1 //===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===// 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 /// \file 9 /// This file implements a TargetTransformInfo analysis pass specific to the 10 /// X86 target machine. It uses the target's detailed information to provide 11 /// more precise answers to certain TTI queries, while letting the target 12 /// independent and default TTI implementations handle the rest. 13 /// 14 //===----------------------------------------------------------------------===// 15 /// About Cost Model numbers used below it's necessary to say the following: 16 /// the numbers correspond to some "generic" X86 CPU instead of usage of 17 /// concrete CPU model. Usually the numbers correspond to CPU where the feature 18 /// apeared at the first time. For example, if we do Subtarget.hasSSE42() in 19 /// the lookups below the cost is based on Nehalem as that was the first CPU 20 /// to support that feature level and thus has most likely the worst case cost. 21 /// Some examples of other technologies/CPUs: 22 /// SSE 3 - Pentium4 / Athlon64 23 /// SSE 4.1 - Penryn 24 /// SSE 4.2 - Nehalem 25 /// AVX - Sandy Bridge 26 /// AVX2 - Haswell 27 /// AVX-512 - Xeon Phi / Skylake 28 /// And some examples of instruction target dependent costs (latency) 29 /// divss sqrtss rsqrtss 30 /// AMD K7 11-16 19 3 31 /// Piledriver 9-24 13-15 5 32 /// Jaguar 14 16 2 33 /// Pentium II,III 18 30 2 34 /// Nehalem 7-14 7-18 3 35 /// Haswell 10-13 11 5 36 /// TODO: Develop and implement the target dependent cost model and 37 /// specialize cost numbers for different Cost Model Targets such as throughput, 38 /// code size, latency and uop count. 39 //===----------------------------------------------------------------------===// 40 41 #include "X86TargetTransformInfo.h" 42 #include "llvm/Analysis/TargetTransformInfo.h" 43 #include "llvm/CodeGen/BasicTTIImpl.h" 44 #include "llvm/CodeGen/CostTable.h" 45 #include "llvm/CodeGen/TargetLowering.h" 46 #include "llvm/IR/IntrinsicInst.h" 47 #include "llvm/Support/Debug.h" 48 49 using namespace llvm; 50 51 #define DEBUG_TYPE "x86tti" 52 53 //===----------------------------------------------------------------------===// 54 // 55 // X86 cost model. 56 // 57 //===----------------------------------------------------------------------===// 58 59 TargetTransformInfo::PopcntSupportKind 60 X86TTIImpl::getPopcntSupport(unsigned TyWidth) { 61 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2"); 62 // TODO: Currently the __builtin_popcount() implementation using SSE3 63 // instructions is inefficient. Once the problem is fixed, we should 64 // call ST->hasSSE3() instead of ST->hasPOPCNT(). 65 return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software; 66 } 67 68 llvm::Optional<unsigned> X86TTIImpl::getCacheSize( 69 TargetTransformInfo::CacheLevel Level) const { 70 switch (Level) { 71 case TargetTransformInfo::CacheLevel::L1D: 72 // - Penryn 73 // - Nehalem 74 // - Westmere 75 // - Sandy Bridge 76 // - Ivy Bridge 77 // - Haswell 78 // - Broadwell 79 // - Skylake 80 // - Kabylake 81 return 32 * 1024; // 32 KByte 82 case TargetTransformInfo::CacheLevel::L2D: 83 // - Penryn 84 // - Nehalem 85 // - Westmere 86 // - Sandy Bridge 87 // - Ivy Bridge 88 // - Haswell 89 // - Broadwell 90 // - Skylake 91 // - Kabylake 92 return 256 * 1024; // 256 KByte 93 } 94 95 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel"); 96 } 97 98 llvm::Optional<unsigned> X86TTIImpl::getCacheAssociativity( 99 TargetTransformInfo::CacheLevel Level) const { 100 // - Penryn 101 // - Nehalem 102 // - Westmere 103 // - Sandy Bridge 104 // - Ivy Bridge 105 // - Haswell 106 // - Broadwell 107 // - Skylake 108 // - Kabylake 109 switch (Level) { 110 case TargetTransformInfo::CacheLevel::L1D: 111 LLVM_FALLTHROUGH; 112 case TargetTransformInfo::CacheLevel::L2D: 113 return 8; 114 } 115 116 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel"); 117 } 118 119 unsigned X86TTIImpl::getNumberOfRegisters(unsigned ClassID) const { 120 bool Vector = (ClassID == 1); 121 if (Vector && !ST->hasSSE1()) 122 return 0; 123 124 if (ST->is64Bit()) { 125 if (Vector && ST->hasAVX512()) 126 return 32; 127 return 16; 128 } 129 return 8; 130 } 131 132 TypeSize 133 X86TTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const { 134 unsigned PreferVectorWidth = ST->getPreferVectorWidth(); 135 switch (K) { 136 case TargetTransformInfo::RGK_Scalar: 137 return TypeSize::getFixed(ST->is64Bit() ? 64 : 32); 138 case TargetTransformInfo::RGK_FixedWidthVector: 139 if (ST->hasAVX512() && PreferVectorWidth >= 512) 140 return TypeSize::getFixed(512); 141 if (ST->hasAVX() && PreferVectorWidth >= 256) 142 return TypeSize::getFixed(256); 143 if (ST->hasSSE1() && PreferVectorWidth >= 128) 144 return TypeSize::getFixed(128); 145 return TypeSize::getFixed(0); 146 case TargetTransformInfo::RGK_ScalableVector: 147 return TypeSize::getScalable(0); 148 } 149 150 llvm_unreachable("Unsupported register kind"); 151 } 152 153 unsigned X86TTIImpl::getLoadStoreVecRegBitWidth(unsigned) const { 154 return getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) 155 .getFixedSize(); 156 } 157 158 unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) { 159 // If the loop will not be vectorized, don't interleave the loop. 160 // Let regular unroll to unroll the loop, which saves the overflow 161 // check and memory check cost. 162 if (VF == 1) 163 return 1; 164 165 if (ST->isAtom()) 166 return 1; 167 168 // Sandybridge and Haswell have multiple execution ports and pipelined 169 // vector units. 170 if (ST->hasAVX()) 171 return 4; 172 173 return 2; 174 } 175 176 InstructionCost X86TTIImpl::getArithmeticInstrCost( 177 unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, 178 TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info, 179 TTI::OperandValueProperties Opd1PropInfo, 180 TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args, 181 const Instruction *CxtI) { 182 // TODO: Handle more cost kinds. 183 if (CostKind != TTI::TCK_RecipThroughput) 184 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, 185 Op2Info, Opd1PropInfo, 186 Opd2PropInfo, Args, CxtI); 187 188 // vXi8 multiplications are always promoted to vXi16. 189 if (Opcode == Instruction::Mul && Ty->isVectorTy() && 190 Ty->getScalarSizeInBits() == 8) { 191 Type *WideVecTy = 192 VectorType::getExtendedElementVectorType(cast<VectorType>(Ty)); 193 return getCastInstrCost(Instruction::ZExt, WideVecTy, Ty, 194 TargetTransformInfo::CastContextHint::None, 195 CostKind) + 196 getCastInstrCost(Instruction::Trunc, Ty, WideVecTy, 197 TargetTransformInfo::CastContextHint::None, 198 CostKind) + 199 getArithmeticInstrCost(Opcode, WideVecTy, CostKind, Op1Info, Op2Info, 200 Opd1PropInfo, Opd2PropInfo); 201 } 202 203 // Legalize the type. 204 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 205 206 int ISD = TLI->InstructionOpcodeToISD(Opcode); 207 assert(ISD && "Invalid opcode"); 208 209 if (ISD == ISD::MUL && Args.size() == 2 && LT.second.isVector() && 210 LT.second.getScalarType() == MVT::i32) { 211 // Check if the operands can be represented as a smaller datatype. 212 bool Op1Signed = false, Op2Signed = false; 213 unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed); 214 unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed); 215 unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize); 216 217 // If both are representable as i15 and at least one is constant, 218 // zero-extended, or sign-extended from vXi16 (or less pre-SSE41) then we 219 // can treat this as PMADDWD which has the same costs as a vXi16 multiply. 220 if (OpMinSize <= 15 && !ST->isPMADDWDSlow()) { 221 bool Op1Constant = 222 isa<ConstantDataVector>(Args[0]) || isa<ConstantVector>(Args[0]); 223 bool Op2Constant = 224 isa<ConstantDataVector>(Args[1]) || isa<ConstantVector>(Args[1]); 225 bool Op1Sext = isa<SExtInst>(Args[0]) && 226 (Op1MinSize == 15 || (Op1MinSize < 15 && !ST->hasSSE41())); 227 bool Op2Sext = isa<SExtInst>(Args[1]) && 228 (Op2MinSize == 15 || (Op2MinSize < 15 && !ST->hasSSE41())); 229 230 bool IsZeroExtended = !Op1Signed || !Op2Signed; 231 bool IsConstant = Op1Constant || Op2Constant; 232 bool IsSext = Op1Sext || Op2Sext; 233 if (IsConstant || IsZeroExtended || IsSext) 234 LT.second = 235 MVT::getVectorVT(MVT::i16, 2 * LT.second.getVectorNumElements()); 236 } 237 } 238 239 if ((ISD == ISD::MUL || ISD == ISD::SDIV || ISD == ISD::SREM || 240 ISD == ISD::UDIV || ISD == ISD::UREM) && 241 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 242 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 243 Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) { 244 // Vector multiply by pow2 will be simplified to shifts. 245 if (ISD == ISD::MUL) { 246 InstructionCost Cost = getArithmeticInstrCost( 247 Instruction::Shl, Ty, CostKind, Op1Info, Op2Info, 248 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 249 return Cost; 250 } 251 252 if (ISD == ISD::SDIV || ISD == ISD::SREM) { 253 // On X86, vector signed division by constants power-of-two are 254 // normally expanded to the sequence SRA + SRL + ADD + SRA. 255 // The OperandValue properties may not be the same as that of the previous 256 // operation; conservatively assume OP_None. 257 InstructionCost Cost = 258 2 * getArithmeticInstrCost(Instruction::AShr, Ty, CostKind, Op1Info, 259 Op2Info, TargetTransformInfo::OP_None, 260 TargetTransformInfo::OP_None); 261 Cost += getArithmeticInstrCost(Instruction::LShr, Ty, CostKind, Op1Info, 262 Op2Info, TargetTransformInfo::OP_None, 263 TargetTransformInfo::OP_None); 264 Cost += getArithmeticInstrCost(Instruction::Add, Ty, CostKind, Op1Info, 265 Op2Info, TargetTransformInfo::OP_None, 266 TargetTransformInfo::OP_None); 267 268 if (ISD == ISD::SREM) { 269 // For SREM: (X % C) is the equivalent of (X - (X/C)*C) 270 Cost += getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, Op1Info, 271 Op2Info); 272 Cost += getArithmeticInstrCost(Instruction::Sub, Ty, CostKind, Op1Info, 273 Op2Info); 274 } 275 276 return Cost; 277 } 278 279 // Vector unsigned division/remainder will be simplified to shifts/masks. 280 if (ISD == ISD::UDIV) 281 return getArithmeticInstrCost(Instruction::LShr, Ty, CostKind, Op1Info, 282 Op2Info, TargetTransformInfo::OP_None, 283 TargetTransformInfo::OP_None); 284 // UREM 285 return getArithmeticInstrCost(Instruction::And, Ty, CostKind, Op1Info, 286 Op2Info, TargetTransformInfo::OP_None, 287 TargetTransformInfo::OP_None); 288 } 289 290 static const CostTblEntry GLMCostTable[] = { 291 { ISD::FDIV, MVT::f32, 18 }, // divss 292 { ISD::FDIV, MVT::v4f32, 35 }, // divps 293 { ISD::FDIV, MVT::f64, 33 }, // divsd 294 { ISD::FDIV, MVT::v2f64, 65 }, // divpd 295 }; 296 297 if (ST->useGLMDivSqrtCosts()) 298 if (const auto *Entry = CostTableLookup(GLMCostTable, ISD, 299 LT.second)) 300 return LT.first * Entry->Cost; 301 302 static const CostTblEntry SLMCostTable[] = { 303 { ISD::MUL, MVT::v4i32, 11 }, // pmulld 304 { ISD::MUL, MVT::v8i16, 2 }, // pmullw 305 { ISD::FMUL, MVT::f64, 2 }, // mulsd 306 { ISD::FMUL, MVT::v2f64, 4 }, // mulpd 307 { ISD::FMUL, MVT::v4f32, 2 }, // mulps 308 { ISD::FDIV, MVT::f32, 17 }, // divss 309 { ISD::FDIV, MVT::v4f32, 39 }, // divps 310 { ISD::FDIV, MVT::f64, 32 }, // divsd 311 { ISD::FDIV, MVT::v2f64, 69 }, // divpd 312 { ISD::FADD, MVT::v2f64, 2 }, // addpd 313 { ISD::FSUB, MVT::v2f64, 2 }, // subpd 314 // v2i64/v4i64 mul is custom lowered as a series of long: 315 // multiplies(3), shifts(3) and adds(2) 316 // slm muldq version throughput is 2 and addq throughput 4 317 // thus: 3X2 (muldq throughput) + 3X1 (shift throughput) + 318 // 3X4 (addq throughput) = 17 319 { ISD::MUL, MVT::v2i64, 17 }, 320 // slm addq\subq throughput is 4 321 { ISD::ADD, MVT::v2i64, 4 }, 322 { ISD::SUB, MVT::v2i64, 4 }, 323 }; 324 325 if (ST->useSLMArithCosts()) { 326 if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) { 327 // Check if the operands can be shrinked into a smaller datatype. 328 // TODO: Merge this into generiic vXi32 MUL patterns above. 329 bool Op1Signed = false; 330 unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed); 331 bool Op2Signed = false; 332 unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed); 333 334 bool SignedMode = Op1Signed || Op2Signed; 335 unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize); 336 337 if (OpMinSize <= 7) 338 return LT.first * 3; // pmullw/sext 339 if (!SignedMode && OpMinSize <= 8) 340 return LT.first * 3; // pmullw/zext 341 if (OpMinSize <= 15) 342 return LT.first * 5; // pmullw/pmulhw/pshuf 343 if (!SignedMode && OpMinSize <= 16) 344 return LT.first * 5; // pmullw/pmulhw/pshuf 345 } 346 347 if (const auto *Entry = CostTableLookup(SLMCostTable, ISD, 348 LT.second)) { 349 return LT.first * Entry->Cost; 350 } 351 } 352 353 static const CostTblEntry AVX512BWUniformConstCostTable[] = { 354 { ISD::SHL, MVT::v64i8, 2 }, // psllw + pand. 355 { ISD::SRL, MVT::v64i8, 2 }, // psrlw + pand. 356 { ISD::SRA, MVT::v64i8, 4 }, // psrlw, pand, pxor, psubb. 357 }; 358 359 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 360 ST->hasBWI()) { 361 if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD, 362 LT.second)) 363 return LT.first * Entry->Cost; 364 } 365 366 static const CostTblEntry AVX512UniformConstCostTable[] = { 367 { ISD::SRA, MVT::v2i64, 1 }, 368 { ISD::SRA, MVT::v4i64, 1 }, 369 { ISD::SRA, MVT::v8i64, 1 }, 370 371 { ISD::SHL, MVT::v64i8, 4 }, // psllw + pand. 372 { ISD::SRL, MVT::v64i8, 4 }, // psrlw + pand. 373 { ISD::SRA, MVT::v64i8, 8 }, // psrlw, pand, pxor, psubb. 374 375 { ISD::SDIV, MVT::v16i32, 6 }, // pmuludq sequence 376 { ISD::SREM, MVT::v16i32, 8 }, // pmuludq+mul+sub sequence 377 { ISD::UDIV, MVT::v16i32, 5 }, // pmuludq sequence 378 { ISD::UREM, MVT::v16i32, 7 }, // pmuludq+mul+sub sequence 379 }; 380 381 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 382 ST->hasAVX512()) { 383 if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD, 384 LT.second)) 385 return LT.first * Entry->Cost; 386 } 387 388 static const CostTblEntry AVX2UniformConstCostTable[] = { 389 { ISD::SHL, MVT::v32i8, 2 }, // psllw + pand. 390 { ISD::SRL, MVT::v32i8, 2 }, // psrlw + pand. 391 { ISD::SRA, MVT::v32i8, 4 }, // psrlw, pand, pxor, psubb. 392 393 { ISD::SRA, MVT::v4i64, 4 }, // 2 x psrad + shuffle. 394 395 { ISD::SDIV, MVT::v8i32, 6 }, // pmuludq sequence 396 { ISD::SREM, MVT::v8i32, 8 }, // pmuludq+mul+sub sequence 397 { ISD::UDIV, MVT::v8i32, 5 }, // pmuludq sequence 398 { ISD::UREM, MVT::v8i32, 7 }, // pmuludq+mul+sub sequence 399 }; 400 401 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 402 ST->hasAVX2()) { 403 if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD, 404 LT.second)) 405 return LT.first * Entry->Cost; 406 } 407 408 static const CostTblEntry SSE2UniformConstCostTable[] = { 409 { ISD::SHL, MVT::v16i8, 2 }, // psllw + pand. 410 { ISD::SRL, MVT::v16i8, 2 }, // psrlw + pand. 411 { ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb. 412 413 { ISD::SHL, MVT::v32i8, 4+2 }, // 2*(psllw + pand) + split. 414 { ISD::SRL, MVT::v32i8, 4+2 }, // 2*(psrlw + pand) + split. 415 { ISD::SRA, MVT::v32i8, 8+2 }, // 2*(psrlw, pand, pxor, psubb) + split. 416 417 { ISD::SDIV, MVT::v8i32, 12+2 }, // 2*pmuludq sequence + split. 418 { ISD::SREM, MVT::v8i32, 16+2 }, // 2*pmuludq+mul+sub sequence + split. 419 { ISD::SDIV, MVT::v4i32, 6 }, // pmuludq sequence 420 { ISD::SREM, MVT::v4i32, 8 }, // pmuludq+mul+sub sequence 421 { ISD::UDIV, MVT::v8i32, 10+2 }, // 2*pmuludq sequence + split. 422 { ISD::UREM, MVT::v8i32, 14+2 }, // 2*pmuludq+mul+sub sequence + split. 423 { ISD::UDIV, MVT::v4i32, 5 }, // pmuludq sequence 424 { ISD::UREM, MVT::v4i32, 7 }, // pmuludq+mul+sub sequence 425 }; 426 427 // XOP has faster vXi8 shifts. 428 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 429 ST->hasSSE2() && !ST->hasXOP()) { 430 if (const auto *Entry = 431 CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second)) 432 return LT.first * Entry->Cost; 433 } 434 435 static const CostTblEntry AVX512BWConstCostTable[] = { 436 { ISD::SDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence 437 { ISD::SREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 438 { ISD::UDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence 439 { ISD::UREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 440 { ISD::SDIV, MVT::v32i16, 6 }, // vpmulhw sequence 441 { ISD::SREM, MVT::v32i16, 8 }, // vpmulhw+mul+sub sequence 442 { ISD::UDIV, MVT::v32i16, 6 }, // vpmulhuw sequence 443 { ISD::UREM, MVT::v32i16, 8 }, // vpmulhuw+mul+sub sequence 444 }; 445 446 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 447 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 448 ST->hasBWI()) { 449 if (const auto *Entry = 450 CostTableLookup(AVX512BWConstCostTable, ISD, LT.second)) 451 return LT.first * Entry->Cost; 452 } 453 454 static const CostTblEntry AVX512ConstCostTable[] = { 455 { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence 456 { ISD::SREM, MVT::v16i32, 17 }, // vpmuldq+mul+sub sequence 457 { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence 458 { ISD::UREM, MVT::v16i32, 17 }, // vpmuludq+mul+sub sequence 459 { ISD::SDIV, MVT::v64i8, 28 }, // 4*ext+4*pmulhw sequence 460 { ISD::SREM, MVT::v64i8, 32 }, // 4*ext+4*pmulhw+mul+sub sequence 461 { ISD::UDIV, MVT::v64i8, 28 }, // 4*ext+4*pmulhw sequence 462 { ISD::UREM, MVT::v64i8, 32 }, // 4*ext+4*pmulhw+mul+sub sequence 463 { ISD::SDIV, MVT::v32i16, 12 }, // 2*vpmulhw sequence 464 { ISD::SREM, MVT::v32i16, 16 }, // 2*vpmulhw+mul+sub sequence 465 { ISD::UDIV, MVT::v32i16, 12 }, // 2*vpmulhuw sequence 466 { ISD::UREM, MVT::v32i16, 16 }, // 2*vpmulhuw+mul+sub sequence 467 }; 468 469 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 470 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 471 ST->hasAVX512()) { 472 if (const auto *Entry = 473 CostTableLookup(AVX512ConstCostTable, ISD, LT.second)) 474 return LT.first * Entry->Cost; 475 } 476 477 static const CostTblEntry AVX2ConstCostTable[] = { 478 { ISD::SDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence 479 { ISD::SREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 480 { ISD::UDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence 481 { ISD::UREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 482 { ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence 483 { ISD::SREM, MVT::v16i16, 8 }, // vpmulhw+mul+sub sequence 484 { ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence 485 { ISD::UREM, MVT::v16i16, 8 }, // vpmulhuw+mul+sub sequence 486 { ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence 487 { ISD::SREM, MVT::v8i32, 19 }, // vpmuldq+mul+sub sequence 488 { ISD::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence 489 { ISD::UREM, MVT::v8i32, 19 }, // vpmuludq+mul+sub sequence 490 }; 491 492 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 493 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 494 ST->hasAVX2()) { 495 if (const auto *Entry = CostTableLookup(AVX2ConstCostTable, ISD, LT.second)) 496 return LT.first * Entry->Cost; 497 } 498 499 static const CostTblEntry SSE2ConstCostTable[] = { 500 { ISD::SDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split. 501 { ISD::SREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split. 502 { ISD::SDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence 503 { ISD::SREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 504 { ISD::UDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split. 505 { ISD::UREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split. 506 { ISD::UDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence 507 { ISD::UREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 508 { ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split. 509 { ISD::SREM, MVT::v16i16, 16+2 }, // 2*pmulhw+mul+sub sequence + split. 510 { ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence 511 { ISD::SREM, MVT::v8i16, 8 }, // pmulhw+mul+sub sequence 512 { ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split. 513 { ISD::UREM, MVT::v16i16, 16+2 }, // 2*pmulhuw+mul+sub sequence + split. 514 { ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence 515 { ISD::UREM, MVT::v8i16, 8 }, // pmulhuw+mul+sub sequence 516 { ISD::SDIV, MVT::v8i32, 38+2 }, // 2*pmuludq sequence + split. 517 { ISD::SREM, MVT::v8i32, 48+2 }, // 2*pmuludq+mul+sub sequence + split. 518 { ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence 519 { ISD::SREM, MVT::v4i32, 24 }, // pmuludq+mul+sub sequence 520 { ISD::UDIV, MVT::v8i32, 30+2 }, // 2*pmuludq sequence + split. 521 { ISD::UREM, MVT::v8i32, 40+2 }, // 2*pmuludq+mul+sub sequence + split. 522 { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence 523 { ISD::UREM, MVT::v4i32, 20 }, // pmuludq+mul+sub sequence 524 }; 525 526 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 527 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 528 ST->hasSSE2()) { 529 // pmuldq sequence. 530 if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX()) 531 return LT.first * 32; 532 if (ISD == ISD::SREM && LT.second == MVT::v8i32 && ST->hasAVX()) 533 return LT.first * 38; 534 if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41()) 535 return LT.first * 15; 536 if (ISD == ISD::SREM && LT.second == MVT::v4i32 && ST->hasSSE41()) 537 return LT.first * 20; 538 539 if (const auto *Entry = CostTableLookup(SSE2ConstCostTable, ISD, LT.second)) 540 return LT.first * Entry->Cost; 541 } 542 543 static const CostTblEntry AVX512BWShiftCostTable[] = { 544 { ISD::SHL, MVT::v16i8, 4 }, // extend/vpsllvw/pack sequence. 545 { ISD::SRL, MVT::v16i8, 4 }, // extend/vpsrlvw/pack sequence. 546 { ISD::SRA, MVT::v16i8, 4 }, // extend/vpsravw/pack sequence. 547 { ISD::SHL, MVT::v32i8, 4 }, // extend/vpsllvw/pack sequence. 548 { ISD::SRL, MVT::v32i8, 4 }, // extend/vpsrlvw/pack sequence. 549 { ISD::SRA, MVT::v32i8, 6 }, // extend/vpsravw/pack sequence. 550 { ISD::SHL, MVT::v64i8, 6 }, // extend/vpsllvw/pack sequence. 551 { ISD::SRL, MVT::v64i8, 7 }, // extend/vpsrlvw/pack sequence. 552 { ISD::SRA, MVT::v64i8, 15 }, // extend/vpsravw/pack sequence. 553 554 { ISD::SHL, MVT::v8i16, 1 }, // vpsllvw 555 { ISD::SRL, MVT::v8i16, 1 }, // vpsrlvw 556 { ISD::SRA, MVT::v8i16, 1 }, // vpsravw 557 { ISD::SHL, MVT::v16i16, 1 }, // vpsllvw 558 { ISD::SRL, MVT::v16i16, 1 }, // vpsrlvw 559 { ISD::SRA, MVT::v16i16, 1 }, // vpsravw 560 { ISD::SHL, MVT::v32i16, 1 }, // vpsllvw 561 { ISD::SRL, MVT::v32i16, 1 }, // vpsrlvw 562 { ISD::SRA, MVT::v32i16, 1 }, // vpsravw 563 }; 564 565 if (ST->hasBWI()) 566 if (const auto *Entry = CostTableLookup(AVX512BWShiftCostTable, ISD, LT.second)) 567 return LT.first * Entry->Cost; 568 569 static const CostTblEntry AVX2UniformCostTable[] = { 570 // Uniform splats are cheaper for the following instructions. 571 { ISD::SHL, MVT::v16i16, 1 }, // psllw. 572 { ISD::SRL, MVT::v16i16, 1 }, // psrlw. 573 { ISD::SRA, MVT::v16i16, 1 }, // psraw. 574 { ISD::SHL, MVT::v32i16, 2 }, // 2*psllw. 575 { ISD::SRL, MVT::v32i16, 2 }, // 2*psrlw. 576 { ISD::SRA, MVT::v32i16, 2 }, // 2*psraw. 577 578 { ISD::SHL, MVT::v8i32, 1 }, // pslld 579 { ISD::SRL, MVT::v8i32, 1 }, // psrld 580 { ISD::SRA, MVT::v8i32, 1 }, // psrad 581 { ISD::SHL, MVT::v4i64, 1 }, // psllq 582 { ISD::SRL, MVT::v4i64, 1 }, // psrlq 583 }; 584 585 if (ST->hasAVX2() && 586 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 587 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 588 if (const auto *Entry = 589 CostTableLookup(AVX2UniformCostTable, ISD, LT.second)) 590 return LT.first * Entry->Cost; 591 } 592 593 static const CostTblEntry SSE2UniformCostTable[] = { 594 // Uniform splats are cheaper for the following instructions. 595 { ISD::SHL, MVT::v8i16, 1 }, // psllw. 596 { ISD::SHL, MVT::v4i32, 1 }, // pslld 597 { ISD::SHL, MVT::v2i64, 1 }, // psllq. 598 599 { ISD::SRL, MVT::v8i16, 1 }, // psrlw. 600 { ISD::SRL, MVT::v4i32, 1 }, // psrld. 601 { ISD::SRL, MVT::v2i64, 1 }, // psrlq. 602 603 { ISD::SRA, MVT::v8i16, 1 }, // psraw. 604 { ISD::SRA, MVT::v4i32, 1 }, // psrad. 605 }; 606 607 if (ST->hasSSE2() && 608 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 609 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 610 if (const auto *Entry = 611 CostTableLookup(SSE2UniformCostTable, ISD, LT.second)) 612 return LT.first * Entry->Cost; 613 } 614 615 static const CostTblEntry AVX512DQCostTable[] = { 616 { ISD::MUL, MVT::v2i64, 2 }, // pmullq 617 { ISD::MUL, MVT::v4i64, 2 }, // pmullq 618 { ISD::MUL, MVT::v8i64, 2 } // pmullq 619 }; 620 621 // Look for AVX512DQ lowering tricks for custom cases. 622 if (ST->hasDQI()) 623 if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second)) 624 return LT.first * Entry->Cost; 625 626 static const CostTblEntry AVX512BWCostTable[] = { 627 { ISD::SHL, MVT::v64i8, 11 }, // vpblendvb sequence. 628 { ISD::SRL, MVT::v64i8, 11 }, // vpblendvb sequence. 629 { ISD::SRA, MVT::v64i8, 24 }, // vpblendvb sequence. 630 }; 631 632 // Look for AVX512BW lowering tricks for custom cases. 633 if (ST->hasBWI()) 634 if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second)) 635 return LT.first * Entry->Cost; 636 637 static const CostTblEntry AVX512CostTable[] = { 638 { ISD::SHL, MVT::v4i32, 1 }, 639 { ISD::SRL, MVT::v4i32, 1 }, 640 { ISD::SRA, MVT::v4i32, 1 }, 641 { ISD::SHL, MVT::v8i32, 1 }, 642 { ISD::SRL, MVT::v8i32, 1 }, 643 { ISD::SRA, MVT::v8i32, 1 }, 644 { ISD::SHL, MVT::v16i32, 1 }, 645 { ISD::SRL, MVT::v16i32, 1 }, 646 { ISD::SRA, MVT::v16i32, 1 }, 647 648 { ISD::SHL, MVT::v2i64, 1 }, 649 { ISD::SRL, MVT::v2i64, 1 }, 650 { ISD::SHL, MVT::v4i64, 1 }, 651 { ISD::SRL, MVT::v4i64, 1 }, 652 { ISD::SHL, MVT::v8i64, 1 }, 653 { ISD::SRL, MVT::v8i64, 1 }, 654 655 { ISD::SRA, MVT::v2i64, 1 }, 656 { ISD::SRA, MVT::v4i64, 1 }, 657 { ISD::SRA, MVT::v8i64, 1 }, 658 659 { ISD::MUL, MVT::v16i32, 1 }, // pmulld (Skylake from agner.org) 660 { ISD::MUL, MVT::v8i32, 1 }, // pmulld (Skylake from agner.org) 661 { ISD::MUL, MVT::v4i32, 1 }, // pmulld (Skylake from agner.org) 662 { ISD::MUL, MVT::v8i64, 6 }, // 3*pmuludq/3*shift/2*add 663 664 { ISD::FNEG, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 665 { ISD::FADD, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 666 { ISD::FSUB, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 667 { ISD::FMUL, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 668 { ISD::FDIV, MVT::f64, 4 }, // Skylake from http://www.agner.org/ 669 { ISD::FDIV, MVT::v2f64, 4 }, // Skylake from http://www.agner.org/ 670 { ISD::FDIV, MVT::v4f64, 8 }, // Skylake from http://www.agner.org/ 671 { ISD::FDIV, MVT::v8f64, 16 }, // Skylake from http://www.agner.org/ 672 673 { ISD::FNEG, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 674 { ISD::FADD, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 675 { ISD::FSUB, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 676 { ISD::FMUL, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 677 { ISD::FDIV, MVT::f32, 3 }, // Skylake from http://www.agner.org/ 678 { ISD::FDIV, MVT::v4f32, 3 }, // Skylake from http://www.agner.org/ 679 { ISD::FDIV, MVT::v8f32, 5 }, // Skylake from http://www.agner.org/ 680 { ISD::FDIV, MVT::v16f32, 10 }, // Skylake from http://www.agner.org/ 681 }; 682 683 if (ST->hasAVX512()) 684 if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second)) 685 return LT.first * Entry->Cost; 686 687 static const CostTblEntry AVX2ShiftCostTable[] = { 688 // Shifts on vXi64/vXi32 on AVX2 is legal even though we declare to 689 // customize them to detect the cases where shift amount is a scalar one. 690 { ISD::SHL, MVT::v4i32, 2 }, // vpsllvd (Haswell from agner.org) 691 { ISD::SRL, MVT::v4i32, 2 }, // vpsrlvd (Haswell from agner.org) 692 { ISD::SRA, MVT::v4i32, 2 }, // vpsravd (Haswell from agner.org) 693 { ISD::SHL, MVT::v8i32, 2 }, // vpsllvd (Haswell from agner.org) 694 { ISD::SRL, MVT::v8i32, 2 }, // vpsrlvd (Haswell from agner.org) 695 { ISD::SRA, MVT::v8i32, 2 }, // vpsravd (Haswell from agner.org) 696 { ISD::SHL, MVT::v2i64, 1 }, // vpsllvq (Haswell from agner.org) 697 { ISD::SRL, MVT::v2i64, 1 }, // vpsrlvq (Haswell from agner.org) 698 { ISD::SHL, MVT::v4i64, 1 }, // vpsllvq (Haswell from agner.org) 699 { ISD::SRL, MVT::v4i64, 1 }, // vpsrlvq (Haswell from agner.org) 700 }; 701 702 if (ST->hasAVX512()) { 703 if (ISD == ISD::SHL && LT.second == MVT::v32i16 && 704 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 705 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 706 // On AVX512, a packed v32i16 shift left by a constant build_vector 707 // is lowered into a vector multiply (vpmullw). 708 return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, 709 Op1Info, Op2Info, 710 TargetTransformInfo::OP_None, 711 TargetTransformInfo::OP_None); 712 } 713 714 // Look for AVX2 lowering tricks (XOP is always better at v4i32 shifts). 715 if (ST->hasAVX2() && !(ST->hasXOP() && LT.second == MVT::v4i32)) { 716 if (ISD == ISD::SHL && LT.second == MVT::v16i16 && 717 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 718 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 719 // On AVX2, a packed v16i16 shift left by a constant build_vector 720 // is lowered into a vector multiply (vpmullw). 721 return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, 722 Op1Info, Op2Info, 723 TargetTransformInfo::OP_None, 724 TargetTransformInfo::OP_None); 725 726 if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second)) 727 return LT.first * Entry->Cost; 728 } 729 730 static const CostTblEntry XOPShiftCostTable[] = { 731 // 128bit shifts take 1cy, but right shifts require negation beforehand. 732 { ISD::SHL, MVT::v16i8, 1 }, 733 { ISD::SRL, MVT::v16i8, 2 }, 734 { ISD::SRA, MVT::v16i8, 2 }, 735 { ISD::SHL, MVT::v8i16, 1 }, 736 { ISD::SRL, MVT::v8i16, 2 }, 737 { ISD::SRA, MVT::v8i16, 2 }, 738 { ISD::SHL, MVT::v4i32, 1 }, 739 { ISD::SRL, MVT::v4i32, 2 }, 740 { ISD::SRA, MVT::v4i32, 2 }, 741 { ISD::SHL, MVT::v2i64, 1 }, 742 { ISD::SRL, MVT::v2i64, 2 }, 743 { ISD::SRA, MVT::v2i64, 2 }, 744 // 256bit shifts require splitting if AVX2 didn't catch them above. 745 { ISD::SHL, MVT::v32i8, 2+2 }, 746 { ISD::SRL, MVT::v32i8, 4+2 }, 747 { ISD::SRA, MVT::v32i8, 4+2 }, 748 { ISD::SHL, MVT::v16i16, 2+2 }, 749 { ISD::SRL, MVT::v16i16, 4+2 }, 750 { ISD::SRA, MVT::v16i16, 4+2 }, 751 { ISD::SHL, MVT::v8i32, 2+2 }, 752 { ISD::SRL, MVT::v8i32, 4+2 }, 753 { ISD::SRA, MVT::v8i32, 4+2 }, 754 { ISD::SHL, MVT::v4i64, 2+2 }, 755 { ISD::SRL, MVT::v4i64, 4+2 }, 756 { ISD::SRA, MVT::v4i64, 4+2 }, 757 }; 758 759 // Look for XOP lowering tricks. 