply. | |||
1497 | unsigned MaxOcc = 0; | |||
1498 | Value *MaxOccVal = nullptr; | |||
1499 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { | |||
1500 | BinaryOperator *BOp = | |||
1501 | isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); | |||
1502 | if (!BOp) | |||
1503 | continue; | |||
1504 | ||||
1505 | // Compute all of the factors of this added value. | |||
1506 | SmallVector<Value*, 8> Factors; | |||
1507 | FindSingleUseMultiplyFactors(BOp, Factors); | |||
1508 | assert(Factors.size() > 1 && "Bad linearize!")(static_cast <bool> (Factors.size() > 1 && "Bad linearize!" ) ? void (0) : __assert_fail ("Factors.size() > 1 && \"Bad linearize!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 1508, __extension__ __PRETTY_FUNCTION__)); | |||
1509 | ||||
1510 | // Add one to FactorOccurrences for each unique factor in this op. | |||
1511 | SmallPtrSet<Value*, 8> Duplicates; | |||
1512 | for (unsigned i = 0, e = Factors.size(); i != e; ++i) { | |||
1513 | Value *Factor = Factors[i]; | |||
1514 | if (!Duplicates.insert(Factor).second) | |||
1515 | continue; | |||
1516 | ||||
1517 | unsigned Occ = ++FactorOccurrences[Factor]; | |||
1518 | if (Occ > MaxOcc) { | |||
1519 | MaxOcc = Occ; | |||
1520 | MaxOccVal = Factor; | |||
1521 | } | |||
1522 | ||||
1523 | // If Factor is a negative constant, add the negated value as a factor | |||
1524 | // because we can percolate the negate out. Watch for minint, which | |||
1525 | // cannot be positivified. | |||
1526 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) { | |||
1527 | if (CI->isNegative() && !CI->isMinValue(true)) { | |||
1528 | Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); | |||
1529 | if (!Duplicates.insert(Factor).second) | |||
1530 | continue; | |||
1531 | unsigned Occ = ++FactorOccurrences[Factor]; | |||
1532 | if (Occ > MaxOcc) { | |||
1533 | MaxOcc = Occ; | |||
1534 | MaxOccVal = Factor; | |||
1535 | } | |||
1536 | } | |||
1537 | } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) { | |||
1538 | if (CF->isNegative()) { | |||
1539 | APFloat F(CF->getValueAPF()); | |||
1540 | F.changeSign(); | |||
1541 | Factor = ConstantFP::get(CF->getContext(), F); | |||
1542 | if (!Duplicates.insert(Factor).second) | |||
1543 | continue; | |||
1544 | unsigned Occ = ++FactorOccurrences[Factor]; | |||
1545 | if (Occ > MaxOcc) { | |||
1546 | MaxOcc = Occ; | |||
1547 | MaxOccVal = Factor; | |||
1548 | } | |||
1549 | } | |||
1550 | } | |||
1551 | } | |||
1552 | } | |||
1553 | ||||
1554 | // If any factor occurred more than one time, we can pull it out. | |||
1555 | if (MaxOcc > 1) { | |||
1556 | DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'; } } while ( false); | |||
1557 | ++NumFactor; | |||
1558 | ||||
1559 | // Create a new instruction that uses the MaxOccVal twice. If we don't do | |||
1560 | // this, we could otherwise run into situations where removing a factor | |||
1561 | // from an expression will drop a use of maxocc, and this can cause | |||
1562 | // RemoveFactorFromExpression on successive values to behave differently. | |||
1563 | Instruction *DummyInst = | |||
1564 | I->getType()->isIntOrIntVectorTy() | |||
1565 | ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal) | |||
1566 | : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal); | |||
1567 | ||||
1568 | SmallVector<WeakTrackingVH, 4> NewMulOps; | |||
1569 | for (unsigned i = 0; i != Ops.size(); ++i) { | |||
1570 | // Only try to remove factors from expressions we're allowed to. | |||
1571 | BinaryOperator *BOp = | |||
1572 | isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); | |||
1573 | if (!BOp) | |||
1574 | continue; | |||
1575 | ||||
1576 | if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { | |||
1577 | // The factorized operand may occur several times. Convert them all in | |||
1578 | // one fell swoop. | |||
1579 | for (unsigned j = Ops.size(); j != i;) { | |||
1580 | --j; | |||
1581 | if (Ops[j].Op == Ops[i].Op) { | |||
1582 | NewMulOps.push_back(V); | |||
1583 | Ops.erase(Ops.begin()+j); | |||
1584 | } | |||
1585 | } | |||
1586 | --i; | |||
1587 | } | |||
1588 | } | |||
1589 | ||||
1590 | // No need for extra uses anymore. | |||
1591 | DummyInst->deleteValue(); | |||
1592 | ||||
1593 | unsigned NumAddedValues = NewMulOps.size(); | |||
1594 | Value *V = EmitAddTreeOfValues(I, NewMulOps); | |||
1595 | ||||
1596 | // Now that we have inserted the add tree, optimize it. This allows us to | |||
1597 | // handle cases that require multiple factoring steps, such as this: | |||
1598 | // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) | |||
1599 | assert(NumAddedValues > 1 && "Each occurrence should contribute a value")(static_cast <bool> (NumAddedValues > 1 && "Each occurrence should contribute a value" ) ? void (0) : __assert_fail ("NumAddedValues > 1 && \"Each occurrence should contribute a value\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 1599, __extension__ __PRETTY_FUNCTION__)); | |||
1600 | (void)NumAddedValues; | |||
1601 | if (Instruction *VI = dyn_cast<Instruction>(V)) | |||
1602 | RedoInsts.insert(VI); | |||
1603 | ||||
1604 | // Create the multiply. | |||
1605 | Instruction *V2 = CreateMul(V, MaxOccVal, "reass.mul", I, I); | |||
1606 | ||||
1607 | // Rerun associate on the multiply in case the inner expression turned into | |||
1608 | // a multiply. We want to make sure that we keep things in canonical form. | |||
1609 | RedoInsts.insert(V2); | |||
1610 | ||||
