Nadav Rotem | cbd9a19 | 2012-10-18 23:22:48 +0000 | [diff] [blame] | 1 | // llvm/Target/TargetTransformImpl.cpp - Target Loop Trans Info ---*- C++ -*-=// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | |
| 10 | #include "llvm/Target/TargetTransformImpl.h" |
| 11 | #include "llvm/Target/TargetLowering.h" |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 12 | #include <utility> |
Nadav Rotem | cbd9a19 | 2012-10-18 23:22:48 +0000 | [diff] [blame] | 13 | |
| 14 | using namespace llvm; |
| 15 | |
Nadav Rotem | 2704834 | 2012-10-24 17:22:41 +0000 | [diff] [blame] | 16 | //===----------------------------------------------------------------------===// |
| 17 | // |
| 18 | // Calls used by scalar transformations. |
| 19 | // |
| 20 | //===----------------------------------------------------------------------===// |
| 21 | |
Nadav Rotem | cbd9a19 | 2012-10-18 23:22:48 +0000 | [diff] [blame] | 22 | bool ScalarTargetTransformImpl::isLegalAddImmediate(int64_t imm) const { |
| 23 | return TLI->isLegalAddImmediate(imm); |
| 24 | } |
| 25 | |
| 26 | bool ScalarTargetTransformImpl::isLegalICmpImmediate(int64_t imm) const { |
| 27 | return TLI->isLegalICmpImmediate(imm); |
| 28 | } |
| 29 | |
| 30 | bool ScalarTargetTransformImpl::isLegalAddressingMode(const AddrMode &AM, |
| 31 | Type *Ty) const { |
| 32 | return TLI->isLegalAddressingMode(AM, Ty); |
| 33 | } |
| 34 | |
| 35 | bool ScalarTargetTransformImpl::isTruncateFree(Type *Ty1, Type *Ty2) const { |
| 36 | return TLI->isTruncateFree(Ty1, Ty2); |
| 37 | } |
| 38 | |
| 39 | bool ScalarTargetTransformImpl::isTypeLegal(Type *Ty) const { |
| 40 | EVT T = TLI->getValueType(Ty); |
| 41 | return TLI->isTypeLegal(T); |
| 42 | } |
| 43 | |
| 44 | unsigned ScalarTargetTransformImpl::getJumpBufAlignment() const { |
| 45 | return TLI->getJumpBufAlignment(); |
| 46 | } |
| 47 | |
| 48 | unsigned ScalarTargetTransformImpl::getJumpBufSize() const { |
| 49 | return TLI->getJumpBufSize(); |
| 50 | } |
Nadav Rotem | 2704834 | 2012-10-24 17:22:41 +0000 | [diff] [blame] | 51 | |
| 52 | //===----------------------------------------------------------------------===// |
| 53 | // |
| 54 | // Calls used by the vectorizers. |
| 55 | // |
| 56 | //===----------------------------------------------------------------------===// |
Renato Golin | e5372d6 | 2012-10-26 12:24:52 +0000 | [diff] [blame] | 57 | static int InstructionOpcodeToISD(unsigned Opcode) { |
| 58 | enum InstructionOpcodes { |
| 59 | #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, |
| 60 | #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM |
| 61 | #include "llvm/Instruction.def" |
| 62 | }; |
| 63 | switch (static_cast<InstructionOpcodes>(Opcode)) { |
| 64 | case Ret: return 0; |
| 65 | case Br: return 0; |
| 66 | case Switch: return 0; |
| 67 | case IndirectBr: return 0; |
| 68 | case Invoke: return 0; |
| 69 | case Resume: return 0; |
| 70 | case Unreachable: return 0; |
| 71 | case Add: return ISD::ADD; |
| 72 | case FAdd: return ISD::FADD; |
| 73 | case Sub: return ISD::SUB; |
| 74 | case FSub: return ISD::FSUB; |
| 75 | case Mul: return ISD::MUL; |
| 76 | case FMul: return ISD::FMUL; |
| 77 | case UDiv: return ISD::UDIV; |
| 78 | case SDiv: return ISD::UDIV; |
| 79 | case FDiv: return ISD::FDIV; |
| 80 | case URem: return ISD::UREM; |
| 81 | case SRem: return ISD::SREM; |
| 82 | case FRem: return ISD::FREM; |
| 83 | case Shl: return ISD::SHL; |
| 84 | case LShr: return ISD::SRL; |
| 85 | case AShr: return ISD::SRA; |
| 86 | case And: return ISD::AND; |
