| File: | build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/lib/Transforms/Scalar/Reassociate.cpp |
| Warning: | line 169, column 8 Dereference of null pointer (loaded from variable '__begin1') |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- Reassociate.cpp - Reassociate binary expressions -------------------===// | |||
| 2 | // | |||
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
| 4 | // See https://llvm.org/LICENSE.txt for license information. | |||
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
| 6 | // | |||
| 7 | //===----------------------------------------------------------------------===// | |||
| 8 | // | |||
| 9 | // This pass reassociates commutative expressions in an order that is designed | |||
| 10 | // to promote better constant propagation, GCSE, LICM, PRE, etc. | |||
| 11 | // | |||
| 12 | // For example: 4 + (x + 5) -> x + (4 + 5) | |||
| 13 | // | |||
| 14 | // In the implementation of this algorithm, constants are assigned rank = 0, | |||
| 15 | // function arguments are rank = 1, and other values are assigned ranks | |||
| 16 | // corresponding to the reverse post order traversal of current function | |||
| 17 | // (starting at 2), which effectively gives values in deep loops higher rank | |||
| 18 | // than values not in loops. | |||
| 19 | // | |||
| 20 | //===----------------------------------------------------------------------===// | |||
| 21 | ||||
| 22 | #include "llvm/Transforms/Scalar/Reassociate.h" | |||
| 23 | #include "llvm/ADT/APFloat.h" | |||
| 24 | #include "llvm/ADT/APInt.h" | |||
| 25 | #include "llvm/ADT/DenseMap.h" | |||
| 26 | #include "llvm/ADT/PostOrderIterator.h" | |||
| 27 | #include "llvm/ADT/SmallPtrSet.h" | |||
| 28 | #include "llvm/ADT/SmallSet.h" | |||
| 29 | #include "llvm/ADT/SmallVector.h" | |||
| 30 | #include "llvm/ADT/Statistic.h" | |||
| 31 | #include "llvm/Analysis/BasicAliasAnalysis.h" | |||
| 32 | #include "llvm/Analysis/GlobalsModRef.h" | |||
| 33 | #include "llvm/Analysis/ValueTracking.h" | |||
| 34 | #include "llvm/IR/Argument.h" | |||
| 35 | #include "llvm/IR/BasicBlock.h" | |||
| 36 | #include "llvm/IR/CFG.h" | |||
| 37 | #include "llvm/IR/Constant.h" | |||
| 38 | #include "llvm/IR/Constants.h" | |||
| 39 | #include "llvm/IR/Function.h" | |||
| 40 | #include "llvm/IR/IRBuilder.h" | |||
| 41 | #include "llvm/IR/InstrTypes.h" | |||
| 42 | #include "llvm/IR/Instruction.h" | |||
| 43 | #include "llvm/IR/Instructions.h" | |||
| 44 | #include "llvm/IR/Operator.h" | |||
| 45 | #include "llvm/IR/PassManager.h" | |||
| 46 | #include "llvm/IR/PatternMatch.h" | |||
| 47 | #include "llvm/IR/Type.h" | |||
| 48 | #include "llvm/IR/User.h" | |||
| 49 | #include "llvm/IR/Value.h" | |||
| 50 | #include "llvm/IR/ValueHandle.h" | |||
| 51 | #include "llvm/InitializePasses.h" | |||
| 52 | #include "llvm/Pass.h" | |||
| 53 | #include "llvm/Support/Casting.h" | |||
| 54 | #include "llvm/Support/Debug.h" | |||
| 55 | #include "llvm/Support/raw_ostream.h" | |||
| 56 | #include "llvm/Transforms/Scalar.h" | |||
| 57 | #include "llvm/Transforms/Utils/Local.h" | |||
| 58 | #include <algorithm> | |||
| 59 | #include <cassert> | |||
| 60 | #include <utility> | |||
| 61 | ||||
| 62 | using namespace llvm; | |||
| 63 | using namespace reassociate; | |||
| 64 | using namespace PatternMatch; | |||
| 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"}; | |||
| 69 | STATISTIC(NumAnnihil, "Number of expr tree annihilated")static llvm::Statistic NumAnnihil = {"reassociate", "NumAnnihil" , "Number of expr tree annihilated"}; | |||
| 70 | STATISTIC(NumFactor , "Number of multiplies factored")static llvm::Statistic NumFactor = {"reassociate", "NumFactor" , "Number of multiplies factored"}; | |||
| 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\"" , "llvm/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, m_APInt(C))) | |||
| 128 | std::swap(V0, V1); | |||
| 129 | ||||
| 130 | if (match(V1, 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::getZero(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 :
| |||
| ||||
| 170 | ValueRankMap[&Arg] = ++Rank; | |||
| 171 | LLVM_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 (mayHaveNonDefUseDependency(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(); i != e && Rank != MaxRank; ++i) | |||
| 204 | Rank = std::max(Rank, getRank(I->getOperand(i))); | |||
| 205 | ||||
| 206 | // If this is a 'not' or 'neg' instruction, do not count it for rank. This | |||
| 207 | // assures us that X and ~X will have the same rank. | |||
| 208 | if (!match(I, m_Not(m_Value())) && !match(I, m_Neg(m_Value())) && | |||
| 209 | !match(I, m_FNeg(m_Value()))) | |||
| 210 | ++Rank; | |||
| 211 | ||||
| 212 | LLVM_DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rankdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n"; } } while (false) | |||
| 213 | << "\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.\"" , "llvm/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.\"" , "llvm/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 Instruction *CreateNeg(Value *S1, const Twine &Name, | |||
| 256 | Instruction *InsertBefore, Value *FlagsOp) { | |||
| 257 | if (S1->getType()->isIntOrIntVectorTy()) | |||
| 258 | return BinaryOperator::CreateNeg(S1, Name, InsertBefore); | |||
| 259 | ||||
| 260 | if (auto *FMFSource = dyn_cast<Instruction>(FlagsOp)) | |||
| 261 | return UnaryOperator::CreateFNegFMF(S1, FMFSource, Name, InsertBefore); | |||
| 262 | ||||
| 263 | return UnaryOperator::CreateFNeg(S1, Name, InsertBefore); | |||
| 264 | } | |||
| 265 | ||||
| 266 | /// Replace 0-X with X*-1. | |||
| 267 | static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) { | |||
| 268 | assert((isa<UnaryOperator>(Neg) || isa<BinaryOperator>(Neg)) &&(static_cast <bool> ((isa<UnaryOperator>(Neg) || isa <BinaryOperator>(Neg)) && "Expected a Negate!") ? void (0) : __assert_fail ("(isa<UnaryOperator>(Neg) || isa<BinaryOperator>(Neg)) && \"Expected a Negate!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 269, __extension__ __PRETTY_FUNCTION__)) | |||
| 269 | "Expected a Negate!")(static_cast <bool> ((isa<UnaryOperator>(Neg) || isa <BinaryOperator>(Neg)) && "Expected a Negate!") ? void (0) : __assert_fail ("(isa<UnaryOperator>(Neg) || isa<BinaryOperator>(Neg)) && \"Expected a Negate!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 269, __extension__ __PRETTY_FUNCTION__)); | |||
| 270 | // FIXME: It's not safe to lower a unary FNeg into a FMul by -1.0. | |||
| 271 | unsigned OpNo = isa<BinaryOperator>(Neg) ? 1 : 0; | |||
| 272 | Type *Ty = Neg->getType(); | |||
| 273 | Constant *NegOne = Ty->isIntOrIntVectorTy() ? | |||
| 274 | ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0); | |||
| 275 | ||||
| 276 | BinaryOperator *Res = CreateMul(Neg->getOperand(OpNo), NegOne, "", Neg, Neg); | |||
| 277 | Neg->setOperand(OpNo, Constant::getNullValue(Ty)); // Drop use of op. | |||
| 278 | Res->takeName(Neg); | |||
| 279 | Neg->replaceAllUsesWith(Res); | |||
| 280 | Res->setDebugLoc(Neg->getDebugLoc()); | |||
| 281 | return Res; | |||
| 282 | } | |||
| 283 | ||||
| 284 | /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael | |||
| 285 | /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for | |||
| 286 | /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic. | |||
| 287 | /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every | |||
| 288 | /// even x in Bitwidth-bit arithmetic. | |||
| 289 | static unsigned CarmichaelShift(unsigned Bitwidth) { | |||
| 290 | if (Bitwidth < 3) | |||
| 291 | return Bitwidth - 1; | |||
| 292 | return Bitwidth - 2; | |||
| 293 | } | |||
| 294 | ||||
| 295 | /// Add the extra weight 'RHS' to the existing weight 'LHS', | |||
| 296 | /// reducing the combined weight using any special properties of the operation. | |||
| 297 | /// The existing weight LHS represents the computation X op X op ... op X where | |||
| 298 | /// X occurs LHS times. The combined weight represents X op X op ... op X with | |||
| 299 | /// X occurring LHS + RHS times. If op is "Xor" for example then the combined | |||
| 300 | /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even; | |||
| 301 | /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second. | |||
| 302 | static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) { | |||
| 303 | // If we were working with infinite precision arithmetic then the combined | |||
| 304 | // weight would be LHS + RHS. But we are using finite precision arithmetic, | |||
| 305 | // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct | |||
| 306 | // for nilpotent operations and addition, but not for idempotent operations | |||
| 307 | // and multiplication), so it is important to correctly reduce the combined | |||
| 308 | // weight back into range if wrapping would be wrong. | |||
| 309 | ||||
| 310 | // If RHS is zero then the weight didn't change. | |||
| 311 | if (RHS.isMinValue()) | |||
| 312 | return; | |||
| 313 | // If LHS is zero then the combined weight is RHS. | |||
| 314 | if (LHS.isMinValue()) { | |||
| 315 | LHS = RHS; | |||
| 316 | return; | |||
| 317 | } | |||
| 318 | // From this point on we know that neither LHS nor RHS is zero. | |||
| 319 | ||||
| 320 | if (Instruction::isIdempotent(Opcode)) { | |||
| 321 | // Idempotent means X op X === X, so any non-zero weight is equivalent to a | |||
| 322 | // weight of 1. Keeping weights at zero or one also means that wrapping is | |||
| 323 | // not a problem. | |||
| 324 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 324, __extension__ __PRETTY_FUNCTION__)); | |||
| 325 | return; // Return a weight of 1. | |||
| 326 | } | |||
| 327 | if (Instruction::isNilpotent(Opcode)) { | |||
| 328 | // Nilpotent means X op X === 0, so reduce weights modulo 2. | |||
| 329 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 329, __extension__ __PRETTY_FUNCTION__)); | |||
| 330 | LHS = 0; // 1 + 1 === 0 modulo 2. | |||
| 331 | return; | |||
| 332 | } | |||
| 333 | if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) { | |||
| 334 | // TODO: Reduce the weight by exploiting nsw/nuw? | |||
| 335 | LHS += RHS; | |||
| 336 | return; | |||
| 337 | } | |||
| 338 | ||||
| 339 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 340, __extension__ __PRETTY_FUNCTION__)) | |||
| 340 | "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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 340, __extension__ __PRETTY_FUNCTION__)); | |||
| 341 | unsigned Bitwidth = LHS.getBitWidth(); | |||
| 342 | // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth | |||
| 343 | // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth | |||
| 344 | // bit number x, since either x is odd in which case x^CM = 1, or x is even in | |||
| 345 | // which case both x^W and x^(W - CM) are zero. By subtracting off multiples | |||
| 346 | // of CM like this weights can always be reduced to the range [0, CM+Bitwidth) | |||
| 347 | // which by a happy accident means that they can always be represented using | |||
| 348 | // Bitwidth bits. | |||
| 349 | // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than | |||
| 350 | // the Carmichael number). | |||
| 351 | if (Bitwidth > 3) { | |||
| 352 | /// CM - The value of Carmichael's lambda function. | |||
| 353 | APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth)); | |||
| 354 | // Any weight W >= Threshold can be replaced with W - CM. | |||
| 355 | APInt Threshold = CM + Bitwidth; | |||
| 356 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 356, __extension__ __PRETTY_FUNCTION__)); | |||
| 357 | // For Bitwidth 4 or more the following sum does not overflow. | |||
| 358 | LHS += RHS; | |||
| 359 | while (LHS.uge(Threshold)) | |||
| 360 | LHS -= CM; | |||
| 361 | } else { | |||
| 362 | // To avoid problems with overflow do everything the same as above but using | |||
| 363 | // a larger type. | |||
| 364 | unsigned CM = 1U << CarmichaelShift(Bitwidth); | |||
| 365 | unsigned Threshold = CM + Bitwidth; | |||
| 366 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 367, __extension__ __PRETTY_FUNCTION__)) | |||
| 367 | "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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 367, __extension__ __PRETTY_FUNCTION__)); | |||
| 368 | unsigned Total = LHS.getZExtValue() + RHS.getZExtValue(); | |||
| 369 | while (Total >= Threshold) | |||
| 370 | Total -= CM; | |||
| 371 | LHS = Total; | |||
| 372 | } | |||
| 373 | } | |||
| 374 | ||||
| 375 | using RepeatedValue = std::pair<Value*, APInt>; | |||
| 376 | ||||
| 377 | /// Given an associative binary expression, return the leaf | |||
| 378 | /// nodes in Ops along with their weights (how many times the leaf occurs). The | |||
| 379 | /// original expression is the same as | |||
| 380 | /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times | |||
| 381 | /// op | |||
| 382 | /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times | |||
| 383 | /// op | |||
| 384 | /// ... | |||
| 385 | /// op | |||
| 386 | /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times | |||
| 387 | /// | |||
| 388 | /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct. | |||
| 389 | /// | |||
| 390 | /// This routine may modify the function, in which case it returns 'true'. The | |||
| 391 | /// changes it makes may well be destructive, changing the value computed by 'I' | |||
| 392 | /// to something completely different. Thus if the routine returns 'true' then | |||
| 393 | /// you MUST either replace I with a new expression computed from the Ops array, | |||
| 394 | /// or use RewriteExprTree to put the values back in. | |||
| 395 | /// | |||
| 396 | /// A leaf node is either not a binary operation of the same kind as the root | |||
| 397 | /// node 'I' (i.e. is not a binary operator at all, or is, but with a different | |||
| 398 | /// opcode), or is the same kind of binary operator but has a use which either | |||
| 399 | /// does not belong to the expression, or does belong to the expression but is | |||
| 400 | /// a leaf node. Every leaf node has at least one use that is a non-leaf node | |||
| 401 | /// of the expression, while for non-leaf nodes (except for the root 'I') every | |||
| 402 | /// use is a non-leaf node of the expression. | |||
| 403 | /// | |||
| 404 | /// For example: | |||
| 405 | /// expression graph node names | |||
| 406 | /// | |||
| 407 | /// + | I | |||
| 408 | /// / \ | | |||
| 409 | /// + + | A, B | |||
| 410 | /// / \ / \ | | |||
| 411 | /// * + * | C, D, E | |||
| 412 | /// / \ / \ / \ | | |||
| 413 | /// + * | F, G | |||
| 414 | /// | |||
| 415 | /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in | |||
| 416 | /// that order) (C, 1), (E, 1), (F, 2), (G, 2). | |||
| 417 | /// | |||
| 418 | /// The expression is maximal: if some instruction is a binary operator of the | |||
| 419 | /// same kind as 'I', and all of its uses are non-leaf nodes of the expression, | |||
| 420 | /// then the instruction also belongs to the expression, is not a leaf node of | |||
| 421 | /// it, and its operands also belong to the expression (but may be leaf nodes). | |||
| 422 | /// | |||
| 423 | /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in | |||
| 424 | /// order to ensure that every non-root node in the expression has *exactly one* | |||
| 425 | /// use by a non-leaf node of the expression. This destruction means that the | |||
| 426 | /// caller MUST either replace 'I' with a new expression or use something like | |||
| 427 | /// RewriteExprTree to put the values back in if the routine indicates that it | |||
| 428 | /// made a change by returning 'true'. | |||
| 429 | /// | |||
| 430 | /// In the above example either the right operand of A or the left operand of B | |||
| 431 | /// will be replaced by undef. If it is B's operand then this gives: | |||
| 432 | /// | |||
| 433 | /// + | I | |||
| 434 | /// / \ | | |||
| 435 | /// + + | A, B - operand of B replaced with undef | |||
| 436 | /// / \ \ | | |||
| 437 | /// * + * | C, D, E | |||
| 438 | /// / \ / \ / \ | | |||
| 439 | /// + * | F, G | |||
| 440 | /// | |||
| 441 | /// Note that such undef operands can only be reached by passing through 'I'. | |||
| 442 | /// For example, if you visit operands recursively starting from a leaf node | |||