1611 | // If every add operand included the factor (e.g. "A*B + A*C"), then the | |||
1612 | // entire result expression is just the multiply "A*(B+C)". | |||
1613 | if (Ops.empty()) | |||
1614 | return V2; | |||
1615 | ||||
1616 | // Otherwise, we had some input that didn't have the factor, such as | |||
1617 | // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of | |||
1618 | // things being added by this operation. | |||
1619 | Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); | |||
1620 | } | |||
1621 | ||||
1622 | return nullptr; | |||
1623 | } | |||
1624 | ||||
1625 | /// \brief Build up a vector of value/power pairs factoring a product. | |||
1626 | /// | |||
1627 | /// Given a series of multiplication operands, build a vector of factors and | |||
1628 | /// the powers each is raised to when forming the final product. Sort them in | |||
1629 | /// the order of descending power. | |||
1630 | /// | |||
1631 | /// (x*x) -> [(x, 2)] | |||
1632 | /// ((x*x)*x) -> [(x, 3)] | |||
1633 | /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)] | |||
1634 | /// | |||
1635 | /// \returns Whether any factors have a power greater than one. | |||
1636 | static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, | |||
1637 | SmallVectorImpl<Factor> &Factors) { | |||
1638 | // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this. | |||
1639 | // Compute the sum of powers of simplifiable factors. | |||
1640 | unsigned FactorPowerSum = 0; | |||
1641 | for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) { | |||
1642 | Value *Op = Ops[Idx-1].Op; | |||
1643 | ||||
1644 | // Count the number of occurrences of this value. | |||
1645 | unsigned Count = 1; | |||
1646 | for (; Idx < Size && Ops[Idx].Op == Op; ++Idx) | |||
1647 | ++Count; | |||
1648 | // Track for simplification all factors which occur 2 or more times. | |||
1649 | if (Count > 1) | |||
1650 | FactorPowerSum += Count; | |||
1651 | } | |||
1652 | ||||
1653 | // We can only simplify factors if the sum of the powers of our simplifiable | |||
1654 | // factors is 4 or higher. When that is the case, we will *always* have | |||
1655 | // a simplification. This is an important invariant to prevent cyclicly | |||
1656 | // trying to simplify already minimal formations. | |||
1657 | if (FactorPowerSum < 4) | |||
1658 | return false; | |||
1659 | ||||
1660 | // Now gather the simplifiable factors, removing them from Ops. | |||
1661 | FactorPowerSum = 0; | |||
1662 | for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) { | |||
1663 | Value *Op = Ops[Idx-1].Op; | |||
1664 | ||||
1665 | // Count the number of occurrences of this value. | |||
1666 | unsigned Count = 1; | |||
1667 | for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx) | |||
1668 | ++Count; | |||
1669 | if (Count == 1) | |||
1670 | continue; | |||
1671 | // Move an even number of occurrences to Factors. | |||
1672 | Count &= ~1U; | |||
1673 | Idx -= Count; | |||
1674 | FactorPowerSum += Count; | |||
1675 | Factors.push_back(Factor(Op, Count)); | |||
1676 | Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count); | |||
1677 | } | |||
1678 | ||||
1679 | // None of the adjustments above should have reduced the sum of factor powers | |||
1680 | // below our mininum of '4'. | |||
1681 | assert(FactorPowerSum >= 4)(static_cast <bool> (FactorPowerSum >= 4) ? void (0) : __assert_fail ("FactorPowerSum >= 4", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 1681, __extension__ __PRETTY_FUNCTION__)); | |||
1682 | ||||
1683 | std::stable_sort(Factors.begin(), Factors.end(), | |||
1684 | [](const Factor &LHS, const Factor &RHS) { | |||
1685 | return LHS.Power > RHS.Power; | |||
1686 | }); | |||
1687 | return true; | |||
1688 | } | |||
1689 | ||||
1690 | /// \brief Build a tree of multiplies, computing the product of Ops. | |||
1691 | static Value *buildMultiplyTree(IRBuilder<> &Builder, | |||
1692 | SmallVectorImpl<Value*> &Ops) { | |||
1693 | if (Ops.size() == 1) | |||
1694 | return Ops.back(); | |||
1695 | ||||
1696 | Value *LHS = Ops.pop_back_val(); | |||
1697 | do { | |||
1698 | if (LHS->getType()->isIntOrIntVectorTy()) | |||
1699 | LHS = Builder.CreateMul(LHS, Ops.pop_back_val()); | |||
1700 | else | |||
1701 | LHS = Builder.CreateFMul(LHS, Ops.pop_back_val()); | |||
1702 | } while (!Ops.empty()); | |||
1703 | ||||
1704 | return LHS; | |||
1705 | } | |||
1706 | ||||
1707 | /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*... | |||
1708 | /// | |||
1709 | /// Given a vector of values raised to various powers, where no two values are | |||
1710 | /// equal and the powers are sorted in decreasing order, compute the minimal | |||
1711 | /// DAG of multiplies to compute the final product, and return that product | |||
1712 | /// value. | |||
1713 | Value * | |||
1714 | ReassociatePass::buildMinimalMultiplyDAG(IRBuilder<> &Builder, | |||
1715 | SmallVectorImpl<Factor> &Factors) { | |||
1716 | assert(Factors[0].Power)(static_cast <bool> (Factors[0].Power) ? void (0) : __assert_fail ("Factors[0].Power", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 1716, __extension__ __PRETTY_FUNCTION__)); | |||
1717 | SmallVector<Value *, 4> OuterProduct; | |||
1718 | for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size(); | |||
1719 | Idx < Size && Factors[Idx].Power > 0; ++Idx) { | |||
1720 | if (Factors[Idx].Power != Factors[LastIdx].Power) { | |||
1721 | LastIdx = Idx; | |||
1722 | continue; | |||
1723 | } | |||
1724 | ||||
1725 | // We want to multiply across all the factors with the same power so that | |||