| 87 | case Or: return ISD::OR; |
| 88 | case Xor: return ISD::XOR; |
| 89 | case Alloca: return 0; |
| 90 | case Load: return ISD::LOAD; |
| 91 | case Store: return ISD::STORE; |
| 92 | case GetElementPtr: return 0; |
| 93 | case Fence: return 0; |
| 94 | case AtomicCmpXchg: return 0; |
| 95 | case AtomicRMW: return 0; |
| 96 | case Trunc: return ISD::TRUNCATE; |
| 97 | case ZExt: return ISD::ZERO_EXTEND; |
| 98 | case SExt: return ISD::SEXTLOAD; |
| 99 | case FPToUI: return ISD::FP_TO_UINT; |
| 100 | case FPToSI: return ISD::FP_TO_SINT; |
| 101 | case UIToFP: return ISD::UINT_TO_FP; |
| 102 | case SIToFP: return ISD::SINT_TO_FP; |
| 103 | case FPTrunc: return ISD::FP_ROUND; |
| 104 | case FPExt: return ISD::FP_EXTEND; |
| 105 | case PtrToInt: return ISD::BITCAST; |
| 106 | case IntToPtr: return ISD::BITCAST; |
| 107 | case BitCast: return ISD::BITCAST; |
| 108 | case ICmp: return ISD::SETCC; |
| 109 | case FCmp: return ISD::SETCC; |
| 110 | case PHI: return 0; |
| 111 | case Call: return 0; |
| 112 | case Select: return ISD::SELECT; |
| 113 | case UserOp1: return 0; |
| 114 | case UserOp2: return 0; |
| 115 | case VAArg: return 0; |
| 116 | case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; |
| 117 | case InsertElement: return ISD::INSERT_VECTOR_ELT; |
| 118 | case ShuffleVector: return ISD::VECTOR_SHUFFLE; |
| 119 | case ExtractValue: return ISD::MERGE_VALUES; |
| 120 | case InsertValue: return ISD::MERGE_VALUES; |
| 121 | case LandingPad: return 0; |
| 122 | } |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 123 | |
Renato Golin | e5372d6 | 2012-10-26 12:24:52 +0000 | [diff] [blame] | 124 | llvm_unreachable("Unknown instruction type encountered!"); |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 125 | } |
| 126 | |
Nadav Rotem | 1e19e46 | 2012-10-25 18:17:48 +0000 | [diff] [blame] | 127 | std::pair<unsigned, EVT> |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 128 | VectorTargetTransformImpl::getTypeLegalizationCost(LLVMContext &C, |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 129 | EVT Ty) const { |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 130 | unsigned Cost = 1; |
| 131 | // We keep legalizing the type until we find a legal kind. We assume that |
| 132 | // the only operation that costs anything is the split. After splitting |
| 133 | // we need to handle two types. |
| 134 | while (true) { |
| 135 | TargetLowering::LegalizeKind LK = TLI->getTypeConversion(C, Ty); |
| 136 | |
| 137 | if (LK.first == TargetLowering::TypeLegal) |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 138 | return std::make_pair(Cost, Ty); |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 139 | |
| 140 | if (LK.first == TargetLowering::TypeSplitVector) |
| 141 | Cost *= 2; |
| 142 | |
| 143 | // Keep legalizing the type. |
| 144 | Ty = LK.second; |
| 145 | } |
| 146 | } |
Nadav Rotem | 2704834 | 2012-10-24 17:22:41 +0000 | [diff] [blame] | 147 | |
| 148 | unsigned |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 149 | VectorTargetTransformImpl::getScalarizationOverhead(Type *Ty, |
| 150 | bool Insert, |
| 151 | bool Extract) const { |
| 152 | assert (Ty->isVectorTy() && "Can only scalarize vectors"); |
| 153 | unsigned Cost = 0; |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 154 | |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 155 | for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) { |
| 156 | if (Insert) |
| 157 | Cost += getVectorInstrCost(Instruction::InsertElement, Ty, i); |
| 158 | if (Extract) |