| 443 | /// then you will never see such an undef operand unless you get back to 'I', | |||
| 444 | /// which requires passing through a phi node. | |||
| 445 | /// | |||
| 446 | /// Note that this routine may also mutate binary operators of the wrong type | |||
| 447 | /// that have all uses inside the expression (i.e. only used by non-leaf nodes | |||
| 448 | /// of the expression) if it can turn them into binary operators of the right | |||
| 449 | /// type and thus make the expression bigger. | |||
| 450 | static bool LinearizeExprTree(Instruction *I, | |||
| 451 | SmallVectorImpl<RepeatedValue> &Ops) { | |||
| 452 | assert((isa<UnaryOperator>(I) || isa<BinaryOperator>(I)) &&(static_cast <bool> ((isa<UnaryOperator>(I) || isa <BinaryOperator>(I)) && "Expected a UnaryOperator or BinaryOperator!" ) ? void (0) : __assert_fail ("(isa<UnaryOperator>(I) || isa<BinaryOperator>(I)) && \"Expected a UnaryOperator or BinaryOperator!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 453, __extension__ __PRETTY_FUNCTION__)) | |||
| 453 | "Expected a UnaryOperator or BinaryOperator!")(static_cast <bool> ((isa<UnaryOperator>(I) || isa <BinaryOperator>(I)) && "Expected a UnaryOperator or BinaryOperator!" ) ? void (0) : __assert_fail ("(isa<UnaryOperator>(I) || isa<BinaryOperator>(I)) && \"Expected a UnaryOperator or BinaryOperator!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 453, __extension__ __PRETTY_FUNCTION__)); | |||
| 454 | LLVM_DEBUG(dbgs() << "LINEARIZE: " << *I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "LINEARIZE: " << *I << '\n'; } } while (false); | |||
| 455 | unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits(); | |||
| 456 | unsigned Opcode = I->getOpcode(); | |||
| 457 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 458, __extension__ __PRETTY_FUNCTION__)) | |||
| 458 | "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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 458, __extension__ __PRETTY_FUNCTION__)); | |||
| 459 | ||||
| 460 | // Visit all operands of the expression, keeping track of their weight (the | |||
| 461 | // number of paths from the expression root to the operand, or if you like | |||
| 462 | // the number of times that operand occurs in the linearized expression). | |||
| 463 | // For example, if I = X + A, where X = A + B, then I, X and B have weight 1 | |||
| 464 | // while A has weight two. | |||
| 465 | ||||
| 466 | // Worklist of non-leaf nodes (their operands are in the expression too) along | |||
| 467 | // with their weights, representing a certain number of paths to the operator. | |||
| 468 | // If an operator occurs in the worklist multiple times then we found multiple | |||
| 469 | // ways to get to it. | |||
| 470 | SmallVector<std::pair<Instruction*, APInt>, 8> Worklist; // (Op, Weight) | |||
| 471 | Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1))); | |||
| 472 | bool Changed = false; | |||
| 473 | ||||
| 474 | // Leaves of the expression are values that either aren't the right kind of | |||
| 475 | // operation (eg: a constant, or a multiply in an add tree), or are, but have | |||
| 476 | // some uses that are not inside the expression. For example, in I = X + X, | |||
| 477 | // X = A + B, the value X has two uses (by I) that are in the expression. If | |||
| 478 | // X has any other uses, for example in a return instruction, then we consider | |||
| 479 | // X to be a leaf, and won't analyze it further. When we first visit a value, | |||
| 480 | // if it has more than one use then at first we conservatively consider it to | |||
| 481 | // be a leaf. Later, as the expression is explored, we may discover some more | |||
| 482 | // uses of the value from inside the expression. If all uses turn out to be | |||
| 483 | // from within the expression (and the value is a binary operator of the right | |||
| 484 | // kind) then the value is no longer considered to be a leaf, and its operands | |||
| 485 | // are explored. | |||
| 486 | ||||
| 487 | // Leaves - Keeps track of the set of putative leaves as well as the number of | |||
| 488 | // paths to each leaf seen so far. | |||
| 489 | using LeafMap = DenseMap<Value *, APInt>; | |||
| 490 | LeafMap Leaves; // Leaf -> Total weight so far. | |||
| 491 | SmallVector<Value *, 8> LeafOrder; // Ensure deterministic leaf output order. | |||
| 492 | ||||
| 493 | #ifndef NDEBUG | |||
| 494 | SmallPtrSet<Value *, 8> Visited; // For checking the iteration scheme. | |||
| 495 | #endif | |||
| 496 | while (!Worklist.empty()) { | |||
| 497 | std::pair<Instruction*, APInt> P = Worklist.pop_back_val(); | |||
| 498 | I = P.first; // We examine the operands of this binary operator. | |||
| 499 | ||||
| 500 | for (unsigned OpIdx = 0; OpIdx < I->getNumOperands(); ++OpIdx) { // Visit operands. | |||
| 501 | Value *Op = I->getOperand(OpIdx); | |||
| 502 | APInt Weight = P.second; // Number of paths to this operand. | |||
| 503 | LLVM_DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n"; } } while (false); | |||
| 504 | 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?!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 504, __extension__ __PRETTY_FUNCTION__)); | |||
| 505 | ||||
| 506 | // If this is a binary operation of the right kind with only one use then | |||
| 507 | // add its operands to the expression. | |||
| 508 | if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { | |||
| 509 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 509, __extension__ __PRETTY_FUNCTION__)); | |||
| 510 | LLVM_DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n"; } } while (false ); | |||
| 511 | Worklist.push_back(std::make_pair(BO, Weight)); | |||
| 512 | continue; | |||
| 513 | } | |||
| 514 | ||||
| 515 | // Appears to be a leaf. Is the operand already in the set of leaves? | |||
| 516 | LeafMap::iterator It = Leaves.find(Op); | |||
| 517 | if (It == Leaves.end()) { | |||
| 518 | // Not in the leaf map. Must be the first time we saw this operand. | |||
| 519 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 519, __extension__ __PRETTY_FUNCTION__)); | |||
| 520 | if (!Op->hasOneUse()) { | |||
| 521 | // This value has uses not accounted for by the expression, so it is | |||
| 522 | // not safe to modify. Mark it as being a leaf. | |||
| 523 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n"; } } while ( false) | |||
| 524 | << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n"; } } while ( false); | |||
| 525 | LeafOrder.push_back(Op); | |||
| 526 | Leaves[Op] = Weight; | |||
| 527 | continue; | |||
| 528 | } | |||
| 529 | // No uses outside the expression, try morphing it. | |||
| 530 | } else { | |||
| 531 | // Already in the leaf map. | |||
| 532 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 533, __extension__ __PRETTY_FUNCTION__)) | |||
| 533 | "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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 533, __extension__ __PRETTY_FUNCTION__)); | |||
| 534 | ||||
| 535 | // Update the number of paths to the leaf. | |||
| 536 | IncorporateWeight(It->second, Weight, Opcode); | |||
| 537 | ||||
| 538 | #if 0 // TODO: Re-enable once PR13021 is fixed. | |||
| 539 | // The leaf already has one use from inside the expression. As we want | |||
| 540 | // exactly one such use, drop this new use of the leaf. | |||
| 541 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 541, __extension__ __PRETTY_FUNCTION__)); | |||
| 542 | I->setOperand(OpIdx, UndefValue::get(I->getType())); | |||
| 543 | Changed = true; | |||
| 544 | ||||
| 545 | // If the leaf is a binary operation of the right kind and we now see | |||
| 546 | // that its multiple original uses were in fact all by nodes belonging | |||
| 547 | // to the expression, then no longer consider it to be a leaf and add | |||
| 548 | // its operands to the expression. | |||
| 549 | if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { | |||
| 550 | LLVM_DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n"; } } while (false ); | |||
| 551 | Worklist.push_back(std::make_pair(BO, It->second)); | |||
| 552 | Leaves.erase(It); | |||
| 553 | continue; | |||
| 554 | } | |||
| 555 | #endif | |||
| 556 | ||||
| 557 | // If we still have uses that are not accounted for by the expression | |||
| 558 | // then it is not safe to modify the value. | |||
| 559 | if (!Op->hasOneUse()) | |||
| 560 | continue; | |||
| 561 | ||||
| 562 | // No uses outside the expression, try morphing it. | |||
| 563 | Weight = It->second; | |||
| 564 | Leaves.erase(It); // Since the value may be morphed below. | |||
| 565 | } | |||
| 566 | ||||
| 567 | // At this point we have a value which, first of all, is not a binary | |||
| 568 | // expression of the right kind, and secondly, is only used inside the | |||
| 569 | // expression. This means that it can safely be modified. See if we | |||
| 570 | // can usefully morph it into an expression of the right kind. | |||
| 571 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 575, __extension__ __PRETTY_FUNCTION__)) | |||
| 572 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 575, __extension__ __PRETTY_FUNCTION__)) | |||
| 573 | || (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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 575, __extension__ __PRETTY_FUNCTION__)) | |||
| 574 | !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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 575, __extension__ __PRETTY_FUNCTION__)) | |||
| 575 | "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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 575, __extension__ __PRETTY_FUNCTION__)); | |||
| 576 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 576, __extension__ __PRETTY_FUNCTION__)); | |||
| 577 | ||||
| 578 | // If this is a multiply expression, turn any internal negations into | |||
| 579 | // multiplies by -1 so they can be reassociated. | |||
| 580 | if (Instruction *Tmp = dyn_cast<Instruction>(Op)) | |||
| 581 | if ((Opcode == Instruction::Mul && match(Tmp, m_Neg(m_Value()))) || | |||
| 582 | (Opcode == Instruction::FMul && match(Tmp, m_FNeg(m_Value())))) { | |||
| 583 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO "; } } while (false ) | |||
| 584 | << "MORPH LEAF: " << *Op << " (" << Weight << ") TO ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO "; } } while (false ); | |||
| 585 | Tmp = LowerNegateToMultiply(Tmp); | |||
| 586 | LLVM_DEBUG(dbgs() << *Tmp << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << *Tmp << '\n'; } } while (false); | |||
| 587 | Worklist.push_back(std::make_pair(Tmp, Weight)); | |||
| 588 | Changed = true; | |||
| 589 | continue; | |||
| 590 | } | |||
| 591 | ||||
| 592 | // Failed to morph into an expression of the right type. This really is | |||
| 593 | // a leaf. | |||
| 594 | LLVM_DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n"; } } while (false); | |||
| 595 | 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?\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 595, __extension__ __PRETTY_FUNCTION__)); | |||
| 596 | LeafOrder.push_back(Op); | |||
| 597 | Leaves[Op] = Weight; | |||
| 598 | } | |||
| 599 | } | |||
| 600 | ||||
| 601 | // The leaves, repeated according to their weights, represent the linearized | |||
| 602 | // form of the expression. | |||
| 603 | for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) { | |||
| 604 | Value *V = LeafOrder[i]; | |||
| 605 | LeafMap::iterator It = Leaves.find(V); | |||
| 606 | if (It == Leaves.end()) | |||
| 607 | // Node initially thought to be a leaf wasn't. | |||
| 608 | continue; | |||
| 609 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 609, __extension__ __PRETTY_FUNCTION__)); | |||
| 610 | APInt Weight = It->second; | |||
| 611 | if (Weight.isMinValue()) | |||
| 612 | // Leaf already output or weight reduction eliminated it. | |||
| 613 | continue; | |||
| 614 | // Ensure the leaf is only output once. | |||
| 615 | It->second = 0; | |||
| 616 | Ops.push_back(std::make_pair(V, Weight)); | |||
| 617 | } | |||
| 618 | ||||
| 619 | // For nilpotent operations or addition there may be no operands, for example | |||
| 620 | // because the expression was "X xor X" or consisted of 2^Bitwidth additions: | |||
| 621 | // in both cases the weight reduces to 0 causing the value to be skipped. | |||
| 622 | if (Ops.empty()) { | |||
| 623 | Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType()); | |||
| 624 | assert(Identity && "Associative operation without identity!")(static_cast <bool> (Identity && "Associative operation without identity!" ) ? void (0) : __assert_fail ("Identity && \"Associative operation without identity!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 624, __extension__ __PRETTY_FUNCTION__)); | |||
| 625 | Ops.emplace_back(Identity, APInt(Bitwidth, 1)); | |||
| 626 | } | |||
| 627 | ||||
| 628 | return Changed; | |||
| 629 | } | |||
| 630 | ||||
| 631 | /// Now that the operands for this expression tree are | |||
| 632 | /// linearized and optimized, emit them in-order. | |||
| 633 | void ReassociatePass::RewriteExprTree(BinaryOperator *I, | |||
| 634 | SmallVectorImpl<ValueEntry> &Ops) { | |||
| 635 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 635, __extension__ __PRETTY_FUNCTION__)); | |||
| 636 | ||||
| 637 | // Since our optimizations should never increase the number of operations, the | |||
| 638 | // new expression can usually be written reusing the existing binary operators | |||
| 639 | // from the original expression tree, without creating any new instructions, | |||
| 640 | // though the rewritten expression may have a completely different topology. | |||
| 641 | // We take care to not change anything if the new expression will be the same | |||
| 642 | // as the original. If more than trivial changes (like commuting operands) | |||
| 643 | // were made then we are obliged to clear out any optional subclass data like | |||
| 644 | // nsw flags. | |||
| 645 | ||||
| 646 | /// NodesToRewrite - Nodes from the original expression available for writing | |||
| 647 | /// the new expression into. | |||
| 648 | SmallVector<BinaryOperator*, 8> NodesToRewrite; | |||
| 649 | unsigned Opcode = I->getOpcode(); | |||
| 650 | BinaryOperator *Op = I; | |||
| 651 | ||||
| 652 | /// NotRewritable - The operands being written will be the leaves of the new | |||
| 653 | /// expression and must not be used as inner nodes (via NodesToRewrite) by | |||
| 654 | /// mistake. Inner nodes are always reassociable, and usually leaves are not | |||
| 655 | /// (if they were they would have been incorporated into the expression and so | |||
| 656 | /// would not be leaves), so most of the time there is no danger of this. But | |||
| 657 | /// in rare cases a leaf may become reassociable if an optimization kills uses | |||
| 658 | /// of it, or it may momentarily become reassociable during rewriting (below) | |||
| 659 | /// due it being removed as an operand of one of its uses. Ensure that misuse | |||
| 660 | /// of leaf nodes as inner nodes cannot occur by remembering all of the future | |||
| 661 | /// leaves and refusing to reuse any of them as inner nodes. | |||
| 662 | SmallPtrSet<Value*, 8> NotRewritable; | |||
| 663 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) | |||
| 664 | NotRewritable.insert(Ops[i].Op); | |||
| 665 | ||||
| 666 | // ExpressionChanged - Non-null if the rewritten expression differs from the | |||
| 667 | // original in some non-trivial way, requiring the clearing of optional flags. | |||
| 668 | // Flags are cleared from the operator in ExpressionChanged up to I inclusive. | |||
| 669 | BinaryOperator *ExpressionChanged = nullptr; | |||
| 670 | for (unsigned i = 0; ; ++i) { | |||
| 671 | // The last operation (which comes earliest in the IR) is special as both | |||
| 672 | // operands will come from Ops, rather than just one with the other being | |||
| 673 | // a subexpression. | |||
| 674 | if (i+2 == Ops.size()) { | |||
| 675 | Value *NewLHS = Ops[i].Op; | |||
| 676 | Value *NewRHS = Ops[i+1].Op; | |||
| 677 | Value *OldLHS = Op->getOperand(0); | |||
| 678 | Value *OldRHS = Op->getOperand(1); | |||
| 679 | ||||