1726 | // we can raise them to that power as a single entity. Build a mini tree | |||
1727 | // for that. | |||
1728 | SmallVector<Value *, 4> InnerProduct; | |||
1729 | InnerProduct.push_back(Factors[LastIdx].Base); | |||
1730 | do { | |||
1731 | InnerProduct.push_back(Factors[Idx].Base); | |||
1732 | ++Idx; | |||
1733 | } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power); | |||
1734 | ||||
1735 | // Reset the base value of the first factor to the new expression tree. | |||
1736 | // We'll remove all the factors with the same power in a second pass. | |||
1737 | Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct); | |||
1738 | if (Instruction *MI = dyn_cast<Instruction>(M)) | |||
1739 | RedoInsts.insert(MI); | |||
1740 | ||||
1741 | LastIdx = Idx; | |||
1742 | } | |||
1743 | // Unique factors with equal powers -- we've folded them into the first one's | |||
1744 | // base. | |||
1745 | Factors.erase(std::unique(Factors.begin(), Factors.end(), | |||
1746 | [](const Factor &LHS, const Factor &RHS) { | |||
1747 | return LHS.Power == RHS.Power; | |||
1748 | }), | |||
1749 | Factors.end()); | |||
1750 | ||||
1751 | // Iteratively collect the base of each factor with an add power into the | |||
1752 | // outer product, and halve each power in preparation for squaring the | |||
1753 | // expression. | |||
1754 | for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) { | |||
1755 | if (Factors[Idx].Power & 1) | |||
1756 | OuterProduct.push_back(Factors[Idx].Base); | |||
1757 | Factors[Idx].Power >>= 1; | |||
1758 | } | |||
1759 | if (Factors[0].Power) { | |||
1760 | Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors); | |||
1761 | OuterProduct.push_back(SquareRoot); | |||
1762 | OuterProduct.push_back(SquareRoot); | |||
1763 | } | |||
1764 | if (OuterProduct.size() == 1) | |||
1765 | return OuterProduct.front(); | |||
1766 | ||||
1767 | Value *V = buildMultiplyTree(Builder, OuterProduct); | |||
1768 | return V; | |||
1769 | } | |||
1770 | ||||
1771 | Value *ReassociatePass::OptimizeMul(BinaryOperator *I, | |||
1772 | SmallVectorImpl<ValueEntry> &Ops) { | |||
1773 | // We can only optimize the multiplies when there is a chain of more than | |||
1774 | // three, such that a balanced tree might require fewer total multiplies. | |||
1775 | if (Ops.size() < 4) | |||
1776 | return nullptr; | |||
1777 | ||||
1778 | // Try to turn linear trees of multiplies without other uses of the | |||
1779 | // intermediate stages into minimal multiply DAGs with perfect sub-expression | |||
1780 | // re-use. | |||
1781 | SmallVector<Factor, 4> Factors; | |||
1782 | if (!collectMultiplyFactors(Ops, Factors)) | |||
1783 | return nullptr; // All distinct factors, so nothing left for us to do. | |||
1784 | ||||
1785 | IRBuilder<> Builder(I); | |||
1786 | // The reassociate transformation for FP operations is performed only | |||
1787 | // if unsafe algebra is permitted by FastMathFlags. Propagate those flags | |||
1788 | // to the newly generated operations. | |||
1789 | if (auto FPI = dyn_cast<FPMathOperator>(I)) | |||
1790 | Builder.setFastMathFlags(FPI->getFastMathFlags()); | |||
1791 | ||||
1792 | Value *V = buildMinimalMultiplyDAG(Builder, Factors); | |||
1793 | if (Ops.empty()) | |||
1794 | return V; | |||
1795 | ||||
1796 | ValueEntry NewEntry = ValueEntry(getRank(V), V); | |||
1797 | Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry); | |||
1798 | return nullptr; | |||
1799 | } | |||
1800 | ||||
1801 | Value *ReassociatePass::OptimizeExpression(BinaryOperator *I, | |||
1802 | SmallVectorImpl<ValueEntry> &Ops) { | |||
1803 | // Now that we have the linearized expression tree, try to optimize it. | |||
1804 | // Start by folding any constants that we found. | |||
1805 | Constant *Cst = nullptr; | |||
1806 | unsigned Opcode = I->getOpcode(); | |||
1807 | while (!Ops.empty() && isa<Constant>(Ops.back().Op)) { | |||
1808 | Constant *C = cast<Constant>(Ops.pop_back_val().Op); | |||
1809 | Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C; | |||
1810 | } | |||
1811 | // If there was nothing but constants then we are done. | |||
1812 | if (Ops.empty()) | |||
1813 | return Cst; | |||
1814 | ||||
1815 | // Put the combined constant back at the end of the operand list, except if | |||
1816 | // there is no point. For example, an add of 0 gets dropped here, while a | |||
1817 | // multiplication by zero turns the whole expression into zero. | |||
1818 | if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) { | |||
1819 | if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType())) | |||
1820 | return Cst; | |||
1821 | Ops.push_back(ValueEntry(0, Cst)); | |||
1822 | } | |||
1823 | ||||
1824 | if (Ops.size() == 1) return Ops[0].Op; | |||
1825 | ||||
1826 | // Handle destructive annihilation due to identities between elements in the | |||
1827 | // argument list here. | |||
1828 | unsigned NumOps = Ops.size(); | |||
1829 | switch (Opcode) { | |||
1830 | default: break; | |||
1831 | case Instruction::And: | |||
1832 | case Instruction::Or: | |||
1833 | if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) | |||
1834 | return Result; | |||
1835 | break; | |||
1836 | ||||
1837 | case Instruction::Xor: | |||
1838 | if (Value *Result = OptimizeXor(I, Ops)) | |||
1839 | return Result; | |||
1840 | break; | |||
1841 | ||||
1842 | case Instruction::Add: | |||
1843 | case Instruction::FAdd: | |||
1844 | if (Value *Result = OptimizeAdd(I, Ops)) | |||