760 if (ST->hasXOP()) { 761 // If the right shift is constant then we'll fold the negation so 762 // it's as cheap as a left shift. 763 int ShiftISD = ISD; 764 if ((ShiftISD == ISD::SRL || ShiftISD == ISD::SRA) && 765 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 766 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 767 ShiftISD = ISD::SHL; 768 if (const auto *Entry = 769 CostTableLookup(XOPShiftCostTable, ShiftISD, LT.second)) 770 return LT.first * Entry->Cost; 771 } 772 773 static const CostTblEntry SSE2UniformShiftCostTable[] = { 774 // Uniform splats are cheaper for the following instructions. 775 { ISD::SHL, MVT::v16i16, 2+2 }, // 2*psllw + split. 776 { ISD::SHL, MVT::v8i32, 2+2 }, // 2*pslld + split. 777 { ISD::SHL, MVT::v4i64, 2+2 }, // 2*psllq + split. 778 779 { ISD::SRL, MVT::v16i16, 2+2 }, // 2*psrlw + split. 780 { ISD::SRL, MVT::v8i32, 2+2 }, // 2*psrld + split. 781 { ISD::SRL, MVT::v4i64, 2+2 }, // 2*psrlq + split. 782 783 { ISD::SRA, MVT::v16i16, 2+2 }, // 2*psraw + split. 784 { ISD::SRA, MVT::v8i32, 2+2 }, // 2*psrad + split. 785 { ISD::SRA, MVT::v2i64, 4 }, // 2*psrad + shuffle. 786 { ISD::SRA, MVT::v4i64, 8+2 }, // 2*(2*psrad + shuffle) + split. 787 }; 788 789 if (ST->hasSSE2() && 790 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 791 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 792 793 // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table. 794 if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2()) 795 return LT.first * 4; // 2*psrad + shuffle. 796 797 if (const auto *Entry = 798 CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second)) 799 return LT.first * Entry->Cost; 800 } 801 802 if (ISD == ISD::SHL && 803 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) { 804 MVT VT = LT.second; 805 // Vector shift left by non uniform constant can be lowered 806 // into vector multiply. 807 if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) || 808 ((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX())) 809 ISD = ISD::MUL; 810 } 811 812 static const CostTblEntry AVX2CostTable[] = { 813 { ISD::SHL, MVT::v16i8, 6 }, // vpblendvb sequence. 814 { ISD::SHL, MVT::v32i8, 6 }, // vpblendvb sequence. 815 { ISD::SHL, MVT::v64i8, 12 }, // 2*vpblendvb sequence. 816 { ISD::SHL, MVT::v8i16, 5 }, // extend/vpsrlvd/pack sequence. 817 { ISD::SHL, MVT::v16i16, 7 }, // extend/vpsrlvd/pack sequence. 818 { ISD::SHL, MVT::v32i16, 14 }, // 2*extend/vpsrlvd/pack sequence. 819 820 { ISD::SRL, MVT::v16i8, 6 }, // vpblendvb sequence. 821 { ISD::SRL, MVT::v32i8, 6 }, // vpblendvb sequence. 822 { ISD::SRL, MVT::v64i8, 12 }, // 2*vpblendvb sequence. 823 { ISD::SRL, MVT::v8i16, 5 }, // extend/vpsrlvd/pack sequence. 824 { ISD::SRL, MVT::v16i16, 7 }, // extend/vpsrlvd/pack sequence. 825 { ISD::SRL, MVT::v32i16, 14 }, // 2*extend/vpsrlvd/pack sequence. 826 827 { ISD::SRA, MVT::v16i8, 17 }, // vpblendvb sequence. 828 { ISD::SRA, MVT::v32i8, 17 }, // vpblendvb sequence. 829 { ISD::SRA, MVT::v64i8, 34 }, // 2*vpblendvb sequence. 830 { ISD::SRA, MVT::v8i16, 5 }, // extend/vpsravd/pack sequence. 831 { ISD::SRA, MVT::v16i16, 7 }, // extend/vpsravd/pack sequence. 832 { ISD::SRA, MVT::v32i16, 14 }, // 2*extend/vpsravd/pack sequence. 833 { ISD::SRA, MVT::v2i64, 2 }, // srl/xor/sub sequence. 834 { ISD::SRA, MVT::v4i64, 2 }, // srl/xor/sub sequence. 835 836 { ISD::SUB, MVT::v32i8, 1 }, // psubb 837 { ISD::ADD, MVT::v32i8, 1 }, // paddb 838 { ISD::SUB, MVT::v16i16, 1 }, // psubw 839 { ISD::ADD, MVT::v16i16, 1 }, // paddw 840 { ISD::SUB, MVT::v8i32, 1 }, // psubd 841 { ISD::ADD, MVT::v8i32, 1 }, // paddd 842 { ISD::SUB, MVT::v4i64, 1 }, // psubq 843 { ISD::ADD, MVT::v4i64, 1 }, // paddq 844 845 { ISD::MUL, MVT::v16i16, 1 }, // pmullw 846 { ISD::MUL, MVT::v8i32, 2 }, // pmulld (Haswell from agner.org) 847 { ISD::MUL, MVT::v4i64, 6 }, // 3*pmuludq/3*shift/2*add 848 849 { ISD::FNEG, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 850 { ISD::FNEG, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 851 { ISD::FADD, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 852 { ISD::FADD, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 853 { ISD::FSUB, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 854 { ISD::FSUB, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 855 { ISD::FMUL, MVT::f64, 1 }, // Haswell from http://www.agner.org/ 856 { ISD::FMUL, MVT::v2f64, 1 }, // Haswell from http://www.agner.org/ 857 { ISD::FMUL, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 858 { ISD::FMUL, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 859 860 { ISD::FDIV, MVT::f32, 7 }, // Haswell from http://www.agner.org/ 861 { ISD::FDIV, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/ 862 { ISD::FDIV, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/ 863 { ISD::FDIV, MVT::f64, 14 }, // Haswell from http://www.agner.org/ 864 { ISD::FDIV, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/ 865 { ISD::FDIV, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/ 866 }; 867 868 // Look for AVX2 lowering tricks for custom cases. 869 if (ST->hasAVX2()) 870 if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second)) 871 return LT.first * Entry->Cost; 872 873 static const CostTblEntry AVX1CostTable[] = { 874 // We don't have to scalarize unsupported ops. We can issue two half-sized 875 // operations and we only need to extract the upper YMM half. 876 // Two ops + 1 extract + 1 insert = 4. 877 { ISD::MUL, MVT::v16i16, 4 }, 878 { ISD::MUL, MVT::v8i32, 5 }, // BTVER2 from http://www.agner.org/ 879 { ISD::MUL, MVT::v4i64, 12 }, 880 881 { ISD::SUB, MVT::v32i8, 4 }, 882 { ISD::ADD, MVT::v32i8, 4 }, 883 { ISD::SUB, MVT::v16i16, 4 }, 884 { ISD::ADD, MVT::v16i16, 4 }, 885 { ISD::SUB, MVT::v8i32, 4 }, 886 { ISD::ADD, MVT::v8i32, 4 }, 887 { ISD::SUB, MVT::v4i64, 4 }, 888 { ISD::ADD, MVT::v4i64, 4 }, 889 890 { ISD::SHL, MVT::v32i8, 22 }, // pblendvb sequence + split. 891 { ISD::SHL, MVT::v8i16, 6 }, // pblendvb sequence. 892 { ISD::SHL, MVT::v16i16, 13 }, // pblendvb sequence + split. 893 { ISD::SHL, MVT::v4i32, 3 }, // pslld/paddd/cvttps2dq/pmulld 894 { ISD::SHL, MVT::v8i32, 9 }, // pslld/paddd/cvttps2dq/pmulld + split 895 { ISD::SHL, MVT::v2i64, 2 }, // Shift each lane + blend. 896 { ISD::SHL, MVT::v4i64, 6 }, // Shift each lane + blend + split. 897 898 { ISD::SRL, MVT::v32i8, 23 }, // pblendvb sequence + split. 899 { ISD::SRL, MVT::v16i16, 28 }, // pblendvb sequence + split. 900 { ISD::SRL, MVT::v4i32, 6 }, // Shift each lane + blend. 901 { ISD::SRL, MVT::v8i32, 14 }, // Shift each lane + blend + split. 902 { ISD::SRL, MVT::v2i64, 2 }, // Shift each lane + blend. 903 { ISD::SRL, MVT::v4i64, 6 }, // Shift each lane + blend + split. 904 905 { ISD::SRA, MVT::v32i8, 44 }, // pblendvb sequence + split. 906 { ISD::SRA, MVT::v16i16, 28 }, // pblendvb sequence + split. 907 { ISD::SRA, MVT::v4i32, 6 }, // Shift each lane + blend. 908 { ISD::SRA, MVT::v8i32, 14 }, // Shift each lane + blend + split. 909 { ISD::SRA, MVT::v2i64, 5 }, // Shift each lane + blend. 910 { ISD::SRA, MVT::v4i64, 12 }, // Shift each lane + blend + split. 911 912 { ISD::FNEG, MVT::v4f64, 2 }, // BTVER2 from http://www.agner.org/ 913 { ISD::FNEG, MVT::v8f32, 2 }, // BTVER2 from http://www.agner.org/ 914 915 { ISD::FMUL, MVT::f64, 2 }, // BTVER2 from http://www.agner.org/ 916 { ISD::FMUL, MVT::v2f64, 2 }, // BTVER2 from http://www.agner.org/ 917 { ISD::FMUL, MVT::v4f64, 4 }, // BTVER2 from http://www.agner.org/ 918 919 { ISD::FDIV, MVT::f32, 14 }, // SNB from http://www.agner.org/ 920 { ISD::FDIV, MVT::v4f32, 14 }, // SNB from http://www.agner.org/ 921 { ISD::FDIV, MVT::v8f32, 28 }, // SNB from http://www.agner.org/ 922 { ISD::FDIV, MVT::f64, 22 }, // SNB from http://www.agner.org/ 923 { ISD::FDIV, MVT::v2f64, 22 }, // SNB from http://www.agner.org/ 924 { ISD::FDIV, MVT::v4f64, 44 }, // SNB from http://www.agner.org/ 925 }; 926 927 if (ST->hasAVX()) 928 if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second)) 929 return LT.first * Entry->Cost; 930 931 static const CostTblEntry SSE42CostTable[] = { 932 { ISD::FADD, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 933 { ISD::FADD, MVT::f32, 1 }, // Nehalem from http://www.agner.org/ 934 { ISD::FADD, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 935 { ISD::FADD, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 936 937 { ISD::FSUB, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 938 { ISD::FSUB, MVT::f32 , 1 }, // Nehalem from http://www.agner.org/ 939 { ISD::FSUB, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 940 { ISD::FSUB, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 941 942 { ISD::FMUL, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 943 { ISD::FMUL, MVT::f32, 1 }, // Nehalem from http://www.agner.org/ 944 { ISD::FMUL, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 945 { ISD::FMUL, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 946 947 { ISD::FDIV, MVT::f32, 14 }, // Nehalem from http://www.agner.org/ 948 { ISD::FDIV, MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/ 949 { ISD::FDIV, MVT::f64, 22 }, // Nehalem from http://www.agner.org/ 950 { ISD::FDIV, MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/ 951 952 { ISD::MUL, MVT::v2i64, 6 } // 3*pmuludq/3*shift/2*add 953 }; 954 955 if (ST->hasSSE42()) 956 if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second)) 957 return LT.first * Entry->Cost; 958 959 static const CostTblEntry SSE41CostTable[] = { 960 { ISD::SHL, MVT::v16i8, 10 }, // pblendvb sequence. 961 { ISD::SHL, MVT::v8i16, 11 }, // pblendvb sequence. 962 { ISD::SHL, MVT::v4i32, 4 }, // pslld/paddd/cvttps2dq/pmulld 963 964 { ISD::SRL, MVT::v16i8, 11 }, // pblendvb sequence. 965 { ISD::SRL, MVT::v8i16, 13 }, // pblendvb sequence. 966 { ISD::SRL, MVT::v4i32, 16 }, // Shift each lane + blend. 967 968 { ISD::SRA, MVT::v16i8, 21 }, // pblendvb sequence. 969 { ISD::SRA, MVT::v8i16, 13 }, // pblendvb sequence. 970 971 { ISD::MUL, MVT::v4i32, 2 } // pmulld (Nehalem from agner.org) 972 }; 973 974 if (ST->hasSSE41()) 975 if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second)) 976 return LT.first * Entry->Cost; 977 978 static const CostTblEntry SSE2CostTable[] = { 979 // We don't correctly identify costs of casts because they are marked as 980 // custom. 981 { ISD::SHL, MVT::v16i8, 13 }, // cmpgtb sequence. 982 { ISD::SHL, MVT::v8i16, 25 }, // cmpgtw sequence. 983 { ISD::SHL, MVT::v4i32, 16 }, // pslld/paddd/cvttps2dq/pmuludq. 984 { ISD::SHL, MVT::v2i64, 4 }, // splat+shuffle sequence. 985 986 { ISD::SRL, MVT::v16i8, 14 }, // cmpgtb sequence. 987 { ISD::SRL, MVT::v8i16, 16 }, // cmpgtw sequence. 988 { ISD::SRL, MVT::v4i32, 12 }, // Shift each lane + blend. 989 { ISD::SRL, MVT::v2i64, 4 }, // splat+shuffle sequence. 990 991 { ISD::SRA, MVT::v16i8, 27 }, // unpacked cmpgtb sequence. 992 { ISD::SRA, MVT::v8i16, 16 }, // cmpgtw sequence. 993 { ISD::SRA, MVT::v4i32, 12 }, // Shift each lane + blend. 994 { ISD::SRA, MVT::v2i64, 8 }, // srl/xor/sub splat+shuffle sequence. 995 996 { ISD::MUL, MVT::v8i16, 1 }, // pmullw 997 { ISD::MUL, MVT::v4i32, 6 }, // 3*pmuludq/4*shuffle 998 { ISD::MUL, MVT::v2i64, 8 }, // 3*pmuludq/3*shift/2*add 999 1000 { ISD::FDIV, MVT::f32, 23 }, // Pentium IV from http://www.agner.org/ 1001 { ISD::FDIV, MVT::v4f32, 39 }, // Pentium IV from http://www.agner.org/ 1002 { ISD::FDIV, MVT::f64, 38 }, // Pentium IV from http://www.agner.org/ 1003 { ISD::FDIV, MVT::v2f64, 69 }, // Pentium IV from http://www.agner.org/ 1004 1005 { ISD::FNEG, MVT::f32, 1 }, // Pentium IV from http://www.agner.org/ 1006 { ISD::FNEG, MVT::f64, 1 }, // Pentium IV from http://www.agner.org/ 1007 { ISD::FNEG, MVT::v4f32, 1 }, // Pentium IV from http://www.agner.org/ 1008 { ISD::FNEG, MVT::v2f64, 1 }, // Pentium IV from http://www.agner.org/ 1009 1010 { ISD::FADD, MVT::f32, 2 }, // Pentium IV from http://www.agner.org/ 1011 { ISD::FADD, MVT::f64, 2 }, // Pentium IV from http://www.agner.org/ 1012 1013 { ISD::FSUB, MVT::f32, 2 }, // Pentium IV from http://www.agner.org/ 1014 { ISD::FSUB, MVT::f64, 2 }, // Pentium IV from http://www.agner.org/ 1015 }; 1016 1017 if (ST->hasSSE2()) 1018 if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second)) 1019 return LT.first * Entry->Cost; 1020 1021 static const CostTblEntry SSE1CostTable[] = { 1022 { ISD::FDIV, MVT::f32, 17 }, // Pentium III from http://www.agner.org/ 1023 { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/ 1024 1025 { ISD::FNEG, MVT::f32, 2 }, // Pentium III from http://www.agner.org/ 1026 { ISD::FNEG, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/ 1027 1028 { ISD::FADD, MVT::f32, 1 }, // Pentium III from http://www.agner.org/ 1029 { ISD::FADD, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/ 1030 1031 { ISD::FSUB, MVT::f32, 1 }, // Pentium III from http://www.agner.org/ 1032 { ISD::FSUB, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/ 1033 }; 1034 1035 if (ST->hasSSE1()) 1036 if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second)) 1037 return LT.first * Entry->Cost; 1038 1039 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 1040 { ISD::ADD, MVT::i64, 1 }, // Core (Merom) from http://www.agner.org/ 1041 { ISD::SUB, MVT::i64, 1 }, // Core (Merom) from http://www.agner.org/ 1042 { ISD::MUL, MVT::i64, 2 }, // Nehalem from http://www.agner.org/ 1043 }; 1044 1045 if (ST->is64Bit()) 1046 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, LT.second)) 1047 return LT.first * Entry->Cost; 1048 1049 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 1050 { ISD::ADD, MVT::i8, 1 }, // Pentium III from http://www.agner.org/ 1051 { ISD::ADD, MVT::i16, 1 }, // Pentium III from http://www.agner.org/ 1052 { ISD::ADD, MVT::i32, 1 }, // Pentium III from http://www.agner.org/ 1053 1054 { ISD::SUB, MVT::i8, 1 }, // Pentium III from http://www.agner.org/ 1055 { ISD::SUB, MVT::i16, 1 }, // Pentium III from http://www.agner.org/ 1056 { ISD::SUB, MVT::i32, 1 }, // Pentium III from http://www.agner.org/ 1057 }; 1058 1059 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, LT.second)) 1060 return LT.first * Entry->Cost; 1061 1062 // It is not a good idea to vectorize division. We have to scalarize it and 1063 // in the process we will often end up having to spilling regular 1064 // registers. The overhead of division is going to dominate most kernels 1065 // anyways so try hard to prevent vectorization of division - it is 1066 // generally a bad idea. Assume somewhat arbitrarily that we have to be able 1067 // to hide "20 cycles" for each lane. 1068 if (LT.second.isVector() && (ISD == ISD::SDIV || ISD == ISD::SREM || 1069 ISD == ISD::UDIV || ISD == ISD::UREM)) { 1070 InstructionCost ScalarCost = getArithmeticInstrCost( 1071 Opcode, Ty->getScalarType(), CostKind, Op1Info, Op2Info, 1072 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 1073 return 20 * LT.first * LT.second.getVectorNumElements() * ScalarCost; 1074 } 1075 1076 // Fallback to the default implementation. 1077 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info); 1078 } 1079 1080 InstructionCost X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, 1081 VectorType *BaseTp, 1082 ArrayRef<int> Mask, int Index, 1083 VectorType *SubTp) { 1084 // 64-bit packed float vectors (v2f32) are widened to type v4f32. 1085 // 64-bit packed integer vectors (v2i32) are widened to type v4i32. 1086 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, BaseTp); 1087 1088 Kind = improveShuffleKindFromMask(Kind, Mask); 1089 // Treat Transpose as 2-op shuffles - there's no difference in lowering. 1090 if (Kind == TTI::SK_Transpose) 1091 Kind = TTI::SK_PermuteTwoSrc; 1092 1093 // For Broadcasts we are splatting the first element from the first input 1094 // register, so only need to reference that input and all the output 1095 // registers are the same. 1096 if (Kind == TTI::SK_Broadcast) 1097 LT.first = 1; 1098 1099 // Subvector extractions are free if they start at the beginning of a 1100 // vector and cheap if the subvectors are aligned. 1101 if (Kind == TTI::SK_ExtractSubvector && LT.second.isVector()) { 1102 int NumElts = LT.second.getVectorNumElements(); 1103 if ((Index % NumElts) == 0) 1104 return 0; 1105 std::pair<InstructionCost, MVT> SubLT = 1106 TLI->getTypeLegalizationCost(DL, SubTp); 1107 if (SubLT.second.isVector()) { 1108 int NumSubElts = SubLT.second.getVectorNumElements(); 1109 if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0) 1110 return SubLT.first; 1111 // Handle some cases for widening legalization. For now we only handle 1112 // cases where the original subvector was naturally aligned and evenly 1113 // fit in its legalized subvector type. 1114 // FIXME: Remove some of the alignment restrictions. 1115 // FIXME: We can use permq for 64-bit or larger extracts from 256-bit 1116 // vectors. 1117 int OrigSubElts = cast<FixedVectorType>(SubTp)->getNumElements(); 1118 if (NumSubElts > OrigSubElts && (Index % OrigSubElts) == 0 && 1119 (NumSubElts % OrigSubElts) == 0 && 1120 LT.second.getVectorElementType() == 1121 SubLT.second.getVectorElementType() && 1122 LT.second.getVectorElementType().getSizeInBits() == 1123 BaseTp->getElementType()->getPrimitiveSizeInBits()) { 1124 assert(NumElts >= NumSubElts && NumElts > OrigSubElts && 1125 "Unexpected number of elements!"); 1126 auto *VecTy = FixedVectorType::get(BaseTp->getElementType(), 1127 LT.second.getVectorNumElements()); 1128 auto *SubTy = FixedVectorType::get(BaseTp->getElementType(), 1129 SubLT.second.getVectorNumElements()); 1130 int ExtractIndex = alignDown((Index % NumElts), NumSubElts); 1131 InstructionCost ExtractCost = getShuffleCost( 1132 TTI::SK_ExtractSubvector, VecTy, None, ExtractIndex, SubTy); 1133 1134 // If the original size is 32-bits or more, we can use pshufd. Otherwise 1135 // if we have SSSE3 we can use pshufb. 1136 if (SubTp->getPrimitiveSizeInBits() >= 32 || ST->hasSSSE3()) 1137 return ExtractCost + 1; // pshufd or pshufb 1138 1139 assert(SubTp->getPrimitiveSizeInBits() == 16 && 1140 "Unexpected vector size"); 1141 1142 return ExtractCost + 2; // worst case pshufhw + pshufd 1143 } 1144 } 1145 } 1146 1147 // Subvector insertions are cheap if the subvectors are aligned. 1148 // Note that in general, the insertion starting at the beginning of a vector 1149 // isn't free, because we need to preserve the rest of the wide vector. 1150 if (Kind == TTI::SK_InsertSubvector && LT.second.isVector()) { 1151 int NumElts = LT.second.getVectorNumElements(); 1152 std::pair<InstructionCost, MVT> SubLT = 1153 TLI->getTypeLegalizationCost(DL, SubTp); 1154 if (SubLT.second.isVector()) { 1155 int NumSubElts = SubLT.second.getVectorNumElements(); 1156 if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0) 1157 return SubLT.first; 1158 } 1159 1160 // If the insertion isn't aligned, treat it like a 2-op shuffle. 1161 Kind = TTI::SK_PermuteTwoSrc; 1162 } 1163 1164 // Handle some common (illegal) sub-vector types as they are often very cheap 1165 // to shuffle even on targets without PSHUFB. 1166 EVT VT = TLI->getValueType(DL, BaseTp); 1167 if (VT.isSimple() && VT.isVector() && VT.getSizeInBits() < 128 && 1168 !ST->hasSSSE3()) { 1169 static const CostTblEntry SSE2SubVectorShuffleTbl[] = { 1170 {TTI::SK_Broadcast, MVT::v4i16, 1}, // pshuflw 1171 {TTI::SK_Broadcast, MVT::v2i16, 1}, // pshuflw 1172 {TTI::SK_Broadcast, MVT::v8i8, 2}, // punpck/pshuflw 1173 {TTI::SK_Broadcast, MVT::v4i8, 2}, // punpck/pshuflw 1174 {TTI::SK_Broadcast, MVT::v2i8, 1}, // punpck 1175 1176 {TTI::SK_Reverse, MVT::v4i16, 1}, // pshuflw 1177 {TTI::SK_Reverse, MVT::v2i16, 1}, // pshuflw 1178 {TTI::SK_Reverse, MVT::v4i8, 3}, // punpck/pshuflw/packus 1179 {TTI::SK_Reverse, MVT::v2i8, 1}, // punpck 1180 1181 {TTI::SK_PermuteTwoSrc, MVT::v4i16, 2}, // punpck/pshuflw 1182 {TTI::SK_PermuteTwoSrc, MVT::v2i16, 2}, // punpck/pshuflw 1183 {TTI::SK_PermuteTwoSrc, MVT::v8i8, 7}, // punpck/pshuflw 1184 {TTI::SK_PermuteTwoSrc, MVT::v4i8, 4}, // punpck/pshuflw 1185 {TTI::SK_PermuteTwoSrc, MVT::v2i8, 2}, // punpck 1186 1187 {TTI::SK_PermuteSingleSrc, MVT::v4i16, 1}, // pshuflw 1188 {TTI::SK_PermuteSingleSrc, MVT::v2i16, 1}, // pshuflw 1189 {TTI::SK_PermuteSingleSrc, MVT::v8i8, 5}, // punpck/pshuflw 1190 {TTI::SK_PermuteSingleSrc, MVT::v4i8, 3}, // punpck/pshuflw 1191 {TTI::SK_PermuteSingleSrc, MVT::v2i8, 1}, // punpck 1192 }; 1193 1194 if (ST->hasSSE2()) 1195 if (const auto *Entry = 1196 CostTableLookup(SSE2SubVectorShuffleTbl, Kind, VT.getSimpleVT())) 1197 return Entry->Cost; 1198 } 1199 1200 // We are going to permute multiple sources and the result will be in multiple 1201 // destinations. Providing an accurate cost only for splits where the element 1202 // type remains the same. 1203 if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) { 1204 MVT LegalVT = LT.second; 1205 if (LegalVT.isVector() && 1206 LegalVT.getVectorElementType().getSizeInBits() == 1207 BaseTp->getElementType()->getPrimitiveSizeInBits() && 1208 LegalVT.getVectorNumElements() < 1209 cast<FixedVectorType>(BaseTp)->getNumElements()) { 1210 1211 unsigned VecTySize = DL.getTypeStoreSize(BaseTp); 1212 unsigned LegalVTSize = LegalVT.getStoreSize(); 1213 // Number of source vectors after legalization: 1214 unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize; 1215 // Number of destination vectors after legalization: 1216 InstructionCost NumOfDests = LT.first; 1217 1218 auto *SingleOpTy = FixedVectorType::get(BaseTp->getElementType(), 1219 LegalVT.getVectorNumElements()); 1220 1221 InstructionCost NumOfShuffles = (NumOfSrcs - 1) * NumOfDests; 1222 return NumOfShuffles * getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy, 1223 None, 0, nullptr); 1224 } 1225 1226 return BaseT::getShuffleCost(Kind, BaseTp, Mask, Index, SubTp); 1227 } 1228 1229 // For 2-input shuffles, we must account for splitting the 2 inputs into many. 1230 if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) { 1231 // We assume that source and destination have the same vector type. 1232 InstructionCost NumOfDests = LT.first; 1233 InstructionCost NumOfShufflesPerDest = LT.first * 2 - 1; 1234 LT.first = NumOfDests * NumOfShufflesPerDest; 1235 } 1236 1237 static const CostTblEntry AVX512FP16ShuffleTbl[] = { 1238 {TTI::SK_Broadcast, MVT::v32f16, 1}, // vpbroadcastw 1239 {TTI::SK_Broadcast, MVT::v16f16, 1}, // vpbroadcastw 1240 {TTI::SK_Broadcast, MVT::v8f16, 1}, // vpbroadcastw 1241 1242 {TTI::SK_Reverse, MVT::v32f16, 2}, // vpermw 1243 {TTI::SK_Reverse, MVT::v16f16, 2}, // vpermw 1244 {TTI::SK_Reverse, MVT::v8f16, 1}, // vpshufb 1245 1246 {TTI::SK_PermuteSingleSrc, MVT::v32f16, 2}, // vpermw 1247 {TTI::SK_PermuteSingleSrc, MVT::v16f16, 2}, // vpermw 1248 {TTI::SK_PermuteSingleSrc, MVT::v8f16, 1}, // vpshufb 1249 1250 {TTI::SK_PermuteTwoSrc, MVT::v32f16, 2}, // vpermt2w 1251 {TTI::SK_PermuteTwoSrc, MVT::v16f16, 2}, // vpermt2w 1252 {TTI::SK_PermuteTwoSrc, MVT::v8f16, 2} // vpermt2w 1253 }; 1254 1255 if (!ST->useSoftFloat() && ST->hasFP16()) 1256 if (const auto *Entry = 1257 CostTableLookup(AVX512FP16ShuffleTbl, Kind, LT.second)) 1258 return LT.first * Entry->Cost; 1259 1260 static const CostTblEntry AVX512VBMIShuffleTbl[] = { 1261 {TTI::SK_Reverse, MVT::v64i8, 1}, // vpermb 1262 {TTI::SK_Reverse, MVT::v32i8, 1}, // vpermb 1263 1264 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 1}, // vpermb 1265 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 1}, // vpermb 1266 1267 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 2}, // vpermt2b 1268 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 2}, // vpermt2b 1269 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 2} // vpermt2b 1270 }; 1271 1272 if (ST->hasVBMI()) 1273 if (const auto *Entry = 1274 CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second)) 1275 return LT.first * Entry->Cost; 1276 1277 static const CostTblEntry AVX512BWShuffleTbl[] = { 1278 {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw 1279 {TTI::SK_Broadcast, MVT::v64i8, 1}, // vpbroadcastb 1280 1281 {TTI::SK_Reverse, MVT::v32i16, 2}, // vpermw 1282 {TTI::SK_Reverse, MVT::v16i16, 2}, // vpermw 1283 {TTI::SK_Reverse, MVT::v64i8, 2}, // pshufb + vshufi64x2 1284 1285 {TTI::SK_PermuteSingleSrc, MVT::v32i16, 2}, // vpermw 1286 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 2}, // vpermw 1287 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 8}, // extend to v32i16 1288 1289 {TTI::SK_PermuteTwoSrc, MVT::v32i16, 2}, // vpermt2w 1290 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 2}, // vpermt2w 1291 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 2}, // vpermt2w 1292 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 19}, // 6 * v32i8 + 1 1293 1294 {TTI::SK_Select, MVT::v32i16, 1}, // vblendmw 1295 {TTI::SK_Select, MVT::v64i8, 1}, // vblendmb 1296 }; 1297 1298 if (ST->hasBWI()) 1299 if (const auto *Entry = 1300 CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second)) 1301 return LT.first * Entry->Cost; 1302 1303 static const CostTblEntry AVX512ShuffleTbl[] = { 1304 {TTI::SK_Broadcast, MVT::v8f64, 1}, // vbroadcastpd 1305 {TTI::SK_Broadcast, MVT::v16f32, 1}, // vbroadcastps 1306 {TTI::SK_Broadcast, MVT::v8i64, 1}, // vpbroadcastq 1307 {TTI::SK_Broadcast, MVT::v16i32, 1}, // vpbroadcastd 1308 {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw 1309 {TTI::SK_Broadcast, MVT::v64i8, 1}, // vpbroadcastb 1310 1311 {TTI::SK_Reverse, MVT::v8f64, 1}, // vpermpd 1312 {TTI::SK_Reverse, MVT::v16f32, 1}, // vpermps 1313 {TTI::SK_Reverse, MVT::v8i64, 1}, // vpermq 1314 {TTI::SK_Reverse, MVT::v16i32, 1}, // vpermd 1315 {TTI::SK_Reverse, MVT::v32i16, 7}, // per mca 1316 {TTI::SK_Reverse, MVT::v64i8, 7}, // per mca 1317 1318 {TTI::SK_PermuteSingleSrc, MVT::v8f64, 1}, // vpermpd 1319 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1}, // vpermpd 1320 {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // vpermpd 1321 {TTI::SK_PermuteSingleSrc, MVT::v16f32, 1}, // vpermps 1322 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1}, // vpermps 1323 {TTI::SK_PermuteSingleSrc, MVT::v4f32, 1}, // vpermps 1324 {TTI::SK_PermuteSingleSrc, MVT::v8i64, 1}, // vpermq 1325 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1}, // vpermq 1326 {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // vpermq 1327 {TTI::SK_PermuteSingleSrc, MVT::v16i32, 1}, // vpermd 1328 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1}, // vpermd 1329 {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // vpermd 1330 {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb 1331 1332 {TTI::SK_PermuteTwoSrc, MVT::v8f64, 1}, // vpermt2pd 1333 {TTI::SK_PermuteTwoSrc, MVT::v16f32, 1}, // vpermt2ps 1334 {TTI::SK_PermuteTwoSrc, MVT::v8i64, 1}, // vpermt2q 1335 {TTI::SK_PermuteTwoSrc, MVT::v16i32, 1}, // vpermt2d 1336 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 1}, // vpermt2pd 1337 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 1}, // vpermt2ps 1338 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 1}, // vpermt2q 1339 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 1}, // vpermt2d 1340 {TTI::SK_PermuteTwoSrc, MVT::v2f64, 1}, // vpermt2pd 1341 {TTI::SK_PermuteTwoSrc, MVT::v4f32, 1}, // vpermt2ps 1342 {TTI::SK_PermuteTwoSrc, MVT::v2i64, 1}, // vpermt2q 1343 {TTI::SK_PermuteTwoSrc, MVT::v4i32, 1}, // vpermt2d 1344 1345 // FIXME: This just applies the type legalization cost rules above 1346 // assuming these completely split. 