| 680 | if (NewLHS == OldLHS && NewRHS == OldRHS) | |||
| 681 | // Nothing changed, leave it alone. | |||
| 682 | break; | |||
| 683 | ||||
| 684 | if (NewLHS == OldRHS && NewRHS == OldLHS) { | |||
| 685 | // The order of the operands was reversed. Swap them. | |||
| 686 | LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "RA: " << *Op << '\n'; } } while (false); | |||
| 687 | Op->swapOperands(); | |||
| 688 | LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "TO: " << *Op << '\n'; } } while (false); | |||
| 689 | MadeChange = true; | |||
| 690 | ++NumChanged; | |||
| 691 | break; | |||
| 692 | } | |||
| 693 | ||||
| 694 | // The new operation differs non-trivially from the original. Overwrite | |||
| 695 | // the old operands with the new ones. | |||
| 696 | LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "RA: " << *Op << '\n'; } } while (false); | |||
| 697 | if (NewLHS != OldLHS) { | |||
| 698 | BinaryOperator *BO = isReassociableOp(OldLHS, Opcode); | |||
| 699 | if (BO && !NotRewritable.count(BO)) | |||
| 700 | NodesToRewrite.push_back(BO); | |||
| 701 | Op->setOperand(0, NewLHS); | |||
| 702 | } | |||
| 703 | if (NewRHS != OldRHS) { | |||
| 704 | BinaryOperator *BO = isReassociableOp(OldRHS, Opcode); | |||
| 705 | if (BO && !NotRewritable.count(BO)) | |||
| 706 | NodesToRewrite.push_back(BO); | |||
| 707 | Op->setOperand(1, NewRHS); | |||
| 708 | } | |||
| 709 | LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "TO: " << *Op << '\n'; } } while (false); | |||
| 710 | ||||
| 711 | ExpressionChanged = Op; | |||
| 712 | MadeChange = true; | |||
| 713 | ++NumChanged; | |||
| 714 | ||||
| 715 | break; | |||
| 716 | } | |||
| 717 | ||||
| 718 | // Not the last operation. The left-hand side will be a sub-expression | |||
| 719 | // while the right-hand side will be the current element of Ops. | |||
| 720 | Value *NewRHS = Ops[i].Op; | |||
| 721 | if (NewRHS != Op->getOperand(1)) { | |||
| 722 | LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "RA: " << *Op << '\n'; } } while (false); | |||
| 723 | if (NewRHS == Op->getOperand(0)) { | |||
| 724 | // The new right-hand side was already present as the left operand. If | |||
| 725 | // we are lucky then swapping the operands will sort out both of them. | |||
| 726 | Op->swapOperands(); | |||
| 727 | } else { | |||
| 728 | // Overwrite with the new right-hand side. | |||
| 729 | BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode); | |||
| 730 | if (BO && !NotRewritable.count(BO)) | |||
| 731 | NodesToRewrite.push_back(BO); | |||
| 732 | Op->setOperand(1, NewRHS); | |||
| 733 | ExpressionChanged = Op; | |||
| 734 | } | |||
| 735 | LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "TO: " << *Op << '\n'; } } while (false); | |||
| 736 | MadeChange = true; | |||
| 737 | ++NumChanged; | |||
| 738 | } | |||
| 739 | ||||
| 740 | // Now deal with the left-hand side. If this is already an operation node | |||
| 741 | // from the original expression then just rewrite the rest of the expression | |||
| 742 | // into it. | |||
| 743 | BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode); | |||
| 744 | if (BO && !NotRewritable.count(BO)) { | |||
| 745 | Op = BO; | |||
| 746 | continue; | |||
| 747 | } | |||
| 748 | ||||
| 749 | // Otherwise, grab a spare node from the original expression and use that as | |||
| 750 | // the left-hand side. If there are no nodes left then the optimizers made | |||
| 751 | // an expression with more nodes than the original! This usually means that | |||
| 752 | // they did something stupid but it might mean that the problem was just too | |||
| 753 | // hard (finding the mimimal number of multiplications needed to realize a | |||
| 754 | // multiplication expression is NP-complete). Whatever the reason, smart or | |||
| 755 | // stupid, create a new node if there are none left. | |||
| 756 | BinaryOperator *NewOp; | |||
| 757 | if (NodesToRewrite.empty()) { | |||
| 758 | Constant *Undef = UndefValue::get(I->getType()); | |||
| 759 | NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode), | |||
| 760 | Undef, Undef, "", I); | |||
| 761 | if (NewOp->getType()->isFPOrFPVectorTy()) | |||
| 762 | NewOp->setFastMathFlags(I->getFastMathFlags()); | |||
| 763 | } else { | |||
| 764 | NewOp = NodesToRewrite.pop_back_val(); | |||
| 765 | } | |||
| 766 | ||||
| 767 | LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "RA: " << *Op << '\n'; } } while (false); | |||
| 768 | Op->setOperand(0, NewOp); | |||
| 769 | LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "TO: " << *Op << '\n'; } } while (false); | |||
| 770 | ExpressionChanged = Op; | |||
| 771 | MadeChange = true; | |||
| 772 | ++NumChanged; | |||
| 773 | Op = NewOp; | |||
| 774 | } | |||
| 775 | ||||
| 776 | // If the expression changed non-trivially then clear out all subclass data | |||
| 777 | // starting from the operator specified in ExpressionChanged, and compactify | |||
| 778 | // the operators to just before the expression root to guarantee that the | |||
| 779 | // expression tree is dominated by all of Ops. | |||
| 780 | if (ExpressionChanged) | |||
| 781 | do { | |||
| 782 | // Preserve FastMathFlags. | |||
| 783 | if (isa<FPMathOperator>(I)) { | |||
| 784 | FastMathFlags Flags = I->getFastMathFlags(); | |||
| 785 | ExpressionChanged->clearSubclassOptionalData(); | |||
| 786 | ExpressionChanged->setFastMathFlags(Flags); | |||
| 787 | } else | |||
| 788 | ExpressionChanged->clearSubclassOptionalData(); | |||
| 789 | ||||
| 790 | if (ExpressionChanged == I) | |||
| 791 | break; | |||
| 792 | ||||
| 793 | // Discard any debug info related to the expressions that has changed (we | |||
| 794 | // can leave debug infor related to the root, since the result of the | |||
| 795 | // expression tree should be the same even after reassociation). | |||
| 796 | replaceDbgUsesWithUndef(ExpressionChanged); | |||
| 797 | ||||
| 798 | ExpressionChanged->moveBefore(I); | |||
| 799 | ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin()); | |||
| 800 | } while (true); | |||
| 801 | ||||
| 802 | // Throw away any left over nodes from the original expression. | |||
| 803 | for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i) | |||
| 804 | RedoInsts.insert(NodesToRewrite[i]); | |||
| 805 | } | |||
| 806 | ||||
| 807 | /// Insert instructions before the instruction pointed to by BI, | |||
| 808 | /// that computes the negative version of the value specified. The negative | |||
| 809 | /// version of the value is returned, and BI is left pointing at the instruction | |||
| 810 | /// that should be processed next by the reassociation pass. | |||
| 811 | /// Also add intermediate instructions to the redo list that are modified while | |||
| 812 | /// pushing the negates through adds. These will be revisited to see if | |||
| 813 | /// additional opportunities have been exposed. | |||
| 814 | static Value *NegateValue(Value *V, Instruction *BI, | |||
| 815 | ReassociatePass::OrderedSet &ToRedo) { | |||
| 816 | if (auto *C = dyn_cast<Constant>(V)) | |||
| 817 | return C->getType()->isFPOrFPVectorTy() ? ConstantExpr::getFNeg(C) : | |||
| 818 | ConstantExpr::getNeg(C); | |||
| 819 | ||||
| 820 | // We are trying to expose opportunity for reassociation. One of the things | |||
| 821 | // that we want to do to achieve this is to push a negation as deep into an | |||
| 822 | // expression chain as possible, to expose the add instructions. In practice, | |||
| 823 | // this means that we turn this: | |||
| 824 | // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D | |||
| 825 | // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate | |||
| 826 | // the constants. We assume that instcombine will clean up the mess later if | |||
| 827 | // we introduce tons of unnecessary negation instructions. | |||
| 828 | // | |||
| 829 | if (BinaryOperator *I = | |||
| 830 | isReassociableOp(V, Instruction::Add, Instruction::FAdd)) { | |||
| 831 | // Push the negates through the add. | |||
| 832 | I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo)); | |||
| 833 | I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo)); | |||
| 834 | if (I->getOpcode() == Instruction::Add) { | |||
| 835 | I->setHasNoUnsignedWrap(false); | |||
| 836 | I->setHasNoSignedWrap(false); | |||
| 837 | } | |||
| 838 | ||||
| 839 | // We must move the add instruction here, because the neg instructions do | |||
| 840 | // not dominate the old add instruction in general. By moving it, we are | |||
| 841 | // assured that the neg instructions we just inserted dominate the | |||
| 842 | // instruction we are about to insert after them. | |||
| 843 | // | |||
| 844 | I->moveBefore(BI); | |||
| 845 | I->setName(I->getName()+".neg"); | |||
| 846 | ||||
| 847 | // Add the intermediate negates to the redo list as processing them later | |||
| 848 | // could expose more reassociating opportunities. | |||
| 849 | ToRedo.insert(I); | |||
| 850 | return I; | |||
| 851 | } | |||
| 852 | ||||
| 853 | // Okay, we need to materialize a negated version of V with an instruction. | |||
| 854 | // Scan the use lists of V to see if we have one already. | |||
| 855 | for (User *U : V->users()) { | |||
| 856 | if (!match(U, m_Neg(m_Value())) && !match(U, m_FNeg(m_Value()))) | |||
| 857 | continue; | |||
| 858 | ||||
| 859 | // We found one! Now we have to make sure that the definition dominates | |||
| 860 | // this use. We do this by moving it to the entry block (if it is a | |||
| 861 | // non-instruction value) or right after the definition. These negates will | |||
| 862 | // be zapped by reassociate later, so we don't need much finesse here. | |||
| 863 | Instruction *TheNeg = cast<Instruction>(U); | |||
| 864 | ||||
| 865 | // Verify that the negate is in this function, V might be a constant expr. | |||
| 866 | if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) | |||
| 867 | continue; | |||
| 868 | ||||
| 869 | bool FoundCatchSwitch = false; | |||
| 870 | ||||
| 871 | BasicBlock::iterator InsertPt; | |||
| 872 | if (Instruction *InstInput = dyn_cast<Instruction>(V)) { | |||
| 873 | if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { | |||
| 874 | InsertPt = II->getNormalDest()->begin(); | |||
| 875 | } else { | |||
| 876 | InsertPt = ++InstInput->getIterator(); | |||
| 877 | } | |||
| 878 | ||||
| 879 | const BasicBlock *BB = InsertPt->getParent(); | |||
| 880 | ||||
| 881 | // Make sure we don't move anything before PHIs or exception | |||
| 882 | // handling pads. | |||
| 883 | while (InsertPt != BB->end() && (isa<PHINode>(InsertPt) || | |||
| 884 | InsertPt->isEHPad())) { | |||
| 885 | if (isa<CatchSwitchInst>(InsertPt)) | |||
| 886 | // A catchswitch cannot have anything in the block except | |||
| 887 | // itself and PHIs. We'll bail out below. | |||
| 888 | FoundCatchSwitch = true; | |||
| 889 | ++InsertPt; | |||
| 890 | } | |||
| 891 | } else { | |||
| 892 | InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); | |||
| 893 | } | |||
| 894 | ||||
| 895 | // We found a catchswitch in the block where we want to move the | |||
| 896 | // neg. We cannot move anything into that block. Bail and just | |||
| 897 | // create the neg before BI, as if we hadn't found an existing | |||
| 898 | // neg. | |||
| 899 | if (FoundCatchSwitch) | |||
| 900 | break; | |||
| 901 | ||||
| 902 | TheNeg->moveBefore(&*InsertPt); | |||
| 903 | if (TheNeg->getOpcode() == Instruction::Sub) { | |||
| 904 | TheNeg->setHasNoUnsignedWrap(false); | |||
| 905 | TheNeg->setHasNoSignedWrap(false); | |||
| 906 | } else { | |||
| 907 | TheNeg->andIRFlags(BI); | |||
| 908 | } | |||
| 909 | ToRedo.insert(TheNeg); | |||
| 910 | return TheNeg; | |||
| 911 | } | |||
| 912 | ||||
| 913 | // Insert a 'neg' instruction that subtracts the value from zero to get the | |||
| 914 | // negation. | |||
| 915 | Instruction *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI); | |||
| 916 | ToRedo.insert(NewNeg); | |||
| 917 | return NewNeg; | |||
| 918 | } | |||
| 919 | ||||
| 920 | // See if this `or` looks like an load widening reduction, i.e. that it | |||
| 921 | // consists of an `or`/`shl`/`zext`/`load` nodes only. Note that we don't | |||
| 922 | // ensure that the pattern is *really* a load widening reduction, | |||
| 923 | // we do not ensure that it can really be replaced with a widened load, | |||
| 924 | // only that it mostly looks like one. | |||
| 925 | static bool isLoadCombineCandidate(Instruction *Or) { | |||
| 926 | SmallVector<Instruction *, 8> Worklist; | |||
| 927 | SmallSet<Instruction *, 8> Visited; | |||
| 928 | ||||
| 929 | auto Enqueue = [&](Value *V) { | |||
| 930 | auto *I = dyn_cast<Instruction>(V); | |||
| 931 | // Each node of an `or` reduction must be an instruction, | |||
| 932 | if (!I) | |||
| 933 | return false; // Node is certainly not part of an `or` load reduction. | |||
| 934 | // Only process instructions we have never processed before. | |||
| 935 | if (Visited.insert(I).second) | |||
| 936 | Worklist.emplace_back(I); | |||
| 937 | return true; // Will need to look at parent nodes. | |||
| 938 | }; | |||
| 939 | ||||
| 940 | if (!Enqueue(Or)) | |||
| 941 | return false; // Not an `or` reduction pattern. | |||
| 942 | ||||
| 943 | while (!Worklist.empty()) { | |||
| 944 | auto *I = Worklist.pop_back_val(); | |||
| 945 | ||||
| 946 | // Okay, which instruction is this node? | |||
| 947 | switch (I->getOpcode()) { | |||
| 948 | case Instruction::Or: | |||
| 949 | // Got an `or` node. That's fine, just recurse into it's operands. | |||
| 950 | for (Value *Op : I->operands()) | |||
| 951 | if (!Enqueue(Op)) | |||
| 952 | return false; // Not an `or` reduction pattern. | |||
| 953 | continue; | |||
| 954 | ||||
| 955 | case Instruction::Shl: | |||
| 956 | case Instruction::ZExt: | |||
| 957 | // `shl`/`zext` nodes are fine, just recurse into their base operand. | |||
| 958 | if (!Enqueue(I->getOperand(0))) | |||
| 959 | return false; // Not an `or` reduction pattern. | |||
| 960 | continue; | |||
| 961 | ||||
| 962 | case Instruction::Load: | |||
| 963 | // Perfect, `load` node means we've reached an edge of the graph. | |||
| 964 | continue; | |||
| 965 | ||||
| 966 | default: // Unknown node. | |||
| 967 | return false; // Not an `or` reduction pattern. | |||
| 968 | } | |||
| 969 | } | |||
| 970 | ||||
| 971 | return true; | |||
| 972 | } | |||
| 973 | ||||
| 974 | /// Return true if it may be profitable to convert this (X|Y) into (X+Y). | |||
| 975 | static bool shouldConvertOrWithNoCommonBitsToAdd(Instruction *Or) { | |||
| 976 | // Don't bother to convert this up unless either the LHS is an associable add | |||
| 977 | // or subtract or mul or if this is only used by one of the above. | |||
| 978 | // This is only a compile-time improvement, it is not needed for correctness! | |||
| 979 | auto isInteresting = [](Value *V) { | |||
| 980 | for (auto Op : {Instruction::Add, Instruction::Sub, Instruction::Mul, | |||
| 981 | Instruction::Shl}) | |||
| 982 | if (isReassociableOp(V, Op)) | |||
| 983 | return true; | |||
| 984 | return false; | |||
| 985 | }; | |||
| 986 | ||||
| 987 | if (any_of(Or->operands(), isInteresting)) | |||
| 988 | return true; | |||
| 989 | ||||
| 990 | Value *VB = Or->user_back(); | |||
| 991 | if (Or->hasOneUse() && isInteresting(VB)) | |||
| 992 | return true; | |||
| 993 | ||||
| 994 | return false; | |||
| 995 | } | |||
| 996 | ||||
| 997 | /// If we have (X|Y), and iff X and Y have no common bits set, | |||
| 998 | /// transform this into (X+Y) to allow arithmetics reassociation. | |||
| 999 | static BinaryOperator *convertOrWithNoCommonBitsToAdd(Instruction *Or) { | |||
| 1000 | // Convert an or into an add. | |||
| 1001 | BinaryOperator *New = | |||
| 1002 | CreateAdd(Or->getOperand(0), Or->getOperand(1), "", Or, Or); | |||
| 1003 | New->setHasNoSignedWrap(); | |||
| 1004 | New->setHasNoUnsignedWrap(); | |||
| 1005 | New->takeName(Or); | |||
| 1006 | ||||
| 1007 | // Everyone now refers to the add instruction. | |||
| 1008 | Or->replaceAllUsesWith(New); | |||
| 1009 | New->setDebugLoc(Or->getDebugLoc()); | |||
| 1010 | ||||
| 1011 | LLVM_DEBUG(dbgs() << "Converted or into an add: " << *New << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Converted or into an add: " << *New << '\n'; } } while (false); | |||