1845 | return Result; | |||
1846 | break; | |||
1847 | ||||
1848 | case Instruction::Mul: | |||
1849 | case Instruction::FMul: | |||
1850 | if (Value *Result = OptimizeMul(I, Ops)) | |||
1851 | return Result; | |||
1852 | break; | |||
1853 | } | |||
1854 | ||||
1855 | if (Ops.size() != NumOps) | |||
1856 | return OptimizeExpression(I, Ops); | |||
1857 | return nullptr; | |||
1858 | } | |||
1859 | ||||
1860 | // Remove dead instructions and if any operands are trivially dead add them to | |||
1861 | // Insts so they will be removed as well. | |||
1862 | void ReassociatePass::RecursivelyEraseDeadInsts( | |||
1863 | Instruction *I, SetVector<AssertingVH<Instruction>> &Insts) { | |||
1864 | assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!")(static_cast <bool> (isInstructionTriviallyDead(I) && "Trivially dead instructions only!") ? void (0) : __assert_fail ("isInstructionTriviallyDead(I) && \"Trivially dead instructions only!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 1864, __extension__ __PRETTY_FUNCTION__)); | |||
1865 | SmallVector<Value *, 4> Ops(I->op_begin(), I->op_end()); | |||
1866 | ValueRankMap.erase(I); | |||
1867 | Insts.remove(I); | |||
1868 | RedoInsts.remove(I); | |||
1869 | I->eraseFromParent(); | |||
1870 | for (auto Op : Ops) | |||
1871 | if (Instruction *OpInst = dyn_cast<Instruction>(Op)) | |||
1872 | if (OpInst->use_empty()) | |||
1873 | Insts.insert(OpInst); | |||
1874 | } | |||
1875 | ||||
1876 | /// Zap the given instruction, adding interesting operands to the work list. | |||
1877 | void ReassociatePass::EraseInst(Instruction *I) { | |||
1878 | assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!")(static_cast <bool> (isInstructionTriviallyDead(I) && "Trivially dead instructions only!") ? void (0) : __assert_fail ("isInstructionTriviallyDead(I) && \"Trivially dead instructions only!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 1878, __extension__ __PRETTY_FUNCTION__)); | |||
1879 | DEBUG(dbgs() << "Erasing dead inst: "; I->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Erasing dead inst: "; I-> dump(); } } while (false); | |||
1880 | ||||
1881 | SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); | |||
1882 | // Erase the dead instruction. | |||
1883 | ValueRankMap.erase(I); | |||
1884 | RedoInsts.remove(I); | |||
1885 | I->eraseFromParent(); | |||
1886 | // Optimize its operands. | |||
1887 | SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes. | |||
1888 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) | |||
1889 | if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) { | |||
1890 | // If this is a node in an expression tree, climb to the expression root | |||
1891 | // and add that since that's where optimization actually happens. | |||
1892 | unsigned Opcode = Op->getOpcode(); | |||
1893 | while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode && | |||
1894 | Visited.insert(Op).second) | |||
1895 | Op = Op->user_back(); | |||
1896 | RedoInsts.insert(Op); | |||
1897 | } | |||
1898 | ||||
1899 | MadeChange = true; | |||
1900 | } | |||
1901 | ||||
1902 | // Canonicalize expressions of the following form: | |||
1903 | // x + (-Constant * y) -> x - (Constant * y) | |||
1904 | // x - (-Constant * y) -> x + (Constant * y) | |||
1905 | Instruction *ReassociatePass::canonicalizeNegConstExpr(Instruction *I) { | |||
1906 | if (!I->hasOneUse() || I->getType()->isVectorTy()) | |||
1907 | return nullptr; | |||
1908 | ||||
1909 | // Must be a fmul or fdiv instruction. | |||
1910 | unsigned Opcode = I->getOpcode(); | |||
1911 | if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv) | |||
1912 | return nullptr; | |||
1913 | ||||
1914 | auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0)); | |||
1915 | auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1)); | |||
1916 | ||||
1917 | // Both operands are constant, let it get constant folded away. | |||
1918 | if (C0 && C1) | |||
1919 | return nullptr; | |||
1920 | ||||
1921 | ConstantFP *CF = C0 ? C0 : C1; | |||
1922 | ||||
1923 | // Must have one constant operand. | |||
1924 | if (!CF) | |||
1925 | return nullptr; | |||
1926 | ||||
1927 | // Must be a negative ConstantFP. | |||
1928 | if (!CF->isNegative()) | |||
1929 | return nullptr; | |||
1930 | ||||
1931 | // User must be a binary operator with one or more uses. | |||
1932 | Instruction *User = I->user_back(); | |||
1933 | if (!isa<BinaryOperator>(User) || User->use_empty()) | |||
1934 | return nullptr; | |||
1935 | ||||
1936 | unsigned UserOpcode = User->getOpcode(); | |||
1937 | if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub) | |||
1938 | return nullptr; | |||
1939 | ||||
1940 | // Subtraction is not commutative. Explicitly, the following transform is | |||
1941 | // not valid: (-Constant * y) - x -> x + (Constant * y) | |||
1942 | if (!User->isCommutative() && User->getOperand(1) != I) | |||
1943 | return nullptr; | |||
1944 | ||||
1945 | // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the | |||
1946 | // resulting subtract will be broken up later. This can get us into an | |||
1947 | // infinite loop during reassociation. | |||
1948 | if (UserOpcode == Instruction::FAdd && ShouldBreakUpSubtract(User)) | |||
1949 | return nullptr; | |||
1950 | ||||
1951 | // Change the sign of the constant. | |||
1952 | APFloat Val = CF->getValueAPF(); | |||
1953 | Val.changeSign(); | |||
1954 | I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val)); | |||
1955 | ||||