| 159 | Cost += getVectorInstrCost(Instruction::ExtractElement, Ty, i); |
Nadav Rotem | 1e19e46 | 2012-10-25 18:17:48 +0000 | [diff] [blame] | 160 | } |
| 161 | |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 162 | return Cost; |
| 163 | } |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 164 | |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 165 | unsigned VectorTargetTransformImpl::getArithmeticInstrCost(unsigned Opcode, |
| 166 | Type *Ty) const { |
| 167 | // Check if any of the operands are vector operands. |
| 168 | int ISD = InstructionOpcodeToISD(Opcode); |
| 169 | assert(ISD && "Invalid opcode"); |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 170 | |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 171 | std::pair<unsigned, EVT> LT = |
| 172 | getTypeLegalizationCost(Ty->getContext(), TLI->getValueType(Ty)); |
| 173 | |
| 174 | if (!TLI->isOperationExpand(ISD, LT.second)) { |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 175 | // The operation is legal. Assume it costs 1. Multiply |
| 176 | // by the type-legalization overhead. |
| 177 | return LT.first * 1; |
| 178 | } |
| 179 | |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 180 | // Else, assume that we need to scalarize this op. |
| 181 | if (Ty->isVectorTy()) { |
| 182 | unsigned Num = Ty->getVectorNumElements(); |
| 183 | unsigned Cost = getArithmeticInstrCost(Opcode, Ty->getScalarType()); |
| 184 | // return the cost of multiple scalar invocation plus the cost of inserting |
| 185 | // and extracting the values. |
| 186 | return getScalarizationOverhead(Ty, true, true) + Num * Cost; |
| 187 | } |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 188 | |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 189 | // We don't know anything about this scalar instruction. |
| 190 | return 1; |
| 191 | } |
| 192 | |
| 193 | unsigned VectorTargetTransformImpl::getBroadcastCost(Type *Tp) const { |
| 194 | return 1; |
| 195 | } |
| 196 | |
| 197 | unsigned VectorTargetTransformImpl::getCastInstrCost(unsigned Opcode, Type *Dst, |
| 198 | Type *Src) const { |
| 199 | assert(Src->isVectorTy() == Dst->isVectorTy() && "Invalid input types"); |
| 200 | int ISD = InstructionOpcodeToISD(Opcode); |
| 201 | assert(ISD && "Invalid opcode"); |
| 202 | |
| 203 | std::pair<unsigned, EVT> SrcLT = |
| 204 | getTypeLegalizationCost(Src->getContext(), TLI->getValueType(Src)); |
| 205 | |
| 206 | std::pair<unsigned, EVT> DstLT = |
| 207 | getTypeLegalizationCost(Dst->getContext(), TLI->getValueType(Dst)); |
| 208 | |
| 209 | // If the cast is between same-sized registers, then the check is simple. |
| 210 | if (SrcLT.first == DstLT.first && |
| 211 | SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) { |
| 212 | // Just check the op cost: |
| 213 | if (!TLI->isOperationExpand(ISD, DstLT.second)) { |
| 214 | // The operation is legal. Assume it costs 1. Multiply |
| 215 | // by the type-legalization overhead. |
| 216 | return SrcLT.first * 1; |
| 217 | } |
| 218 | } |
| 219 | |
| 220 | // Otherwise, assume that the cast is scalarized. |
| 221 | if (Dst->isVectorTy()) { |
| 222 | unsigned Num = Dst->getVectorNumElements(); |
| 223 | unsigned Cost = getCastInstrCost(Opcode, Src->getScalarType(), |
| 224 | Dst->getScalarType()); |
| 225 | // return the cost of multiple scalar invocation plus the cost of inserting |
| 226 | // and extracting the values. |
| 227 | return getScalarizationOverhead(Dst, true, true) + Num * Cost; |
| 228 | } |
| 229 | |
| 230 | // Unknown scalar opcode. |