1347 {TTI::SK_PermuteSingleSrc, MVT::v32i16, 14}, 1348 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 14}, 1349 {TTI::SK_PermuteTwoSrc, MVT::v32i16, 42}, 1350 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 42}, 1351 1352 {TTI::SK_Select, MVT::v32i16, 1}, // vpternlogq 1353 {TTI::SK_Select, MVT::v64i8, 1}, // vpternlogq 1354 {TTI::SK_Select, MVT::v8f64, 1}, // vblendmpd 1355 {TTI::SK_Select, MVT::v16f32, 1}, // vblendmps 1356 {TTI::SK_Select, MVT::v8i64, 1}, // vblendmq 1357 {TTI::SK_Select, MVT::v16i32, 1}, // vblendmd 1358 }; 1359 1360 if (ST->hasAVX512()) 1361 if (const auto *Entry = CostTableLookup(AVX512ShuffleTbl, Kind, LT.second)) 1362 return LT.first * Entry->Cost; 1363 1364 static const CostTblEntry AVX2ShuffleTbl[] = { 1365 {TTI::SK_Broadcast, MVT::v4f64, 1}, // vbroadcastpd 1366 {TTI::SK_Broadcast, MVT::v8f32, 1}, // vbroadcastps 1367 {TTI::SK_Broadcast, MVT::v4i64, 1}, // vpbroadcastq 1368 {TTI::SK_Broadcast, MVT::v8i32, 1}, // vpbroadcastd 1369 {TTI::SK_Broadcast, MVT::v16i16, 1}, // vpbroadcastw 1370 {TTI::SK_Broadcast, MVT::v32i8, 1}, // vpbroadcastb 1371 1372 {TTI::SK_Reverse, MVT::v4f64, 1}, // vpermpd 1373 {TTI::SK_Reverse, MVT::v8f32, 1}, // vpermps 1374 {TTI::SK_Reverse, MVT::v4i64, 1}, // vpermq 1375 {TTI::SK_Reverse, MVT::v8i32, 1}, // vpermd 1376 {TTI::SK_Reverse, MVT::v16i16, 2}, // vperm2i128 + pshufb 1377 {TTI::SK_Reverse, MVT::v32i8, 2}, // vperm2i128 + pshufb 1378 1379 {TTI::SK_Select, MVT::v16i16, 1}, // vpblendvb 1380 {TTI::SK_Select, MVT::v32i8, 1}, // vpblendvb 1381 1382 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1}, // vpermpd 1383 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1}, // vpermps 1384 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1}, // vpermq 1385 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1}, // vpermd 1386 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vperm2i128 + 2*vpshufb 1387 // + vpblendvb 1388 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4}, // vperm2i128 + 2*vpshufb 1389 // + vpblendvb 1390 1391 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3}, // 2*vpermpd + vblendpd 1392 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 3}, // 2*vpermps + vblendps 1393 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3}, // 2*vpermq + vpblendd 1394 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 3}, // 2*vpermd + vpblendd 1395 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 7}, // 2*vperm2i128 + 4*vpshufb 1396 // + vpblendvb 1397 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 7}, // 2*vperm2i128 + 4*vpshufb 1398 // + vpblendvb 1399 }; 1400 1401 if (ST->hasAVX2()) 1402 if (const auto *Entry = CostTableLookup(AVX2ShuffleTbl, Kind, LT.second)) 1403 return LT.first * Entry->Cost; 1404 1405 static const CostTblEntry XOPShuffleTbl[] = { 1406 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2}, // vperm2f128 + vpermil2pd 1407 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 2}, // vperm2f128 + vpermil2ps 1408 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2}, // vperm2f128 + vpermil2pd 1409 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 2}, // vperm2f128 + vpermil2ps 1410 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vextractf128 + 2*vpperm 1411 // + vinsertf128 1412 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4}, // vextractf128 + 2*vpperm 1413 // + vinsertf128 1414 1415 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 9}, // 2*vextractf128 + 6*vpperm 1416 // + vinsertf128 1417 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 1}, // vpperm 1418 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 9}, // 2*vextractf128 + 6*vpperm 1419 // + vinsertf128 1420 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 1}, // vpperm 1421 }; 1422 1423 if (ST->hasXOP()) 1424 if (const auto *Entry = CostTableLookup(XOPShuffleTbl, Kind, LT.second)) 1425 return LT.first * Entry->Cost; 1426 1427 static const CostTblEntry AVX1ShuffleTbl[] = { 1428 {TTI::SK_Broadcast, MVT::v4f64, 2}, // vperm2f128 + vpermilpd 1429 {TTI::SK_Broadcast, MVT::v8f32, 2}, // vperm2f128 + vpermilps 1430 {TTI::SK_Broadcast, MVT::v4i64, 2}, // vperm2f128 + vpermilpd 1431 {TTI::SK_Broadcast, MVT::v8i32, 2}, // vperm2f128 + vpermilps 1432 {TTI::SK_Broadcast, MVT::v16i16, 3}, // vpshuflw + vpshufd + vinsertf128 1433 {TTI::SK_Broadcast, MVT::v32i8, 2}, // vpshufb + vinsertf128 1434 1435 {TTI::SK_Reverse, MVT::v4f64, 2}, // vperm2f128 + vpermilpd 1436 {TTI::SK_Reverse, MVT::v8f32, 2}, // vperm2f128 + vpermilps 1437 {TTI::SK_Reverse, MVT::v4i64, 2}, // vperm2f128 + vpermilpd 1438 {TTI::SK_Reverse, MVT::v8i32, 2}, // vperm2f128 + vpermilps 1439 {TTI::SK_Reverse, MVT::v16i16, 4}, // vextractf128 + 2*pshufb 1440 // + vinsertf128 1441 {TTI::SK_Reverse, MVT::v32i8, 4}, // vextractf128 + 2*pshufb 1442 // + vinsertf128 1443 1444 {TTI::SK_Select, MVT::v4i64, 1}, // vblendpd 1445 {TTI::SK_Select, MVT::v4f64, 1}, // vblendpd 1446 {TTI::SK_Select, MVT::v8i32, 1}, // vblendps 1447 {TTI::SK_Select, MVT::v8f32, 1}, // vblendps 1448 {TTI::SK_Select, MVT::v16i16, 3}, // vpand + vpandn + vpor 1449 {TTI::SK_Select, MVT::v32i8, 3}, // vpand + vpandn + vpor 1450 1451 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2}, // vperm2f128 + vshufpd 1452 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2}, // vperm2f128 + vshufpd 1453 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 4}, // 2*vperm2f128 + 2*vshufps 1454 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 4}, // 2*vperm2f128 + 2*vshufps 1455 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 8}, // vextractf128 + 4*pshufb 1456 // + 2*por + vinsertf128 1457 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 8}, // vextractf128 + 4*pshufb 1458 // + 2*por + vinsertf128 1459 1460 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3}, // 2*vperm2f128 + vshufpd 1461 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3}, // 2*vperm2f128 + vshufpd 1462 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 4}, // 2*vperm2f128 + 2*vshufps 1463 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 4}, // 2*vperm2f128 + 2*vshufps 1464 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 15}, // 2*vextractf128 + 8*pshufb 1465 // + 4*por + vinsertf128 1466 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 15}, // 2*vextractf128 + 8*pshufb 1467 // + 4*por + vinsertf128 1468 }; 1469 1470 if (ST->hasAVX()) 1471 if (const auto *Entry = CostTableLookup(AVX1ShuffleTbl, Kind, LT.second)) 1472 return LT.first * Entry->Cost; 1473 1474 static const CostTblEntry SSE41ShuffleTbl[] = { 1475 {TTI::SK_Select, MVT::v2i64, 1}, // pblendw 1476 {TTI::SK_Select, MVT::v2f64, 1}, // movsd 1477 {TTI::SK_Select, MVT::v4i32, 1}, // pblendw 1478 {TTI::SK_Select, MVT::v4f32, 1}, // blendps 1479 {TTI::SK_Select, MVT::v8i16, 1}, // pblendw 1480 {TTI::SK_Select, MVT::v16i8, 1} // pblendvb 1481 }; 1482 1483 if (ST->hasSSE41()) 1484 if (const auto *Entry = CostTableLookup(SSE41ShuffleTbl, Kind, LT.second)) 1485 return LT.first * Entry->Cost; 1486 1487 static const CostTblEntry SSSE3ShuffleTbl[] = { 1488 {TTI::SK_Broadcast, MVT::v8i16, 1}, // pshufb 1489 {TTI::SK_Broadcast, MVT::v16i8, 1}, // pshufb 1490 1491 {TTI::SK_Reverse, MVT::v8i16, 1}, // pshufb 1492 {TTI::SK_Reverse, MVT::v16i8, 1}, // pshufb 1493 1494 {TTI::SK_Select, MVT::v8i16, 3}, // 2*pshufb + por 1495 {TTI::SK_Select, MVT::v16i8, 3}, // 2*pshufb + por 1496 1497 {TTI::SK_PermuteSingleSrc, MVT::v8i16, 1}, // pshufb 1498 {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb 1499 1500 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 3}, // 2*pshufb + por 1501 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 3}, // 2*pshufb + por 1502 }; 1503 1504 if (ST->hasSSSE3()) 1505 if (const auto *Entry = CostTableLookup(SSSE3ShuffleTbl, Kind, LT.second)) 1506 return LT.first * Entry->Cost; 1507 1508 static const CostTblEntry SSE2ShuffleTbl[] = { 1509 {TTI::SK_Broadcast, MVT::v2f64, 1}, // shufpd 1510 {TTI::SK_Broadcast, MVT::v2i64, 1}, // pshufd 1511 {TTI::SK_Broadcast, MVT::v4i32, 1}, // pshufd 1512 {TTI::SK_Broadcast, MVT::v8i16, 2}, // pshuflw + pshufd 1513 {TTI::SK_Broadcast, MVT::v16i8, 3}, // unpck + pshuflw + pshufd 1514 1515 {TTI::SK_Reverse, MVT::v2f64, 1}, // shufpd 1516 {TTI::SK_Reverse, MVT::v2i64, 1}, // pshufd 1517 {TTI::SK_Reverse, MVT::v4i32, 1}, // pshufd 1518 {TTI::SK_Reverse, MVT::v8i16, 3}, // pshuflw + pshufhw + pshufd 1519 {TTI::SK_Reverse, MVT::v16i8, 9}, // 2*pshuflw + 2*pshufhw 1520 // + 2*pshufd + 2*unpck + packus 1521 1522 {TTI::SK_Select, MVT::v2i64, 1}, // movsd 1523 {TTI::SK_Select, MVT::v2f64, 1}, // movsd 1524 {TTI::SK_Select, MVT::v4i32, 2}, // 2*shufps 1525 {TTI::SK_Select, MVT::v8i16, 3}, // pand + pandn + por 1526 {TTI::SK_Select, MVT::v16i8, 3}, // pand + pandn + por 1527 1528 {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // shufpd 1529 {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // pshufd 1530 {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // pshufd 1531 {TTI::SK_PermuteSingleSrc, MVT::v8i16, 5}, // 2*pshuflw + 2*pshufhw 1532 // + pshufd/unpck 1533 { TTI::SK_PermuteSingleSrc, MVT::v16i8, 10 }, // 2*pshuflw + 2*pshufhw 1534 // + 2*pshufd + 2*unpck + 2*packus 1535 1536 { TTI::SK_PermuteTwoSrc, MVT::v2f64, 1 }, // shufpd 1537 { TTI::SK_PermuteTwoSrc, MVT::v2i64, 1 }, // shufpd 1538 { TTI::SK_PermuteTwoSrc, MVT::v4i32, 2 }, // 2*{unpck,movsd,pshufd} 1539 { TTI::SK_PermuteTwoSrc, MVT::v8i16, 8 }, // blend+permute 1540 { TTI::SK_PermuteTwoSrc, MVT::v16i8, 13 }, // blend+permute 1541 }; 1542 1543 if (ST->hasSSE2()) 1544 if (const auto *Entry = CostTableLookup(SSE2ShuffleTbl, Kind, LT.second)) 1545 return LT.first * Entry->Cost; 1546 1547 static const CostTblEntry SSE1ShuffleTbl[] = { 1548 { TTI::SK_Broadcast, MVT::v4f32, 1 }, // shufps 1549 { TTI::SK_Reverse, MVT::v4f32, 1 }, // shufps 1550 { TTI::SK_Select, MVT::v4f32, 2 }, // 2*shufps 1551 { TTI::SK_PermuteSingleSrc, MVT::v4f32, 1 }, // shufps 1552 { TTI::SK_PermuteTwoSrc, MVT::v4f32, 2 }, // 2*shufps 1553 }; 1554 1555 if (ST->hasSSE1()) 1556 if (const auto *Entry = CostTableLookup(SSE1ShuffleTbl, Kind, LT.second)) 1557 return LT.first * Entry->Cost; 1558 1559 return BaseT::getShuffleCost(Kind, BaseTp, Mask, Index, SubTp); 1560 } 1561 1562 InstructionCost X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, 1563 Type *Src, 1564 TTI::CastContextHint CCH, 1565 TTI::TargetCostKind CostKind, 1566 const Instruction *I) { 1567 int ISD = TLI->InstructionOpcodeToISD(Opcode); 1568 assert(ISD && "Invalid opcode"); 1569 1570 // TODO: Allow non-throughput costs that aren't binary. 1571 auto AdjustCost = [&CostKind](InstructionCost Cost) -> InstructionCost { 1572 if (CostKind != TTI::TCK_RecipThroughput) 1573 return Cost == 0 ? 0 : 1; 1574 return Cost; 1575 }; 1576 1577 // The cost tables include both specific, custom (non-legal) src/dst type 1578 // conversions and generic, legalized types. We test for customs first, before 1579 // falling back to legalization. 1580 // FIXME: Need a better design of the cost table to handle non-simple types of 1581 // potential massive combinations (elem_num x src_type x dst_type). 1582 static const TypeConversionCostTblEntry AVX512BWConversionTbl[] { 1583 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 1 }, 1584 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 1 }, 1585 1586 // Mask sign extend has an instruction. 1587 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 1 }, 1588 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 1 }, 1589 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 1 }, 1590 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 1 }, 1591 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 1 }, 1592 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 1 }, 1593 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 1 }, 1594 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1595 { ISD::SIGN_EXTEND, MVT::v32i8, MVT::v32i1, 1 }, 1596 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i1, 1 }, 1597 { ISD::SIGN_EXTEND, MVT::v64i8, MVT::v64i1, 1 }, 1598 1599 // Mask zero extend is a sext + shift. 1600 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 2 }, 1601 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 2 }, 1602 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 2 }, 1603 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 2 }, 1604 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 2 }, 1605 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 2 }, 1606 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 2 }, 1607 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 }, 1608 { ISD::ZERO_EXTEND, MVT::v32i8, MVT::v32i1, 2 }, 1609 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i1, 2 }, 1610 { ISD::ZERO_EXTEND, MVT::v64i8, MVT::v64i1, 2 }, 1611 1612 { ISD::TRUNCATE, MVT::v32i8, MVT::v32i16, 2 }, 1613 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 2 }, // widen to zmm 1614 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 2 }, // widen to zmm 1615 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // widen to zmm 1616 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i16, 2 }, // vpmovwb 1617 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // widen to zmm 1618 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 2 }, // widen to zmm 1619 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i16, 2 }, // vpmovwb 1620 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 2 }, // widen to zmm 1621 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 2 }, // widen to zmm 1622 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 2 }, // vpmovwb 1623 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 2 }, // widen to zmm 1624 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 2 }, // widen to zmm 1625 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i8, 2 }, // widen to zmm 1626 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i16, 2 }, 1627 { ISD::TRUNCATE, MVT::v64i1, MVT::v64i8, 2 }, 1628 }; 1629 1630 static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = { 1631 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i64, 1 }, 1632 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i64, 1 }, 1633 1634 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i64, 1 }, 1635 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 1 }, 1636 1637 { ISD::FP_TO_SINT, MVT::v8i64, MVT::v8f32, 1 }, 1638 { ISD::FP_TO_SINT, MVT::v8i64, MVT::v8f64, 1 }, 1639 1640 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f32, 1 }, 1641 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f64, 1 }, 1642 }; 1643 1644 // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and 1645 // 256-bit wide vectors. 1646 1647 static const TypeConversionCostTblEntry AVX512FConversionTbl[] = { 1648 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 1 }, 1649 { ISD::FP_EXTEND, MVT::v8f64, MVT::v16f32, 3 }, 1650 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 1 }, 1651 1652 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // sext+vpslld+vptestmd 1653 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 3 }, // sext+vpslld+vptestmd 1654 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 3 }, // sext+vpslld+vptestmd 1655 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 3 }, // sext+vpslld+vptestmd 1656 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 3 }, // sext+vpsllq+vptestmq 1657 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 3 }, // sext+vpsllq+vptestmq 1658 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 3 }, // sext+vpsllq+vptestmq 1659 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 3 }, // sext+vpslld+vptestmd 1660 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 2 }, // zmm vpslld+vptestmd 1661 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i32, 2 }, // zmm vpslld+vptestmd 1662 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 }, // zmm vpslld+vptestmd 1663 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i32, 2 }, // vpslld+vptestmd 1664 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i64, 2 }, // zmm vpsllq+vptestmq 1665 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 2 }, // zmm vpsllq+vptestmq 1666 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i64, 2 }, // vpsllq+vptestmq 1667 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i32, 2 }, // vpmovdb 1668 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 2 }, // vpmovdb 1669 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 2 }, // vpmovdb 1670 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 2 }, // vpmovdb 1671 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i64, 2 }, // vpmovqb 1672 { ISD::TRUNCATE, MVT::v2i16, MVT::v2i64, 1 }, // vpshufb 1673 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i64, 2 }, // vpmovqb 1674 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 2 }, // vpmovqw 1675 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 1 }, // vpmovqd 1676 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 }, // zmm vpmovqd 1677 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i64, 5 },// 2*vpmovqd+concat+vpmovdb 1678 1679 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 3 }, // extend to v16i32 1680 { ISD::TRUNCATE, MVT::v32i8, MVT::v32i16, 8 }, 1681 1682 // Sign extend is zmm vpternlogd+vptruncdb. 1683 // Zero extend is zmm broadcast load+vptruncdw. 1684 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 3 }, 1685 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 4 }, 1686 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 3 }, 1687 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 4 }, 1688 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 3 }, 1689 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 4 }, 1690 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 3 }, 1691 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 4 }, 1692 1693 // Sign extend is zmm vpternlogd+vptruncdw. 1694 // Zero extend is zmm vpternlogd+vptruncdw+vpsrlw. 1695 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 3 }, 1696 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 4 }, 1697 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 3 }, 1698 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 4 }, 1699 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 3 }, 1700 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 4 }, 1701 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 3 }, 1702 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 4 }, 1703 1704 { ISD::SIGN_EXTEND, MVT::v2i32, MVT::v2i1, 1 }, // zmm vpternlogd 1705 { ISD::ZERO_EXTEND, MVT::v2i32, MVT::v2i1, 2 }, // zmm vpternlogd+psrld 1706 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i1, 1 }, // zmm vpternlogd 1707 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i1, 2 }, // zmm vpternlogd+psrld 1708 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 1 }, // zmm vpternlogd 1709 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 2 }, // zmm vpternlogd+psrld 1710 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i1, 1 }, // zmm vpternlogq 1711 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i1, 2 }, // zmm vpternlogq+psrlq 1712 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 1 }, // zmm vpternlogq 1713 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 2 }, // zmm vpternlogq+psrlq 1714 1715 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1, 1 }, // vpternlogd 1716 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1, 2 }, // vpternlogd+psrld 1717 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i1, 1 }, // vpternlogq 1718 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i1, 2 }, // vpternlogq+psrlq 1719 1720 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 1 }, 1721 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 1 }, 1722 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 }, 1723 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 }, 1724 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i8, 1 }, 1725 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i8, 1 }, 1726 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 1 }, 1727 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 1 }, 1728 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i32, 1 }, 1729 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i32, 1 }, 1730 1731 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 3 }, // FIXME: May not be right 1732 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 3 }, // FIXME: May not be right 1733 1734 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 }, 1735 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 }, 1736 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v16i8, 2 }, 1737 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i8, 1 }, 1738 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 }, 1739 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i16, 1 }, 1740 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 }, 1741 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 }, 1742 1743 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 }, 1744 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 }, 1745 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v16i8, 2 }, 1746 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i8, 1 }, 1747 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 }, 1748 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i16, 1 }, 1749 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 }, 1750 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 }, 1751 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i64, 26 }, 1752 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 5 }, 1753 1754 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f32, 2 }, 1755 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f64, 7 }, 1756 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v32f64,15 }, 1757 { ISD::FP_TO_SINT, MVT::v64i8, MVT::v64f32,11 }, 1758 { ISD::FP_TO_SINT, MVT::v64i8, MVT::v64f64,31 }, 1759 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f64, 3 }, 1760 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v16f64, 7 }, 1761 { ISD::FP_TO_SINT, MVT::v32i16, MVT::v32f32, 5 }, 1762 { ISD::FP_TO_SINT, MVT::v32i16, MVT::v32f64,15 }, 1763 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f64, 1 }, 1764 { ISD::FP_TO_SINT, MVT::v16i32, MVT::v16f64, 3 }, 1765 1766 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f64, 1 }, 1767 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f64, 3 }, 1768 { ISD::FP_TO_UINT, MVT::v8i8, MVT::v8f64, 3 }, 1769 { ISD::FP_TO_UINT, MVT::v16i32, MVT::v16f32, 1 }, 1770 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v16f32, 3 }, 1771 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v16f32, 3 }, 1772 }; 1773 1774 static const TypeConversionCostTblEntry AVX512BWVLConversionTbl[] { 1775 // Mask sign extend has an instruction. 1776 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 1 }, 1777 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 1 }, 1778 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 1 }, 1779 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 1 }, 1780 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 1 }, 1781 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 1 }, 1782 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 1 }, 1783 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1784 { ISD::SIGN_EXTEND, MVT::v32i8, MVT::v32i1, 1 }, 1785 1786 // Mask zero extend is a sext + shift. 1787 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 2 }, 1788 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 2 }, 1789 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 2 }, 1790 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 2 }, 1791 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 2 }, 1792 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 2 }, 1793 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 2 }, 1794 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 }, 1795 { ISD::ZERO_EXTEND, MVT::v32i8, MVT::v32i1, 2 }, 1796 1797 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 2 }, 1798 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 2 }, // vpsllw+vptestmb 1799 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // vpsllw+vptestmw 1800 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // vpsllw+vptestmb 1801 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 2 }, // vpsllw+vptestmw 1802 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 2 }, // vpsllw+vptestmb 1803 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 2 }, // vpsllw+vptestmw 1804 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 2 }, // vpsllw+vptestmb 1805 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 2 }, // vpsllw+vptestmw 1806 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i8, 2 }, // vpsllw+vptestmb 1807 }; 1808 1809 static const TypeConversionCostTblEntry AVX512DQVLConversionTbl[] = { 1810 { ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i64, 1 }, 1811 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 }, 1812 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i64, 1 }, 1813 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i64, 1 }, 1814 1815 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 1 }, 1816 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 }, 1817 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 1 }, 1818 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 1 }, 1819 1820 { ISD::FP_TO_SINT, MVT::v2i64, MVT::v4f32, 1 }, 1821 { ISD::FP_TO_SINT, MVT::v4i64, MVT::v4f32, 1 }, 1822 { ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f64, 1 }, 1823 { ISD::FP_TO_SINT, MVT::v4i64, MVT::v4f64, 1 }, 1824 1825 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v4f32, 1 }, 1826 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f32, 1 }, 1827 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f64, 1 }, 1828 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f64, 1 }, 1829 }; 1830 1831 static const TypeConversionCostTblEntry AVX512VLConversionTbl[] = { 1832 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // sext+vpslld+vptestmd 1833 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 3 }, // sext+vpslld+vptestmd 1834 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 3 }, // sext+vpslld+vptestmd 1835 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 8 }, // split+2*v8i8 1836 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 3 }, // sext+vpsllq+vptestmq 1837 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 3 }, // sext+vpsllq+vptestmq 1838 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 3 }, // sext+vpsllq+vptestmq 1839 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 8 }, // split+2*v8i16 1840 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 2 }, // vpslld+vptestmd 1841 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i32, 2 }, // vpslld+vptestmd 1842 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 }, // vpslld+vptestmd 1843 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i64, 2 }, // vpsllq+vptestmq 1844 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 2 }, // vpsllq+vptestmq 1845 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 }, // vpmovqd 1846 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 2 }, // vpmovqb 1847 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2 }, // vpmovqw 1848 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2 }, // vpmovwb 1849 1850 // sign extend is vpcmpeq+maskedmove+vpmovdw+vpacksswb 1851 // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw+vpackuswb 1852 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 5 }, 1853 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 6 }, 1854 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 5 }, 1855 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 6 }, 1856 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 5 }, 1857 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 6 }, 1858 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 10 }, 1859 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 12 }, 1860 1861 // sign extend is vpcmpeq+maskedmove+vpmovdw 1862 // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw 1863 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 4 }, 1864 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 5 }, 1865 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 4 }, 1866 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 5 }, 1867 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 4 }, 1868 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 5 }, 1869 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 10 }, 1870 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 12 }, 1871 1872 { ISD::SIGN_EXTEND, MVT::v2i32, MVT::v2i1, 1 }, // vpternlogd 1873 { ISD::ZERO_EXTEND, MVT::v2i32, MVT::v2i1, 2 }, // vpternlogd+psrld 1874 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i1, 1 }, // vpternlogd 1875 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i1, 2 }, // vpternlogd+psrld 1876 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 1 }, // vpternlogd 1877 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 2 }, // vpternlogd+psrld 1878 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i1, 1 }, // vpternlogq 1879 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i1, 2 }, // vpternlogq+psrlq 1880 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 1 }, // vpternlogq 1881 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 2 }, // vpternlogq+psrlq 1882 1883 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v16i8, 1 }, 1884 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v16i8, 1 }, 1885 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v16i8, 1 }, 1886 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v16i8, 1 }, 1887 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1 }, 1888 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1 }, 1889 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v8i16, 1 }, 1890 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v8i16, 1 }, 1891 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 1892 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 1893 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 1894 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 1895 1896 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 }, 1897 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v16i8, 1 }, 1898 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 }, 1899 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 1 }, 1900 1901 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 1 }, 1902 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 1 }, 1903 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 }, 1904 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v16i8, 1 }, 1905 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 }, 1906 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 1 }, 1907 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 1 }, 1908 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 1909 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 }, 1910 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 }, 1911 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 5 }, 1912 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 5 }, 1913 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 5 }, 1914 1915 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v8f32, 2 }, 1916 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f32, 2 }, 1917 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v32f32, 5 }, 1918 1919 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 1 }, 1920 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 1 }, 1921 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1 }, 1922 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 1 }, 1923 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 1 }, 1924 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 1 }, 1925 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f64, 1 }, 1926 }; 1927 1928 static const TypeConversionCostTblEntry AVX2ConversionTbl[] = { 1929 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 3 }, 1930 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 3 }, 1931 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 3 }, 1932 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 3 }, 1933 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1934 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1935 1936 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v16i8, 2 }, 1937 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v16i8, 2 }, 1938 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v16i8, 2 }, 1939 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v16i8, 2 }, 1940 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 2 }, 1941 