1956 | // Canonicalize I to RHS to simplify the next bit of logic. E.g., | |||
1957 | // ((-Const*y) + x) -> (x + (-Const*y)). | |||
1958 | if (User->getOperand(0) == I && User->isCommutative()) | |||
1959 | cast<BinaryOperator>(User)->swapOperands(); | |||
1960 | ||||
1961 | Value *Op0 = User->getOperand(0); | |||
1962 | Value *Op1 = User->getOperand(1); | |||
1963 | BinaryOperator *NI; | |||
1964 | switch (UserOpcode) { | |||
1965 | default: | |||
1966 | llvm_unreachable("Unexpected Opcode!")::llvm::llvm_unreachable_internal("Unexpected Opcode!", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 1966); | |||
1967 | case Instruction::FAdd: | |||
1968 | NI = BinaryOperator::CreateFSub(Op0, Op1); | |||
1969 | NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags()); | |||
1970 | break; | |||
1971 | case Instruction::FSub: | |||
1972 | NI = BinaryOperator::CreateFAdd(Op0, Op1); | |||
1973 | NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags()); | |||
1974 | break; | |||
1975 | } | |||
1976 | ||||
1977 | NI->insertBefore(User); | |||
1978 | NI->setName(User->getName()); | |||
1979 | User->replaceAllUsesWith(NI); | |||
1980 | NI->setDebugLoc(I->getDebugLoc()); | |||
1981 | RedoInsts.insert(I); | |||
1982 | MadeChange = true; | |||
1983 | return NI; | |||
1984 | } | |||
1985 | ||||
1986 | /// Inspect and optimize the given instruction. Note that erasing | |||
1987 | /// instructions is not allowed. | |||
1988 | void ReassociatePass::OptimizeInst(Instruction *I) { | |||
1989 | // Only consider operations that we understand. | |||
1990 | if (!isa<BinaryOperator>(I)) | |||
1991 | return; | |||
1992 | ||||
1993 | if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1))) | |||
1994 | // If an operand of this shift is a reassociable multiply, or if the shift | |||
1995 | // is used by a reassociable multiply or add, turn into a multiply. | |||
1996 | if (isReassociableOp(I->getOperand(0), Instruction::Mul) || | |||
1997 | (I->hasOneUse() && | |||
1998 | (isReassociableOp(I->user_back(), Instruction::Mul) || | |||
1999 | isReassociableOp(I->user_back(), Instruction::Add)))) { | |||
2000 | Instruction *NI = ConvertShiftToMul(I); | |||
2001 | RedoInsts.insert(I); | |||
2002 | MadeChange = true; | |||
2003 | I = NI; | |||
2004 | } | |||
2005 | ||||
2006 | // Canonicalize negative constants out of expressions. | |||
2007 | if (Instruction *Res = canonicalizeNegConstExpr(I)) | |||
2008 | I = Res; | |||
2009 | ||||
2010 | // Commute binary operators, to canonicalize the order of their operands. | |||
2011 | // This can potentially expose more CSE opportunities, and makes writing other | |||
2012 | // transformations simpler. | |||
2013 | if (I->isCommutative()) | |||
2014 | canonicalizeOperands(I); | |||
2015 | ||||
2016 | // Don't optimize floating-point instructions unless they are 'fast'. | |||
2017 | if (I->getType()->isFPOrFPVectorTy() && !I->isFast()) | |||
2018 | return; | |||
2019 | ||||
2020 | // Do not reassociate boolean (i1) expressions. We want to preserve the | |||
2021 | // original order of evaluation for short-circuited comparisons that | |||
2022 | // SimplifyCFG has folded to AND/OR expressions. If the expression | |||
2023 | // is not further optimized, it is likely to be transformed back to a | |||
2024 | // short-circuited form for code gen, and the source order may have been | |||
2025 | // optimized for the most likely conditions. | |||
2026 | if (I->getType()->isIntegerTy(1)) | |||
2027 | return; | |||
2028 | ||||
2029 | // If this is a subtract instruction which is not already in negate form, | |||
2030 | // see if we can convert it to X+-Y. | |||
2031 | if (I->getOpcode() == Instruction::Sub) { | |||
2032 | if (ShouldBreakUpSubtract(I)) { | |||
2033 | Instruction *NI = BreakUpSubtract(I, RedoInsts); | |||
2034 | RedoInsts.insert(I); | |||
2035 | MadeChange = true; | |||
2036 | I = NI; | |||
2037 | } else if (BinaryOperator::isNeg(I)) { | |||
2038 | // Otherwise, this is a negation. See if the operand is a multiply tree | |||
2039 | // and if this is not an inner node of a multiply tree. | |||
2040 | if (isReassociableOp(I->getOperand(1), Instruction::Mul) && | |||
2041 | (!I->hasOneUse() || | |||
2042 | !isReassociableOp(I->user_back(), Instruction::Mul))) { | |||
2043 | Instruction *NI = LowerNegateToMultiply(I); | |||
2044 | // If the negate was simplified, revisit the users to see if we can | |||
2045 | // reassociate further. | |||
2046 | for (User *U : NI->users()) { | |||
2047 | if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U)) | |||
2048 | RedoInsts.insert(Tmp); | |||
2049 | } | |||
2050 | RedoInsts.insert(I); | |||
2051 | MadeChange = true; | |||
2052 | I = NI; | |||
2053 | } | |||
2054 | } | |||
2055 | } else if (I->getOpcode() == Instruction::FSub) { | |||
2056 | if (ShouldBreakUpSubtract(I)) { | |||
2057 | Instruction *NI = BreakUpSubtract(I, RedoInsts); | |||
2058 | RedoInsts.insert(I); | |||
2059 | MadeChange = true; | |||
2060 | I = NI; | |||
2061 | } else if (BinaryOperator::isFNeg(I)) { | |||
2062 | // Otherwise, this is a negation. See if the operand is a multiply tree | |||
2063 | // and if this is not an inner node of a multiply tree. | |||
2064 | if (isReassociableOp(I->getOperand(1), Instruction::FMul) && | |||
2065 | (!I->hasOneUse() || | |||
2066 | !isReassociableOp(I->user_back(), Instruction::FMul))) { | |||
2067 | // If the negate was simplified, revisit the users to see if we can | |||
2068 | // reassociate further. | |||
2069 | Instruction *NI = LowerNegateToMultiply(I); | |||