| 1012 | return New; | |||
| 1013 | } | |||
| 1014 | ||||
| 1015 | /// Return true if we should break up this subtract of X-Y into (X + -Y). | |||
| 1016 | static bool ShouldBreakUpSubtract(Instruction *Sub) { | |||
| 1017 | // If this is a negation, we can't split it up! | |||
| 1018 | if (match(Sub, m_Neg(m_Value())) || match(Sub, m_FNeg(m_Value()))) | |||
| 1019 | return false; | |||
| 1020 | ||||
| 1021 | // Don't breakup X - undef. | |||
| 1022 | if (isa<UndefValue>(Sub->getOperand(1))) | |||
| 1023 | return false; | |||
| 1024 | ||||
| 1025 | // Don't bother to break this up unless either the LHS is an associable add or | |||
| 1026 | // subtract or if this is only used by one. | |||
| 1027 | Value *V0 = Sub->getOperand(0); | |||
| 1028 | if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) || | |||
| 1029 | isReassociableOp(V0, Instruction::Sub, Instruction::FSub)) | |||
| 1030 | return true; | |||
| 1031 | Value *V1 = Sub->getOperand(1); | |||
| 1032 | if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) || | |||
| 1033 | isReassociableOp(V1, Instruction::Sub, Instruction::FSub)) | |||
| 1034 | return true; | |||
| 1035 | Value *VB = Sub->user_back(); | |||
| 1036 | if (Sub->hasOneUse() && | |||
| 1037 | (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) || | |||
| 1038 | isReassociableOp(VB, Instruction::Sub, Instruction::FSub))) | |||
| 1039 | return true; | |||
| 1040 | ||||
| 1041 | return false; | |||
| 1042 | } | |||
| 1043 | ||||
| 1044 | /// If we have (X-Y), and if either X is an add, or if this is only used by an | |||
| 1045 | /// add, transform this into (X+(0-Y)) to promote better reassociation. | |||
| 1046 | static BinaryOperator *BreakUpSubtract(Instruction *Sub, | |||
| 1047 | ReassociatePass::OrderedSet &ToRedo) { | |||
| 1048 | // Convert a subtract into an add and a neg instruction. This allows sub | |||
| 1049 | // instructions to be commuted with other add instructions. | |||
| 1050 | // | |||
| 1051 | // Calculate the negative value of Operand 1 of the sub instruction, | |||
| 1052 | // and set it as the RHS of the add instruction we just made. | |||
| 1053 | Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo); | |||
| 1054 | BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub); | |||
| 1055 | Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op. | |||
| 1056 | Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op. | |||
| 1057 | New->takeName(Sub); | |||
| 1058 | ||||
| 1059 | // Everyone now refers to the add instruction. | |||
| 1060 | Sub->replaceAllUsesWith(New); | |||
| 1061 | New->setDebugLoc(Sub->getDebugLoc()); | |||
| 1062 | ||||
| 1063 | LLVM_DEBUG(dbgs() << "Negated: " << *New << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Negated: " << *New << '\n'; } } while (false); | |||
| 1064 | return New; | |||
| 1065 | } | |||
| 1066 | ||||
| 1067 | /// If this is a shift of a reassociable multiply or is used by one, change | |||
| 1068 | /// this into a multiply by a constant to assist with further reassociation. | |||
| 1069 | static BinaryOperator *ConvertShiftToMul(Instruction *Shl) { | |||
| 1070 | Constant *MulCst = ConstantInt::get(Shl->getType(), 1); | |||
| 1071 | auto *SA = cast<ConstantInt>(Shl->getOperand(1)); | |||
| 1072 | MulCst = ConstantExpr::getShl(MulCst, SA); | |||
| 1073 | ||||
| 1074 | BinaryOperator *Mul = | |||
| 1075 | BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); | |||
| 1076 | Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op. | |||
| 1077 | Mul->takeName(Shl); | |||
| 1078 | ||||
| 1079 | // Everyone now refers to the mul instruction. | |||
| 1080 | Shl->replaceAllUsesWith(Mul); | |||
| 1081 | Mul->setDebugLoc(Shl->getDebugLoc()); | |||
| 1082 | ||||
| 1083 | // We can safely preserve the nuw flag in all cases. It's also safe to turn a | |||
| 1084 | // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special | |||
| 1085 | // handling. It can be preserved as long as we're not left shifting by | |||
| 1086 | // bitwidth - 1. | |||
| 1087 | bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap(); | |||
| 1088 | bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap(); | |||
| 1089 | unsigned BitWidth = Shl->getType()->getIntegerBitWidth(); | |||
| 1090 | if (NSW && (NUW || SA->getValue().ult(BitWidth - 1))) | |||
| 1091 | Mul->setHasNoSignedWrap(true); | |||
| 1092 | Mul->setHasNoUnsignedWrap(NUW); | |||
| 1093 | return Mul; | |||
| 1094 | } | |||
| 1095 | ||||
| 1096 | /// Scan backwards and forwards among values with the same rank as element i | |||
| 1097 | /// to see if X exists. If X does not exist, return i. This is useful when | |||
| 1098 | /// scanning for 'x' when we see '-x' because they both get the same rank. | |||
| 1099 | static unsigned FindInOperandList(const SmallVectorImpl<ValueEntry> &Ops, | |||
| 1100 | unsigned i, Value *X) { | |||
| 1101 | unsigned XRank = Ops[i].Rank; | |||
| 1102 | unsigned e = Ops.size(); | |||
| 1103 | for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) { | |||
| 1104 | if (Ops[j].Op == X) | |||
| 1105 | return j; | |||
| 1106 | if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op)) | |||
| 1107 | if (Instruction *I2 = dyn_cast<Instruction>(X)) | |||
| 1108 | if (I1->isIdenticalTo(I2)) | |||
| 1109 | return j; | |||
| 1110 | } | |||
| 1111 | // Scan backwards. | |||
| 1112 | for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) { | |||
| 1113 | if (Ops[j].Op == X) | |||
| 1114 | return j; | |||
| 1115 | if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op)) | |||
| 1116 | if (Instruction *I2 = dyn_cast<Instruction>(X)) | |||
| 1117 | if (I1->isIdenticalTo(I2)) | |||
| 1118 | return j; | |||
| 1119 | } | |||
| 1120 | return i; | |||
| 1121 | } | |||
| 1122 | ||||
| 1123 | /// Emit a tree of add instructions, summing Ops together | |||
| 1124 | /// and returning the result. Insert the tree before I. | |||
| 1125 | static Value *EmitAddTreeOfValues(Instruction *I, | |||
| 1126 | SmallVectorImpl<WeakTrackingVH> &Ops) { | |||
| 1127 | if (Ops.size() == 1) return Ops.back(); | |||
| 1128 | ||||
| 1129 | Value *V1 = Ops.pop_back_val(); | |||
| 1130 | Value *V2 = EmitAddTreeOfValues(I, Ops); | |||
| 1131 | return CreateAdd(V2, V1, "reass.add", I, I); | |||
| 1132 | } | |||
| 1133 | ||||
| 1134 | /// If V is an expression tree that is a multiplication sequence, | |||
| 1135 | /// and if this sequence contains a multiply by Factor, | |||
| 1136 | /// remove Factor from the tree and return the new tree. | |||
| 1137 | Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) { | |||
| 1138 | BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); | |||
| 1139 | if (!BO) | |||
| 1140 | return nullptr; | |||
| 1141 | ||||
| 1142 | SmallVector<RepeatedValue, 8> Tree; | |||
| 1143 | MadeChange |= LinearizeExprTree(BO, Tree); | |||
| 1144 | SmallVector<ValueEntry, 8> Factors; | |||
| 1145 | Factors.reserve(Tree.size()); | |||
| 1146 | for (unsigned i = 0, e = Tree.size(); i != e; ++i) { | |||
| 1147 | RepeatedValue E = Tree[i]; | |||
| 1148 | Factors.append(E.second.getZExtValue(), | |||
| 1149 | ValueEntry(getRank(E.first), E.first)); | |||
| 1150 | } | |||
| 1151 | ||||
| 1152 | bool FoundFactor = false; | |||
| 1153 | bool NeedsNegate = false; | |||
| 1154 | for (unsigned i = 0, e = Factors.size(); i != e; ++i) { | |||
| 1155 | if (Factors[i].Op == Factor) { | |||
| 1156 | FoundFactor = true; | |||
| 1157 | Factors.erase(Factors.begin()+i); | |||
| 1158 | break; | |||
| 1159 | } | |||
| 1160 | ||||
| 1161 | // If this is a negative version of this factor, remove it. | |||
| 1162 | if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) { | |||
| 1163 | if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) | |||
| 1164 | if (FC1->getValue() == -FC2->getValue()) { | |||
| 1165 | FoundFactor = NeedsNegate = true; | |||
| 1166 | Factors.erase(Factors.begin()+i); | |||
| 1167 | break; | |||
| 1168 | } | |||
| 1169 | } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) { | |||
| 1170 | if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) { | |||
| 1171 | const APFloat &F1 = FC1->getValueAPF(); | |||
| 1172 | APFloat F2(FC2->getValueAPF()); | |||
| 1173 | F2.changeSign(); | |||
| 1174 | if (F1 == F2) { | |||
| 1175 | FoundFactor = NeedsNegate = true; | |||
| 1176 | Factors.erase(Factors.begin() + i); | |||
| 1177 | break; | |||
| 1178 | } | |||
| 1179 | } | |||
| 1180 | } | |||
| 1181 | } | |||
| 1182 | ||||
| 1183 | if (!FoundFactor) { | |||
| 1184 | // Make sure to restore the operands to the expression tree. | |||
| 1185 | RewriteExprTree(BO, Factors); | |||
| 1186 | return nullptr; | |||
| 1187 | } | |||
| 1188 | ||||
| 1189 | BasicBlock::iterator InsertPt = ++BO->getIterator(); | |||
| 1190 | ||||
| 1191 | // If this was just a single multiply, remove the multiply and return the only | |||
| 1192 | // remaining operand. | |||
| 1193 | if (Factors.size() == 1) { | |||
| 1194 | RedoInsts.insert(BO); | |||
| 1195 | V = Factors[0].Op; | |||
| 1196 | } else { | |||
| 1197 | RewriteExprTree(BO, Factors); | |||
| 1198 | V = BO; | |||
| 1199 | } | |||
| 1200 | ||||
| 1201 | if (NeedsNegate) | |||
| 1202 | V = CreateNeg(V, "neg", &*InsertPt, BO); | |||
| 1203 | ||||
| 1204 | return V; | |||
| 1205 | } | |||
| 1206 | ||||
| 1207 | /// If V is a single-use multiply, recursively add its operands as factors, | |||
| 1208 | /// otherwise add V to the list of factors. | |||
| 1209 | /// | |||
| 1210 | /// Ops is the top-level list of add operands we're trying to factor. | |||
| 1211 | static void FindSingleUseMultiplyFactors(Value *V, | |||
| 1212 | SmallVectorImpl<Value*> &Factors) { | |||
| 1213 | BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); | |||
| 1214 | if (!BO) { | |||
| 1215 | Factors.push_back(V); | |||
| 1216 | return; | |||
| 1217 | } | |||
| 1218 | ||||
| 1219 | // Otherwise, add the LHS and RHS to the list of factors. | |||
| 1220 | FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); | |||
| 1221 | FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); | |||
| 1222 | } | |||
| 1223 | ||||
| 1224 | /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction. | |||
| 1225 | /// This optimizes based on identities. If it can be reduced to a single Value, | |||
| 1226 | /// it is returned, otherwise the Ops list is mutated as necessary. | |||
| 1227 | static Value *OptimizeAndOrXor(unsigned Opcode, | |||
| 1228 | SmallVectorImpl<ValueEntry> &Ops) { | |||
| 1229 | // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. | |||
| 1230 | // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. | |||
| 1231 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { | |||
| 1232 | // First, check for X and ~X in the operand list. | |||
| 1233 | assert(i < Ops.size())(static_cast <bool> (i < Ops.size()) ? void (0) : __assert_fail ("i < Ops.size()", "llvm/lib/Transforms/Scalar/Reassociate.cpp" , 1233, __extension__ __PRETTY_FUNCTION__)); | |||
| 1234 | Value *X; | |||
| 1235 | if (match(Ops[i].Op, m_Not(m_Value(X)))) { // Cannot occur for ^. | |||
| 1236 | unsigned FoundX = FindInOperandList(Ops, i, X); | |||
| 1237 | if (FoundX != i) { | |||
| 1238 | if (Opcode == Instruction::And) // ...&X&~X = 0 | |||
| 1239 | return Constant::getNullValue(X->getType()); | |||
| 1240 | ||||
| 1241 | if (Opcode == Instruction::Or) // ...|X|~X = -1 | |||
| 1242 | return Constant::getAllOnesValue(X->getType()); | |||
| 1243 | } | |||
| 1244 | } | |||
| 1245 | ||||
| 1246 | // Next, check for duplicate pairs of values, which we assume are next to | |||
| 1247 | // each other, due to our sorting criteria. | |||
| 1248 | assert(i < Ops.size())(static_cast <bool> (i < Ops.size()) ? void (0) : __assert_fail ("i < Ops.size()", "llvm/lib/Transforms/Scalar/Reassociate.cpp" , 1248, __extension__ __PRETTY_FUNCTION__)); | |||
| 1249 | if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { | |||
| 1250 | if (Opcode == Instruction::And || Opcode == Instruction::Or) { | |||
| 1251 | // Drop duplicate values for And and Or. | |||
| 1252 | Ops.erase(Ops.begin()+i); | |||
| 1253 | --i; --e; | |||
| 1254 | ++NumAnnihil; | |||
| 1255 | continue; | |||
| 1256 | } | |||
| 1257 | ||||
| 1258 | // Drop pairs of values for Xor. | |||
| 1259 | assert(Opcode == Instruction::Xor)(static_cast <bool> (Opcode == Instruction::Xor) ? void (0) : __assert_fail ("Opcode == Instruction::Xor", "llvm/lib/Transforms/Scalar/Reassociate.cpp" , 1259, __extension__ __PRETTY_FUNCTION__)); | |||
| 1260 | if (e == 2) | |||
| 1261 | return Constant::getNullValue(Ops[0].Op->getType()); | |||
| 1262 | ||||
| 1263 | // Y ^ X^X -> Y | |||
| 1264 | Ops.erase(Ops.begin()+i, Ops.begin()+i+2); | |||
| 1265 | i -= 1; e -= 2; | |||
| 1266 | ++NumAnnihil; | |||
| 1267 | } | |||
| 1268 | } | |||
| 1269 | return nullptr; | |||
| 1270 | } | |||
| 1271 | ||||
| 1272 | /// Helper function of CombineXorOpnd(). It creates a bitwise-and | |||
| 1273 | /// instruction with the given two operands, and return the resulting | |||
| 1274 | /// instruction. There are two special cases: 1) if the constant operand is 0, | |||
| 1275 | /// it will return NULL. 2) if the constant is ~0, the symbolic operand will | |||
| 1276 | /// be returned. | |||
| 1277 | static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd, | |||
| 1278 | const APInt &ConstOpnd) { | |||
| 1279 | if (ConstOpnd.isZero()) | |||
| 1280 | return nullptr; | |||
| 1281 | ||||
| 1282 | if (ConstOpnd.isAllOnes()) | |||
| 1283 | return Opnd; | |||
| 1284 | ||||
| 1285 | Instruction *I = BinaryOperator::CreateAnd( | |||
| 1286 | Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra", | |||
| 1287 | InsertBefore); | |||
| 1288 | I->setDebugLoc(InsertBefore->getDebugLoc()); | |||
| 1289 | return I; | |||
| 1290 | } | |||
| 1291 | ||||
| 1292 | // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd" | |||
| 1293 | // into "R ^ C", where C would be 0, and R is a symbolic value. | |||
| 1294 | // | |||
| 1295 | // If it was successful, true is returned, and the "R" and "C" is returned | |||
| 1296 | // via "Res" and "ConstOpnd", respectively; otherwise, false is returned, | |||
| 1297 | // and both "Res" and "ConstOpnd" remain unchanged. | |||
| 1298 | bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, | |||
| 1299 | APInt &ConstOpnd, Value *&Res) { | |||
| 1300 | // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2 | |||
| 1301 | // = ((x | c1) ^ c1) ^ (c1 ^ c2) | |||
| 1302 | // = (x & ~c1) ^ (c1 ^ c2) | |||
| 1303 | // It is useful only when c1 == c2. | |||
| 1304 | if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isZero()) | |||
| 1305 | return false; | |||
| 1306 | ||||
| 1307 | if (!Opnd1->getValue()->hasOneUse()) | |||
| 1308 | return false; | |||
| 1309 | ||||
| 1310 | const APInt &C1 = Opnd1->getConstPart(); | |||
| 1311 | if (C1 != ConstOpnd) | |||
| 1312 | return false; | |||
| 1313 | ||||
| 1314 | Value *X = Opnd1->getSymbolicPart(); | |||
| 1315 | Res = createAndInstr(I, X, ~C1); | |||
| 1316 | // ConstOpnd was C2, now C1 ^ C2. | |||
| 1317 | ConstOpnd ^= C1; | |||
| 1318 | ||||
| 1319 | if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue())) | |||
| 1320 | RedoInsts.insert(T); | |||
| 1321 | return true; | |||
| 1322 | } | |||
| 1323 | ||||
| 1324 | // Helper function of OptimizeXor(). It tries to simplify | |||
| 1325 | // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a | |||
| 1326 | // symbolic value. | |||
| 1327 | // | |||
| 1328 | // If it was successful, true is returned, and the "R" and "C" is returned | |||
| 1329 | // via "Res" and "ConstOpnd", respectively (If the entire expression is | |||
| 1330 | // evaluated to a constant, the Res is set to NULL); otherwise, false is | |||
| 1331 | // returned, and both "Res" and "ConstOpnd" remain unchanged. | |||
| 1332 | bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, | |||
| 1333 | XorOpnd *Opnd2, APInt &ConstOpnd, | |||
| 1334 | Value *&Res) { | |||
| 1335 | Value *X = Opnd1->getSymbolicPart(); | |||
| 1336 | if (X != Opnd2->getSymbolicPart()) | |||
| 1337 | return false; | |||
| 1338 | ||||
| 1339 | // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.) | |||
| 1340 | int DeadInstNum = 1; | |||
| 1341 | if (Opnd1->getValue()->hasOneUse()) | |||
| 1342 | DeadInstNum++; | |||