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 2 }, 1942 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v8i16, 2 }, 1943 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v8i16, 2 }, 1944 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 2 }, 1945 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 2 }, 1946 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 3 }, 1947 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 3 }, 1948 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 2 }, 1949 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 2 }, 1950 1951 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 }, 1952 1953 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 4 }, 1954 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 4 }, 1955 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i16, 1 }, 1956 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i32, 1 }, 1957 { ISD::TRUNCATE, MVT::v16i8, MVT::v2i64, 1 }, 1958 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i32, 4 }, 1959 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i64, 4 }, 1960 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i32, 1 }, 1961 { ISD::TRUNCATE, MVT::v8i16, MVT::v2i64, 1 }, 1962 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i64, 5 }, 1963 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 }, 1964 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 2 }, 1965 1966 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 3 }, 1967 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 3 }, 1968 1969 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v8f32, 1 }, 1970 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f64, 1 }, 1971 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f32, 1 }, 1972 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f64, 3 }, 1973 1974 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 3 }, 1975 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 3 }, 1976 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v8f32, 1 }, 1977 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 3 }, 1978 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 4 }, 1979 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 4 }, 1980 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 3 }, 1981 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v4f64, 4 }, 1982 1983 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 2 }, 1984 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v16i8, 2 }, 1985 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 2 }, 1986 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 2 }, 1987 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 }, 1988 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 }, 1989 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 3 }, 1990 1991 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 2 }, 1992 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v16i8, 2 }, 1993 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 2 }, 1994 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 2 }, 1995 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 2 }, 1996 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 1 }, 1997 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 2 }, 1998 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 2 }, 1999 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 }, 2000 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 4 }, 2001 }; 2002 2003 static const TypeConversionCostTblEntry AVXConversionTbl[] = { 2004 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 6 }, 2005 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 4 }, 2006 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 7 }, 2007 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 4 }, 2008 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 4 }, 2009 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 4 }, 2010 2011 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v16i8, 3 }, 2012 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v16i8, 3 }, 2013 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v16i8, 3 }, 2014 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v16i8, 3 }, 2015 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 3 }, 2016 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 3 }, 2017 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v8i16, 3 }, 2018 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v8i16, 3 }, 2019 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 3 }, 2020 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 3 }, 2021 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 3 }, 2022 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 3 }, 2023 2024 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 4 }, 2025 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 5 }, 2026 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 4 }, 2027 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i64, 9 }, 2028 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i64, 11 }, 2029 2030 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 6 }, 2031 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 6 }, 2032 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 2 }, // and+extract+packuswb 2033 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i32, 5 }, 2034 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 }, 2035 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i64, 5 }, 2036 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i64, 3 }, // and+extract+2*packusdw 2037 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 2 }, 2038 2039 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 }, 2040 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i1, 3 }, 2041 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i1, 8 }, 2042 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v16i8, 4 }, 2043 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v16i8, 2 }, 2044 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 }, 2045 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v8i16, 2 }, 2046 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 2 }, 2047 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 }, 2048 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 4 }, 2049 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 5 }, 2050 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i64, 8 }, 2051 2052 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 7 }, 2053 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i1, 7 }, 2054 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i1, 6 }, 2055 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v16i8, 4 }, 2056 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v16i8, 2 }, 2057 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 }, 2058 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v8i16, 2 }, 2059 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 4 }, 2060 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 4 }, 2061 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 }, 2062 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 6 }, 2063 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 8 }, 2064 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 10 }, 2065 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 10 }, 2066 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 18 }, 2067 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 5 }, 2068 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 10 }, 2069 2070 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v8f32, 2 }, 2071 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v4f64, 2 }, 2072 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v8f32, 2 }, 2073 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v4f64, 2 }, 2074 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f32, 2 }, 2075 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v4f64, 2 }, 2076 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v8f32, 2 }, 2077 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v4f64, 2 }, 2078 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f64, 2 }, 2079 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f32, 2 }, 2080 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f64, 5 }, 2081 2082 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v8f32, 2 }, 2083 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v4f64, 2 }, 2084 { ISD::FP_TO_UINT, MVT::v32i8, MVT::v8f32, 2 }, 2085 { ISD::FP_TO_UINT, MVT::v32i8, MVT::v4f64, 2 }, 2086 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f32, 2 }, 2087 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v4f64, 2 }, 2088 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v8f32, 2 }, 2089 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v4f64, 2 }, 2090 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 3 }, 2091 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 4 }, 2092 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 6 }, 2093 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 7 }, 2094 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v4f64, 7 }, 2095 2096 { ISD::FP_EXTEND, MVT::v4f64, MVT::v4f32, 1 }, 2097 { ISD::FP_ROUND, MVT::v4f32, MVT::v4f64, 1 }, 2098 }; 2099 2100 static const TypeConversionCostTblEntry SSE41ConversionTbl[] = { 2101 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v16i8, 1 }, 2102 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v16i8, 1 }, 2103 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v16i8, 1 }, 2104 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v16i8, 1 }, 2105 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v16i8, 1 }, 2106 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v16i8, 1 }, 2107 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v8i16, 1 }, 2108 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v8i16, 1 }, 2109 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v8i16, 1 }, 2110 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v8i16, 1 }, 2111 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v4i32, 1 }, 2112 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v4i32, 1 }, 2113 2114 // These truncates end up widening elements. 2115 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 1 }, // PMOVXZBQ 2116 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 1 }, // PMOVXZWQ 2117 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 1 }, // PMOVXZBD 2118 2119 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i32, 2 }, 2120 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i32, 2 }, 2121 { ISD::TRUNCATE, MVT::v16i8, MVT::v2i64, 2 }, 2122 2123 { ISD::SINT_TO_FP, MVT::f32, MVT::i32, 1 }, 2124 { ISD::SINT_TO_FP, MVT::f64, MVT::i32, 1 }, 2125 { ISD::SINT_TO_FP, MVT::f32, MVT::i64, 1 }, 2126 { ISD::SINT_TO_FP, MVT::f64, MVT::i64, 1 }, 2127 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 1 }, 2128 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 }, 2129 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 1 }, 2130 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 }, 2131 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 2132 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 1 }, 2133 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 2 }, 2134 2135 { ISD::UINT_TO_FP, MVT::f32, MVT::i32, 1 }, 2136 { ISD::UINT_TO_FP, MVT::f64, MVT::i32, 1 }, 2137 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 4 }, 2138 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 4 }, 2139 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 1 }, 2140 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 }, 2141 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 1 }, 2142 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 }, 2143 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 3 }, 2144 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 3 }, 2145 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 2 }, 2146 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 12 }, 2147 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 22 }, 2148 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 4 }, 2149 2150 { ISD::FP_TO_SINT, MVT::i32, MVT::f32, 1 }, 2151 { ISD::FP_TO_SINT, MVT::i64, MVT::f32, 1 }, 2152 { ISD::FP_TO_SINT, MVT::i32, MVT::f64, 1 }, 2153 { ISD::FP_TO_SINT, MVT::i64, MVT::f64, 1 }, 2154 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v4f32, 2 }, 2155 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v2f64, 2 }, 2156 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v4f32, 1 }, 2157 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v2f64, 1 }, 2158 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 1 }, 2159 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v2f64, 1 }, 2160 2161 { ISD::FP_TO_UINT, MVT::i32, MVT::f32, 1 }, 2162 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 4 }, 2163 { ISD::FP_TO_UINT, MVT::i32, MVT::f64, 1 }, 2164 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 4 }, 2165 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v4f32, 2 }, 2166 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v2f64, 2 }, 2167 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v4f32, 1 }, 2168 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v2f64, 1 }, 2169 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 4 }, 2170 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 4 }, 2171 }; 2172 2173 static const TypeConversionCostTblEntry SSE2ConversionTbl[] = { 2174 // These are somewhat magic numbers justified by comparing the 2175 // output of llvm-mca for our various supported scheduler models 2176 // and basing it off the worst case scenario. 2177 { ISD::SINT_TO_FP, MVT::f32, MVT::i32, 3 }, 2178 { ISD::SINT_TO_FP, MVT::f64, MVT::i32, 3 }, 2179 { ISD::SINT_TO_FP, MVT::f32, MVT::i64, 3 }, 2180 { ISD::SINT_TO_FP, MVT::f64, MVT::i64, 3 }, 2181 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 3 }, 2182 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 4 }, 2183 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 3 }, 2184 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 4 }, 2185 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 3 }, 2186 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4 }, 2187 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 8 }, 2188 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 8 }, 2189 2190 { ISD::UINT_TO_FP, MVT::f32, MVT::i32, 3 }, 2191 { ISD::UINT_TO_FP, MVT::f64, MVT::i32, 3 }, 2192 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 8 }, 2193 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 9 }, 2194 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 4 }, 2195 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 4 }, 2196 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 4 }, 2197 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 4 }, 2198 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 7 }, 2199 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 7 }, 2200 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 }, 2201 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 15 }, 2202 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 18 }, 2203 2204 { ISD::FP_TO_SINT, MVT::i32, MVT::f32, 4 }, 2205 { ISD::FP_TO_SINT, MVT::i64, MVT::f32, 4 }, 2206 { ISD::FP_TO_SINT, MVT::i32, MVT::f64, 4 }, 2207 { ISD::FP_TO_SINT, MVT::i64, MVT::f64, 4 }, 2208 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v4f32, 6 }, 2209 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v2f64, 6 }, 2210 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v4f32, 5 }, 2211 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v2f64, 5 }, 2212 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 4 }, 2213 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v2f64, 4 }, 2214 2215 { ISD::FP_TO_UINT, MVT::i32, MVT::f32, 4 }, 2216 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 4 }, 2217 { ISD::FP_TO_UINT, MVT::i32, MVT::f64, 4 }, 2218 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 15 }, 2219 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v4f32, 6 }, 2220 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v2f64, 6 }, 2221 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v4f32, 5 }, 2222 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v2f64, 5 }, 2223 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 8 }, 2224 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 8 }, 2225 2226 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v16i8, 4 }, 2227 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v16i8, 4 }, 2228 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v16i8, 2 }, 2229 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v16i8, 3 }, 2230 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v16i8, 1 }, 2231 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v16i8, 2 }, 2232 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v8i16, 2 }, 2233 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v8i16, 3 }, 2234 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v8i16, 1 }, 2235 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v8i16, 2 }, 2236 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v4i32, 1 }, 2237 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v4i32, 2 }, 2238 2239 // These truncates are really widening elements. 2240 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 1 }, // PSHUFD 2241 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // PUNPCKLWD+DQ 2242 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // PUNPCKLBW+WD+PSHUFD 2243 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 1 }, // PUNPCKLWD 2244 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // PUNPCKLBW+WD 2245 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 1 }, // PUNPCKLBW 2246 2247 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i16, 2 }, // PAND+PACKUSWB 2248 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 3 }, 2249 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i32, 3 }, // PAND+2*PACKUSWB 2250 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 7 }, 2251 { ISD::TRUNCATE, MVT::v2i16, MVT::v2i32, 1 }, 2252 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i32, 3 }, 2253 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 }, 2254 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32,10 }, 2255 { ISD::TRUNCATE, MVT::v16i8, MVT::v2i64, 4 }, // PAND+3*PACKUSWB 2256 { ISD::TRUNCATE, MVT::v8i16, MVT::v2i64, 2 }, // PSHUFD+PSHUFLW 2257 { ISD::TRUNCATE, MVT::v4i32, MVT::v2i64, 1 }, // PSHUFD 2258 }; 2259 2260 // Attempt to map directly to (simple) MVT types to let us match custom entries. 2261 EVT SrcTy = TLI->getValueType(DL, Src); 2262 EVT DstTy = TLI->getValueType(DL, Dst); 2263 2264 // The function getSimpleVT only handles simple value types. 2265 if (SrcTy.isSimple() && DstTy.isSimple()) { 2266 MVT SimpleSrcTy = SrcTy.getSimpleVT(); 2267 MVT SimpleDstTy = DstTy.getSimpleVT(); 2268 2269 if (ST->useAVX512Regs()) { 2270 if (ST->hasBWI()) 2271 if (const auto *Entry = ConvertCostTableLookup( 2272 AVX512BWConversionTbl, ISD, SimpleDstTy, SimpleSrcTy)) 2273 return AdjustCost(Entry->Cost); 2274 2275 if (ST->hasDQI()) 2276 if (const auto *Entry = ConvertCostTableLookup( 2277 AVX512DQConversionTbl, ISD, SimpleDstTy, SimpleSrcTy)) 2278 return AdjustCost(Entry->Cost); 2279 2280 if (ST->hasAVX512()) 2281 if (const auto *Entry = ConvertCostTableLookup( 2282 AVX512FConversionTbl, ISD, SimpleDstTy, SimpleSrcTy)) 2283 return AdjustCost(Entry->Cost); 2284 } 2285 2286 if (ST->hasBWI()) 2287 if (const auto *Entry = ConvertCostTableLookup( 2288 AVX512BWVLConversionTbl, ISD, SimpleDstTy, SimpleSrcTy)) 2289 return AdjustCost(Entry->Cost); 2290 2291 if (ST->hasDQI()) 2292 if (const auto *Entry = ConvertCostTableLookup( 2293 AVX512DQVLConversionTbl, ISD, SimpleDstTy, SimpleSrcTy)) 2294 return AdjustCost(Entry->Cost); 2295 2296 if (ST->hasAVX512()) 2297 if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD, 2298 SimpleDstTy, SimpleSrcTy)) 2299 return AdjustCost(Entry->Cost); 2300 2301 if (ST->hasAVX2()) { 2302 if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD, 2303 SimpleDstTy, SimpleSrcTy)) 2304 return AdjustCost(Entry->Cost); 2305 } 2306 2307 if (ST->hasAVX()) { 2308 if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD, 2309 SimpleDstTy, SimpleSrcTy)) 2310 return AdjustCost(Entry->Cost); 2311 } 2312 2313 if (ST->hasSSE41()) { 2314 if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD, 2315 SimpleDstTy, SimpleSrcTy)) 2316 return AdjustCost(Entry->Cost); 2317 } 2318 2319 if (ST->hasSSE2()) { 2320 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD, 2321 SimpleDstTy, SimpleSrcTy)) 2322 return AdjustCost(Entry->Cost); 2323 } 2324 } 2325 2326 // Fall back to legalized types. 2327 std::pair<InstructionCost, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src); 2328 std::pair<InstructionCost, MVT> LTDest = 2329 TLI->getTypeLegalizationCost(DL, Dst); 2330 2331 if (ST->useAVX512Regs()) { 2332 if (ST->hasBWI()) 2333 if (const auto *Entry = ConvertCostTableLookup( 2334 AVX512BWConversionTbl, ISD, LTDest.second, LTSrc.second)) 2335 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2336 2337 if (ST->hasDQI()) 2338 if (const auto *Entry = ConvertCostTableLookup( 2339 AVX512DQConversionTbl, ISD, LTDest.second, LTSrc.second)) 2340 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2341 2342 if (ST->hasAVX512()) 2343 if (const auto *Entry = ConvertCostTableLookup( 2344 AVX512FConversionTbl, ISD, LTDest.second, LTSrc.second)) 2345 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2346 } 2347 2348 if (ST->hasBWI()) 2349 if (const auto *Entry = ConvertCostTableLookup(AVX512BWVLConversionTbl, ISD, 2350 LTDest.second, LTSrc.second)) 2351 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2352 2353 if (ST->hasDQI()) 2354 if (const auto *Entry = ConvertCostTableLookup(AVX512DQVLConversionTbl, ISD, 2355 LTDest.second, LTSrc.second)) 2356 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2357 2358 if (ST->hasAVX512()) 2359 if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD, 2360 LTDest.second, LTSrc.second)) 2361 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2362 2363 if (ST->hasAVX2()) 2364 if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD, 2365 LTDest.second, LTSrc.second)) 2366 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2367 2368 if (ST->hasAVX()) 2369 if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD, 2370 LTDest.second, LTSrc.second)) 2371 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2372 2373 if (ST->hasSSE41()) 2374 if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD, 2375 LTDest.second, LTSrc.second)) 2376 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2377 2378 if (ST->hasSSE2()) 2379 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD, 2380 LTDest.second, LTSrc.second)) 2381 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2382 2383 // Fallback, for i8/i16 sitofp/uitofp cases we need to extend to i32 for 2384 // sitofp. 2385 if ((ISD == ISD::SINT_TO_FP || ISD == ISD::UINT_TO_FP) && 2386 1 < Src->getScalarSizeInBits() && Src->getScalarSizeInBits() < 32) { 2387 Type *ExtSrc = Src->getWithNewBitWidth(32); 2388 unsigned ExtOpc = 2389 (ISD == ISD::SINT_TO_FP) ? Instruction::SExt : Instruction::ZExt; 2390 2391 // For scalar loads the extend would be free. 2392 InstructionCost ExtCost = 0; 2393 if (!(Src->isIntegerTy() && I && isa<LoadInst>(I->getOperand(0)))) 2394 ExtCost = getCastInstrCost(ExtOpc, ExtSrc, Src, CCH, CostKind); 2395 2396 return ExtCost + getCastInstrCost(Instruction::SIToFP, Dst, ExtSrc, 2397 TTI::CastContextHint::None, CostKind); 2398 } 2399 2400 // Fallback for fptosi/fptoui i8/i16 cases we need to truncate from fptosi 2401 // i32. 2402 if ((ISD == ISD::FP_TO_SINT || ISD == ISD::FP_TO_UINT) && 2403 1 < Dst->getScalarSizeInBits() && Dst->getScalarSizeInBits() < 32) { 2404 Type *TruncDst = Dst->getWithNewBitWidth(32); 2405 return getCastInstrCost(Instruction::FPToSI, TruncDst, Src, CCH, CostKind) + 2406 getCastInstrCost(Instruction::Trunc, Dst, TruncDst, 2407 TTI::CastContextHint::None, CostKind); 2408 } 2409 2410 return AdjustCost( 2411 BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I)); 2412 } 2413 2414 InstructionCost X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 2415 Type *CondTy, 2416 CmpInst::Predicate VecPred, 2417 TTI::TargetCostKind CostKind, 2418 const Instruction *I) { 2419 // TODO: Handle other cost kinds. 2420 if (CostKind != TTI::TCK_RecipThroughput) 2421 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, 2422 I); 2423 2424 // Legalize the type. 2425 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 2426 2427 MVT MTy = LT.second; 2428 2429 int ISD = TLI->InstructionOpcodeToISD(Opcode); 2430 assert(ISD && "Invalid opcode"); 2431 2432 unsigned ExtraCost = 0; 2433 if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) { 2434 // Some vector comparison predicates cost extra instructions. 2435 // TODO: Should we invert this and assume worst case cmp costs 2436 // and reduce for particular predicates? 2437 if (MTy.isVector() && 2438 !((ST->hasXOP() && (!ST->hasAVX2() || MTy.is128BitVector())) || 2439 (ST->hasAVX512() && 32 <= MTy.getScalarSizeInBits()) || 2440 ST->hasBWI())) { 2441 // Fallback to I if a specific predicate wasn't specified. 2442 CmpInst::Predicate Pred = VecPred; 2443 if (I && (Pred == CmpInst::BAD_ICMP_PREDICATE || 2444 Pred == CmpInst::BAD_FCMP_PREDICATE)) 2445 Pred = cast<CmpInst>(I)->getPredicate(); 2446 2447 switch (Pred) { 2448 case CmpInst::Predicate::ICMP_NE: 2449 // xor(cmpeq(x,y),-1) 2450 ExtraCost = 1; 2451 break; 2452 case CmpInst::Predicate::ICMP_SGE: 2453 case CmpInst::Predicate::ICMP_SLE: 2454 // xor(cmpgt(x,y),-1) 2455 ExtraCost = 1; 2456 break; 2457 case CmpInst::Predicate::ICMP_ULT: 2458 case CmpInst::Predicate::ICMP_UGT: 2459 // cmpgt(xor(x,signbit),xor(y,signbit)) 2460 // xor(cmpeq(pmaxu(x,y),x),-1) 2461 ExtraCost = 2; 2462 break; 2463 case CmpInst::Predicate::ICMP_ULE: 2464 case CmpInst::Predicate::ICMP_UGE: 2465 if ((ST->hasSSE41() && MTy.getScalarSizeInBits() == 32) || 2466 (ST->hasSSE2() && MTy.getScalarSizeInBits() < 32)) { 2467 // cmpeq(psubus(x,y),0) 2468 // cmpeq(pminu(x,y),x) 2469 ExtraCost = 1; 2470 } else { 2471 // xor(cmpgt(xor(x,signbit),xor(y,signbit)),-1) 2472 ExtraCost = 3; 2473 } 2474 break; 2475 case CmpInst::Predicate::BAD_ICMP_PREDICATE: 2476 case CmpInst::Predicate::BAD_FCMP_PREDICATE: 2477 // Assume worst case scenario and add the maximum extra cost. 2478 ExtraCost = 3; 2479 break; 2480 default: 2481 break; 2482 } 2483 } 2484 } 2485 2486 static const CostTblEntry SLMCostTbl[] = { 2487 // slm pcmpeq/pcmpgt throughput is 2 2488 { ISD::SETCC, MVT::v2i64, 2 }, 2489 }; 2490 2491 static const CostTblEntry AVX512BWCostTbl[] = { 2492 { ISD::SETCC, MVT::v32i16, 1 }, 2493 { ISD::SETCC, MVT::v64i8, 1 }, 2494 2495 { ISD::SELECT, MVT::v32i16, 1 }, 2496 { ISD::SELECT, MVT::v64i8, 1 }, 2497 }; 2498 2499 static const CostTblEntry AVX512CostTbl[] = { 2500 { ISD::SETCC, MVT::v8i64, 1 }, 2501 { ISD::SETCC, MVT::v16i32, 1 }, 2502 { ISD::SETCC, MVT::v8f64, 1 }, 2503 { ISD::SETCC, MVT::v16f32, 1 }, 2504 2505 { ISD::SELECT, MVT::v8i64, 1 }, 2506 { ISD::SELECT, MVT::v16i32, 1 }, 2507 { ISD::SELECT, MVT::v8f64, 1 }, 2508 { ISD::SELECT, MVT::v16f32, 1 }, 2509 2510 { ISD::SETCC, MVT::v32i16, 2 }, // FIXME: should probably be 4 2511 { ISD::SETCC, MVT::v64i8, 2 }, // FIXME: should probably be 4 2512 2513 { ISD::SELECT, MVT::v32i16, 2 }, // FIXME: should be 3 2514 { ISD::SELECT, MVT::v64i8, 2 }, // FIXME: should be 3 2515 }; 2516 2517 static const CostTblEntry AVX2CostTbl[] = { 2518 { ISD::SETCC, MVT::v4i64, 1 }, 2519 { ISD::SETCC, MVT::v8i32, 1 }, 2520 { ISD::SETCC, MVT::v16i16, 1 }, 2521 { ISD::SETCC, MVT::v32i8, 1 }, 2522 2523 { ISD::SELECT, MVT::v4i64, 1 }, // pblendvb 2524 { ISD::SELECT, MVT::v8i32, 1 }, // pblendvb 2525 { ISD::SELECT, MVT::v16i16, 1 }, // pblendvb 2526 { ISD::SELECT, MVT::v32i8, 1 }, // pblendvb 2527 }; 2528 2529 static const CostTblEntry AVX1CostTbl[] = { 2530 { ISD::SETCC, MVT::v4f64, 1 }, 2531 { ISD::SETCC, MVT::v8f32, 1 }, 2532 // AVX1 does not support 8-wide integer compare. 