2070 | for (User *U : NI->users()) { | |||
2071 | if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U)) | |||
2072 | RedoInsts.insert(Tmp); | |||
2073 | } | |||
2074 | RedoInsts.insert(I); | |||
2075 | MadeChange = true; | |||
2076 | I = NI; | |||
2077 | } | |||
2078 | } | |||
2079 | } | |||
2080 | ||||
2081 | // If this instruction is an associative binary operator, process it. | |||
2082 | if (!I->isAssociative()) return; | |||
2083 | BinaryOperator *BO = cast<BinaryOperator>(I); | |||
2084 | ||||
2085 | // If this is an interior node of a reassociable tree, ignore it until we | |||
2086 | // get to the root of the tree, to avoid N^2 analysis. | |||
2087 | unsigned Opcode = BO->getOpcode(); | |||
2088 | if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) { | |||
2089 | // During the initial run we will get to the root of the tree. | |||
2090 | // But if we get here while we are redoing instructions, there is no | |||
2091 | // guarantee that the root will be visited. So Redo later | |||
2092 | if (BO->user_back() != BO && | |||
2093 | BO->getParent() == BO->user_back()->getParent()) | |||
2094 | RedoInsts.insert(BO->user_back()); | |||
2095 | return; | |||
2096 | } | |||
2097 | ||||
2098 | // If this is an add tree that is used by a sub instruction, ignore it | |||
2099 | // until we process the subtract. | |||
2100 | if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add && | |||
2101 | cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub) | |||
2102 | return; | |||
2103 | if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd && | |||
2104 | cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub) | |||
2105 | return; | |||
2106 | ||||
2107 | ReassociateExpression(BO); | |||
2108 | } | |||
2109 | ||||
2110 | void ReassociatePass::ReassociateExpression(BinaryOperator *I) { | |||
2111 | // First, walk the expression tree, linearizing the tree, collecting the | |||
2112 | // operand information. | |||
2113 | SmallVector<RepeatedValue, 8> Tree; | |||
2114 | MadeChange |= LinearizeExprTree(I, Tree); | |||
2115 | SmallVector<ValueEntry, 8> Ops; | |||
2116 | Ops.reserve(Tree.size()); | |||
2117 | for (unsigned i = 0, e = Tree.size(); i != e; ++i) { | |||
2118 | RepeatedValue E = Tree[i]; | |||
2119 | Ops.append(E.second.getZExtValue(), | |||
2120 | ValueEntry(getRank(E.first), E.first)); | |||
2121 | } | |||
2122 | ||||
2123 | DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "RAIn:\t"; PrintOps(I, Ops ); dbgs() << '\n'; } } while (false); | |||
2124 | ||||
2125 | // Now that we have linearized the tree to a list and have gathered all of | |||
2126 | // the operands and their ranks, sort the operands by their rank. Use a | |||
2127 | // stable_sort so that values with equal ranks will have their relative | |||
2128 | // positions maintained (and so the compiler is deterministic). Note that | |||
2129 | // this sorts so that the highest ranking values end up at the beginning of | |||
2130 | // the vector. | |||
2131 | std::stable_sort(Ops.begin(), Ops.end()); | |||
2132 | ||||
2133 | // Now that we have the expression tree in a convenient | |||
2134 | // sorted form, optimize it globally if possible. | |||
2135 | if (Value *V = OptimizeExpression(I, Ops)) { | |||
2136 | if (V == I) | |||
2137 | // Self-referential expression in unreachable code. | |||
2138 | return; | |||
2139 | // This expression tree simplified to something that isn't a tree, | |||
2140 | // eliminate it. | |||
2141 | DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Reassoc to scalar: " << *V << '\n'; } } while (false); | |||
2142 | I->replaceAllUsesWith(V); | |||
2143 | if (Instruction *VI = dyn_cast<Instruction>(V)) | |||
2144 | if (I->getDebugLoc()) | |||
2145 | VI->setDebugLoc(I->getDebugLoc()); | |||
2146 | RedoInsts.insert(I); | |||
2147 | ++NumAnnihil; | |||
2148 | return; | |||
2149 | } | |||
2150 | ||||
2151 | // We want to sink immediates as deeply as possible except in the case where | |||
2152 | // this is a multiply tree used only by an add, and the immediate is a -1. | |||
2153 | // In this case we reassociate to put the negation on the outside so that we | |||
2154 | // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y | |||
2155 | if (I->hasOneUse()) { | |||
2156 | if (I->getOpcode() == Instruction::Mul && | |||
2157 | cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add && | |||
2158 | isa<ConstantInt>(Ops.back().Op) && | |||
2159 | cast<ConstantInt>(Ops.back().Op)->isMinusOne()) { | |||
2160 | ValueEntry Tmp = Ops.pop_back_val(); | |||
2161 | Ops.insert(Ops.begin(), Tmp); | |||
2162 | } else if (I->getOpcode() == Instruction::FMul && | |||
2163 | cast<Instruction>(I->user_back())->getOpcode() == | |||
2164 | Instruction::FAdd && | |||
2165 | isa<ConstantFP>(Ops.back().Op) && | |||
2166 | cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) { | |||
2167 | ValueEntry Tmp = Ops.pop_back_val(); | |||
2168 | Ops.insert(Ops.begin(), Tmp); | |||
2169 | } | |||
2170 | } | |||
2171 | ||||
2172 | DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "RAOut:\t"; PrintOps(I, Ops ); dbgs() << '\n'; } } while (false); | |||
2173 | ||||
2174 | if (Ops.size() == 1) { | |||
2175 | if (Ops[0].Op == I) | |||
2176 | // Self-referential expression in unreachable code. | |||
2177 | return; | |||
2178 | ||||
2179 | // This expression tree simplified to something that isn't a tree, | |||
2180 | // eliminate it. | |||
2181 | I->replaceAllUsesWith(Ops[0].Op); | |||