| 231 | return 1; |
| 232 | } |
| 233 | |
| 234 | unsigned VectorTargetTransformImpl::getCFInstrCost(unsigned Opcode) const { |
| 235 | return 1; |
| 236 | } |
| 237 | |
| 238 | unsigned VectorTargetTransformImpl::getCmpSelInstrCost(unsigned Opcode, |
| 239 | Type *ValTy, |
| 240 | Type *CondTy) const { |
| 241 | int ISD = InstructionOpcodeToISD(Opcode); |
| 242 | assert(ISD && "Invalid opcode"); |
| 243 | |
| 244 | // Selects on vectors are actually vector selects. |
| 245 | if (ISD == ISD::SELECT) { |
| 246 | assert(CondTy && "CondTy must exist"); |
| 247 | if (CondTy->isVectorTy()) |
| 248 | ISD = ISD::VSELECT; |
| 249 | } |
| 250 | |
| 251 | std::pair<unsigned, EVT> LT = |
| 252 | getTypeLegalizationCost(ValTy->getContext(), TLI->getValueType(ValTy)); |
| 253 | |
| 254 | if (!TLI->isOperationExpand(ISD, LT.second)) { |
| 255 | // The operation is legal. Assume it costs 1. Multiply |
| 256 | // by the type-legalization overhead. |
| 257 | return LT.first * 1; |
| 258 | } |
| 259 | |
| 260 | // Otherwise, assume that the cast is scalarized. |
| 261 | if (ValTy->isVectorTy()) { |
| 262 | unsigned Num = ValTy->getVectorNumElements(); |
| 263 | if (CondTy) |
| 264 | CondTy = CondTy->getScalarType(); |
| 265 | unsigned Cost = getCmpSelInstrCost(Opcode, ValTy->getScalarType(), |
| 266 | CondTy); |
| 267 | |
| 268 | // return the cost of multiple scalar invocation plus the cost of inserting |
| 269 | // and extracting the values. |
| 270 | return getScalarizationOverhead(ValTy, true, false) + Num * Cost; |
| 271 | } |
| 272 | |
Nadav Rotem | e498160 | 2012-10-29 05:28:35 +0000 | [diff] [blame^] | 273 | // Unknown scalar opcode. |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 274 | return 1; |
| 275 | } |
| 276 | |
| 277 | /// Returns the expected cost of Vector Insert and Extract. |
| 278 | unsigned VectorTargetTransformImpl::getVectorInstrCost(unsigned Opcode, |
| 279 | Type *Val, |
| 280 | unsigned Index) const { |
| 281 | return 1; |
Nadav Rotem | 2704834 | 2012-10-24 17:22:41 +0000 | [diff] [blame] | 282 | } |
| 283 | |
| 284 | unsigned |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 285 | VectorTargetTransformImpl::getInstrCost(unsigned Opcode, Type *Ty1, |
| 286 | Type *Ty2) const { |
Nadav Rotem | 2704834 | 2012-10-24 17:22:41 +0000 | [diff] [blame] | 287 | return 1; |
| 288 | } |
| 289 | |
| 290 | unsigned |
| 291 | VectorTargetTransformImpl::getMemoryOpCost(unsigned Opcode, Type *Src, |
| 292 | unsigned Alignment, |
| 293 | unsigned AddressSpace) const { |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 294 | std::pair<unsigned, EVT> LT = |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 295 | getTypeLegalizationCost(Src->getContext(), TLI->getValueType(Src)); |
Nadav Rotem | a5a3a61 | 2012-10-26 23:49:28 +0000 | [diff] [blame] | 296 | |
Nadav Rotem | 2652c50 | 2012-10-24 23:47:38 +0000 | [diff] [blame] | 297 | // Assume that all loads of legal types cost 1. |
| 298 | return LT.first; |
Nadav Rotem | 2704834 | 2012-10-24 17:22:41 +0000 | [diff] [blame] | 299 | } |
Hal Finkel | 102a7c0 | 2012-10-26 04:28:02 +0000 | [diff] [blame] | 300 | |
| 301 | unsigned |
| 302 | VectorTargetTransformImpl::getNumberOfParts(Type *Tp) const { |
Nadav Rotem | e498160 | 2012-10-29 05:28:35 +0000 | [diff] [blame^] | 303 | std::pair<unsigned, EVT> LT = |
| 304 | getTypeLegalizationCost(Tp->getContext(), TLI->getValueType(Tp)); |
| 305 | return LT.first; |
Hal Finkel | 102a7c0 | 2012-10-26 04:28:02 +0000 | [diff] [blame] | 306 | } |
| 307 | |