2533 { ISD::SETCC, MVT::v4i64, 4 }, 2534 { ISD::SETCC, MVT::v8i32, 4 }, 2535 { ISD::SETCC, MVT::v16i16, 4 }, 2536 { ISD::SETCC, MVT::v32i8, 4 }, 2537 2538 { ISD::SELECT, MVT::v4f64, 1 }, // vblendvpd 2539 { ISD::SELECT, MVT::v8f32, 1 }, // vblendvps 2540 { ISD::SELECT, MVT::v4i64, 1 }, // vblendvpd 2541 { ISD::SELECT, MVT::v8i32, 1 }, // vblendvps 2542 { ISD::SELECT, MVT::v16i16, 3 }, // vandps + vandnps + vorps 2543 { ISD::SELECT, MVT::v32i8, 3 }, // vandps + vandnps + vorps 2544 }; 2545 2546 static const CostTblEntry SSE42CostTbl[] = { 2547 { ISD::SETCC, MVT::v2f64, 1 }, 2548 { ISD::SETCC, MVT::v4f32, 1 }, 2549 { ISD::SETCC, MVT::v2i64, 1 }, 2550 }; 2551 2552 static const CostTblEntry SSE41CostTbl[] = { 2553 { ISD::SELECT, MVT::v2f64, 1 }, // blendvpd 2554 { ISD::SELECT, MVT::v4f32, 1 }, // blendvps 2555 { ISD::SELECT, MVT::v2i64, 1 }, // pblendvb 2556 { ISD::SELECT, MVT::v4i32, 1 }, // pblendvb 2557 { ISD::SELECT, MVT::v8i16, 1 }, // pblendvb 2558 { ISD::SELECT, MVT::v16i8, 1 }, // pblendvb 2559 }; 2560 2561 static const CostTblEntry SSE2CostTbl[] = { 2562 { ISD::SETCC, MVT::v2f64, 2 }, 2563 { ISD::SETCC, MVT::f64, 1 }, 2564 { ISD::SETCC, MVT::v2i64, 8 }, 2565 { ISD::SETCC, MVT::v4i32, 1 }, 2566 { ISD::SETCC, MVT::v8i16, 1 }, 2567 { ISD::SETCC, MVT::v16i8, 1 }, 2568 2569 { ISD::SELECT, MVT::v2f64, 3 }, // andpd + andnpd + orpd 2570 { ISD::SELECT, MVT::v2i64, 3 }, // pand + pandn + por 2571 { ISD::SELECT, MVT::v4i32, 3 }, // pand + pandn + por 2572 { ISD::SELECT, MVT::v8i16, 3 }, // pand + pandn + por 2573 { ISD::SELECT, MVT::v16i8, 3 }, // pand + pandn + por 2574 }; 2575 2576 static const CostTblEntry SSE1CostTbl[] = { 2577 { ISD::SETCC, MVT::v4f32, 2 }, 2578 { ISD::SETCC, MVT::f32, 1 }, 2579 2580 { ISD::SELECT, MVT::v4f32, 3 }, // andps + andnps + orps 2581 }; 2582 2583 if (ST->useSLMArithCosts()) 2584 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy)) 2585 return LT.first * (ExtraCost + Entry->Cost); 2586 2587 if (ST->hasBWI()) 2588 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 2589 return LT.first * (ExtraCost + Entry->Cost); 2590 2591 if (ST->hasAVX512()) 2592 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 2593 return LT.first * (ExtraCost + Entry->Cost); 2594 2595 if (ST->hasAVX2()) 2596 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 2597 return LT.first * (ExtraCost + Entry->Cost); 2598 2599 if (ST->hasAVX()) 2600 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 2601 return LT.first * (ExtraCost + Entry->Cost); 2602 2603 if (ST->hasSSE42()) 2604 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 2605 return LT.first * (ExtraCost + Entry->Cost); 2606 2607 if (ST->hasSSE41()) 2608 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 2609 return LT.first * (ExtraCost + Entry->Cost); 2610 2611 if (ST->hasSSE2()) 2612 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 2613 return LT.first * (ExtraCost + Entry->Cost); 2614 2615 if (ST->hasSSE1()) 2616 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 2617 return LT.first * (ExtraCost + Entry->Cost); 2618 2619 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I); 2620 } 2621 2622 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; } 2623 2624 InstructionCost 2625 X86TTIImpl::getTypeBasedIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, 2626 TTI::TargetCostKind CostKind) { 2627 2628 // Costs should match the codegen from: 2629 // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll 2630 // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll 2631 // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll 2632 // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll 2633 // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll 2634 2635 // TODO: Overflow intrinsics (*ADDO, *SUBO, *MULO) with vector types are not 2636 // specialized in these tables yet. 2637 static const CostTblEntry AVX512BITALGCostTbl[] = { 2638 { ISD::CTPOP, MVT::v32i16, 1 }, 2639 { ISD::CTPOP, MVT::v64i8, 1 }, 2640 { ISD::CTPOP, MVT::v16i16, 1 }, 2641 { ISD::CTPOP, MVT::v32i8, 1 }, 2642 { ISD::CTPOP, MVT::v8i16, 1 }, 2643 { ISD::CTPOP, MVT::v16i8, 1 }, 2644 }; 2645 static const CostTblEntry AVX512VPOPCNTDQCostTbl[] = { 2646 { ISD::CTPOP, MVT::v8i64, 1 }, 2647 { ISD::CTPOP, MVT::v16i32, 1 }, 2648 { ISD::CTPOP, MVT::v4i64, 1 }, 2649 { ISD::CTPOP, MVT::v8i32, 1 }, 2650 { ISD::CTPOP, MVT::v2i64, 1 }, 2651 { ISD::CTPOP, MVT::v4i32, 1 }, 2652 }; 2653 static const CostTblEntry AVX512CDCostTbl[] = { 2654 { ISD::CTLZ, MVT::v8i64, 1 }, 2655 { ISD::CTLZ, MVT::v16i32, 1 }, 2656 { ISD::CTLZ, MVT::v32i16, 8 }, 2657 { ISD::CTLZ, MVT::v64i8, 20 }, 2658 { ISD::CTLZ, MVT::v4i64, 1 }, 2659 { ISD::CTLZ, MVT::v8i32, 1 }, 2660 { ISD::CTLZ, MVT::v16i16, 4 }, 2661 { ISD::CTLZ, MVT::v32i8, 10 }, 2662 { ISD::CTLZ, MVT::v2i64, 1 }, 2663 { ISD::CTLZ, MVT::v4i32, 1 }, 2664 { ISD::CTLZ, MVT::v8i16, 4 }, 2665 { ISD::CTLZ, MVT::v16i8, 4 }, 2666 }; 2667 static const CostTblEntry AVX512BWCostTbl[] = { 2668 { ISD::ABS, MVT::v32i16, 1 }, 2669 { ISD::ABS, MVT::v64i8, 1 }, 2670 { ISD::BITREVERSE, MVT::v8i64, 3 }, 2671 { ISD::BITREVERSE, MVT::v16i32, 3 }, 2672 { ISD::BITREVERSE, MVT::v32i16, 3 }, 2673 { ISD::BITREVERSE, MVT::v64i8, 2 }, 2674 { ISD::BSWAP, MVT::v8i64, 1 }, 2675 { ISD::BSWAP, MVT::v16i32, 1 }, 2676 { ISD::BSWAP, MVT::v32i16, 1 }, 2677 { ISD::CTLZ, MVT::v8i64, 23 }, 2678 { ISD::CTLZ, MVT::v16i32, 22 }, 2679 { ISD::CTLZ, MVT::v32i16, 18 }, 2680 { ISD::CTLZ, MVT::v64i8, 17 }, 2681 { ISD::CTPOP, MVT::v8i64, 7 }, 2682 { ISD::CTPOP, MVT::v16i32, 11 }, 2683 { ISD::CTPOP, MVT::v32i16, 9 }, 2684 { ISD::CTPOP, MVT::v64i8, 6 }, 2685 { ISD::CTTZ, MVT::v8i64, 10 }, 2686 { ISD::CTTZ, MVT::v16i32, 14 }, 2687 { ISD::CTTZ, MVT::v32i16, 12 }, 2688 { ISD::CTTZ, MVT::v64i8, 9 }, 2689 { ISD::SADDSAT, MVT::v32i16, 1 }, 2690 { ISD::SADDSAT, MVT::v64i8, 1 }, 2691 { ISD::SMAX, MVT::v32i16, 1 }, 2692 { ISD::SMAX, MVT::v64i8, 1 }, 2693 { ISD::SMIN, MVT::v32i16, 1 }, 2694 { ISD::SMIN, MVT::v64i8, 1 }, 2695 { ISD::SSUBSAT, MVT::v32i16, 1 }, 2696 { ISD::SSUBSAT, MVT::v64i8, 1 }, 2697 { ISD::UADDSAT, MVT::v32i16, 1 }, 2698 { ISD::UADDSAT, MVT::v64i8, 1 }, 2699 { ISD::UMAX, MVT::v32i16, 1 }, 2700 { ISD::UMAX, MVT::v64i8, 1 }, 2701 { ISD::UMIN, MVT::v32i16, 1 }, 2702 { ISD::UMIN, MVT::v64i8, 1 }, 2703 { ISD::USUBSAT, MVT::v32i16, 1 }, 2704 { ISD::USUBSAT, MVT::v64i8, 1 }, 2705 }; 2706 static const CostTblEntry AVX512CostTbl[] = { 2707 { ISD::ABS, MVT::v8i64, 1 }, 2708 { ISD::ABS, MVT::v16i32, 1 }, 2709 { ISD::ABS, MVT::v32i16, 2 }, 2710 { ISD::ABS, MVT::v64i8, 2 }, 2711 { ISD::ABS, MVT::v4i64, 1 }, 2712 { ISD::ABS, MVT::v2i64, 1 }, 2713 { ISD::BITREVERSE, MVT::v8i64, 36 }, 2714 { ISD::BITREVERSE, MVT::v16i32, 24 }, 2715 { ISD::BITREVERSE, MVT::v32i16, 10 }, 2716 { ISD::BITREVERSE, MVT::v64i8, 10 }, 2717 { ISD::BSWAP, MVT::v8i64, 4 }, 2718 { ISD::BSWAP, MVT::v16i32, 4 }, 2719 { ISD::BSWAP, MVT::v32i16, 4 }, 2720 { ISD::CTLZ, MVT::v8i64, 29 }, 2721 { ISD::CTLZ, MVT::v16i32, 35 }, 2722 { ISD::CTLZ, MVT::v32i16, 28 }, 2723 { ISD::CTLZ, MVT::v64i8, 18 }, 2724 { ISD::CTPOP, MVT::v8i64, 16 }, 2725 { ISD::CTPOP, MVT::v16i32, 24 }, 2726 { ISD::CTPOP, MVT::v32i16, 18 }, 2727 { ISD::CTPOP, MVT::v64i8, 12 }, 2728 { ISD::CTTZ, MVT::v8i64, 20 }, 2729 { ISD::CTTZ, MVT::v16i32, 28 }, 2730 { ISD::CTTZ, MVT::v32i16, 24 }, 2731 { ISD::CTTZ, MVT::v64i8, 18 }, 2732 { ISD::SMAX, MVT::v8i64, 1 }, 2733 { ISD::SMAX, MVT::v16i32, 1 }, 2734 { ISD::SMAX, MVT::v32i16, 2 }, 2735 { ISD::SMAX, MVT::v64i8, 2 }, 2736 { ISD::SMAX, MVT::v4i64, 1 }, 2737 { ISD::SMAX, MVT::v2i64, 1 }, 2738 { ISD::SMIN, MVT::v8i64, 1 }, 2739 { ISD::SMIN, MVT::v16i32, 1 }, 2740 { ISD::SMIN, MVT::v32i16, 2 }, 2741 { ISD::SMIN, MVT::v64i8, 2 }, 2742 { ISD::SMIN, MVT::v4i64, 1 }, 2743 { ISD::SMIN, MVT::v2i64, 1 }, 2744 { ISD::UMAX, MVT::v8i64, 1 }, 2745 { ISD::UMAX, MVT::v16i32, 1 }, 2746 { ISD::UMAX, MVT::v32i16, 2 }, 2747 { ISD::UMAX, MVT::v64i8, 2 }, 2748 { ISD::UMAX, MVT::v4i64, 1 }, 2749 { ISD::UMAX, MVT::v2i64, 1 }, 2750 { ISD::UMIN, MVT::v8i64, 1 }, 2751 { ISD::UMIN, MVT::v16i32, 1 }, 2752 { ISD::UMIN, MVT::v32i16, 2 }, 2753 { ISD::UMIN, MVT::v64i8, 2 }, 2754 { ISD::UMIN, MVT::v4i64, 1 }, 2755 { ISD::UMIN, MVT::v2i64, 1 }, 2756 { ISD::USUBSAT, MVT::v16i32, 2 }, // pmaxud + psubd 2757 { ISD::USUBSAT, MVT::v2i64, 2 }, // pmaxuq + psubq 2758 { ISD::USUBSAT, MVT::v4i64, 2 }, // pmaxuq + psubq 2759 { ISD::USUBSAT, MVT::v8i64, 2 }, // pmaxuq + psubq 2760 { ISD::UADDSAT, MVT::v16i32, 3 }, // not + pminud + paddd 2761 { ISD::UADDSAT, MVT::v2i64, 3 }, // not + pminuq + paddq 2762 { ISD::UADDSAT, MVT::v4i64, 3 }, // not + pminuq + paddq 2763 { ISD::UADDSAT, MVT::v8i64, 3 }, // not + pminuq + paddq 2764 { ISD::SADDSAT, MVT::v32i16, 2 }, 2765 { ISD::SADDSAT, MVT::v64i8, 2 }, 2766 { ISD::SSUBSAT, MVT::v32i16, 2 }, 2767 { ISD::SSUBSAT, MVT::v64i8, 2 }, 2768 { ISD::UADDSAT, MVT::v32i16, 2 }, 2769 { ISD::UADDSAT, MVT::v64i8, 2 }, 2770 { ISD::USUBSAT, MVT::v32i16, 2 }, 2771 { ISD::USUBSAT, MVT::v64i8, 2 }, 2772 { ISD::FMAXNUM, MVT::f32, 2 }, 2773 { ISD::FMAXNUM, MVT::v4f32, 2 }, 2774 { ISD::FMAXNUM, MVT::v8f32, 2 }, 2775 { ISD::FMAXNUM, MVT::v16f32, 2 }, 2776 { ISD::FMAXNUM, MVT::f64, 2 }, 2777 { ISD::FMAXNUM, MVT::v2f64, 2 }, 2778 { ISD::FMAXNUM, MVT::v4f64, 2 }, 2779 { ISD::FMAXNUM, MVT::v8f64, 2 }, 2780 }; 2781 static const CostTblEntry XOPCostTbl[] = { 2782 { ISD::BITREVERSE, MVT::v4i64, 4 }, 2783 { ISD::BITREVERSE, MVT::v8i32, 4 }, 2784 { ISD::BITREVERSE, MVT::v16i16, 4 }, 2785 { ISD::BITREVERSE, MVT::v32i8, 4 }, 2786 { ISD::BITREVERSE, MVT::v2i64, 1 }, 2787 { ISD::BITREVERSE, MVT::v4i32, 1 }, 2788 { ISD::BITREVERSE, MVT::v8i16, 1 }, 2789 { ISD::BITREVERSE, MVT::v16i8, 1 }, 2790 { ISD::BITREVERSE, MVT::i64, 3 }, 2791 { ISD::BITREVERSE, MVT::i32, 3 }, 2792 { ISD::BITREVERSE, MVT::i16, 3 }, 2793 { ISD::BITREVERSE, MVT::i8, 3 } 2794 }; 2795 static const CostTblEntry AVX2CostTbl[] = { 2796 { ISD::ABS, MVT::v4i64, 2 }, // VBLENDVPD(X,VPSUBQ(0,X),X) 2797 { ISD::ABS, MVT::v8i32, 1 }, 2798 { ISD::ABS, MVT::v16i16, 1 }, 2799 { ISD::ABS, MVT::v32i8, 1 }, 2800 { ISD::BITREVERSE, MVT::v2i64, 3 }, 2801 { ISD::BITREVERSE, MVT::v4i64, 3 }, 2802 { ISD::BITREVERSE, MVT::v4i32, 3 }, 2803 { ISD::BITREVERSE, MVT::v8i32, 3 }, 2804 { ISD::BITREVERSE, MVT::v8i16, 3 }, 2805 { ISD::BITREVERSE, MVT::v16i16, 3 }, 2806 { ISD::BITREVERSE, MVT::v16i8, 3 }, 2807 { ISD::BITREVERSE, MVT::v32i8, 3 }, 2808 { ISD::BSWAP, MVT::v4i64, 1 }, 2809 { ISD::BSWAP, MVT::v8i32, 1 }, 2810 { ISD::BSWAP, MVT::v16i16, 1 }, 2811 { ISD::CTLZ, MVT::v2i64, 7 }, 2812 { ISD::CTLZ, MVT::v4i64, 7 }, 2813 { ISD::CTLZ, MVT::v4i32, 5 }, 2814 { ISD::CTLZ, MVT::v8i32, 5 }, 2815 { ISD::CTLZ, MVT::v8i16, 4 }, 2816 { ISD::CTLZ, MVT::v16i16, 4 }, 2817 { ISD::CTLZ, MVT::v16i8, 3 }, 2818 { ISD::CTLZ, MVT::v32i8, 3 }, 2819 { ISD::CTPOP, MVT::v2i64, 3 }, 2820 { ISD::CTPOP, MVT::v4i64, 3 }, 2821 { ISD::CTPOP, MVT::v4i32, 7 }, 2822 { ISD::CTPOP, MVT::v8i32, 7 }, 2823 { ISD::CTPOP, MVT::v8i16, 3 }, 2824 { ISD::CTPOP, MVT::v16i16, 3 }, 2825 { ISD::CTPOP, MVT::v16i8, 2 }, 2826 { ISD::CTPOP, MVT::v32i8, 2 }, 2827 { ISD::CTTZ, MVT::v2i64, 4 }, 2828 { ISD::CTTZ, MVT::v4i64, 4 }, 2829 { ISD::CTTZ, MVT::v4i32, 7 }, 2830 { ISD::CTTZ, MVT::v8i32, 7 }, 2831 { ISD::CTTZ, MVT::v8i16, 4 }, 2832 { ISD::CTTZ, MVT::v16i16, 4 }, 2833 { ISD::CTTZ, MVT::v16i8, 3 }, 2834 { ISD::CTTZ, MVT::v32i8, 3 }, 2835 { ISD::SADDSAT, MVT::v16i16, 1 }, 2836 { ISD::SADDSAT, MVT::v32i8, 1 }, 2837 { ISD::SMAX, MVT::v8i32, 1 }, 2838 { ISD::SMAX, MVT::v16i16, 1 }, 2839 { ISD::SMAX, MVT::v32i8, 1 }, 2840 { ISD::SMIN, MVT::v8i32, 1 }, 2841 { ISD::SMIN, MVT::v16i16, 1 }, 2842 { ISD::SMIN, MVT::v32i8, 1 }, 2843 { ISD::SSUBSAT, MVT::v16i16, 1 }, 2844 { ISD::SSUBSAT, MVT::v32i8, 1 }, 2845 { ISD::UADDSAT, MVT::v16i16, 1 }, 2846 { ISD::UADDSAT, MVT::v32i8, 1 }, 2847 { ISD::UADDSAT, MVT::v8i32, 3 }, // not + pminud + paddd 2848 { ISD::UMAX, MVT::v8i32, 1 }, 2849 { ISD::UMAX, MVT::v16i16, 1 }, 2850 { ISD::UMAX, MVT::v32i8, 1 }, 2851 { ISD::UMIN, MVT::v8i32, 1 }, 2852 { ISD::UMIN, MVT::v16i16, 1 }, 2853 { ISD::UMIN, MVT::v32i8, 1 }, 2854 { ISD::USUBSAT, MVT::v16i16, 1 }, 2855 { ISD::USUBSAT, MVT::v32i8, 1 }, 2856 { ISD::USUBSAT, MVT::v8i32, 2 }, // pmaxud + psubd 2857 { ISD::FMAXNUM, MVT::v8f32, 3 }, // MAXPS + CMPUNORDPS + BLENDVPS 2858 { ISD::FMAXNUM, MVT::v4f64, 3 }, // MAXPD + CMPUNORDPD + BLENDVPD 2859 { ISD::FSQRT, MVT::f32, 7 }, // Haswell from http://www.agner.org/ 2860 { ISD::FSQRT, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/ 2861 { ISD::FSQRT, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/ 2862 { ISD::FSQRT, MVT::f64, 14 }, // Haswell from http://www.agner.org/ 2863 { ISD::FSQRT, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/ 2864 { ISD::FSQRT, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/ 2865 }; 2866 static const CostTblEntry AVX1CostTbl[] = { 2867 { ISD::ABS, MVT::v4i64, 5 }, // VBLENDVPD(X,VPSUBQ(0,X),X) 2868 { ISD::ABS, MVT::v8i32, 3 }, 2869 { ISD::ABS, MVT::v16i16, 3 }, 2870 { ISD::ABS, MVT::v32i8, 3 }, 2871 { ISD::BITREVERSE, MVT::v4i64, 12 }, // 2 x 128-bit Op + extract/insert 2872 { ISD::BITREVERSE, MVT::v8i32, 12 }, // 2 x 128-bit Op + extract/insert 2873 { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert 2874 { ISD::BITREVERSE, MVT::v32i8, 12 }, // 2 x 128-bit Op + extract/insert 2875 { ISD::BSWAP, MVT::v4i64, 4 }, 2876 { ISD::BSWAP, MVT::v8i32, 4 }, 2877 { ISD::BSWAP, MVT::v16i16, 4 }, 2878 { ISD::CTLZ, MVT::v4i64, 48 }, // 2 x 128-bit Op + extract/insert 2879 { ISD::CTLZ, MVT::v8i32, 38 }, // 2 x 128-bit Op + extract/insert 2880 { ISD::CTLZ, MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert 2881 { ISD::CTLZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert 2882 { ISD::CTPOP, MVT::v4i64, 16 }, // 2 x 128-bit Op + extract/insert 2883 { ISD::CTPOP, MVT::v8i32, 24 }, // 2 x 128-bit Op + extract/insert 2884 { ISD::CTPOP, MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert 2885 { ISD::CTPOP, MVT::v32i8, 14 }, // 2 x 128-bit Op + extract/insert 2886 { ISD::CTTZ, MVT::v4i64, 22 }, // 2 x 128-bit Op + extract/insert 2887 { ISD::CTTZ, MVT::v8i32, 30 }, // 2 x 128-bit Op + extract/insert 2888 { ISD::CTTZ, MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert 2889 { ISD::CTTZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert 2890 { ISD::SADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2891 { ISD::SADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2892 { ISD::SMAX, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2893 { ISD::SMAX, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2894 { ISD::SMAX, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2895 { ISD::SMIN, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2896 { ISD::SMIN, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2897 { ISD::SMIN, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2898 { ISD::SSUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2899 { ISD::SSUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2900 { ISD::UADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2901 { ISD::UADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2902 { ISD::UADDSAT, MVT::v8i32, 8 }, // 2 x 128-bit Op + extract/insert 2903 { ISD::UMAX, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2904 { ISD::UMAX, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2905 { ISD::UMAX, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2906 { ISD::UMIN, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2907 { ISD::UMIN, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2908 { ISD::UMIN, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2909 { ISD::USUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2910 { ISD::USUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2911 { ISD::USUBSAT, MVT::v8i32, 6 }, // 2 x 128-bit Op + extract/insert 2912 { ISD::FMAXNUM, MVT::f32, 3 }, // MAXSS + CMPUNORDSS + BLENDVPS 2913 { ISD::FMAXNUM, MVT::v4f32, 3 }, // MAXPS + CMPUNORDPS + BLENDVPS 2914 { ISD::FMAXNUM, MVT::v8f32, 5 }, // MAXPS + CMPUNORDPS + BLENDVPS + ? 2915 { ISD::FMAXNUM, MVT::f64, 3 }, // MAXSD + CMPUNORDSD + BLENDVPD 2916 { ISD::FMAXNUM, MVT::v2f64, 3 }, // MAXPD + CMPUNORDPD + BLENDVPD 2917 { ISD::FMAXNUM, MVT::v4f64, 5 }, // MAXPD + CMPUNORDPD + BLENDVPD + ? 2918 { ISD::FSQRT, MVT::f32, 14 }, // SNB from http://www.agner.org/ 2919 { ISD::FSQRT, MVT::v4f32, 14 }, // SNB from http://www.agner.org/ 2920 { ISD::FSQRT, MVT::v8f32, 28 }, // SNB from http://www.agner.org/ 2921 { ISD::FSQRT, MVT::f64, 21 }, // SNB from http://www.agner.org/ 2922 { ISD::FSQRT, MVT::v2f64, 21 }, // SNB from http://www.agner.org/ 2923 { ISD::FSQRT, MVT::v4f64, 43 }, // SNB from http://www.agner.org/ 2924 }; 2925 static const CostTblEntry GLMCostTbl[] = { 2926 { ISD::FSQRT, MVT::f32, 19 }, // sqrtss 2927 { ISD::FSQRT, MVT::v4f32, 37 }, // sqrtps 2928 { ISD::FSQRT, MVT::f64, 34 }, // sqrtsd 2929 { ISD::FSQRT, MVT::v2f64, 67 }, // sqrtpd 2930 }; 2931 static const CostTblEntry SLMCostTbl[] = { 2932 { ISD::FSQRT, MVT::f32, 20 }, // sqrtss 2933 { ISD::FSQRT, MVT::v4f32, 40 }, // sqrtps 2934 { ISD::FSQRT, MVT::f64, 35 }, // sqrtsd 2935 { ISD::FSQRT, MVT::v2f64, 70 }, // sqrtpd 2936 }; 2937 static const CostTblEntry SSE42CostTbl[] = { 2938 { ISD::USUBSAT, MVT::v4i32, 2 }, // pmaxud + psubd 2939 { ISD::UADDSAT, MVT::v4i32, 3 }, // not + pminud + paddd 2940 { ISD::FSQRT, MVT::f32, 18 }, // Nehalem from http://www.agner.org/ 2941 { ISD::FSQRT, MVT::v4f32, 18 }, // Nehalem from http://www.agner.org/ 2942 }; 2943 static const CostTblEntry SSE41CostTbl[] = { 2944 { ISD::ABS, MVT::v2i64, 2 }, // BLENDVPD(X,PSUBQ(0,X),X) 2945 { ISD::SMAX, MVT::v4i32, 1 }, 2946 { ISD::SMAX, MVT::v16i8, 1 }, 2947 { ISD::SMIN, MVT::v4i32, 1 }, 2948 { ISD::SMIN, MVT::v16i8, 1 }, 2949 { ISD::UMAX, MVT::v4i32, 1 }, 2950 { ISD::UMAX, MVT::v8i16, 1 }, 2951 { ISD::UMIN, MVT::v4i32, 1 }, 2952 { ISD::UMIN, MVT::v8i16, 1 }, 2953 }; 2954 static const CostTblEntry SSSE3CostTbl[] = { 2955 { ISD::ABS, MVT::v4i32, 1 }, 2956 { ISD::ABS, MVT::v8i16, 1 }, 2957 { ISD::ABS, MVT::v16i8, 1 }, 2958 { ISD::BITREVERSE, MVT::v2i64, 5 }, 2959 { ISD::BITREVERSE, MVT::v4i32, 5 }, 2960 { ISD::BITREVERSE, MVT::v8i16, 5 }, 2961 { ISD::BITREVERSE, MVT::v16i8, 5 }, 2962 { ISD::BSWAP, MVT::v2i64, 1 }, 2963 { ISD::BSWAP, MVT::v4i32, 1 }, 2964 { ISD::BSWAP, MVT::v8i16, 1 }, 2965 { ISD::CTLZ, MVT::v2i64, 23 }, 2966 { ISD::CTLZ, MVT::v4i32, 18 }, 2967 { ISD::CTLZ, MVT::v8i16, 14 }, 2968 { ISD::CTLZ, MVT::v16i8, 9 }, 2969 { ISD::CTPOP, MVT::v2i64, 7 }, 2970 { ISD::CTPOP, MVT::v4i32, 11 }, 2971 { ISD::CTPOP, MVT::v8i16, 9 }, 2972 { ISD::CTPOP, MVT::v16i8, 6 }, 2973 { ISD::CTTZ, MVT::v2i64, 10 }, 2974 { ISD::CTTZ, MVT::v4i32, 14 }, 2975 { ISD::CTTZ, MVT::v8i16, 12 }, 2976 { ISD::CTTZ, MVT::v16i8, 9 } 2977 }; 2978 static const CostTblEntry SSE2CostTbl[] = { 2979 { ISD::ABS, MVT::v2i64, 4 }, 2980 { ISD::ABS, MVT::v4i32, 3 }, 2981 { ISD::ABS, MVT::v8i16, 2 }, 2982 { ISD::ABS, MVT::v16i8, 2 }, 2983 { ISD::BITREVERSE, MVT::v2i64, 29 }, 2984 { ISD::BITREVERSE, MVT::v4i32, 27 }, 2985 { ISD::BITREVERSE, MVT::v8i16, 27 }, 2986 { ISD::BITREVERSE, MVT::v16i8, 20 }, 2987 { ISD::BSWAP, MVT::v2i64, 7 }, 2988 { ISD::BSWAP, MVT::v4i32, 7 }, 2989 { ISD::BSWAP, MVT::v8i16, 7 }, 2990 { ISD::CTLZ, MVT::v2i64, 25 }, 2991 { ISD::CTLZ, MVT::v4i32, 26 }, 2992 { ISD::CTLZ, MVT::v8i16, 20 }, 2993 { ISD::CTLZ, MVT::v16i8, 17 }, 2994 { ISD::CTPOP, MVT::v2i64, 12 }, 2995 { ISD::CTPOP, MVT::v4i32, 15 }, 2996 { ISD::CTPOP, MVT::v8i16, 13 }, 2997 { ISD::CTPOP, MVT::v16i8, 10 }, 2998 { ISD::CTTZ, MVT::v2i64, 14 }, 2999 { ISD::CTTZ, MVT::v4i32, 18 }, 3000 { ISD::CTTZ, MVT::v8i16, 16 }, 3001 { ISD::CTTZ, MVT::v16i8, 13 }, 3002 { ISD::SADDSAT, MVT::v8i16, 1 }, 3003 { ISD::SADDSAT, MVT::v16i8, 1 }, 3004 { ISD::SMAX, MVT::v8i16, 1 }, 3005 { ISD::SMIN, MVT::v8i16, 1 }, 3006 { ISD::SSUBSAT, MVT::v8i16, 1 }, 3007 { ISD::SSUBSAT, MVT::v16i8, 1 }, 3008 { ISD::UADDSAT, MVT::v8i16, 1 }, 3009 { ISD::UADDSAT, MVT::v16i8, 1 }, 3010 { ISD::UMAX, MVT::v8i16, 2 }, 3011 { ISD::UMAX, MVT::v16i8, 1 }, 3012 { ISD::UMIN, MVT::v8i16, 2 }, 3013 { ISD::UMIN, MVT::v16i8, 1 }, 3014 { ISD::USUBSAT, MVT::v8i16, 1 }, 3015 { ISD::USUBSAT, MVT::v16i8, 1 }, 3016 { ISD::FMAXNUM, MVT::f64, 4 }, 3017 { ISD::FMAXNUM, MVT::v2f64, 4 }, 3018 { ISD::FSQRT, MVT::f64, 32 }, // Nehalem from http://www.agner.org/ 3019 { ISD::FSQRT, MVT::v2f64, 32 }, // Nehalem from http://www.agner.org/ 3020 }; 3021 static const CostTblEntry SSE1CostTbl[] = { 3022 { ISD::FMAXNUM, MVT::f32, 4 }, 3023 { ISD::FMAXNUM, MVT::v4f32, 4 }, 3024 { ISD::FSQRT, MVT::f32, 28 }, // Pentium III from http://www.agner.org/ 3025 { ISD::FSQRT, MVT::v4f32, 56 }, // Pentium III from http://www.agner.org/ 3026 }; 3027 static const CostTblEntry BMI64CostTbl[] = { // 64-bit targets 3028 { ISD::CTTZ, MVT::i64, 1 }, 3029 }; 3030 static const CostTblEntry BMI32CostTbl[] = { // 32 or 64-bit targets 3031 { ISD::CTTZ, MVT::i32, 1 }, 3032 { ISD::CTTZ, MVT::i16, 1 }, 3033 { ISD::CTTZ, MVT::i8, 1 }, 3034 }; 3035 static const CostTblEntry LZCNT64CostTbl[] = { // 64-bit targets 3036 { ISD::CTLZ, MVT::i64, 1 }, 3037 }; 3038 static const CostTblEntry LZCNT32CostTbl[] = { // 32 or 64-bit targets 3039 { ISD::CTLZ, MVT::i32, 1 }, 3040 { ISD::CTLZ, MVT::i16, 1 }, 3041 { ISD::CTLZ, MVT::i8, 1 }, 3042 }; 3043 static const CostTblEntry POPCNT64CostTbl[] = { // 64-bit targets 3044 { ISD::CTPOP, MVT::i64, 1 }, 3045 }; 3046 static const CostTblEntry POPCNT32CostTbl[] = { // 32 or 64-bit targets 3047 { ISD::CTPOP, MVT::i32, 1 }, 3048 { ISD::CTPOP, MVT::i16, 1 }, 3049 { ISD::CTPOP, MVT::i8, 1 }, 3050 }; 3051 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 3052 { ISD::ABS, MVT::i64, 2 }, // SUB+CMOV 3053 { ISD::BITREVERSE, MVT::i64, 14 }, 3054 { ISD::BSWAP, MVT::i64, 1 }, 3055 { ISD::CTLZ, MVT::i64, 4 }, // BSR+XOR or BSR+XOR+CMOV 3056 { ISD::CTTZ, MVT::i64, 3 }, // TEST+BSF+CMOV/BRANCH 3057 { ISD::CTPOP, MVT::i64, 10 }, 3058 { ISD::SADDO, MVT::i64, 1 }, 3059 { ISD::UADDO, MVT::i64, 1 }, 3060 { ISD::UMULO, MVT::i64, 2 }, // mulq + seto 3061 }; 3062 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 3063 { ISD::ABS, MVT::i32, 2 }, // SUB+CMOV 3064 { ISD::ABS, MVT::i16, 2 }, // SUB+CMOV 3065 { ISD::BITREVERSE, MVT::i32, 14 }, 3066 { ISD::BITREVERSE, MVT::i16, 14 }, 3067 { ISD::BITREVERSE, MVT::i8, 11 }, 3068 { ISD::BSWAP, MVT::i32, 1 }, 3069 { ISD::BSWAP, MVT::i16, 1 }, // ROL 3070 { ISD::CTLZ, MVT::i32, 4 }, // BSR+XOR or BSR+XOR+CMOV 3071 { ISD::CTLZ, MVT::i16, 4 }, // BSR+XOR or BSR+XOR+CMOV 3072 { ISD::CTLZ, MVT::i8, 4 }, // BSR+XOR or BSR+XOR+CMOV 3073 { ISD::CTTZ, MVT::i32, 3 }, // TEST+BSF+CMOV/BRANCH 3074 { ISD::CTTZ, MVT::i16, 3 }, // TEST+BSF+CMOV/BRANCH 3075 { ISD::CTTZ, MVT::i8, 3 }, // TEST+BSF+CMOV/BRANCH 3076 { ISD::CTPOP, MVT::i32, 8 }, 3077 { ISD::CTPOP, MVT::i16, 9 }, 3078 { ISD::CTPOP, MVT::i8, 7 }, 3079 { ISD::SADDO, MVT::i32, 1 }, 3080 { ISD::SADDO, MVT::i16, 1 }, 3081 { ISD::SADDO, MVT::i8, 1 }, 3082 { ISD::UADDO, MVT::i32, 1 }, 3083 { ISD::UADDO, MVT::i16, 1 }, 3084 { ISD::UADDO, MVT::i8, 1 }, 3085 { ISD::UMULO, MVT::i32, 2 }, // mul + seto 3086 { ISD::UMULO, MVT::i16, 2 }, 3087 { ISD::UMULO, MVT::i8, 2 }, 3088 }; 3089 3090 Type *RetTy = ICA.getReturnType(); 3091 Type *OpTy = RetTy; 3092 Intrinsic::ID IID = ICA.getID(); 3093 unsigned ISD = ISD::DELETED_NODE; 3094 switch (IID) { 3095 default: 3096 break; 3097 case Intrinsic::abs: 3098 ISD = ISD::ABS; 3099 break; 3100 case Intrinsic::bitreverse: 3101 ISD = ISD::BITREVERSE; 3102 break; 3103 case Intrinsic::bswap: 3104 ISD = ISD::BSWAP; 3105 break; 3106 case Intrinsic::ctlz: 3107 ISD = ISD::CTLZ; 3108 break; 3109 case Intrinsic::ctpop: 3110 ISD = ISD::CTPOP; 3111 break; 3112 case Intrinsic::cttz: 3113 ISD = ISD::CTTZ; 3114 break; 3115 case Intrinsic::maxnum: 3116 case Intrinsic::minnum: 3117 // FMINNUM has same costs so don't duplicate. 3118 ISD = ISD::FMAXNUM; 3119 break; 3120 case Intrinsic::sadd_sat: 3121 ISD = ISD::SADDSAT; 3122 break; 3123 case Intrinsic::smax: 3124 ISD = ISD::SMAX; 3125 break; 3126 case Intrinsic::smin: 3127 ISD = ISD::SMIN; 3128 break; 3129 case Intrinsic::ssub_sat: 3130 ISD = ISD::SSUBSAT; 3131 break; 3132 case Intrinsic::uadd_sat: 3133 ISD = ISD::UADDSAT; 3134 break; 3135 case Intrinsic::umax: 3136 ISD = ISD::UMAX; 3137 break; 3138 case Intrinsic::umin: 3139 ISD = ISD::UMIN; 3140 break; 3141 case Intrinsic::usub_sat: 3142 ISD = ISD::USUBSAT; 3143 break; 3144 case Intrinsic::sqrt: 3145 ISD = ISD::FSQRT; 3146 break; 3147 case Intrinsic::sadd_with_overflow: 3148 case Intrinsic::ssub_with_overflow: 3149 // SSUBO has same costs so don't duplicate. 3150 ISD = ISD::SADDO; 3151 OpTy = RetTy->getContainedType(0); 3152 break; 3153 case Intrinsic::uadd_with_overflow: 3154 case Intrinsic::usub_with_overflow: 3155 // USUBO has same costs so don't duplicate. 3156 ISD = ISD::UADDO; 3157 OpTy = RetTy->getContainedType(0); 3158 break; 3159 case Intrinsic::umul_with_overflow: 3160 case Intrinsic::smul_with_overflow: 3161 // SMULO has same costs so don't duplicate. 3162 ISD = ISD::UMULO; 3163 OpTy = RetTy->getContainedType(0); 3164 break; 3165 } 3166 3167 if (ISD != ISD::DELETED_NODE) { 3168 // Legalize the type. 3169 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, OpTy); 3170 MVT MTy = LT.second; 3171 3172 // Attempt to lookup cost. 3173 if (ISD == ISD::BITREVERSE && ST->hasGFNI() && ST->hasSSSE3() && 3174 MTy.isVector()) { 3175 // With PSHUFB the code is very similar for all types. If we have integer 3176 // byte operations, we just need a GF2P8AFFINEQB for vXi8. For other types 3177 // we also need a PSHUFB. 3178 unsigned Cost = MTy.getVectorElementType() == MVT::i8 ? 1 : 2; 3179 3180 // Without byte operations, we need twice as many GF2P8AFFINEQB and PSHUFB 3181 // instructions. We also need an extract and an insert. 3182 if (!(MTy.is128BitVector() || (ST->hasAVX2() && MTy.is256BitVector()) || 3183 (ST->hasBWI() && MTy.is512BitVector()))) 3184 Cost = Cost * 2 + 2; 3185 3186 return LT.first * Cost; 3187 } 3188 3189 auto adjustTableCost = [](const CostTblEntry &Entry, 3190 InstructionCost LegalizationCost, 3191 FastMathFlags FMF) { 3192 // If there are no NANs to deal with, then these are reduced to a 3193 // single MIN** or MAX** instruction instead of the MIN/CMP/SELECT that we 3194 // assume is used in the non-fast case. 3195 if (Entry.ISD == ISD::FMAXNUM || Entry.ISD == ISD::FMINNUM) { 3196 if (FMF.noNaNs()) 3197 return LegalizationCost * 1; 3198 } 3199 return LegalizationCost * (int)Entry.Cost; 3200 }; 3201 3202 if (ST->useGLMDivSqrtCosts()) 3203 if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy)) 3204 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3205 3206 if (ST->useSLMArithCosts()) 3207 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy)) 3208 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3209 3210 if (ST->hasBITALG()) 3211 if (const auto *Entry = CostTableLookup(AVX512BITALGCostTbl, ISD, MTy)) 3212 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3213 3214 if (ST->hasVPOPCNTDQ()) 3215 if (const auto *Entry = CostTableLookup(AVX512VPOPCNTDQCostTbl, ISD, MTy)) 3216 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3217 3218 if (ST->hasCDI()) 3219 if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy)) 3220 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3221 3222 if (ST->hasBWI()) 3223 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 3224 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3225 3226 if (ST->hasAVX512()) 3227 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 3228 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3229 3230 if (ST->hasXOP()) 3231 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy)) 3232 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3233 3234 if (ST->hasAVX2()) 3235 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 3236 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3237 3238 if (ST->hasAVX()) 3239 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 3240 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3241 3242 if (ST->hasSSE42()) 3243 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 3244 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3245 3246 if (ST->hasSSE41()) 3247 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 3248 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3249 3250 if (ST->hasSSSE3()) 3251 if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy)) 3252 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3253 3254 if (ST->hasSSE2()) 3255 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 3256 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3257 3258 if (ST->hasSSE1()) 3259 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 3260 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3261 3262 if (ST->hasBMI()) { 3263 if (ST->is64Bit()) 3264 if (const auto *Entry = CostTableLookup(BMI64CostTbl, ISD, MTy)) 3265 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3266 3267 if (const auto *Entry = CostTableLookup(BMI32CostTbl, ISD, MTy)) 3268 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3269 } 3270 3271 if (ST->hasLZCNT()) { 3272 if (ST->is64Bit()) 3273 if (const auto *Entry = CostTableLookup(LZCNT64CostTbl, ISD, MTy)) 3274 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3275 3276 if (const auto *Entry = CostTableLookup(LZCNT32CostTbl, ISD, MTy)) 3277 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3278 } 3279 3280 if (ST->hasPOPCNT()) { 3281 if (ST->is64Bit()) 3282 if (const auto *Entry = CostTableLookup(POPCNT64CostTbl, ISD, MTy)) 3283 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3284 3285 if (const auto *Entry = CostTableLookup(POPCNT32CostTbl, ISD, MTy)) 3286 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3287 } 3288 3289 if (ISD == ISD::BSWAP && ST->hasMOVBE() && ST->hasFastMOVBE()) { 3290 if (const Instruction *II = ICA.getInst()) { 3291 if (II->hasOneUse() && isa<StoreInst>(II->user_back())) 3292 return TTI::TCC_Free; 3293 if (auto *LI = dyn_cast<LoadInst>(II->getOperand(0))) { 3294 if (LI->hasOneUse()) 3295 return TTI::TCC_Free; 3296 } 3297 } 3298 } 3299 3300 if (ST->is64Bit()) 3301 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy)) 3302 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3303 3304 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy)) 3305 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3306 } 3307 3308 return BaseT::getIntrinsicInstrCost(ICA, CostKind); 3309 } 3310 3311 InstructionCost 3312 X86TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, 3313 TTI::TargetCostKind CostKind) { 3314 if (ICA.isTypeBasedOnly()) 3315 return getTypeBasedIntrinsicInstrCost(ICA, CostKind); 3316 3317 static const CostTblEntry AVX512CostTbl[] = { 3318 { ISD::ROTL, MVT::v8i64, 1 }, 3319 { ISD::ROTL, MVT::v4i64, 1 }, 3320 { ISD::ROTL, MVT::v2i64, 1 }, 3321 { ISD::ROTL, MVT::v16i32, 1 }, 3322 { ISD::ROTL, MVT::v8i32, 1 }, 3323 { ISD::ROTL, MVT::v4i32, 1 }, 3324 { ISD::ROTR, MVT::v8i64, 1 }, 3325 { ISD::ROTR, MVT::v4i64, 1 }, 3326 { ISD::ROTR, MVT::v2i64, 1 }, 3327 { ISD::ROTR, MVT::v16i32, 1 }, 3328 { ISD::ROTR, MVT::v8i32, 1 }, 3329 { ISD::ROTR, MVT::v4i32, 1 } 3330 }; 3331 // XOP: ROTL = VPROT(X,Y), ROTR = VPROT(X,SUB(0,Y)) 3332 static const CostTblEntry XOPCostTbl[] = { 3333 { ISD::ROTL, MVT::v4i64, 4 }, 3334 { ISD::ROTL, MVT::v8i32, 4 }, 3335 { ISD::ROTL, MVT::v16i16, 4 }, 3336 { ISD::ROTL, MVT::v32i8, 4 }, 3337 { ISD::ROTL, MVT::v2i64, 1 }, 3338 { ISD::ROTL, MVT::v4i32, 1 }, 3339 { ISD::ROTL, MVT::v8i16, 1 }, 3340 { ISD::ROTL, MVT::v16i8, 1 }, 3341 { ISD::ROTR, MVT::v4i64, 6 }, 3342 { ISD::ROTR, MVT::v8i32, 6 }, 3343 { ISD::ROTR, MVT::v16i16, 6 }, 3344 { ISD::ROTR, MVT::v32i8, 6 }, 3345 { ISD::ROTR, MVT::v2i64, 2 }, 3346 { ISD::ROTR, MVT::v4i32, 2 }, 3347 { ISD::ROTR, MVT::v8i16, 2 }, 3348 { ISD::ROTR, MVT::v16i8, 2 } 3349 }; 3350 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 3351 { ISD::ROTL, MVT::i64, 1 }, 3352 { ISD::ROTR, MVT::i64, 1 }, 3353 { ISD::FSHL, MVT::i64, 4 } 3354 }; 3355 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 3356 { ISD::ROTL, MVT::i32, 1 }, 3357 { ISD::ROTL, MVT::i16, 1 }, 3358 { ISD::ROTL, MVT::i8, 1 }, 3359 { ISD::ROTR, MVT::i32, 1 }, 3360 { ISD::ROTR, MVT::i16, 1 }, 3361 { ISD::ROTR, MVT::i8, 1 }, 3362 { ISD::FSHL, MVT::i32, 4 }, 3363 { ISD::FSHL, MVT::i16, 4 }, 3364 { ISD::FSHL, MVT::i8, 4 } 3365 }; 3366 3367 Intrinsic::ID IID = ICA.getID(); 3368 Type *RetTy = ICA.getReturnType(); 3369 const SmallVectorImpl<const Value *> &Args = ICA.getArgs(); 3370 unsigned ISD = ISD::DELETED_NODE; 3371 switch (IID) { 3372 default: 3373 break; 3374 case Intrinsic::fshl: 3375 ISD = ISD::FSHL; 3376 if (Args[0] == Args[1]) 3377 ISD = ISD::ROTL; 3378 break; 3379 case Intrinsic::fshr: 3380 // FSHR has same costs so don't duplicate. 