2182 | if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op)) | |||
2183 | OI->setDebugLoc(I->getDebugLoc()); | |||
2184 | RedoInsts.insert(I); | |||
2185 | return; | |||
2186 | } | |||
2187 | ||||
2188 | if (Ops.size() > 2 && Ops.size() <= GlobalReassociateLimit) { | |||
2189 | // Find the pair with the highest count in the pairmap and move it to the | |||
2190 | // back of the list so that it can later be CSE'd. | |||
2191 | // example: | |||
2192 | // a*b*c*d*e | |||
2193 | // if c*e is the most "popular" pair, we can express this as | |||
2194 | // (((c*e)*d)*b)*a | |||
2195 | unsigned Max = 1; | |||
2196 | unsigned BestRank = 0; | |||
2197 | std::pair<unsigned, unsigned> BestPair; | |||
2198 | unsigned Idx = I->getOpcode() - Instruction::BinaryOpsBegin; | |||
2199 | for (unsigned i = 0; i < Ops.size() - 1; ++i) | |||
2200 | for (unsigned j = i + 1; j < Ops.size(); ++j) { | |||
2201 | unsigned Score = 0; | |||
2202 | Value *Op0 = Ops[i].Op; | |||
2203 | Value *Op1 = Ops[j].Op; | |||
2204 | if (std::less<Value *>()(Op1, Op0)) | |||
2205 | std::swap(Op0, Op1); | |||
2206 | auto it = PairMap[Idx].find({Op0, Op1}); | |||
2207 | if (it != PairMap[Idx].end()) | |||
2208 | Score += it->second; | |||
2209 | ||||
2210 | unsigned MaxRank = std::max(Ops[i].Rank, Ops[j].Rank); | |||
2211 | if (Score > Max || (Score == Max && MaxRank < BestRank)) { | |||
2212 | BestPair = {i, j}; | |||
2213 | Max = Score; | |||
2214 | BestRank = MaxRank; | |||
2215 | } | |||
2216 | } | |||
2217 | if (Max > 1) { | |||
2218 | auto Op0 = Ops[BestPair.first]; | |||
2219 | auto Op1 = Ops[BestPair.second]; | |||
2220 | Ops.erase(&Ops[BestPair.second]); | |||
2221 | Ops.erase(&Ops[BestPair.first]); | |||
2222 | Ops.push_back(Op0); | |||
2223 | Ops.push_back(Op1); | |||
2224 | } | |||
2225 | } | |||
2226 | // Now that we ordered and optimized the expressions, splat them back into | |||
2227 | // the expression tree, removing any unneeded nodes. | |||
2228 | RewriteExprTree(I, Ops); | |||
2229 | } | |||
2230 | ||||
2231 | void | |||
2232 | ReassociatePass::BuildPairMap(ReversePostOrderTraversal<Function *> &RPOT) { | |||
2233 | // Make a "pairmap" of how often each operand pair occurs. | |||
2234 | for (BasicBlock *BI : RPOT) { | |||
2235 | for (Instruction &I : *BI) { | |||
2236 | if (!I.isAssociative()) | |||
2237 | continue; | |||
2238 | ||||
2239 | // Ignore nodes that aren't at the root of trees. | |||
2240 | if (I.hasOneUse() && I.user_back()->getOpcode() == I.getOpcode()) | |||
2241 | continue; | |||
2242 | ||||
2243 | // Collect all operands in a single reassociable expression. | |||
2244 | // Since Reassociate has already been run once, we can assume things | |||
2245 | // are already canonical according to Reassociation's regime. | |||
2246 | SmallVector<Value *, 8> Worklist = { I.getOperand(0), I.getOperand(1) }; | |||
2247 | SmallVector<Value *, 8> Ops; | |||
2248 | while (!Worklist.empty() && Ops.size() <= GlobalReassociateLimit) { | |||
2249 | Value *Op = Worklist.pop_back_val(); | |||
2250 | Instruction *OpI = dyn_cast<Instruction>(Op); | |||
2251 | if (!OpI || OpI->getOpcode() != I.getOpcode() || !OpI->hasOneUse()) { | |||
2252 | Ops.push_back(Op); | |||
2253 | continue; | |||
2254 | } | |||
2255 | // Be paranoid about self-referencing expressions in unreachable code. | |||
2256 | if (OpI->getOperand(0) != OpI) | |||
2257 | Worklist.push_back(OpI->getOperand(0)); | |||
2258 | if (OpI->getOperand(1) != OpI) | |||
2259 | Worklist.push_back(OpI->getOperand(1)); | |||
2260 | } | |||
2261 | // Skip extremely long expressions. | |||
2262 | if (Ops.size() > GlobalReassociateLimit) | |||
2263 | continue; | |||
2264 | ||||
2265 | // Add all pairwise combinations of operands to the pair map. | |||
2266 | unsigned BinaryIdx = I.getOpcode() - Instruction::BinaryOpsBegin; | |||
2267 | SmallSet<std::pair<Value *, Value*>, 32> Visited; | |||
2268 | for (unsigned i = 0; i < Ops.size() - 1; ++i) { | |||
2269 | for (unsigned j = i + 1; j < Ops.size(); ++j) { | |||
2270 | // Canonicalize operand orderings. | |||
2271 | Value *Op0 = Ops[i]; | |||
2272 | Value *Op1 = Ops[j]; | |||
2273 | if (std::less<Value *>()(Op1, Op0)) | |||
2274 | std::swap(Op0, Op1); | |||
2275 | if (!Visited.insert({Op0, Op1}).second) | |||
2276 | continue; | |||
2277 | auto res = PairMap[BinaryIdx].insert({{Op0, Op1}, 1}); | |||
2278 | if (!res.second) | |||
2279 | ++res.first->second; | |||
2280 | } | |||
2281 | } | |||
2282 | } | |||
2283 | } | |||
2284 | } | |||
2285 | ||||
2286 | PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) { | |||
2287 | // Get the functions basic blocks in Reverse Post Order. This order is used by | |||
2288 | // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic | |||
2289 | // blocks (it has been seen that the analysis in this pass could hang when | |||
2290 | // analysing dead basic blocks). | |||
2291 | ReversePostOrderTraversal<Function *> RPOT(&F); | |||
2292 | ||||
2293 | // Calculate the rank map for F. | |||
2294 | BuildRankMap(F, RPOT); | |||
2295 | ||||
2296 | // Build the pair map before running reassociate. | |||
2297 | // Technically this would be more accurate if we did it after one round | |||
2298 | // of reassociation, but in practice it doesn't seem to help much on | |||
2299 | // real-world code, so don't waste the compile time running reassociate | |||
2300 | // twice. | |||
2301 | // If a user wants, they could expicitly run reassociate twice in their | |||