| 1343 | if (Opnd2->getValue()->hasOneUse()) | |||
| 1344 | DeadInstNum++; | |||
| 1345 | ||||
| 1346 | // Xor-Rule 2: | |||
| 1347 | // (x | c1) ^ (x & c2) | |||
| 1348 | // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1 | |||
| 1349 | // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1 | |||
| 1350 | // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3 | |||
| 1351 | // | |||
| 1352 | if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) { | |||
| 1353 | if (Opnd2->isOrExpr()) | |||
| 1354 | std::swap(Opnd1, Opnd2); | |||
| 1355 | ||||
| 1356 | const APInt &C1 = Opnd1->getConstPart(); | |||
| 1357 | const APInt &C2 = Opnd2->getConstPart(); | |||
| 1358 | APInt C3((~C1) ^ C2); | |||
| 1359 | ||||
| 1360 | // Do not increase code size! | |||
| 1361 | if (!C3.isZero() && !C3.isAllOnes()) { | |||
| 1362 | int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2; | |||
| 1363 | if (NewInstNum > DeadInstNum) | |||
| 1364 | return false; | |||
| 1365 | } | |||
| 1366 | ||||
| 1367 | Res = createAndInstr(I, X, C3); | |||
| 1368 | ConstOpnd ^= C1; | |||
| 1369 | } else if (Opnd1->isOrExpr()) { | |||
| 1370 | // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2 | |||
| 1371 | // | |||
| 1372 | const APInt &C1 = Opnd1->getConstPart(); | |||
| 1373 | const APInt &C2 = Opnd2->getConstPart(); | |||
| 1374 | APInt C3 = C1 ^ C2; | |||
| 1375 | ||||
| 1376 | // Do not increase code size | |||
| 1377 | if (!C3.isZero() && !C3.isAllOnes()) { | |||
| 1378 | int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2; | |||
| 1379 | if (NewInstNum > DeadInstNum) | |||
| 1380 | return false; | |||
| 1381 | } | |||
| 1382 | ||||
| 1383 | Res = createAndInstr(I, X, C3); | |||
| 1384 | ConstOpnd ^= C3; | |||
| 1385 | } else { | |||
| 1386 | // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2)) | |||
| 1387 | // | |||
| 1388 | const APInt &C1 = Opnd1->getConstPart(); | |||
| 1389 | const APInt &C2 = Opnd2->getConstPart(); | |||
| 1390 | APInt C3 = C1 ^ C2; | |||
| 1391 | Res = createAndInstr(I, X, C3); | |||
| 1392 | } | |||
| 1393 | ||||
| 1394 | // Put the original operands in the Redo list; hope they will be deleted | |||
| 1395 | // as dead code. | |||
| 1396 | if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue())) | |||
| 1397 | RedoInsts.insert(T); | |||
| 1398 | if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue())) | |||
| 1399 | RedoInsts.insert(T); | |||
| 1400 | ||||
| 1401 | return true; | |||
| 1402 | } | |||
| 1403 | ||||
| 1404 | /// Optimize a series of operands to an 'xor' instruction. If it can be reduced | |||
| 1405 | /// to a single Value, it is returned, otherwise the Ops list is mutated as | |||
| 1406 | /// necessary. | |||
| 1407 | Value *ReassociatePass::OptimizeXor(Instruction *I, | |||
| 1408 | SmallVectorImpl<ValueEntry> &Ops) { | |||
| 1409 | if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops)) | |||
| 1410 | return V; | |||
| 1411 | ||||
| 1412 | if (Ops.size() == 1) | |||
| 1413 | return nullptr; | |||
| 1414 | ||||
| 1415 | SmallVector<XorOpnd, 8> Opnds; | |||
| 1416 | SmallVector<XorOpnd*, 8> OpndPtrs; | |||
| 1417 | Type *Ty = Ops[0].Op->getType(); | |||
| 1418 | APInt ConstOpnd(Ty->getScalarSizeInBits(), 0); | |||
| 1419 | ||||
| 1420 | // Step 1: Convert ValueEntry to XorOpnd | |||
| 1421 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { | |||
| 1422 | Value *V = Ops[i].Op; | |||
| 1423 | const APInt *C; | |||
| 1424 | // TODO: Support non-splat vectors. | |||
| 1425 | if (match(V, m_APInt(C))) { | |||
| 1426 | ConstOpnd ^= *C; | |||
| 1427 | } else { | |||
| 1428 | XorOpnd O(V); | |||
| 1429 | O.setSymbolicRank(getRank(O.getSymbolicPart())); | |||
| 1430 | Opnds.push_back(O); | |||
| 1431 | } | |||
| 1432 | } | |||
| 1433 | ||||
| 1434 | // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds". | |||
| 1435 | // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate | |||
| 1436 | // the "OpndPtrs" as well. For the similar reason, do not fuse this loop | |||
| 1437 | // with the previous loop --- the iterator of the "Opnds" may be invalidated | |||
| 1438 | // when new elements are added to the vector. | |||
| 1439 | for (unsigned i = 0, e = Opnds.size(); i != e; ++i) | |||
| 1440 | OpndPtrs.push_back(&Opnds[i]); | |||
| 1441 | ||||
| 1442 | // Step 2: Sort the Xor-Operands in a way such that the operands containing | |||
| 1443 | // the same symbolic value cluster together. For instance, the input operand | |||
| 1444 | // sequence ("x | 123", "y & 456", "x & 789") will be sorted into: | |||
| 1445 | // ("x | 123", "x & 789", "y & 456"). | |||
| 1446 | // | |||
| 1447 | // The purpose is twofold: | |||
| 1448 | // 1) Cluster together the operands sharing the same symbolic-value. | |||
| 1449 | // 2) Operand having smaller symbolic-value-rank is permuted earlier, which | |||
| 1450 | // could potentially shorten crital path, and expose more loop-invariants. | |||
| 1451 | // Note that values' rank are basically defined in RPO order (FIXME). | |||
| 1452 | // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier | |||
| 1453 | // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2", | |||
| 1454 | // "z" in the order of X-Y-Z is better than any other orders. | |||
| 1455 | llvm::stable_sort(OpndPtrs, [](XorOpnd *LHS, XorOpnd *RHS) { | |||
| 1456 | return LHS->getSymbolicRank() < RHS->getSymbolicRank(); | |||
| 1457 | }); | |||
| 1458 | ||||
| 1459 | // Step 3: Combine adjacent operands | |||
| 1460 | XorOpnd *PrevOpnd = nullptr; | |||
| 1461 | bool Changed = false; | |||
| 1462 | for (unsigned i = 0, e = Opnds.size(); i < e; i++) { | |||
| 1463 | XorOpnd *CurrOpnd = OpndPtrs[i]; | |||
| 1464 | // The combined value | |||
| 1465 | Value *CV; | |||
| 1466 | ||||
| 1467 | // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd" | |||
| 1468 | if (!ConstOpnd.isZero() && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) { | |||
| 1469 | Changed = true; | |||
| 1470 | if (CV) | |||
| 1471 | *CurrOpnd = XorOpnd(CV); | |||
| 1472 | else { | |||
| 1473 | CurrOpnd->Invalidate(); | |||
| 1474 | continue; | |||
| 1475 | } | |||
| 1476 | } | |||
| 1477 | ||||
| 1478 | if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) { | |||
| 1479 | PrevOpnd = CurrOpnd; | |||
| 1480 | continue; | |||
| 1481 | } | |||
| 1482 | ||||
| 1483 | // step 3.2: When previous and current operands share the same symbolic | |||
| 1484 | // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd" | |||
| 1485 | if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) { | |||
| 1486 | // Remove previous operand | |||
| 1487 | PrevOpnd->Invalidate(); | |||
| 1488 | if (CV) { | |||
| 1489 | *CurrOpnd = XorOpnd(CV); | |||
| 1490 | PrevOpnd = CurrOpnd; | |||
| 1491 | } else { | |||
| 1492 | CurrOpnd->Invalidate(); | |||
| 1493 | PrevOpnd = nullptr; | |||
| 1494 | } | |||
| 1495 | Changed = true; | |||
| 1496 | } | |||
| 1497 | } | |||
| 1498 | ||||
| 1499 | // Step 4: Reassemble the Ops | |||
| 1500 | if (Changed) { | |||
| 1501 | Ops.clear(); | |||
| 1502 | for (unsigned int i = 0, e = Opnds.size(); i < e; i++) { | |||
| 1503 | XorOpnd &O = Opnds[i]; | |||
| 1504 | if (O.isInvalid()) | |||
| 1505 | continue; | |||
| 1506 | ValueEntry VE(getRank(O.getValue()), O.getValue()); | |||
| 1507 | Ops.push_back(VE); | |||
| 1508 | } | |||
| 1509 | if (!ConstOpnd.isZero()) { | |||
| 1510 | Value *C = ConstantInt::get(Ty, ConstOpnd); | |||
| 1511 | ValueEntry VE(getRank(C), C); | |||
| 1512 | Ops.push_back(VE); | |||
| 1513 | } | |||
| 1514 | unsigned Sz = Ops.size(); | |||
| 1515 | if (Sz == 1) | |||
| 1516 | return Ops.back().Op; | |||
| 1517 | if (Sz == 0) { | |||
| 1518 | assert(ConstOpnd.isZero())(static_cast <bool> (ConstOpnd.isZero()) ? void (0) : __assert_fail ("ConstOpnd.isZero()", "llvm/lib/Transforms/Scalar/Reassociate.cpp" , 1518, __extension__ __PRETTY_FUNCTION__)); | |||
| 1519 | return ConstantInt::get(Ty, ConstOpnd); | |||
| 1520 | } | |||
| 1521 | } | |||
| 1522 | ||||
| 1523 | return nullptr; | |||
| 1524 | } | |||
| 1525 | ||||
| 1526 | /// Optimize a series of operands to an 'add' instruction. This | |||
| 1527 | /// optimizes based on identities. If it can be reduced to a single Value, it | |||
| 1528 | /// is returned, otherwise the Ops list is mutated as necessary. | |||
| 1529 | Value *ReassociatePass::OptimizeAdd(Instruction *I, | |||
| 1530 | SmallVectorImpl<ValueEntry> &Ops) { | |||
| 1531 | // Scan the operand lists looking for X and -X pairs. If we find any, we | |||
| 1532 | // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it, | |||
| 1533 | // scan for any | |||
| 1534 | // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. | |||
| 1535 | ||||
| 1536 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { | |||
| 1537 | Value *TheOp = Ops[i].Op; | |||
| 1538 | // Check to see if we've seen this operand before. If so, we factor all | |||
| 1539 | // instances of the operand together. Due to our sorting criteria, we know | |||
| 1540 | // that these need to be next to each other in the vector. | |||
| 1541 | if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { | |||
| 1542 | // Rescan the list, remove all instances of this operand from the expr. | |||
| 1543 | unsigned NumFound = 0; | |||
| 1544 | do { | |||
| 1545 | Ops.erase(Ops.begin()+i); | |||
| 1546 | ++NumFound; | |||
| 1547 | } while (i != Ops.size() && Ops[i].Op == TheOp); | |||
| 1548 | ||||
| 1549 | LLVM_DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOpdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'; } } while (false ) | |||
| 1550 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'; } } while (false ); | |||
| 1551 | ++NumFactor; | |||
| 1552 | ||||
| 1553 | // Insert a new multiply. | |||
| 1554 | Type *Ty = TheOp->getType(); | |||
| 1555 | Constant *C = Ty->isIntOrIntVectorTy() ? | |||
| 1556 | ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound); | |||
| 1557 | Instruction *Mul = CreateMul(TheOp, C, "factor", I, I); | |||
| 1558 | ||||
| 1559 | // Now that we have inserted a multiply, optimize it. This allows us to | |||
| 1560 | // handle cases that require multiple factoring steps, such as this: | |||
| 1561 | // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 | |||
| 1562 | RedoInsts.insert(Mul); | |||
| 1563 | ||||
| 1564 | // If every add operand was a duplicate, return the multiply. | |||
| 1565 | if (Ops.empty()) | |||
| 1566 | return Mul; | |||
| 1567 | ||||
| 1568 | // Otherwise, we had some input that didn't have the dupe, such as | |||
| 1569 | // "A + A + B" -> "A*2 + B". Add the new multiply to the list of | |||
| 1570 | // things being added by this operation. | |||
| 1571 | Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); | |||
| 1572 | ||||
| 1573 | --i; | |||
| 1574 | e = Ops.size(); | |||
| 1575 | continue; | |||
| 1576 | } | |||
| 1577 | ||||
| 1578 | // Check for X and -X or X and ~X in the operand list. | |||
| 1579 | Value *X; | |||
| 1580 | if (!match(TheOp, m_Neg(m_Value(X))) && !match(TheOp, m_Not(m_Value(X))) && | |||
| 1581 | !match(TheOp, m_FNeg(m_Value(X)))) | |||
| 1582 | continue; | |||
| 1583 | ||||
| 1584 | unsigned FoundX = FindInOperandList(Ops, i, X); | |||
| 1585 | if (FoundX == i) | |||
| 1586 | continue; | |||
| 1587 | ||||
| 1588 | // Remove X and -X from the operand list. | |||
| 1589 | if (Ops.size() == 2 && | |||
| 1590 | (match(TheOp, m_Neg(m_Value())) || match(TheOp, m_FNeg(m_Value())))) | |||
| 1591 | return Constant::getNullValue(X->getType()); | |||
| 1592 | ||||
| 1593 | // Remove X and ~X from the operand list. | |||
| 1594 | if (Ops.size() == 2 && match(TheOp, m_Not(m_Value()))) | |||
| 1595 | return Constant::getAllOnesValue(X->getType()); | |||
| 1596 | ||||
| 1597 | Ops.erase(Ops.begin()+i); | |||
| 1598 | if (i < FoundX) | |||
| 1599 | --FoundX; | |||
| 1600 | else | |||
| 1601 | --i; // Need to back up an extra one. | |||
| 1602 | Ops.erase(Ops.begin()+FoundX); | |||
| 1603 | ++NumAnnihil; | |||
| 1604 | --i; // Revisit element. | |||
| 1605 | e -= 2; // Removed two elements. | |||
| 1606 | ||||
| 1607 | // if X and ~X we append -1 to the operand list. | |||
| 1608 | if (match(TheOp, m_Not(m_Value()))) { | |||
| 1609 | Value *V = Constant::getAllOnesValue(X->getType()); | |||
| 1610 | Ops.insert(Ops.end(), ValueEntry(getRank(V), V)); | |||
| 1611 | e += 1; | |||
| 1612 | } | |||
| 1613 | } | |||
| 1614 | ||||
| 1615 | // Scan the operand list, checking to see if there are any common factors | |||
| 1616 | // between operands. Consider something like A*A+A*B*C+D. We would like to | |||
| 1617 | // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. | |||
| 1618 | // To efficiently find this, we count the number of times a factor occurs | |||
| 1619 | // for any ADD operands that are MULs. | |||
| 1620 | DenseMap<Value*, unsigned> FactorOccurrences; | |||
| 1621 | ||||
| 1622 | // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) | |||
| 1623 | // where they are actually the same multiply. | |||
| 1624 | unsigned MaxOcc = 0; | |||
| 1625 | Value *MaxOccVal = nullptr; | |||
| 1626 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { | |||
| 1627 | BinaryOperator *BOp = | |||
| 1628 | isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); | |||
| 1629 | if (!BOp) | |||
| 1630 | continue; | |||
| 1631 | ||||
| 1632 | // Compute all of the factors of this added value. | |||
| 1633 | SmallVector<Value*, 8> Factors; | |||
| 1634 | FindSingleUseMultiplyFactors(BOp, Factors); | |||
| 1635 | assert(Factors.size() > 1 && "Bad linearize!")(static_cast <bool> (Factors.size() > 1 && "Bad linearize!" ) ? void (0) : __assert_fail ("Factors.size() > 1 && \"Bad linearize!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 1635, __extension__ __PRETTY_FUNCTION__)); | |||
| 1636 | ||||
| 1637 | // Add one to FactorOccurrences for each unique factor in this op. | |||
| 1638 | SmallPtrSet<Value*, 8> Duplicates; | |||
| 1639 | for (unsigned i = 0, e = Factors.size(); i != e; ++i) { | |||
| 1640 | Value *Factor = Factors[i]; | |||
| 1641 | if (!Duplicates.insert(Factor).second) | |||
| 1642 | continue; | |||
| 1643 | ||||
| 1644 | unsigned Occ = ++FactorOccurrences[Factor]; | |||
| 1645 | if (Occ > MaxOcc) { | |||
| 1646 | MaxOcc = Occ; | |||
| 1647 | MaxOccVal = Factor; | |||
| 1648 | } | |||
| 1649 | ||||
| 1650 | // If Factor is a negative constant, add the negated value as a factor | |||
| 1651 | // because we can percolate the negate out. Watch for minint, which | |||
| 1652 | // cannot be positivified. | |||
| 1653 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) { | |||
| 1654 | if (CI->isNegative() && !CI->isMinValue(true)) { | |||
| 1655 | Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); | |||
| 1656 | if (!Duplicates.insert(Factor).second) | |||
| 1657 | continue; | |||
| 1658 | unsigned Occ = ++FactorOccurrences[Factor]; | |||
| 1659 | if (Occ > MaxOcc) { | |||
| 1660 | MaxOcc = Occ; | |||
| 1661 | MaxOccVal = Factor; | |||
| 1662 | } | |||
| 1663 | } | |||
| 1664 | } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) { | |||
| 1665 | if (CF->isNegative()) { | |||
| 1666 | APFloat F(CF->getValueAPF()); | |||
| 1667 | F.changeSign(); | |||
| 1668 | Factor = ConstantFP::get(CF->getContext(), F); | |||
| 1669 | if (!Duplicates.insert(Factor).second) | |||
| 1670 | continue; | |||
| 1671 | unsigned Occ = ++FactorOccurrences[Factor]; | |||
| 1672 | if (Occ > MaxOcc) { | |||
| 1673 | MaxOcc = Occ; | |||
| 1674 | MaxOccVal = Factor; | |||
| 1675 | } | |||
| 1676 | } | |||
| 1677 | } | |||
| 1678 | } | |||
| 1679 | } | |||
| 1680 | ||||
| 1681 | // If any factor occurred more than one time, we can pull it out. | |||
| 1682 | if (MaxOcc > 1) { | |||
| 1683 | LLVM_DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccValdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'; } } while ( false) | |||
| 1684 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'; } } while ( false); | |||
| 1685 | ++NumFactor; | |||
| 1686 | ||||
| 1687 | // Create a new instruction that uses the MaxOccVal twice. If we don't do | |||
| 1688 | // this, we could otherwise run into situations where removing a factor | |||
| 1689 | // from an expression will drop a use of maxocc, and this can cause | |||