3381 ISD = ISD::FSHL; 3382 if (Args[0] == Args[1]) 3383 ISD = ISD::ROTR; 3384 break; 3385 } 3386 3387 if (ISD != ISD::DELETED_NODE) { 3388 // Legalize the type. 3389 std::pair<InstructionCost, MVT> LT = 3390 TLI->getTypeLegalizationCost(DL, RetTy); 3391 MVT MTy = LT.second; 3392 3393 // Attempt to lookup cost. 3394 if (ST->hasAVX512()) 3395 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 3396 return LT.first * Entry->Cost; 3397 3398 if (ST->hasXOP()) 3399 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy)) 3400 return LT.first * Entry->Cost; 3401 3402 if (ST->is64Bit()) 3403 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy)) 3404 return LT.first * Entry->Cost; 3405 3406 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy)) 3407 return LT.first * Entry->Cost; 3408 } 3409 3410 return BaseT::getIntrinsicInstrCost(ICA, CostKind); 3411 } 3412 3413 InstructionCost X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, 3414 unsigned Index) { 3415 static const CostTblEntry SLMCostTbl[] = { 3416 { ISD::EXTRACT_VECTOR_ELT, MVT::i8, 4 }, 3417 { ISD::EXTRACT_VECTOR_ELT, MVT::i16, 4 }, 3418 { ISD::EXTRACT_VECTOR_ELT, MVT::i32, 4 }, 3419 { ISD::EXTRACT_VECTOR_ELT, MVT::i64, 7 } 3420 }; 3421 3422 assert(Val->isVectorTy() && "This must be a vector type"); 3423 Type *ScalarType = Val->getScalarType(); 3424 int RegisterFileMoveCost = 0; 3425 3426 // Non-immediate extraction/insertion can be handled as a sequence of 3427 // aliased loads+stores via the stack. 3428 if (Index == -1U && (Opcode == Instruction::ExtractElement || 3429 Opcode == Instruction::InsertElement)) { 3430 // TODO: On some SSE41+ targets, we expand to cmp+splat+select patterns: 3431 // inselt N0, N1, N2 --> select (SplatN2 == {0,1,2...}) ? SplatN1 : N0. 3432 3433 // TODO: Move this to BasicTTIImpl.h? We'd need better gep + index handling. 3434 assert(isa<FixedVectorType>(Val) && "Fixed vector type expected"); 3435 Align VecAlign = DL.getPrefTypeAlign(Val); 3436 Align SclAlign = DL.getPrefTypeAlign(ScalarType); 3437 3438 // Extract - store vector to stack, load scalar. 3439 if (Opcode == Instruction::ExtractElement) { 3440 return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0, 3441 TTI::TargetCostKind::TCK_RecipThroughput) + 3442 getMemoryOpCost(Instruction::Load, ScalarType, SclAlign, 0, 3443 TTI::TargetCostKind::TCK_RecipThroughput); 3444 } 3445 // Insert - store vector to stack, store scalar, load vector. 3446 if (Opcode == Instruction::InsertElement) { 3447 return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0, 3448 TTI::TargetCostKind::TCK_RecipThroughput) + 3449 getMemoryOpCost(Instruction::Store, ScalarType, SclAlign, 0, 3450 TTI::TargetCostKind::TCK_RecipThroughput) + 3451 getMemoryOpCost(Instruction::Load, Val, VecAlign, 0, 3452 TTI::TargetCostKind::TCK_RecipThroughput); 3453 } 3454 } 3455 3456 if (Index != -1U && (Opcode == Instruction::ExtractElement || 3457 Opcode == Instruction::InsertElement)) { 3458 // Legalize the type. 3459 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Val); 3460 3461 // This type is legalized to a scalar type. 3462 if (!LT.second.isVector()) 3463 return 0; 3464 3465 // The type may be split. Normalize the index to the new type. 3466 unsigned NumElts = LT.second.getVectorNumElements(); 3467 unsigned SubNumElts = NumElts; 3468 Index = Index % NumElts; 3469 3470 // For >128-bit vectors, we need to extract higher 128-bit subvectors. 3471 // For inserts, we also need to insert the subvector back. 3472 if (LT.second.getSizeInBits() > 128) { 3473 assert((LT.second.getSizeInBits() % 128) == 0 && "Illegal vector"); 3474 unsigned NumSubVecs = LT.second.getSizeInBits() / 128; 3475 SubNumElts = NumElts / NumSubVecs; 3476 if (SubNumElts <= Index) { 3477 RegisterFileMoveCost += (Opcode == Instruction::InsertElement ? 2 : 1); 3478 Index %= SubNumElts; 3479 } 3480 } 3481 3482 if (Index == 0) { 3483 // Floating point scalars are already located in index #0. 3484 // Many insertions to #0 can fold away for scalar fp-ops, so let's assume 3485 // true for all. 3486 if (ScalarType->isFloatingPointTy()) 3487 return RegisterFileMoveCost; 3488 3489 // Assume movd/movq XMM -> GPR is relatively cheap on all targets. 3490 if (ScalarType->isIntegerTy() && Opcode == Instruction::ExtractElement) 3491 return 1 + RegisterFileMoveCost; 3492 } 3493 3494 int ISD = TLI->InstructionOpcodeToISD(Opcode); 3495 assert(ISD && "Unexpected vector opcode"); 3496 MVT MScalarTy = LT.second.getScalarType(); 3497 if (ST->useSLMArithCosts()) 3498 if (auto *Entry = CostTableLookup(SLMCostTbl, ISD, MScalarTy)) 3499 return Entry->Cost + RegisterFileMoveCost; 3500 3501 // Assume pinsr/pextr XMM <-> GPR is relatively cheap on all targets. 3502 if ((MScalarTy == MVT::i16 && ST->hasSSE2()) || 3503 (MScalarTy.isInteger() && ST->hasSSE41())) 3504 return 1 + RegisterFileMoveCost; 3505 3506 // Assume insertps is relatively cheap on all targets. 3507 if (MScalarTy == MVT::f32 && ST->hasSSE41() && 3508 Opcode == Instruction::InsertElement) 3509 return 1 + RegisterFileMoveCost; 3510 3511 // For extractions we just need to shuffle the element to index 0, which 3512 // should be very cheap (assume cost = 1). For insertions we need to shuffle 3513 // the elements to its destination. In both cases we must handle the 3514 // subvector move(s). 3515 // If the vector type is already less than 128-bits then don't reduce it. 3516 // TODO: Under what circumstances should we shuffle using the full width? 3517 InstructionCost ShuffleCost = 1; 3518 if (Opcode == Instruction::InsertElement) { 3519 auto *SubTy = cast<VectorType>(Val); 3520 EVT VT = TLI->getValueType(DL, Val); 3521 if (VT.getScalarType() != MScalarTy || VT.getSizeInBits() >= 128) 3522 SubTy = FixedVectorType::get(ScalarType, SubNumElts); 3523 ShuffleCost = 3524 getShuffleCost(TTI::SK_PermuteTwoSrc, SubTy, None, 0, SubTy); 3525 } 3526 int IntOrFpCost = ScalarType->isFloatingPointTy() ? 0 : 1; 3527 return ShuffleCost + IntOrFpCost + RegisterFileMoveCost; 3528 } 3529 3530 // Add to the base cost if we know that the extracted element of a vector is 3531 // destined to be moved to and used in the integer register file. 3532 if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy()) 3533 RegisterFileMoveCost += 1; 3534 3535 return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost; 3536 } 3537 3538 InstructionCost X86TTIImpl::getScalarizationOverhead(VectorType *Ty, 3539 const APInt &DemandedElts, 3540 bool Insert, 3541 bool Extract) { 3542 InstructionCost Cost = 0; 3543 3544 // For insertions, a ISD::BUILD_VECTOR style vector initialization can be much 3545 // cheaper than an accumulation of ISD::INSERT_VECTOR_ELT. 3546 if (Insert) { 3547 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 3548 MVT MScalarTy = LT.second.getScalarType(); 3549 3550 if ((MScalarTy == MVT::i16 && ST->hasSSE2()) || 3551 (MScalarTy.isInteger() && ST->hasSSE41()) || 3552 (MScalarTy == MVT::f32 && ST->hasSSE41())) { 3553 // For types we can insert directly, insertion into 128-bit sub vectors is 3554 // cheap, followed by a cheap chain of concatenations. 3555 if (LT.second.getSizeInBits() <= 128) { 3556 Cost += 3557 BaseT::getScalarizationOverhead(Ty, DemandedElts, Insert, false); 3558 } else { 3559 // In each 128-lane, if at least one index is demanded but not all 3560 // indices are demanded and this 128-lane is not the first 128-lane of 3561 // the legalized-vector, then this 128-lane needs a extracti128; If in 3562 // each 128-lane, there is at least one demanded index, this 128-lane 3563 // needs a inserti128. 3564 3565 // The following cases will help you build a better understanding: 3566 // Assume we insert several elements into a v8i32 vector in avx2, 3567 // Case#1: inserting into 1th index needs vpinsrd + inserti128. 3568 // Case#2: inserting into 5th index needs extracti128 + vpinsrd + 3569 // inserti128. 3570 // Case#3: inserting into 4,5,6,7 index needs 4*vpinsrd + inserti128. 3571 const int CostValue = *LT.first.getValue(); 3572 assert(CostValue >= 0 && "Negative cost!"); 3573 unsigned Num128Lanes = LT.second.getSizeInBits() / 128 * CostValue; 3574 unsigned NumElts = LT.second.getVectorNumElements() * CostValue; 3575 APInt WidenedDemandedElts = DemandedElts.zextOrSelf(NumElts); 3576 unsigned Scale = NumElts / Num128Lanes; 3577 // We iterate each 128-lane, and check if we need a 3578 // extracti128/inserti128 for this 128-lane. 3579 for (unsigned I = 0; I < NumElts; I += Scale) { 3580 APInt Mask = WidenedDemandedElts.getBitsSet(NumElts, I, I + Scale); 3581 APInt MaskedDE = Mask & WidenedDemandedElts; 3582 unsigned Population = MaskedDE.countPopulation(); 3583 Cost += (Population > 0 && Population != Scale && 3584 I % LT.second.getVectorNumElements() != 0); 3585 Cost += Population > 0; 3586 } 3587 Cost += DemandedElts.countPopulation(); 3588 3589 // For vXf32 cases, insertion into the 0'th index in each v4f32 3590 // 128-bit vector is free. 3591 // NOTE: This assumes legalization widens vXf32 vectors. 3592 if (MScalarTy == MVT::f32) 3593 for (unsigned i = 0, e = cast<FixedVectorType>(Ty)->getNumElements(); 3594 i < e; i += 4) 3595 if (DemandedElts[i]) 3596 Cost--; 3597 } 3598 } else if (LT.second.isVector()) { 3599 // Without fast insertion, we need to use MOVD/MOVQ to pass each demanded 3600 // integer element as a SCALAR_TO_VECTOR, then we build the vector as a 3601 // series of UNPCK followed by CONCAT_VECTORS - all of these can be 3602 // considered cheap. 3603 if (Ty->isIntOrIntVectorTy()) 3604 Cost += DemandedElts.countPopulation(); 3605 3606 // Get the smaller of the legalized or original pow2-extended number of 3607 // vector elements, which represents the number of unpacks we'll end up 3608 // performing. 3609 unsigned NumElts = LT.second.getVectorNumElements(); 3610 unsigned Pow2Elts = 3611 PowerOf2Ceil(cast<FixedVectorType>(Ty)->getNumElements()); 3612 Cost += (std::min<unsigned>(NumElts, Pow2Elts) - 1) * LT.first; 3613 } 3614 } 3615 3616 // TODO: Use default extraction for now, but we should investigate extending this 3617 // to handle repeated subvector extraction. 3618 if (Extract) 3619 Cost += BaseT::getScalarizationOverhead(Ty, DemandedElts, false, Extract); 3620 3621 return Cost; 3622 } 3623 3624 InstructionCost X86TTIImpl::getReplicationShuffleCost( 3625 Type *EltTy, int ReplicationFactor, int VF, 3626 const APInt &DemandedReplicatedElts, TTI::TargetCostKind CostKind) { 3627 const unsigned EltTyBits = DL.getTypeSizeInBits(EltTy); 3628 3629 auto bailout = [&]() { 3630 return BaseT::getReplicationShuffleCost(EltTy, ReplicationFactor, VF, 3631 DemandedReplicatedElts, CostKind); 3632 }; 3633 3634 // For now, only deal with AVX512 cases. 3635 if (!ST->hasAVX512()) 3636 return bailout(); 3637 3638 switch (EltTyBits) { 3639 case 32: 3640 case 64: 3641 break; // AVX512F. 3642 case 16: 3643 if (!ST->hasBWI()) 3644 return bailout(); 3645 break; 3646 case 8: 3647 if (!ST->hasVBMI()) 3648 return bailout(); 3649 break; 3650 default: 3651 return bailout(); 3652 } 3653 3654 auto *SrcVecTy = FixedVectorType::get(EltTy, VF); 3655 int NumReplicatedElements = VF * ReplicationFactor; 3656 auto *ReplicatedVecTy = FixedVectorType::get(EltTy, NumReplicatedElements); 3657 3658 // Legalize the types. 3659 MVT LegalSrcVecTy = TLI->getTypeLegalizationCost(DL, SrcVecTy).second; 3660 MVT LegalReplicatedVecTy = 3661 TLI->getTypeLegalizationCost(DL, ReplicatedVecTy).second; 3662 3663 // They both should have legalized into vector types. 3664 if (!LegalSrcVecTy.isVector() || !LegalReplicatedVecTy.isVector()) 3665 return bailout(); 3666 3667 assert(LegalSrcVecTy.getScalarSizeInBits() == EltTyBits && 3668 LegalSrcVecTy.getScalarType() == 3669 LegalReplicatedVecTy.getScalarType() && 3670 "We expect that the legalization doesn't affect the element width, " 3671 "doesn't coalesce/split elements."); 3672 3673 unsigned NumEltsPerReplicatedVec = 3674 LegalReplicatedVecTy.getVectorNumElements(); 3675 unsigned NumReplicatedVectors = 3676 divideCeil(ReplicatedVecTy->getNumElements(), NumEltsPerReplicatedVec); 3677 3678 auto *SingleReplicatedVecTy = 3679 FixedVectorType::get(EltTy, NumEltsPerReplicatedVec); 3680 3681 APInt DemandedReplicatedVectors = APIntOps::ScaleBitMask( 3682 DemandedReplicatedElts.zextOrSelf(NumReplicatedVectors * 3683 NumEltsPerReplicatedVec), 3684 NumReplicatedVectors); 3685 unsigned NumReplicatedVectorsDemanded = 3686 DemandedReplicatedVectors.countPopulation(); 3687 3688 InstructionCost SingleShuffleCost = 3689 getShuffleCost(TTI::SK_PermuteSingleSrc, SingleReplicatedVecTy, 3690 /*Mask=*/None, /*Index=*/0, /*SubTp=*/nullptr); 3691 return NumReplicatedVectorsDemanded * SingleShuffleCost; 3692 } 3693 3694 InstructionCost X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, 3695 MaybeAlign Alignment, 3696 unsigned AddressSpace, 3697 TTI::TargetCostKind CostKind, 3698 const Instruction *I) { 3699 // TODO: Handle other cost kinds. 3700 if (CostKind != TTI::TCK_RecipThroughput) { 3701 if (auto *SI = dyn_cast_or_null<StoreInst>(I)) { 3702 // Store instruction with index and scale costs 2 Uops. 3703 // Check the preceding GEP to identify non-const indices. 3704 if (auto *GEP = dyn_cast<GetElementPtrInst>(SI->getPointerOperand())) { 3705 if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); })) 3706 return TTI::TCC_Basic * 2; 3707 } 3708 } 3709 return TTI::TCC_Basic; 3710 } 3711 3712 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) && 3713 "Invalid Opcode"); 3714 // Type legalization can't handle structs 3715 if (TLI->getValueType(DL, Src, true) == MVT::Other) 3716 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, 3717 CostKind); 3718 3719 // Legalize the type. 3720 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Src); 3721 3722 auto *VTy = dyn_cast<FixedVectorType>(Src); 3723 3724 // Handle the simple case of non-vectors. 3725 // NOTE: this assumes that legalization never creates vector from scalars! 3726 if (!VTy || !LT.second.isVector()) 3727 // Each load/store unit costs 1. 3728 return LT.first * 1; 3729 3730 bool IsLoad = Opcode == Instruction::Load; 3731 3732 Type *EltTy = VTy->getElementType(); 3733 3734 const int EltTyBits = DL.getTypeSizeInBits(EltTy); 3735 3736 InstructionCost Cost = 0; 3737 3738 // Source of truth: how many elements were there in the original IR vector? 3739 const unsigned SrcNumElt = VTy->getNumElements(); 3740 3741 // How far have we gotten? 3742 int NumEltRemaining = SrcNumElt; 3743 // Note that we intentionally capture by-reference, NumEltRemaining changes. 3744 auto NumEltDone = [&]() { return SrcNumElt - NumEltRemaining; }; 3745 3746 const int MaxLegalOpSizeBytes = divideCeil(LT.second.getSizeInBits(), 8); 3747 3748 // Note that even if we can store 64 bits of an XMM, we still operate on XMM. 3749 const unsigned XMMBits = 128; 3750 if (XMMBits % EltTyBits != 0) 3751 // Vector size must be a multiple of the element size. I.e. no padding. 3752 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, 3753 CostKind); 3754 const int NumEltPerXMM = XMMBits / EltTyBits; 3755 3756 auto *XMMVecTy = FixedVectorType::get(EltTy, NumEltPerXMM); 3757 3758 for (int CurrOpSizeBytes = MaxLegalOpSizeBytes, SubVecEltsLeft = 0; 3759 NumEltRemaining > 0; CurrOpSizeBytes /= 2) { 3760 // How many elements would a single op deal with at once? 3761 if ((8 * CurrOpSizeBytes) % EltTyBits != 0) 3762 // Vector size must be a multiple of the element size. I.e. no padding. 3763 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, 3764 CostKind); 3765 int CurrNumEltPerOp = (8 * CurrOpSizeBytes) / EltTyBits; 3766 3767 assert(CurrOpSizeBytes > 0 && CurrNumEltPerOp > 0 && "How'd we get here?"); 3768 assert((((NumEltRemaining * EltTyBits) < (2 * 8 * CurrOpSizeBytes)) || 3769 (CurrOpSizeBytes == MaxLegalOpSizeBytes)) && 3770 "Unless we haven't halved the op size yet, " 3771 "we have less than two op's sized units of work left."); 3772 3773 auto *CurrVecTy = CurrNumEltPerOp > NumEltPerXMM 3774 ? FixedVectorType::get(EltTy, CurrNumEltPerOp) 3775 : XMMVecTy; 3776 3777 assert(CurrVecTy->getNumElements() % CurrNumEltPerOp == 0 && 3778 "After halving sizes, the vector elt count is no longer a multiple " 3779 "of number of elements per operation?"); 3780 auto *CoalescedVecTy = 3781 CurrNumEltPerOp == 1 3782 ? CurrVecTy 3783 : FixedVectorType::get( 3784 IntegerType::get(Src->getContext(), 3785 EltTyBits * CurrNumEltPerOp), 3786 CurrVecTy->getNumElements() / CurrNumEltPerOp); 3787 assert(DL.getTypeSizeInBits(CoalescedVecTy) == 3788 DL.getTypeSizeInBits(CurrVecTy) && 3789 "coalesciing elements doesn't change vector width."); 3790 3791 while (NumEltRemaining > 0) { 3792 assert(SubVecEltsLeft >= 0 && "Subreg element count overconsumtion?"); 3793 3794 // Can we use this vector size, as per the remaining element count? 3795 // Iff the vector is naturally aligned, we can do a wide load regardless. 3796 if (NumEltRemaining < CurrNumEltPerOp && 3797 (!IsLoad || Alignment.valueOrOne() < CurrOpSizeBytes) && 3798 CurrOpSizeBytes != 1) 3799 break; // Try smalled vector size. 3800 3801 bool Is0thSubVec = (NumEltDone() % LT.second.getVectorNumElements()) == 0; 3802 3803 // If we have fully processed the previous reg, we need to replenish it. 3804 if (SubVecEltsLeft == 0) { 3805 SubVecEltsLeft += CurrVecTy->getNumElements(); 3806 // And that's free only for the 0'th subvector of a legalized vector. 3807 if (!Is0thSubVec) 3808 Cost += getShuffleCost(IsLoad ? TTI::ShuffleKind::SK_InsertSubvector 3809 : TTI::ShuffleKind::SK_ExtractSubvector, 3810 VTy, None, NumEltDone(), CurrVecTy); 3811 } 3812 3813 // While we can directly load/store ZMM, YMM, and 64-bit halves of XMM, 3814 // for smaller widths (32/16/8) we have to insert/extract them separately. 3815 // Again, it's free for the 0'th subreg (if op is 32/64 bit wide, 3816 // but let's pretend that it is also true for 16/8 bit wide ops...) 3817 if (CurrOpSizeBytes <= 32 / 8 && !Is0thSubVec) { 3818 int NumEltDoneInCurrXMM = NumEltDone() % NumEltPerXMM; 3819 assert(NumEltDoneInCurrXMM % CurrNumEltPerOp == 0 && ""); 3820 int CoalescedVecEltIdx = NumEltDoneInCurrXMM / CurrNumEltPerOp; 3821 APInt DemandedElts = 3822 APInt::getBitsSet(CoalescedVecTy->getNumElements(), 3823 CoalescedVecEltIdx, CoalescedVecEltIdx + 1); 3824 assert(DemandedElts.countPopulation() == 1 && "Inserting single value"); 3825 Cost += getScalarizationOverhead(CoalescedVecTy, DemandedElts, IsLoad, 3826 !IsLoad); 3827 } 3828 3829 // This isn't exactly right. We're using slow unaligned 32-byte accesses 3830 // as a proxy for a double-pumped AVX memory interface such as on 3831 // Sandybridge. 3832 if (CurrOpSizeBytes == 32 && ST->isUnalignedMem32Slow()) 3833 Cost += 2; 3834 else 3835 Cost += 1; 3836 3837 SubVecEltsLeft -= CurrNumEltPerOp; 3838 NumEltRemaining -= CurrNumEltPerOp; 3839 Alignment = commonAlignment(Alignment.valueOrOne(), CurrOpSizeBytes); 3840 } 3841 } 3842 3843 assert(NumEltRemaining <= 0 && "Should have processed all the elements."); 3844 3845 return Cost; 3846 } 3847 3848 InstructionCost 3849 X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy, Align Alignment, 3850 unsigned AddressSpace, 3851 TTI::TargetCostKind CostKind) { 3852 bool IsLoad = (Instruction::Load == Opcode); 3853 bool IsStore = (Instruction::Store == Opcode); 3854 3855 auto *SrcVTy = dyn_cast<FixedVectorType>(SrcTy); 3856 if (!SrcVTy) 3857 // To calculate scalar take the regular cost, without mask 3858 return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace, CostKind); 3859 3860 unsigned NumElem = SrcVTy->getNumElements(); 3861 auto *MaskTy = 3862 FixedVectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem); 3863 if ((IsLoad && !isLegalMaskedLoad(SrcVTy, Alignment)) || 3864 (IsStore && !isLegalMaskedStore(SrcVTy, Alignment))) { 3865 // Scalarization 3866 APInt DemandedElts = APInt::getAllOnes(NumElem); 3867 InstructionCost MaskSplitCost = 3868 getScalarizationOverhead(MaskTy, DemandedElts, false, true); 3869 InstructionCost ScalarCompareCost = getCmpSelInstrCost( 3870 Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr, 3871 CmpInst::BAD_ICMP_PREDICATE, CostKind); 3872 InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind); 3873 InstructionCost MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost); 3874 InstructionCost ValueSplitCost = 3875 getScalarizationOverhead(SrcVTy, DemandedElts, IsLoad, IsStore); 3876 InstructionCost MemopCost = 3877 NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 3878 Alignment, AddressSpace, CostKind); 3879 return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost; 3880 } 3881 3882 // Legalize the type. 3883 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy); 3884 auto VT = TLI->getValueType(DL, SrcVTy); 3885 InstructionCost Cost = 0; 3886 if (VT.isSimple() && LT.second != VT.getSimpleVT() && 3887 LT.second.getVectorNumElements() == NumElem) 3888 // Promotion requires extend/truncate for data and a shuffle for mask. 3889 Cost += getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVTy, None, 0, nullptr) + 3890 getShuffleCost(TTI::SK_PermuteTwoSrc, MaskTy, None, 0, nullptr); 3891 3892 else if (LT.first * LT.second.getVectorNumElements() > NumElem) { 3893 auto *NewMaskTy = FixedVectorType::get(MaskTy->getElementType(), 3894 LT.second.getVectorNumElements()); 3895 // Expanding requires fill mask with zeroes 3896 Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, None, 0, MaskTy); 3897 } 3898 3899 // Pre-AVX512 - each maskmov load costs 2 + store costs ~8. 3900 if (!ST->hasAVX512()) 3901 return Cost + LT.first * (IsLoad ? 2 : 8); 3902 3903 // AVX-512 masked load/store is cheapper 3904 return Cost + LT.first; 3905 } 3906 3907 InstructionCost X86TTIImpl::getAddressComputationCost(Type *Ty, 3908 ScalarEvolution *SE, 3909 const SCEV *Ptr) { 3910 // Address computations in vectorized code with non-consecutive addresses will 3911 // likely result in more instructions compared to scalar code where the 3912 // computation can more often be merged into the index mode. The resulting 3913 // extra micro-ops can significantly decrease throughput. 3914 const unsigned NumVectorInstToHideOverhead = 10; 3915 3916 // Cost modeling of Strided Access Computation is hidden by the indexing 3917 // modes of X86 regardless of the stride value. We dont believe that there 3918 // is a difference between constant strided access in gerenal and constant 3919 // strided value which is less than or equal to 64. 3920 // Even in the case of (loop invariant) stride whose value is not known at 3921 // compile time, the address computation will not incur more than one extra 3922 // ADD instruction. 3923 if (Ty->isVectorTy() && SE) { 3924 if (!BaseT::isStridedAccess(Ptr)) 3925 return NumVectorInstToHideOverhead; 3926 if (!BaseT::getConstantStrideStep(SE, Ptr)) 3927 return 1; 3928 } 3929 3930 return BaseT::getAddressComputationCost(Ty, SE, Ptr); 3931 } 3932 3933 InstructionCost 3934 X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy, 3935 Optional<FastMathFlags> FMF, 3936 TTI::TargetCostKind CostKind) { 3937 if (TTI::requiresOrderedReduction(FMF)) 3938 return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind); 3939 3940 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput 3941 // and make it as the cost. 3942 3943 static const CostTblEntry SLMCostTblNoPairWise[] = { 3944 { ISD::FADD, MVT::v2f64, 3 }, 3945 { ISD::ADD, MVT::v2i64, 5 }, 3946 }; 3947 3948 static const CostTblEntry SSE2CostTblNoPairWise[] = { 3949 { ISD::FADD, MVT::v2f64, 2 }, 3950 { ISD::FADD, MVT::v2f32, 2 }, 3951 { ISD::FADD, MVT::v4f32, 4 }, 3952 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6". 3953 { ISD::ADD, MVT::v2i32, 2 }, // FIXME: chosen to be less than v4i32 3954 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3". 3955 { ISD::ADD, MVT::v2i16, 2 }, // The data reported by the IACA tool is "4.3". 3956 { ISD::ADD, MVT::v4i16, 3 }, // The data reported by the IACA tool is "4.3". 3957 { ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3". 3958 { ISD::ADD, MVT::v2i8, 2 }, 3959 { ISD::ADD, MVT::v4i8, 2 }, 3960 { ISD::ADD, MVT::v8i8, 2 }, 3961 { ISD::ADD, MVT::v16i8, 3 }, 3962 }; 3963 3964 static const CostTblEntry AVX1CostTblNoPairWise[] = { 3965 { ISD::FADD, MVT::v4f64, 3 }, 3966 { ISD::FADD, MVT::v4f32, 3 }, 3967 { ISD::FADD, MVT::v8f32, 4 }, 3968 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5". 3969 { ISD::ADD, MVT::v4i64, 3 }, 3970 { ISD::ADD, MVT::v8i32, 5 }, 3971 { ISD::ADD, MVT::v16i16, 5 }, 3972 { ISD::ADD, MVT::v32i8, 4 }, 3973 }; 3974 3975 int ISD = TLI->InstructionOpcodeToISD(Opcode); 3976 assert(ISD && "Invalid opcode"); 3977 3978 // Before legalizing the type, give a chance to look up illegal narrow types 3979 // in the table. 3980 // FIXME: Is there a better way to do this? 3981 EVT VT = TLI->getValueType(DL, ValTy); 3982 if (VT.isSimple()) { 3983 MVT MTy = VT.getSimpleVT(); 3984 if (ST->useSLMArithCosts()) 3985 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy)) 3986 return Entry->Cost; 3987 3988 if (ST->hasAVX()) 3989 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3990 return Entry->Cost; 3991 3992 if (ST->hasSSE2()) 3993 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3994 return Entry->Cost; 3995 } 3996 3997 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 3998 3999 MVT MTy = LT.second; 4000 4001 auto *ValVTy = cast<FixedVectorType>(ValTy); 4002 4003 // Special case: vXi8 mul reductions are performed as vXi16. 4004 if (ISD == ISD::MUL && MTy.getScalarType() == MVT::i8) { 4005 auto *WideSclTy = IntegerType::get(ValVTy->getContext(), 16); 4006 auto *WideVecTy = FixedVectorType::get(WideSclTy, ValVTy->getNumElements()); 4007 return getCastInstrCost(Instruction::ZExt, WideVecTy, ValTy, 4008 TargetTransformInfo::CastContextHint::None, 4009 CostKind) + 4010 getArithmeticReductionCost(Opcode, WideVecTy, FMF, CostKind); 4011 } 4012 4013 InstructionCost ArithmeticCost = 0; 4014 if (LT.first != 1 && MTy.isVector() && 4015 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 4016 // Type needs to be split. We need LT.first - 1 arithmetic ops. 4017 auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(), 4018 MTy.getVectorNumElements()); 4019 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind); 4020 ArithmeticCost *= LT.first - 1; 4021 } 4022 4023 if (ST->useSLMArithCosts()) 4024 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy)) 4025 return ArithmeticCost + Entry->Cost; 4026 4027 if (ST->hasAVX()) 4028 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 4029 return ArithmeticCost + Entry->Cost; 4030 4031 if (ST->hasSSE2()) 4032 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 4033 return ArithmeticCost + Entry->Cost; 4034 4035 // FIXME: These assume a naive kshift+binop lowering, which is probably 4036 // conservative in most cases. 4037 static const CostTblEntry AVX512BoolReduction[] = { 4038 { ISD::AND, MVT::v2i1, 3 }, 4039 { ISD::AND, MVT::v4i1, 5 }, 4040 { ISD::AND, MVT::v8i1, 7 }, 4041 { ISD::AND, MVT::v16i1, 9 }, 4042 { ISD::AND, MVT::v32i1, 11 }, 4043 { ISD::AND, MVT::v64i1, 13 }, 4044 { ISD::OR, MVT::v2i1, 3 }, 4045 { ISD::OR, MVT::v4i1, 5 }, 4046 { ISD::OR, MVT::v8i1, 7 }, 4047 { ISD::OR, MVT::v16i1, 9 }, 4048 { ISD::OR, MVT::v32i1, 11 }, 4049 { ISD::OR, MVT::v64i1, 13 }, 4050 }; 4051 4052 static const CostTblEntry AVX2BoolReduction[] = { 4053 { ISD::AND, MVT::v16i16, 2 }, // vpmovmskb + cmp 4054 { ISD::AND, MVT::v32i8, 2 }, // vpmovmskb + cmp 4055 { ISD::OR, MVT::v16i16, 2 }, // vpmovmskb + cmp 4056 { ISD::OR, MVT::v32i8, 2 }, // vpmovmskb + cmp 4057 }; 4058 4059 static const CostTblEntry AVX1BoolReduction[] = { 4060 { ISD::AND, MVT::v4i64, 2 }, // vmovmskpd + cmp 4061 { ISD::AND, MVT::v8i32, 2 }, // vmovmskps + cmp 4062 { ISD::AND, MVT::v16i16, 4 }, // vextractf128 + vpand + vpmovmskb + cmp 4063 { ISD::AND, MVT::v32i8, 4 }, // vextractf128 + vpand + vpmovmskb + cmp 4064 { ISD::OR, MVT::v4i64, 2 }, // vmovmskpd + cmp 4065 { ISD::OR, MVT::v8i32, 2 }, // vmovmskps + cmp 4066 { ISD::OR, MVT::v16i16, 4 }, // vextractf128 + vpor + vpmovmskb + cmp 4067 { ISD::OR, MVT::v32i8, 4 }, // vextractf128 + vpor + vpmovmskb + cmp 4068 }; 4069 4070 static const CostTblEntry SSE2BoolReduction[] = { 4071 { ISD::AND, MVT::v2i64, 2 }, // movmskpd + cmp 4072 { ISD::AND, MVT::v4i32, 2 }, // movmskps + cmp 4073 { ISD::AND, MVT::v8i16, 2 }, // pmovmskb + cmp 4074 { ISD::AND, MVT::v16i8, 2 }, // pmovmskb + cmp 4075 { ISD::OR, MVT::v2i64, 2 }, // movmskpd + cmp 4076 { ISD::OR, MVT::v4i32, 2 }, // movmskps + cmp 4077 { ISD::OR, MVT::v8i16, 2 }, // pmovmskb + cmp 4078 { ISD::OR, MVT::v16i8, 2 }, // pmovmskb + cmp 4079 }; 4080 4081 // Handle bool allof/anyof patterns. 4082 if (ValVTy->getElementType()->isIntegerTy(1)) { 4083 InstructionCost ArithmeticCost = 0; 4084 if (LT.first != 1 && MTy.isVector() && 4085 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 4086 // Type needs to be split. We need LT.first - 1 arithmetic ops. 4087 auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(), 4088 MTy.getVectorNumElements()); 4089 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind); 4090 ArithmeticCost *= LT.first - 1; 4091 } 4092 4093 if (ST->hasAVX512()) 4094 if (const auto *Entry = CostTableLookup(AVX512BoolReduction, ISD, MTy)) 4095 return ArithmeticCost + Entry->Cost; 4096 if (ST->hasAVX2()) 4097 if (const auto *Entry = CostTableLookup(AVX2BoolReduction, ISD, MTy)) 4098 return ArithmeticCost + Entry->Cost; 4099 if (ST->hasAVX()) 4100 if (const auto *Entry = CostTableLookup(AVX1BoolReduction, ISD, MTy)) 4101 return ArithmeticCost + Entry->Cost; 4102 if (ST->hasSSE2()) 4103 if (const auto *Entry = CostTableLookup(SSE2BoolReduction, ISD, MTy)) 4104 return ArithmeticCost + Entry->Cost; 4105 4106 return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind); 4107 } 4108 4109 unsigned NumVecElts = ValVTy->getNumElements(); 4110 unsigned ScalarSize = ValVTy->getScalarSizeInBits(); 4111 4112 // Special case power of 2 reductions where the scalar type isn't changed 4113 // by type legalization. 