2302 | // pass pipeline for further potential gains. | |||
2303 | // It might also be possible to update the pair map during runtime, but the | |||
2304 | // overhead of that may be large if there's many reassociable chains. | |||
2305 | BuildPairMap(RPOT); | |||
2306 | ||||
2307 | MadeChange = false; | |||
2308 | ||||
2309 | // Traverse the same blocks that were analysed by BuildRankMap. | |||
2310 | for (BasicBlock *BI : RPOT) { | |||
2311 | assert(RankMap.count(&*BI) && "BB should be ranked.")(static_cast <bool> (RankMap.count(&*BI) && "BB should be ranked.") ? void (0) : __assert_fail ("RankMap.count(&*BI) && \"BB should be ranked.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 2311, __extension__ __PRETTY_FUNCTION__)); | |||
2312 | // Optimize every instruction in the basic block. | |||
2313 | for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;) | |||
2314 | if (isInstructionTriviallyDead(&*II)) { | |||
2315 | EraseInst(&*II++); | |||
2316 | } else { | |||
2317 | OptimizeInst(&*II); | |||
2318 | assert(II->getParent() == &*BI && "Moved to a different block!")(static_cast <bool> (II->getParent() == &*BI && "Moved to a different block!") ? void (0) : __assert_fail ("II->getParent() == &*BI && \"Moved to a different block!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/Reassociate.cpp" , 2318, __extension__ __PRETTY_FUNCTION__)); | |||
2319 | ++II; | |||
2320 | } | |||
2321 | ||||
2322 | // Make a copy of all the instructions to be redone so we can remove dead | |||
2323 | // instructions. | |||
2324 | SetVector<AssertingVH<Instruction>> ToRedo(RedoInsts); | |||
2325 | // Iterate over all instructions to be reevaluated and remove trivially dead | |||
2326 | // instructions. If any operand of the trivially dead instruction becomes | |||
2327 | // dead mark it for deletion as well. Continue this process until all | |||
2328 | // trivially dead instructions have been removed. | |||
2329 | while (!ToRedo.empty()) { | |||
2330 | Instruction *I = ToRedo.pop_back_val(); | |||
2331 | if (isInstructionTriviallyDead(I)) { | |||
2332 | RecursivelyEraseDeadInsts(I, ToRedo); | |||
2333 | MadeChange = true; | |||
2334 | } | |||
2335 | } | |||
2336 | ||||
2337 | // Now that we have removed dead instructions, we can reoptimize the | |||
2338 | // remaining instructions. | |||
2339 | while (!RedoInsts.empty()) { | |||
2340 | Instruction *I = RedoInsts.pop_back_val(); | |||
2341 | if (isInstructionTriviallyDead(I)) | |||
2342 | EraseInst(I); | |||
2343 | else | |||
2344 | OptimizeInst(I); | |||
2345 | } | |||
2346 | } | |||
2347 | ||||
2348 | // We are done with the rank map and pair map. | |||
2349 | RankMap.clear(); | |||
2350 | ValueRankMap.clear(); | |||
2351 | for (auto &Entry : PairMap) | |||
2352 | Entry.clear(); | |||
2353 | ||||
2354 | if (MadeChange) { | |||
2355 | PreservedAnalyses PA; | |||
2356 | PA.preserveSet<CFGAnalyses>(); | |||
2357 | PA.preserve<GlobalsAA>(); | |||
2358 | return PA; | |||
2359 | } | |||
2360 | ||||
2361 | return PreservedAnalyses::all(); | |||
2362 | } | |||
2363 | ||||
2364 | namespace { | |||
2365 | ||||
2366 | class ReassociateLegacyPass : public FunctionPass { | |||
2367 | ReassociatePass Impl; | |||
2368 | ||||
2369 | public: | |||
2370 | static char ID; // Pass identification, replacement for typeid | |||
2371 | ||||
2372 | ReassociateLegacyPass() : FunctionPass(ID) { | |||
2373 | initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry()); | |||
2374 | } | |||
2375 | ||||
2376 | bool runOnFunction(Function &F) override { | |||
2377 | if (skipFunction(F)) | |||
2378 | return false; | |||
2379 | ||||
2380 | FunctionAnalysisManager DummyFAM; | |||
2381 | auto PA = Impl.run(F, DummyFAM); | |||
2382 | return !PA.areAllPreserved(); | |||
2383 | } | |||
2384 | ||||
2385 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
2386 | AU.setPreservesCFG(); | |||
2387 | AU.addPreserved<GlobalsAAWrapperPass>(); | |||
2388 | } | |||
2389 | }; | |||
2390 | ||||
2391 | } // end anonymous namespace | |||
2392 | ||||
2393 | char ReassociateLegacyPass::ID = 0; | |||
2394 | ||||
2395 | INITIALIZE_PASS(ReassociateLegacyPass, "reassociate",static void *initializeReassociateLegacyPassPassOnce(PassRegistry &Registry) { PassInfo *PI = new PassInfo( "Reassociate expressions" , "reassociate", &ReassociateLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<ReassociateLegacyPass>), false, false) ; Registry.registerPass(*PI, true); return PI; } static llvm:: once_flag InitializeReassociateLegacyPassPassFlag; void llvm:: initializeReassociateLegacyPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeReassociateLegacyPassPassFlag, initializeReassociateLegacyPassPassOnce , std::ref(Registry)); } | |||
2396 | "Reassociate expressions", false, false)static void *initializeReassociateLegacyPassPassOnce(PassRegistry &Registry) { PassInfo *PI = new PassInfo( "Reassociate expressions" , "reassociate", &ReassociateLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<ReassociateLegacyPass>), false, false) ; Registry.registerPass(*PI, true); return PI; } static llvm:: once_flag InitializeReassociateLegacyPassPassFlag; void llvm:: initializeReassociateLegacyPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeReassociateLegacyPassPassFlag, initializeReassociateLegacyPassPassOnce , std::ref(Registry)); } | |||
2397 | ||||
2398 | // Public interface to the Reassociate pass | |||
2399 | FunctionPass *llvm::createReassociatePass() { | |||
2400 | return new ReassociateLegacyPass(); | |||
2401 | } |