| 1690 | // RemoveFactorFromExpression on successive values to behave differently. | |||
| 1691 | Instruction *DummyInst = | |||
| 1692 | I->getType()->isIntOrIntVectorTy() | |||
| 1693 | ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal) | |||
| 1694 | : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal); | |||
| 1695 | ||||
| 1696 | SmallVector<WeakTrackingVH, 4> NewMulOps; | |||
| 1697 | for (unsigned i = 0; i != Ops.size(); ++i) { | |||
| 1698 | // Only try to remove factors from expressions we're allowed to. | |||
| 1699 | BinaryOperator *BOp = | |||
| 1700 | isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); | |||
| 1701 | if (!BOp) | |||
| 1702 | continue; | |||
| 1703 | ||||
| 1704 | if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { | |||
| 1705 | // The factorized operand may occur several times. Convert them all in | |||
| 1706 | // one fell swoop. | |||
| 1707 | for (unsigned j = Ops.size(); j != i;) { | |||
| 1708 | --j; | |||
| 1709 | if (Ops[j].Op == Ops[i].Op) { | |||
| 1710 | NewMulOps.push_back(V); | |||
| 1711 | Ops.erase(Ops.begin()+j); | |||
| 1712 | } | |||
| 1713 | } | |||
| 1714 | --i; | |||
| 1715 | } | |||
| 1716 | } | |||
| 1717 | ||||
| 1718 | // No need for extra uses anymore. | |||
| 1719 | DummyInst->deleteValue(); | |||
| 1720 | ||||
| 1721 | unsigned NumAddedValues = NewMulOps.size(); | |||
| 1722 | Value *V = EmitAddTreeOfValues(I, NewMulOps); | |||
| 1723 | ||||
| 1724 | // Now that we have inserted the add tree, optimize it. This allows us to | |||
| 1725 | // handle cases that require multiple factoring steps, such as this: | |||
| 1726 | // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) | |||
| 1727 | 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\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 1727, __extension__ __PRETTY_FUNCTION__)); | |||
| 1728 | (void)NumAddedValues; | |||
| 1729 | if (Instruction *VI = dyn_cast<Instruction>(V)) | |||
| 1730 | RedoInsts.insert(VI); | |||
| 1731 | ||||
| 1732 | // Create the multiply. | |||
| 1733 | Instruction *V2 = CreateMul(V, MaxOccVal, "reass.mul", I, I); | |||
| 1734 | ||||
| 1735 | // Rerun associate on the multiply in case the inner expression turned into | |||
| 1736 | // a multiply. We want to make sure that we keep things in canonical form. | |||
| 1737 | RedoInsts.insert(V2); | |||
| 1738 | ||||
| 1739 | // If every add operand included the factor (e.g. "A*B + A*C"), then the | |||
| 1740 | // entire result expression is just the multiply "A*(B+C)". | |||
| 1741 | if (Ops.empty()) | |||
| 1742 | return V2; | |||
| 1743 | ||||
| 1744 | // Otherwise, we had some input that didn't have the factor, such as | |||
| 1745 | // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of | |||
| 1746 | // things being added by this operation. | |||
| 1747 | Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); | |||
| 1748 | } | |||
| 1749 | ||||
| 1750 | return nullptr; | |||
| 1751 | } | |||
| 1752 | ||||
| 1753 | /// Build up a vector of value/power pairs factoring a product. | |||
| 1754 | /// | |||
| 1755 | /// Given a series of multiplication operands, build a vector of factors and | |||
| 1756 | /// the powers each is raised to when forming the final product. Sort them in | |||
| 1757 | /// the order of descending power. | |||
| 1758 | /// | |||
| 1759 | /// (x*x) -> [(x, 2)] | |||
| 1760 | /// ((x*x)*x) -> [(x, 3)] | |||
| 1761 | /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)] | |||
| 1762 | /// | |||
| 1763 | /// \returns Whether any factors have a power greater than one. | |||
| 1764 | static bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, | |||
| 1765 | SmallVectorImpl<Factor> &Factors) { | |||
| 1766 | // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this. | |||
| 1767 | // Compute the sum of powers of simplifiable factors. | |||
| 1768 | unsigned FactorPowerSum = 0; | |||
| 1769 | for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) { | |||
| 1770 | Value *Op = Ops[Idx-1].Op; | |||
| 1771 | ||||
| 1772 | // Count the number of occurrences of this value. | |||
| 1773 | unsigned Count = 1; | |||
| 1774 | for (; Idx < Size && Ops[Idx].Op == Op; ++Idx) | |||
| 1775 | ++Count; | |||
| 1776 | // Track for simplification all factors which occur 2 or more times. | |||
| 1777 | if (Count > 1) | |||
| 1778 | FactorPowerSum += Count; | |||
| 1779 | } | |||
| 1780 | ||||
| 1781 | // We can only simplify factors if the sum of the powers of our simplifiable | |||
| 1782 | // factors is 4 or higher. When that is the case, we will *always* have | |||
| 1783 | // a simplification. This is an important invariant to prevent cyclicly | |||
| 1784 | // trying to simplify already minimal formations. | |||
| 1785 | if (FactorPowerSum < 4) | |||
| 1786 | return false; | |||
| 1787 | ||||
| 1788 | // Now gather the simplifiable factors, removing them from Ops. | |||
| 1789 | FactorPowerSum = 0; | |||
| 1790 | for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) { | |||
| 1791 | Value *Op = Ops[Idx-1].Op; | |||
| 1792 | ||||
| 1793 | // Count the number of occurrences of this value. | |||
| 1794 | unsigned Count = 1; | |||
| 1795 | for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx) | |||
| 1796 | ++Count; | |||
| 1797 | if (Count == 1) | |||
| 1798 | continue; | |||
| 1799 | // Move an even number of occurrences to Factors. | |||
| 1800 | Count &= ~1U; | |||
| 1801 | Idx -= Count; | |||
| 1802 | FactorPowerSum += Count; | |||
| 1803 | Factors.push_back(Factor(Op, Count)); | |||
| 1804 | Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count); | |||
| 1805 | } | |||
| 1806 | ||||
| 1807 | // None of the adjustments above should have reduced the sum of factor powers | |||
| 1808 | // below our mininum of '4'. | |||
| 1809 | assert(FactorPowerSum >= 4)(static_cast <bool> (FactorPowerSum >= 4) ? void (0) : __assert_fail ("FactorPowerSum >= 4", "llvm/lib/Transforms/Scalar/Reassociate.cpp" , 1809, __extension__ __PRETTY_FUNCTION__)); | |||
| 1810 | ||||
| 1811 | llvm::stable_sort(Factors, [](const Factor &LHS, const Factor &RHS) { | |||
| 1812 | return LHS.Power > RHS.Power; | |||
| 1813 | }); | |||
| 1814 | return true; | |||
| 1815 | } | |||
| 1816 | ||||
| 1817 | /// Build a tree of multiplies, computing the product of Ops. | |||
| 1818 | static Value *buildMultiplyTree(IRBuilderBase &Builder, | |||
| 1819 | SmallVectorImpl<Value*> &Ops) { | |||
| 1820 | if (Ops.size() == 1) | |||
| 1821 | return Ops.back(); | |||
| 1822 | ||||
| 1823 | Value *LHS = Ops.pop_back_val(); | |||
| 1824 | do { | |||
| 1825 | if (LHS->getType()->isIntOrIntVectorTy()) | |||
| 1826 | LHS = Builder.CreateMul(LHS, Ops.pop_back_val()); | |||
| 1827 | else | |||
| 1828 | LHS = Builder.CreateFMul(LHS, Ops.pop_back_val()); | |||
| 1829 | } while (!Ops.empty()); | |||
| 1830 | ||||
| 1831 | return LHS; | |||
| 1832 | } | |||
| 1833 | ||||
| 1834 | /// Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*... | |||
| 1835 | /// | |||
| 1836 | /// Given a vector of values raised to various powers, where no two values are | |||
| 1837 | /// equal and the powers are sorted in decreasing order, compute the minimal | |||
| 1838 | /// DAG of multiplies to compute the final product, and return that product | |||
| 1839 | /// value. | |||
| 1840 | Value * | |||
| 1841 | ReassociatePass::buildMinimalMultiplyDAG(IRBuilderBase &Builder, | |||
| 1842 | SmallVectorImpl<Factor> &Factors) { | |||
| 1843 | assert(Factors[0].Power)(static_cast <bool> (Factors[0].Power) ? void (0) : __assert_fail ("Factors[0].Power", "llvm/lib/Transforms/Scalar/Reassociate.cpp" , 1843, __extension__ __PRETTY_FUNCTION__)); | |||
| 1844 | SmallVector<Value *, 4> OuterProduct; | |||
| 1845 | for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size(); | |||
| 1846 | Idx < Size && Factors[Idx].Power > 0; ++Idx) { | |||
| 1847 | if (Factors[Idx].Power != Factors[LastIdx].Power) { | |||
| 1848 | LastIdx = Idx; | |||
| 1849 | continue; | |||
| 1850 | } | |||
| 1851 | ||||
| 1852 | // We want to multiply across all the factors with the same power so that | |||
| 1853 | // we can raise them to that power as a single entity. Build a mini tree | |||
| 1854 | // for that. | |||
| 1855 | SmallVector<Value *, 4> InnerProduct; | |||
| 1856 | InnerProduct.push_back(Factors[LastIdx].Base); | |||
| 1857 | do { | |||
| 1858 | InnerProduct.push_back(Factors[Idx].Base); | |||
| 1859 | ++Idx; | |||
| 1860 | } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power); | |||
| 1861 | ||||
| 1862 | // Reset the base value of the first factor to the new expression tree. | |||
| 1863 | // We'll remove all the factors with the same power in a second pass. | |||
| 1864 | Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct); | |||
| 1865 | if (Instruction *MI = dyn_cast<Instruction>(M)) | |||
| 1866 | RedoInsts.insert(MI); | |||
| 1867 | ||||
| 1868 | LastIdx = Idx; | |||
| 1869 | } | |||
| 1870 | // Unique factors with equal powers -- we've folded them into the first one's | |||
| 1871 | // base. | |||
| 1872 | Factors.erase(std::unique(Factors.begin(), Factors.end(), | |||
| 1873 | [](const Factor &LHS, const Factor &RHS) { | |||
| 1874 | return LHS.Power == RHS.Power; | |||
| 1875 | }), | |||
| 1876 | Factors.end()); | |||
| 1877 | ||||
| 1878 | // Iteratively collect the base of each factor with an add power into the | |||
| 1879 | // outer product, and halve each power in preparation for squaring the | |||
| 1880 | // expression. | |||
| 1881 | for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) { | |||
| 1882 | if (Factors[Idx].Power & 1) | |||
| 1883 | OuterProduct.push_back(Factors[Idx].Base); | |||
| 1884 | Factors[Idx].Power >>= 1; | |||
| 1885 | } | |||
| 1886 | if (Factors[0].Power) { | |||
| 1887 | Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors); | |||
| 1888 | OuterProduct.push_back(SquareRoot); | |||
| 1889 | OuterProduct.push_back(SquareRoot); | |||
| 1890 | } | |||
| 1891 | if (OuterProduct.size() == 1) | |||
| 1892 | return OuterProduct.front(); | |||
| 1893 | ||||
| 1894 | Value *V = buildMultiplyTree(Builder, OuterProduct); | |||
| 1895 | return V; | |||
| 1896 | } | |||
| 1897 | ||||
| 1898 | Value *ReassociatePass::OptimizeMul(BinaryOperator *I, | |||
| 1899 | SmallVectorImpl<ValueEntry> &Ops) { | |||
| 1900 | // We can only optimize the multiplies when there is a chain of more than | |||
| 1901 | // three, such that a balanced tree might require fewer total multiplies. | |||
| 1902 | if (Ops.size() < 4) | |||
| 1903 | return nullptr; | |||
| 1904 | ||||
| 1905 | // Try to turn linear trees of multiplies without other uses of the | |||
| 1906 | // intermediate stages into minimal multiply DAGs with perfect sub-expression | |||
| 1907 | // re-use. | |||
| 1908 | SmallVector<Factor, 4> Factors; | |||
| 1909 | if (!collectMultiplyFactors(Ops, Factors)) | |||
| 1910 | return nullptr; // All distinct factors, so nothing left for us to do. | |||
| 1911 | ||||
| 1912 | IRBuilder<> Builder(I); | |||
| 1913 | // The reassociate transformation for FP operations is performed only | |||
| 1914 | // if unsafe algebra is permitted by FastMathFlags. Propagate those flags | |||
| 1915 | // to the newly generated operations. | |||
| 1916 | if (auto FPI = dyn_cast<FPMathOperator>(I)) | |||
| 1917 | Builder.setFastMathFlags(FPI->getFastMathFlags()); | |||
| 1918 | ||||
| 1919 | Value *V = buildMinimalMultiplyDAG(Builder, Factors); | |||
| 1920 | if (Ops.empty()) | |||
| 1921 | return V; | |||
| 1922 | ||||
| 1923 | ValueEntry NewEntry = ValueEntry(getRank(V), V); | |||
| 1924 | Ops.insert(llvm::lower_bound(Ops, NewEntry), NewEntry); | |||
| 1925 | return nullptr; | |||
| 1926 | } | |||
| 1927 | ||||
| 1928 | Value *ReassociatePass::OptimizeExpression(BinaryOperator *I, | |||
| 1929 | SmallVectorImpl<ValueEntry> &Ops) { | |||
| 1930 | // Now that we have the linearized expression tree, try to optimize it. | |||
| 1931 | // Start by folding any constants that we found. | |||
| 1932 | Constant *Cst = nullptr; | |||
| 1933 | unsigned Opcode = I->getOpcode(); | |||
| 1934 | while (!Ops.empty() && isa<Constant>(Ops.back().Op)) { | |||
| 1935 | Constant *C = cast<Constant>(Ops.pop_back_val().Op); | |||
| 1936 | Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C; | |||
| 1937 | } | |||
| 1938 | // If there was nothing but constants then we are done. | |||
| 1939 | if (Ops.empty()) | |||
| 1940 | return Cst; | |||
| 1941 | ||||
| 1942 | // Put the combined constant back at the end of the operand list, except if | |||
| 1943 | // there is no point. For example, an add of 0 gets dropped here, while a | |||
| 1944 | // multiplication by zero turns the whole expression into zero. | |||
| 1945 | if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) { | |||
| 1946 | if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType())) | |||
| 1947 | return Cst; | |||
| 1948 | Ops.push_back(ValueEntry(0, Cst)); | |||
| 1949 | } | |||
| 1950 | ||||
| 1951 | if (Ops.size() == 1) return Ops[0].Op; | |||
| 1952 | ||||
| 1953 | // Handle destructive annihilation due to identities between elements in the | |||
| 1954 | // argument list here. | |||
| 1955 | unsigned NumOps = Ops.size(); | |||
| 1956 | switch (Opcode) { | |||
| 1957 | default: break; | |||
| 1958 | case Instruction::And: | |||
| 1959 | case Instruction::Or: | |||
| 1960 | if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) | |||
| 1961 | return Result; | |||
| 1962 | break; | |||
| 1963 | ||||
| 1964 | case Instruction::Xor: | |||
| 1965 | if (Value *Result = OptimizeXor(I, Ops)) | |||
| 1966 | return Result; | |||
| 1967 | break; | |||
| 1968 | ||||
| 1969 | case Instruction::Add: | |||
| 1970 | case Instruction::FAdd: | |||
| 1971 | if (Value *Result = OptimizeAdd(I, Ops)) | |||
| 1972 | return Result; | |||
| 1973 | break; | |||
| 1974 | ||||
| 1975 | case Instruction::Mul: | |||
| 1976 | case Instruction::FMul: | |||
| 1977 | if (Value *Result = OptimizeMul(I, Ops)) | |||
| 1978 | return Result; | |||
| 1979 | break; | |||
| 1980 | } | |||
| 1981 | ||||
| 1982 | if (Ops.size() != NumOps) | |||
| 1983 | return OptimizeExpression(I, Ops); | |||
| 1984 | return nullptr; | |||
| 1985 | } | |||
| 1986 | ||||
| 1987 | // Remove dead instructions and if any operands are trivially dead add them to | |||
| 1988 | // Insts so they will be removed as well. | |||
| 1989 | void ReassociatePass::RecursivelyEraseDeadInsts(Instruction *I, | |||
| 1990 | OrderedSet &Insts) { | |||
| 1991 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 1991, __extension__ __PRETTY_FUNCTION__)); | |||
| 1992 | SmallVector<Value *, 4> Ops(I->operands()); | |||
| 1993 | ValueRankMap.erase(I); | |||
| 1994 | Insts.remove(I); | |||
| 1995 | RedoInsts.remove(I); | |||
| 1996 | llvm::salvageDebugInfo(*I); | |||
| 1997 | I->eraseFromParent(); | |||
| 1998 | for (auto Op : Ops) | |||
| 1999 | if (Instruction *OpInst = dyn_cast<Instruction>(Op)) | |||
| 2000 | if (OpInst->use_empty()) | |||
| 2001 | Insts.insert(OpInst); | |||
| 2002 | } | |||
| 2003 | ||||
| 2004 | /// Zap the given instruction, adding interesting operands to the work list. | |||
| 2005 | void ReassociatePass::EraseInst(Instruction *I) { | |||
| 2006 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2006, __extension__ __PRETTY_FUNCTION__)); | |||
| 2007 | LLVM_DEBUG(dbgs() << "Erasing dead inst: "; I->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Erasing dead inst: "; I-> dump(); } } while (false); | |||
| 2008 | ||||
| 2009 | SmallVector<Value *, 8> Ops(I->operands()); | |||
| 2010 | // Erase the dead instruction. | |||
| 2011 | ValueRankMap.erase(I); | |||
| 2012 | RedoInsts.remove(I); | |||
| 2013 | llvm::salvageDebugInfo(*I); | |||
| 2014 | I->eraseFromParent(); | |||
| 2015 | // Optimize its operands. | |||
| 2016 | SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes. | |||
| 2017 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) | |||
| 2018 | if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) { | |||
| 2019 | // If this is a node in an expression tree, climb to the expression root | |||
| 2020 | // and add that since that's where optimization actually happens. | |||
| 2021 | unsigned Opcode = Op->getOpcode(); | |||
| 2022 | while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode && | |||