4114 if (!isPowerOf2_32(NumVecElts) || ScalarSize != MTy.getScalarSizeInBits()) 4115 return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind); 4116 4117 InstructionCost ReductionCost = 0; 4118 4119 auto *Ty = ValVTy; 4120 if (LT.first != 1 && MTy.isVector() && 4121 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 4122 // Type needs to be split. We need LT.first - 1 arithmetic ops. 4123 Ty = FixedVectorType::get(ValVTy->getElementType(), 4124 MTy.getVectorNumElements()); 4125 ReductionCost = getArithmeticInstrCost(Opcode, Ty, CostKind); 4126 ReductionCost *= LT.first - 1; 4127 NumVecElts = MTy.getVectorNumElements(); 4128 } 4129 4130 // Now handle reduction with the legal type, taking into account size changes 4131 // at each level. 4132 while (NumVecElts > 1) { 4133 // Determine the size of the remaining vector we need to reduce. 4134 unsigned Size = NumVecElts * ScalarSize; 4135 NumVecElts /= 2; 4136 // If we're reducing from 256/512 bits, use an extract_subvector. 4137 if (Size > 128) { 4138 auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts); 4139 ReductionCost += 4140 getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy); 4141 Ty = SubTy; 4142 } else if (Size == 128) { 4143 // Reducing from 128 bits is a permute of v2f64/v2i64. 4144 FixedVectorType *ShufTy; 4145 if (ValVTy->isFloatingPointTy()) 4146 ShufTy = 4147 FixedVectorType::get(Type::getDoubleTy(ValVTy->getContext()), 2); 4148 else 4149 ShufTy = 4150 FixedVectorType::get(Type::getInt64Ty(ValVTy->getContext()), 2); 4151 ReductionCost += 4152 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr); 4153 } else if (Size == 64) { 4154 // Reducing from 64 bits is a shuffle of v4f32/v4i32. 4155 FixedVectorType *ShufTy; 4156 if (ValVTy->isFloatingPointTy()) 4157 ShufTy = 4158 FixedVectorType::get(Type::getFloatTy(ValVTy->getContext()), 4); 4159 else 4160 ShufTy = 4161 FixedVectorType::get(Type::getInt32Ty(ValVTy->getContext()), 4); 4162 ReductionCost += 4163 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr); 4164 } else { 4165 // Reducing from smaller size is a shift by immediate. 4166 auto *ShiftTy = FixedVectorType::get( 4167 Type::getIntNTy(ValVTy->getContext(), Size), 128 / Size); 4168 ReductionCost += getArithmeticInstrCost( 4169 Instruction::LShr, ShiftTy, CostKind, 4170 TargetTransformInfo::OK_AnyValue, 4171 TargetTransformInfo::OK_UniformConstantValue, 4172 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 4173 } 4174 4175 // Add the arithmetic op for this level. 4176 ReductionCost += getArithmeticInstrCost(Opcode, Ty, CostKind); 4177 } 4178 4179 // Add the final extract element to the cost. 4180 return ReductionCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0); 4181 } 4182 4183 InstructionCost X86TTIImpl::getMinMaxCost(Type *Ty, Type *CondTy, 4184 bool IsUnsigned) { 4185 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 4186 4187 MVT MTy = LT.second; 4188 4189 int ISD; 4190 if (Ty->isIntOrIntVectorTy()) { 4191 ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN; 4192 } else { 4193 assert(Ty->isFPOrFPVectorTy() && 4194 "Expected float point or integer vector type."); 4195 ISD = ISD::FMINNUM; 4196 } 4197 4198 static const CostTblEntry SSE1CostTbl[] = { 4199 {ISD::FMINNUM, MVT::v4f32, 1}, 4200 }; 4201 4202 static const CostTblEntry SSE2CostTbl[] = { 4203 {ISD::FMINNUM, MVT::v2f64, 1}, 4204 {ISD::SMIN, MVT::v8i16, 1}, 4205 {ISD::UMIN, MVT::v16i8, 1}, 4206 }; 4207 4208 static const CostTblEntry SSE41CostTbl[] = { 4209 {ISD::SMIN, MVT::v4i32, 1}, 4210 {ISD::UMIN, MVT::v4i32, 1}, 4211 {ISD::UMIN, MVT::v8i16, 1}, 4212 {ISD::SMIN, MVT::v16i8, 1}, 4213 }; 4214 4215 static const CostTblEntry SSE42CostTbl[] = { 4216 {ISD::UMIN, MVT::v2i64, 3}, // xor+pcmpgtq+blendvpd 4217 }; 4218 4219 static const CostTblEntry AVX1CostTbl[] = { 4220 {ISD::FMINNUM, MVT::v8f32, 1}, 4221 {ISD::FMINNUM, MVT::v4f64, 1}, 4222 {ISD::SMIN, MVT::v8i32, 3}, 4223 {ISD::UMIN, MVT::v8i32, 3}, 4224 {ISD::SMIN, MVT::v16i16, 3}, 4225 {ISD::UMIN, MVT::v16i16, 3}, 4226 {ISD::SMIN, MVT::v32i8, 3}, 4227 {ISD::UMIN, MVT::v32i8, 3}, 4228 }; 4229 4230 static const CostTblEntry AVX2CostTbl[] = { 4231 {ISD::SMIN, MVT::v8i32, 1}, 4232 {ISD::UMIN, MVT::v8i32, 1}, 4233 {ISD::SMIN, MVT::v16i16, 1}, 4234 {ISD::UMIN, MVT::v16i16, 1}, 4235 {ISD::SMIN, MVT::v32i8, 1}, 4236 {ISD::UMIN, MVT::v32i8, 1}, 4237 }; 4238 4239 static const CostTblEntry AVX512CostTbl[] = { 4240 {ISD::FMINNUM, MVT::v16f32, 1}, 4241 {ISD::FMINNUM, MVT::v8f64, 1}, 4242 {ISD::SMIN, MVT::v2i64, 1}, 4243 {ISD::UMIN, MVT::v2i64, 1}, 4244 {ISD::SMIN, MVT::v4i64, 1}, 4245 {ISD::UMIN, MVT::v4i64, 1}, 4246 {ISD::SMIN, MVT::v8i64, 1}, 4247 {ISD::UMIN, MVT::v8i64, 1}, 4248 {ISD::SMIN, MVT::v16i32, 1}, 4249 {ISD::UMIN, MVT::v16i32, 1}, 4250 }; 4251 4252 static const CostTblEntry AVX512BWCostTbl[] = { 4253 {ISD::SMIN, MVT::v32i16, 1}, 4254 {ISD::UMIN, MVT::v32i16, 1}, 4255 {ISD::SMIN, MVT::v64i8, 1}, 4256 {ISD::UMIN, MVT::v64i8, 1}, 4257 }; 4258 4259 // If we have a native MIN/MAX instruction for this type, use it. 4260 if (ST->hasBWI()) 4261 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 4262 return LT.first * Entry->Cost; 4263 4264 if (ST->hasAVX512()) 4265 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 4266 return LT.first * Entry->Cost; 4267 4268 if (ST->hasAVX2()) 4269 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 4270 return LT.first * Entry->Cost; 4271 4272 if (ST->hasAVX()) 4273 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 4274 return LT.first * Entry->Cost; 4275 4276 if (ST->hasSSE42()) 4277 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 4278 return LT.first * Entry->Cost; 4279 4280 if (ST->hasSSE41()) 4281 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 4282 return LT.first * Entry->Cost; 4283 4284 if (ST->hasSSE2()) 4285 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 4286 return LT.first * Entry->Cost; 4287 4288 if (ST->hasSSE1()) 4289 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 4290 return LT.first * Entry->Cost; 4291 4292 unsigned CmpOpcode; 4293 if (Ty->isFPOrFPVectorTy()) { 4294 CmpOpcode = Instruction::FCmp; 4295 } else { 4296 assert(Ty->isIntOrIntVectorTy() && 4297 "expecting floating point or integer type for min/max reduction"); 4298 CmpOpcode = Instruction::ICmp; 4299 } 4300 4301 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; 4302 // Otherwise fall back to cmp+select. 4303 InstructionCost Result = 4304 getCmpSelInstrCost(CmpOpcode, Ty, CondTy, CmpInst::BAD_ICMP_PREDICATE, 4305 CostKind) + 4306 getCmpSelInstrCost(Instruction::Select, Ty, CondTy, 4307 CmpInst::BAD_ICMP_PREDICATE, CostKind); 4308 return Result; 4309 } 4310 4311 InstructionCost 4312 X86TTIImpl::getMinMaxReductionCost(VectorType *ValTy, VectorType *CondTy, 4313 bool IsUnsigned, 4314 TTI::TargetCostKind CostKind) { 4315 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 4316 4317 MVT MTy = LT.second; 4318 4319 int ISD; 4320 if (ValTy->isIntOrIntVectorTy()) { 4321 ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN; 4322 } else { 4323 assert(ValTy->isFPOrFPVectorTy() && 4324 "Expected float point or integer vector type."); 4325 ISD = ISD::FMINNUM; 4326 } 4327 4328 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput 4329 // and make it as the cost. 4330 4331 static const CostTblEntry SSE2CostTblNoPairWise[] = { 4332 {ISD::UMIN, MVT::v2i16, 5}, // need pxors to use pminsw/pmaxsw 4333 {ISD::UMIN, MVT::v4i16, 7}, // need pxors to use pminsw/pmaxsw 4334 {ISD::UMIN, MVT::v8i16, 9}, // need pxors to use pminsw/pmaxsw 4335 }; 4336 4337 static const CostTblEntry SSE41CostTblNoPairWise[] = { 4338 {ISD::SMIN, MVT::v2i16, 3}, // same as sse2 4339 {ISD::SMIN, MVT::v4i16, 5}, // same as sse2 4340 {ISD::UMIN, MVT::v2i16, 5}, // same as sse2 4341 {ISD::UMIN, MVT::v4i16, 7}, // same as sse2 4342 {ISD::SMIN, MVT::v8i16, 4}, // phminposuw+xor 4343 {ISD::UMIN, MVT::v8i16, 4}, // FIXME: umin is cheaper than umax 4344 {ISD::SMIN, MVT::v2i8, 3}, // pminsb 4345 {ISD::SMIN, MVT::v4i8, 5}, // pminsb 4346 {ISD::SMIN, MVT::v8i8, 7}, // pminsb 4347 {ISD::SMIN, MVT::v16i8, 6}, 4348 {ISD::UMIN, MVT::v2i8, 3}, // same as sse2 4349 {ISD::UMIN, MVT::v4i8, 5}, // same as sse2 4350 {ISD::UMIN, MVT::v8i8, 7}, // same as sse2 4351 {ISD::UMIN, MVT::v16i8, 6}, // FIXME: umin is cheaper than umax 4352 }; 4353 4354 static const CostTblEntry AVX1CostTblNoPairWise[] = { 4355 {ISD::SMIN, MVT::v16i16, 6}, 4356 {ISD::UMIN, MVT::v16i16, 6}, // FIXME: umin is cheaper than umax 4357 {ISD::SMIN, MVT::v32i8, 8}, 4358 {ISD::UMIN, MVT::v32i8, 8}, 4359 }; 4360 4361 static const CostTblEntry AVX512BWCostTblNoPairWise[] = { 4362 {ISD::SMIN, MVT::v32i16, 8}, 4363 {ISD::UMIN, MVT::v32i16, 8}, // FIXME: umin is cheaper than umax 4364 {ISD::SMIN, MVT::v64i8, 10}, 4365 {ISD::UMIN, MVT::v64i8, 10}, 4366 }; 4367 4368 // Before legalizing the type, give a chance to look up illegal narrow types 4369 // in the table. 4370 // FIXME: Is there a better way to do this? 4371 EVT VT = TLI->getValueType(DL, ValTy); 4372 if (VT.isSimple()) { 4373 MVT MTy = VT.getSimpleVT(); 4374 if (ST->hasBWI()) 4375 if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy)) 4376 return Entry->Cost; 4377 4378 if (ST->hasAVX()) 4379 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 4380 return Entry->Cost; 4381 4382 if (ST->hasSSE41()) 4383 if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy)) 4384 return Entry->Cost; 4385 4386 if (ST->hasSSE2()) 4387 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 4388 return Entry->Cost; 4389 } 4390 4391 auto *ValVTy = cast<FixedVectorType>(ValTy); 4392 unsigned NumVecElts = ValVTy->getNumElements(); 4393 4394 auto *Ty = ValVTy; 4395 InstructionCost MinMaxCost = 0; 4396 if (LT.first != 1 && MTy.isVector() && 4397 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 4398 // Type needs to be split. We need LT.first - 1 operations ops. 4399 Ty = FixedVectorType::get(ValVTy->getElementType(), 4400 MTy.getVectorNumElements()); 4401 auto *SubCondTy = FixedVectorType::get(CondTy->getElementType(), 4402 MTy.getVectorNumElements()); 4403 MinMaxCost = getMinMaxCost(Ty, SubCondTy, IsUnsigned); 4404 MinMaxCost *= LT.first - 1; 4405 NumVecElts = MTy.getVectorNumElements(); 4406 } 4407 4408 if (ST->hasBWI()) 4409 if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy)) 4410 return MinMaxCost + Entry->Cost; 4411 4412 if (ST->hasAVX()) 4413 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 4414 return MinMaxCost + Entry->Cost; 4415 4416 if (ST->hasSSE41()) 4417 if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy)) 4418 return MinMaxCost + Entry->Cost; 4419 4420 if (ST->hasSSE2()) 4421 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 4422 return MinMaxCost + Entry->Cost; 4423 4424 unsigned ScalarSize = ValTy->getScalarSizeInBits(); 4425 4426 // Special case power of 2 reductions where the scalar type isn't changed 4427 // by type legalization. 4428 if (!isPowerOf2_32(ValVTy->getNumElements()) || 4429 ScalarSize != MTy.getScalarSizeInBits()) 4430 return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsUnsigned, CostKind); 4431 4432 // Now handle reduction with the legal type, taking into account size changes 4433 // at each level. 4434 while (NumVecElts > 1) { 4435 // Determine the size of the remaining vector we need to reduce. 4436 unsigned Size = NumVecElts * ScalarSize; 4437 NumVecElts /= 2; 4438 // If we're reducing from 256/512 bits, use an extract_subvector. 4439 if (Size > 128) { 4440 auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts); 4441 MinMaxCost += 4442 getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy); 4443 Ty = SubTy; 4444 } else if (Size == 128) { 4445 // Reducing from 128 bits is a permute of v2f64/v2i64. 4446 VectorType *ShufTy; 4447 if (ValTy->isFloatingPointTy()) 4448 ShufTy = 4449 FixedVectorType::get(Type::getDoubleTy(ValTy->getContext()), 2); 4450 else 4451 ShufTy = FixedVectorType::get(Type::getInt64Ty(ValTy->getContext()), 2); 4452 MinMaxCost += 4453 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr); 4454 } else if (Size == 64) { 4455 // Reducing from 64 bits is a shuffle of v4f32/v4i32. 4456 FixedVectorType *ShufTy; 4457 if (ValTy->isFloatingPointTy()) 4458 ShufTy = FixedVectorType::get(Type::getFloatTy(ValTy->getContext()), 4); 4459 else 4460 ShufTy = FixedVectorType::get(Type::getInt32Ty(ValTy->getContext()), 4); 4461 MinMaxCost += 4462 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr); 4463 } else { 4464 // Reducing from smaller size is a shift by immediate. 4465 auto *ShiftTy = FixedVectorType::get( 4466 Type::getIntNTy(ValTy->getContext(), Size), 128 / Size); 4467 MinMaxCost += getArithmeticInstrCost( 4468 Instruction::LShr, ShiftTy, TTI::TCK_RecipThroughput, 4469 TargetTransformInfo::OK_AnyValue, 4470 TargetTransformInfo::OK_UniformConstantValue, 4471 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 4472 } 4473 4474 // Add the arithmetic op for this level. 4475 auto *SubCondTy = 4476 FixedVectorType::get(CondTy->getElementType(), Ty->getNumElements()); 4477 MinMaxCost += getMinMaxCost(Ty, SubCondTy, IsUnsigned); 4478 } 4479 4480 // Add the final extract element to the cost. 4481 return MinMaxCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0); 4482 } 4483 4484 /// Calculate the cost of materializing a 64-bit value. This helper 4485 /// method might only calculate a fraction of a larger immediate. Therefore it 4486 /// is valid to return a cost of ZERO. 4487 InstructionCost X86TTIImpl::getIntImmCost(int64_t Val) { 4488 if (Val == 0) 4489 return TTI::TCC_Free; 4490 4491 if (isInt<32>(Val)) 4492 return TTI::TCC_Basic; 4493 4494 return 2 * TTI::TCC_Basic; 4495 } 4496 4497 InstructionCost X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty, 4498 TTI::TargetCostKind CostKind) { 4499 assert(Ty->isIntegerTy()); 4500 4501 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 4502 if (BitSize == 0) 4503 return ~0U; 4504 4505 // Never hoist constants larger than 128bit, because this might lead to 4506 // incorrect code generation or assertions in codegen. 4507 // Fixme: Create a cost model for types larger than i128 once the codegen 4508 // issues have been fixed. 4509 if (BitSize > 128) 4510 return TTI::TCC_Free; 4511 4512 if (Imm == 0) 4513 return TTI::TCC_Free; 4514 4515 // Sign-extend all constants to a multiple of 64-bit. 4516 APInt ImmVal = Imm; 4517 if (BitSize % 64 != 0) 4518 ImmVal = Imm.sext(alignTo(BitSize, 64)); 4519 4520 // Split the constant into 64-bit chunks and calculate the cost for each 4521 // chunk. 4522 InstructionCost Cost = 0; 4523 for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) { 4524 APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64); 4525 int64_t Val = Tmp.getSExtValue(); 4526 Cost += getIntImmCost(Val); 4527 } 4528 // We need at least one instruction to materialize the constant. 4529 return std::max<InstructionCost>(1, Cost); 4530 } 4531 4532 InstructionCost X86TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx, 4533 const APInt &Imm, Type *Ty, 4534 TTI::TargetCostKind CostKind, 4535 Instruction *Inst) { 4536 assert(Ty->isIntegerTy()); 4537 4538 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 4539 // There is no cost model for constants with a bit size of 0. Return TCC_Free 4540 // here, so that constant hoisting will ignore this constant. 4541 if (BitSize == 0) 4542 return TTI::TCC_Free; 4543 4544 unsigned ImmIdx = ~0U; 4545 switch (Opcode) { 4546 default: 4547 return TTI::TCC_Free; 4548 case Instruction::GetElementPtr: 4549 // Always hoist the base address of a GetElementPtr. This prevents the 4550 // creation of new constants for every base constant that gets constant 4551 // folded with the offset. 4552 if (Idx == 0) 4553 return 2 * TTI::TCC_Basic; 4554 return TTI::TCC_Free; 4555 case Instruction::Store: 4556 ImmIdx = 0; 4557 break; 4558 case Instruction::ICmp: 4559 // This is an imperfect hack to prevent constant hoisting of 4560 // compares that might be trying to check if a 64-bit value fits in 4561 // 32-bits. The backend can optimize these cases using a right shift by 32. 4562 // Ideally we would check the compare predicate here. There also other 4563 // similar immediates the backend can use shifts for. 4564 if (Idx == 1 && Imm.getBitWidth() == 64) { 4565 uint64_t ImmVal = Imm.getZExtValue(); 4566 if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff) 4567 return TTI::TCC_Free; 4568 } 4569 ImmIdx = 1; 4570 break; 4571 case Instruction::And: 4572 // We support 64-bit ANDs with immediates with 32-bits of leading zeroes 4573 // by using a 32-bit operation with implicit zero extension. Detect such 4574 // immediates here as the normal path expects bit 31 to be sign extended. 4575 if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue())) 4576 return TTI::TCC_Free; 4577 ImmIdx = 1; 4578 break; 4579 case Instruction::Add: 4580 case Instruction::Sub: 4581 // For add/sub, we can use the opposite instruction for INT32_MIN. 4582 if (Idx == 1 && Imm.getBitWidth() == 64 && Imm.getZExtValue() == 0x80000000) 4583 return TTI::TCC_Free; 4584 ImmIdx = 1; 4585 break; 4586 case Instruction::UDiv: 4587 case Instruction::SDiv: 4588 case Instruction::URem: 4589 case Instruction::SRem: 4590 // Division by constant is typically expanded later into a different 4591 // instruction sequence. This completely changes the constants. 4592 // Report them as "free" to stop ConstantHoist from marking them as opaque. 4593 return TTI::TCC_Free; 4594 case Instruction::Mul: 4595 case Instruction::Or: 4596 case Instruction::Xor: 4597 ImmIdx = 1; 4598 break; 4599 // Always return TCC_Free for the shift value of a shift instruction. 4600 case Instruction::Shl: 4601 case Instruction::LShr: 4602 case Instruction::AShr: 4603 if (Idx == 1) 4604 return TTI::TCC_Free; 4605 break; 4606 case Instruction::Trunc: 4607 case Instruction::ZExt: 4608 case Instruction::SExt: 4609 case Instruction::IntToPtr: 4610 case Instruction::PtrToInt: 4611 case Instruction::BitCast: 4612 case Instruction::PHI: 4613 case Instruction::Call: 4614 case Instruction::Select: 4615 case Instruction::Ret: 4616 case Instruction::Load: 4617 break; 4618 } 4619 4620 if (Idx == ImmIdx) { 4621 int NumConstants = divideCeil(BitSize, 64); 4622 InstructionCost Cost = X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 4623 return (Cost <= NumConstants * TTI::TCC_Basic) 4624 ? static_cast<int>(TTI::TCC_Free) 4625 : Cost; 4626 } 4627 4628 return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 4629 } 4630 4631 InstructionCost X86TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, 4632 const APInt &Imm, Type *Ty, 4633 TTI::TargetCostKind CostKind) { 4634 assert(Ty->isIntegerTy()); 4635 4636 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 4637 // There is no cost model for constants with a bit size of 0. Return TCC_Free 4638 // here, so that constant hoisting will ignore this constant. 4639 if (BitSize == 0) 4640 return TTI::TCC_Free; 4641 4642 switch (IID) { 4643 default: 4644 return TTI::TCC_Free; 4645 case Intrinsic::sadd_with_overflow: 4646 case Intrinsic::uadd_with_overflow: 4647 case Intrinsic::ssub_with_overflow: 4648 case Intrinsic::usub_with_overflow: 4649 case Intrinsic::smul_with_overflow: 4650 case Intrinsic::umul_with_overflow: 4651 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue())) 4652 return TTI::TCC_Free; 4653 break; 4654 case Intrinsic::experimental_stackmap: 4655 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 4656 return TTI::TCC_Free; 4657 break; 4658 case Intrinsic::experimental_patchpoint_void: 4659 case Intrinsic::experimental_patchpoint_i64: 4660 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 4661 return TTI::TCC_Free; 4662 break; 4663 } 4664 return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 4665 } 4666 4667 InstructionCost X86TTIImpl::getCFInstrCost(unsigned Opcode, 4668 TTI::TargetCostKind CostKind, 4669 const Instruction *I) { 4670 if (CostKind != TTI::TCK_RecipThroughput) 4671 return Opcode == Instruction::PHI ? 0 : 1; 4672 // Branches are assumed to be predicted. 4673 return 0; 4674 } 4675 4676 int X86TTIImpl::getGatherOverhead() const { 4677 // Some CPUs have more overhead for gather. The specified overhead is relative 4678 // to the Load operation. "2" is the number provided by Intel architects. This 4679 // parameter is used for cost estimation of Gather Op and comparison with 4680 // other alternatives. 4681 // TODO: Remove the explicit hasAVX512()?, That would mean we would only 4682 // enable gather with a -march. 4683 if (ST->hasAVX512() || (ST->hasAVX2() && ST->hasFastGather())) 4684 return 2; 4685 4686 return 1024; 4687 } 4688 4689 int X86TTIImpl::getScatterOverhead() const { 4690 if (ST->hasAVX512()) 4691 return 2; 4692 4693 return 1024; 4694 } 4695 4696 // Return an average cost of Gather / Scatter instruction, maybe improved later. 4697 // FIXME: Add TargetCostKind support. 4698 InstructionCost X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, 4699 const Value *Ptr, Align Alignment, 4700 unsigned AddressSpace) { 4701 4702 assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost"); 4703 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements(); 4704 4705 // Try to reduce index size from 64 bit (default for GEP) 4706 // to 32. It is essential for VF 16. If the index can't be reduced to 32, the 4707 // operation will use 16 x 64 indices which do not fit in a zmm and needs 4708 // to split. Also check that the base pointer is the same for all lanes, 4709 // and that there's at most one variable index. 4710 auto getIndexSizeInBits = [](const Value *Ptr, const DataLayout &DL) { 4711 unsigned IndexSize = DL.getPointerSizeInBits(); 4712 const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr); 4713 if (IndexSize < 64 || !GEP) 4714 return IndexSize; 4715 4716 unsigned NumOfVarIndices = 0; 4717 const Value *Ptrs = GEP->getPointerOperand(); 4718 if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs)) 4719 return IndexSize; 4720 for (unsigned i = 1; i < GEP->getNumOperands(); ++i) { 4721 if (isa<Constant>(GEP->getOperand(i))) 4722 continue; 4723 Type *IndxTy = GEP->getOperand(i)->getType(); 4724 if (auto *IndexVTy = dyn_cast<VectorType>(IndxTy)) 4725 IndxTy = IndexVTy->getElementType(); 4726 if ((IndxTy->getPrimitiveSizeInBits() == 64 && 4727 !isa<SExtInst>(GEP->getOperand(i))) || 4728 ++NumOfVarIndices > 1) 4729 return IndexSize; // 64 4730 } 4731 return (unsigned)32; 4732 }; 4733 4734 // Trying to reduce IndexSize to 32 bits for vector 16. 4735 // By default the IndexSize is equal to pointer size. 4736 unsigned IndexSize = (ST->hasAVX512() && VF >= 16) 4737 ? getIndexSizeInBits(Ptr, DL) 4738 : DL.getPointerSizeInBits(); 4739 4740 auto *IndexVTy = FixedVectorType::get( 4741 IntegerType::get(SrcVTy->getContext(), IndexSize), VF); 4742 std::pair<InstructionCost, MVT> IdxsLT = 4743 TLI->getTypeLegalizationCost(DL, IndexVTy); 4744 std::pair<InstructionCost, MVT> SrcLT = 4745 TLI->getTypeLegalizationCost(DL, SrcVTy); 4746 InstructionCost::CostType SplitFactor = 4747 *std::max(IdxsLT.first, SrcLT.first).getValue(); 4748 if (SplitFactor > 1) { 4749 // Handle splitting of vector of pointers 4750 auto *SplitSrcTy = 4751 FixedVectorType::get(SrcVTy->getScalarType(), VF / SplitFactor); 4752 return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment, 4753 AddressSpace); 4754 } 4755 4756 // The gather / scatter cost is given by Intel architects. It is a rough 4757 // number since we are looking at one instruction in a time. 4758 const int GSOverhead = (Opcode == Instruction::Load) 4759 ? getGatherOverhead() 4760 : getScatterOverhead(); 4761 return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 4762 MaybeAlign(Alignment), AddressSpace, 4763 TTI::TCK_RecipThroughput); 4764 } 4765 4766 /// Return the cost of full scalarization of gather / scatter operation. 4767 /// 4768 /// Opcode - Load or Store instruction. 4769 /// SrcVTy - The type of the data vector that should be gathered or scattered. 4770 /// VariableMask - The mask is non-constant at compile time. 4771 /// Alignment - Alignment for one element. 4772 /// AddressSpace - pointer[s] address space. 4773 /// 4774 /// FIXME: Add TargetCostKind support. 4775 InstructionCost X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy, 4776 bool VariableMask, Align Alignment, 4777 unsigned AddressSpace) { 4778 Type *ScalarTy = SrcVTy->getScalarType(); 4779 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements(); 4780 APInt DemandedElts = APInt::getAllOnes(VF); 4781 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; 4782 4783 InstructionCost MaskUnpackCost = 0; 4784 if (VariableMask) { 4785 auto *MaskTy = 4786 FixedVectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF); 4787 MaskUnpackCost = getScalarizationOverhead( 4788 MaskTy, DemandedElts, /*Insert=*/false, /*Extract=*/true); 4789 InstructionCost ScalarCompareCost = getCmpSelInstrCost( 4790 Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()), nullptr, 4791 CmpInst::BAD_ICMP_PREDICATE, CostKind); 4792 InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind); 4793 MaskUnpackCost += VF * (BranchCost + ScalarCompareCost); 4794 } 4795 4796 InstructionCost AddressUnpackCost = getScalarizationOverhead( 4797 FixedVectorType::get(ScalarTy->getPointerTo(), VF), DemandedElts, 4798 /*Insert=*/false, /*Extract=*/true); 4799 4800 // The cost of the scalar loads/stores. 4801 InstructionCost MemoryOpCost = 4802 VF * getMemoryOpCost(Opcode, ScalarTy, MaybeAlign(Alignment), 4803 AddressSpace, CostKind); 4804 4805 // The cost of forming the vector from loaded scalars/ 4806 // scalarizing the vector to perform scalar stores. 4807 InstructionCost InsertExtractCost = 4808 getScalarizationOverhead(cast<FixedVectorType>(SrcVTy), DemandedElts, 4809 /*Insert=*/Opcode == Instruction::Load, 4810 /*Extract=*/Opcode == Instruction::Store); 4811 4812 return AddressUnpackCost + MemoryOpCost + MaskUnpackCost + InsertExtractCost; 4813 } 4814 4815 /// Calculate the cost of Gather / Scatter operation 4816 InstructionCost X86TTIImpl::getGatherScatterOpCost( 4817 unsigned Opcode, Type *SrcVTy, const Value *Ptr, bool VariableMask, 4818 Align Alignment, TTI::TargetCostKind CostKind, 4819 const Instruction *I = nullptr) { 4820 if (CostKind != TTI::TCK_RecipThroughput) { 4821 if ((Opcode == Instruction::Load && 4822 isLegalMaskedGather(SrcVTy, Align(Alignment))) || 4823 (Opcode == Instruction::Store && 4824 isLegalMaskedScatter(SrcVTy, Align(Alignment)))) 4825 return 1; 4826 return BaseT::getGatherScatterOpCost(Opcode, SrcVTy, Ptr, VariableMask, 4827 Alignment, CostKind, I); 4828 } 4829 4830 assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter"); 4831 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType()); 4832 if (!PtrTy && Ptr->getType()->isVectorTy()) 4833 PtrTy = dyn_cast<PointerType>( 4834 cast<VectorType>(Ptr->getType())->getElementType()); 4835 assert(PtrTy && "Unexpected type for Ptr argument"); 4836 unsigned AddressSpace = PtrTy->getAddressSpace(); 4837 4838 if ((Opcode == Instruction::Load && 4839 !isLegalMaskedGather(SrcVTy, Align(Alignment))) || 4840 (Opcode == Instruction::Store && 4841 !isLegalMaskedScatter(SrcVTy, Align(Alignment)))) 4842 return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment, 4843 AddressSpace); 4844 4845 return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace); 4846 } 4847 4848 bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1, 4849 TargetTransformInfo::LSRCost &C2) { 4850 // X86 specific here are "instruction number 1st priority". 4851 return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost, 4852 C1.NumIVMuls, C1.NumBaseAdds, 4853 C1.ScaleCost, C1.ImmCost, C1.SetupCost) < 4854 std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost, 4855 C2.NumIVMuls, C2.NumBaseAdds, 4856 C2.ScaleCost, C2.ImmCost, C2.SetupCost); 4857 } 4858 4859 bool X86TTIImpl::canMacroFuseCmp() { 4860 return ST->hasMacroFusion() || ST->hasBranchFusion(); 4861 } 4862 4863 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy, Align Alignment) { 4864 if (!ST->hasAVX()) 4865 return false; 4866 4867 // The backend can't handle a single element vector. 4868 if (isa<VectorType>(DataTy) && 4869 cast<FixedVectorType>(DataTy)->getNumElements() == 1) 4870 return false; 4871 Type *ScalarTy = DataTy->getScalarType(); 4872 4873 if (ScalarTy->isPointerTy()) 4874 return true; 4875 4876 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4877 return true; 4878 4879 if (ScalarTy->isHalfTy() && ST->hasBWI() && ST->hasFP16()) 4880 return true; 4881 4882 if (!ScalarTy->isIntegerTy()) 4883 return false; 4884 4885 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 4886 return IntWidth == 32 || IntWidth == 64 || 4887 ((IntWidth == 8 || IntWidth == 16) && ST->hasBWI()); 4888 } 4889 4890 bool X86TTIImpl::isLegalMaskedStore(Type *DataType, Align Alignment) { 4891 return isLegalMaskedLoad(DataType, Alignment); 4892 } 4893 4894 bool X86TTIImpl::isLegalNTLoad(Type *DataType, Align Alignment) { 4895 unsigned DataSize = DL.getTypeStoreSize(DataType); 4896 // The only supported nontemporal loads are for aligned vectors of 16 or 32 4897 // bytes. Note that 32-byte nontemporal vector loads are supported by AVX2 4898 // (the equivalent stores only require AVX). 4899 if (Alignment >= DataSize && (DataSize == 16 || DataSize == 32)) 4900 return DataSize == 16 ? ST->hasSSE1() : ST->hasAVX2(); 4901 4902 return false; 4903 } 4904 4905 bool X86TTIImpl::isLegalNTStore(Type *DataType, Align Alignment) { 4906 unsigned DataSize = DL.getTypeStoreSize(DataType); 4907 4908 // SSE4A supports nontemporal stores of float and double at arbitrary 4909 // alignment. 4910 if (ST->hasSSE4A() && (DataType->isFloatTy() || DataType->isDoubleTy())) 4911 return true; 4912 4913 // Besides the SSE4A subtarget exception above, only aligned stores are 4914 // available nontemporaly on any other subtarget. And only stores with a size 4915 // of 4..32 bytes (powers of 2, only) are permitted. 4916 if (Alignment < DataSize || DataSize < 4 || DataSize > 32 || 4917 !isPowerOf2_32(DataSize)) 4918 return false; 4919 4920 // 32-byte vector nontemporal stores are supported by AVX (the equivalent 4921 // loads require AVX2). 4922 if (DataSize == 32) 4923 return ST->hasAVX(); 4924 if (DataSize == 16) 4925 return ST->hasSSE1(); 4926 return true; 4927 } 4928 4929 bool X86TTIImpl::isLegalMaskedExpandLoad(Type *DataTy) { 4930 if (!isa<VectorType>(DataTy)) 4931 return false; 4932 4933 if (!ST->hasAVX512()) 4934 return false; 4935 4936 // The backend can't handle a single element vector. 4937 if (cast<FixedVectorType>(DataTy)->getNumElements() == 1) 4938 return false; 4939 4940 Type *ScalarTy = cast<VectorType>(DataTy)->getElementType(); 4941 4942 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4943 return true; 4944 4945 if (!ScalarTy->isIntegerTy()) 4946 return false; 4947 4948 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 4949 return IntWidth == 32 || IntWidth == 64 || 4950 ((IntWidth == 8 || IntWidth == 16) && ST->hasVBMI2()); 4951 } 4952 4953 bool X86TTIImpl::isLegalMaskedCompressStore(Type *DataTy) { 4954 return isLegalMaskedExpandLoad(DataTy); 4955 } 4956 4957 bool X86TTIImpl::supportsGather() const { 4958 // Some CPUs have better gather performance than others. 4959 // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only 4960 // enable gather with a -march. 4961 return ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2()); 4962 } 4963 4964 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy, Align Alignment) { 4965 if (!supportsGather()) 4966 return false; 4967 4968 // This function is called now in two cases: from the Loop Vectorizer 4969 // and from the Scalarizer. 4970 // When the Loop Vectorizer asks about legality of the feature, 4971 // the vectorization factor is not calculated yet. The Loop Vectorizer 4972 // sends a scalar type and the decision is based on the width of the 4973 // scalar element. 4974 // Later on, the cost model will estimate usage this intrinsic based on 4975 // the vector type. 4976 // The Scalarizer asks again about legality. It sends a vector type. 4977 // In this case we can reject non-power-of-2 vectors. 4978 // We also reject single element vectors as the type legalizer can't 4979 // scalarize it. 4980 if (auto *DataVTy = dyn_cast<FixedVectorType>(DataTy)) { 4981 unsigned NumElts = DataVTy->getNumElements(); 4982 if (NumElts == 1) 4983 return false; 4984 // Gather / Scatter for vector 2 is not profitable on KNL / SKX 4985 // Vector-4 of gather/scatter instruction does not exist on KNL. 4986 // We can extend it to 8 elements, but zeroing upper bits of 4987 // the mask vector will add more instructions. Right now we give the scalar 4988 // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter 4989 // instruction is better in the VariableMask case. 4990 if (ST->hasAVX512() && (NumElts == 2 || (NumElts == 4 && !ST->hasVLX()))) 4991 return false; 4992 } 4993 Type *ScalarTy = DataTy->getScalarType(); 4994 if (ScalarTy->isPointerTy()) 4995 return true; 4996 4997 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4998 return true; 4999 5000 if (!ScalarTy->isIntegerTy()) 5001 return false; 5002 5003 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 5004 return IntWidth == 32 || IntWidth == 64; 5005 } 5006 5007 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType, Align Alignment) { 5008 // AVX2 doesn't support scatter 5009 if (!ST->hasAVX512()) 5010 return false; 5011 return isLegalMaskedGather(DataType, Alignment); 5012 } 5013 5014 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) { 5015 EVT VT = TLI->getValueType(DL, DataType); 5016 return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT); 5017 } 5018 5019 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) { 5020 return false; 5021 } 5022 5023 bool X86TTIImpl::areInlineCompatible(const Function *Caller, 5024 const Function *Callee) const { 5025 const TargetMachine &TM = getTLI()->getTargetMachine(); 5026 5027 // Work this as a subsetting of subtarget features. 5028 const FeatureBitset &CallerBits = 5029 TM.getSubtargetImpl(*Caller)->getFeatureBits(); 5030 const FeatureBitset &CalleeBits = 5031 TM.getSubtargetImpl(*Callee)->getFeatureBits(); 5032 5033 FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList; 5034 FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList; 5035 return (RealCallerBits & RealCalleeBits) == RealCalleeBits; 5036 } 5037 5038 bool X86TTIImpl::areFunctionArgsABICompatible( 5039 const Function *Caller, const Function *Callee, 5040 SmallPtrSetImpl<Argument *> &Args) const { 5041 if (!BaseT::areFunctionArgsABICompatible(Caller, Callee, Args)) 5042 return false; 5043 5044 // If we get here, we know the target features match. If one function 5045 // considers 512-bit vectors legal and the other does not, consider them 5046 // incompatible. 5047 const TargetMachine &TM = getTLI()->getTargetMachine(); 5048 5049 if (TM.getSubtarget<X86Subtarget>(*Caller).useAVX512Regs() == 5050 TM.getSubtarget<X86Subtarget>(*Callee).useAVX512Regs()) 5051 return true; 5052 5053 // Consider the arguments compatible if they aren't vectors or aggregates. 5054 // FIXME: Look at the size of vectors. 5055 // FIXME: Look at the element types of aggregates to see if there are vectors. 5056 // FIXME: The API of this function seems intended to allow arguments 5057 // to be removed from the set, but the caller doesn't check if the set 5058 // becomes empty so that may not work in practice. 5059 return llvm::none_of(Args, [](Argument *A) { 5060 auto *EltTy = cast<PointerType>(A->getType())->getElementType(); 5061 return EltTy->isVectorTy() || EltTy->isAggregateType(); 5062 }); 5063 } 5064 5065 X86TTIImpl::TTI::MemCmpExpansionOptions 5066 X86TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { 5067 TTI::MemCmpExpansionOptions Options; 5068 Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize); 5069 Options.NumLoadsPerBlock = 2; 5070 // All GPR and vector loads can be unaligned. 5071 Options.AllowOverlappingLoads = true; 5072 if (IsZeroCmp) { 5073 // Only enable vector loads for equality comparison. Right now the vector 5074 // version is not as fast for three way compare (see #33329). 5075 const unsigned PreferredWidth = ST->getPreferVectorWidth(); 5076 if (PreferredWidth >= 512 && ST->hasAVX512()) Options.LoadSizes.push_back(64); 5077 if (PreferredWidth >= 256 && ST->hasAVX()) Options.LoadSizes.push_back(32); 5078 if (PreferredWidth >= 128 && ST->hasSSE2()) Options.LoadSizes.push_back(16); 5079 } 5080 if (ST->is64Bit()) { 5081 Options.LoadSizes.push_back(8); 5082 } 5083 Options.LoadSizes.push_back(4); 5084 Options.LoadSizes.push_back(2); 5085 Options.LoadSizes.push_back(1); 5086 return Options; 5087 } 5088 5089 bool X86TTIImpl::prefersVectorizedAddressing() const { 5090 return supportsGather(); 5091 } 5092 5093 bool X86TTIImpl::supportsEfficientVectorElementLoadStore() const { 5094 return false; 5095 } 5096 5097 bool X86TTIImpl::enableInterleavedAccessVectorization() { 5098 // TODO: We expect this to be beneficial regardless of arch, 5099 // but there are currently some unexplained performance artifacts on Atom. 5100 // As a temporary solution, disable on Atom. 5101 return !(ST->isAtom()); 5102 } 5103 5104 // Get estimation for interleaved load/store operations and strided load. 5105 // \p Indices contains indices for strided load. 5106 // \p Factor - the factor of interleaving. 5107 // AVX-512 provides 3-src shuffles that significantly reduces the cost. 5108 InstructionCost X86TTIImpl::getInterleavedMemoryOpCostAVX512( 5109 unsigned Opcode, FixedVectorType *VecTy, unsigned Factor, 5110 ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace, 5111 TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) { 5112 // VecTy for interleave memop is <VF*Factor x Elt>. 5113 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have 5114 // VecTy = <12 x i32>. 5115 5116 // Calculate the number of memory operations (NumOfMemOps), required 5117 // for load/store the VecTy. 5118 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second; 5119 unsigned VecTySize = DL.getTypeStoreSize(VecTy); 5120 unsigned LegalVTSize = LegalVT.getStoreSize(); 5121 unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize; 5122 5123 // Get the cost of one memory operation. 5124 auto *SingleMemOpTy = FixedVectorType::get(VecTy->getElementType(), 5125 LegalVT.getVectorNumElements()); 5126 InstructionCost MemOpCost; 5127 if (UseMaskForCond || UseMaskForGaps) 5128 MemOpCost = getMaskedMemoryOpCost(Opcode, SingleMemOpTy, Alignment, 5129 AddressSpace, CostKind); 5130 else 5131 MemOpCost = getMemoryOpCost(Opcode, SingleMemOpTy, MaybeAlign(Alignment), 5132 AddressSpace, CostKind); 5133 5134 unsigned VF = VecTy->getNumElements() / Factor; 5135 MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF); 5136 5137 // FIXME: this is the most conservative estimate for the mask cost. 5138 InstructionCost MaskCost; 5139 if (UseMaskForCond || UseMaskForGaps) { 5140 APInt DemandedLoadStoreElts = APInt::getZero(VecTy->getNumElements()); 5141 for (unsigned Index : Indices) { 5142 assert(Index < Factor && "Invalid index for interleaved memory op"); 5143 for (unsigned Elm = 0; Elm < VF; Elm++) 5144 DemandedLoadStoreElts.setBit(Index + Elm * Factor); 5145 } 5146 5147 Type *I8Type = Type::getInt8Ty(VecTy->getContext()); 5148 5149 MaskCost = getReplicationShuffleCost( 5150 I8Type, Factor, VF, 5151 UseMaskForGaps ? DemandedLoadStoreElts 5152 : APInt::getAllOnes(VecTy->getNumElements()), 5153 CostKind); 5154 5155 // The Gaps mask is invariant and created outside the loop, therefore the 5156 // cost of creating it is not accounted for here. However if we have both 5157 // a MaskForGaps and some other mask that guards the execution of the 5158 // memory access, we need to account for the cost of And-ing the two masks 5159 // inside the loop. 5160 if (UseMaskForGaps) { 5161 auto *MaskVT = FixedVectorType::get(I8Type, VecTy->getNumElements()); 5162 MaskCost += getArithmeticInstrCost(BinaryOperator::And, MaskVT, CostKind); 5163 } 5164 } 5165 5166 if (Opcode == Instruction::Load) { 5167 // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl) 5168 // contain the cost of the optimized shuffle sequence that the 5169 // X86InterleavedAccess pass will generate. 5170 // The cost of loads and stores are computed separately from the table. 5171 5172 // X86InterleavedAccess support only the following interleaved-access group. 5173 static const CostTblEntry AVX512InterleavedLoadTbl[] = { 5174 {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8 5175 {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8 5176 {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8 5177 }; 5178 5179 if (const auto *Entry = 5180 CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT)) 5181 return MaskCost + NumOfMemOps * MemOpCost + Entry->Cost; 5182 //If an entry does not exist, fallback to the default implementation. 5183 5184 // Kind of shuffle depends on number of loaded values. 5185 // If we load the entire data in one register, we can use a 1-src shuffle. 5186 // Otherwise, we'll merge 2 sources in each operation. 5187 TTI::ShuffleKind ShuffleKind = 5188 (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc; 5189 5190 InstructionCost ShuffleCost = 5191 getShuffleCost(ShuffleKind, SingleMemOpTy, None, 0, nullptr); 5192 5193 unsigned NumOfLoadsInInterleaveGrp = 5194 Indices.size() ? Indices.size() : Factor; 5195 auto *ResultTy = FixedVectorType::get(VecTy->getElementType(), 5196 VecTy->getNumElements() / Factor); 5197 InstructionCost NumOfResults = 5198 getTLI()->getTypeLegalizationCost(DL, ResultTy).first * 5199 NumOfLoadsInInterleaveGrp; 5200 5201 // About a half of the loads may be folded in shuffles when we have only 5202 // one result. If we have more than one result, we do not fold loads at all. 5203 unsigned NumOfUnfoldedLoads = 5204 NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2; 5205 5206 // Get a number of shuffle operations per result. 5207 unsigned NumOfShufflesPerResult = 5208 std::max((unsigned)1, (unsigned)(NumOfMemOps - 1)); 5209 5210 // The SK_MergeTwoSrc shuffle clobbers one of src operands. 5211 // When we have more than one destination, we need additional instructions 5212 // to keep sources. 5213 InstructionCost NumOfMoves = 0; 5214 if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc) 5215 NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2; 5216 5217 InstructionCost Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost + 5218 MaskCost + NumOfUnfoldedLoads * MemOpCost + 5219 NumOfMoves; 5220 5221 return Cost; 5222 } 5223 5224 // Store. 5225 assert(Opcode == Instruction::Store && 5226 "Expected Store Instruction at this point"); 5227 // X86InterleavedAccess support only the following interleaved-access group. 5228 static const CostTblEntry AVX512InterleavedStoreTbl[] = { 5229 {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store) 5230 {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store) 5231 {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store) 5232 5233 {4, MVT::v8i8, 10}, // interleave 4 x 8i8 into 32i8 (and store) 5234 {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8 (and store) 5235 {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store) 5236 {4, MVT::v64i8, 24} // interleave 4 x 32i8 into 256i8 (and store) 5237 }; 5238 5239 if (const auto *Entry = 5240 CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT)) 5241 return MaskCost + NumOfMemOps * MemOpCost + Entry->Cost; 5242 //If an entry does not exist, fallback to the default implementation. 5243 5244 // There is no strided stores meanwhile. And store can't be folded in 5245 // shuffle. 5246 unsigned NumOfSources = Factor; // The number of values to be merged. 5247 InstructionCost ShuffleCost = 5248 getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, None, 0, nullptr); 5249 unsigned NumOfShufflesPerStore = NumOfSources - 1; 5250 5251 // The SK_MergeTwoSrc shuffle clobbers one of src operands. 5252 // We need additional instructions to keep sources. 5253 unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2; 5254 InstructionCost Cost = 5255 MaskCost + 5256 NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) + 5257 NumOfMoves; 5258 return Cost; 5259 } 5260 5261 InstructionCost X86TTIImpl::getInterleavedMemoryOpCost( 5262 unsigned Opcode, Type *BaseTy, unsigned Factor, ArrayRef<unsigned> Indices, 5263 Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, 5264 bool UseMaskForCond, bool UseMaskForGaps) { 5265 auto *VecTy = cast<FixedVectorType>(BaseTy); 5266 5267 auto isSupportedOnAVX512 = [&](Type *VecTy, bool HasBW) { 5268 Type *EltTy = cast<VectorType>(VecTy)->getElementType(); 5269 if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) || 5270 EltTy->isIntegerTy(32) || EltTy->isPointerTy()) 5271 return true; 5272 if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8) || 5273 (!ST->useSoftFloat() && ST->hasFP16() && EltTy->isHalfTy())) 5274 return HasBW; 5275 return false; 5276 }; 5277 if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI())) 5278 return getInterleavedMemoryOpCostAVX512( 5279 Opcode, VecTy, Factor, Indices, Alignment, 5280 AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps); 5281 5282 if (UseMaskForCond || UseMaskForGaps) 5283 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 5284 Alignment, AddressSpace, CostKind, 5285 UseMaskForCond, UseMaskForGaps); 5286 5287 // Get estimation for interleaved load/store operations for SSE-AVX2. 5288 // As opposed to AVX-512, SSE-AVX2 do not have generic shuffles that allow 5289 // computing the cost using a generic formula as a function of generic 5290 // shuffles. We therefore use a lookup table instead, filled according to 5291 // the instruction sequences that codegen currently generates. 5292 5293 // VecTy for interleave memop is <VF*Factor x Elt>. 5294 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have 5295 // VecTy = <12 x i32>. 5296 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second; 5297 5298 // This function can be called with VecTy=<6xi128>, Factor=3, in which case 5299 // the VF=2, while v2i128 is an unsupported MVT vector type 5300 // (see MachineValueType.h::getVectorVT()). 5301 if (!LegalVT.isVector()) 5302 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 5303 Alignment, AddressSpace, CostKind); 5304 5305 unsigned VF = VecTy->getNumElements() / Factor; 5306 Type *ScalarTy = VecTy->getElementType(); 5307 // Deduplicate entries, model floats/pointers as appropriately-sized integers. 5308 if (!ScalarTy->isIntegerTy()) 5309 ScalarTy = 5310 Type::getIntNTy(ScalarTy->getContext(), DL.getTypeSizeInBits(ScalarTy)); 5311 5312 // Get the cost of all the memory operations. 5313 // FIXME: discount dead loads. 5314 InstructionCost MemOpCosts = getMemoryOpCost( 5315 Opcode, VecTy, MaybeAlign(Alignment), AddressSpace, CostKind); 5316 5317 auto *VT = FixedVectorType::get(ScalarTy, VF); 5318 EVT ETy = TLI->getValueType(DL, VT); 5319 if (!ETy.isSimple()) 5320 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 5321 Alignment, AddressSpace, CostKind); 5322 5323 // TODO: Complete for other data-types and strides. 5324 // Each combination of Stride, element bit width and VF results in a different 5325 // sequence; The cost tables are therefore accessed with: 5326 // Factor (stride) and VectorType=VFxiN. 5327 // The Cost accounts only for the shuffle sequence; 5328 // The cost of the loads/stores is accounted for separately. 5329 // 5330 static const CostTblEntry AVX2InterleavedLoadTbl[] = { 5331 {2, MVT::v2i8, 2}, // (load 4i8 and) deinterleave into 2 x 2i8 5332 {2, MVT::v4i8, 2}, // (load 8i8 and) deinterleave into 2 x 4i8 5333 {2, MVT::v8i8, 2}, // (load 16i8 and) deinterleave into 2 x 8i8 5334 {2, MVT::v16i8, 4}, // (load 32i8 and) deinterleave into 2 x 16i8 5335 {2, MVT::v32i8, 6}, // (load 64i8 and) deinterleave into 2 x 32i8 5336 5337 {2, MVT::v8i16, 6}, // (load 16i16 and) deinterleave into 2 x 8i16 5338 {2, MVT::v16i16, 9}, // (load 32i16 and) deinterleave into 2 x 16i16 5339 {2, MVT::v32i16, 18}, // (load 64i16 and) deinterleave into 2 x 32i16 5340 5341 {2, MVT::v8i32, 4}, // (load 16i32 and) deinterleave into 2 x 8i32 5342 {2, MVT::v16i32, 8}, // (load 32i32 and) deinterleave into 2 x 16i32 5343 {2, MVT::v32i32, 16}, // (load 64i32 and) deinterleave into 2 x 32i32 5344 5345 {2, MVT::v4i64, 4}, // (load 8i64 and) deinterleave into 2 x 4i64 5346 {2, MVT::v8i64, 8}, // (load 16i64 and) deinterleave into 2 x 8i64 5347 {2, MVT::v16i64, 16}, // (load 32i64 and) deinterleave into 2 x 16i64 5348 {2, MVT::v32i64, 32}, // (load 64i64 and) deinterleave into 2 x 32i64 5349 5350 {3, MVT::v2i8, 3}, // (load 6i8 and) deinterleave into 3 x 2i8 5351 {3, MVT::v4i8, 3}, // (load 12i8 and) deinterleave into 3 x 4i8 5352 {3, MVT::v8i8, 6}, // (load 24i8 and) deinterleave into 3 x 8i8 5353 {3, MVT::v16i8, 11}, // (load 48i8 and) deinterleave into 3 x 16i8 5354 {3, MVT::v32i8, 14}, // (load 96i8 and) deinterleave into 3 x 32i8 5355 5356 {3, MVT::v2i16, 5}, // (load 6i16 and) deinterleave into 3 x 2i16 5357 {3, MVT::v4i16, 7}, // (load 12i16 and) deinterleave into 3 x 4i16 5358 {3, MVT::v8i16, 9}, // (load 24i16 and) deinterleave into 3 x 8i16 5359 {3, MVT::v16i16, 28}, // (load 48i16 and) deinterleave into 3 x 16i16 5360 {3, MVT::v32i16, 56}, // (load 96i16 and) deinterleave into 3 x 32i16 5361 5362 {3, MVT::v2i32, 3}, // (load 6i32 and) deinterleave into 3 x 2i32 5363 {3, MVT::v4i32, 3}, // (load 12i32 and) deinterleave into 3 x 4i32 5364 {3, MVT::v8i32, 7}, // (load 24i32 and) deinterleave into 3 x 8i32 5365 {3, MVT::v16i32, 14}, // (load 48i32 and) deinterleave into 3 x 16i32 5366 {3, MVT::v32i32, 32}, // (load 96i32 and) deinterleave into 3 x 32i32 5367 5368 {3, MVT::v2i64, 1}, // (load 6i64 and) deinterleave into 3 x 2i64 5369 {3, MVT::v4i64, 5}, // (load 12i64 and) deinterleave into 3 x 4i64 5370 {3, MVT::v8i64, 10}, // (load 24i64 and) deinterleave into 3 x 8i64 5371 {3, MVT::v16i64, 20}, // (load 48i64 and) deinterleave into 3 x 16i64 5372 5373 {4, MVT::v2i8, 4}, // (load 8i8 and) deinterleave into 4 x 2i8 5374 {4, MVT::v4i8, 4}, // (load 16i8 and) deinterleave into 4 x 4i8 5375 {4, MVT::v8i8, 12}, // (load 32i8 and) deinterleave into 4 x 8i8 5376 {4, MVT::v16i8, 24}, // (load 64i8 and) deinterleave into 4 x 16i8 5377 {4, MVT::v32i8, 56}, // (load 128i8 and) deinterleave into 4 x 32i8 5378 5379 {4, MVT::v2i16, 6}, // (load 8i16 and) deinterleave into 4 x 2i16 5380 {4, MVT::v4i16, 17}, // (load 16i16 and) deinterleave into 4 x 4i16 5381 {4, MVT::v8i16, 33}, // (load 32i16 and) deinterleave into 4 x 8i16 5382 {4, MVT::v16i16, 75}, // (load 64i16 and) deinterleave into 4 x 16i16 5383 {4, MVT::v32i16, 150}, // (load 128i16 and) deinterleave into 4 x 32i16 5384 5385 {4, MVT::v2i32, 4}, // (load 8i32 and) deinterleave into 4 x 2i32 5386 {4, MVT::v4i32, 8}, // (load 16i32 and) deinterleave into 4 x 4i32 5387 {4, MVT::v8i32, 16}, // (load 32i32 and) deinterleave into 4 x 8i32 5388 {4, MVT::v16i32, 32}, // (load 64i32 and) deinterleave into 4 x 16i32 5389 {4, MVT::v32i32, 68}, // (load 128i32 and) deinterleave into 4 x 32i32 5390 5391 {4, MVT::v2i64, 6}, // (load 8i64 and) deinterleave into 4 x 2i64 5392 {4, MVT::v4i64, 8}, // (load 16i64 and) deinterleave into 4 x 4i64 5393 {4, MVT::v8i64, 20}, // (load 32i64 and) deinterleave into 4 x 8i64 5394 {4, MVT::v16i64, 40}, // (load 64i64 and) deinterleave into 4 x 16i64 5395 5396 {6, MVT::v2i8, 6}, // (load 12i8 and) deinterleave into 6 x 2i8 5397 {6, MVT::v4i8, 14}, // (load 24i8 and) deinterleave into 6 x 4i8 5398 {6, MVT::v8i8, 18}, // (load 48i8 and) deinterleave into 6 x 8i8 5399 {6, MVT::v16i8, 43}, // (load 96i8 and) deinterleave into 6 x 16i8 5400 {6, MVT::v32i8, 82}, // (load 192i8 and) deinterleave into 6 x 32i8 5401 5402 {6, MVT::v2i16, 13}, // (load 12i16 and) deinterleave into 6 x 2i16 5403 {6, MVT::v4i16, 9}, // (load 24i16 and) deinterleave into 6 x 4i16 5404 {6, MVT::v8i16, 39}, // (load 48i16 and) deinterleave into 6 x 8i16 5405 {6, MVT::v16i16, 106}, // (load 96i16 and) deinterleave into 6 x 16i16 5406 {6, MVT::v32i16, 212}, // (load 192i16 and) deinterleave into 6 x 32i16 5407 5408 {6, MVT::v2i32, 6}, // (load 12i32 and) deinterleave into 6 x 2i32 5409 {6, MVT::v4i32, 15}, // (load 24i32 and) deinterleave into 6 x 4i32 5410 {6, MVT::v8i32, 31}, // (load 48i32 and) deinterleave into 6 x 8i32 5411 {6, MVT::v16i32, 64}, // (load 96i32 and) deinterleave into 6 x 16i32 5412 5413 {6, MVT::v2i64, 6}, // (load 12i64 and) deinterleave into 6 x 2i64 5414 {6, MVT::v4i64, 18}, // (load 24i64 and) deinterleave into 6 x 4i64 5415 {6, MVT::v8i64, 36}, // (load 48i64 and) deinterleave into 6 x 8i64 5416 5417 {8, MVT::v8i32, 40} // (load 64i32 and) deinterleave into 8 x 8i32 5418 }; 5419 5420 static const CostTblEntry SSSE3InterleavedLoadTbl[] = { 5421 {2, MVT::v4i16, 2}, // (load 8i16 and) deinterleave into 2 x 4i16 5422 }; 5423 5424 static const CostTblEntry SSE2InterleavedLoadTbl[] = { 5425 {2, MVT::v2i16, 2}, // (load 4i16 and) deinterleave into 2 x 2i16 5426 {2, MVT::v4i16, 7}, // (load 8i16 and) deinterleave into 2 x 4i16 5427 5428 {2, MVT::v2i32, 2}, // (load 4i32 and) deinterleave into 2 x 2i32 5429 {2, MVT::v4i32, 2}, // (load 8i32 and) deinterleave into 2 x 4i32 5430 5431 {2, MVT::v2i64, 2}, // (load 4i64 and) deinterleave into 2 x 2i64 5432 }; 5433 5434 static const CostTblEntry AVX2InterleavedStoreTbl[] = { 5435 {2, MVT::v16i8, 3}, // interleave 2 x 16i8 into 32i8 (and store) 5436 {2, MVT::v32i8, 4}, // interleave 2 x 32i8 into 64i8 (and store) 5437 5438 {2, MVT::v8i16, 3}, // interleave 2 x 8i16 into 16i16 (and store) 5439 {2, MVT::v16i16, 4}, // interleave 2 x 16i16 into 32i16 (and store) 5440 {2, MVT::v32i16, 8}, // interleave 2 x 32i16 into 64i16 (and store) 5441 5442 {2, MVT::v4i32, 2}, // interleave 2 x 4i32 into 8i32 (and store) 5443 {2, MVT::v8i32, 4}, // interleave 2 x 8i32 into 16i32 (and store) 5444 {2, MVT::v16i32, 8}, // interleave 2 x 16i32 into 32i32 (and store) 5445 {2, MVT::v32i32, 16}, // interleave 2 x 32i32 into 64i32 (and store) 5446 5447 {2, MVT::v2i64, 2}, // interleave 2 x 2i64 into 4i64 (and store) 5448 {2, MVT::v4i64, 4}, // interleave 2 x 4i64 into 8i64 (and store) 5449 {2, MVT::v8i64, 8}, // interleave 2 x 8i64 into 16i64 (and store) 5450 {2, MVT::v16i64, 16}, // interleave 2 x 16i64 into 32i64 (and store) 5451 {2, MVT::v32i64, 32}, // interleave 2 x 32i64 into 64i64 (and store) 5452 5453 {3, MVT::v2i8, 4}, // interleave 3 x 2i8 into 6i8 (and store) 5454 {3, MVT::v4i8, 4}, // interleave 3 x 4i8 into 12i8 (and store) 5455 {3, MVT::v8i8, 6}, // interleave 3 x 8i8 into 24i8 (and store) 5456 {3, MVT::v16i8, 11}, // interleave 3 x 16i8 into 48i8 (and store) 5457 {3, MVT::v32i8, 13}, // interleave 3 x 32i8 into 96i8 (and store) 5458 5459 {3, MVT::v2i16, 4}, // interleave 3 x 2i16 into 6i16 (and store) 5460 {3, MVT::v4i16, 6}, // interleave 3 x 4i16 into 12i16 (and store) 5461 {3, MVT::v8i16, 12}, // interleave 3 x 8i16 into 24i16 (and store) 5462 {3, MVT::v16i16, 27}, // interleave 3 x 16i16 into 48i16 (and store) 5463 {3, MVT::v32i16, 54}, // interleave 3 x 32i16 into 96i16 (and store) 5464 5465 {3, MVT::v2i32, 4}, // interleave 3 x 2i32 into 6i32 (and store) 5466 {3, MVT::v4i32, 5}, // interleave 3 x 4i32 into 12i32 (and store) 5467 {3, MVT::v8i32, 11}, // interleave 3 x 8i32 into 24i32 (and store) 5468 {3, MVT::v16i32, 22}, // interleave 3 x 16i32 into 48i32 (and store) 5469 {3, MVT::v32i32, 48}, // interleave 3 x 32i32 into 96i32 (and store) 5470 5471 {3, MVT::v2i64, 4}, // interleave 3 x 2i64 into 6i64 (and store) 5472 {3, MVT::v4i64, 6}, // interleave 3 x 4i64 into 12i64 (and store) 5473 {3, MVT::v8i64, 12}, // interleave 3 x 8i64 into 24i64 (and store) 5474 {3, MVT::v16i64, 24}, // interleave 3 x 16i64 into 48i64 (and store) 5475 5476 {4, MVT::v2i8, 4}, // interleave 4 x 2i8 into 8i8 (and store) 5477 {4, MVT::v4i8, 4}, // interleave 4 x 4i8 into 16i8 (and store) 5478 {4, MVT::v8i8, 4}, // interleave 4 x 8i8 into 32i8 (and store) 5479 {4, MVT::v16i8, 8}, // interleave 4 x 16i8 into 64i8 (and store) 5480 {4, MVT::v32i8, 12}, // interleave 4 x 32i8 into 128i8 (and store) 5481 5482 {4, MVT::v2i16, 2}, // interleave 4 x 2i16 into 8i16 (and store) 5483 {4, MVT::v4i16, 6}, // interleave 4 x 4i16 into 16i16 (and store) 5484 {4, MVT::v8i16, 10}, // interleave 4 x 8i16 into 32i16 (and store) 5485 {4, MVT::v16i16, 32}, // interleave 4 x 16i16 into 64i16 (and store) 5486 {4, MVT::v32i16, 64}, // interleave 4 x 32i16 into 128i16 (and store) 5487 5488 {4, MVT::v2i32, 5}, // interleave 4 x 2i32 into 8i32 (and store) 5489 {4, MVT::v4i32, 6}, // interleave 4 x 4i32 into 16i32 (and store) 5490 {4, MVT::v8i32, 16}, // interleave 4 x 8i32 into 32i32 (and store) 5491 {4, MVT::v16i32, 32}, // interleave 4 x 16i32 into 64i32 (and store) 5492 {4, MVT::v32i32, 64}, // interleave 4 x 32i32 into 128i32 (and store) 5493 5494 {4, MVT::v2i64, 6}, // interleave 4 x 2i64 into 8i64 (and store) 5495 {4, MVT::v4i64, 8}, // interleave 4 x 4i64 into 16i64 (and store) 5496 {4, MVT::v8i64, 20}, // interleave 4 x 8i64 into 32i64 (and store) 5497 {4, MVT::v16i64, 40}, // interleave 4 x 16i64 into 64i64 (and store) 5498 5499 {6, MVT::v2i8, 7}, // interleave 6 x 2i8 into 12i8 (and store) 5500 {6, MVT::v4i8, 9}, // interleave 6 x 4i8 into 24i8 (and store) 5501 {6, MVT::v8i8, 16}, // interleave 6 x 8i8 into 48i8 (and store) 5502 {6, MVT::v16i8, 27}, // interleave 6 x 16i8 into 96i8 (and store) 5503 {6, MVT::v32i8, 90}, // interleave 6 x 32i8 into 192i8 (and store) 5504 5505 {6, MVT::v2i16, 10}, // interleave 6 x 2i16 into 12i16 (and store) 5506 {6, MVT::v4i16, 15}, // interleave 6 x 4i16 into 24i16 (and store) 5507 {6, MVT::v8i16, 21}, // interleave 6 x 8i16 into 48i16 (and store) 5508 {6, MVT::v16i16, 58}, // interleave 6 x 16i16 into 96i16 (and store) 5509 {6, MVT::v32i16, 90}, // interleave 6 x 32i16 into 192i16 (and store) 5510 5511 {6, MVT::v2i32, 9}, // interleave 6 x 2i32 into 12i32 (and store) 5512 {6, MVT::v4i32, 12}, // interleave 6 x 4i32 into 24i32 (and store) 5513 {6, MVT::v8i32, 33}, // interleave 6 x 8i32 into 48i32 (and store) 5514 {6, MVT::v16i32, 66}, // interleave 6 x 16i32 into 96i32 (and store) 5515 5516 {6, MVT::v2i64, 8}, // interleave 6 x 2i64 into 12i64 (and store) 5517 {6, MVT::v4i64, 15}, // interleave 6 x 4i64 into 24i64 (and store) 5518 {6, MVT::v8i64, 30}, // interleave 6 x 8i64 into 48i64 (and store) 5519 }; 5520 5521 static const CostTblEntry SSE2InterleavedStoreTbl[] = { 5522 {2, MVT::v2i8, 1}, // interleave 2 x 2i8 into 4i8 (and store) 5523 {2, MVT::v4i8, 1}, // interleave 2 x 4i8 into 8i8 (and store) 5524 {2, MVT::v8i8, 1}, // interleave 2 x 8i8 into 16i8 (and store) 5525 5526 {2, MVT::v2i16, 1}, // interleave 2 x 2i16 into 4i16 (and store) 5527 {2, MVT::v4i16, 1}, // interleave 2 x 4i16 into 8i16 (and store) 5528 5529 {2, MVT::v2i32, 1}, // interleave 2 x 2i32 into 4i32 (and store) 5530 }; 5531 5532 if (Opcode == Instruction::Load) { 5533 auto GetDiscountedCost = [Factor, NumMembers = Indices.size(), 5534 MemOpCosts](const CostTblEntry *Entry) { 5535 // NOTE: this is just an approximation! 5536 // It can over/under -estimate the cost! 5537 return MemOpCosts + divideCeil(NumMembers * Entry->Cost, Factor); 5538 }; 5539 5540 if (ST->hasAVX2()) 5541 if (const auto *Entry = CostTableLookup(AVX2InterleavedLoadTbl, Factor, 5542 ETy.getSimpleVT())) 5543 return GetDiscountedCost(Entry); 5544 5545 if (ST->hasSSSE3()) 5546 if (const auto *Entry = CostTableLookup(SSSE3InterleavedLoadTbl, Factor, 5547 ETy.getSimpleVT())) 5548 return GetDiscountedCost(Entry); 5549 5550 if (ST->hasSSE2()) 5551 if (const auto *Entry = CostTableLookup(SSE2InterleavedLoadTbl, Factor, 5552 ETy.getSimpleVT())) 5553 return GetDiscountedCost(Entry); 5554 } else { 5555 assert(Opcode == Instruction::Store && 5556 "Expected Store Instruction at this point"); 5557 assert((!Indices.size() || Indices.size() == Factor) && 5558 "Interleaved store only supports fully-interleaved groups."); 5559 if (ST->hasAVX2()) 5560 if (const auto *Entry = CostTableLookup(AVX2InterleavedStoreTbl, Factor, 5561 ETy.getSimpleVT())) 5562 return MemOpCosts + Entry->Cost; 5563 5564 if (ST->hasSSE2()) 5565 if (const auto *Entry = CostTableLookup(SSE2InterleavedStoreTbl, Factor, 5566 ETy.getSimpleVT())) 5567 return MemOpCosts + Entry->Cost; 5568 } 5569 5570 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 5571 Alignment, AddressSpace, CostKind, 5572 UseMaskForCond, UseMaskForGaps); 5573 } 5574