| 2023 | Visited.insert(Op).second) | |||
| 2024 | Op = Op->user_back(); | |||
| 2025 | ||||
| 2026 | // The instruction we're going to push may be coming from a | |||
| 2027 | // dead block, and Reassociate skips the processing of unreachable | |||
| 2028 | // blocks because it's a waste of time and also because it can | |||
| 2029 | // lead to infinite loop due to LLVM's non-standard definition | |||
| 2030 | // of dominance. | |||
| 2031 | if (ValueRankMap.find(Op) != ValueRankMap.end()) | |||
| 2032 | RedoInsts.insert(Op); | |||
| 2033 | } | |||
| 2034 | ||||
| 2035 | MadeChange = true; | |||
| 2036 | } | |||
| 2037 | ||||
| 2038 | /// Recursively analyze an expression to build a list of instructions that have | |||
| 2039 | /// negative floating-point constant operands. The caller can then transform | |||
| 2040 | /// the list to create positive constants for better reassociation and CSE. | |||
| 2041 | static void getNegatibleInsts(Value *V, | |||
| 2042 | SmallVectorImpl<Instruction *> &Candidates) { | |||
| 2043 | // Handle only one-use instructions. Combining negations does not justify | |||
| 2044 | // replicating instructions. | |||
| 2045 | Instruction *I; | |||
| 2046 | if (!match(V, m_OneUse(m_Instruction(I)))) | |||
| 2047 | return; | |||
| 2048 | ||||
| 2049 | // Handle expressions of multiplications and divisions. | |||
| 2050 | // TODO: This could look through floating-point casts. | |||
| 2051 | const APFloat *C; | |||
| 2052 | switch (I->getOpcode()) { | |||
| 2053 | case Instruction::FMul: | |||
| 2054 | // Not expecting non-canonical code here. Bail out and wait. | |||
| 2055 | if (match(I->getOperand(0), m_Constant())) | |||
| 2056 | break; | |||
| 2057 | ||||
| 2058 | if (match(I->getOperand(1), m_APFloat(C)) && C->isNegative()) { | |||
| 2059 | Candidates.push_back(I); | |||
| 2060 | LLVM_DEBUG(dbgs() << "FMul with negative constant: " << *I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "FMul with negative constant: " << *I << '\n'; } } while (false); | |||
| 2061 | } | |||
| 2062 | getNegatibleInsts(I->getOperand(0), Candidates); | |||
| 2063 | getNegatibleInsts(I->getOperand(1), Candidates); | |||
| 2064 | break; | |||
| 2065 | case Instruction::FDiv: | |||
| 2066 | // Not expecting non-canonical code here. Bail out and wait. | |||
| 2067 | if (match(I->getOperand(0), m_Constant()) && | |||
| 2068 | match(I->getOperand(1), m_Constant())) | |||
| 2069 | break; | |||
| 2070 | ||||
| 2071 | if ((match(I->getOperand(0), m_APFloat(C)) && C->isNegative()) || | |||
| 2072 | (match(I->getOperand(1), m_APFloat(C)) && C->isNegative())) { | |||
| 2073 | Candidates.push_back(I); | |||
| 2074 | LLVM_DEBUG(dbgs() << "FDiv with negative constant: " << *I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "FDiv with negative constant: " << *I << '\n'; } } while (false); | |||
| 2075 | } | |||
| 2076 | getNegatibleInsts(I->getOperand(0), Candidates); | |||
| 2077 | getNegatibleInsts(I->getOperand(1), Candidates); | |||
| 2078 | break; | |||
| 2079 | default: | |||
| 2080 | break; | |||
| 2081 | } | |||
| 2082 | } | |||
| 2083 | ||||
| 2084 | /// Given an fadd/fsub with an operand that is a one-use instruction | |||
| 2085 | /// (the fadd/fsub), try to change negative floating-point constants into | |||
| 2086 | /// positive constants to increase potential for reassociation and CSE. | |||
| 2087 | Instruction *ReassociatePass::canonicalizeNegFPConstantsForOp(Instruction *I, | |||
| 2088 | Instruction *Op, | |||
| 2089 | Value *OtherOp) { | |||
| 2090 | assert((I->getOpcode() == Instruction::FAdd ||(static_cast <bool> ((I->getOpcode() == Instruction:: FAdd || I->getOpcode() == Instruction::FSub) && "Expected fadd/fsub" ) ? void (0) : __assert_fail ("(I->getOpcode() == Instruction::FAdd || I->getOpcode() == Instruction::FSub) && \"Expected fadd/fsub\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2091, __extension__ __PRETTY_FUNCTION__)) | |||
| 2091 | I->getOpcode() == Instruction::FSub) && "Expected fadd/fsub")(static_cast <bool> ((I->getOpcode() == Instruction:: FAdd || I->getOpcode() == Instruction::FSub) && "Expected fadd/fsub" ) ? void (0) : __assert_fail ("(I->getOpcode() == Instruction::FAdd || I->getOpcode() == Instruction::FSub) && \"Expected fadd/fsub\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2091, __extension__ __PRETTY_FUNCTION__)); | |||
| 2092 | ||||
| 2093 | // Collect instructions with negative FP constants from the subtree that ends | |||
| 2094 | // in Op. | |||
| 2095 | SmallVector<Instruction *, 4> Candidates; | |||
| 2096 | getNegatibleInsts(Op, Candidates); | |||
| 2097 | if (Candidates.empty()) | |||
| 2098 | return nullptr; | |||
| 2099 | ||||
| 2100 | // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the | |||
| 2101 | // resulting subtract will be broken up later. This can get us into an | |||
| 2102 | // infinite loop during reassociation. | |||
| 2103 | bool IsFSub = I->getOpcode() == Instruction::FSub; | |||
| 2104 | bool NeedsSubtract = !IsFSub && Candidates.size() % 2 == 1; | |||
| 2105 | if (NeedsSubtract && ShouldBreakUpSubtract(I)) | |||
| 2106 | return nullptr; | |||
| 2107 | ||||
| 2108 | for (Instruction *Negatible : Candidates) { | |||
| 2109 | const APFloat *C; | |||
| 2110 | if (match(Negatible->getOperand(0), m_APFloat(C))) { | |||
| 2111 | assert(!match(Negatible->getOperand(1), m_Constant()) &&(static_cast <bool> (!match(Negatible->getOperand(1) , m_Constant()) && "Expecting only 1 constant operand" ) ? void (0) : __assert_fail ("!match(Negatible->getOperand(1), m_Constant()) && \"Expecting only 1 constant operand\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2112, __extension__ __PRETTY_FUNCTION__)) | |||
| 2112 | "Expecting only 1 constant operand")(static_cast <bool> (!match(Negatible->getOperand(1) , m_Constant()) && "Expecting only 1 constant operand" ) ? void (0) : __assert_fail ("!match(Negatible->getOperand(1), m_Constant()) && \"Expecting only 1 constant operand\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2112, __extension__ __PRETTY_FUNCTION__)); | |||
| 2113 | assert(C->isNegative() && "Expected negative FP constant")(static_cast <bool> (C->isNegative() && "Expected negative FP constant" ) ? void (0) : __assert_fail ("C->isNegative() && \"Expected negative FP constant\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2113, __extension__ __PRETTY_FUNCTION__)); | |||
| 2114 | Negatible->setOperand(0, ConstantFP::get(Negatible->getType(), abs(*C))); | |||
| 2115 | MadeChange = true; | |||
| 2116 | } | |||
| 2117 | if (match(Negatible->getOperand(1), m_APFloat(C))) { | |||
| 2118 | assert(!match(Negatible->getOperand(0), m_Constant()) &&(static_cast <bool> (!match(Negatible->getOperand(0) , m_Constant()) && "Expecting only 1 constant operand" ) ? void (0) : __assert_fail ("!match(Negatible->getOperand(0), m_Constant()) && \"Expecting only 1 constant operand\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2119, __extension__ __PRETTY_FUNCTION__)) | |||
| 2119 | "Expecting only 1 constant operand")(static_cast <bool> (!match(Negatible->getOperand(0) , m_Constant()) && "Expecting only 1 constant operand" ) ? void (0) : __assert_fail ("!match(Negatible->getOperand(0), m_Constant()) && \"Expecting only 1 constant operand\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2119, __extension__ __PRETTY_FUNCTION__)); | |||
| 2120 | assert(C->isNegative() && "Expected negative FP constant")(static_cast <bool> (C->isNegative() && "Expected negative FP constant" ) ? void (0) : __assert_fail ("C->isNegative() && \"Expected negative FP constant\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2120, __extension__ __PRETTY_FUNCTION__)); | |||
| 2121 | Negatible->setOperand(1, ConstantFP::get(Negatible->getType(), abs(*C))); | |||
| 2122 | MadeChange = true; | |||
| 2123 | } | |||
| 2124 | } | |||
| 2125 | assert(MadeChange == true && "Negative constant candidate was not changed")(static_cast <bool> (MadeChange == true && "Negative constant candidate was not changed" ) ? void (0) : __assert_fail ("MadeChange == true && \"Negative constant candidate was not changed\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2125, __extension__ __PRETTY_FUNCTION__)); | |||
| 2126 | ||||
| 2127 | // Negations cancelled out. | |||
| 2128 | if (Candidates.size() % 2 == 0) | |||
| 2129 | return I; | |||
| 2130 | ||||
| 2131 | // Negate the final operand in the expression by flipping the opcode of this | |||
| 2132 | // fadd/fsub. | |||
| 2133 | assert(Candidates.size() % 2 == 1 && "Expected odd number")(static_cast <bool> (Candidates.size() % 2 == 1 && "Expected odd number") ? void (0) : __assert_fail ("Candidates.size() % 2 == 1 && \"Expected odd number\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2133, __extension__ __PRETTY_FUNCTION__)); | |||
| 2134 | IRBuilder<> Builder(I); | |||
| 2135 | Value *NewInst = IsFSub ? Builder.CreateFAddFMF(OtherOp, Op, I) | |||
| 2136 | : Builder.CreateFSubFMF(OtherOp, Op, I); | |||
| 2137 | I->replaceAllUsesWith(NewInst); | |||
| 2138 | RedoInsts.insert(I); | |||
| 2139 | return dyn_cast<Instruction>(NewInst); | |||
| 2140 | } | |||
| 2141 | ||||
| 2142 | /// Canonicalize expressions that contain a negative floating-point constant | |||
| 2143 | /// of the following form: | |||
| 2144 | /// OtherOp + (subtree) -> OtherOp {+/-} (canonical subtree) | |||
| 2145 | /// (subtree) + OtherOp -> OtherOp {+/-} (canonical subtree) | |||
| 2146 | /// OtherOp - (subtree) -> OtherOp {+/-} (canonical subtree) | |||
| 2147 | /// | |||
| 2148 | /// The fadd/fsub opcode may be switched to allow folding a negation into the | |||
| 2149 | /// input instruction. | |||
| 2150 | Instruction *ReassociatePass::canonicalizeNegFPConstants(Instruction *I) { | |||
| 2151 | LLVM_DEBUG(dbgs() << "Combine negations for: " << *I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Combine negations for: " << *I << '\n'; } } while (false); | |||
| 2152 | Value *X; | |||
| 2153 | Instruction *Op; | |||
| 2154 | if (match(I, m_FAdd(m_Value(X), m_OneUse(m_Instruction(Op))))) | |||
| 2155 | if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X)) | |||
| 2156 | I = R; | |||
| 2157 | if (match(I, m_FAdd(m_OneUse(m_Instruction(Op)), m_Value(X)))) | |||
| 2158 | if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X)) | |||
| 2159 | I = R; | |||
| 2160 | if (match(I, m_FSub(m_Value(X), m_OneUse(m_Instruction(Op))))) | |||
| 2161 | if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X)) | |||
| 2162 | I = R; | |||
| 2163 | return I; | |||
| 2164 | } | |||
| 2165 | ||||
| 2166 | /// Inspect and optimize the given instruction. Note that erasing | |||
| 2167 | /// instructions is not allowed. | |||
| 2168 | void ReassociatePass::OptimizeInst(Instruction *I) { | |||
| 2169 | // Only consider operations that we understand. | |||
| 2170 | if (!isa<UnaryOperator>(I) && !isa<BinaryOperator>(I)) | |||
| 2171 | return; | |||
| 2172 | ||||
| 2173 | if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1))) | |||
| 2174 | // If an operand of this shift is a reassociable multiply, or if the shift | |||
| 2175 | // is used by a reassociable multiply or add, turn into a multiply. | |||
| 2176 | if (isReassociableOp(I->getOperand(0), Instruction::Mul) || | |||
| 2177 | (I->hasOneUse() && | |||
| 2178 | (isReassociableOp(I->user_back(), Instruction::Mul) || | |||
| 2179 | isReassociableOp(I->user_back(), Instruction::Add)))) { | |||
| 2180 | Instruction *NI = ConvertShiftToMul(I); | |||
| 2181 | RedoInsts.insert(I); | |||
| 2182 | MadeChange = true; | |||
| 2183 | I = NI; | |||
| 2184 | } | |||
| 2185 | ||||
| 2186 | // Commute binary operators, to canonicalize the order of their operands. | |||
| 2187 | // This can potentially expose more CSE opportunities, and makes writing other | |||
| 2188 | // transformations simpler. | |||
| 2189 | if (I->isCommutative()) | |||
| 2190 | canonicalizeOperands(I); | |||
| 2191 | ||||
| 2192 | // Canonicalize negative constants out of expressions. | |||
| 2193 | if (Instruction *Res = canonicalizeNegFPConstants(I)) | |||
| 2194 | I = Res; | |||
| 2195 | ||||
| 2196 | // Don't optimize floating-point instructions unless they are 'fast'. | |||
| 2197 | if (I->getType()->isFPOrFPVectorTy() && !I->isFast()) | |||
| 2198 | return; | |||
| 2199 | ||||
| 2200 | // Do not reassociate boolean (i1) expressions. We want to preserve the | |||
| 2201 | // original order of evaluation for short-circuited comparisons that | |||
| 2202 | // SimplifyCFG has folded to AND/OR expressions. If the expression | |||
| 2203 | // is not further optimized, it is likely to be transformed back to a | |||
| 2204 | // short-circuited form for code gen, and the source order may have been | |||
| 2205 | // optimized for the most likely conditions. | |||
| 2206 | if (I->getType()->isIntegerTy(1)) | |||
| 2207 | return; | |||
| 2208 | ||||
| 2209 | // If this is a bitwise or instruction of operands | |||
| 2210 | // with no common bits set, convert it to X+Y. | |||
| 2211 | if (I->getOpcode() == Instruction::Or && | |||
| 2212 | shouldConvertOrWithNoCommonBitsToAdd(I) && !isLoadCombineCandidate(I) && | |||
| 2213 | haveNoCommonBitsSet(I->getOperand(0), I->getOperand(1), | |||
| 2214 | I->getModule()->getDataLayout(), /*AC=*/nullptr, I, | |||
| 2215 | /*DT=*/nullptr)) { | |||
| 2216 | Instruction *NI = convertOrWithNoCommonBitsToAdd(I); | |||
| 2217 | RedoInsts.insert(I); | |||
| 2218 | MadeChange = true; | |||
| 2219 | I = NI; | |||
| 2220 | } | |||
| 2221 | ||||
| 2222 | // If this is a subtract instruction which is not already in negate form, | |||
| 2223 | // see if we can convert it to X+-Y. | |||
| 2224 | if (I->getOpcode() == Instruction::Sub) { | |||
| 2225 | if (ShouldBreakUpSubtract(I)) { | |||
| 2226 | Instruction *NI = BreakUpSubtract(I, RedoInsts); | |||
| 2227 | RedoInsts.insert(I); | |||
| 2228 | MadeChange = true; | |||
| 2229 | I = NI; | |||
| 2230 | } else if (match(I, m_Neg(m_Value()))) { | |||
| 2231 | // Otherwise, this is a negation. See if the operand is a multiply tree | |||
| 2232 | // and if this is not an inner node of a multiply tree. | |||
| 2233 | if (isReassociableOp(I->getOperand(1), Instruction::Mul) && | |||
| 2234 | (!I->hasOneUse() || | |||
| 2235 | !isReassociableOp(I->user_back(), Instruction::Mul))) { | |||
| 2236 | Instruction *NI = LowerNegateToMultiply(I); | |||
| 2237 | // If the negate was simplified, revisit the users to see if we can | |||
| 2238 | // reassociate further. | |||
| 2239 | for (User *U : NI->users()) { | |||
| 2240 | if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U)) | |||
| 2241 | RedoInsts.insert(Tmp); | |||
| 2242 | } | |||
| 2243 | RedoInsts.insert(I); | |||
| 2244 | MadeChange = true; | |||
| 2245 | I = NI; | |||
| 2246 | } | |||
| 2247 | } | |||
| 2248 | } else if (I->getOpcode() == Instruction::FNeg || | |||
| 2249 | I->getOpcode() == Instruction::FSub) { | |||
| 2250 | if (ShouldBreakUpSubtract(I)) { | |||
| 2251 | Instruction *NI = BreakUpSubtract(I, RedoInsts); | |||
| 2252 | RedoInsts.insert(I); | |||
| 2253 | MadeChange = true; | |||
| 2254 | I = NI; | |||
| 2255 | } else if (match(I, m_FNeg(m_Value()))) { | |||
| 2256 | // Otherwise, this is a negation. See if the operand is a multiply tree | |||
| 2257 | // and if this is not an inner node of a multiply tree. | |||
| 2258 | Value *Op = isa<BinaryOperator>(I) ? I->getOperand(1) : | |||
| 2259 | I->getOperand(0); | |||
| 2260 | if (isReassociableOp(Op, Instruction::FMul) && | |||
| 2261 | (!I->hasOneUse() || | |||
| 2262 | !isReassociableOp(I->user_back(), Instruction::FMul))) { | |||
| 2263 | // If the negate was simplified, revisit the users to see if we can | |||
| 2264 | // reassociate further. | |||
| 2265 | Instruction *NI = LowerNegateToMultiply(I); | |||
| 2266 | for (User *U : NI->users()) { | |||
| 2267 | if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U)) | |||
| 2268 | RedoInsts.insert(Tmp); | |||
| 2269 | } | |||
| 2270 | RedoInsts.insert(I); | |||
| 2271 | MadeChange = true; | |||
| 2272 | I = NI; | |||
| 2273 | } | |||
| 2274 | } | |||
| 2275 | } | |||
| 2276 | ||||
| 2277 | // If this instruction is an associative binary operator, process it. | |||
| 2278 | if (!I->isAssociative()) return; | |||
| 2279 | BinaryOperator *BO = cast<BinaryOperator>(I); | |||
| 2280 | ||||
| 2281 | // If this is an interior node of a reassociable tree, ignore it until we | |||
| 2282 | // get to the root of the tree, to avoid N^2 analysis. | |||
| 2283 | unsigned Opcode = BO->getOpcode(); | |||
| 2284 | if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) { | |||
| 2285 | // During the initial run we will get to the root of the tree. | |||
| 2286 | // But if we get here while we are redoing instructions, there is no | |||
| 2287 | // guarantee that the root will be visited. So Redo later | |||
| 2288 | if (BO->user_back() != BO && | |||
| 2289 | BO->getParent() == BO->user_back()->getParent()) | |||
| 2290 | RedoInsts.insert(BO->user_back()); | |||
| 2291 | return; | |||
| 2292 | } | |||
| 2293 | ||||
| 2294 | // If this is an add tree that is used by a sub instruction, ignore it | |||
| 2295 | // until we process the subtract. | |||
| 2296 | if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add && | |||
| 2297 | cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub) | |||
| 2298 | return; | |||
| 2299 | if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd && | |||
| 2300 | cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub) | |||
| 2301 | return; | |||
| 2302 | ||||
| 2303 | ReassociateExpression(BO); | |||
| 2304 | } | |||
| 2305 | ||||
| 2306 | void ReassociatePass::ReassociateExpression(BinaryOperator *I) { | |||
| 2307 | // First, walk the expression tree, linearizing the tree, collecting the | |||
| 2308 | // operand information. | |||
| 2309 | SmallVector<RepeatedValue, 8> Tree; | |||
| 2310 | MadeChange |= LinearizeExprTree(I, Tree); | |||
| 2311 | SmallVector<ValueEntry, 8> Ops; | |||
| 2312 | Ops.reserve(Tree.size()); | |||
| 2313 | for (const RepeatedValue &E : Tree) | |||
| 2314 | Ops.append(E.second.getZExtValue(), ValueEntry(getRank(E.first), E.first)); | |||
| 2315 | ||||
| 2316 | LLVM_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); | |||
| 2317 | ||||
| 2318 | // Now that we have linearized the tree to a list and have gathered all of | |||
| 2319 | // the operands and their ranks, sort the operands by their rank. Use a | |||
| 2320 | // stable_sort so that values with equal ranks will have their relative | |||
| 2321 | // positions maintained (and so the compiler is deterministic). Note that | |||
| 2322 | // this sorts so that the highest ranking values end up at the beginning of | |||
| 2323 | // the vector. | |||
| 2324 | llvm::stable_sort(Ops); | |||
| 2325 | ||||
| 2326 | // Now that we have the expression tree in a convenient | |||
| 2327 | // sorted form, optimize it globally if possible. | |||
| 2328 | if (Value *V = OptimizeExpression(I, Ops)) { | |||
| 2329 | if (V == I) | |||
| 2330 | // Self-referential expression in unreachable code. | |||
| 2331 | return; | |||
| 2332 | // This expression tree simplified to something that isn't a tree, | |||
| 2333 | // eliminate it. | |||
| 2334 | LLVM_DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("reassociate")) { dbgs() << "Reassoc to scalar: " << *V << '\n'; } } while (false); | |||
| 2335 | I->replaceAllUsesWith(V); | |||
| 2336 | if (Instruction *VI = dyn_cast<Instruction>(V)) | |||
| 2337 | if (I->getDebugLoc()) | |||
| 2338 | VI->setDebugLoc(I->getDebugLoc()); | |||
| 2339 | RedoInsts.insert(I); | |||
| 2340 | ++NumAnnihil; | |||
| 2341 | return; | |||
| 2342 | } | |||
| 2343 | ||||
| 2344 | // We want to sink immediates as deeply as possible except in the case where | |||
| 2345 | // this is a multiply tree used only by an add, and the immediate is a -1. | |||
| 2346 | // In this case we reassociate to put the negation on the outside so that we | |||
| 2347 | // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y | |||
| 2348 | if (I->hasOneUse()) { | |||
| 2349 | if (I->getOpcode() == Instruction::Mul && | |||
| 2350 | cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add && | |||
| 2351 | isa<ConstantInt>(Ops.back().Op) && | |||
| 2352 | cast<ConstantInt>(Ops.back().Op)->isMinusOne()) { | |||
| 2353 | ValueEntry Tmp = Ops.pop_back_val(); | |||
| 2354 | Ops.insert(Ops.begin(), Tmp); | |||
| 2355 | } else if (I->getOpcode() == Instruction::FMul && | |||
| 2356 | cast<Instruction>(I->user_back())->getOpcode() == | |||
| 2357 | Instruction::FAdd && | |||
| 2358 | isa<ConstantFP>(Ops.back().Op) && | |||
| 2359 | cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) { | |||
| 2360 | ValueEntry Tmp = Ops.pop_back_val(); | |||
| 2361 | Ops.insert(Ops.begin(), Tmp); | |||
| 2362 | } | |||
| 2363 | } | |||
| 2364 | ||||
| 2365 | LLVM_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); | |||
| 2366 | ||||
| 2367 | if (Ops.size() == 1) { | |||
| 2368 | if (Ops[0].Op == I) | |||
| 2369 | // Self-referential expression in unreachable code. | |||
| 2370 | return; | |||
| 2371 | ||||
| 2372 | // This expression tree simplified to something that isn't a tree, | |||
| 2373 | // eliminate it. | |||
| 2374 | I->replaceAllUsesWith(Ops[0].Op); | |||
| 2375 | if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op)) | |||
| 2376 | OI->setDebugLoc(I->getDebugLoc()); | |||
| 2377 | RedoInsts.insert(I); | |||
| 2378 | return; | |||
| 2379 | } | |||
| 2380 | ||||
| 2381 | if (Ops.size() > 2 && Ops.size() <= GlobalReassociateLimit) { | |||
| 2382 | // Find the pair with the highest count in the pairmap and move it to the | |||
| 2383 | // back of the list so that it can later be CSE'd. | |||
| 2384 | // example: | |||
| 2385 | // a*b*c*d*e | |||
| 2386 | // if c*e is the most "popular" pair, we can express this as | |||
| 2387 | // (((c*e)*d)*b)*a | |||
| 2388 | unsigned Max = 1; | |||
| 2389 | unsigned BestRank = 0; | |||
| 2390 | std::pair<unsigned, unsigned> BestPair; | |||
| 2391 | unsigned Idx = I->getOpcode() - Instruction::BinaryOpsBegin; | |||
| 2392 | for (unsigned i = 0; i < Ops.size() - 1; ++i) | |||
| 2393 | for (unsigned j = i + 1; j < Ops.size(); ++j) { | |||
| 2394 | unsigned Score = 0; | |||
| 2395 | Value *Op0 = Ops[i].Op; | |||
| 2396 | Value *Op1 = Ops[j].Op; | |||
| 2397 | if (std::less<Value *>()(Op1, Op0)) | |||
| 2398 | std::swap(Op0, Op1); | |||
| 2399 | auto it = PairMap[Idx].find({Op0, Op1}); | |||
| 2400 | if (it != PairMap[Idx].end()) { | |||
| 2401 | // Functions like BreakUpSubtract() can erase the Values we're using | |||
| 2402 | // as keys and create new Values after we built the PairMap. There's a | |||
| 2403 | // small chance that the new nodes can have the same address as | |||
| 2404 | // something already in the table. We shouldn't accumulate the stored | |||
| 2405 | // score in that case as it refers to the wrong Value. | |||
| 2406 | if (it->second.isValid()) | |||
| 2407 | Score += it->second.Score; | |||
| 2408 | } | |||
| 2409 | ||||
| 2410 | unsigned MaxRank = std::max(Ops[i].Rank, Ops[j].Rank); | |||
| 2411 | if (Score > Max || (Score == Max && MaxRank < BestRank)) { | |||
| 2412 | BestPair = {i, j}; | |||
| 2413 | Max = Score; | |||
| 2414 | BestRank = MaxRank; | |||
| 2415 | } | |||
| 2416 | } | |||
| 2417 | if (Max > 1) { | |||
| 2418 | auto Op0 = Ops[BestPair.first]; | |||
| 2419 | auto Op1 = Ops[BestPair.second]; | |||
| 2420 | Ops.erase(&Ops[BestPair.second]); | |||
| 2421 | Ops.erase(&Ops[BestPair.first]); | |||
| 2422 | Ops.push_back(Op0); | |||
| 2423 | Ops.push_back(Op1); | |||
| 2424 | } | |||
| 2425 | } | |||
| 2426 | // Now that we ordered and optimized the expressions, splat them back into | |||
| 2427 | // the expression tree, removing any unneeded nodes. | |||
| 2428 | RewriteExprTree(I, Ops); | |||
| 2429 | } | |||
| 2430 | ||||
| 2431 | void | |||
| 2432 | ReassociatePass::BuildPairMap(ReversePostOrderTraversal<Function *> &RPOT) { | |||
| 2433 | // Make a "pairmap" of how often each operand pair occurs. | |||
| 2434 | for (BasicBlock *BI : RPOT) { | |||
| 2435 | for (Instruction &I : *BI) { | |||
| 2436 | if (!I.isAssociative()) | |||
| 2437 | continue; | |||
| 2438 | ||||
| 2439 | // Ignore nodes that aren't at the root of trees. | |||
| 2440 | if (I.hasOneUse() && I.user_back()->getOpcode() == I.getOpcode()) | |||
| 2441 | continue; | |||
| 2442 | ||||
| 2443 | // Collect all operands in a single reassociable expression. | |||
| 2444 | // Since Reassociate has already been run once, we can assume things | |||
| 2445 | // are already canonical according to Reassociation's regime. | |||
| 2446 | SmallVector<Value *, 8> Worklist = { I.getOperand(0), I.getOperand(1) }; | |||
| 2447 | SmallVector<Value *, 8> Ops; | |||
| 2448 | while (!Worklist.empty() && Ops.size() <= GlobalReassociateLimit) { | |||
| 2449 | Value *Op = Worklist.pop_back_val(); | |||
| 2450 | Instruction *OpI = dyn_cast<Instruction>(Op); | |||
| 2451 | if (!OpI || OpI->getOpcode() != I.getOpcode() || !OpI->hasOneUse()) { | |||
| 2452 | Ops.push_back(Op); | |||
| 2453 | continue; | |||
| 2454 | } | |||
| 2455 | // Be paranoid about self-referencing expressions in unreachable code. | |||
| 2456 | if (OpI->getOperand(0) != OpI) | |||
| 2457 | Worklist.push_back(OpI->getOperand(0)); | |||
| 2458 | if (OpI->getOperand(1) != OpI) | |||
| 2459 | Worklist.push_back(OpI->getOperand(1)); | |||
| 2460 | } | |||
| 2461 | // Skip extremely long expressions. | |||
| 2462 | if (Ops.size() > GlobalReassociateLimit) | |||
| 2463 | continue; | |||
| 2464 | ||||
| 2465 | // Add all pairwise combinations of operands to the pair map. | |||
| 2466 | unsigned BinaryIdx = I.getOpcode() - Instruction::BinaryOpsBegin; | |||
| 2467 | SmallSet<std::pair<Value *, Value*>, 32> Visited; | |||
| 2468 | for (unsigned i = 0; i < Ops.size() - 1; ++i) { | |||
| 2469 | for (unsigned j = i + 1; j < Ops.size(); ++j) { | |||
| 2470 | // Canonicalize operand orderings. | |||
| 2471 | Value *Op0 = Ops[i]; | |||
| 2472 | Value *Op1 = Ops[j]; | |||
| 2473 | if (std::less<Value *>()(Op1, Op0)) | |||
| 2474 | std::swap(Op0, Op1); | |||
| 2475 | if (!Visited.insert({Op0, Op1}).second) | |||
| 2476 | continue; | |||
| 2477 | auto res = PairMap[BinaryIdx].insert({{Op0, Op1}, {Op0, Op1, 1}}); | |||
| 2478 | if (!res.second) { | |||
| 2479 | // If either key value has been erased then we've got the same | |||
| 2480 | // address by coincidence. That can't happen here because nothing is | |||
| 2481 | // erasing values but it can happen by the time we're querying the | |||
| 2482 | // map. | |||
| 2483 | assert(res.first->second.isValid() && "WeakVH invalidated")(static_cast <bool> (res.first->second.isValid() && "WeakVH invalidated") ? void (0) : __assert_fail ("res.first->second.isValid() && \"WeakVH invalidated\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2483, __extension__ __PRETTY_FUNCTION__)); | |||
| 2484 | ++res.first->second.Score; | |||
| 2485 | } | |||
| 2486 | } | |||
| 2487 | } | |||
| 2488 | } | |||
| 2489 | } | |||
| 2490 | } | |||
| 2491 | ||||
| 2492 | PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) { | |||
| 2493 | // Get the functions basic blocks in Reverse Post Order. This order is used by | |||
| 2494 | // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic | |||
| 2495 | // blocks (it has been seen that the analysis in this pass could hang when | |||
| 2496 | // analysing dead basic blocks). | |||
| 2497 | ReversePostOrderTraversal<Function *> RPOT(&F); | |||
| 2498 | ||||
| 2499 | // Calculate the rank map for F. | |||
| 2500 | BuildRankMap(F, RPOT); | |||
| 2501 | ||||
| 2502 | // Build the pair map before running reassociate. | |||
| 2503 | // Technically this would be more accurate if we did it after one round | |||
| 2504 | // of reassociation, but in practice it doesn't seem to help much on | |||
| 2505 | // real-world code, so don't waste the compile time running reassociate | |||
| 2506 | // twice. | |||
| 2507 | // If a user wants, they could expicitly run reassociate twice in their | |||
| 2508 | // pass pipeline for further potential gains. | |||
| 2509 | // It might also be possible to update the pair map during runtime, but the | |||
| 2510 | // overhead of that may be large if there's many reassociable chains. | |||
| 2511 | BuildPairMap(RPOT); | |||
| 2512 | ||||
| 2513 | MadeChange = false; | |||
| 2514 | ||||
| 2515 | // Traverse the same blocks that were analysed by BuildRankMap. | |||
| 2516 | for (BasicBlock *BI : RPOT) { | |||
| 2517 | 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.\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2517, __extension__ __PRETTY_FUNCTION__)); | |||
| 2518 | // Optimize every instruction in the basic block. | |||
| 2519 | for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;) | |||
| 2520 | if (isInstructionTriviallyDead(&*II)) { | |||
| 2521 | EraseInst(&*II++); | |||
| 2522 | } else { | |||
| 2523 | OptimizeInst(&*II); | |||
| 2524 | 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!\"" , "llvm/lib/Transforms/Scalar/Reassociate.cpp", 2524, __extension__ __PRETTY_FUNCTION__)); | |||
| 2525 | ++II; | |||
| 2526 | } | |||
| 2527 | ||||
| 2528 | // Make a copy of all the instructions to be redone so we can remove dead | |||
| 2529 | // instructions. | |||
| 2530 | OrderedSet ToRedo(RedoInsts); | |||
| 2531 | // Iterate over all instructions to be reevaluated and remove trivially dead | |||
| 2532 | // instructions. If any operand of the trivially dead instruction becomes | |||
| 2533 | // dead mark it for deletion as well. Continue this process until all | |||
| 2534 | // trivially dead instructions have been removed. | |||
| 2535 | while (!ToRedo.empty()) { | |||
| 2536 | Instruction *I = ToRedo.pop_back_val(); | |||
| 2537 | if (isInstructionTriviallyDead(I)) { | |||
| 2538 | RecursivelyEraseDeadInsts(I, ToRedo); | |||
| 2539 | MadeChange = true; | |||
| 2540 | } | |||
| 2541 | } | |||
| 2542 | ||||
| 2543 | // Now that we have removed dead instructions, we can reoptimize the | |||
| 2544 | // remaining instructions. | |||
| 2545 | while (!RedoInsts.empty()) { | |||
| 2546 | Instruction *I = RedoInsts.front(); | |||
| 2547 | RedoInsts.erase(RedoInsts.begin()); | |||
| 2548 | if (isInstructionTriviallyDead(I)) | |||
| 2549 | EraseInst(I); | |||
| 2550 | else | |||
| 2551 | OptimizeInst(I); | |||
| 2552 | } | |||
| 2553 | } | |||
| 2554 | ||||
| 2555 | // We are done with the rank map and pair map. | |||
| 2556 | RankMap.clear(); | |||
| 2557 | ValueRankMap.clear(); | |||
| 2558 | for (auto &Entry : PairMap) | |||
| 2559 | Entry.clear(); | |||
| 2560 | ||||
| 2561 | if (MadeChange) { | |||
| 2562 | PreservedAnalyses PA; | |||
| 2563 | PA.preserveSet<CFGAnalyses>(); | |||
| 2564 | return PA; | |||
| 2565 | } | |||
| 2566 | ||||
| 2567 | return PreservedAnalyses::all(); | |||
| 2568 | } | |||
| 2569 | ||||
| 2570 | namespace { | |||
| 2571 | ||||
| 2572 | class ReassociateLegacyPass : public FunctionPass { | |||
| 2573 | ReassociatePass Impl; | |||
| 2574 | ||||
| 2575 | public: | |||
| 2576 | static char ID; // Pass identification, replacement for typeid | |||
| 2577 | ||||
| 2578 | ReassociateLegacyPass() : FunctionPass(ID) { | |||
| 2579 | initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry()); | |||
| 2580 | } | |||
| 2581 | ||||
| 2582 | bool runOnFunction(Function &F) override { | |||
| 2583 | if (skipFunction(F)) | |||
| ||||
| 2584 | return false; | |||
| 2585 | ||||
| 2586 | FunctionAnalysisManager DummyFAM; | |||
| 2587 | auto PA = Impl.run(F, DummyFAM); | |||
| 2588 | return !PA.areAllPreserved(); | |||
| 2589 | } | |||
| 2590 | ||||
| 2591 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
| 2592 | AU.setPreservesCFG(); | |||
| 2593 | AU.addPreserved<AAResultsWrapperPass>(); | |||
| 2594 | AU.addPreserved<BasicAAWrapperPass>(); | |||
| 2595 | AU.addPreserved<GlobalsAAWrapperPass>(); | |||
| 2596 | } | |||
| 2597 | }; | |||
| 2598 | ||||
| 2599 | } // end anonymous namespace | |||
| 2600 | ||||
| 2601 | char ReassociateLegacyPass::ID = 0; | |||
| 2602 | ||||
| 2603 | 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)); } | |||
| 2604 | "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)); } | |||
| 2605 | ||||
| 2606 | // Public interface to the Reassociate pass | |||
| 2607 | FunctionPass *llvm::createReassociatePass() { | |||
| 2608 | return new ReassociateLegacyPass(); | |||
| 2609 | } |